Tue, 20 Dec 2011 16:56:50 +0100
7121140: Allocation paths require explicit memory synchronization operations for RMO systems
Summary: adds store store barrier after initialization of header and body of objects.
Reviewed-by: never, kvn
1 //
2 // Copyright (c) 1997, 2011, Oracle and/or its affiliates. All rights reserved.
3 // DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 //
5 // This code is free software; you can redistribute it and/or modify it
6 // under the terms of the GNU General Public License version 2 only, as
7 // published by the Free Software Foundation.
8 //
9 // This code is distributed in the hope that it will be useful, but WITHOUT
10 // ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 // FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 // version 2 for more details (a copy is included in the LICENSE file that
13 // accompanied this code).
14 //
15 // You should have received a copy of the GNU General Public License version
16 // 2 along with this work; if not, write to the Free Software Foundation,
17 // Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 //
19 // Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 // or visit www.oracle.com if you need additional information or have any
21 // questions.
22 //
23 //
25 // X86 Architecture Description File
27 //----------REGISTER DEFINITION BLOCK------------------------------------------
28 // This information is used by the matcher and the register allocator to
29 // describe individual registers and classes of registers within the target
30 // archtecture.
32 register %{
33 //----------Architecture Description Register Definitions----------------------
34 // General Registers
35 // "reg_def" name ( register save type, C convention save type,
36 // ideal register type, encoding );
37 // Register Save Types:
38 //
39 // NS = No-Save: The register allocator assumes that these registers
40 // can be used without saving upon entry to the method, &
41 // that they do not need to be saved at call sites.
42 //
43 // SOC = Save-On-Call: The register allocator assumes that these registers
44 // can be used without saving upon entry to the method,
45 // but that they must be saved at call sites.
46 //
47 // SOE = Save-On-Entry: The register allocator assumes that these registers
48 // must be saved before using them upon entry to the
49 // method, but they do not need to be saved at call
50 // sites.
51 //
52 // AS = Always-Save: The register allocator assumes that these registers
53 // must be saved before using them upon entry to the
54 // method, & that they must be saved at call sites.
55 //
56 // Ideal Register Type is used to determine how to save & restore a
57 // register. Op_RegI will get spilled with LoadI/StoreI, Op_RegP will get
58 // spilled with LoadP/StoreP. If the register supports both, use Op_RegI.
59 //
60 // The encoding number is the actual bit-pattern placed into the opcodes.
62 // General Registers
63 // Previously set EBX, ESI, and EDI as save-on-entry for java code
64 // Turn off SOE in java-code due to frequent use of uncommon-traps.
65 // Now that allocator is better, turn on ESI and EDI as SOE registers.
67 reg_def EBX(SOC, SOE, Op_RegI, 3, rbx->as_VMReg());
68 reg_def ECX(SOC, SOC, Op_RegI, 1, rcx->as_VMReg());
69 reg_def ESI(SOC, SOE, Op_RegI, 6, rsi->as_VMReg());
70 reg_def EDI(SOC, SOE, Op_RegI, 7, rdi->as_VMReg());
71 // now that adapter frames are gone EBP is always saved and restored by the prolog/epilog code
72 reg_def EBP(NS, SOE, Op_RegI, 5, rbp->as_VMReg());
73 reg_def EDX(SOC, SOC, Op_RegI, 2, rdx->as_VMReg());
74 reg_def EAX(SOC, SOC, Op_RegI, 0, rax->as_VMReg());
75 reg_def ESP( NS, NS, Op_RegI, 4, rsp->as_VMReg());
77 // Special Registers
78 reg_def EFLAGS(SOC, SOC, 0, 8, VMRegImpl::Bad());
80 // Float registers. We treat TOS/FPR0 special. It is invisible to the
81 // allocator, and only shows up in the encodings.
82 reg_def FPR0L( SOC, SOC, Op_RegF, 0, VMRegImpl::Bad());
83 reg_def FPR0H( SOC, SOC, Op_RegF, 0, VMRegImpl::Bad());
84 // Ok so here's the trick FPR1 is really st(0) except in the midst
85 // of emission of assembly for a machnode. During the emission the fpu stack
86 // is pushed making FPR1 == st(1) temporarily. However at any safepoint
87 // the stack will not have this element so FPR1 == st(0) from the
88 // oopMap viewpoint. This same weirdness with numbering causes
89 // instruction encoding to have to play games with the register
90 // encode to correct for this 0/1 issue. See MachSpillCopyNode::implementation
91 // where it does flt->flt moves to see an example
92 //
93 reg_def FPR1L( SOC, SOC, Op_RegF, 1, as_FloatRegister(0)->as_VMReg());
94 reg_def FPR1H( SOC, SOC, Op_RegF, 1, as_FloatRegister(0)->as_VMReg()->next());
95 reg_def FPR2L( SOC, SOC, Op_RegF, 2, as_FloatRegister(1)->as_VMReg());
96 reg_def FPR2H( SOC, SOC, Op_RegF, 2, as_FloatRegister(1)->as_VMReg()->next());
97 reg_def FPR3L( SOC, SOC, Op_RegF, 3, as_FloatRegister(2)->as_VMReg());
98 reg_def FPR3H( SOC, SOC, Op_RegF, 3, as_FloatRegister(2)->as_VMReg()->next());
99 reg_def FPR4L( SOC, SOC, Op_RegF, 4, as_FloatRegister(3)->as_VMReg());
100 reg_def FPR4H( SOC, SOC, Op_RegF, 4, as_FloatRegister(3)->as_VMReg()->next());
101 reg_def FPR5L( SOC, SOC, Op_RegF, 5, as_FloatRegister(4)->as_VMReg());
102 reg_def FPR5H( SOC, SOC, Op_RegF, 5, as_FloatRegister(4)->as_VMReg()->next());
103 reg_def FPR6L( SOC, SOC, Op_RegF, 6, as_FloatRegister(5)->as_VMReg());
104 reg_def FPR6H( SOC, SOC, Op_RegF, 6, as_FloatRegister(5)->as_VMReg()->next());
105 reg_def FPR7L( SOC, SOC, Op_RegF, 7, as_FloatRegister(6)->as_VMReg());
106 reg_def FPR7H( SOC, SOC, Op_RegF, 7, as_FloatRegister(6)->as_VMReg()->next());
108 // XMM registers. 128-bit registers or 4 words each, labeled a-d.
109 // Word a in each register holds a Float, words ab hold a Double.
110 // We currently do not use the SIMD capabilities, so registers cd
111 // are unused at the moment.
112 reg_def XMM0a( SOC, SOC, Op_RegF, 0, xmm0->as_VMReg());
113 reg_def XMM0b( SOC, SOC, Op_RegF, 0, xmm0->as_VMReg()->next());
114 reg_def XMM1a( SOC, SOC, Op_RegF, 1, xmm1->as_VMReg());
115 reg_def XMM1b( SOC, SOC, Op_RegF, 1, xmm1->as_VMReg()->next());
116 reg_def XMM2a( SOC, SOC, Op_RegF, 2, xmm2->as_VMReg());
117 reg_def XMM2b( SOC, SOC, Op_RegF, 2, xmm2->as_VMReg()->next());
118 reg_def XMM3a( SOC, SOC, Op_RegF, 3, xmm3->as_VMReg());
119 reg_def XMM3b( SOC, SOC, Op_RegF, 3, xmm3->as_VMReg()->next());
120 reg_def XMM4a( SOC, SOC, Op_RegF, 4, xmm4->as_VMReg());
121 reg_def XMM4b( SOC, SOC, Op_RegF, 4, xmm4->as_VMReg()->next());
122 reg_def XMM5a( SOC, SOC, Op_RegF, 5, xmm5->as_VMReg());
123 reg_def XMM5b( SOC, SOC, Op_RegF, 5, xmm5->as_VMReg()->next());
124 reg_def XMM6a( SOC, SOC, Op_RegF, 6, xmm6->as_VMReg());
125 reg_def XMM6b( SOC, SOC, Op_RegF, 6, xmm6->as_VMReg()->next());
126 reg_def XMM7a( SOC, SOC, Op_RegF, 7, xmm7->as_VMReg());
127 reg_def XMM7b( SOC, SOC, Op_RegF, 7, xmm7->as_VMReg()->next());
129 // Specify priority of register selection within phases of register
130 // allocation. Highest priority is first. A useful heuristic is to
131 // give registers a low priority when they are required by machine
132 // instructions, like EAX and EDX. Registers which are used as
133 // pairs must fall on an even boundary (witness the FPR#L's in this list).
134 // For the Intel integer registers, the equivalent Long pairs are
135 // EDX:EAX, EBX:ECX, and EDI:EBP.
136 alloc_class chunk0( ECX, EBX, EBP, EDI, EAX, EDX, ESI, ESP,
137 FPR0L, FPR0H, FPR1L, FPR1H, FPR2L, FPR2H,
138 FPR3L, FPR3H, FPR4L, FPR4H, FPR5L, FPR5H,
139 FPR6L, FPR6H, FPR7L, FPR7H );
141 alloc_class chunk1( XMM0a, XMM0b,
142 XMM1a, XMM1b,
143 XMM2a, XMM2b,
144 XMM3a, XMM3b,
145 XMM4a, XMM4b,
146 XMM5a, XMM5b,
147 XMM6a, XMM6b,
148 XMM7a, XMM7b, EFLAGS);
151 //----------Architecture Description Register Classes--------------------------
152 // Several register classes are automatically defined based upon information in
153 // this architecture description.
154 // 1) reg_class inline_cache_reg ( /* as def'd in frame section */ )
155 // 2) reg_class compiler_method_oop_reg ( /* as def'd in frame section */ )
156 // 2) reg_class interpreter_method_oop_reg ( /* as def'd in frame section */ )
157 // 3) reg_class stack_slots( /* one chunk of stack-based "registers" */ )
158 //
159 // Class for all registers
160 reg_class any_reg(EAX, EDX, EBP, EDI, ESI, ECX, EBX, ESP);
161 // Class for general registers
162 reg_class e_reg(EAX, EDX, EBP, EDI, ESI, ECX, EBX);
163 // Class for general registers which may be used for implicit null checks on win95
164 // Also safe for use by tailjump. We don't want to allocate in rbp,
165 reg_class e_reg_no_rbp(EAX, EDX, EDI, ESI, ECX, EBX);
166 // Class of "X" registers
167 reg_class x_reg(EBX, ECX, EDX, EAX);
168 // Class of registers that can appear in an address with no offset.
169 // EBP and ESP require an extra instruction byte for zero offset.
170 // Used in fast-unlock
171 reg_class p_reg(EDX, EDI, ESI, EBX);
172 // Class for general registers not including ECX
173 reg_class ncx_reg(EAX, EDX, EBP, EDI, ESI, EBX);
174 // Class for general registers not including EAX
175 reg_class nax_reg(EDX, EDI, ESI, ECX, EBX);
176 // Class for general registers not including EAX or EBX.
177 reg_class nabx_reg(EDX, EDI, ESI, ECX, EBP);
178 // Class of EAX (for multiply and divide operations)
179 reg_class eax_reg(EAX);
180 // Class of EBX (for atomic add)
181 reg_class ebx_reg(EBX);
182 // Class of ECX (for shift and JCXZ operations and cmpLTMask)
183 reg_class ecx_reg(ECX);
184 // Class of EDX (for multiply and divide operations)
185 reg_class edx_reg(EDX);
186 // Class of EDI (for synchronization)
187 reg_class edi_reg(EDI);
188 // Class of ESI (for synchronization)
189 reg_class esi_reg(ESI);
190 // Singleton class for interpreter's stack pointer
191 reg_class ebp_reg(EBP);
192 // Singleton class for stack pointer
193 reg_class sp_reg(ESP);
194 // Singleton class for instruction pointer
195 // reg_class ip_reg(EIP);
196 // Singleton class for condition codes
197 reg_class int_flags(EFLAGS);
198 // Class of integer register pairs
199 reg_class long_reg( EAX,EDX, ECX,EBX, EBP,EDI );
200 // Class of integer register pairs that aligns with calling convention
201 reg_class eadx_reg( EAX,EDX );
202 reg_class ebcx_reg( ECX,EBX );
203 // Not AX or DX, used in divides
204 reg_class nadx_reg( EBX,ECX,ESI,EDI,EBP );
206 // Floating point registers. Notice FPR0 is not a choice.
207 // FPR0 is not ever allocated; we use clever encodings to fake
208 // a 2-address instructions out of Intels FP stack.
209 reg_class flt_reg( FPR1L,FPR2L,FPR3L,FPR4L,FPR5L,FPR6L,FPR7L );
211 // make a register class for SSE registers
212 reg_class xmm_reg(XMM0a, XMM1a, XMM2a, XMM3a, XMM4a, XMM5a, XMM6a, XMM7a);
214 // make a double register class for SSE2 registers
215 reg_class xdb_reg(XMM0a,XMM0b, XMM1a,XMM1b, XMM2a,XMM2b, XMM3a,XMM3b,
216 XMM4a,XMM4b, XMM5a,XMM5b, XMM6a,XMM6b, XMM7a,XMM7b );
218 reg_class dbl_reg( FPR1L,FPR1H, FPR2L,FPR2H, FPR3L,FPR3H,
219 FPR4L,FPR4H, FPR5L,FPR5H, FPR6L,FPR6H,
220 FPR7L,FPR7H );
222 reg_class flt_reg0( FPR1L );
223 reg_class dbl_reg0( FPR1L,FPR1H );
224 reg_class dbl_reg1( FPR2L,FPR2H );
225 reg_class dbl_notreg0( FPR2L,FPR2H, FPR3L,FPR3H, FPR4L,FPR4H,
226 FPR5L,FPR5H, FPR6L,FPR6H, FPR7L,FPR7H );
228 // XMM6 and XMM7 could be used as temporary registers for long, float and
229 // double values for SSE2.
230 reg_class xdb_reg6( XMM6a,XMM6b );
231 reg_class xdb_reg7( XMM7a,XMM7b );
232 %}
235 //----------SOURCE BLOCK-------------------------------------------------------
236 // This is a block of C++ code which provides values, functions, and
237 // definitions necessary in the rest of the architecture description
238 source_hpp %{
239 // Must be visible to the DFA in dfa_x86_32.cpp
240 extern bool is_operand_hi32_zero(Node* n);
241 %}
243 source %{
244 #define RELOC_IMM32 Assembler::imm_operand
245 #define RELOC_DISP32 Assembler::disp32_operand
247 #define __ _masm.
249 // How to find the high register of a Long pair, given the low register
250 #define HIGH_FROM_LOW(x) ((x)+2)
252 // These masks are used to provide 128-bit aligned bitmasks to the XMM
253 // instructions, to allow sign-masking or sign-bit flipping. They allow
254 // fast versions of NegF/NegD and AbsF/AbsD.
256 // Note: 'double' and 'long long' have 32-bits alignment on x86.
257 static jlong* double_quadword(jlong *adr, jlong lo, jlong hi) {
258 // Use the expression (adr)&(~0xF) to provide 128-bits aligned address
259 // of 128-bits operands for SSE instructions.
260 jlong *operand = (jlong*)(((uintptr_t)adr)&((uintptr_t)(~0xF)));
261 // Store the value to a 128-bits operand.
262 operand[0] = lo;
263 operand[1] = hi;
264 return operand;
265 }
267 // Buffer for 128-bits masks used by SSE instructions.
268 static jlong fp_signmask_pool[(4+1)*2]; // 4*128bits(data) + 128bits(alignment)
270 // Static initialization during VM startup.
271 static jlong *float_signmask_pool = double_quadword(&fp_signmask_pool[1*2], CONST64(0x7FFFFFFF7FFFFFFF), CONST64(0x7FFFFFFF7FFFFFFF));
272 static jlong *double_signmask_pool = double_quadword(&fp_signmask_pool[2*2], CONST64(0x7FFFFFFFFFFFFFFF), CONST64(0x7FFFFFFFFFFFFFFF));
273 static jlong *float_signflip_pool = double_quadword(&fp_signmask_pool[3*2], CONST64(0x8000000080000000), CONST64(0x8000000080000000));
274 static jlong *double_signflip_pool = double_quadword(&fp_signmask_pool[4*2], CONST64(0x8000000000000000), CONST64(0x8000000000000000));
276 // Offset hacking within calls.
277 static int pre_call_FPU_size() {
278 if (Compile::current()->in_24_bit_fp_mode())
279 return 6; // fldcw
280 return 0;
281 }
283 static int preserve_SP_size() {
284 return 2; // op, rm(reg/reg)
285 }
287 // !!!!! Special hack to get all type of calls to specify the byte offset
288 // from the start of the call to the point where the return address
289 // will point.
290 int MachCallStaticJavaNode::ret_addr_offset() {
291 int offset = 5 + pre_call_FPU_size(); // 5 bytes from start of call to where return address points
292 if (_method_handle_invoke)
293 offset += preserve_SP_size();
294 return offset;
295 }
297 int MachCallDynamicJavaNode::ret_addr_offset() {
298 return 10 + pre_call_FPU_size(); // 10 bytes from start of call to where return address points
299 }
301 static int sizeof_FFree_Float_Stack_All = -1;
303 int MachCallRuntimeNode::ret_addr_offset() {
304 assert(sizeof_FFree_Float_Stack_All != -1, "must have been emitted already");
305 return sizeof_FFree_Float_Stack_All + 5 + pre_call_FPU_size();
306 }
308 // Indicate if the safepoint node needs the polling page as an input.
309 // Since x86 does have absolute addressing, it doesn't.
310 bool SafePointNode::needs_polling_address_input() {
311 return false;
312 }
314 //
315 // Compute padding required for nodes which need alignment
316 //
318 // The address of the call instruction needs to be 4-byte aligned to
319 // ensure that it does not span a cache line so that it can be patched.
320 int CallStaticJavaDirectNode::compute_padding(int current_offset) const {
321 current_offset += pre_call_FPU_size(); // skip fldcw, if any
322 current_offset += 1; // skip call opcode byte
323 return round_to(current_offset, alignment_required()) - current_offset;
324 }
326 // The address of the call instruction needs to be 4-byte aligned to
327 // ensure that it does not span a cache line so that it can be patched.
328 int CallStaticJavaHandleNode::compute_padding(int current_offset) const {
329 current_offset += pre_call_FPU_size(); // skip fldcw, if any
330 current_offset += preserve_SP_size(); // skip mov rbp, rsp
331 current_offset += 1; // skip call opcode byte
332 return round_to(current_offset, alignment_required()) - current_offset;
333 }
335 // The address of the call instruction needs to be 4-byte aligned to
336 // ensure that it does not span a cache line so that it can be patched.
337 int CallDynamicJavaDirectNode::compute_padding(int current_offset) const {
338 current_offset += pre_call_FPU_size(); // skip fldcw, if any
339 current_offset += 5; // skip MOV instruction
340 current_offset += 1; // skip call opcode byte
341 return round_to(current_offset, alignment_required()) - current_offset;
342 }
344 #ifndef PRODUCT
345 void MachBreakpointNode::format( PhaseRegAlloc *, outputStream* st ) const {
346 st->print("INT3");
347 }
348 #endif
350 // EMIT_RM()
351 void emit_rm(CodeBuffer &cbuf, int f1, int f2, int f3) {
352 unsigned char c = (unsigned char)((f1 << 6) | (f2 << 3) | f3);
353 cbuf.insts()->emit_int8(c);
354 }
356 // EMIT_CC()
357 void emit_cc(CodeBuffer &cbuf, int f1, int f2) {
358 unsigned char c = (unsigned char)( f1 | f2 );
359 cbuf.insts()->emit_int8(c);
360 }
362 // EMIT_OPCODE()
363 void emit_opcode(CodeBuffer &cbuf, int code) {
364 cbuf.insts()->emit_int8((unsigned char) code);
365 }
367 // EMIT_OPCODE() w/ relocation information
368 void emit_opcode(CodeBuffer &cbuf, int code, relocInfo::relocType reloc, int offset = 0) {
369 cbuf.relocate(cbuf.insts_mark() + offset, reloc);
370 emit_opcode(cbuf, code);
371 }
373 // EMIT_D8()
374 void emit_d8(CodeBuffer &cbuf, int d8) {
375 cbuf.insts()->emit_int8((unsigned char) d8);
376 }
378 // EMIT_D16()
379 void emit_d16(CodeBuffer &cbuf, int d16) {
380 cbuf.insts()->emit_int16(d16);
381 }
383 // EMIT_D32()
384 void emit_d32(CodeBuffer &cbuf, int d32) {
385 cbuf.insts()->emit_int32(d32);
386 }
388 // emit 32 bit value and construct relocation entry from relocInfo::relocType
389 void emit_d32_reloc(CodeBuffer &cbuf, int d32, relocInfo::relocType reloc,
390 int format) {
391 cbuf.relocate(cbuf.insts_mark(), reloc, format);
392 cbuf.insts()->emit_int32(d32);
393 }
395 // emit 32 bit value and construct relocation entry from RelocationHolder
396 void emit_d32_reloc(CodeBuffer &cbuf, int d32, RelocationHolder const& rspec,
397 int format) {
398 #ifdef ASSERT
399 if (rspec.reloc()->type() == relocInfo::oop_type && d32 != 0 && d32 != (int)Universe::non_oop_word()) {
400 assert(oop(d32)->is_oop() && (ScavengeRootsInCode || !oop(d32)->is_scavengable()), "cannot embed scavengable oops in code");
401 }
402 #endif
403 cbuf.relocate(cbuf.insts_mark(), rspec, format);
404 cbuf.insts()->emit_int32(d32);
405 }
407 // Access stack slot for load or store
408 void store_to_stackslot(CodeBuffer &cbuf, int opcode, int rm_field, int disp) {
409 emit_opcode( cbuf, opcode ); // (e.g., FILD [ESP+src])
410 if( -128 <= disp && disp <= 127 ) {
411 emit_rm( cbuf, 0x01, rm_field, ESP_enc ); // R/M byte
412 emit_rm( cbuf, 0x00, ESP_enc, ESP_enc); // SIB byte
413 emit_d8 (cbuf, disp); // Displacement // R/M byte
414 } else {
415 emit_rm( cbuf, 0x02, rm_field, ESP_enc ); // R/M byte
416 emit_rm( cbuf, 0x00, ESP_enc, ESP_enc); // SIB byte
417 emit_d32(cbuf, disp); // Displacement // R/M byte
418 }
419 }
421 // eRegI ereg, memory mem) %{ // emit_reg_mem
422 void encode_RegMem( CodeBuffer &cbuf, int reg_encoding, int base, int index, int scale, int displace, bool displace_is_oop ) {
423 // There is no index & no scale, use form without SIB byte
424 if ((index == 0x4) &&
425 (scale == 0) && (base != ESP_enc)) {
426 // If no displacement, mode is 0x0; unless base is [EBP]
427 if ( (displace == 0) && (base != EBP_enc) ) {
428 emit_rm(cbuf, 0x0, reg_encoding, base);
429 }
430 else { // If 8-bit displacement, mode 0x1
431 if ((displace >= -128) && (displace <= 127)
432 && !(displace_is_oop) ) {
433 emit_rm(cbuf, 0x1, reg_encoding, base);
434 emit_d8(cbuf, displace);
435 }
436 else { // If 32-bit displacement
437 if (base == -1) { // Special flag for absolute address
438 emit_rm(cbuf, 0x0, reg_encoding, 0x5);
439 // (manual lies; no SIB needed here)
440 if ( displace_is_oop ) {
441 emit_d32_reloc(cbuf, displace, relocInfo::oop_type, 1);
442 } else {
443 emit_d32 (cbuf, displace);
444 }
445 }
446 else { // Normal base + offset
447 emit_rm(cbuf, 0x2, reg_encoding, base);
448 if ( displace_is_oop ) {
449 emit_d32_reloc(cbuf, displace, relocInfo::oop_type, 1);
450 } else {
451 emit_d32 (cbuf, displace);
452 }
453 }
454 }
455 }
456 }
457 else { // Else, encode with the SIB byte
458 // If no displacement, mode is 0x0; unless base is [EBP]
459 if (displace == 0 && (base != EBP_enc)) { // If no displacement
460 emit_rm(cbuf, 0x0, reg_encoding, 0x4);
461 emit_rm(cbuf, scale, index, base);
462 }
463 else { // If 8-bit displacement, mode 0x1
464 if ((displace >= -128) && (displace <= 127)
465 && !(displace_is_oop) ) {
466 emit_rm(cbuf, 0x1, reg_encoding, 0x4);
467 emit_rm(cbuf, scale, index, base);
468 emit_d8(cbuf, displace);
469 }
470 else { // If 32-bit displacement
471 if (base == 0x04 ) {
472 emit_rm(cbuf, 0x2, reg_encoding, 0x4);
473 emit_rm(cbuf, scale, index, 0x04);
474 } else {
475 emit_rm(cbuf, 0x2, reg_encoding, 0x4);
476 emit_rm(cbuf, scale, index, base);
477 }
478 if ( displace_is_oop ) {
479 emit_d32_reloc(cbuf, displace, relocInfo::oop_type, 1);
480 } else {
481 emit_d32 (cbuf, displace);
482 }
483 }
484 }
485 }
486 }
489 void encode_Copy( CodeBuffer &cbuf, int dst_encoding, int src_encoding ) {
490 if( dst_encoding == src_encoding ) {
491 // reg-reg copy, use an empty encoding
492 } else {
493 emit_opcode( cbuf, 0x8B );
494 emit_rm(cbuf, 0x3, dst_encoding, src_encoding );
495 }
496 }
498 void emit_cmpfp_fixup(MacroAssembler& _masm) {
499 Label exit;
500 __ jccb(Assembler::noParity, exit);
501 __ pushf();
502 //
503 // comiss/ucomiss instructions set ZF,PF,CF flags and
504 // zero OF,AF,SF for NaN values.
505 // Fixup flags by zeroing ZF,PF so that compare of NaN
506 // values returns 'less than' result (CF is set).
507 // Leave the rest of flags unchanged.
508 //
509 // 7 6 5 4 3 2 1 0
510 // |S|Z|r|A|r|P|r|C| (r - reserved bit)
511 // 0 0 1 0 1 0 1 1 (0x2B)
512 //
513 __ andl(Address(rsp, 0), 0xffffff2b);
514 __ popf();
515 __ bind(exit);
516 }
518 void emit_cmpfp3(MacroAssembler& _masm, Register dst) {
519 Label done;
520 __ movl(dst, -1);
521 __ jcc(Assembler::parity, done);
522 __ jcc(Assembler::below, done);
523 __ setb(Assembler::notEqual, dst);
524 __ movzbl(dst, dst);
525 __ bind(done);
526 }
529 //=============================================================================
530 const RegMask& MachConstantBaseNode::_out_RegMask = RegMask::Empty;
532 int Compile::ConstantTable::calculate_table_base_offset() const {
533 return 0; // absolute addressing, no offset
534 }
536 void MachConstantBaseNode::emit(CodeBuffer& cbuf, PhaseRegAlloc* ra_) const {
537 // Empty encoding
538 }
540 uint MachConstantBaseNode::size(PhaseRegAlloc* ra_) const {
541 return 0;
542 }
544 #ifndef PRODUCT
545 void MachConstantBaseNode::format(PhaseRegAlloc* ra_, outputStream* st) const {
546 st->print("# MachConstantBaseNode (empty encoding)");
547 }
548 #endif
551 //=============================================================================
552 #ifndef PRODUCT
553 void MachPrologNode::format( PhaseRegAlloc *ra_, outputStream* st ) const {
554 Compile* C = ra_->C;
555 if( C->in_24_bit_fp_mode() ) {
556 st->print("FLDCW 24 bit fpu control word");
557 st->print_cr(""); st->print("\t");
558 }
560 int framesize = C->frame_slots() << LogBytesPerInt;
561 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
562 // Remove two words for return addr and rbp,
563 framesize -= 2*wordSize;
565 // Calls to C2R adapters often do not accept exceptional returns.
566 // We require that their callers must bang for them. But be careful, because
567 // some VM calls (such as call site linkage) can use several kilobytes of
568 // stack. But the stack safety zone should account for that.
569 // See bugs 4446381, 4468289, 4497237.
570 if (C->need_stack_bang(framesize)) {
571 st->print_cr("# stack bang"); st->print("\t");
572 }
573 st->print_cr("PUSHL EBP"); st->print("\t");
575 if( VerifyStackAtCalls ) { // Majik cookie to verify stack depth
576 st->print("PUSH 0xBADB100D\t# Majik cookie for stack depth check");
577 st->print_cr(""); st->print("\t");
578 framesize -= wordSize;
579 }
581 if ((C->in_24_bit_fp_mode() || VerifyStackAtCalls ) && framesize < 128 ) {
582 if (framesize) {
583 st->print("SUB ESP,%d\t# Create frame",framesize);
584 }
585 } else {
586 st->print("SUB ESP,%d\t# Create frame",framesize);
587 }
588 }
589 #endif
592 void MachPrologNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
593 Compile* C = ra_->C;
595 if (UseSSE >= 2 && VerifyFPU) {
596 MacroAssembler masm(&cbuf);
597 masm.verify_FPU(0, "FPU stack must be clean on entry");
598 }
600 // WARNING: Initial instruction MUST be 5 bytes or longer so that
601 // NativeJump::patch_verified_entry will be able to patch out the entry
602 // code safely. The fldcw is ok at 6 bytes, the push to verify stack
603 // depth is ok at 5 bytes, the frame allocation can be either 3 or
604 // 6 bytes. So if we don't do the fldcw or the push then we must
605 // use the 6 byte frame allocation even if we have no frame. :-(
606 // If method sets FPU control word do it now
607 if( C->in_24_bit_fp_mode() ) {
608 MacroAssembler masm(&cbuf);
609 masm.fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
610 }
612 int framesize = C->frame_slots() << LogBytesPerInt;
613 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
614 // Remove two words for return addr and rbp,
615 framesize -= 2*wordSize;
617 // Calls to C2R adapters often do not accept exceptional returns.
618 // We require that their callers must bang for them. But be careful, because
619 // some VM calls (such as call site linkage) can use several kilobytes of
620 // stack. But the stack safety zone should account for that.
621 // See bugs 4446381, 4468289, 4497237.
622 if (C->need_stack_bang(framesize)) {
623 MacroAssembler masm(&cbuf);
624 masm.generate_stack_overflow_check(framesize);
625 }
627 // We always push rbp, so that on return to interpreter rbp, will be
628 // restored correctly and we can correct the stack.
629 emit_opcode(cbuf, 0x50 | EBP_enc);
631 if( VerifyStackAtCalls ) { // Majik cookie to verify stack depth
632 emit_opcode(cbuf, 0x68); // push 0xbadb100d
633 emit_d32(cbuf, 0xbadb100d);
634 framesize -= wordSize;
635 }
637 if ((C->in_24_bit_fp_mode() || VerifyStackAtCalls ) && framesize < 128 ) {
638 if (framesize) {
639 emit_opcode(cbuf, 0x83); // sub SP,#framesize
640 emit_rm(cbuf, 0x3, 0x05, ESP_enc);
641 emit_d8(cbuf, framesize);
642 }
643 } else {
644 emit_opcode(cbuf, 0x81); // sub SP,#framesize
645 emit_rm(cbuf, 0x3, 0x05, ESP_enc);
646 emit_d32(cbuf, framesize);
647 }
648 C->set_frame_complete(cbuf.insts_size());
650 #ifdef ASSERT
651 if (VerifyStackAtCalls) {
652 Label L;
653 MacroAssembler masm(&cbuf);
654 masm.push(rax);
655 masm.mov(rax, rsp);
656 masm.andptr(rax, StackAlignmentInBytes-1);
657 masm.cmpptr(rax, StackAlignmentInBytes-wordSize);
658 masm.pop(rax);
659 masm.jcc(Assembler::equal, L);
660 masm.stop("Stack is not properly aligned!");
661 masm.bind(L);
662 }
663 #endif
665 if (C->has_mach_constant_base_node()) {
666 // NOTE: We set the table base offset here because users might be
667 // emitted before MachConstantBaseNode.
668 Compile::ConstantTable& constant_table = C->constant_table();
669 constant_table.set_table_base_offset(constant_table.calculate_table_base_offset());
670 }
671 }
673 uint MachPrologNode::size(PhaseRegAlloc *ra_) const {
674 return MachNode::size(ra_); // too many variables; just compute it the hard way
675 }
677 int MachPrologNode::reloc() const {
678 return 0; // a large enough number
679 }
681 //=============================================================================
682 #ifndef PRODUCT
683 void MachEpilogNode::format( PhaseRegAlloc *ra_, outputStream* st ) const {
684 Compile *C = ra_->C;
685 int framesize = C->frame_slots() << LogBytesPerInt;
686 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
687 // Remove two words for return addr and rbp,
688 framesize -= 2*wordSize;
690 if( C->in_24_bit_fp_mode() ) {
691 st->print("FLDCW standard control word");
692 st->cr(); st->print("\t");
693 }
694 if( framesize ) {
695 st->print("ADD ESP,%d\t# Destroy frame",framesize);
696 st->cr(); st->print("\t");
697 }
698 st->print_cr("POPL EBP"); st->print("\t");
699 if( do_polling() && C->is_method_compilation() ) {
700 st->print("TEST PollPage,EAX\t! Poll Safepoint");
701 st->cr(); st->print("\t");
702 }
703 }
704 #endif
706 void MachEpilogNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
707 Compile *C = ra_->C;
709 // If method set FPU control word, restore to standard control word
710 if( C->in_24_bit_fp_mode() ) {
711 MacroAssembler masm(&cbuf);
712 masm.fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
713 }
715 int framesize = C->frame_slots() << LogBytesPerInt;
716 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
717 // Remove two words for return addr and rbp,
718 framesize -= 2*wordSize;
720 // Note that VerifyStackAtCalls' Majik cookie does not change the frame size popped here
722 if( framesize >= 128 ) {
723 emit_opcode(cbuf, 0x81); // add SP, #framesize
724 emit_rm(cbuf, 0x3, 0x00, ESP_enc);
725 emit_d32(cbuf, framesize);
726 }
727 else if( framesize ) {
728 emit_opcode(cbuf, 0x83); // add SP, #framesize
729 emit_rm(cbuf, 0x3, 0x00, ESP_enc);
730 emit_d8(cbuf, framesize);
731 }
733 emit_opcode(cbuf, 0x58 | EBP_enc);
735 if( do_polling() && C->is_method_compilation() ) {
736 cbuf.relocate(cbuf.insts_end(), relocInfo::poll_return_type, 0);
737 emit_opcode(cbuf,0x85);
738 emit_rm(cbuf, 0x0, EAX_enc, 0x5); // EAX
739 emit_d32(cbuf, (intptr_t)os::get_polling_page());
740 }
741 }
743 uint MachEpilogNode::size(PhaseRegAlloc *ra_) const {
744 Compile *C = ra_->C;
745 // If method set FPU control word, restore to standard control word
746 int size = C->in_24_bit_fp_mode() ? 6 : 0;
747 if( do_polling() && C->is_method_compilation() ) size += 6;
749 int framesize = C->frame_slots() << LogBytesPerInt;
750 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
751 // Remove two words for return addr and rbp,
752 framesize -= 2*wordSize;
754 size++; // popl rbp,
756 if( framesize >= 128 ) {
757 size += 6;
758 } else {
759 size += framesize ? 3 : 0;
760 }
761 return size;
762 }
764 int MachEpilogNode::reloc() const {
765 return 0; // a large enough number
766 }
768 const Pipeline * MachEpilogNode::pipeline() const {
769 return MachNode::pipeline_class();
770 }
772 int MachEpilogNode::safepoint_offset() const { return 0; }
774 //=============================================================================
776 enum RC { rc_bad, rc_int, rc_float, rc_xmm, rc_stack };
777 static enum RC rc_class( OptoReg::Name reg ) {
779 if( !OptoReg::is_valid(reg) ) return rc_bad;
780 if (OptoReg::is_stack(reg)) return rc_stack;
782 VMReg r = OptoReg::as_VMReg(reg);
783 if (r->is_Register()) return rc_int;
784 if (r->is_FloatRegister()) {
785 assert(UseSSE < 2, "shouldn't be used in SSE2+ mode");
786 return rc_float;
787 }
788 assert(r->is_XMMRegister(), "must be");
789 return rc_xmm;
790 }
792 static int impl_helper( CodeBuffer *cbuf, bool do_size, bool is_load, int offset, int reg,
793 int opcode, const char *op_str, int size, outputStream* st ) {
794 if( cbuf ) {
795 emit_opcode (*cbuf, opcode );
796 encode_RegMem(*cbuf, Matcher::_regEncode[reg], ESP_enc, 0x4, 0, offset, false);
797 #ifndef PRODUCT
798 } else if( !do_size ) {
799 if( size != 0 ) st->print("\n\t");
800 if( opcode == 0x8B || opcode == 0x89 ) { // MOV
801 if( is_load ) st->print("%s %s,[ESP + #%d]",op_str,Matcher::regName[reg],offset);
802 else st->print("%s [ESP + #%d],%s",op_str,offset,Matcher::regName[reg]);
803 } else { // FLD, FST, PUSH, POP
804 st->print("%s [ESP + #%d]",op_str,offset);
805 }
806 #endif
807 }
808 int offset_size = (offset == 0) ? 0 : ((offset <= 127) ? 1 : 4);
809 return size+3+offset_size;
810 }
812 // Helper for XMM registers. Extra opcode bits, limited syntax.
813 static int impl_x_helper( CodeBuffer *cbuf, bool do_size, bool is_load,
814 int offset, int reg_lo, int reg_hi, int size, outputStream* st ) {
815 if (cbuf) {
816 MacroAssembler _masm(cbuf);
817 if (reg_lo+1 == reg_hi) { // double move?
818 if (is_load) {
819 __ movdbl(as_XMMRegister(Matcher::_regEncode[reg_lo]), Address(rsp, offset));
820 } else {
821 __ movdbl(Address(rsp, offset), as_XMMRegister(Matcher::_regEncode[reg_lo]));
822 }
823 } else {
824 if (is_load) {
825 __ movflt(as_XMMRegister(Matcher::_regEncode[reg_lo]), Address(rsp, offset));
826 } else {
827 __ movflt(Address(rsp, offset), as_XMMRegister(Matcher::_regEncode[reg_lo]));
828 }
829 }
830 #ifndef PRODUCT
831 } else if (!do_size) {
832 if (size != 0) st->print("\n\t");
833 if (reg_lo+1 == reg_hi) { // double move?
834 if (is_load) st->print("%s %s,[ESP + #%d]",
835 UseXmmLoadAndClearUpper ? "MOVSD " : "MOVLPD",
836 Matcher::regName[reg_lo], offset);
837 else st->print("MOVSD [ESP + #%d],%s",
838 offset, Matcher::regName[reg_lo]);
839 } else {
840 if (is_load) st->print("MOVSS %s,[ESP + #%d]",
841 Matcher::regName[reg_lo], offset);
842 else st->print("MOVSS [ESP + #%d],%s",
843 offset, Matcher::regName[reg_lo]);
844 }
845 #endif
846 }
847 int offset_size = (offset == 0) ? 0 : ((offset <= 127) ? 1 : 4);
848 // VEX_2bytes prefix is used if UseAVX > 0, so it takes the same 2 bytes.
849 return size+5+offset_size;
850 }
853 static int impl_movx_helper( CodeBuffer *cbuf, bool do_size, int src_lo, int dst_lo,
854 int src_hi, int dst_hi, int size, outputStream* st ) {
855 if (cbuf) {
856 MacroAssembler _masm(cbuf);
857 if (src_lo+1 == src_hi && dst_lo+1 == dst_hi) { // double move?
858 __ movdbl(as_XMMRegister(Matcher::_regEncode[dst_lo]),
859 as_XMMRegister(Matcher::_regEncode[src_lo]));
860 } else {
861 __ movflt(as_XMMRegister(Matcher::_regEncode[dst_lo]),
862 as_XMMRegister(Matcher::_regEncode[src_lo]));
863 }
864 #ifndef PRODUCT
865 } else if (!do_size) {
866 if (size != 0) st->print("\n\t");
867 if (UseXmmRegToRegMoveAll) {//Use movaps,movapd to move between xmm registers
868 if (src_lo+1 == src_hi && dst_lo+1 == dst_hi) { // double move?
869 st->print("MOVAPD %s,%s",Matcher::regName[dst_lo],Matcher::regName[src_lo]);
870 } else {
871 st->print("MOVAPS %s,%s",Matcher::regName[dst_lo],Matcher::regName[src_lo]);
872 }
873 } else {
874 if( src_lo+1 == src_hi && dst_lo+1 == dst_hi ) { // double move?
875 st->print("MOVSD %s,%s",Matcher::regName[dst_lo],Matcher::regName[src_lo]);
876 } else {
877 st->print("MOVSS %s,%s",Matcher::regName[dst_lo],Matcher::regName[src_lo]);
878 }
879 }
880 #endif
881 }
882 // VEX_2bytes prefix is used if UseAVX > 0, and it takes the same 2 bytes.
883 // Only MOVAPS SSE prefix uses 1 byte.
884 int sz = 4;
885 if (!(src_lo+1 == src_hi && dst_lo+1 == dst_hi) &&
886 UseXmmRegToRegMoveAll && (UseAVX == 0)) sz = 3;
887 return size + sz;
888 }
890 static int impl_movgpr2x_helper( CodeBuffer *cbuf, bool do_size, int src_lo, int dst_lo,
891 int src_hi, int dst_hi, int size, outputStream* st ) {
892 // 32-bit
893 if (cbuf) {
894 MacroAssembler _masm(cbuf);
895 __ movdl(as_XMMRegister(Matcher::_regEncode[dst_lo]),
896 as_Register(Matcher::_regEncode[src_lo]));
897 #ifndef PRODUCT
898 } else if (!do_size) {
899 st->print("movdl %s, %s\t# spill", Matcher::regName[dst_lo], Matcher::regName[src_lo]);
900 #endif
901 }
902 return 4;
903 }
906 static int impl_movx2gpr_helper( CodeBuffer *cbuf, bool do_size, int src_lo, int dst_lo,
907 int src_hi, int dst_hi, int size, outputStream* st ) {
908 // 32-bit
909 if (cbuf) {
910 MacroAssembler _masm(cbuf);
911 __ movdl(as_Register(Matcher::_regEncode[dst_lo]),
912 as_XMMRegister(Matcher::_regEncode[src_lo]));
913 #ifndef PRODUCT
914 } else if (!do_size) {
915 st->print("movdl %s, %s\t# spill", Matcher::regName[dst_lo], Matcher::regName[src_lo]);
916 #endif
917 }
918 return 4;
919 }
921 static int impl_mov_helper( CodeBuffer *cbuf, bool do_size, int src, int dst, int size, outputStream* st ) {
922 if( cbuf ) {
923 emit_opcode(*cbuf, 0x8B );
924 emit_rm (*cbuf, 0x3, Matcher::_regEncode[dst], Matcher::_regEncode[src] );
925 #ifndef PRODUCT
926 } else if( !do_size ) {
927 if( size != 0 ) st->print("\n\t");
928 st->print("MOV %s,%s",Matcher::regName[dst],Matcher::regName[src]);
929 #endif
930 }
931 return size+2;
932 }
934 static int impl_fp_store_helper( CodeBuffer *cbuf, bool do_size, int src_lo, int src_hi, int dst_lo, int dst_hi,
935 int offset, int size, outputStream* st ) {
936 if( src_lo != FPR1L_num ) { // Move value to top of FP stack, if not already there
937 if( cbuf ) {
938 emit_opcode( *cbuf, 0xD9 ); // FLD (i.e., push it)
939 emit_d8( *cbuf, 0xC0-1+Matcher::_regEncode[src_lo] );
940 #ifndef PRODUCT
941 } else if( !do_size ) {
942 if( size != 0 ) st->print("\n\t");
943 st->print("FLD %s",Matcher::regName[src_lo]);
944 #endif
945 }
946 size += 2;
947 }
949 int st_op = (src_lo != FPR1L_num) ? EBX_num /*store & pop*/ : EDX_num /*store no pop*/;
950 const char *op_str;
951 int op;
952 if( src_lo+1 == src_hi && dst_lo+1 == dst_hi ) { // double store?
953 op_str = (src_lo != FPR1L_num) ? "FSTP_D" : "FST_D ";
954 op = 0xDD;
955 } else { // 32-bit store
956 op_str = (src_lo != FPR1L_num) ? "FSTP_S" : "FST_S ";
957 op = 0xD9;
958 assert( !OptoReg::is_valid(src_hi) && !OptoReg::is_valid(dst_hi), "no non-adjacent float-stores" );
959 }
961 return impl_helper(cbuf,do_size,false,offset,st_op,op,op_str,size, st);
962 }
964 uint MachSpillCopyNode::implementation( CodeBuffer *cbuf, PhaseRegAlloc *ra_, bool do_size, outputStream* st ) const {
965 // Get registers to move
966 OptoReg::Name src_second = ra_->get_reg_second(in(1));
967 OptoReg::Name src_first = ra_->get_reg_first(in(1));
968 OptoReg::Name dst_second = ra_->get_reg_second(this );
969 OptoReg::Name dst_first = ra_->get_reg_first(this );
971 enum RC src_second_rc = rc_class(src_second);
972 enum RC src_first_rc = rc_class(src_first);
973 enum RC dst_second_rc = rc_class(dst_second);
974 enum RC dst_first_rc = rc_class(dst_first);
976 assert( OptoReg::is_valid(src_first) && OptoReg::is_valid(dst_first), "must move at least 1 register" );
978 // Generate spill code!
979 int size = 0;
981 if( src_first == dst_first && src_second == dst_second )
982 return size; // Self copy, no move
984 // --------------------------------------
985 // Check for mem-mem move. push/pop to move.
986 if( src_first_rc == rc_stack && dst_first_rc == rc_stack ) {
987 if( src_second == dst_first ) { // overlapping stack copy ranges
988 assert( src_second_rc == rc_stack && dst_second_rc == rc_stack, "we only expect a stk-stk copy here" );
989 size = impl_helper(cbuf,do_size,true ,ra_->reg2offset(src_second),ESI_num,0xFF,"PUSH ",size, st);
990 size = impl_helper(cbuf,do_size,false,ra_->reg2offset(dst_second),EAX_num,0x8F,"POP ",size, st);
991 src_second_rc = dst_second_rc = rc_bad; // flag as already moved the second bits
992 }
993 // move low bits
994 size = impl_helper(cbuf,do_size,true ,ra_->reg2offset(src_first),ESI_num,0xFF,"PUSH ",size, st);
995 size = impl_helper(cbuf,do_size,false,ra_->reg2offset(dst_first),EAX_num,0x8F,"POP ",size, st);
996 if( src_second_rc == rc_stack && dst_second_rc == rc_stack ) { // mov second bits
997 size = impl_helper(cbuf,do_size,true ,ra_->reg2offset(src_second),ESI_num,0xFF,"PUSH ",size, st);
998 size = impl_helper(cbuf,do_size,false,ra_->reg2offset(dst_second),EAX_num,0x8F,"POP ",size, st);
999 }
1000 return size;
1001 }
1003 // --------------------------------------
1004 // Check for integer reg-reg copy
1005 if( src_first_rc == rc_int && dst_first_rc == rc_int )
1006 size = impl_mov_helper(cbuf,do_size,src_first,dst_first,size, st);
1008 // Check for integer store
1009 if( src_first_rc == rc_int && dst_first_rc == rc_stack )
1010 size = impl_helper(cbuf,do_size,false,ra_->reg2offset(dst_first),src_first,0x89,"MOV ",size, st);
1012 // Check for integer load
1013 if( dst_first_rc == rc_int && src_first_rc == rc_stack )
1014 size = impl_helper(cbuf,do_size,true ,ra_->reg2offset(src_first),dst_first,0x8B,"MOV ",size, st);
1016 // Check for integer reg-xmm reg copy
1017 if( src_first_rc == rc_int && dst_first_rc == rc_xmm ) {
1018 assert( (src_second_rc == rc_bad && dst_second_rc == rc_bad),
1019 "no 64 bit integer-float reg moves" );
1020 return impl_movgpr2x_helper(cbuf,do_size,src_first,dst_first,src_second, dst_second, size, st);
1021 }
1022 // --------------------------------------
1023 // Check for float reg-reg copy
1024 if( src_first_rc == rc_float && dst_first_rc == rc_float ) {
1025 assert( (src_second_rc == rc_bad && dst_second_rc == rc_bad) ||
1026 (src_first+1 == src_second && dst_first+1 == dst_second), "no non-adjacent float-moves" );
1027 if( cbuf ) {
1029 // Note the mucking with the register encode to compensate for the 0/1
1030 // indexing issue mentioned in a comment in the reg_def sections
1031 // for FPR registers many lines above here.
1033 if( src_first != FPR1L_num ) {
1034 emit_opcode (*cbuf, 0xD9 ); // FLD ST(i)
1035 emit_d8 (*cbuf, 0xC0+Matcher::_regEncode[src_first]-1 );
1036 emit_opcode (*cbuf, 0xDD ); // FSTP ST(i)
1037 emit_d8 (*cbuf, 0xD8+Matcher::_regEncode[dst_first] );
1038 } else {
1039 emit_opcode (*cbuf, 0xDD ); // FST ST(i)
1040 emit_d8 (*cbuf, 0xD0+Matcher::_regEncode[dst_first]-1 );
1041 }
1042 #ifndef PRODUCT
1043 } else if( !do_size ) {
1044 if( size != 0 ) st->print("\n\t");
1045 if( src_first != FPR1L_num ) st->print("FLD %s\n\tFSTP %s",Matcher::regName[src_first],Matcher::regName[dst_first]);
1046 else st->print( "FST %s", Matcher::regName[dst_first]);
1047 #endif
1048 }
1049 return size + ((src_first != FPR1L_num) ? 2+2 : 2);
1050 }
1052 // Check for float store
1053 if( src_first_rc == rc_float && dst_first_rc == rc_stack ) {
1054 return impl_fp_store_helper(cbuf,do_size,src_first,src_second,dst_first,dst_second,ra_->reg2offset(dst_first),size, st);
1055 }
1057 // Check for float load
1058 if( dst_first_rc == rc_float && src_first_rc == rc_stack ) {
1059 int offset = ra_->reg2offset(src_first);
1060 const char *op_str;
1061 int op;
1062 if( src_first+1 == src_second && dst_first+1 == dst_second ) { // double load?
1063 op_str = "FLD_D";
1064 op = 0xDD;
1065 } else { // 32-bit load
1066 op_str = "FLD_S";
1067 op = 0xD9;
1068 assert( src_second_rc == rc_bad && dst_second_rc == rc_bad, "no non-adjacent float-loads" );
1069 }
1070 if( cbuf ) {
1071 emit_opcode (*cbuf, op );
1072 encode_RegMem(*cbuf, 0x0, ESP_enc, 0x4, 0, offset, false);
1073 emit_opcode (*cbuf, 0xDD ); // FSTP ST(i)
1074 emit_d8 (*cbuf, 0xD8+Matcher::_regEncode[dst_first] );
1075 #ifndef PRODUCT
1076 } else if( !do_size ) {
1077 if( size != 0 ) st->print("\n\t");
1078 st->print("%s ST,[ESP + #%d]\n\tFSTP %s",op_str, offset,Matcher::regName[dst_first]);
1079 #endif
1080 }
1081 int offset_size = (offset == 0) ? 0 : ((offset <= 127) ? 1 : 4);
1082 return size + 3+offset_size+2;
1083 }
1085 // Check for xmm reg-reg copy
1086 if( src_first_rc == rc_xmm && dst_first_rc == rc_xmm ) {
1087 assert( (src_second_rc == rc_bad && dst_second_rc == rc_bad) ||
1088 (src_first+1 == src_second && dst_first+1 == dst_second),
1089 "no non-adjacent float-moves" );
1090 return impl_movx_helper(cbuf,do_size,src_first,dst_first,src_second, dst_second, size, st);
1091 }
1093 // Check for xmm reg-integer reg copy
1094 if( src_first_rc == rc_xmm && dst_first_rc == rc_int ) {
1095 assert( (src_second_rc == rc_bad && dst_second_rc == rc_bad),
1096 "no 64 bit float-integer reg moves" );
1097 return impl_movx2gpr_helper(cbuf,do_size,src_first,dst_first,src_second, dst_second, size, st);
1098 }
1100 // Check for xmm store
1101 if( src_first_rc == rc_xmm && dst_first_rc == rc_stack ) {
1102 return impl_x_helper(cbuf,do_size,false,ra_->reg2offset(dst_first),src_first, src_second, size, st);
1103 }
1105 // Check for float xmm load
1106 if( dst_first_rc == rc_xmm && src_first_rc == rc_stack ) {
1107 return impl_x_helper(cbuf,do_size,true ,ra_->reg2offset(src_first),dst_first, dst_second, size, st);
1108 }
1110 // Copy from float reg to xmm reg
1111 if( dst_first_rc == rc_xmm && src_first_rc == rc_float ) {
1112 // copy to the top of stack from floating point reg
1113 // and use LEA to preserve flags
1114 if( cbuf ) {
1115 emit_opcode(*cbuf,0x8D); // LEA ESP,[ESP-8]
1116 emit_rm(*cbuf, 0x1, ESP_enc, 0x04);
1117 emit_rm(*cbuf, 0x0, 0x04, ESP_enc);
1118 emit_d8(*cbuf,0xF8);
1119 #ifndef PRODUCT
1120 } else if( !do_size ) {
1121 if( size != 0 ) st->print("\n\t");
1122 st->print("LEA ESP,[ESP-8]");
1123 #endif
1124 }
1125 size += 4;
1127 size = impl_fp_store_helper(cbuf,do_size,src_first,src_second,dst_first,dst_second,0,size, st);
1129 // Copy from the temp memory to the xmm reg.
1130 size = impl_x_helper(cbuf,do_size,true ,0,dst_first, dst_second, size, st);
1132 if( cbuf ) {
1133 emit_opcode(*cbuf,0x8D); // LEA ESP,[ESP+8]
1134 emit_rm(*cbuf, 0x1, ESP_enc, 0x04);
1135 emit_rm(*cbuf, 0x0, 0x04, ESP_enc);
1136 emit_d8(*cbuf,0x08);
1137 #ifndef PRODUCT
1138 } else if( !do_size ) {
1139 if( size != 0 ) st->print("\n\t");
1140 st->print("LEA ESP,[ESP+8]");
1141 #endif
1142 }
1143 size += 4;
1144 return size;
1145 }
1147 assert( size > 0, "missed a case" );
1149 // --------------------------------------------------------------------
1150 // Check for second bits still needing moving.
1151 if( src_second == dst_second )
1152 return size; // Self copy; no move
1153 assert( src_second_rc != rc_bad && dst_second_rc != rc_bad, "src_second & dst_second cannot be Bad" );
1155 // Check for second word int-int move
1156 if( src_second_rc == rc_int && dst_second_rc == rc_int )
1157 return impl_mov_helper(cbuf,do_size,src_second,dst_second,size, st);
1159 // Check for second word integer store
1160 if( src_second_rc == rc_int && dst_second_rc == rc_stack )
1161 return impl_helper(cbuf,do_size,false,ra_->reg2offset(dst_second),src_second,0x89,"MOV ",size, st);
1163 // Check for second word integer load
1164 if( dst_second_rc == rc_int && src_second_rc == rc_stack )
1165 return impl_helper(cbuf,do_size,true ,ra_->reg2offset(src_second),dst_second,0x8B,"MOV ",size, st);
1168 Unimplemented();
1169 }
1171 #ifndef PRODUCT
1172 void MachSpillCopyNode::format( PhaseRegAlloc *ra_, outputStream* st ) const {
1173 implementation( NULL, ra_, false, st );
1174 }
1175 #endif
1177 void MachSpillCopyNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
1178 implementation( &cbuf, ra_, false, NULL );
1179 }
1181 uint MachSpillCopyNode::size(PhaseRegAlloc *ra_) const {
1182 return implementation( NULL, ra_, true, NULL );
1183 }
1185 //=============================================================================
1186 #ifndef PRODUCT
1187 void MachNopNode::format( PhaseRegAlloc *, outputStream* st ) const {
1188 st->print("NOP \t# %d bytes pad for loops and calls", _count);
1189 }
1190 #endif
1192 void MachNopNode::emit(CodeBuffer &cbuf, PhaseRegAlloc * ) const {
1193 MacroAssembler _masm(&cbuf);
1194 __ nop(_count);
1195 }
1197 uint MachNopNode::size(PhaseRegAlloc *) const {
1198 return _count;
1199 }
1202 //=============================================================================
1203 #ifndef PRODUCT
1204 void BoxLockNode::format( PhaseRegAlloc *ra_, outputStream* st ) const {
1205 int offset = ra_->reg2offset(in_RegMask(0).find_first_elem());
1206 int reg = ra_->get_reg_first(this);
1207 st->print("LEA %s,[ESP + #%d]",Matcher::regName[reg],offset);
1208 }
1209 #endif
1211 void BoxLockNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
1212 int offset = ra_->reg2offset(in_RegMask(0).find_first_elem());
1213 int reg = ra_->get_encode(this);
1214 if( offset >= 128 ) {
1215 emit_opcode(cbuf, 0x8D); // LEA reg,[SP+offset]
1216 emit_rm(cbuf, 0x2, reg, 0x04);
1217 emit_rm(cbuf, 0x0, 0x04, ESP_enc);
1218 emit_d32(cbuf, offset);
1219 }
1220 else {
1221 emit_opcode(cbuf, 0x8D); // LEA reg,[SP+offset]
1222 emit_rm(cbuf, 0x1, reg, 0x04);
1223 emit_rm(cbuf, 0x0, 0x04, ESP_enc);
1224 emit_d8(cbuf, offset);
1225 }
1226 }
1228 uint BoxLockNode::size(PhaseRegAlloc *ra_) const {
1229 int offset = ra_->reg2offset(in_RegMask(0).find_first_elem());
1230 if( offset >= 128 ) {
1231 return 7;
1232 }
1233 else {
1234 return 4;
1235 }
1236 }
1238 //=============================================================================
1240 // emit call stub, compiled java to interpreter
1241 void emit_java_to_interp(CodeBuffer &cbuf ) {
1242 // Stub is fixed up when the corresponding call is converted from calling
1243 // compiled code to calling interpreted code.
1244 // mov rbx,0
1245 // jmp -1
1247 address mark = cbuf.insts_mark(); // get mark within main instrs section
1249 // Note that the code buffer's insts_mark is always relative to insts.
1250 // That's why we must use the macroassembler to generate a stub.
1251 MacroAssembler _masm(&cbuf);
1253 address base =
1254 __ start_a_stub(Compile::MAX_stubs_size);
1255 if (base == NULL) return; // CodeBuffer::expand failed
1256 // static stub relocation stores the instruction address of the call
1257 __ relocate(static_stub_Relocation::spec(mark), RELOC_IMM32);
1258 // static stub relocation also tags the methodOop in the code-stream.
1259 __ movoop(rbx, (jobject)NULL); // method is zapped till fixup time
1260 // This is recognized as unresolved by relocs/nativeInst/ic code
1261 __ jump(RuntimeAddress(__ pc()));
1263 __ end_a_stub();
1264 // Update current stubs pointer and restore insts_end.
1265 }
1266 // size of call stub, compiled java to interpretor
1267 uint size_java_to_interp() {
1268 return 10; // movl; jmp
1269 }
1270 // relocation entries for call stub, compiled java to interpretor
1271 uint reloc_java_to_interp() {
1272 return 4; // 3 in emit_java_to_interp + 1 in Java_Static_Call
1273 }
1275 //=============================================================================
1276 #ifndef PRODUCT
1277 void MachUEPNode::format( PhaseRegAlloc *ra_, outputStream* st ) const {
1278 st->print_cr( "CMP EAX,[ECX+4]\t# Inline cache check");
1279 st->print_cr("\tJNE SharedRuntime::handle_ic_miss_stub");
1280 st->print_cr("\tNOP");
1281 st->print_cr("\tNOP");
1282 if( !OptoBreakpoint )
1283 st->print_cr("\tNOP");
1284 }
1285 #endif
1287 void MachUEPNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
1288 MacroAssembler masm(&cbuf);
1289 #ifdef ASSERT
1290 uint insts_size = cbuf.insts_size();
1291 #endif
1292 masm.cmpptr(rax, Address(rcx, oopDesc::klass_offset_in_bytes()));
1293 masm.jump_cc(Assembler::notEqual,
1294 RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1295 /* WARNING these NOPs are critical so that verified entry point is properly
1296 aligned for patching by NativeJump::patch_verified_entry() */
1297 int nops_cnt = 2;
1298 if( !OptoBreakpoint ) // Leave space for int3
1299 nops_cnt += 1;
1300 masm.nop(nops_cnt);
1302 assert(cbuf.insts_size() - insts_size == size(ra_), "checking code size of inline cache node");
1303 }
1305 uint MachUEPNode::size(PhaseRegAlloc *ra_) const {
1306 return OptoBreakpoint ? 11 : 12;
1307 }
1310 //=============================================================================
1311 uint size_exception_handler() {
1312 // NativeCall instruction size is the same as NativeJump.
1313 // exception handler starts out as jump and can be patched to
1314 // a call be deoptimization. (4932387)
1315 // Note that this value is also credited (in output.cpp) to
1316 // the size of the code section.
1317 return NativeJump::instruction_size;
1318 }
1320 // Emit exception handler code. Stuff framesize into a register
1321 // and call a VM stub routine.
1322 int emit_exception_handler(CodeBuffer& cbuf) {
1324 // Note that the code buffer's insts_mark is always relative to insts.
1325 // That's why we must use the macroassembler to generate a handler.
1326 MacroAssembler _masm(&cbuf);
1327 address base =
1328 __ start_a_stub(size_exception_handler());
1329 if (base == NULL) return 0; // CodeBuffer::expand failed
1330 int offset = __ offset();
1331 __ jump(RuntimeAddress(OptoRuntime::exception_blob()->entry_point()));
1332 assert(__ offset() - offset <= (int) size_exception_handler(), "overflow");
1333 __ end_a_stub();
1334 return offset;
1335 }
1337 uint size_deopt_handler() {
1338 // NativeCall instruction size is the same as NativeJump.
1339 // exception handler starts out as jump and can be patched to
1340 // a call be deoptimization. (4932387)
1341 // Note that this value is also credited (in output.cpp) to
1342 // the size of the code section.
1343 return 5 + NativeJump::instruction_size; // pushl(); jmp;
1344 }
1346 // Emit deopt handler code.
1347 int emit_deopt_handler(CodeBuffer& cbuf) {
1349 // Note that the code buffer's insts_mark is always relative to insts.
1350 // That's why we must use the macroassembler to generate a handler.
1351 MacroAssembler _masm(&cbuf);
1352 address base =
1353 __ start_a_stub(size_exception_handler());
1354 if (base == NULL) return 0; // CodeBuffer::expand failed
1355 int offset = __ offset();
1356 InternalAddress here(__ pc());
1357 __ pushptr(here.addr());
1359 __ jump(RuntimeAddress(SharedRuntime::deopt_blob()->unpack()));
1360 assert(__ offset() - offset <= (int) size_deopt_handler(), "overflow");
1361 __ end_a_stub();
1362 return offset;
1363 }
1366 const bool Matcher::match_rule_supported(int opcode) {
1367 if (!has_match_rule(opcode))
1368 return false;
1370 return true; // Per default match rules are supported.
1371 }
1373 int Matcher::regnum_to_fpu_offset(int regnum) {
1374 return regnum - 32; // The FP registers are in the second chunk
1375 }
1377 // This is UltraSparc specific, true just means we have fast l2f conversion
1378 const bool Matcher::convL2FSupported(void) {
1379 return true;
1380 }
1382 // Vector width in bytes
1383 const uint Matcher::vector_width_in_bytes(void) {
1384 return UseSSE >= 2 ? 8 : 0;
1385 }
1387 // Vector ideal reg
1388 const uint Matcher::vector_ideal_reg(void) {
1389 return Op_RegD;
1390 }
1392 // Is this branch offset short enough that a short branch can be used?
1393 //
1394 // NOTE: If the platform does not provide any short branch variants, then
1395 // this method should return false for offset 0.
1396 bool Matcher::is_short_branch_offset(int rule, int br_size, int offset) {
1397 // The passed offset is relative to address of the branch.
1398 // On 86 a branch displacement is calculated relative to address
1399 // of a next instruction.
1400 offset -= br_size;
1402 // the short version of jmpConUCF2 contains multiple branches,
1403 // making the reach slightly less
1404 if (rule == jmpConUCF2_rule)
1405 return (-126 <= offset && offset <= 125);
1406 return (-128 <= offset && offset <= 127);
1407 }
1409 const bool Matcher::isSimpleConstant64(jlong value) {
1410 // Will one (StoreL ConL) be cheaper than two (StoreI ConI)?.
1411 return false;
1412 }
1414 // The ecx parameter to rep stos for the ClearArray node is in dwords.
1415 const bool Matcher::init_array_count_is_in_bytes = false;
1417 // Threshold size for cleararray.
1418 const int Matcher::init_array_short_size = 8 * BytesPerLong;
1420 // Needs 2 CMOV's for longs.
1421 const int Matcher::long_cmove_cost() { return 1; }
1423 // No CMOVF/CMOVD with SSE/SSE2
1424 const int Matcher::float_cmove_cost() { return (UseSSE>=1) ? ConditionalMoveLimit : 0; }
1426 // Should the Matcher clone shifts on addressing modes, expecting them to
1427 // be subsumed into complex addressing expressions or compute them into
1428 // registers? True for Intel but false for most RISCs
1429 const bool Matcher::clone_shift_expressions = true;
1431 // Do we need to mask the count passed to shift instructions or does
1432 // the cpu only look at the lower 5/6 bits anyway?
1433 const bool Matcher::need_masked_shift_count = false;
1435 bool Matcher::narrow_oop_use_complex_address() {
1436 ShouldNotCallThis();
1437 return true;
1438 }
1441 // Is it better to copy float constants, or load them directly from memory?
1442 // Intel can load a float constant from a direct address, requiring no
1443 // extra registers. Most RISCs will have to materialize an address into a
1444 // register first, so they would do better to copy the constant from stack.
1445 const bool Matcher::rematerialize_float_constants = true;
1447 // If CPU can load and store mis-aligned doubles directly then no fixup is
1448 // needed. Else we split the double into 2 integer pieces and move it
1449 // piece-by-piece. Only happens when passing doubles into C code as the
1450 // Java calling convention forces doubles to be aligned.
1451 const bool Matcher::misaligned_doubles_ok = true;
1454 void Matcher::pd_implicit_null_fixup(MachNode *node, uint idx) {
1455 // Get the memory operand from the node
1456 uint numopnds = node->num_opnds(); // Virtual call for number of operands
1457 uint skipped = node->oper_input_base(); // Sum of leaves skipped so far
1458 assert( idx >= skipped, "idx too low in pd_implicit_null_fixup" );
1459 uint opcnt = 1; // First operand
1460 uint num_edges = node->_opnds[1]->num_edges(); // leaves for first operand
1461 while( idx >= skipped+num_edges ) {
1462 skipped += num_edges;
1463 opcnt++; // Bump operand count
1464 assert( opcnt < numopnds, "Accessing non-existent operand" );
1465 num_edges = node->_opnds[opcnt]->num_edges(); // leaves for next operand
1466 }
1468 MachOper *memory = node->_opnds[opcnt];
1469 MachOper *new_memory = NULL;
1470 switch (memory->opcode()) {
1471 case DIRECT:
1472 case INDOFFSET32X:
1473 // No transformation necessary.
1474 return;
1475 case INDIRECT:
1476 new_memory = new (C) indirect_win95_safeOper( );
1477 break;
1478 case INDOFFSET8:
1479 new_memory = new (C) indOffset8_win95_safeOper(memory->disp(NULL, NULL, 0));
1480 break;
1481 case INDOFFSET32:
1482 new_memory = new (C) indOffset32_win95_safeOper(memory->disp(NULL, NULL, 0));
1483 break;
1484 case INDINDEXOFFSET:
1485 new_memory = new (C) indIndexOffset_win95_safeOper(memory->disp(NULL, NULL, 0));
1486 break;
1487 case INDINDEXSCALE:
1488 new_memory = new (C) indIndexScale_win95_safeOper(memory->scale());
1489 break;
1490 case INDINDEXSCALEOFFSET:
1491 new_memory = new (C) indIndexScaleOffset_win95_safeOper(memory->scale(), memory->disp(NULL, NULL, 0));
1492 break;
1493 case LOAD_LONG_INDIRECT:
1494 case LOAD_LONG_INDOFFSET32:
1495 // Does not use EBP as address register, use { EDX, EBX, EDI, ESI}
1496 return;
1497 default:
1498 assert(false, "unexpected memory operand in pd_implicit_null_fixup()");
1499 return;
1500 }
1501 node->_opnds[opcnt] = new_memory;
1502 }
1504 // Advertise here if the CPU requires explicit rounding operations
1505 // to implement the UseStrictFP mode.
1506 const bool Matcher::strict_fp_requires_explicit_rounding = true;
1508 // Are floats conerted to double when stored to stack during deoptimization?
1509 // On x32 it is stored with convertion only when FPU is used for floats.
1510 bool Matcher::float_in_double() { return (UseSSE == 0); }
1512 // Do ints take an entire long register or just half?
1513 const bool Matcher::int_in_long = false;
1515 // Return whether or not this register is ever used as an argument. This
1516 // function is used on startup to build the trampoline stubs in generateOptoStub.
1517 // Registers not mentioned will be killed by the VM call in the trampoline, and
1518 // arguments in those registers not be available to the callee.
1519 bool Matcher::can_be_java_arg( int reg ) {
1520 if( reg == ECX_num || reg == EDX_num ) return true;
1521 if( (reg == XMM0a_num || reg == XMM1a_num) && UseSSE>=1 ) return true;
1522 if( (reg == XMM0b_num || reg == XMM1b_num) && UseSSE>=2 ) return true;
1523 return false;
1524 }
1526 bool Matcher::is_spillable_arg( int reg ) {
1527 return can_be_java_arg(reg);
1528 }
1530 bool Matcher::use_asm_for_ldiv_by_con( jlong divisor ) {
1531 // Use hardware integer DIV instruction when
1532 // it is faster than a code which use multiply.
1533 // Only when constant divisor fits into 32 bit
1534 // (min_jint is excluded to get only correct
1535 // positive 32 bit values from negative).
1536 return VM_Version::has_fast_idiv() &&
1537 (divisor == (int)divisor && divisor != min_jint);
1538 }
1540 // Register for DIVI projection of divmodI
1541 RegMask Matcher::divI_proj_mask() {
1542 return EAX_REG_mask();
1543 }
1545 // Register for MODI projection of divmodI
1546 RegMask Matcher::modI_proj_mask() {
1547 return EDX_REG_mask();
1548 }
1550 // Register for DIVL projection of divmodL
1551 RegMask Matcher::divL_proj_mask() {
1552 ShouldNotReachHere();
1553 return RegMask();
1554 }
1556 // Register for MODL projection of divmodL
1557 RegMask Matcher::modL_proj_mask() {
1558 ShouldNotReachHere();
1559 return RegMask();
1560 }
1562 const RegMask Matcher::method_handle_invoke_SP_save_mask() {
1563 return EBP_REG_mask();
1564 }
1566 // Returns true if the high 32 bits of the value is known to be zero.
1567 bool is_operand_hi32_zero(Node* n) {
1568 int opc = n->Opcode();
1569 if (opc == Op_LoadUI2L) {
1570 return true;
1571 }
1572 if (opc == Op_AndL) {
1573 Node* o2 = n->in(2);
1574 if (o2->is_Con() && (o2->get_long() & 0xFFFFFFFF00000000LL) == 0LL) {
1575 return true;
1576 }
1577 }
1578 if (opc == Op_ConL && (n->get_long() & 0xFFFFFFFF00000000LL) == 0LL) {
1579 return true;
1580 }
1581 return false;
1582 }
1584 %}
1586 //----------ENCODING BLOCK-----------------------------------------------------
1587 // This block specifies the encoding classes used by the compiler to output
1588 // byte streams. Encoding classes generate functions which are called by
1589 // Machine Instruction Nodes in order to generate the bit encoding of the
1590 // instruction. Operands specify their base encoding interface with the
1591 // interface keyword. There are currently supported four interfaces,
1592 // REG_INTER, CONST_INTER, MEMORY_INTER, & COND_INTER. REG_INTER causes an
1593 // operand to generate a function which returns its register number when
1594 // queried. CONST_INTER causes an operand to generate a function which
1595 // returns the value of the constant when queried. MEMORY_INTER causes an
1596 // operand to generate four functions which return the Base Register, the
1597 // Index Register, the Scale Value, and the Offset Value of the operand when
1598 // queried. COND_INTER causes an operand to generate six functions which
1599 // return the encoding code (ie - encoding bits for the instruction)
1600 // associated with each basic boolean condition for a conditional instruction.
1601 // Instructions specify two basic values for encoding. They use the
1602 // ins_encode keyword to specify their encoding class (which must be one of
1603 // the class names specified in the encoding block), and they use the
1604 // opcode keyword to specify, in order, their primary, secondary, and
1605 // tertiary opcode. Only the opcode sections which a particular instruction
1606 // needs for encoding need to be specified.
1607 encode %{
1608 // Build emit functions for each basic byte or larger field in the intel
1609 // encoding scheme (opcode, rm, sib, immediate), and call them from C++
1610 // code in the enc_class source block. Emit functions will live in the
1611 // main source block for now. In future, we can generalize this by
1612 // adding a syntax that specifies the sizes of fields in an order,
1613 // so that the adlc can build the emit functions automagically
1615 // Emit primary opcode
1616 enc_class OpcP %{
1617 emit_opcode(cbuf, $primary);
1618 %}
1620 // Emit secondary opcode
1621 enc_class OpcS %{
1622 emit_opcode(cbuf, $secondary);
1623 %}
1625 // Emit opcode directly
1626 enc_class Opcode(immI d8) %{
1627 emit_opcode(cbuf, $d8$$constant);
1628 %}
1630 enc_class SizePrefix %{
1631 emit_opcode(cbuf,0x66);
1632 %}
1634 enc_class RegReg (eRegI dst, eRegI src) %{ // RegReg(Many)
1635 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
1636 %}
1638 enc_class OpcRegReg (immI opcode, eRegI dst, eRegI src) %{ // OpcRegReg(Many)
1639 emit_opcode(cbuf,$opcode$$constant);
1640 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
1641 %}
1643 enc_class mov_r32_imm0( eRegI dst ) %{
1644 emit_opcode( cbuf, 0xB8 + $dst$$reg ); // 0xB8+ rd -- MOV r32 ,imm32
1645 emit_d32 ( cbuf, 0x0 ); // imm32==0x0
1646 %}
1648 enc_class cdq_enc %{
1649 // Full implementation of Java idiv and irem; checks for
1650 // special case as described in JVM spec., p.243 & p.271.
1651 //
1652 // normal case special case
1653 //
1654 // input : rax,: dividend min_int
1655 // reg: divisor -1
1656 //
1657 // output: rax,: quotient (= rax, idiv reg) min_int
1658 // rdx: remainder (= rax, irem reg) 0
1659 //
1660 // Code sequnce:
1661 //
1662 // 81 F8 00 00 00 80 cmp rax,80000000h
1663 // 0F 85 0B 00 00 00 jne normal_case
1664 // 33 D2 xor rdx,edx
1665 // 83 F9 FF cmp rcx,0FFh
1666 // 0F 84 03 00 00 00 je done
1667 // normal_case:
1668 // 99 cdq
1669 // F7 F9 idiv rax,ecx
1670 // done:
1671 //
1672 emit_opcode(cbuf,0x81); emit_d8(cbuf,0xF8);
1673 emit_opcode(cbuf,0x00); emit_d8(cbuf,0x00);
1674 emit_opcode(cbuf,0x00); emit_d8(cbuf,0x80); // cmp rax,80000000h
1675 emit_opcode(cbuf,0x0F); emit_d8(cbuf,0x85);
1676 emit_opcode(cbuf,0x0B); emit_d8(cbuf,0x00);
1677 emit_opcode(cbuf,0x00); emit_d8(cbuf,0x00); // jne normal_case
1678 emit_opcode(cbuf,0x33); emit_d8(cbuf,0xD2); // xor rdx,edx
1679 emit_opcode(cbuf,0x83); emit_d8(cbuf,0xF9); emit_d8(cbuf,0xFF); // cmp rcx,0FFh
1680 emit_opcode(cbuf,0x0F); emit_d8(cbuf,0x84);
1681 emit_opcode(cbuf,0x03); emit_d8(cbuf,0x00);
1682 emit_opcode(cbuf,0x00); emit_d8(cbuf,0x00); // je done
1683 // normal_case:
1684 emit_opcode(cbuf,0x99); // cdq
1685 // idiv (note: must be emitted by the user of this rule)
1686 // normal:
1687 %}
1689 // Dense encoding for older common ops
1690 enc_class Opc_plus(immI opcode, eRegI reg) %{
1691 emit_opcode(cbuf, $opcode$$constant + $reg$$reg);
1692 %}
1695 // Opcde enc_class for 8/32 bit immediate instructions with sign-extension
1696 enc_class OpcSE (immI imm) %{ // Emit primary opcode and set sign-extend bit
1697 // Check for 8-bit immediate, and set sign extend bit in opcode
1698 if (($imm$$constant >= -128) && ($imm$$constant <= 127)) {
1699 emit_opcode(cbuf, $primary | 0x02);
1700 }
1701 else { // If 32-bit immediate
1702 emit_opcode(cbuf, $primary);
1703 }
1704 %}
1706 enc_class OpcSErm (eRegI dst, immI imm) %{ // OpcSEr/m
1707 // Emit primary opcode and set sign-extend bit
1708 // Check for 8-bit immediate, and set sign extend bit in opcode
1709 if (($imm$$constant >= -128) && ($imm$$constant <= 127)) {
1710 emit_opcode(cbuf, $primary | 0x02); }
1711 else { // If 32-bit immediate
1712 emit_opcode(cbuf, $primary);
1713 }
1714 // Emit r/m byte with secondary opcode, after primary opcode.
1715 emit_rm(cbuf, 0x3, $secondary, $dst$$reg);
1716 %}
1718 enc_class Con8or32 (immI imm) %{ // Con8or32(storeImmI), 8 or 32 bits
1719 // Check for 8-bit immediate, and set sign extend bit in opcode
1720 if (($imm$$constant >= -128) && ($imm$$constant <= 127)) {
1721 $$$emit8$imm$$constant;
1722 }
1723 else { // If 32-bit immediate
1724 // Output immediate
1725 $$$emit32$imm$$constant;
1726 }
1727 %}
1729 enc_class Long_OpcSErm_Lo(eRegL dst, immL imm) %{
1730 // Emit primary opcode and set sign-extend bit
1731 // Check for 8-bit immediate, and set sign extend bit in opcode
1732 int con = (int)$imm$$constant; // Throw away top bits
1733 emit_opcode(cbuf, ((con >= -128) && (con <= 127)) ? ($primary | 0x02) : $primary);
1734 // Emit r/m byte with secondary opcode, after primary opcode.
1735 emit_rm(cbuf, 0x3, $secondary, $dst$$reg);
1736 if ((con >= -128) && (con <= 127)) emit_d8 (cbuf,con);
1737 else emit_d32(cbuf,con);
1738 %}
1740 enc_class Long_OpcSErm_Hi(eRegL dst, immL imm) %{
1741 // Emit primary opcode and set sign-extend bit
1742 // Check for 8-bit immediate, and set sign extend bit in opcode
1743 int con = (int)($imm$$constant >> 32); // Throw away bottom bits
1744 emit_opcode(cbuf, ((con >= -128) && (con <= 127)) ? ($primary | 0x02) : $primary);
1745 // Emit r/m byte with tertiary opcode, after primary opcode.
1746 emit_rm(cbuf, 0x3, $tertiary, HIGH_FROM_LOW($dst$$reg));
1747 if ((con >= -128) && (con <= 127)) emit_d8 (cbuf,con);
1748 else emit_d32(cbuf,con);
1749 %}
1751 enc_class OpcSReg (eRegI dst) %{ // BSWAP
1752 emit_cc(cbuf, $secondary, $dst$$reg );
1753 %}
1755 enc_class bswap_long_bytes(eRegL dst) %{ // BSWAP
1756 int destlo = $dst$$reg;
1757 int desthi = HIGH_FROM_LOW(destlo);
1758 // bswap lo
1759 emit_opcode(cbuf, 0x0F);
1760 emit_cc(cbuf, 0xC8, destlo);
1761 // bswap hi
1762 emit_opcode(cbuf, 0x0F);
1763 emit_cc(cbuf, 0xC8, desthi);
1764 // xchg lo and hi
1765 emit_opcode(cbuf, 0x87);
1766 emit_rm(cbuf, 0x3, destlo, desthi);
1767 %}
1769 enc_class RegOpc (eRegI div) %{ // IDIV, IMOD, JMP indirect, ...
1770 emit_rm(cbuf, 0x3, $secondary, $div$$reg );
1771 %}
1773 enc_class enc_cmov(cmpOp cop ) %{ // CMOV
1774 $$$emit8$primary;
1775 emit_cc(cbuf, $secondary, $cop$$cmpcode);
1776 %}
1778 enc_class enc_cmov_dpr(cmpOp cop, regDPR src ) %{ // CMOV
1779 int op = 0xDA00 + $cop$$cmpcode + ($src$$reg-1);
1780 emit_d8(cbuf, op >> 8 );
1781 emit_d8(cbuf, op & 255);
1782 %}
1784 // emulate a CMOV with a conditional branch around a MOV
1785 enc_class enc_cmov_branch( cmpOp cop, immI brOffs ) %{ // CMOV
1786 // Invert sense of branch from sense of CMOV
1787 emit_cc( cbuf, 0x70, ($cop$$cmpcode^1) );
1788 emit_d8( cbuf, $brOffs$$constant );
1789 %}
1791 enc_class enc_PartialSubtypeCheck( ) %{
1792 Register Redi = as_Register(EDI_enc); // result register
1793 Register Reax = as_Register(EAX_enc); // super class
1794 Register Recx = as_Register(ECX_enc); // killed
1795 Register Resi = as_Register(ESI_enc); // sub class
1796 Label miss;
1798 MacroAssembler _masm(&cbuf);
1799 __ check_klass_subtype_slow_path(Resi, Reax, Recx, Redi,
1800 NULL, &miss,
1801 /*set_cond_codes:*/ true);
1802 if ($primary) {
1803 __ xorptr(Redi, Redi);
1804 }
1805 __ bind(miss);
1806 %}
1808 enc_class FFree_Float_Stack_All %{ // Free_Float_Stack_All
1809 MacroAssembler masm(&cbuf);
1810 int start = masm.offset();
1811 if (UseSSE >= 2) {
1812 if (VerifyFPU) {
1813 masm.verify_FPU(0, "must be empty in SSE2+ mode");
1814 }
1815 } else {
1816 // External c_calling_convention expects the FPU stack to be 'clean'.
1817 // Compiled code leaves it dirty. Do cleanup now.
1818 masm.empty_FPU_stack();
1819 }
1820 if (sizeof_FFree_Float_Stack_All == -1) {
1821 sizeof_FFree_Float_Stack_All = masm.offset() - start;
1822 } else {
1823 assert(masm.offset() - start == sizeof_FFree_Float_Stack_All, "wrong size");
1824 }
1825 %}
1827 enc_class Verify_FPU_For_Leaf %{
1828 if( VerifyFPU ) {
1829 MacroAssembler masm(&cbuf);
1830 masm.verify_FPU( -3, "Returning from Runtime Leaf call");
1831 }
1832 %}
1834 enc_class Java_To_Runtime (method meth) %{ // CALL Java_To_Runtime, Java_To_Runtime_Leaf
1835 // This is the instruction starting address for relocation info.
1836 cbuf.set_insts_mark();
1837 $$$emit8$primary;
1838 // CALL directly to the runtime
1839 emit_d32_reloc(cbuf, ($meth$$method - (int)(cbuf.insts_end()) - 4),
1840 runtime_call_Relocation::spec(), RELOC_IMM32 );
1842 if (UseSSE >= 2) {
1843 MacroAssembler _masm(&cbuf);
1844 BasicType rt = tf()->return_type();
1846 if ((rt == T_FLOAT || rt == T_DOUBLE) && !return_value_is_used()) {
1847 // A C runtime call where the return value is unused. In SSE2+
1848 // mode the result needs to be removed from the FPU stack. It's
1849 // likely that this function call could be removed by the
1850 // optimizer if the C function is a pure function.
1851 __ ffree(0);
1852 } else if (rt == T_FLOAT) {
1853 __ lea(rsp, Address(rsp, -4));
1854 __ fstp_s(Address(rsp, 0));
1855 __ movflt(xmm0, Address(rsp, 0));
1856 __ lea(rsp, Address(rsp, 4));
1857 } else if (rt == T_DOUBLE) {
1858 __ lea(rsp, Address(rsp, -8));
1859 __ fstp_d(Address(rsp, 0));
1860 __ movdbl(xmm0, Address(rsp, 0));
1861 __ lea(rsp, Address(rsp, 8));
1862 }
1863 }
1864 %}
1867 enc_class pre_call_FPU %{
1868 // If method sets FPU control word restore it here
1869 debug_only(int off0 = cbuf.insts_size());
1870 if( Compile::current()->in_24_bit_fp_mode() ) {
1871 MacroAssembler masm(&cbuf);
1872 masm.fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
1873 }
1874 debug_only(int off1 = cbuf.insts_size());
1875 assert(off1 - off0 == pre_call_FPU_size(), "correct size prediction");
1876 %}
1878 enc_class post_call_FPU %{
1879 // If method sets FPU control word do it here also
1880 if( Compile::current()->in_24_bit_fp_mode() ) {
1881 MacroAssembler masm(&cbuf);
1882 masm.fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
1883 }
1884 %}
1886 enc_class preserve_SP %{
1887 debug_only(int off0 = cbuf.insts_size());
1888 MacroAssembler _masm(&cbuf);
1889 // RBP is preserved across all calls, even compiled calls.
1890 // Use it to preserve RSP in places where the callee might change the SP.
1891 __ movptr(rbp_mh_SP_save, rsp);
1892 debug_only(int off1 = cbuf.insts_size());
1893 assert(off1 - off0 == preserve_SP_size(), "correct size prediction");
1894 %}
1896 enc_class restore_SP %{
1897 MacroAssembler _masm(&cbuf);
1898 __ movptr(rsp, rbp_mh_SP_save);
1899 %}
1901 enc_class Java_Static_Call (method meth) %{ // JAVA STATIC CALL
1902 // CALL to fixup routine. Fixup routine uses ScopeDesc info to determine
1903 // who we intended to call.
1904 cbuf.set_insts_mark();
1905 $$$emit8$primary;
1906 if ( !_method ) {
1907 emit_d32_reloc(cbuf, ($meth$$method - (int)(cbuf.insts_end()) - 4),
1908 runtime_call_Relocation::spec(), RELOC_IMM32 );
1909 } else if(_optimized_virtual) {
1910 emit_d32_reloc(cbuf, ($meth$$method - (int)(cbuf.insts_end()) - 4),
1911 opt_virtual_call_Relocation::spec(), RELOC_IMM32 );
1912 } else {
1913 emit_d32_reloc(cbuf, ($meth$$method - (int)(cbuf.insts_end()) - 4),
1914 static_call_Relocation::spec(), RELOC_IMM32 );
1915 }
1916 if( _method ) { // Emit stub for static call
1917 emit_java_to_interp(cbuf);
1918 }
1919 %}
1921 enc_class Java_Dynamic_Call (method meth) %{ // JAVA DYNAMIC CALL
1922 // !!!!!
1923 // Generate "Mov EAX,0x00", placeholder instruction to load oop-info
1924 // emit_call_dynamic_prologue( cbuf );
1925 cbuf.set_insts_mark();
1926 emit_opcode(cbuf, 0xB8 + EAX_enc); // mov EAX,-1
1927 emit_d32_reloc(cbuf, (int)Universe::non_oop_word(), oop_Relocation::spec_for_immediate(), RELOC_IMM32);
1928 address virtual_call_oop_addr = cbuf.insts_mark();
1929 // CALL to fixup routine. Fixup routine uses ScopeDesc info to determine
1930 // who we intended to call.
1931 cbuf.set_insts_mark();
1932 $$$emit8$primary;
1933 emit_d32_reloc(cbuf, ($meth$$method - (int)(cbuf.insts_end()) - 4),
1934 virtual_call_Relocation::spec(virtual_call_oop_addr), RELOC_IMM32 );
1935 %}
1937 enc_class Java_Compiled_Call (method meth) %{ // JAVA COMPILED CALL
1938 int disp = in_bytes(methodOopDesc::from_compiled_offset());
1939 assert( -128 <= disp && disp <= 127, "compiled_code_offset isn't small");
1941 // CALL *[EAX+in_bytes(methodOopDesc::from_compiled_code_entry_point_offset())]
1942 cbuf.set_insts_mark();
1943 $$$emit8$primary;
1944 emit_rm(cbuf, 0x01, $secondary, EAX_enc ); // R/M byte
1945 emit_d8(cbuf, disp); // Displacement
1947 %}
1949 // Following encoding is no longer used, but may be restored if calling
1950 // convention changes significantly.
1951 // Became: Xor_Reg(EBP), Java_To_Runtime( labl )
1952 //
1953 // enc_class Java_Interpreter_Call (label labl) %{ // JAVA INTERPRETER CALL
1954 // // int ic_reg = Matcher::inline_cache_reg();
1955 // // int ic_encode = Matcher::_regEncode[ic_reg];
1956 // // int imo_reg = Matcher::interpreter_method_oop_reg();
1957 // // int imo_encode = Matcher::_regEncode[imo_reg];
1958 //
1959 // // // Interpreter expects method_oop in EBX, currently a callee-saved register,
1960 // // // so we load it immediately before the call
1961 // // emit_opcode(cbuf, 0x8B); // MOV imo_reg,ic_reg # method_oop
1962 // // emit_rm(cbuf, 0x03, imo_encode, ic_encode ); // R/M byte
1963 //
1964 // // xor rbp,ebp
1965 // emit_opcode(cbuf, 0x33);
1966 // emit_rm(cbuf, 0x3, EBP_enc, EBP_enc);
1967 //
1968 // // CALL to interpreter.
1969 // cbuf.set_insts_mark();
1970 // $$$emit8$primary;
1971 // emit_d32_reloc(cbuf, ($labl$$label - (int)(cbuf.insts_end()) - 4),
1972 // runtime_call_Relocation::spec(), RELOC_IMM32 );
1973 // %}
1975 enc_class RegOpcImm (eRegI dst, immI8 shift) %{ // SHL, SAR, SHR
1976 $$$emit8$primary;
1977 emit_rm(cbuf, 0x3, $secondary, $dst$$reg);
1978 $$$emit8$shift$$constant;
1979 %}
1981 enc_class LdImmI (eRegI dst, immI src) %{ // Load Immediate
1982 // Load immediate does not have a zero or sign extended version
1983 // for 8-bit immediates
1984 emit_opcode(cbuf, 0xB8 + $dst$$reg);
1985 $$$emit32$src$$constant;
1986 %}
1988 enc_class LdImmP (eRegI dst, immI src) %{ // Load Immediate
1989 // Load immediate does not have a zero or sign extended version
1990 // for 8-bit immediates
1991 emit_opcode(cbuf, $primary + $dst$$reg);
1992 $$$emit32$src$$constant;
1993 %}
1995 enc_class LdImmL_Lo( eRegL dst, immL src) %{ // Load Immediate
1996 // Load immediate does not have a zero or sign extended version
1997 // for 8-bit immediates
1998 int dst_enc = $dst$$reg;
1999 int src_con = $src$$constant & 0x0FFFFFFFFL;
2000 if (src_con == 0) {
2001 // xor dst, dst
2002 emit_opcode(cbuf, 0x33);
2003 emit_rm(cbuf, 0x3, dst_enc, dst_enc);
2004 } else {
2005 emit_opcode(cbuf, $primary + dst_enc);
2006 emit_d32(cbuf, src_con);
2007 }
2008 %}
2010 enc_class LdImmL_Hi( eRegL dst, immL src) %{ // Load Immediate
2011 // Load immediate does not have a zero or sign extended version
2012 // for 8-bit immediates
2013 int dst_enc = $dst$$reg + 2;
2014 int src_con = ((julong)($src$$constant)) >> 32;
2015 if (src_con == 0) {
2016 // xor dst, dst
2017 emit_opcode(cbuf, 0x33);
2018 emit_rm(cbuf, 0x3, dst_enc, dst_enc);
2019 } else {
2020 emit_opcode(cbuf, $primary + dst_enc);
2021 emit_d32(cbuf, src_con);
2022 }
2023 %}
2026 // Encode a reg-reg copy. If it is useless, then empty encoding.
2027 enc_class enc_Copy( eRegI dst, eRegI src ) %{
2028 encode_Copy( cbuf, $dst$$reg, $src$$reg );
2029 %}
2031 enc_class enc_CopyL_Lo( eRegI dst, eRegL src ) %{
2032 encode_Copy( cbuf, $dst$$reg, $src$$reg );
2033 %}
2035 enc_class RegReg (eRegI dst, eRegI src) %{ // RegReg(Many)
2036 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2037 %}
2039 enc_class RegReg_Lo(eRegL dst, eRegL src) %{ // RegReg(Many)
2040 $$$emit8$primary;
2041 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2042 %}
2044 enc_class RegReg_Hi(eRegL dst, eRegL src) %{ // RegReg(Many)
2045 $$$emit8$secondary;
2046 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), HIGH_FROM_LOW($src$$reg));
2047 %}
2049 enc_class RegReg_Lo2(eRegL dst, eRegL src) %{ // RegReg(Many)
2050 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2051 %}
2053 enc_class RegReg_Hi2(eRegL dst, eRegL src) %{ // RegReg(Many)
2054 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), HIGH_FROM_LOW($src$$reg));
2055 %}
2057 enc_class RegReg_HiLo( eRegL src, eRegI dst ) %{
2058 emit_rm(cbuf, 0x3, $dst$$reg, HIGH_FROM_LOW($src$$reg));
2059 %}
2061 enc_class Con32 (immI src) %{ // Con32(storeImmI)
2062 // Output immediate
2063 $$$emit32$src$$constant;
2064 %}
2066 enc_class Con32FPR_as_bits(immFPR src) %{ // storeF_imm
2067 // Output Float immediate bits
2068 jfloat jf = $src$$constant;
2069 int jf_as_bits = jint_cast( jf );
2070 emit_d32(cbuf, jf_as_bits);
2071 %}
2073 enc_class Con32F_as_bits(immF src) %{ // storeX_imm
2074 // Output Float immediate bits
2075 jfloat jf = $src$$constant;
2076 int jf_as_bits = jint_cast( jf );
2077 emit_d32(cbuf, jf_as_bits);
2078 %}
2080 enc_class Con16 (immI src) %{ // Con16(storeImmI)
2081 // Output immediate
2082 $$$emit16$src$$constant;
2083 %}
2085 enc_class Con_d32(immI src) %{
2086 emit_d32(cbuf,$src$$constant);
2087 %}
2089 enc_class conmemref (eRegP t1) %{ // Con32(storeImmI)
2090 // Output immediate memory reference
2091 emit_rm(cbuf, 0x00, $t1$$reg, 0x05 );
2092 emit_d32(cbuf, 0x00);
2093 %}
2095 enc_class lock_prefix( ) %{
2096 if( os::is_MP() )
2097 emit_opcode(cbuf,0xF0); // [Lock]
2098 %}
2100 // Cmp-xchg long value.
2101 // Note: we need to swap rbx, and rcx before and after the
2102 // cmpxchg8 instruction because the instruction uses
2103 // rcx as the high order word of the new value to store but
2104 // our register encoding uses rbx,.
2105 enc_class enc_cmpxchg8(eSIRegP mem_ptr) %{
2107 // XCHG rbx,ecx
2108 emit_opcode(cbuf,0x87);
2109 emit_opcode(cbuf,0xD9);
2110 // [Lock]
2111 if( os::is_MP() )
2112 emit_opcode(cbuf,0xF0);
2113 // CMPXCHG8 [Eptr]
2114 emit_opcode(cbuf,0x0F);
2115 emit_opcode(cbuf,0xC7);
2116 emit_rm( cbuf, 0x0, 1, $mem_ptr$$reg );
2117 // XCHG rbx,ecx
2118 emit_opcode(cbuf,0x87);
2119 emit_opcode(cbuf,0xD9);
2120 %}
2122 enc_class enc_cmpxchg(eSIRegP mem_ptr) %{
2123 // [Lock]
2124 if( os::is_MP() )
2125 emit_opcode(cbuf,0xF0);
2127 // CMPXCHG [Eptr]
2128 emit_opcode(cbuf,0x0F);
2129 emit_opcode(cbuf,0xB1);
2130 emit_rm( cbuf, 0x0, 1, $mem_ptr$$reg );
2131 %}
2133 enc_class enc_flags_ne_to_boolean( iRegI res ) %{
2134 int res_encoding = $res$$reg;
2136 // MOV res,0
2137 emit_opcode( cbuf, 0xB8 + res_encoding);
2138 emit_d32( cbuf, 0 );
2139 // JNE,s fail
2140 emit_opcode(cbuf,0x75);
2141 emit_d8(cbuf, 5 );
2142 // MOV res,1
2143 emit_opcode( cbuf, 0xB8 + res_encoding);
2144 emit_d32( cbuf, 1 );
2145 // fail:
2146 %}
2148 enc_class set_instruction_start( ) %{
2149 cbuf.set_insts_mark(); // Mark start of opcode for reloc info in mem operand
2150 %}
2152 enc_class RegMem (eRegI ereg, memory mem) %{ // emit_reg_mem
2153 int reg_encoding = $ereg$$reg;
2154 int base = $mem$$base;
2155 int index = $mem$$index;
2156 int scale = $mem$$scale;
2157 int displace = $mem$$disp;
2158 bool disp_is_oop = $mem->disp_is_oop();
2159 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, disp_is_oop);
2160 %}
2162 enc_class RegMem_Hi(eRegL ereg, memory mem) %{ // emit_reg_mem
2163 int reg_encoding = HIGH_FROM_LOW($ereg$$reg); // Hi register of pair, computed from lo
2164 int base = $mem$$base;
2165 int index = $mem$$index;
2166 int scale = $mem$$scale;
2167 int displace = $mem$$disp + 4; // Offset is 4 further in memory
2168 assert( !$mem->disp_is_oop(), "Cannot add 4 to oop" );
2169 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, false/*disp_is_oop*/);
2170 %}
2172 enc_class move_long_small_shift( eRegL dst, immI_1_31 cnt ) %{
2173 int r1, r2;
2174 if( $tertiary == 0xA4 ) { r1 = $dst$$reg; r2 = HIGH_FROM_LOW($dst$$reg); }
2175 else { r2 = $dst$$reg; r1 = HIGH_FROM_LOW($dst$$reg); }
2176 emit_opcode(cbuf,0x0F);
2177 emit_opcode(cbuf,$tertiary);
2178 emit_rm(cbuf, 0x3, r1, r2);
2179 emit_d8(cbuf,$cnt$$constant);
2180 emit_d8(cbuf,$primary);
2181 emit_rm(cbuf, 0x3, $secondary, r1);
2182 emit_d8(cbuf,$cnt$$constant);
2183 %}
2185 enc_class move_long_big_shift_sign( eRegL dst, immI_32_63 cnt ) %{
2186 emit_opcode( cbuf, 0x8B ); // Move
2187 emit_rm(cbuf, 0x3, $dst$$reg, HIGH_FROM_LOW($dst$$reg));
2188 if( $cnt$$constant > 32 ) { // Shift, if not by zero
2189 emit_d8(cbuf,$primary);
2190 emit_rm(cbuf, 0x3, $secondary, $dst$$reg);
2191 emit_d8(cbuf,$cnt$$constant-32);
2192 }
2193 emit_d8(cbuf,$primary);
2194 emit_rm(cbuf, 0x3, $secondary, HIGH_FROM_LOW($dst$$reg));
2195 emit_d8(cbuf,31);
2196 %}
2198 enc_class move_long_big_shift_clr( eRegL dst, immI_32_63 cnt ) %{
2199 int r1, r2;
2200 if( $secondary == 0x5 ) { r1 = $dst$$reg; r2 = HIGH_FROM_LOW($dst$$reg); }
2201 else { r2 = $dst$$reg; r1 = HIGH_FROM_LOW($dst$$reg); }
2203 emit_opcode( cbuf, 0x8B ); // Move r1,r2
2204 emit_rm(cbuf, 0x3, r1, r2);
2205 if( $cnt$$constant > 32 ) { // Shift, if not by zero
2206 emit_opcode(cbuf,$primary);
2207 emit_rm(cbuf, 0x3, $secondary, r1);
2208 emit_d8(cbuf,$cnt$$constant-32);
2209 }
2210 emit_opcode(cbuf,0x33); // XOR r2,r2
2211 emit_rm(cbuf, 0x3, r2, r2);
2212 %}
2214 // Clone of RegMem but accepts an extra parameter to access each
2215 // half of a double in memory; it never needs relocation info.
2216 enc_class Mov_MemD_half_to_Reg (immI opcode, memory mem, immI disp_for_half, eRegI rm_reg) %{
2217 emit_opcode(cbuf,$opcode$$constant);
2218 int reg_encoding = $rm_reg$$reg;
2219 int base = $mem$$base;
2220 int index = $mem$$index;
2221 int scale = $mem$$scale;
2222 int displace = $mem$$disp + $disp_for_half$$constant;
2223 bool disp_is_oop = false;
2224 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, disp_is_oop);
2225 %}
2227 // !!!!! Special Custom Code used by MemMove, and stack access instructions !!!!!
2228 //
2229 // Clone of RegMem except the RM-byte's reg/opcode field is an ADLC-time constant
2230 // and it never needs relocation information.
2231 // Frequently used to move data between FPU's Stack Top and memory.
2232 enc_class RMopc_Mem_no_oop (immI rm_opcode, memory mem) %{
2233 int rm_byte_opcode = $rm_opcode$$constant;
2234 int base = $mem$$base;
2235 int index = $mem$$index;
2236 int scale = $mem$$scale;
2237 int displace = $mem$$disp;
2238 assert( !$mem->disp_is_oop(), "No oops here because no relo info allowed" );
2239 encode_RegMem(cbuf, rm_byte_opcode, base, index, scale, displace, false);
2240 %}
2242 enc_class RMopc_Mem (immI rm_opcode, memory mem) %{
2243 int rm_byte_opcode = $rm_opcode$$constant;
2244 int base = $mem$$base;
2245 int index = $mem$$index;
2246 int scale = $mem$$scale;
2247 int displace = $mem$$disp;
2248 bool disp_is_oop = $mem->disp_is_oop(); // disp-as-oop when working with static globals
2249 encode_RegMem(cbuf, rm_byte_opcode, base, index, scale, displace, disp_is_oop);
2250 %}
2252 enc_class RegLea (eRegI dst, eRegI src0, immI src1 ) %{ // emit_reg_lea
2253 int reg_encoding = $dst$$reg;
2254 int base = $src0$$reg; // 0xFFFFFFFF indicates no base
2255 int index = 0x04; // 0x04 indicates no index
2256 int scale = 0x00; // 0x00 indicates no scale
2257 int displace = $src1$$constant; // 0x00 indicates no displacement
2258 bool disp_is_oop = false;
2259 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, disp_is_oop);
2260 %}
2262 enc_class min_enc (eRegI dst, eRegI src) %{ // MIN
2263 // Compare dst,src
2264 emit_opcode(cbuf,0x3B);
2265 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2266 // jmp dst < src around move
2267 emit_opcode(cbuf,0x7C);
2268 emit_d8(cbuf,2);
2269 // move dst,src
2270 emit_opcode(cbuf,0x8B);
2271 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2272 %}
2274 enc_class max_enc (eRegI dst, eRegI src) %{ // MAX
2275 // Compare dst,src
2276 emit_opcode(cbuf,0x3B);
2277 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2278 // jmp dst > src around move
2279 emit_opcode(cbuf,0x7F);
2280 emit_d8(cbuf,2);
2281 // move dst,src
2282 emit_opcode(cbuf,0x8B);
2283 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2284 %}
2286 enc_class enc_FPR_store(memory mem, regDPR src) %{
2287 // If src is FPR1, we can just FST to store it.
2288 // Else we need to FLD it to FPR1, then FSTP to store/pop it.
2289 int reg_encoding = 0x2; // Just store
2290 int base = $mem$$base;
2291 int index = $mem$$index;
2292 int scale = $mem$$scale;
2293 int displace = $mem$$disp;
2294 bool disp_is_oop = $mem->disp_is_oop(); // disp-as-oop when working with static globals
2295 if( $src$$reg != FPR1L_enc ) {
2296 reg_encoding = 0x3; // Store & pop
2297 emit_opcode( cbuf, 0xD9 ); // FLD (i.e., push it)
2298 emit_d8( cbuf, 0xC0-1+$src$$reg );
2299 }
2300 cbuf.set_insts_mark(); // Mark start of opcode for reloc info in mem operand
2301 emit_opcode(cbuf,$primary);
2302 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, disp_is_oop);
2303 %}
2305 enc_class neg_reg(eRegI dst) %{
2306 // NEG $dst
2307 emit_opcode(cbuf,0xF7);
2308 emit_rm(cbuf, 0x3, 0x03, $dst$$reg );
2309 %}
2311 enc_class setLT_reg(eCXRegI dst) %{
2312 // SETLT $dst
2313 emit_opcode(cbuf,0x0F);
2314 emit_opcode(cbuf,0x9C);
2315 emit_rm( cbuf, 0x3, 0x4, $dst$$reg );
2316 %}
2318 enc_class enc_cmpLTP(ncxRegI p, ncxRegI q, ncxRegI y, eCXRegI tmp) %{ // cadd_cmpLT
2319 int tmpReg = $tmp$$reg;
2321 // SUB $p,$q
2322 emit_opcode(cbuf,0x2B);
2323 emit_rm(cbuf, 0x3, $p$$reg, $q$$reg);
2324 // SBB $tmp,$tmp
2325 emit_opcode(cbuf,0x1B);
2326 emit_rm(cbuf, 0x3, tmpReg, tmpReg);
2327 // AND $tmp,$y
2328 emit_opcode(cbuf,0x23);
2329 emit_rm(cbuf, 0x3, tmpReg, $y$$reg);
2330 // ADD $p,$tmp
2331 emit_opcode(cbuf,0x03);
2332 emit_rm(cbuf, 0x3, $p$$reg, tmpReg);
2333 %}
2335 enc_class enc_cmpLTP_mem(eRegI p, eRegI q, memory mem, eCXRegI tmp) %{ // cadd_cmpLT
2336 int tmpReg = $tmp$$reg;
2338 // SUB $p,$q
2339 emit_opcode(cbuf,0x2B);
2340 emit_rm(cbuf, 0x3, $p$$reg, $q$$reg);
2341 // SBB $tmp,$tmp
2342 emit_opcode(cbuf,0x1B);
2343 emit_rm(cbuf, 0x3, tmpReg, tmpReg);
2344 // AND $tmp,$y
2345 cbuf.set_insts_mark(); // Mark start of opcode for reloc info in mem operand
2346 emit_opcode(cbuf,0x23);
2347 int reg_encoding = tmpReg;
2348 int base = $mem$$base;
2349 int index = $mem$$index;
2350 int scale = $mem$$scale;
2351 int displace = $mem$$disp;
2352 bool disp_is_oop = $mem->disp_is_oop();
2353 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, disp_is_oop);
2354 // ADD $p,$tmp
2355 emit_opcode(cbuf,0x03);
2356 emit_rm(cbuf, 0x3, $p$$reg, tmpReg);
2357 %}
2359 enc_class shift_left_long( eRegL dst, eCXRegI shift ) %{
2360 // TEST shift,32
2361 emit_opcode(cbuf,0xF7);
2362 emit_rm(cbuf, 0x3, 0, ECX_enc);
2363 emit_d32(cbuf,0x20);
2364 // JEQ,s small
2365 emit_opcode(cbuf, 0x74);
2366 emit_d8(cbuf, 0x04);
2367 // MOV $dst.hi,$dst.lo
2368 emit_opcode( cbuf, 0x8B );
2369 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), $dst$$reg );
2370 // CLR $dst.lo
2371 emit_opcode(cbuf, 0x33);
2372 emit_rm(cbuf, 0x3, $dst$$reg, $dst$$reg);
2373 // small:
2374 // SHLD $dst.hi,$dst.lo,$shift
2375 emit_opcode(cbuf,0x0F);
2376 emit_opcode(cbuf,0xA5);
2377 emit_rm(cbuf, 0x3, $dst$$reg, HIGH_FROM_LOW($dst$$reg));
2378 // SHL $dst.lo,$shift"
2379 emit_opcode(cbuf,0xD3);
2380 emit_rm(cbuf, 0x3, 0x4, $dst$$reg );
2381 %}
2383 enc_class shift_right_long( eRegL dst, eCXRegI shift ) %{
2384 // TEST shift,32
2385 emit_opcode(cbuf,0xF7);
2386 emit_rm(cbuf, 0x3, 0, ECX_enc);
2387 emit_d32(cbuf,0x20);
2388 // JEQ,s small
2389 emit_opcode(cbuf, 0x74);
2390 emit_d8(cbuf, 0x04);
2391 // MOV $dst.lo,$dst.hi
2392 emit_opcode( cbuf, 0x8B );
2393 emit_rm(cbuf, 0x3, $dst$$reg, HIGH_FROM_LOW($dst$$reg) );
2394 // CLR $dst.hi
2395 emit_opcode(cbuf, 0x33);
2396 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), HIGH_FROM_LOW($dst$$reg));
2397 // small:
2398 // SHRD $dst.lo,$dst.hi,$shift
2399 emit_opcode(cbuf,0x0F);
2400 emit_opcode(cbuf,0xAD);
2401 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), $dst$$reg);
2402 // SHR $dst.hi,$shift"
2403 emit_opcode(cbuf,0xD3);
2404 emit_rm(cbuf, 0x3, 0x5, HIGH_FROM_LOW($dst$$reg) );
2405 %}
2407 enc_class shift_right_arith_long( eRegL dst, eCXRegI shift ) %{
2408 // TEST shift,32
2409 emit_opcode(cbuf,0xF7);
2410 emit_rm(cbuf, 0x3, 0, ECX_enc);
2411 emit_d32(cbuf,0x20);
2412 // JEQ,s small
2413 emit_opcode(cbuf, 0x74);
2414 emit_d8(cbuf, 0x05);
2415 // MOV $dst.lo,$dst.hi
2416 emit_opcode( cbuf, 0x8B );
2417 emit_rm(cbuf, 0x3, $dst$$reg, HIGH_FROM_LOW($dst$$reg) );
2418 // SAR $dst.hi,31
2419 emit_opcode(cbuf, 0xC1);
2420 emit_rm(cbuf, 0x3, 7, HIGH_FROM_LOW($dst$$reg) );
2421 emit_d8(cbuf, 0x1F );
2422 // small:
2423 // SHRD $dst.lo,$dst.hi,$shift
2424 emit_opcode(cbuf,0x0F);
2425 emit_opcode(cbuf,0xAD);
2426 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), $dst$$reg);
2427 // SAR $dst.hi,$shift"
2428 emit_opcode(cbuf,0xD3);
2429 emit_rm(cbuf, 0x3, 0x7, HIGH_FROM_LOW($dst$$reg) );
2430 %}
2433 // ----------------- Encodings for floating point unit -----------------
2434 // May leave result in FPU-TOS or FPU reg depending on opcodes
2435 enc_class OpcReg_FPR(regFPR src) %{ // FMUL, FDIV
2436 $$$emit8$primary;
2437 emit_rm(cbuf, 0x3, $secondary, $src$$reg );
2438 %}
2440 // Pop argument in FPR0 with FSTP ST(0)
2441 enc_class PopFPU() %{
2442 emit_opcode( cbuf, 0xDD );
2443 emit_d8( cbuf, 0xD8 );
2444 %}
2446 // !!!!! equivalent to Pop_Reg_F
2447 enc_class Pop_Reg_DPR( regDPR dst ) %{
2448 emit_opcode( cbuf, 0xDD ); // FSTP ST(i)
2449 emit_d8( cbuf, 0xD8+$dst$$reg );
2450 %}
2452 enc_class Push_Reg_DPR( regDPR dst ) %{
2453 emit_opcode( cbuf, 0xD9 );
2454 emit_d8( cbuf, 0xC0-1+$dst$$reg ); // FLD ST(i-1)
2455 %}
2457 enc_class strictfp_bias1( regDPR dst ) %{
2458 emit_opcode( cbuf, 0xDB ); // FLD m80real
2459 emit_opcode( cbuf, 0x2D );
2460 emit_d32( cbuf, (int)StubRoutines::addr_fpu_subnormal_bias1() );
2461 emit_opcode( cbuf, 0xDE ); // FMULP ST(dst), ST0
2462 emit_opcode( cbuf, 0xC8+$dst$$reg );
2463 %}
2465 enc_class strictfp_bias2( regDPR dst ) %{
2466 emit_opcode( cbuf, 0xDB ); // FLD m80real
2467 emit_opcode( cbuf, 0x2D );
2468 emit_d32( cbuf, (int)StubRoutines::addr_fpu_subnormal_bias2() );
2469 emit_opcode( cbuf, 0xDE ); // FMULP ST(dst), ST0
2470 emit_opcode( cbuf, 0xC8+$dst$$reg );
2471 %}
2473 // Special case for moving an integer register to a stack slot.
2474 enc_class OpcPRegSS( stackSlotI dst, eRegI src ) %{ // RegSS
2475 store_to_stackslot( cbuf, $primary, $src$$reg, $dst$$disp );
2476 %}
2478 // Special case for moving a register to a stack slot.
2479 enc_class RegSS( stackSlotI dst, eRegI src ) %{ // RegSS
2480 // Opcode already emitted
2481 emit_rm( cbuf, 0x02, $src$$reg, ESP_enc ); // R/M byte
2482 emit_rm( cbuf, 0x00, ESP_enc, ESP_enc); // SIB byte
2483 emit_d32(cbuf, $dst$$disp); // Displacement
2484 %}
2486 // Push the integer in stackSlot 'src' onto FP-stack
2487 enc_class Push_Mem_I( memory src ) %{ // FILD [ESP+src]
2488 store_to_stackslot( cbuf, $primary, $secondary, $src$$disp );
2489 %}
2491 // Push FPU's TOS float to a stack-slot, and pop FPU-stack
2492 enc_class Pop_Mem_FPR( stackSlotF dst ) %{ // FSTP_S [ESP+dst]
2493 store_to_stackslot( cbuf, 0xD9, 0x03, $dst$$disp );
2494 %}
2496 // Same as Pop_Mem_F except for opcode
2497 // Push FPU's TOS double to a stack-slot, and pop FPU-stack
2498 enc_class Pop_Mem_DPR( stackSlotD dst ) %{ // FSTP_D [ESP+dst]
2499 store_to_stackslot( cbuf, 0xDD, 0x03, $dst$$disp );
2500 %}
2502 enc_class Pop_Reg_FPR( regFPR dst ) %{
2503 emit_opcode( cbuf, 0xDD ); // FSTP ST(i)
2504 emit_d8( cbuf, 0xD8+$dst$$reg );
2505 %}
2507 enc_class Push_Reg_FPR( regFPR dst ) %{
2508 emit_opcode( cbuf, 0xD9 ); // FLD ST(i-1)
2509 emit_d8( cbuf, 0xC0-1+$dst$$reg );
2510 %}
2512 // Push FPU's float to a stack-slot, and pop FPU-stack
2513 enc_class Pop_Mem_Reg_FPR( stackSlotF dst, regFPR src ) %{
2514 int pop = 0x02;
2515 if ($src$$reg != FPR1L_enc) {
2516 emit_opcode( cbuf, 0xD9 ); // FLD ST(i-1)
2517 emit_d8( cbuf, 0xC0-1+$src$$reg );
2518 pop = 0x03;
2519 }
2520 store_to_stackslot( cbuf, 0xD9, pop, $dst$$disp ); // FST<P>_S [ESP+dst]
2521 %}
2523 // Push FPU's double to a stack-slot, and pop FPU-stack
2524 enc_class Pop_Mem_Reg_DPR( stackSlotD dst, regDPR src ) %{
2525 int pop = 0x02;
2526 if ($src$$reg != FPR1L_enc) {
2527 emit_opcode( cbuf, 0xD9 ); // FLD ST(i-1)
2528 emit_d8( cbuf, 0xC0-1+$src$$reg );
2529 pop = 0x03;
2530 }
2531 store_to_stackslot( cbuf, 0xDD, pop, $dst$$disp ); // FST<P>_D [ESP+dst]
2532 %}
2534 // Push FPU's double to a FPU-stack-slot, and pop FPU-stack
2535 enc_class Pop_Reg_Reg_DPR( regDPR dst, regFPR src ) %{
2536 int pop = 0xD0 - 1; // -1 since we skip FLD
2537 if ($src$$reg != FPR1L_enc) {
2538 emit_opcode( cbuf, 0xD9 ); // FLD ST(src-1)
2539 emit_d8( cbuf, 0xC0-1+$src$$reg );
2540 pop = 0xD8;
2541 }
2542 emit_opcode( cbuf, 0xDD );
2543 emit_d8( cbuf, pop+$dst$$reg ); // FST<P> ST(i)
2544 %}
2547 enc_class Push_Reg_Mod_DPR( regDPR dst, regDPR src) %{
2548 // load dst in FPR0
2549 emit_opcode( cbuf, 0xD9 );
2550 emit_d8( cbuf, 0xC0-1+$dst$$reg );
2551 if ($src$$reg != FPR1L_enc) {
2552 // fincstp
2553 emit_opcode (cbuf, 0xD9);
2554 emit_opcode (cbuf, 0xF7);
2555 // swap src with FPR1:
2556 // FXCH FPR1 with src
2557 emit_opcode(cbuf, 0xD9);
2558 emit_d8(cbuf, 0xC8-1+$src$$reg );
2559 // fdecstp
2560 emit_opcode (cbuf, 0xD9);
2561 emit_opcode (cbuf, 0xF6);
2562 }
2563 %}
2565 enc_class Push_ModD_encoding(regD src0, regD src1) %{
2566 MacroAssembler _masm(&cbuf);
2567 __ subptr(rsp, 8);
2568 __ movdbl(Address(rsp, 0), $src1$$XMMRegister);
2569 __ fld_d(Address(rsp, 0));
2570 __ movdbl(Address(rsp, 0), $src0$$XMMRegister);
2571 __ fld_d(Address(rsp, 0));
2572 %}
2574 enc_class Push_ModF_encoding(regF src0, regF src1) %{
2575 MacroAssembler _masm(&cbuf);
2576 __ subptr(rsp, 4);
2577 __ movflt(Address(rsp, 0), $src1$$XMMRegister);
2578 __ fld_s(Address(rsp, 0));
2579 __ movflt(Address(rsp, 0), $src0$$XMMRegister);
2580 __ fld_s(Address(rsp, 0));
2581 %}
2583 enc_class Push_ResultD(regD dst) %{
2584 MacroAssembler _masm(&cbuf);
2585 __ fstp_d(Address(rsp, 0));
2586 __ movdbl($dst$$XMMRegister, Address(rsp, 0));
2587 __ addptr(rsp, 8);
2588 %}
2590 enc_class Push_ResultF(regF dst, immI d8) %{
2591 MacroAssembler _masm(&cbuf);
2592 __ fstp_s(Address(rsp, 0));
2593 __ movflt($dst$$XMMRegister, Address(rsp, 0));
2594 __ addptr(rsp, $d8$$constant);
2595 %}
2597 enc_class Push_SrcD(regD src) %{
2598 MacroAssembler _masm(&cbuf);
2599 __ subptr(rsp, 8);
2600 __ movdbl(Address(rsp, 0), $src$$XMMRegister);
2601 __ fld_d(Address(rsp, 0));
2602 %}
2604 enc_class push_stack_temp_qword() %{
2605 MacroAssembler _masm(&cbuf);
2606 __ subptr(rsp, 8);
2607 %}
2609 enc_class pop_stack_temp_qword() %{
2610 MacroAssembler _masm(&cbuf);
2611 __ addptr(rsp, 8);
2612 %}
2614 enc_class push_xmm_to_fpr1(regD src) %{
2615 MacroAssembler _masm(&cbuf);
2616 __ movdbl(Address(rsp, 0), $src$$XMMRegister);
2617 __ fld_d(Address(rsp, 0));
2618 %}
2620 // Compute X^Y using Intel's fast hardware instructions, if possible.
2621 // Otherwise return a NaN.
2622 enc_class pow_exp_core_encoding %{
2623 // FPR1 holds Y*ln2(X). Compute FPR1 = 2^(Y*ln2(X))
2624 emit_opcode(cbuf,0xD9); emit_opcode(cbuf,0xC0); // fdup = fld st(0) Q Q
2625 emit_opcode(cbuf,0xD9); emit_opcode(cbuf,0xFC); // frndint int(Q) Q
2626 emit_opcode(cbuf,0xDC); emit_opcode(cbuf,0xE9); // fsub st(1) -= st(0); int(Q) frac(Q)
2627 emit_opcode(cbuf,0xDB); // FISTP [ESP] frac(Q)
2628 emit_opcode(cbuf,0x1C);
2629 emit_d8(cbuf,0x24);
2630 emit_opcode(cbuf,0xD9); emit_opcode(cbuf,0xF0); // f2xm1 2^frac(Q)-1
2631 emit_opcode(cbuf,0xD9); emit_opcode(cbuf,0xE8); // fld1 1 2^frac(Q)-1
2632 emit_opcode(cbuf,0xDE); emit_opcode(cbuf,0xC1); // faddp 2^frac(Q)
2633 emit_opcode(cbuf,0x8B); // mov rax,[esp+0]=int(Q)
2634 encode_RegMem(cbuf, EAX_enc, ESP_enc, 0x4, 0, 0, false);
2635 emit_opcode(cbuf,0xC7); // mov rcx,0xFFFFF800 - overflow mask
2636 emit_rm(cbuf, 0x3, 0x0, ECX_enc);
2637 emit_d32(cbuf,0xFFFFF800);
2638 emit_opcode(cbuf,0x81); // add rax,1023 - the double exponent bias
2639 emit_rm(cbuf, 0x3, 0x0, EAX_enc);
2640 emit_d32(cbuf,1023);
2641 emit_opcode(cbuf,0x8B); // mov rbx,eax
2642 emit_rm(cbuf, 0x3, EBX_enc, EAX_enc);
2643 emit_opcode(cbuf,0xC1); // shl rax,20 - Slide to exponent position
2644 emit_rm(cbuf,0x3,0x4,EAX_enc);
2645 emit_d8(cbuf,20);
2646 emit_opcode(cbuf,0x85); // test rbx,ecx - check for overflow
2647 emit_rm(cbuf, 0x3, EBX_enc, ECX_enc);
2648 emit_opcode(cbuf,0x0F); emit_opcode(cbuf,0x45); // CMOVne rax,ecx - overflow; stuff NAN into EAX
2649 emit_rm(cbuf, 0x3, EAX_enc, ECX_enc);
2650 emit_opcode(cbuf,0x89); // mov [esp+4],eax - Store as part of double word
2651 encode_RegMem(cbuf, EAX_enc, ESP_enc, 0x4, 0, 4, false);
2652 emit_opcode(cbuf,0xC7); // mov [esp+0],0 - [ESP] = (double)(1<<int(Q)) = 2^int(Q)
2653 encode_RegMem(cbuf, 0x0, ESP_enc, 0x4, 0, 0, false);
2654 emit_d32(cbuf,0);
2655 emit_opcode(cbuf,0xDC); // fmul dword st(0),[esp+0]; FPR1 = 2^int(Q)*2^frac(Q) = 2^Q
2656 encode_RegMem(cbuf, 0x1, ESP_enc, 0x4, 0, 0, false);
2657 %}
2659 enc_class Push_Result_Mod_DPR( regDPR src) %{
2660 if ($src$$reg != FPR1L_enc) {
2661 // fincstp
2662 emit_opcode (cbuf, 0xD9);
2663 emit_opcode (cbuf, 0xF7);
2664 // FXCH FPR1 with src
2665 emit_opcode(cbuf, 0xD9);
2666 emit_d8(cbuf, 0xC8-1+$src$$reg );
2667 // fdecstp
2668 emit_opcode (cbuf, 0xD9);
2669 emit_opcode (cbuf, 0xF6);
2670 }
2671 // // following asm replaced with Pop_Reg_F or Pop_Mem_F
2672 // // FSTP FPR$dst$$reg
2673 // emit_opcode( cbuf, 0xDD );
2674 // emit_d8( cbuf, 0xD8+$dst$$reg );
2675 %}
2677 enc_class fnstsw_sahf_skip_parity() %{
2678 // fnstsw ax
2679 emit_opcode( cbuf, 0xDF );
2680 emit_opcode( cbuf, 0xE0 );
2681 // sahf
2682 emit_opcode( cbuf, 0x9E );
2683 // jnp ::skip
2684 emit_opcode( cbuf, 0x7B );
2685 emit_opcode( cbuf, 0x05 );
2686 %}
2688 enc_class emitModDPR() %{
2689 // fprem must be iterative
2690 // :: loop
2691 // fprem
2692 emit_opcode( cbuf, 0xD9 );
2693 emit_opcode( cbuf, 0xF8 );
2694 // wait
2695 emit_opcode( cbuf, 0x9b );
2696 // fnstsw ax
2697 emit_opcode( cbuf, 0xDF );
2698 emit_opcode( cbuf, 0xE0 );
2699 // sahf
2700 emit_opcode( cbuf, 0x9E );
2701 // jp ::loop
2702 emit_opcode( cbuf, 0x0F );
2703 emit_opcode( cbuf, 0x8A );
2704 emit_opcode( cbuf, 0xF4 );
2705 emit_opcode( cbuf, 0xFF );
2706 emit_opcode( cbuf, 0xFF );
2707 emit_opcode( cbuf, 0xFF );
2708 %}
2710 enc_class fpu_flags() %{
2711 // fnstsw_ax
2712 emit_opcode( cbuf, 0xDF);
2713 emit_opcode( cbuf, 0xE0);
2714 // test ax,0x0400
2715 emit_opcode( cbuf, 0x66 ); // operand-size prefix for 16-bit immediate
2716 emit_opcode( cbuf, 0xA9 );
2717 emit_d16 ( cbuf, 0x0400 );
2718 // // // This sequence works, but stalls for 12-16 cycles on PPro
2719 // // test rax,0x0400
2720 // emit_opcode( cbuf, 0xA9 );
2721 // emit_d32 ( cbuf, 0x00000400 );
2722 //
2723 // jz exit (no unordered comparison)
2724 emit_opcode( cbuf, 0x74 );
2725 emit_d8 ( cbuf, 0x02 );
2726 // mov ah,1 - treat as LT case (set carry flag)
2727 emit_opcode( cbuf, 0xB4 );
2728 emit_d8 ( cbuf, 0x01 );
2729 // sahf
2730 emit_opcode( cbuf, 0x9E);
2731 %}
2733 enc_class cmpF_P6_fixup() %{
2734 // Fixup the integer flags in case comparison involved a NaN
2735 //
2736 // JNP exit (no unordered comparison, P-flag is set by NaN)
2737 emit_opcode( cbuf, 0x7B );
2738 emit_d8 ( cbuf, 0x03 );
2739 // MOV AH,1 - treat as LT case (set carry flag)
2740 emit_opcode( cbuf, 0xB4 );
2741 emit_d8 ( cbuf, 0x01 );
2742 // SAHF
2743 emit_opcode( cbuf, 0x9E);
2744 // NOP // target for branch to avoid branch to branch
2745 emit_opcode( cbuf, 0x90);
2746 %}
2748 // fnstsw_ax();
2749 // sahf();
2750 // movl(dst, nan_result);
2751 // jcc(Assembler::parity, exit);
2752 // movl(dst, less_result);
2753 // jcc(Assembler::below, exit);
2754 // movl(dst, equal_result);
2755 // jcc(Assembler::equal, exit);
2756 // movl(dst, greater_result);
2758 // less_result = 1;
2759 // greater_result = -1;
2760 // equal_result = 0;
2761 // nan_result = -1;
2763 enc_class CmpF_Result(eRegI dst) %{
2764 // fnstsw_ax();
2765 emit_opcode( cbuf, 0xDF);
2766 emit_opcode( cbuf, 0xE0);
2767 // sahf
2768 emit_opcode( cbuf, 0x9E);
2769 // movl(dst, nan_result);
2770 emit_opcode( cbuf, 0xB8 + $dst$$reg);
2771 emit_d32( cbuf, -1 );
2772 // jcc(Assembler::parity, exit);
2773 emit_opcode( cbuf, 0x7A );
2774 emit_d8 ( cbuf, 0x13 );
2775 // movl(dst, less_result);
2776 emit_opcode( cbuf, 0xB8 + $dst$$reg);
2777 emit_d32( cbuf, -1 );
2778 // jcc(Assembler::below, exit);
2779 emit_opcode( cbuf, 0x72 );
2780 emit_d8 ( cbuf, 0x0C );
2781 // movl(dst, equal_result);
2782 emit_opcode( cbuf, 0xB8 + $dst$$reg);
2783 emit_d32( cbuf, 0 );
2784 // jcc(Assembler::equal, exit);
2785 emit_opcode( cbuf, 0x74 );
2786 emit_d8 ( cbuf, 0x05 );
2787 // movl(dst, greater_result);
2788 emit_opcode( cbuf, 0xB8 + $dst$$reg);
2789 emit_d32( cbuf, 1 );
2790 %}
2793 // Compare the longs and set flags
2794 // BROKEN! Do Not use as-is
2795 enc_class cmpl_test( eRegL src1, eRegL src2 ) %{
2796 // CMP $src1.hi,$src2.hi
2797 emit_opcode( cbuf, 0x3B );
2798 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($src1$$reg), HIGH_FROM_LOW($src2$$reg) );
2799 // JNE,s done
2800 emit_opcode(cbuf,0x75);
2801 emit_d8(cbuf, 2 );
2802 // CMP $src1.lo,$src2.lo
2803 emit_opcode( cbuf, 0x3B );
2804 emit_rm(cbuf, 0x3, $src1$$reg, $src2$$reg );
2805 // done:
2806 %}
2808 enc_class convert_int_long( regL dst, eRegI src ) %{
2809 // mov $dst.lo,$src
2810 int dst_encoding = $dst$$reg;
2811 int src_encoding = $src$$reg;
2812 encode_Copy( cbuf, dst_encoding , src_encoding );
2813 // mov $dst.hi,$src
2814 encode_Copy( cbuf, HIGH_FROM_LOW(dst_encoding), src_encoding );
2815 // sar $dst.hi,31
2816 emit_opcode( cbuf, 0xC1 );
2817 emit_rm(cbuf, 0x3, 7, HIGH_FROM_LOW(dst_encoding) );
2818 emit_d8(cbuf, 0x1F );
2819 %}
2821 enc_class convert_long_double( eRegL src ) %{
2822 // push $src.hi
2823 emit_opcode(cbuf, 0x50+HIGH_FROM_LOW($src$$reg));
2824 // push $src.lo
2825 emit_opcode(cbuf, 0x50+$src$$reg );
2826 // fild 64-bits at [SP]
2827 emit_opcode(cbuf,0xdf);
2828 emit_d8(cbuf, 0x6C);
2829 emit_d8(cbuf, 0x24);
2830 emit_d8(cbuf, 0x00);
2831 // pop stack
2832 emit_opcode(cbuf, 0x83); // add SP, #8
2833 emit_rm(cbuf, 0x3, 0x00, ESP_enc);
2834 emit_d8(cbuf, 0x8);
2835 %}
2837 enc_class multiply_con_and_shift_high( eDXRegI dst, nadxRegI src1, eADXRegL_low_only src2, immI_32_63 cnt, eFlagsReg cr ) %{
2838 // IMUL EDX:EAX,$src1
2839 emit_opcode( cbuf, 0xF7 );
2840 emit_rm( cbuf, 0x3, 0x5, $src1$$reg );
2841 // SAR EDX,$cnt-32
2842 int shift_count = ((int)$cnt$$constant) - 32;
2843 if (shift_count > 0) {
2844 emit_opcode(cbuf, 0xC1);
2845 emit_rm(cbuf, 0x3, 7, $dst$$reg );
2846 emit_d8(cbuf, shift_count);
2847 }
2848 %}
2850 // this version doesn't have add sp, 8
2851 enc_class convert_long_double2( eRegL src ) %{
2852 // push $src.hi
2853 emit_opcode(cbuf, 0x50+HIGH_FROM_LOW($src$$reg));
2854 // push $src.lo
2855 emit_opcode(cbuf, 0x50+$src$$reg );
2856 // fild 64-bits at [SP]
2857 emit_opcode(cbuf,0xdf);
2858 emit_d8(cbuf, 0x6C);
2859 emit_d8(cbuf, 0x24);
2860 emit_d8(cbuf, 0x00);
2861 %}
2863 enc_class long_int_multiply( eADXRegL dst, nadxRegI src) %{
2864 // Basic idea: long = (long)int * (long)int
2865 // IMUL EDX:EAX, src
2866 emit_opcode( cbuf, 0xF7 );
2867 emit_rm( cbuf, 0x3, 0x5, $src$$reg);
2868 %}
2870 enc_class long_uint_multiply( eADXRegL dst, nadxRegI src) %{
2871 // Basic Idea: long = (int & 0xffffffffL) * (int & 0xffffffffL)
2872 // MUL EDX:EAX, src
2873 emit_opcode( cbuf, 0xF7 );
2874 emit_rm( cbuf, 0x3, 0x4, $src$$reg);
2875 %}
2877 enc_class long_multiply( eADXRegL dst, eRegL src, eRegI tmp ) %{
2878 // Basic idea: lo(result) = lo(x_lo * y_lo)
2879 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi)
2880 // MOV $tmp,$src.lo
2881 encode_Copy( cbuf, $tmp$$reg, $src$$reg );
2882 // IMUL $tmp,EDX
2883 emit_opcode( cbuf, 0x0F );
2884 emit_opcode( cbuf, 0xAF );
2885 emit_rm( cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($dst$$reg) );
2886 // MOV EDX,$src.hi
2887 encode_Copy( cbuf, HIGH_FROM_LOW($dst$$reg), HIGH_FROM_LOW($src$$reg) );
2888 // IMUL EDX,EAX
2889 emit_opcode( cbuf, 0x0F );
2890 emit_opcode( cbuf, 0xAF );
2891 emit_rm( cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), $dst$$reg );
2892 // ADD $tmp,EDX
2893 emit_opcode( cbuf, 0x03 );
2894 emit_rm( cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($dst$$reg) );
2895 // MUL EDX:EAX,$src.lo
2896 emit_opcode( cbuf, 0xF7 );
2897 emit_rm( cbuf, 0x3, 0x4, $src$$reg );
2898 // ADD EDX,ESI
2899 emit_opcode( cbuf, 0x03 );
2900 emit_rm( cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), $tmp$$reg );
2901 %}
2903 enc_class long_multiply_con( eADXRegL dst, immL_127 src, eRegI tmp ) %{
2904 // Basic idea: lo(result) = lo(src * y_lo)
2905 // hi(result) = hi(src * y_lo) + lo(src * y_hi)
2906 // IMUL $tmp,EDX,$src
2907 emit_opcode( cbuf, 0x6B );
2908 emit_rm( cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($dst$$reg) );
2909 emit_d8( cbuf, (int)$src$$constant );
2910 // MOV EDX,$src
2911 emit_opcode(cbuf, 0xB8 + EDX_enc);
2912 emit_d32( cbuf, (int)$src$$constant );
2913 // MUL EDX:EAX,EDX
2914 emit_opcode( cbuf, 0xF7 );
2915 emit_rm( cbuf, 0x3, 0x4, EDX_enc );
2916 // ADD EDX,ESI
2917 emit_opcode( cbuf, 0x03 );
2918 emit_rm( cbuf, 0x3, EDX_enc, $tmp$$reg );
2919 %}
2921 enc_class long_div( eRegL src1, eRegL src2 ) %{
2922 // PUSH src1.hi
2923 emit_opcode(cbuf, HIGH_FROM_LOW(0x50+$src1$$reg) );
2924 // PUSH src1.lo
2925 emit_opcode(cbuf, 0x50+$src1$$reg );
2926 // PUSH src2.hi
2927 emit_opcode(cbuf, HIGH_FROM_LOW(0x50+$src2$$reg) );
2928 // PUSH src2.lo
2929 emit_opcode(cbuf, 0x50+$src2$$reg );
2930 // CALL directly to the runtime
2931 cbuf.set_insts_mark();
2932 emit_opcode(cbuf,0xE8); // Call into runtime
2933 emit_d32_reloc(cbuf, (CAST_FROM_FN_PTR(address, SharedRuntime::ldiv) - cbuf.insts_end()) - 4, runtime_call_Relocation::spec(), RELOC_IMM32 );
2934 // Restore stack
2935 emit_opcode(cbuf, 0x83); // add SP, #framesize
2936 emit_rm(cbuf, 0x3, 0x00, ESP_enc);
2937 emit_d8(cbuf, 4*4);
2938 %}
2940 enc_class long_mod( eRegL src1, eRegL src2 ) %{
2941 // PUSH src1.hi
2942 emit_opcode(cbuf, HIGH_FROM_LOW(0x50+$src1$$reg) );
2943 // PUSH src1.lo
2944 emit_opcode(cbuf, 0x50+$src1$$reg );
2945 // PUSH src2.hi
2946 emit_opcode(cbuf, HIGH_FROM_LOW(0x50+$src2$$reg) );
2947 // PUSH src2.lo
2948 emit_opcode(cbuf, 0x50+$src2$$reg );
2949 // CALL directly to the runtime
2950 cbuf.set_insts_mark();
2951 emit_opcode(cbuf,0xE8); // Call into runtime
2952 emit_d32_reloc(cbuf, (CAST_FROM_FN_PTR(address, SharedRuntime::lrem ) - cbuf.insts_end()) - 4, runtime_call_Relocation::spec(), RELOC_IMM32 );
2953 // Restore stack
2954 emit_opcode(cbuf, 0x83); // add SP, #framesize
2955 emit_rm(cbuf, 0x3, 0x00, ESP_enc);
2956 emit_d8(cbuf, 4*4);
2957 %}
2959 enc_class long_cmp_flags0( eRegL src, eRegI tmp ) %{
2960 // MOV $tmp,$src.lo
2961 emit_opcode(cbuf, 0x8B);
2962 emit_rm(cbuf, 0x3, $tmp$$reg, $src$$reg);
2963 // OR $tmp,$src.hi
2964 emit_opcode(cbuf, 0x0B);
2965 emit_rm(cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($src$$reg));
2966 %}
2968 enc_class long_cmp_flags1( eRegL src1, eRegL src2 ) %{
2969 // CMP $src1.lo,$src2.lo
2970 emit_opcode( cbuf, 0x3B );
2971 emit_rm(cbuf, 0x3, $src1$$reg, $src2$$reg );
2972 // JNE,s skip
2973 emit_cc(cbuf, 0x70, 0x5);
2974 emit_d8(cbuf,2);
2975 // CMP $src1.hi,$src2.hi
2976 emit_opcode( cbuf, 0x3B );
2977 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($src1$$reg), HIGH_FROM_LOW($src2$$reg) );
2978 %}
2980 enc_class long_cmp_flags2( eRegL src1, eRegL src2, eRegI tmp ) %{
2981 // CMP $src1.lo,$src2.lo\t! Long compare; set flags for low bits
2982 emit_opcode( cbuf, 0x3B );
2983 emit_rm(cbuf, 0x3, $src1$$reg, $src2$$reg );
2984 // MOV $tmp,$src1.hi
2985 emit_opcode( cbuf, 0x8B );
2986 emit_rm(cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($src1$$reg) );
2987 // SBB $tmp,$src2.hi\t! Compute flags for long compare
2988 emit_opcode( cbuf, 0x1B );
2989 emit_rm(cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($src2$$reg) );
2990 %}
2992 enc_class long_cmp_flags3( eRegL src, eRegI tmp ) %{
2993 // XOR $tmp,$tmp
2994 emit_opcode(cbuf,0x33); // XOR
2995 emit_rm(cbuf,0x3, $tmp$$reg, $tmp$$reg);
2996 // CMP $tmp,$src.lo
2997 emit_opcode( cbuf, 0x3B );
2998 emit_rm(cbuf, 0x3, $tmp$$reg, $src$$reg );
2999 // SBB $tmp,$src.hi
3000 emit_opcode( cbuf, 0x1B );
3001 emit_rm(cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($src$$reg) );
3002 %}
3004 // Sniff, sniff... smells like Gnu Superoptimizer
3005 enc_class neg_long( eRegL dst ) %{
3006 emit_opcode(cbuf,0xF7); // NEG hi
3007 emit_rm (cbuf,0x3, 0x3, HIGH_FROM_LOW($dst$$reg));
3008 emit_opcode(cbuf,0xF7); // NEG lo
3009 emit_rm (cbuf,0x3, 0x3, $dst$$reg );
3010 emit_opcode(cbuf,0x83); // SBB hi,0
3011 emit_rm (cbuf,0x3, 0x3, HIGH_FROM_LOW($dst$$reg));
3012 emit_d8 (cbuf,0 );
3013 %}
3016 // Because the transitions from emitted code to the runtime
3017 // monitorenter/exit helper stubs are so slow it's critical that
3018 // we inline both the stack-locking fast-path and the inflated fast path.
3019 //
3020 // See also: cmpFastLock and cmpFastUnlock.
3021 //
3022 // What follows is a specialized inline transliteration of the code
3023 // in slow_enter() and slow_exit(). If we're concerned about I$ bloat
3024 // another option would be to emit TrySlowEnter and TrySlowExit methods
3025 // at startup-time. These methods would accept arguments as
3026 // (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure
3027 // indications in the icc.ZFlag. Fast_Lock and Fast_Unlock would simply
3028 // marshal the arguments and emit calls to TrySlowEnter and TrySlowExit.
3029 // In practice, however, the # of lock sites is bounded and is usually small.
3030 // Besides the call overhead, TrySlowEnter and TrySlowExit might suffer
3031 // if the processor uses simple bimodal branch predictors keyed by EIP
3032 // Since the helper routines would be called from multiple synchronization
3033 // sites.
3034 //
3035 // An even better approach would be write "MonitorEnter()" and "MonitorExit()"
3036 // in java - using j.u.c and unsafe - and just bind the lock and unlock sites
3037 // to those specialized methods. That'd give us a mostly platform-independent
3038 // implementation that the JITs could optimize and inline at their pleasure.
3039 // Done correctly, the only time we'd need to cross to native could would be
3040 // to park() or unpark() threads. We'd also need a few more unsafe operators
3041 // to (a) prevent compiler-JIT reordering of non-volatile accesses, and
3042 // (b) explicit barriers or fence operations.
3043 //
3044 // TODO:
3045 //
3046 // * Arrange for C2 to pass "Self" into Fast_Lock and Fast_Unlock in one of the registers (scr).
3047 // This avoids manifesting the Self pointer in the Fast_Lock and Fast_Unlock terminals.
3048 // Given TLAB allocation, Self is usually manifested in a register, so passing it into
3049 // the lock operators would typically be faster than reifying Self.
3050 //
3051 // * Ideally I'd define the primitives as:
3052 // fast_lock (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED.
3053 // fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED
3054 // Unfortunately ADLC bugs prevent us from expressing the ideal form.
3055 // Instead, we're stuck with a rather awkward and brittle register assignments below.
3056 // Furthermore the register assignments are overconstrained, possibly resulting in
3057 // sub-optimal code near the synchronization site.
3058 //
3059 // * Eliminate the sp-proximity tests and just use "== Self" tests instead.
3060 // Alternately, use a better sp-proximity test.
3061 //
3062 // * Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value.
3063 // Either one is sufficient to uniquely identify a thread.
3064 // TODO: eliminate use of sp in _owner and use get_thread(tr) instead.
3065 //
3066 // * Intrinsify notify() and notifyAll() for the common cases where the
3067 // object is locked by the calling thread but the waitlist is empty.
3068 // avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll().
3069 //
3070 // * use jccb and jmpb instead of jcc and jmp to improve code density.
3071 // But beware of excessive branch density on AMD Opterons.
3072 //
3073 // * Both Fast_Lock and Fast_Unlock set the ICC.ZF to indicate success
3074 // or failure of the fast-path. If the fast-path fails then we pass
3075 // control to the slow-path, typically in C. In Fast_Lock and
3076 // Fast_Unlock we often branch to DONE_LABEL, just to find that C2
3077 // will emit a conditional branch immediately after the node.
3078 // So we have branches to branches and lots of ICC.ZF games.
3079 // Instead, it might be better to have C2 pass a "FailureLabel"
3080 // into Fast_Lock and Fast_Unlock. In the case of success, control
3081 // will drop through the node. ICC.ZF is undefined at exit.
3082 // In the case of failure, the node will branch directly to the
3083 // FailureLabel
3086 // obj: object to lock
3087 // box: on-stack box address (displaced header location) - KILLED
3088 // rax,: tmp -- KILLED
3089 // scr: tmp -- KILLED
3090 enc_class Fast_Lock( eRegP obj, eRegP box, eAXRegI tmp, eRegP scr ) %{
3092 Register objReg = as_Register($obj$$reg);
3093 Register boxReg = as_Register($box$$reg);
3094 Register tmpReg = as_Register($tmp$$reg);
3095 Register scrReg = as_Register($scr$$reg);
3097 // Ensure the register assignents are disjoint
3098 guarantee (objReg != boxReg, "") ;
3099 guarantee (objReg != tmpReg, "") ;
3100 guarantee (objReg != scrReg, "") ;
3101 guarantee (boxReg != tmpReg, "") ;
3102 guarantee (boxReg != scrReg, "") ;
3103 guarantee (tmpReg == as_Register(EAX_enc), "") ;
3105 MacroAssembler masm(&cbuf);
3107 if (_counters != NULL) {
3108 masm.atomic_incl(ExternalAddress((address) _counters->total_entry_count_addr()));
3109 }
3110 if (EmitSync & 1) {
3111 // set box->dhw = unused_mark (3)
3112 // Force all sync thru slow-path: slow_enter() and slow_exit()
3113 masm.movptr (Address(boxReg, 0), int32_t(markOopDesc::unused_mark())) ;
3114 masm.cmpptr (rsp, (int32_t)0) ;
3115 } else
3116 if (EmitSync & 2) {
3117 Label DONE_LABEL ;
3118 if (UseBiasedLocking) {
3119 // Note: tmpReg maps to the swap_reg argument and scrReg to the tmp_reg argument.
3120 masm.biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, _counters);
3121 }
3123 masm.movptr(tmpReg, Address(objReg, 0)) ; // fetch markword
3124 masm.orptr (tmpReg, 0x1);
3125 masm.movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS
3126 if (os::is_MP()) { masm.lock(); }
3127 masm.cmpxchgptr(boxReg, Address(objReg, 0)); // Updates tmpReg
3128 masm.jcc(Assembler::equal, DONE_LABEL);
3129 // Recursive locking
3130 masm.subptr(tmpReg, rsp);
3131 masm.andptr(tmpReg, (int32_t) 0xFFFFF003 );
3132 masm.movptr(Address(boxReg, 0), tmpReg);
3133 masm.bind(DONE_LABEL) ;
3134 } else {
3135 // Possible cases that we'll encounter in fast_lock
3136 // ------------------------------------------------
3137 // * Inflated
3138 // -- unlocked
3139 // -- Locked
3140 // = by self
3141 // = by other
3142 // * biased
3143 // -- by Self
3144 // -- by other
3145 // * neutral
3146 // * stack-locked
3147 // -- by self
3148 // = sp-proximity test hits
3149 // = sp-proximity test generates false-negative
3150 // -- by other
3151 //
3153 Label IsInflated, DONE_LABEL, PopDone ;
3155 // TODO: optimize away redundant LDs of obj->mark and improve the markword triage
3156 // order to reduce the number of conditional branches in the most common cases.
3157 // Beware -- there's a subtle invariant that fetch of the markword
3158 // at [FETCH], below, will never observe a biased encoding (*101b).
3159 // If this invariant is not held we risk exclusion (safety) failure.
3160 if (UseBiasedLocking && !UseOptoBiasInlining) {
3161 masm.biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, _counters);
3162 }
3164 masm.movptr(tmpReg, Address(objReg, 0)) ; // [FETCH]
3165 masm.testptr(tmpReg, 0x02) ; // Inflated v (Stack-locked or neutral)
3166 masm.jccb (Assembler::notZero, IsInflated) ;
3168 // Attempt stack-locking ...
3169 masm.orptr (tmpReg, 0x1);
3170 masm.movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS
3171 if (os::is_MP()) { masm.lock(); }
3172 masm.cmpxchgptr(boxReg, Address(objReg, 0)); // Updates tmpReg
3173 if (_counters != NULL) {
3174 masm.cond_inc32(Assembler::equal,
3175 ExternalAddress((address)_counters->fast_path_entry_count_addr()));
3176 }
3177 masm.jccb (Assembler::equal, DONE_LABEL);
3179 // Recursive locking
3180 masm.subptr(tmpReg, rsp);
3181 masm.andptr(tmpReg, 0xFFFFF003 );
3182 masm.movptr(Address(boxReg, 0), tmpReg);
3183 if (_counters != NULL) {
3184 masm.cond_inc32(Assembler::equal,
3185 ExternalAddress((address)_counters->fast_path_entry_count_addr()));
3186 }
3187 masm.jmp (DONE_LABEL) ;
3189 masm.bind (IsInflated) ;
3191 // The object is inflated.
3192 //
3193 // TODO-FIXME: eliminate the ugly use of manifest constants:
3194 // Use markOopDesc::monitor_value instead of "2".
3195 // use markOop::unused_mark() instead of "3".
3196 // The tmpReg value is an objectMonitor reference ORed with
3197 // markOopDesc::monitor_value (2). We can either convert tmpReg to an
3198 // objectmonitor pointer by masking off the "2" bit or we can just
3199 // use tmpReg as an objectmonitor pointer but bias the objectmonitor
3200 // field offsets with "-2" to compensate for and annul the low-order tag bit.
3201 //
3202 // I use the latter as it avoids AGI stalls.
3203 // As such, we write "mov r, [tmpReg+OFFSETOF(Owner)-2]"
3204 // instead of "mov r, [tmpReg+OFFSETOF(Owner)]".
3205 //
3206 #define OFFSET_SKEWED(f) ((ObjectMonitor::f ## _offset_in_bytes())-2)
3208 // boxReg refers to the on-stack BasicLock in the current frame.
3209 // We'd like to write:
3210 // set box->_displaced_header = markOop::unused_mark(). Any non-0 value suffices.
3211 // This is convenient but results a ST-before-CAS penalty. The following CAS suffers
3212 // additional latency as we have another ST in the store buffer that must drain.
3214 if (EmitSync & 8192) {
3215 masm.movptr(Address(boxReg, 0), 3) ; // results in ST-before-CAS penalty
3216 masm.get_thread (scrReg) ;
3217 masm.movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2]
3218 masm.movptr(tmpReg, NULL_WORD); // consider: xor vs mov
3219 if (os::is_MP()) { masm.lock(); }
3220 masm.cmpxchgptr(scrReg, Address(boxReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
3221 } else
3222 if ((EmitSync & 128) == 0) { // avoid ST-before-CAS
3223 masm.movptr(scrReg, boxReg) ;
3224 masm.movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2]
3226 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
3227 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
3228 // prefetchw [eax + Offset(_owner)-2]
3229 masm.prefetchw(Address(rax, ObjectMonitor::owner_offset_in_bytes()-2));
3230 }
3232 if ((EmitSync & 64) == 0) {
3233 // Optimistic form: consider XORL tmpReg,tmpReg
3234 masm.movptr(tmpReg, NULL_WORD) ;
3235 } else {
3236 // Can suffer RTS->RTO upgrades on shared or cold $ lines
3237 // Test-And-CAS instead of CAS
3238 masm.movptr(tmpReg, Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ; // rax, = m->_owner
3239 masm.testptr(tmpReg, tmpReg) ; // Locked ?
3240 masm.jccb (Assembler::notZero, DONE_LABEL) ;
3241 }
3243 // Appears unlocked - try to swing _owner from null to non-null.
3244 // Ideally, I'd manifest "Self" with get_thread and then attempt
3245 // to CAS the register containing Self into m->Owner.
3246 // But we don't have enough registers, so instead we can either try to CAS
3247 // rsp or the address of the box (in scr) into &m->owner. If the CAS succeeds
3248 // we later store "Self" into m->Owner. Transiently storing a stack address
3249 // (rsp or the address of the box) into m->owner is harmless.
3250 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
3251 if (os::is_MP()) { masm.lock(); }
3252 masm.cmpxchgptr(scrReg, Address(boxReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
3253 masm.movptr(Address(scrReg, 0), 3) ; // box->_displaced_header = 3
3254 masm.jccb (Assembler::notZero, DONE_LABEL) ;
3255 masm.get_thread (scrReg) ; // beware: clobbers ICCs
3256 masm.movptr(Address(boxReg, ObjectMonitor::owner_offset_in_bytes()-2), scrReg) ;
3257 masm.xorptr(boxReg, boxReg) ; // set icc.ZFlag = 1 to indicate success
3259 // If the CAS fails we can either retry or pass control to the slow-path.
3260 // We use the latter tactic.
3261 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
3262 // If the CAS was successful ...
3263 // Self has acquired the lock
3264 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
3265 // Intentional fall-through into DONE_LABEL ...
3266 } else {
3267 masm.movptr(Address(boxReg, 0), 3) ; // results in ST-before-CAS penalty
3268 masm.movptr(boxReg, tmpReg) ;
3270 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
3271 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
3272 // prefetchw [eax + Offset(_owner)-2]
3273 masm.prefetchw(Address(rax, ObjectMonitor::owner_offset_in_bytes()-2));
3274 }
3276 if ((EmitSync & 64) == 0) {
3277 // Optimistic form
3278 masm.xorptr (tmpReg, tmpReg) ;
3279 } else {
3280 // Can suffer RTS->RTO upgrades on shared or cold $ lines
3281 masm.movptr(tmpReg, Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ; // rax, = m->_owner
3282 masm.testptr(tmpReg, tmpReg) ; // Locked ?
3283 masm.jccb (Assembler::notZero, DONE_LABEL) ;
3284 }
3286 // Appears unlocked - try to swing _owner from null to non-null.
3287 // Use either "Self" (in scr) or rsp as thread identity in _owner.
3288 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
3289 masm.get_thread (scrReg) ;
3290 if (os::is_MP()) { masm.lock(); }
3291 masm.cmpxchgptr(scrReg, Address(boxReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
3293 // If the CAS fails we can either retry or pass control to the slow-path.
3294 // We use the latter tactic.
3295 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
3296 // If the CAS was successful ...
3297 // Self has acquired the lock
3298 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
3299 // Intentional fall-through into DONE_LABEL ...
3300 }
3302 // DONE_LABEL is a hot target - we'd really like to place it at the
3303 // start of cache line by padding with NOPs.
3304 // See the AMD and Intel software optimization manuals for the
3305 // most efficient "long" NOP encodings.
3306 // Unfortunately none of our alignment mechanisms suffice.
3307 masm.bind(DONE_LABEL);
3309 // Avoid branch-to-branch on AMD processors
3310 // This appears to be superstition.
3311 if (EmitSync & 32) masm.nop() ;
3314 // At DONE_LABEL the icc ZFlag is set as follows ...
3315 // Fast_Unlock uses the same protocol.
3316 // ZFlag == 1 -> Success
3317 // ZFlag == 0 -> Failure - force control through the slow-path
3318 }
3319 %}
3321 // obj: object to unlock
3322 // box: box address (displaced header location), killed. Must be EAX.
3323 // rbx,: killed tmp; cannot be obj nor box.
3324 //
3325 // Some commentary on balanced locking:
3326 //
3327 // Fast_Lock and Fast_Unlock are emitted only for provably balanced lock sites.
3328 // Methods that don't have provably balanced locking are forced to run in the
3329 // interpreter - such methods won't be compiled to use fast_lock and fast_unlock.
3330 // The interpreter provides two properties:
3331 // I1: At return-time the interpreter automatically and quietly unlocks any
3332 // objects acquired the current activation (frame). Recall that the
3333 // interpreter maintains an on-stack list of locks currently held by
3334 // a frame.
3335 // I2: If a method attempts to unlock an object that is not held by the
3336 // the frame the interpreter throws IMSX.
3337 //
3338 // Lets say A(), which has provably balanced locking, acquires O and then calls B().
3339 // B() doesn't have provably balanced locking so it runs in the interpreter.
3340 // Control returns to A() and A() unlocks O. By I1 and I2, above, we know that O
3341 // is still locked by A().
3342 //
3343 // The only other source of unbalanced locking would be JNI. The "Java Native Interface:
3344 // Programmer's Guide and Specification" claims that an object locked by jni_monitorenter
3345 // should not be unlocked by "normal" java-level locking and vice-versa. The specification
3346 // doesn't specify what will occur if a program engages in such mixed-mode locking, however.
3348 enc_class Fast_Unlock( nabxRegP obj, eAXRegP box, eRegP tmp) %{
3350 Register objReg = as_Register($obj$$reg);
3351 Register boxReg = as_Register($box$$reg);
3352 Register tmpReg = as_Register($tmp$$reg);
3354 guarantee (objReg != boxReg, "") ;
3355 guarantee (objReg != tmpReg, "") ;
3356 guarantee (boxReg != tmpReg, "") ;
3357 guarantee (boxReg == as_Register(EAX_enc), "") ;
3358 MacroAssembler masm(&cbuf);
3360 if (EmitSync & 4) {
3361 // Disable - inhibit all inlining. Force control through the slow-path
3362 masm.cmpptr (rsp, 0) ;
3363 } else
3364 if (EmitSync & 8) {
3365 Label DONE_LABEL ;
3366 if (UseBiasedLocking) {
3367 masm.biased_locking_exit(objReg, tmpReg, DONE_LABEL);
3368 }
3369 // classic stack-locking code ...
3370 masm.movptr(tmpReg, Address(boxReg, 0)) ;
3371 masm.testptr(tmpReg, tmpReg) ;
3372 masm.jcc (Assembler::zero, DONE_LABEL) ;
3373 if (os::is_MP()) { masm.lock(); }
3374 masm.cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses EAX which is box
3375 masm.bind(DONE_LABEL);
3376 } else {
3377 Label DONE_LABEL, Stacked, CheckSucc, Inflated ;
3379 // Critically, the biased locking test must have precedence over
3380 // and appear before the (box->dhw == 0) recursive stack-lock test.
3381 if (UseBiasedLocking && !UseOptoBiasInlining) {
3382 masm.biased_locking_exit(objReg, tmpReg, DONE_LABEL);
3383 }
3385 masm.cmpptr(Address(boxReg, 0), 0) ; // Examine the displaced header
3386 masm.movptr(tmpReg, Address(objReg, 0)) ; // Examine the object's markword
3387 masm.jccb (Assembler::zero, DONE_LABEL) ; // 0 indicates recursive stack-lock
3389 masm.testptr(tmpReg, 0x02) ; // Inflated?
3390 masm.jccb (Assembler::zero, Stacked) ;
3392 masm.bind (Inflated) ;
3393 // It's inflated.
3394 // Despite our balanced locking property we still check that m->_owner == Self
3395 // as java routines or native JNI code called by this thread might
3396 // have released the lock.
3397 // Refer to the comments in synchronizer.cpp for how we might encode extra
3398 // state in _succ so we can avoid fetching EntryList|cxq.
3399 //
3400 // I'd like to add more cases in fast_lock() and fast_unlock() --
3401 // such as recursive enter and exit -- but we have to be wary of
3402 // I$ bloat, T$ effects and BP$ effects.
3403 //
3404 // If there's no contention try a 1-0 exit. That is, exit without
3405 // a costly MEMBAR or CAS. See synchronizer.cpp for details on how
3406 // we detect and recover from the race that the 1-0 exit admits.
3407 //
3408 // Conceptually Fast_Unlock() must execute a STST|LDST "release" barrier
3409 // before it STs null into _owner, releasing the lock. Updates
3410 // to data protected by the critical section must be visible before
3411 // we drop the lock (and thus before any other thread could acquire
3412 // the lock and observe the fields protected by the lock).
3413 // IA32's memory-model is SPO, so STs are ordered with respect to
3414 // each other and there's no need for an explicit barrier (fence).
3415 // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
3417 masm.get_thread (boxReg) ;
3418 if ((EmitSync & 4096) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
3419 // prefetchw [ebx + Offset(_owner)-2]
3420 masm.prefetchw(Address(rbx, ObjectMonitor::owner_offset_in_bytes()-2));
3421 }
3423 // Note that we could employ various encoding schemes to reduce
3424 // the number of loads below (currently 4) to just 2 or 3.
3425 // Refer to the comments in synchronizer.cpp.
3426 // In practice the chain of fetches doesn't seem to impact performance, however.
3427 if ((EmitSync & 65536) == 0 && (EmitSync & 256)) {
3428 // Attempt to reduce branch density - AMD's branch predictor.
3429 masm.xorptr(boxReg, Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
3430 masm.orptr(boxReg, Address (tmpReg, ObjectMonitor::recursions_offset_in_bytes()-2)) ;
3431 masm.orptr(boxReg, Address (tmpReg, ObjectMonitor::EntryList_offset_in_bytes()-2)) ;
3432 masm.orptr(boxReg, Address (tmpReg, ObjectMonitor::cxq_offset_in_bytes()-2)) ;
3433 masm.jccb (Assembler::notZero, DONE_LABEL) ;
3434 masm.movptr(Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2), NULL_WORD) ;
3435 masm.jmpb (DONE_LABEL) ;
3436 } else {
3437 masm.xorptr(boxReg, Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
3438 masm.orptr(boxReg, Address (tmpReg, ObjectMonitor::recursions_offset_in_bytes()-2)) ;
3439 masm.jccb (Assembler::notZero, DONE_LABEL) ;
3440 masm.movptr(boxReg, Address (tmpReg, ObjectMonitor::EntryList_offset_in_bytes()-2)) ;
3441 masm.orptr(boxReg, Address (tmpReg, ObjectMonitor::cxq_offset_in_bytes()-2)) ;
3442 masm.jccb (Assembler::notZero, CheckSucc) ;
3443 masm.movptr(Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2), NULL_WORD) ;
3444 masm.jmpb (DONE_LABEL) ;
3445 }
3447 // The Following code fragment (EmitSync & 65536) improves the performance of
3448 // contended applications and contended synchronization microbenchmarks.
3449 // Unfortunately the emission of the code - even though not executed - causes regressions
3450 // in scimark and jetstream, evidently because of $ effects. Replacing the code
3451 // with an equal number of never-executed NOPs results in the same regression.
3452 // We leave it off by default.
3454 if ((EmitSync & 65536) != 0) {
3455 Label LSuccess, LGoSlowPath ;
3457 masm.bind (CheckSucc) ;
3459 // Optional pre-test ... it's safe to elide this
3460 if ((EmitSync & 16) == 0) {
3461 masm.cmpptr(Address (tmpReg, ObjectMonitor::succ_offset_in_bytes()-2), 0) ;
3462 masm.jccb (Assembler::zero, LGoSlowPath) ;
3463 }
3465 // We have a classic Dekker-style idiom:
3466 // ST m->_owner = 0 ; MEMBAR; LD m->_succ
3467 // There are a number of ways to implement the barrier:
3468 // (1) lock:andl &m->_owner, 0
3469 // is fast, but mask doesn't currently support the "ANDL M,IMM32" form.
3470 // LOCK: ANDL [ebx+Offset(_Owner)-2], 0
3471 // Encodes as 81 31 OFF32 IMM32 or 83 63 OFF8 IMM8
3472 // (2) If supported, an explicit MFENCE is appealing.
3473 // In older IA32 processors MFENCE is slower than lock:add or xchg
3474 // particularly if the write-buffer is full as might be the case if
3475 // if stores closely precede the fence or fence-equivalent instruction.
3476 // In more modern implementations MFENCE appears faster, however.
3477 // (3) In lieu of an explicit fence, use lock:addl to the top-of-stack
3478 // The $lines underlying the top-of-stack should be in M-state.
3479 // The locked add instruction is serializing, of course.
3480 // (4) Use xchg, which is serializing
3481 // mov boxReg, 0; xchgl boxReg, [tmpReg + Offset(_owner)-2] also works
3482 // (5) ST m->_owner = 0 and then execute lock:orl &m->_succ, 0.
3483 // The integer condition codes will tell us if succ was 0.
3484 // Since _succ and _owner should reside in the same $line and
3485 // we just stored into _owner, it's likely that the $line
3486 // remains in M-state for the lock:orl.
3487 //
3488 // We currently use (3), although it's likely that switching to (2)
3489 // is correct for the future.
3491 masm.movptr(Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2), NULL_WORD) ;
3492 if (os::is_MP()) {
3493 if (VM_Version::supports_sse2() && 1 == FenceInstruction) {
3494 masm.mfence();
3495 } else {
3496 masm.lock () ; masm.addptr(Address(rsp, 0), 0) ;
3497 }
3498 }
3499 // Ratify _succ remains non-null
3500 masm.cmpptr(Address (tmpReg, ObjectMonitor::succ_offset_in_bytes()-2), 0) ;
3501 masm.jccb (Assembler::notZero, LSuccess) ;
3503 masm.xorptr(boxReg, boxReg) ; // box is really EAX
3504 if (os::is_MP()) { masm.lock(); }
3505 masm.cmpxchgptr(rsp, Address(tmpReg, ObjectMonitor::owner_offset_in_bytes()-2));
3506 masm.jccb (Assembler::notEqual, LSuccess) ;
3507 // Since we're low on registers we installed rsp as a placeholding in _owner.
3508 // Now install Self over rsp. This is safe as we're transitioning from
3509 // non-null to non=null
3510 masm.get_thread (boxReg) ;
3511 masm.movptr(Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2), boxReg) ;
3512 // Intentional fall-through into LGoSlowPath ...
3514 masm.bind (LGoSlowPath) ;
3515 masm.orptr(boxReg, 1) ; // set ICC.ZF=0 to indicate failure
3516 masm.jmpb (DONE_LABEL) ;
3518 masm.bind (LSuccess) ;
3519 masm.xorptr(boxReg, boxReg) ; // set ICC.ZF=1 to indicate success
3520 masm.jmpb (DONE_LABEL) ;
3521 }
3523 masm.bind (Stacked) ;
3524 // It's not inflated and it's not recursively stack-locked and it's not biased.
3525 // It must be stack-locked.
3526 // Try to reset the header to displaced header.
3527 // The "box" value on the stack is stable, so we can reload
3528 // and be assured we observe the same value as above.
3529 masm.movptr(tmpReg, Address(boxReg, 0)) ;
3530 if (os::is_MP()) { masm.lock(); }
3531 masm.cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses EAX which is box
3532 // Intention fall-thru into DONE_LABEL
3535 // DONE_LABEL is a hot target - we'd really like to place it at the
3536 // start of cache line by padding with NOPs.
3537 // See the AMD and Intel software optimization manuals for the
3538 // most efficient "long" NOP encodings.
3539 // Unfortunately none of our alignment mechanisms suffice.
3540 if ((EmitSync & 65536) == 0) {
3541 masm.bind (CheckSucc) ;
3542 }
3543 masm.bind(DONE_LABEL);
3545 // Avoid branch to branch on AMD processors
3546 if (EmitSync & 32768) { masm.nop() ; }
3547 }
3548 %}
3551 enc_class enc_pop_rdx() %{
3552 emit_opcode(cbuf,0x5A);
3553 %}
3555 enc_class enc_rethrow() %{
3556 cbuf.set_insts_mark();
3557 emit_opcode(cbuf, 0xE9); // jmp entry
3558 emit_d32_reloc(cbuf, (int)OptoRuntime::rethrow_stub() - ((int)cbuf.insts_end())-4,
3559 runtime_call_Relocation::spec(), RELOC_IMM32 );
3560 %}
3563 // Convert a double to an int. Java semantics require we do complex
3564 // manglelations in the corner cases. So we set the rounding mode to
3565 // 'zero', store the darned double down as an int, and reset the
3566 // rounding mode to 'nearest'. The hardware throws an exception which
3567 // patches up the correct value directly to the stack.
3568 enc_class DPR2I_encoding( regDPR src ) %{
3569 // Flip to round-to-zero mode. We attempted to allow invalid-op
3570 // exceptions here, so that a NAN or other corner-case value will
3571 // thrown an exception (but normal values get converted at full speed).
3572 // However, I2C adapters and other float-stack manglers leave pending
3573 // invalid-op exceptions hanging. We would have to clear them before
3574 // enabling them and that is more expensive than just testing for the
3575 // invalid value Intel stores down in the corner cases.
3576 emit_opcode(cbuf,0xD9); // FLDCW trunc
3577 emit_opcode(cbuf,0x2D);
3578 emit_d32(cbuf,(int)StubRoutines::addr_fpu_cntrl_wrd_trunc());
3579 // Allocate a word
3580 emit_opcode(cbuf,0x83); // SUB ESP,4
3581 emit_opcode(cbuf,0xEC);
3582 emit_d8(cbuf,0x04);
3583 // Encoding assumes a double has been pushed into FPR0.
3584 // Store down the double as an int, popping the FPU stack
3585 emit_opcode(cbuf,0xDB); // FISTP [ESP]
3586 emit_opcode(cbuf,0x1C);
3587 emit_d8(cbuf,0x24);
3588 // Restore the rounding mode; mask the exception
3589 emit_opcode(cbuf,0xD9); // FLDCW std/24-bit mode
3590 emit_opcode(cbuf,0x2D);
3591 emit_d32( cbuf, Compile::current()->in_24_bit_fp_mode()
3592 ? (int)StubRoutines::addr_fpu_cntrl_wrd_24()
3593 : (int)StubRoutines::addr_fpu_cntrl_wrd_std());
3595 // Load the converted int; adjust CPU stack
3596 emit_opcode(cbuf,0x58); // POP EAX
3597 emit_opcode(cbuf,0x3D); // CMP EAX,imm
3598 emit_d32 (cbuf,0x80000000); // 0x80000000
3599 emit_opcode(cbuf,0x75); // JNE around_slow_call
3600 emit_d8 (cbuf,0x07); // Size of slow_call
3601 // Push src onto stack slow-path
3602 emit_opcode(cbuf,0xD9 ); // FLD ST(i)
3603 emit_d8 (cbuf,0xC0-1+$src$$reg );
3604 // CALL directly to the runtime
3605 cbuf.set_insts_mark();
3606 emit_opcode(cbuf,0xE8); // Call into runtime
3607 emit_d32_reloc(cbuf, (StubRoutines::d2i_wrapper() - cbuf.insts_end()) - 4, runtime_call_Relocation::spec(), RELOC_IMM32 );
3608 // Carry on here...
3609 %}
3611 enc_class DPR2L_encoding( regDPR src ) %{
3612 emit_opcode(cbuf,0xD9); // FLDCW trunc
3613 emit_opcode(cbuf,0x2D);
3614 emit_d32(cbuf,(int)StubRoutines::addr_fpu_cntrl_wrd_trunc());
3615 // Allocate a word
3616 emit_opcode(cbuf,0x83); // SUB ESP,8
3617 emit_opcode(cbuf,0xEC);
3618 emit_d8(cbuf,0x08);
3619 // Encoding assumes a double has been pushed into FPR0.
3620 // Store down the double as a long, popping the FPU stack
3621 emit_opcode(cbuf,0xDF); // FISTP [ESP]
3622 emit_opcode(cbuf,0x3C);
3623 emit_d8(cbuf,0x24);
3624 // Restore the rounding mode; mask the exception
3625 emit_opcode(cbuf,0xD9); // FLDCW std/24-bit mode
3626 emit_opcode(cbuf,0x2D);
3627 emit_d32( cbuf, Compile::current()->in_24_bit_fp_mode()
3628 ? (int)StubRoutines::addr_fpu_cntrl_wrd_24()
3629 : (int)StubRoutines::addr_fpu_cntrl_wrd_std());
3631 // Load the converted int; adjust CPU stack
3632 emit_opcode(cbuf,0x58); // POP EAX
3633 emit_opcode(cbuf,0x5A); // POP EDX
3634 emit_opcode(cbuf,0x81); // CMP EDX,imm
3635 emit_d8 (cbuf,0xFA); // rdx
3636 emit_d32 (cbuf,0x80000000); // 0x80000000
3637 emit_opcode(cbuf,0x75); // JNE around_slow_call
3638 emit_d8 (cbuf,0x07+4); // Size of slow_call
3639 emit_opcode(cbuf,0x85); // TEST EAX,EAX
3640 emit_opcode(cbuf,0xC0); // 2/rax,/rax,
3641 emit_opcode(cbuf,0x75); // JNE around_slow_call
3642 emit_d8 (cbuf,0x07); // Size of slow_call
3643 // Push src onto stack slow-path
3644 emit_opcode(cbuf,0xD9 ); // FLD ST(i)
3645 emit_d8 (cbuf,0xC0-1+$src$$reg );
3646 // CALL directly to the runtime
3647 cbuf.set_insts_mark();
3648 emit_opcode(cbuf,0xE8); // Call into runtime
3649 emit_d32_reloc(cbuf, (StubRoutines::d2l_wrapper() - cbuf.insts_end()) - 4, runtime_call_Relocation::spec(), RELOC_IMM32 );
3650 // Carry on here...
3651 %}
3653 enc_class FMul_ST_reg( eRegFPR src1 ) %{
3654 // Operand was loaded from memory into fp ST (stack top)
3655 // FMUL ST,$src /* D8 C8+i */
3656 emit_opcode(cbuf, 0xD8);
3657 emit_opcode(cbuf, 0xC8 + $src1$$reg);
3658 %}
3660 enc_class FAdd_ST_reg( eRegFPR src2 ) %{
3661 // FADDP ST,src2 /* D8 C0+i */
3662 emit_opcode(cbuf, 0xD8);
3663 emit_opcode(cbuf, 0xC0 + $src2$$reg);
3664 //could use FADDP src2,fpST /* DE C0+i */
3665 %}
3667 enc_class FAddP_reg_ST( eRegFPR src2 ) %{
3668 // FADDP src2,ST /* DE C0+i */
3669 emit_opcode(cbuf, 0xDE);
3670 emit_opcode(cbuf, 0xC0 + $src2$$reg);
3671 %}
3673 enc_class subFPR_divFPR_encode( eRegFPR src1, eRegFPR src2) %{
3674 // Operand has been loaded into fp ST (stack top)
3675 // FSUB ST,$src1
3676 emit_opcode(cbuf, 0xD8);
3677 emit_opcode(cbuf, 0xE0 + $src1$$reg);
3679 // FDIV
3680 emit_opcode(cbuf, 0xD8);
3681 emit_opcode(cbuf, 0xF0 + $src2$$reg);
3682 %}
3684 enc_class MulFAddF (eRegFPR src1, eRegFPR src2) %{
3685 // Operand was loaded from memory into fp ST (stack top)
3686 // FADD ST,$src /* D8 C0+i */
3687 emit_opcode(cbuf, 0xD8);
3688 emit_opcode(cbuf, 0xC0 + $src1$$reg);
3690 // FMUL ST,src2 /* D8 C*+i */
3691 emit_opcode(cbuf, 0xD8);
3692 emit_opcode(cbuf, 0xC8 + $src2$$reg);
3693 %}
3696 enc_class MulFAddFreverse (eRegFPR src1, eRegFPR src2) %{
3697 // Operand was loaded from memory into fp ST (stack top)
3698 // FADD ST,$src /* D8 C0+i */
3699 emit_opcode(cbuf, 0xD8);
3700 emit_opcode(cbuf, 0xC0 + $src1$$reg);
3702 // FMULP src2,ST /* DE C8+i */
3703 emit_opcode(cbuf, 0xDE);
3704 emit_opcode(cbuf, 0xC8 + $src2$$reg);
3705 %}
3707 // Atomically load the volatile long
3708 enc_class enc_loadL_volatile( memory mem, stackSlotL dst ) %{
3709 emit_opcode(cbuf,0xDF);
3710 int rm_byte_opcode = 0x05;
3711 int base = $mem$$base;
3712 int index = $mem$$index;
3713 int scale = $mem$$scale;
3714 int displace = $mem$$disp;
3715 bool disp_is_oop = $mem->disp_is_oop(); // disp-as-oop when working with static globals
3716 encode_RegMem(cbuf, rm_byte_opcode, base, index, scale, displace, disp_is_oop);
3717 store_to_stackslot( cbuf, 0x0DF, 0x07, $dst$$disp );
3718 %}
3720 // Volatile Store Long. Must be atomic, so move it into
3721 // the FP TOS and then do a 64-bit FIST. Has to probe the
3722 // target address before the store (for null-ptr checks)
3723 // so the memory operand is used twice in the encoding.
3724 enc_class enc_storeL_volatile( memory mem, stackSlotL src ) %{
3725 store_to_stackslot( cbuf, 0x0DF, 0x05, $src$$disp );
3726 cbuf.set_insts_mark(); // Mark start of FIST in case $mem has an oop
3727 emit_opcode(cbuf,0xDF);
3728 int rm_byte_opcode = 0x07;
3729 int base = $mem$$base;
3730 int index = $mem$$index;
3731 int scale = $mem$$scale;
3732 int displace = $mem$$disp;
3733 bool disp_is_oop = $mem->disp_is_oop(); // disp-as-oop when working with static globals
3734 encode_RegMem(cbuf, rm_byte_opcode, base, index, scale, displace, disp_is_oop);
3735 %}
3737 // Safepoint Poll. This polls the safepoint page, and causes an
3738 // exception if it is not readable. Unfortunately, it kills the condition code
3739 // in the process
3740 // We current use TESTL [spp],EDI
3741 // A better choice might be TESTB [spp + pagesize() - CacheLineSize()],0
3743 enc_class Safepoint_Poll() %{
3744 cbuf.relocate(cbuf.insts_mark(), relocInfo::poll_type, 0);
3745 emit_opcode(cbuf,0x85);
3746 emit_rm (cbuf, 0x0, 0x7, 0x5);
3747 emit_d32(cbuf, (intptr_t)os::get_polling_page());
3748 %}
3749 %}
3752 //----------FRAME--------------------------------------------------------------
3753 // Definition of frame structure and management information.
3754 //
3755 // S T A C K L A Y O U T Allocators stack-slot number
3756 // | (to get allocators register number
3757 // G Owned by | | v add OptoReg::stack0())
3758 // r CALLER | |
3759 // o | +--------+ pad to even-align allocators stack-slot
3760 // w V | pad0 | numbers; owned by CALLER
3761 // t -----------+--------+----> Matcher::_in_arg_limit, unaligned
3762 // h ^ | in | 5
3763 // | | args | 4 Holes in incoming args owned by SELF
3764 // | | | | 3
3765 // | | +--------+
3766 // V | | old out| Empty on Intel, window on Sparc
3767 // | old |preserve| Must be even aligned.
3768 // | SP-+--------+----> Matcher::_old_SP, even aligned
3769 // | | in | 3 area for Intel ret address
3770 // Owned by |preserve| Empty on Sparc.
3771 // SELF +--------+
3772 // | | pad2 | 2 pad to align old SP
3773 // | +--------+ 1
3774 // | | locks | 0
3775 // | +--------+----> OptoReg::stack0(), even aligned
3776 // | | pad1 | 11 pad to align new SP
3777 // | +--------+
3778 // | | | 10
3779 // | | spills | 9 spills
3780 // V | | 8 (pad0 slot for callee)
3781 // -----------+--------+----> Matcher::_out_arg_limit, unaligned
3782 // ^ | out | 7
3783 // | | args | 6 Holes in outgoing args owned by CALLEE
3784 // Owned by +--------+
3785 // CALLEE | new out| 6 Empty on Intel, window on Sparc
3786 // | new |preserve| Must be even-aligned.
3787 // | SP-+--------+----> Matcher::_new_SP, even aligned
3788 // | | |
3789 //
3790 // Note 1: Only region 8-11 is determined by the allocator. Region 0-5 is
3791 // known from SELF's arguments and the Java calling convention.
3792 // Region 6-7 is determined per call site.
3793 // Note 2: If the calling convention leaves holes in the incoming argument
3794 // area, those holes are owned by SELF. Holes in the outgoing area
3795 // are owned by the CALLEE. Holes should not be nessecary in the
3796 // incoming area, as the Java calling convention is completely under
3797 // the control of the AD file. Doubles can be sorted and packed to
3798 // avoid holes. Holes in the outgoing arguments may be nessecary for
3799 // varargs C calling conventions.
3800 // Note 3: Region 0-3 is even aligned, with pad2 as needed. Region 3-5 is
3801 // even aligned with pad0 as needed.
3802 // Region 6 is even aligned. Region 6-7 is NOT even aligned;
3803 // region 6-11 is even aligned; it may be padded out more so that
3804 // the region from SP to FP meets the minimum stack alignment.
3806 frame %{
3807 // What direction does stack grow in (assumed to be same for C & Java)
3808 stack_direction(TOWARDS_LOW);
3810 // These three registers define part of the calling convention
3811 // between compiled code and the interpreter.
3812 inline_cache_reg(EAX); // Inline Cache Register
3813 interpreter_method_oop_reg(EBX); // Method Oop Register when calling interpreter
3815 // Optional: name the operand used by cisc-spilling to access [stack_pointer + offset]
3816 cisc_spilling_operand_name(indOffset32);
3818 // Number of stack slots consumed by locking an object
3819 sync_stack_slots(1);
3821 // Compiled code's Frame Pointer
3822 frame_pointer(ESP);
3823 // Interpreter stores its frame pointer in a register which is
3824 // stored to the stack by I2CAdaptors.
3825 // I2CAdaptors convert from interpreted java to compiled java.
3826 interpreter_frame_pointer(EBP);
3828 // Stack alignment requirement
3829 // Alignment size in bytes (128-bit -> 16 bytes)
3830 stack_alignment(StackAlignmentInBytes);
3832 // Number of stack slots between incoming argument block and the start of
3833 // a new frame. The PROLOG must add this many slots to the stack. The
3834 // EPILOG must remove this many slots. Intel needs one slot for
3835 // return address and one for rbp, (must save rbp)
3836 in_preserve_stack_slots(2+VerifyStackAtCalls);
3838 // Number of outgoing stack slots killed above the out_preserve_stack_slots
3839 // for calls to C. Supports the var-args backing area for register parms.
3840 varargs_C_out_slots_killed(0);
3842 // The after-PROLOG location of the return address. Location of
3843 // return address specifies a type (REG or STACK) and a number
3844 // representing the register number (i.e. - use a register name) or
3845 // stack slot.
3846 // Ret Addr is on stack in slot 0 if no locks or verification or alignment.
3847 // Otherwise, it is above the locks and verification slot and alignment word
3848 return_addr(STACK - 1 +
3849 round_to(1+VerifyStackAtCalls+
3850 Compile::current()->fixed_slots(),
3851 (StackAlignmentInBytes/wordSize)));
3853 // Body of function which returns an integer array locating
3854 // arguments either in registers or in stack slots. Passed an array
3855 // of ideal registers called "sig" and a "length" count. Stack-slot
3856 // offsets are based on outgoing arguments, i.e. a CALLER setting up
3857 // arguments for a CALLEE. Incoming stack arguments are
3858 // automatically biased by the preserve_stack_slots field above.
3859 calling_convention %{
3860 // No difference between ingoing/outgoing just pass false
3861 SharedRuntime::java_calling_convention(sig_bt, regs, length, false);
3862 %}
3865 // Body of function which returns an integer array locating
3866 // arguments either in registers or in stack slots. Passed an array
3867 // of ideal registers called "sig" and a "length" count. Stack-slot
3868 // offsets are based on outgoing arguments, i.e. a CALLER setting up
3869 // arguments for a CALLEE. Incoming stack arguments are
3870 // automatically biased by the preserve_stack_slots field above.
3871 c_calling_convention %{
3872 // This is obviously always outgoing
3873 (void) SharedRuntime::c_calling_convention(sig_bt, regs, length);
3874 %}
3876 // Location of C & interpreter return values
3877 c_return_value %{
3878 assert( ideal_reg >= Op_RegI && ideal_reg <= Op_RegL, "only return normal values" );
3879 static int lo[Op_RegL+1] = { 0, 0, OptoReg::Bad, EAX_num, EAX_num, FPR1L_num, FPR1L_num, EAX_num };
3880 static int hi[Op_RegL+1] = { 0, 0, OptoReg::Bad, OptoReg::Bad, OptoReg::Bad, OptoReg::Bad, FPR1H_num, EDX_num };
3882 // in SSE2+ mode we want to keep the FPU stack clean so pretend
3883 // that C functions return float and double results in XMM0.
3884 if( ideal_reg == Op_RegD && UseSSE>=2 )
3885 return OptoRegPair(XMM0b_num,XMM0a_num);
3886 if( ideal_reg == Op_RegF && UseSSE>=2 )
3887 return OptoRegPair(OptoReg::Bad,XMM0a_num);
3889 return OptoRegPair(hi[ideal_reg],lo[ideal_reg]);
3890 %}
3892 // Location of return values
3893 return_value %{
3894 assert( ideal_reg >= Op_RegI && ideal_reg <= Op_RegL, "only return normal values" );
3895 static int lo[Op_RegL+1] = { 0, 0, OptoReg::Bad, EAX_num, EAX_num, FPR1L_num, FPR1L_num, EAX_num };
3896 static int hi[Op_RegL+1] = { 0, 0, OptoReg::Bad, OptoReg::Bad, OptoReg::Bad, OptoReg::Bad, FPR1H_num, EDX_num };
3897 if( ideal_reg == Op_RegD && UseSSE>=2 )
3898 return OptoRegPair(XMM0b_num,XMM0a_num);
3899 if( ideal_reg == Op_RegF && UseSSE>=1 )
3900 return OptoRegPair(OptoReg::Bad,XMM0a_num);
3901 return OptoRegPair(hi[ideal_reg],lo[ideal_reg]);
3902 %}
3904 %}
3906 //----------ATTRIBUTES---------------------------------------------------------
3907 //----------Operand Attributes-------------------------------------------------
3908 op_attrib op_cost(0); // Required cost attribute
3910 //----------Instruction Attributes---------------------------------------------
3911 ins_attrib ins_cost(100); // Required cost attribute
3912 ins_attrib ins_size(8); // Required size attribute (in bits)
3913 ins_attrib ins_short_branch(0); // Required flag: is this instruction a
3914 // non-matching short branch variant of some
3915 // long branch?
3916 ins_attrib ins_alignment(1); // Required alignment attribute (must be a power of 2)
3917 // specifies the alignment that some part of the instruction (not
3918 // necessarily the start) requires. If > 1, a compute_padding()
3919 // function must be provided for the instruction
3921 //----------OPERANDS-----------------------------------------------------------
3922 // Operand definitions must precede instruction definitions for correct parsing
3923 // in the ADLC because operands constitute user defined types which are used in
3924 // instruction definitions.
3926 //----------Simple Operands----------------------------------------------------
3927 // Immediate Operands
3928 // Integer Immediate
3929 operand immI() %{
3930 match(ConI);
3932 op_cost(10);
3933 format %{ %}
3934 interface(CONST_INTER);
3935 %}
3937 // Constant for test vs zero
3938 operand immI0() %{
3939 predicate(n->get_int() == 0);
3940 match(ConI);
3942 op_cost(0);
3943 format %{ %}
3944 interface(CONST_INTER);
3945 %}
3947 // Constant for increment
3948 operand immI1() %{
3949 predicate(n->get_int() == 1);
3950 match(ConI);
3952 op_cost(0);
3953 format %{ %}
3954 interface(CONST_INTER);
3955 %}
3957 // Constant for decrement
3958 operand immI_M1() %{
3959 predicate(n->get_int() == -1);
3960 match(ConI);
3962 op_cost(0);
3963 format %{ %}
3964 interface(CONST_INTER);
3965 %}
3967 // Valid scale values for addressing modes
3968 operand immI2() %{
3969 predicate(0 <= n->get_int() && (n->get_int() <= 3));
3970 match(ConI);
3972 format %{ %}
3973 interface(CONST_INTER);
3974 %}
3976 operand immI8() %{
3977 predicate((-128 <= n->get_int()) && (n->get_int() <= 127));
3978 match(ConI);
3980 op_cost(5);
3981 format %{ %}
3982 interface(CONST_INTER);
3983 %}
3985 operand immI16() %{
3986 predicate((-32768 <= n->get_int()) && (n->get_int() <= 32767));
3987 match(ConI);
3989 op_cost(10);
3990 format %{ %}
3991 interface(CONST_INTER);
3992 %}
3994 // Constant for long shifts
3995 operand immI_32() %{
3996 predicate( n->get_int() == 32 );
3997 match(ConI);
3999 op_cost(0);
4000 format %{ %}
4001 interface(CONST_INTER);
4002 %}
4004 operand immI_1_31() %{
4005 predicate( n->get_int() >= 1 && n->get_int() <= 31 );
4006 match(ConI);
4008 op_cost(0);
4009 format %{ %}
4010 interface(CONST_INTER);
4011 %}
4013 operand immI_32_63() %{
4014 predicate( n->get_int() >= 32 && n->get_int() <= 63 );
4015 match(ConI);
4016 op_cost(0);
4018 format %{ %}
4019 interface(CONST_INTER);
4020 %}
4022 operand immI_1() %{
4023 predicate( n->get_int() == 1 );
4024 match(ConI);
4026 op_cost(0);
4027 format %{ %}
4028 interface(CONST_INTER);
4029 %}
4031 operand immI_2() %{
4032 predicate( n->get_int() == 2 );
4033 match(ConI);
4035 op_cost(0);
4036 format %{ %}
4037 interface(CONST_INTER);
4038 %}
4040 operand immI_3() %{
4041 predicate( n->get_int() == 3 );
4042 match(ConI);
4044 op_cost(0);
4045 format %{ %}
4046 interface(CONST_INTER);
4047 %}
4049 // Pointer Immediate
4050 operand immP() %{
4051 match(ConP);
4053 op_cost(10);
4054 format %{ %}
4055 interface(CONST_INTER);
4056 %}
4058 // NULL Pointer Immediate
4059 operand immP0() %{
4060 predicate( n->get_ptr() == 0 );
4061 match(ConP);
4062 op_cost(0);
4064 format %{ %}
4065 interface(CONST_INTER);
4066 %}
4068 // Long Immediate
4069 operand immL() %{
4070 match(ConL);
4072 op_cost(20);
4073 format %{ %}
4074 interface(CONST_INTER);
4075 %}
4077 // Long Immediate zero
4078 operand immL0() %{
4079 predicate( n->get_long() == 0L );
4080 match(ConL);
4081 op_cost(0);
4083 format %{ %}
4084 interface(CONST_INTER);
4085 %}
4087 // Long Immediate zero
4088 operand immL_M1() %{
4089 predicate( n->get_long() == -1L );
4090 match(ConL);
4091 op_cost(0);
4093 format %{ %}
4094 interface(CONST_INTER);
4095 %}
4097 // Long immediate from 0 to 127.
4098 // Used for a shorter form of long mul by 10.
4099 operand immL_127() %{
4100 predicate((0 <= n->get_long()) && (n->get_long() <= 127));
4101 match(ConL);
4102 op_cost(0);
4104 format %{ %}
4105 interface(CONST_INTER);
4106 %}
4108 // Long Immediate: low 32-bit mask
4109 operand immL_32bits() %{
4110 predicate(n->get_long() == 0xFFFFFFFFL);
4111 match(ConL);
4112 op_cost(0);
4114 format %{ %}
4115 interface(CONST_INTER);
4116 %}
4118 // Long Immediate: low 32-bit mask
4119 operand immL32() %{
4120 predicate(n->get_long() == (int)(n->get_long()));
4121 match(ConL);
4122 op_cost(20);
4124 format %{ %}
4125 interface(CONST_INTER);
4126 %}
4128 //Double Immediate zero
4129 operand immDPR0() %{
4130 // Do additional (and counter-intuitive) test against NaN to work around VC++
4131 // bug that generates code such that NaNs compare equal to 0.0
4132 predicate( UseSSE<=1 && n->getd() == 0.0 && !g_isnan(n->getd()) );
4133 match(ConD);
4135 op_cost(5);
4136 format %{ %}
4137 interface(CONST_INTER);
4138 %}
4140 // Double Immediate one
4141 operand immDPR1() %{
4142 predicate( UseSSE<=1 && n->getd() == 1.0 );
4143 match(ConD);
4145 op_cost(5);
4146 format %{ %}
4147 interface(CONST_INTER);
4148 %}
4150 // Double Immediate
4151 operand immDPR() %{
4152 predicate(UseSSE<=1);
4153 match(ConD);
4155 op_cost(5);
4156 format %{ %}
4157 interface(CONST_INTER);
4158 %}
4160 operand immD() %{
4161 predicate(UseSSE>=2);
4162 match(ConD);
4164 op_cost(5);
4165 format %{ %}
4166 interface(CONST_INTER);
4167 %}
4169 // Double Immediate zero
4170 operand immD0() %{
4171 // Do additional (and counter-intuitive) test against NaN to work around VC++
4172 // bug that generates code such that NaNs compare equal to 0.0 AND do not
4173 // compare equal to -0.0.
4174 predicate( UseSSE>=2 && jlong_cast(n->getd()) == 0 );
4175 match(ConD);
4177 format %{ %}
4178 interface(CONST_INTER);
4179 %}
4181 // Float Immediate zero
4182 operand immFPR0() %{
4183 predicate(UseSSE == 0 && n->getf() == 0.0F);
4184 match(ConF);
4186 op_cost(5);
4187 format %{ %}
4188 interface(CONST_INTER);
4189 %}
4191 // Float Immediate one
4192 operand immFPR1() %{
4193 predicate(UseSSE == 0 && n->getf() == 1.0F);
4194 match(ConF);
4196 op_cost(5);
4197 format %{ %}
4198 interface(CONST_INTER);
4199 %}
4201 // Float Immediate
4202 operand immFPR() %{
4203 predicate( UseSSE == 0 );
4204 match(ConF);
4206 op_cost(5);
4207 format %{ %}
4208 interface(CONST_INTER);
4209 %}
4211 // Float Immediate
4212 operand immF() %{
4213 predicate(UseSSE >= 1);
4214 match(ConF);
4216 op_cost(5);
4217 format %{ %}
4218 interface(CONST_INTER);
4219 %}
4221 // Float Immediate zero. Zero and not -0.0
4222 operand immF0() %{
4223 predicate( UseSSE >= 1 && jint_cast(n->getf()) == 0 );
4224 match(ConF);
4226 op_cost(5);
4227 format %{ %}
4228 interface(CONST_INTER);
4229 %}
4231 // Immediates for special shifts (sign extend)
4233 // Constants for increment
4234 operand immI_16() %{
4235 predicate( n->get_int() == 16 );
4236 match(ConI);
4238 format %{ %}
4239 interface(CONST_INTER);
4240 %}
4242 operand immI_24() %{
4243 predicate( n->get_int() == 24 );
4244 match(ConI);
4246 format %{ %}
4247 interface(CONST_INTER);
4248 %}
4250 // Constant for byte-wide masking
4251 operand immI_255() %{
4252 predicate( n->get_int() == 255 );
4253 match(ConI);
4255 format %{ %}
4256 interface(CONST_INTER);
4257 %}
4259 // Constant for short-wide masking
4260 operand immI_65535() %{
4261 predicate(n->get_int() == 65535);
4262 match(ConI);
4264 format %{ %}
4265 interface(CONST_INTER);
4266 %}
4268 // Register Operands
4269 // Integer Register
4270 operand eRegI() %{
4271 constraint(ALLOC_IN_RC(e_reg));
4272 match(RegI);
4273 match(xRegI);
4274 match(eAXRegI);
4275 match(eBXRegI);
4276 match(eCXRegI);
4277 match(eDXRegI);
4278 match(eDIRegI);
4279 match(eSIRegI);
4281 format %{ %}
4282 interface(REG_INTER);
4283 %}
4285 // Subset of Integer Register
4286 operand xRegI(eRegI reg) %{
4287 constraint(ALLOC_IN_RC(x_reg));
4288 match(reg);
4289 match(eAXRegI);
4290 match(eBXRegI);
4291 match(eCXRegI);
4292 match(eDXRegI);
4294 format %{ %}
4295 interface(REG_INTER);
4296 %}
4298 // Special Registers
4299 operand eAXRegI(xRegI reg) %{
4300 constraint(ALLOC_IN_RC(eax_reg));
4301 match(reg);
4302 match(eRegI);
4304 format %{ "EAX" %}
4305 interface(REG_INTER);
4306 %}
4308 // Special Registers
4309 operand eBXRegI(xRegI reg) %{
4310 constraint(ALLOC_IN_RC(ebx_reg));
4311 match(reg);
4312 match(eRegI);
4314 format %{ "EBX" %}
4315 interface(REG_INTER);
4316 %}
4318 operand eCXRegI(xRegI reg) %{
4319 constraint(ALLOC_IN_RC(ecx_reg));
4320 match(reg);
4321 match(eRegI);
4323 format %{ "ECX" %}
4324 interface(REG_INTER);
4325 %}
4327 operand eDXRegI(xRegI reg) %{
4328 constraint(ALLOC_IN_RC(edx_reg));
4329 match(reg);
4330 match(eRegI);
4332 format %{ "EDX" %}
4333 interface(REG_INTER);
4334 %}
4336 operand eDIRegI(xRegI reg) %{
4337 constraint(ALLOC_IN_RC(edi_reg));
4338 match(reg);
4339 match(eRegI);
4341 format %{ "EDI" %}
4342 interface(REG_INTER);
4343 %}
4345 operand naxRegI() %{
4346 constraint(ALLOC_IN_RC(nax_reg));
4347 match(RegI);
4348 match(eCXRegI);
4349 match(eDXRegI);
4350 match(eSIRegI);
4351 match(eDIRegI);
4353 format %{ %}
4354 interface(REG_INTER);
4355 %}
4357 operand nadxRegI() %{
4358 constraint(ALLOC_IN_RC(nadx_reg));
4359 match(RegI);
4360 match(eBXRegI);
4361 match(eCXRegI);
4362 match(eSIRegI);
4363 match(eDIRegI);
4365 format %{ %}
4366 interface(REG_INTER);
4367 %}
4369 operand ncxRegI() %{
4370 constraint(ALLOC_IN_RC(ncx_reg));
4371 match(RegI);
4372 match(eAXRegI);
4373 match(eDXRegI);
4374 match(eSIRegI);
4375 match(eDIRegI);
4377 format %{ %}
4378 interface(REG_INTER);
4379 %}
4381 // // This operand was used by cmpFastUnlock, but conflicted with 'object' reg
4382 // //
4383 operand eSIRegI(xRegI reg) %{
4384 constraint(ALLOC_IN_RC(esi_reg));
4385 match(reg);
4386 match(eRegI);
4388 format %{ "ESI" %}
4389 interface(REG_INTER);
4390 %}
4392 // Pointer Register
4393 operand anyRegP() %{
4394 constraint(ALLOC_IN_RC(any_reg));
4395 match(RegP);
4396 match(eAXRegP);
4397 match(eBXRegP);
4398 match(eCXRegP);
4399 match(eDIRegP);
4400 match(eRegP);
4402 format %{ %}
4403 interface(REG_INTER);
4404 %}
4406 operand eRegP() %{
4407 constraint(ALLOC_IN_RC(e_reg));
4408 match(RegP);
4409 match(eAXRegP);
4410 match(eBXRegP);
4411 match(eCXRegP);
4412 match(eDIRegP);
4414 format %{ %}
4415 interface(REG_INTER);
4416 %}
4418 // On windows95, EBP is not safe to use for implicit null tests.
4419 operand eRegP_no_EBP() %{
4420 constraint(ALLOC_IN_RC(e_reg_no_rbp));
4421 match(RegP);
4422 match(eAXRegP);
4423 match(eBXRegP);
4424 match(eCXRegP);
4425 match(eDIRegP);
4427 op_cost(100);
4428 format %{ %}
4429 interface(REG_INTER);
4430 %}
4432 operand naxRegP() %{
4433 constraint(ALLOC_IN_RC(nax_reg));
4434 match(RegP);
4435 match(eBXRegP);
4436 match(eDXRegP);
4437 match(eCXRegP);
4438 match(eSIRegP);
4439 match(eDIRegP);
4441 format %{ %}
4442 interface(REG_INTER);
4443 %}
4445 operand nabxRegP() %{
4446 constraint(ALLOC_IN_RC(nabx_reg));
4447 match(RegP);
4448 match(eCXRegP);
4449 match(eDXRegP);
4450 match(eSIRegP);
4451 match(eDIRegP);
4453 format %{ %}
4454 interface(REG_INTER);
4455 %}
4457 operand pRegP() %{
4458 constraint(ALLOC_IN_RC(p_reg));
4459 match(RegP);
4460 match(eBXRegP);
4461 match(eDXRegP);
4462 match(eSIRegP);
4463 match(eDIRegP);
4465 format %{ %}
4466 interface(REG_INTER);
4467 %}
4469 // Special Registers
4470 // Return a pointer value
4471 operand eAXRegP(eRegP reg) %{
4472 constraint(ALLOC_IN_RC(eax_reg));
4473 match(reg);
4474 format %{ "EAX" %}
4475 interface(REG_INTER);
4476 %}
4478 // Used in AtomicAdd
4479 operand eBXRegP(eRegP reg) %{
4480 constraint(ALLOC_IN_RC(ebx_reg));
4481 match(reg);
4482 format %{ "EBX" %}
4483 interface(REG_INTER);
4484 %}
4486 // Tail-call (interprocedural jump) to interpreter
4487 operand eCXRegP(eRegP reg) %{
4488 constraint(ALLOC_IN_RC(ecx_reg));
4489 match(reg);
4490 format %{ "ECX" %}
4491 interface(REG_INTER);
4492 %}
4494 operand eSIRegP(eRegP reg) %{
4495 constraint(ALLOC_IN_RC(esi_reg));
4496 match(reg);
4497 format %{ "ESI" %}
4498 interface(REG_INTER);
4499 %}
4501 // Used in rep stosw
4502 operand eDIRegP(eRegP reg) %{
4503 constraint(ALLOC_IN_RC(edi_reg));
4504 match(reg);
4505 format %{ "EDI" %}
4506 interface(REG_INTER);
4507 %}
4509 operand eBPRegP() %{
4510 constraint(ALLOC_IN_RC(ebp_reg));
4511 match(RegP);
4512 format %{ "EBP" %}
4513 interface(REG_INTER);
4514 %}
4516 operand eRegL() %{
4517 constraint(ALLOC_IN_RC(long_reg));
4518 match(RegL);
4519 match(eADXRegL);
4521 format %{ %}
4522 interface(REG_INTER);
4523 %}
4525 operand eADXRegL( eRegL reg ) %{
4526 constraint(ALLOC_IN_RC(eadx_reg));
4527 match(reg);
4529 format %{ "EDX:EAX" %}
4530 interface(REG_INTER);
4531 %}
4533 operand eBCXRegL( eRegL reg ) %{
4534 constraint(ALLOC_IN_RC(ebcx_reg));
4535 match(reg);
4537 format %{ "EBX:ECX" %}
4538 interface(REG_INTER);
4539 %}
4541 // Special case for integer high multiply
4542 operand eADXRegL_low_only() %{
4543 constraint(ALLOC_IN_RC(eadx_reg));
4544 match(RegL);
4546 format %{ "EAX" %}
4547 interface(REG_INTER);
4548 %}
4550 // Flags register, used as output of compare instructions
4551 operand eFlagsReg() %{
4552 constraint(ALLOC_IN_RC(int_flags));
4553 match(RegFlags);
4555 format %{ "EFLAGS" %}
4556 interface(REG_INTER);
4557 %}
4559 // Flags register, used as output of FLOATING POINT compare instructions
4560 operand eFlagsRegU() %{
4561 constraint(ALLOC_IN_RC(int_flags));
4562 match(RegFlags);
4564 format %{ "EFLAGS_U" %}
4565 interface(REG_INTER);
4566 %}
4568 operand eFlagsRegUCF() %{
4569 constraint(ALLOC_IN_RC(int_flags));
4570 match(RegFlags);
4571 predicate(false);
4573 format %{ "EFLAGS_U_CF" %}
4574 interface(REG_INTER);
4575 %}
4577 // Condition Code Register used by long compare
4578 operand flagsReg_long_LTGE() %{
4579 constraint(ALLOC_IN_RC(int_flags));
4580 match(RegFlags);
4581 format %{ "FLAGS_LTGE" %}
4582 interface(REG_INTER);
4583 %}
4584 operand flagsReg_long_EQNE() %{
4585 constraint(ALLOC_IN_RC(int_flags));
4586 match(RegFlags);
4587 format %{ "FLAGS_EQNE" %}
4588 interface(REG_INTER);
4589 %}
4590 operand flagsReg_long_LEGT() %{
4591 constraint(ALLOC_IN_RC(int_flags));
4592 match(RegFlags);
4593 format %{ "FLAGS_LEGT" %}
4594 interface(REG_INTER);
4595 %}
4597 // Float register operands
4598 operand regDPR() %{
4599 predicate( UseSSE < 2 );
4600 constraint(ALLOC_IN_RC(dbl_reg));
4601 match(RegD);
4602 match(regDPR1);
4603 match(regDPR2);
4604 format %{ %}
4605 interface(REG_INTER);
4606 %}
4608 operand regDPR1(regDPR reg) %{
4609 predicate( UseSSE < 2 );
4610 constraint(ALLOC_IN_RC(dbl_reg0));
4611 match(reg);
4612 format %{ "FPR1" %}
4613 interface(REG_INTER);
4614 %}
4616 operand regDPR2(regDPR reg) %{
4617 predicate( UseSSE < 2 );
4618 constraint(ALLOC_IN_RC(dbl_reg1));
4619 match(reg);
4620 format %{ "FPR2" %}
4621 interface(REG_INTER);
4622 %}
4624 operand regnotDPR1(regDPR reg) %{
4625 predicate( UseSSE < 2 );
4626 constraint(ALLOC_IN_RC(dbl_notreg0));
4627 match(reg);
4628 format %{ %}
4629 interface(REG_INTER);
4630 %}
4632 // XMM Double register operands
4633 operand regD() %{
4634 predicate( UseSSE>=2 );
4635 constraint(ALLOC_IN_RC(xdb_reg));
4636 match(RegD);
4637 match(regD6);
4638 match(regD7);
4639 format %{ %}
4640 interface(REG_INTER);
4641 %}
4643 // XMM6 double register operands
4644 operand regD6(regD reg) %{
4645 predicate( UseSSE>=2 );
4646 constraint(ALLOC_IN_RC(xdb_reg6));
4647 match(reg);
4648 format %{ "XMM6" %}
4649 interface(REG_INTER);
4650 %}
4652 // XMM7 double register operands
4653 operand regD7(regD reg) %{
4654 predicate( UseSSE>=2 );
4655 constraint(ALLOC_IN_RC(xdb_reg7));
4656 match(reg);
4657 format %{ "XMM7" %}
4658 interface(REG_INTER);
4659 %}
4661 // Float register operands
4662 operand regFPR() %{
4663 predicate( UseSSE < 2 );
4664 constraint(ALLOC_IN_RC(flt_reg));
4665 match(RegF);
4666 match(regFPR1);
4667 format %{ %}
4668 interface(REG_INTER);
4669 %}
4671 // Float register operands
4672 operand regFPR1(regFPR reg) %{
4673 predicate( UseSSE < 2 );
4674 constraint(ALLOC_IN_RC(flt_reg0));
4675 match(reg);
4676 format %{ "FPR1" %}
4677 interface(REG_INTER);
4678 %}
4680 // XMM register operands
4681 operand regF() %{
4682 predicate( UseSSE>=1 );
4683 constraint(ALLOC_IN_RC(xmm_reg));
4684 match(RegF);
4685 format %{ %}
4686 interface(REG_INTER);
4687 %}
4690 //----------Memory Operands----------------------------------------------------
4691 // Direct Memory Operand
4692 operand direct(immP addr) %{
4693 match(addr);
4695 format %{ "[$addr]" %}
4696 interface(MEMORY_INTER) %{
4697 base(0xFFFFFFFF);
4698 index(0x4);
4699 scale(0x0);
4700 disp($addr);
4701 %}
4702 %}
4704 // Indirect Memory Operand
4705 operand indirect(eRegP reg) %{
4706 constraint(ALLOC_IN_RC(e_reg));
4707 match(reg);
4709 format %{ "[$reg]" %}
4710 interface(MEMORY_INTER) %{
4711 base($reg);
4712 index(0x4);
4713 scale(0x0);
4714 disp(0x0);
4715 %}
4716 %}
4718 // Indirect Memory Plus Short Offset Operand
4719 operand indOffset8(eRegP reg, immI8 off) %{
4720 match(AddP reg off);
4722 format %{ "[$reg + $off]" %}
4723 interface(MEMORY_INTER) %{
4724 base($reg);
4725 index(0x4);
4726 scale(0x0);
4727 disp($off);
4728 %}
4729 %}
4731 // Indirect Memory Plus Long Offset Operand
4732 operand indOffset32(eRegP reg, immI off) %{
4733 match(AddP reg off);
4735 format %{ "[$reg + $off]" %}
4736 interface(MEMORY_INTER) %{
4737 base($reg);
4738 index(0x4);
4739 scale(0x0);
4740 disp($off);
4741 %}
4742 %}
4744 // Indirect Memory Plus Long Offset Operand
4745 operand indOffset32X(eRegI reg, immP off) %{
4746 match(AddP off reg);
4748 format %{ "[$reg + $off]" %}
4749 interface(MEMORY_INTER) %{
4750 base($reg);
4751 index(0x4);
4752 scale(0x0);
4753 disp($off);
4754 %}
4755 %}
4757 // Indirect Memory Plus Index Register Plus Offset Operand
4758 operand indIndexOffset(eRegP reg, eRegI ireg, immI off) %{
4759 match(AddP (AddP reg ireg) off);
4761 op_cost(10);
4762 format %{"[$reg + $off + $ireg]" %}
4763 interface(MEMORY_INTER) %{
4764 base($reg);
4765 index($ireg);
4766 scale(0x0);
4767 disp($off);
4768 %}
4769 %}
4771 // Indirect Memory Plus Index Register Plus Offset Operand
4772 operand indIndex(eRegP reg, eRegI ireg) %{
4773 match(AddP reg ireg);
4775 op_cost(10);
4776 format %{"[$reg + $ireg]" %}
4777 interface(MEMORY_INTER) %{
4778 base($reg);
4779 index($ireg);
4780 scale(0x0);
4781 disp(0x0);
4782 %}
4783 %}
4785 // // -------------------------------------------------------------------------
4786 // // 486 architecture doesn't support "scale * index + offset" with out a base
4787 // // -------------------------------------------------------------------------
4788 // // Scaled Memory Operands
4789 // // Indirect Memory Times Scale Plus Offset Operand
4790 // operand indScaleOffset(immP off, eRegI ireg, immI2 scale) %{
4791 // match(AddP off (LShiftI ireg scale));
4792 //
4793 // op_cost(10);
4794 // format %{"[$off + $ireg << $scale]" %}
4795 // interface(MEMORY_INTER) %{
4796 // base(0x4);
4797 // index($ireg);
4798 // scale($scale);
4799 // disp($off);
4800 // %}
4801 // %}
4803 // Indirect Memory Times Scale Plus Index Register
4804 operand indIndexScale(eRegP reg, eRegI ireg, immI2 scale) %{
4805 match(AddP reg (LShiftI ireg scale));
4807 op_cost(10);
4808 format %{"[$reg + $ireg << $scale]" %}
4809 interface(MEMORY_INTER) %{
4810 base($reg);
4811 index($ireg);
4812 scale($scale);
4813 disp(0x0);
4814 %}
4815 %}
4817 // Indirect Memory Times Scale Plus Index Register Plus Offset Operand
4818 operand indIndexScaleOffset(eRegP reg, immI off, eRegI ireg, immI2 scale) %{
4819 match(AddP (AddP reg (LShiftI ireg scale)) off);
4821 op_cost(10);
4822 format %{"[$reg + $off + $ireg << $scale]" %}
4823 interface(MEMORY_INTER) %{
4824 base($reg);
4825 index($ireg);
4826 scale($scale);
4827 disp($off);
4828 %}
4829 %}
4831 //----------Load Long Memory Operands------------------------------------------
4832 // The load-long idiom will use it's address expression again after loading
4833 // the first word of the long. If the load-long destination overlaps with
4834 // registers used in the addressing expression, the 2nd half will be loaded
4835 // from a clobbered address. Fix this by requiring that load-long use
4836 // address registers that do not overlap with the load-long target.
4838 // load-long support
4839 operand load_long_RegP() %{
4840 constraint(ALLOC_IN_RC(esi_reg));
4841 match(RegP);
4842 match(eSIRegP);
4843 op_cost(100);
4844 format %{ %}
4845 interface(REG_INTER);
4846 %}
4848 // Indirect Memory Operand Long
4849 operand load_long_indirect(load_long_RegP reg) %{
4850 constraint(ALLOC_IN_RC(esi_reg));
4851 match(reg);
4853 format %{ "[$reg]" %}
4854 interface(MEMORY_INTER) %{
4855 base($reg);
4856 index(0x4);
4857 scale(0x0);
4858 disp(0x0);
4859 %}
4860 %}
4862 // Indirect Memory Plus Long Offset Operand
4863 operand load_long_indOffset32(load_long_RegP reg, immI off) %{
4864 match(AddP reg off);
4866 format %{ "[$reg + $off]" %}
4867 interface(MEMORY_INTER) %{
4868 base($reg);
4869 index(0x4);
4870 scale(0x0);
4871 disp($off);
4872 %}
4873 %}
4875 opclass load_long_memory(load_long_indirect, load_long_indOffset32);
4878 //----------Special Memory Operands--------------------------------------------
4879 // Stack Slot Operand - This operand is used for loading and storing temporary
4880 // values on the stack where a match requires a value to
4881 // flow through memory.
4882 operand stackSlotP(sRegP reg) %{
4883 constraint(ALLOC_IN_RC(stack_slots));
4884 // No match rule because this operand is only generated in matching
4885 format %{ "[$reg]" %}
4886 interface(MEMORY_INTER) %{
4887 base(0x4); // ESP
4888 index(0x4); // No Index
4889 scale(0x0); // No Scale
4890 disp($reg); // Stack Offset
4891 %}
4892 %}
4894 operand stackSlotI(sRegI reg) %{
4895 constraint(ALLOC_IN_RC(stack_slots));
4896 // No match rule because this operand is only generated in matching
4897 format %{ "[$reg]" %}
4898 interface(MEMORY_INTER) %{
4899 base(0x4); // ESP
4900 index(0x4); // No Index
4901 scale(0x0); // No Scale
4902 disp($reg); // Stack Offset
4903 %}
4904 %}
4906 operand stackSlotF(sRegF reg) %{
4907 constraint(ALLOC_IN_RC(stack_slots));
4908 // No match rule because this operand is only generated in matching
4909 format %{ "[$reg]" %}
4910 interface(MEMORY_INTER) %{
4911 base(0x4); // ESP
4912 index(0x4); // No Index
4913 scale(0x0); // No Scale
4914 disp($reg); // Stack Offset
4915 %}
4916 %}
4918 operand stackSlotD(sRegD reg) %{
4919 constraint(ALLOC_IN_RC(stack_slots));
4920 // No match rule because this operand is only generated in matching
4921 format %{ "[$reg]" %}
4922 interface(MEMORY_INTER) %{
4923 base(0x4); // ESP
4924 index(0x4); // No Index
4925 scale(0x0); // No Scale
4926 disp($reg); // Stack Offset
4927 %}
4928 %}
4930 operand stackSlotL(sRegL reg) %{
4931 constraint(ALLOC_IN_RC(stack_slots));
4932 // No match rule because this operand is only generated in matching
4933 format %{ "[$reg]" %}
4934 interface(MEMORY_INTER) %{
4935 base(0x4); // ESP
4936 index(0x4); // No Index
4937 scale(0x0); // No Scale
4938 disp($reg); // Stack Offset
4939 %}
4940 %}
4942 //----------Memory Operands - Win95 Implicit Null Variants----------------
4943 // Indirect Memory Operand
4944 operand indirect_win95_safe(eRegP_no_EBP reg)
4945 %{
4946 constraint(ALLOC_IN_RC(e_reg));
4947 match(reg);
4949 op_cost(100);
4950 format %{ "[$reg]" %}
4951 interface(MEMORY_INTER) %{
4952 base($reg);
4953 index(0x4);
4954 scale(0x0);
4955 disp(0x0);
4956 %}
4957 %}
4959 // Indirect Memory Plus Short Offset Operand
4960 operand indOffset8_win95_safe(eRegP_no_EBP reg, immI8 off)
4961 %{
4962 match(AddP reg off);
4964 op_cost(100);
4965 format %{ "[$reg + $off]" %}
4966 interface(MEMORY_INTER) %{
4967 base($reg);
4968 index(0x4);
4969 scale(0x0);
4970 disp($off);
4971 %}
4972 %}
4974 // Indirect Memory Plus Long Offset Operand
4975 operand indOffset32_win95_safe(eRegP_no_EBP reg, immI off)
4976 %{
4977 match(AddP reg off);
4979 op_cost(100);
4980 format %{ "[$reg + $off]" %}
4981 interface(MEMORY_INTER) %{
4982 base($reg);
4983 index(0x4);
4984 scale(0x0);
4985 disp($off);
4986 %}
4987 %}
4989 // Indirect Memory Plus Index Register Plus Offset Operand
4990 operand indIndexOffset_win95_safe(eRegP_no_EBP reg, eRegI ireg, immI off)
4991 %{
4992 match(AddP (AddP reg ireg) off);
4994 op_cost(100);
4995 format %{"[$reg + $off + $ireg]" %}
4996 interface(MEMORY_INTER) %{
4997 base($reg);
4998 index($ireg);
4999 scale(0x0);
5000 disp($off);
5001 %}
5002 %}
5004 // Indirect Memory Times Scale Plus Index Register
5005 operand indIndexScale_win95_safe(eRegP_no_EBP reg, eRegI ireg, immI2 scale)
5006 %{
5007 match(AddP reg (LShiftI ireg scale));
5009 op_cost(100);
5010 format %{"[$reg + $ireg << $scale]" %}
5011 interface(MEMORY_INTER) %{
5012 base($reg);
5013 index($ireg);
5014 scale($scale);
5015 disp(0x0);
5016 %}
5017 %}
5019 // Indirect Memory Times Scale Plus Index Register Plus Offset Operand
5020 operand indIndexScaleOffset_win95_safe(eRegP_no_EBP reg, immI off, eRegI ireg, immI2 scale)
5021 %{
5022 match(AddP (AddP reg (LShiftI ireg scale)) off);
5024 op_cost(100);
5025 format %{"[$reg + $off + $ireg << $scale]" %}
5026 interface(MEMORY_INTER) %{
5027 base($reg);
5028 index($ireg);
5029 scale($scale);
5030 disp($off);
5031 %}
5032 %}
5034 //----------Conditional Branch Operands----------------------------------------
5035 // Comparison Op - This is the operation of the comparison, and is limited to
5036 // the following set of codes:
5037 // L (<), LE (<=), G (>), GE (>=), E (==), NE (!=)
5038 //
5039 // Other attributes of the comparison, such as unsignedness, are specified
5040 // by the comparison instruction that sets a condition code flags register.
5041 // That result is represented by a flags operand whose subtype is appropriate
5042 // to the unsignedness (etc.) of the comparison.
5043 //
5044 // Later, the instruction which matches both the Comparison Op (a Bool) and
5045 // the flags (produced by the Cmp) specifies the coding of the comparison op
5046 // by matching a specific subtype of Bool operand below, such as cmpOpU.
5048 // Comparision Code
5049 operand cmpOp() %{
5050 match(Bool);
5052 format %{ "" %}
5053 interface(COND_INTER) %{
5054 equal(0x4, "e");
5055 not_equal(0x5, "ne");
5056 less(0xC, "l");
5057 greater_equal(0xD, "ge");
5058 less_equal(0xE, "le");
5059 greater(0xF, "g");
5060 %}
5061 %}
5063 // Comparison Code, unsigned compare. Used by FP also, with
5064 // C2 (unordered) turned into GT or LT already. The other bits
5065 // C0 and C3 are turned into Carry & Zero flags.
5066 operand cmpOpU() %{
5067 match(Bool);
5069 format %{ "" %}
5070 interface(COND_INTER) %{
5071 equal(0x4, "e");
5072 not_equal(0x5, "ne");
5073 less(0x2, "b");
5074 greater_equal(0x3, "nb");
5075 less_equal(0x6, "be");
5076 greater(0x7, "nbe");
5077 %}
5078 %}
5080 // Floating comparisons that don't require any fixup for the unordered case
5081 operand cmpOpUCF() %{
5082 match(Bool);
5083 predicate(n->as_Bool()->_test._test == BoolTest::lt ||
5084 n->as_Bool()->_test._test == BoolTest::ge ||
5085 n->as_Bool()->_test._test == BoolTest::le ||
5086 n->as_Bool()->_test._test == BoolTest::gt);
5087 format %{ "" %}
5088 interface(COND_INTER) %{
5089 equal(0x4, "e");
5090 not_equal(0x5, "ne");
5091 less(0x2, "b");
5092 greater_equal(0x3, "nb");
5093 less_equal(0x6, "be");
5094 greater(0x7, "nbe");
5095 %}
5096 %}
5099 // Floating comparisons that can be fixed up with extra conditional jumps
5100 operand cmpOpUCF2() %{
5101 match(Bool);
5102 predicate(n->as_Bool()->_test._test == BoolTest::ne ||
5103 n->as_Bool()->_test._test == BoolTest::eq);
5104 format %{ "" %}
5105 interface(COND_INTER) %{
5106 equal(0x4, "e");
5107 not_equal(0x5, "ne");
5108 less(0x2, "b");
5109 greater_equal(0x3, "nb");
5110 less_equal(0x6, "be");
5111 greater(0x7, "nbe");
5112 %}
5113 %}
5115 // Comparison Code for FP conditional move
5116 operand cmpOp_fcmov() %{
5117 match(Bool);
5119 format %{ "" %}
5120 interface(COND_INTER) %{
5121 equal (0x0C8);
5122 not_equal (0x1C8);
5123 less (0x0C0);
5124 greater_equal(0x1C0);
5125 less_equal (0x0D0);
5126 greater (0x1D0);
5127 %}
5128 %}
5130 // Comparision Code used in long compares
5131 operand cmpOp_commute() %{
5132 match(Bool);
5134 format %{ "" %}
5135 interface(COND_INTER) %{
5136 equal(0x4, "e");
5137 not_equal(0x5, "ne");
5138 less(0xF, "g");
5139 greater_equal(0xE, "le");
5140 less_equal(0xD, "ge");
5141 greater(0xC, "l");
5142 %}
5143 %}
5145 //----------OPERAND CLASSES----------------------------------------------------
5146 // Operand Classes are groups of operands that are used as to simplify
5147 // instruction definitions by not requiring the AD writer to specify separate
5148 // instructions for every form of operand when the instruction accepts
5149 // multiple operand types with the same basic encoding and format. The classic
5150 // case of this is memory operands.
5152 opclass memory(direct, indirect, indOffset8, indOffset32, indOffset32X, indIndexOffset,
5153 indIndex, indIndexScale, indIndexScaleOffset);
5155 // Long memory operations are encoded in 2 instructions and a +4 offset.
5156 // This means some kind of offset is always required and you cannot use
5157 // an oop as the offset (done when working on static globals).
5158 opclass long_memory(direct, indirect, indOffset8, indOffset32, indIndexOffset,
5159 indIndex, indIndexScale, indIndexScaleOffset);
5162 //----------PIPELINE-----------------------------------------------------------
5163 // Rules which define the behavior of the target architectures pipeline.
5164 pipeline %{
5166 //----------ATTRIBUTES---------------------------------------------------------
5167 attributes %{
5168 variable_size_instructions; // Fixed size instructions
5169 max_instructions_per_bundle = 3; // Up to 3 instructions per bundle
5170 instruction_unit_size = 1; // An instruction is 1 bytes long
5171 instruction_fetch_unit_size = 16; // The processor fetches one line
5172 instruction_fetch_units = 1; // of 16 bytes
5174 // List of nop instructions
5175 nops( MachNop );
5176 %}
5178 //----------RESOURCES----------------------------------------------------------
5179 // Resources are the functional units available to the machine
5181 // Generic P2/P3 pipeline
5182 // 3 decoders, only D0 handles big operands; a "bundle" is the limit of
5183 // 3 instructions decoded per cycle.
5184 // 2 load/store ops per cycle, 1 branch, 1 FPU,
5185 // 2 ALU op, only ALU0 handles mul/div instructions.
5186 resources( D0, D1, D2, DECODE = D0 | D1 | D2,
5187 MS0, MS1, MEM = MS0 | MS1,
5188 BR, FPU,
5189 ALU0, ALU1, ALU = ALU0 | ALU1 );
5191 //----------PIPELINE DESCRIPTION-----------------------------------------------
5192 // Pipeline Description specifies the stages in the machine's pipeline
5194 // Generic P2/P3 pipeline
5195 pipe_desc(S0, S1, S2, S3, S4, S5);
5197 //----------PIPELINE CLASSES---------------------------------------------------
5198 // Pipeline Classes describe the stages in which input and output are
5199 // referenced by the hardware pipeline.
5201 // Naming convention: ialu or fpu
5202 // Then: _reg
5203 // Then: _reg if there is a 2nd register
5204 // Then: _long if it's a pair of instructions implementing a long
5205 // Then: _fat if it requires the big decoder
5206 // Or: _mem if it requires the big decoder and a memory unit.
5208 // Integer ALU reg operation
5209 pipe_class ialu_reg(eRegI dst) %{
5210 single_instruction;
5211 dst : S4(write);
5212 dst : S3(read);
5213 DECODE : S0; // any decoder
5214 ALU : S3; // any alu
5215 %}
5217 // Long ALU reg operation
5218 pipe_class ialu_reg_long(eRegL dst) %{
5219 instruction_count(2);
5220 dst : S4(write);
5221 dst : S3(read);
5222 DECODE : S0(2); // any 2 decoders
5223 ALU : S3(2); // both alus
5224 %}
5226 // Integer ALU reg operation using big decoder
5227 pipe_class ialu_reg_fat(eRegI dst) %{
5228 single_instruction;
5229 dst : S4(write);
5230 dst : S3(read);
5231 D0 : S0; // big decoder only
5232 ALU : S3; // any alu
5233 %}
5235 // Long ALU reg operation using big decoder
5236 pipe_class ialu_reg_long_fat(eRegL dst) %{
5237 instruction_count(2);
5238 dst : S4(write);
5239 dst : S3(read);
5240 D0 : S0(2); // big decoder only; twice
5241 ALU : S3(2); // any 2 alus
5242 %}
5244 // Integer ALU reg-reg operation
5245 pipe_class ialu_reg_reg(eRegI dst, eRegI src) %{
5246 single_instruction;
5247 dst : S4(write);
5248 src : S3(read);
5249 DECODE : S0; // any decoder
5250 ALU : S3; // any alu
5251 %}
5253 // Long ALU reg-reg operation
5254 pipe_class ialu_reg_reg_long(eRegL dst, eRegL src) %{
5255 instruction_count(2);
5256 dst : S4(write);
5257 src : S3(read);
5258 DECODE : S0(2); // any 2 decoders
5259 ALU : S3(2); // both alus
5260 %}
5262 // Integer ALU reg-reg operation
5263 pipe_class ialu_reg_reg_fat(eRegI dst, memory src) %{
5264 single_instruction;
5265 dst : S4(write);
5266 src : S3(read);
5267 D0 : S0; // big decoder only
5268 ALU : S3; // any alu
5269 %}
5271 // Long ALU reg-reg operation
5272 pipe_class ialu_reg_reg_long_fat(eRegL dst, eRegL src) %{
5273 instruction_count(2);
5274 dst : S4(write);
5275 src : S3(read);
5276 D0 : S0(2); // big decoder only; twice
5277 ALU : S3(2); // both alus
5278 %}
5280 // Integer ALU reg-mem operation
5281 pipe_class ialu_reg_mem(eRegI dst, memory mem) %{
5282 single_instruction;
5283 dst : S5(write);
5284 mem : S3(read);
5285 D0 : S0; // big decoder only
5286 ALU : S4; // any alu
5287 MEM : S3; // any mem
5288 %}
5290 // Long ALU reg-mem operation
5291 pipe_class ialu_reg_long_mem(eRegL dst, load_long_memory mem) %{
5292 instruction_count(2);
5293 dst : S5(write);
5294 mem : S3(read);
5295 D0 : S0(2); // big decoder only; twice
5296 ALU : S4(2); // any 2 alus
5297 MEM : S3(2); // both mems
5298 %}
5300 // Integer mem operation (prefetch)
5301 pipe_class ialu_mem(memory mem)
5302 %{
5303 single_instruction;
5304 mem : S3(read);
5305 D0 : S0; // big decoder only
5306 MEM : S3; // any mem
5307 %}
5309 // Integer Store to Memory
5310 pipe_class ialu_mem_reg(memory mem, eRegI src) %{
5311 single_instruction;
5312 mem : S3(read);
5313 src : S5(read);
5314 D0 : S0; // big decoder only
5315 ALU : S4; // any alu
5316 MEM : S3;
5317 %}
5319 // Long Store to Memory
5320 pipe_class ialu_mem_long_reg(memory mem, eRegL src) %{
5321 instruction_count(2);
5322 mem : S3(read);
5323 src : S5(read);
5324 D0 : S0(2); // big decoder only; twice
5325 ALU : S4(2); // any 2 alus
5326 MEM : S3(2); // Both mems
5327 %}
5329 // Integer Store to Memory
5330 pipe_class ialu_mem_imm(memory mem) %{
5331 single_instruction;
5332 mem : S3(read);
5333 D0 : S0; // big decoder only
5334 ALU : S4; // any alu
5335 MEM : S3;
5336 %}
5338 // Integer ALU0 reg-reg operation
5339 pipe_class ialu_reg_reg_alu0(eRegI dst, eRegI src) %{
5340 single_instruction;
5341 dst : S4(write);
5342 src : S3(read);
5343 D0 : S0; // Big decoder only
5344 ALU0 : S3; // only alu0
5345 %}
5347 // Integer ALU0 reg-mem operation
5348 pipe_class ialu_reg_mem_alu0(eRegI dst, memory mem) %{
5349 single_instruction;
5350 dst : S5(write);
5351 mem : S3(read);
5352 D0 : S0; // big decoder only
5353 ALU0 : S4; // ALU0 only
5354 MEM : S3; // any mem
5355 %}
5357 // Integer ALU reg-reg operation
5358 pipe_class ialu_cr_reg_reg(eFlagsReg cr, eRegI src1, eRegI src2) %{
5359 single_instruction;
5360 cr : S4(write);
5361 src1 : S3(read);
5362 src2 : S3(read);
5363 DECODE : S0; // any decoder
5364 ALU : S3; // any alu
5365 %}
5367 // Integer ALU reg-imm operation
5368 pipe_class ialu_cr_reg_imm(eFlagsReg cr, eRegI src1) %{
5369 single_instruction;
5370 cr : S4(write);
5371 src1 : S3(read);
5372 DECODE : S0; // any decoder
5373 ALU : S3; // any alu
5374 %}
5376 // Integer ALU reg-mem operation
5377 pipe_class ialu_cr_reg_mem(eFlagsReg cr, eRegI src1, memory src2) %{
5378 single_instruction;
5379 cr : S4(write);
5380 src1 : S3(read);
5381 src2 : S3(read);
5382 D0 : S0; // big decoder only
5383 ALU : S4; // any alu
5384 MEM : S3;
5385 %}
5387 // Conditional move reg-reg
5388 pipe_class pipe_cmplt( eRegI p, eRegI q, eRegI y ) %{
5389 instruction_count(4);
5390 y : S4(read);
5391 q : S3(read);
5392 p : S3(read);
5393 DECODE : S0(4); // any decoder
5394 %}
5396 // Conditional move reg-reg
5397 pipe_class pipe_cmov_reg( eRegI dst, eRegI src, eFlagsReg cr ) %{
5398 single_instruction;
5399 dst : S4(write);
5400 src : S3(read);
5401 cr : S3(read);
5402 DECODE : S0; // any decoder
5403 %}
5405 // Conditional move reg-mem
5406 pipe_class pipe_cmov_mem( eFlagsReg cr, eRegI dst, memory src) %{
5407 single_instruction;
5408 dst : S4(write);
5409 src : S3(read);
5410 cr : S3(read);
5411 DECODE : S0; // any decoder
5412 MEM : S3;
5413 %}
5415 // Conditional move reg-reg long
5416 pipe_class pipe_cmov_reg_long( eFlagsReg cr, eRegL dst, eRegL src) %{
5417 single_instruction;
5418 dst : S4(write);
5419 src : S3(read);
5420 cr : S3(read);
5421 DECODE : S0(2); // any 2 decoders
5422 %}
5424 // Conditional move double reg-reg
5425 pipe_class pipe_cmovDPR_reg( eFlagsReg cr, regDPR1 dst, regDPR src) %{
5426 single_instruction;
5427 dst : S4(write);
5428 src : S3(read);
5429 cr : S3(read);
5430 DECODE : S0; // any decoder
5431 %}
5433 // Float reg-reg operation
5434 pipe_class fpu_reg(regDPR dst) %{
5435 instruction_count(2);
5436 dst : S3(read);
5437 DECODE : S0(2); // any 2 decoders
5438 FPU : S3;
5439 %}
5441 // Float reg-reg operation
5442 pipe_class fpu_reg_reg(regDPR dst, regDPR src) %{
5443 instruction_count(2);
5444 dst : S4(write);
5445 src : S3(read);
5446 DECODE : S0(2); // any 2 decoders
5447 FPU : S3;
5448 %}
5450 // Float reg-reg operation
5451 pipe_class fpu_reg_reg_reg(regDPR dst, regDPR src1, regDPR src2) %{
5452 instruction_count(3);
5453 dst : S4(write);
5454 src1 : S3(read);
5455 src2 : S3(read);
5456 DECODE : S0(3); // any 3 decoders
5457 FPU : S3(2);
5458 %}
5460 // Float reg-reg operation
5461 pipe_class fpu_reg_reg_reg_reg(regDPR dst, regDPR src1, regDPR src2, regDPR src3) %{
5462 instruction_count(4);
5463 dst : S4(write);
5464 src1 : S3(read);
5465 src2 : S3(read);
5466 src3 : S3(read);
5467 DECODE : S0(4); // any 3 decoders
5468 FPU : S3(2);
5469 %}
5471 // Float reg-reg operation
5472 pipe_class fpu_reg_mem_reg_reg(regDPR dst, memory src1, regDPR src2, regDPR src3) %{
5473 instruction_count(4);
5474 dst : S4(write);
5475 src1 : S3(read);
5476 src2 : S3(read);
5477 src3 : S3(read);
5478 DECODE : S1(3); // any 3 decoders
5479 D0 : S0; // Big decoder only
5480 FPU : S3(2);
5481 MEM : S3;
5482 %}
5484 // Float reg-mem operation
5485 pipe_class fpu_reg_mem(regDPR dst, memory mem) %{
5486 instruction_count(2);
5487 dst : S5(write);
5488 mem : S3(read);
5489 D0 : S0; // big decoder only
5490 DECODE : S1; // any decoder for FPU POP
5491 FPU : S4;
5492 MEM : S3; // any mem
5493 %}
5495 // Float reg-mem operation
5496 pipe_class fpu_reg_reg_mem(regDPR dst, regDPR src1, memory mem) %{
5497 instruction_count(3);
5498 dst : S5(write);
5499 src1 : S3(read);
5500 mem : S3(read);
5501 D0 : S0; // big decoder only
5502 DECODE : S1(2); // any decoder for FPU POP
5503 FPU : S4;
5504 MEM : S3; // any mem
5505 %}
5507 // Float mem-reg operation
5508 pipe_class fpu_mem_reg(memory mem, regDPR src) %{
5509 instruction_count(2);
5510 src : S5(read);
5511 mem : S3(read);
5512 DECODE : S0; // any decoder for FPU PUSH
5513 D0 : S1; // big decoder only
5514 FPU : S4;
5515 MEM : S3; // any mem
5516 %}
5518 pipe_class fpu_mem_reg_reg(memory mem, regDPR src1, regDPR src2) %{
5519 instruction_count(3);
5520 src1 : S3(read);
5521 src2 : S3(read);
5522 mem : S3(read);
5523 DECODE : S0(2); // any decoder for FPU PUSH
5524 D0 : S1; // big decoder only
5525 FPU : S4;
5526 MEM : S3; // any mem
5527 %}
5529 pipe_class fpu_mem_reg_mem(memory mem, regDPR src1, memory src2) %{
5530 instruction_count(3);
5531 src1 : S3(read);
5532 src2 : S3(read);
5533 mem : S4(read);
5534 DECODE : S0; // any decoder for FPU PUSH
5535 D0 : S0(2); // big decoder only
5536 FPU : S4;
5537 MEM : S3(2); // any mem
5538 %}
5540 pipe_class fpu_mem_mem(memory dst, memory src1) %{
5541 instruction_count(2);
5542 src1 : S3(read);
5543 dst : S4(read);
5544 D0 : S0(2); // big decoder only
5545 MEM : S3(2); // any mem
5546 %}
5548 pipe_class fpu_mem_mem_mem(memory dst, memory src1, memory src2) %{
5549 instruction_count(3);
5550 src1 : S3(read);
5551 src2 : S3(read);
5552 dst : S4(read);
5553 D0 : S0(3); // big decoder only
5554 FPU : S4;
5555 MEM : S3(3); // any mem
5556 %}
5558 pipe_class fpu_mem_reg_con(memory mem, regDPR src1) %{
5559 instruction_count(3);
5560 src1 : S4(read);
5561 mem : S4(read);
5562 DECODE : S0; // any decoder for FPU PUSH
5563 D0 : S0(2); // big decoder only
5564 FPU : S4;
5565 MEM : S3(2); // any mem
5566 %}
5568 // Float load constant
5569 pipe_class fpu_reg_con(regDPR dst) %{
5570 instruction_count(2);
5571 dst : S5(write);
5572 D0 : S0; // big decoder only for the load
5573 DECODE : S1; // any decoder for FPU POP
5574 FPU : S4;
5575 MEM : S3; // any mem
5576 %}
5578 // Float load constant
5579 pipe_class fpu_reg_reg_con(regDPR dst, regDPR src) %{
5580 instruction_count(3);
5581 dst : S5(write);
5582 src : S3(read);
5583 D0 : S0; // big decoder only for the load
5584 DECODE : S1(2); // any decoder for FPU POP
5585 FPU : S4;
5586 MEM : S3; // any mem
5587 %}
5589 // UnConditional branch
5590 pipe_class pipe_jmp( label labl ) %{
5591 single_instruction;
5592 BR : S3;
5593 %}
5595 // Conditional branch
5596 pipe_class pipe_jcc( cmpOp cmp, eFlagsReg cr, label labl ) %{
5597 single_instruction;
5598 cr : S1(read);
5599 BR : S3;
5600 %}
5602 // Allocation idiom
5603 pipe_class pipe_cmpxchg( eRegP dst, eRegP heap_ptr ) %{
5604 instruction_count(1); force_serialization;
5605 fixed_latency(6);
5606 heap_ptr : S3(read);
5607 DECODE : S0(3);
5608 D0 : S2;
5609 MEM : S3;
5610 ALU : S3(2);
5611 dst : S5(write);
5612 BR : S5;
5613 %}
5615 // Generic big/slow expanded idiom
5616 pipe_class pipe_slow( ) %{
5617 instruction_count(10); multiple_bundles; force_serialization;
5618 fixed_latency(100);
5619 D0 : S0(2);
5620 MEM : S3(2);
5621 %}
5623 // The real do-nothing guy
5624 pipe_class empty( ) %{
5625 instruction_count(0);
5626 %}
5628 // Define the class for the Nop node
5629 define %{
5630 MachNop = empty;
5631 %}
5633 %}
5635 //----------INSTRUCTIONS-------------------------------------------------------
5636 //
5637 // match -- States which machine-independent subtree may be replaced
5638 // by this instruction.
5639 // ins_cost -- The estimated cost of this instruction is used by instruction
5640 // selection to identify a minimum cost tree of machine
5641 // instructions that matches a tree of machine-independent
5642 // instructions.
5643 // format -- A string providing the disassembly for this instruction.
5644 // The value of an instruction's operand may be inserted
5645 // by referring to it with a '$' prefix.
5646 // opcode -- Three instruction opcodes may be provided. These are referred
5647 // to within an encode class as $primary, $secondary, and $tertiary
5648 // respectively. The primary opcode is commonly used to
5649 // indicate the type of machine instruction, while secondary
5650 // and tertiary are often used for prefix options or addressing
5651 // modes.
5652 // ins_encode -- A list of encode classes with parameters. The encode class
5653 // name must have been defined in an 'enc_class' specification
5654 // in the encode section of the architecture description.
5656 //----------BSWAP-Instruction--------------------------------------------------
5657 instruct bytes_reverse_int(eRegI dst) %{
5658 match(Set dst (ReverseBytesI dst));
5660 format %{ "BSWAP $dst" %}
5661 opcode(0x0F, 0xC8);
5662 ins_encode( OpcP, OpcSReg(dst) );
5663 ins_pipe( ialu_reg );
5664 %}
5666 instruct bytes_reverse_long(eRegL dst) %{
5667 match(Set dst (ReverseBytesL dst));
5669 format %{ "BSWAP $dst.lo\n\t"
5670 "BSWAP $dst.hi\n\t"
5671 "XCHG $dst.lo $dst.hi" %}
5673 ins_cost(125);
5674 ins_encode( bswap_long_bytes(dst) );
5675 ins_pipe( ialu_reg_reg);
5676 %}
5678 instruct bytes_reverse_unsigned_short(eRegI dst) %{
5679 match(Set dst (ReverseBytesUS dst));
5681 format %{ "BSWAP $dst\n\t"
5682 "SHR $dst,16\n\t" %}
5683 ins_encode %{
5684 __ bswapl($dst$$Register);
5685 __ shrl($dst$$Register, 16);
5686 %}
5687 ins_pipe( ialu_reg );
5688 %}
5690 instruct bytes_reverse_short(eRegI dst) %{
5691 match(Set dst (ReverseBytesS dst));
5693 format %{ "BSWAP $dst\n\t"
5694 "SAR $dst,16\n\t" %}
5695 ins_encode %{
5696 __ bswapl($dst$$Register);
5697 __ sarl($dst$$Register, 16);
5698 %}
5699 ins_pipe( ialu_reg );
5700 %}
5703 //---------- Zeros Count Instructions ------------------------------------------
5705 instruct countLeadingZerosI(eRegI dst, eRegI src, eFlagsReg cr) %{
5706 predicate(UseCountLeadingZerosInstruction);
5707 match(Set dst (CountLeadingZerosI src));
5708 effect(KILL cr);
5710 format %{ "LZCNT $dst, $src\t# count leading zeros (int)" %}
5711 ins_encode %{
5712 __ lzcntl($dst$$Register, $src$$Register);
5713 %}
5714 ins_pipe(ialu_reg);
5715 %}
5717 instruct countLeadingZerosI_bsr(eRegI dst, eRegI src, eFlagsReg cr) %{
5718 predicate(!UseCountLeadingZerosInstruction);
5719 match(Set dst (CountLeadingZerosI src));
5720 effect(KILL cr);
5722 format %{ "BSR $dst, $src\t# count leading zeros (int)\n\t"
5723 "JNZ skip\n\t"
5724 "MOV $dst, -1\n"
5725 "skip:\n\t"
5726 "NEG $dst\n\t"
5727 "ADD $dst, 31" %}
5728 ins_encode %{
5729 Register Rdst = $dst$$Register;
5730 Register Rsrc = $src$$Register;
5731 Label skip;
5732 __ bsrl(Rdst, Rsrc);
5733 __ jccb(Assembler::notZero, skip);
5734 __ movl(Rdst, -1);
5735 __ bind(skip);
5736 __ negl(Rdst);
5737 __ addl(Rdst, BitsPerInt - 1);
5738 %}
5739 ins_pipe(ialu_reg);
5740 %}
5742 instruct countLeadingZerosL(eRegI dst, eRegL src, eFlagsReg cr) %{
5743 predicate(UseCountLeadingZerosInstruction);
5744 match(Set dst (CountLeadingZerosL src));
5745 effect(TEMP dst, KILL cr);
5747 format %{ "LZCNT $dst, $src.hi\t# count leading zeros (long)\n\t"
5748 "JNC done\n\t"
5749 "LZCNT $dst, $src.lo\n\t"
5750 "ADD $dst, 32\n"
5751 "done:" %}
5752 ins_encode %{
5753 Register Rdst = $dst$$Register;
5754 Register Rsrc = $src$$Register;
5755 Label done;
5756 __ lzcntl(Rdst, HIGH_FROM_LOW(Rsrc));
5757 __ jccb(Assembler::carryClear, done);
5758 __ lzcntl(Rdst, Rsrc);
5759 __ addl(Rdst, BitsPerInt);
5760 __ bind(done);
5761 %}
5762 ins_pipe(ialu_reg);
5763 %}
5765 instruct countLeadingZerosL_bsr(eRegI dst, eRegL src, eFlagsReg cr) %{
5766 predicate(!UseCountLeadingZerosInstruction);
5767 match(Set dst (CountLeadingZerosL src));
5768 effect(TEMP dst, KILL cr);
5770 format %{ "BSR $dst, $src.hi\t# count leading zeros (long)\n\t"
5771 "JZ msw_is_zero\n\t"
5772 "ADD $dst, 32\n\t"
5773 "JMP not_zero\n"
5774 "msw_is_zero:\n\t"
5775 "BSR $dst, $src.lo\n\t"
5776 "JNZ not_zero\n\t"
5777 "MOV $dst, -1\n"
5778 "not_zero:\n\t"
5779 "NEG $dst\n\t"
5780 "ADD $dst, 63\n" %}
5781 ins_encode %{
5782 Register Rdst = $dst$$Register;
5783 Register Rsrc = $src$$Register;
5784 Label msw_is_zero;
5785 Label not_zero;
5786 __ bsrl(Rdst, HIGH_FROM_LOW(Rsrc));
5787 __ jccb(Assembler::zero, msw_is_zero);
5788 __ addl(Rdst, BitsPerInt);
5789 __ jmpb(not_zero);
5790 __ bind(msw_is_zero);
5791 __ bsrl(Rdst, Rsrc);
5792 __ jccb(Assembler::notZero, not_zero);
5793 __ movl(Rdst, -1);
5794 __ bind(not_zero);
5795 __ negl(Rdst);
5796 __ addl(Rdst, BitsPerLong - 1);
5797 %}
5798 ins_pipe(ialu_reg);
5799 %}
5801 instruct countTrailingZerosI(eRegI dst, eRegI src, eFlagsReg cr) %{
5802 match(Set dst (CountTrailingZerosI src));
5803 effect(KILL cr);
5805 format %{ "BSF $dst, $src\t# count trailing zeros (int)\n\t"
5806 "JNZ done\n\t"
5807 "MOV $dst, 32\n"
5808 "done:" %}
5809 ins_encode %{
5810 Register Rdst = $dst$$Register;
5811 Label done;
5812 __ bsfl(Rdst, $src$$Register);
5813 __ jccb(Assembler::notZero, done);
5814 __ movl(Rdst, BitsPerInt);
5815 __ bind(done);
5816 %}
5817 ins_pipe(ialu_reg);
5818 %}
5820 instruct countTrailingZerosL(eRegI dst, eRegL src, eFlagsReg cr) %{
5821 match(Set dst (CountTrailingZerosL src));
5822 effect(TEMP dst, KILL cr);
5824 format %{ "BSF $dst, $src.lo\t# count trailing zeros (long)\n\t"
5825 "JNZ done\n\t"
5826 "BSF $dst, $src.hi\n\t"
5827 "JNZ msw_not_zero\n\t"
5828 "MOV $dst, 32\n"
5829 "msw_not_zero:\n\t"
5830 "ADD $dst, 32\n"
5831 "done:" %}
5832 ins_encode %{
5833 Register Rdst = $dst$$Register;
5834 Register Rsrc = $src$$Register;
5835 Label msw_not_zero;
5836 Label done;
5837 __ bsfl(Rdst, Rsrc);
5838 __ jccb(Assembler::notZero, done);
5839 __ bsfl(Rdst, HIGH_FROM_LOW(Rsrc));
5840 __ jccb(Assembler::notZero, msw_not_zero);
5841 __ movl(Rdst, BitsPerInt);
5842 __ bind(msw_not_zero);
5843 __ addl(Rdst, BitsPerInt);
5844 __ bind(done);
5845 %}
5846 ins_pipe(ialu_reg);
5847 %}
5850 //---------- Population Count Instructions -------------------------------------
5852 instruct popCountI(eRegI dst, eRegI src) %{
5853 predicate(UsePopCountInstruction);
5854 match(Set dst (PopCountI src));
5856 format %{ "POPCNT $dst, $src" %}
5857 ins_encode %{
5858 __ popcntl($dst$$Register, $src$$Register);
5859 %}
5860 ins_pipe(ialu_reg);
5861 %}
5863 instruct popCountI_mem(eRegI dst, memory mem) %{
5864 predicate(UsePopCountInstruction);
5865 match(Set dst (PopCountI (LoadI mem)));
5867 format %{ "POPCNT $dst, $mem" %}
5868 ins_encode %{
5869 __ popcntl($dst$$Register, $mem$$Address);
5870 %}
5871 ins_pipe(ialu_reg);
5872 %}
5874 // Note: Long.bitCount(long) returns an int.
5875 instruct popCountL(eRegI dst, eRegL src, eRegI tmp, eFlagsReg cr) %{
5876 predicate(UsePopCountInstruction);
5877 match(Set dst (PopCountL src));
5878 effect(KILL cr, TEMP tmp, TEMP dst);
5880 format %{ "POPCNT $dst, $src.lo\n\t"
5881 "POPCNT $tmp, $src.hi\n\t"
5882 "ADD $dst, $tmp" %}
5883 ins_encode %{
5884 __ popcntl($dst$$Register, $src$$Register);
5885 __ popcntl($tmp$$Register, HIGH_FROM_LOW($src$$Register));
5886 __ addl($dst$$Register, $tmp$$Register);
5887 %}
5888 ins_pipe(ialu_reg);
5889 %}
5891 // Note: Long.bitCount(long) returns an int.
5892 instruct popCountL_mem(eRegI dst, memory mem, eRegI tmp, eFlagsReg cr) %{
5893 predicate(UsePopCountInstruction);
5894 match(Set dst (PopCountL (LoadL mem)));
5895 effect(KILL cr, TEMP tmp, TEMP dst);
5897 format %{ "POPCNT $dst, $mem\n\t"
5898 "POPCNT $tmp, $mem+4\n\t"
5899 "ADD $dst, $tmp" %}
5900 ins_encode %{
5901 //__ popcntl($dst$$Register, $mem$$Address$$first);
5902 //__ popcntl($tmp$$Register, $mem$$Address$$second);
5903 __ popcntl($dst$$Register, Address::make_raw($mem$$base, $mem$$index, $mem$$scale, $mem$$disp, false));
5904 __ popcntl($tmp$$Register, Address::make_raw($mem$$base, $mem$$index, $mem$$scale, $mem$$disp + 4, false));
5905 __ addl($dst$$Register, $tmp$$Register);
5906 %}
5907 ins_pipe(ialu_reg);
5908 %}
5911 //----------Load/Store/Move Instructions---------------------------------------
5912 //----------Load Instructions--------------------------------------------------
5913 // Load Byte (8bit signed)
5914 instruct loadB(xRegI dst, memory mem) %{
5915 match(Set dst (LoadB mem));
5917 ins_cost(125);
5918 format %{ "MOVSX8 $dst,$mem\t# byte" %}
5920 ins_encode %{
5921 __ movsbl($dst$$Register, $mem$$Address);
5922 %}
5924 ins_pipe(ialu_reg_mem);
5925 %}
5927 // Load Byte (8bit signed) into Long Register
5928 instruct loadB2L(eRegL dst, memory mem, eFlagsReg cr) %{
5929 match(Set dst (ConvI2L (LoadB mem)));
5930 effect(KILL cr);
5932 ins_cost(375);
5933 format %{ "MOVSX8 $dst.lo,$mem\t# byte -> long\n\t"
5934 "MOV $dst.hi,$dst.lo\n\t"
5935 "SAR $dst.hi,7" %}
5937 ins_encode %{
5938 __ movsbl($dst$$Register, $mem$$Address);
5939 __ movl(HIGH_FROM_LOW($dst$$Register), $dst$$Register); // This is always a different register.
5940 __ sarl(HIGH_FROM_LOW($dst$$Register), 7); // 24+1 MSB are already signed extended.
5941 %}
5943 ins_pipe(ialu_reg_mem);
5944 %}
5946 // Load Unsigned Byte (8bit UNsigned)
5947 instruct loadUB(xRegI dst, memory mem) %{
5948 match(Set dst (LoadUB mem));
5950 ins_cost(125);
5951 format %{ "MOVZX8 $dst,$mem\t# ubyte -> int" %}
5953 ins_encode %{
5954 __ movzbl($dst$$Register, $mem$$Address);
5955 %}
5957 ins_pipe(ialu_reg_mem);
5958 %}
5960 // Load Unsigned Byte (8 bit UNsigned) into Long Register
5961 instruct loadUB2L(eRegL dst, memory mem, eFlagsReg cr) %{
5962 match(Set dst (ConvI2L (LoadUB mem)));
5963 effect(KILL cr);
5965 ins_cost(250);
5966 format %{ "MOVZX8 $dst.lo,$mem\t# ubyte -> long\n\t"
5967 "XOR $dst.hi,$dst.hi" %}
5969 ins_encode %{
5970 Register Rdst = $dst$$Register;
5971 __ movzbl(Rdst, $mem$$Address);
5972 __ xorl(HIGH_FROM_LOW(Rdst), HIGH_FROM_LOW(Rdst));
5973 %}
5975 ins_pipe(ialu_reg_mem);
5976 %}
5978 // Load Unsigned Byte (8 bit UNsigned) with mask into Long Register
5979 instruct loadUB2L_immI8(eRegL dst, memory mem, immI8 mask, eFlagsReg cr) %{
5980 match(Set dst (ConvI2L (AndI (LoadUB mem) mask)));
5981 effect(KILL cr);
5983 format %{ "MOVZX8 $dst.lo,$mem\t# ubyte & 8-bit mask -> long\n\t"
5984 "XOR $dst.hi,$dst.hi\n\t"
5985 "AND $dst.lo,$mask" %}
5986 ins_encode %{
5987 Register Rdst = $dst$$Register;
5988 __ movzbl(Rdst, $mem$$Address);
5989 __ xorl(HIGH_FROM_LOW(Rdst), HIGH_FROM_LOW(Rdst));
5990 __ andl(Rdst, $mask$$constant);
5991 %}
5992 ins_pipe(ialu_reg_mem);
5993 %}
5995 // Load Short (16bit signed)
5996 instruct loadS(eRegI dst, memory mem) %{
5997 match(Set dst (LoadS mem));
5999 ins_cost(125);
6000 format %{ "MOVSX $dst,$mem\t# short" %}
6002 ins_encode %{
6003 __ movswl($dst$$Register, $mem$$Address);
6004 %}
6006 ins_pipe(ialu_reg_mem);
6007 %}
6009 // Load Short (16 bit signed) to Byte (8 bit signed)
6010 instruct loadS2B(eRegI dst, memory mem, immI_24 twentyfour) %{
6011 match(Set dst (RShiftI (LShiftI (LoadS mem) twentyfour) twentyfour));
6013 ins_cost(125);
6014 format %{ "MOVSX $dst, $mem\t# short -> byte" %}
6015 ins_encode %{
6016 __ movsbl($dst$$Register, $mem$$Address);
6017 %}
6018 ins_pipe(ialu_reg_mem);
6019 %}
6021 // Load Short (16bit signed) into Long Register
6022 instruct loadS2L(eRegL dst, memory mem, eFlagsReg cr) %{
6023 match(Set dst (ConvI2L (LoadS mem)));
6024 effect(KILL cr);
6026 ins_cost(375);
6027 format %{ "MOVSX $dst.lo,$mem\t# short -> long\n\t"
6028 "MOV $dst.hi,$dst.lo\n\t"
6029 "SAR $dst.hi,15" %}
6031 ins_encode %{
6032 __ movswl($dst$$Register, $mem$$Address);
6033 __ movl(HIGH_FROM_LOW($dst$$Register), $dst$$Register); // This is always a different register.
6034 __ sarl(HIGH_FROM_LOW($dst$$Register), 15); // 16+1 MSB are already signed extended.
6035 %}
6037 ins_pipe(ialu_reg_mem);
6038 %}
6040 // Load Unsigned Short/Char (16bit unsigned)
6041 instruct loadUS(eRegI dst, memory mem) %{
6042 match(Set dst (LoadUS mem));
6044 ins_cost(125);
6045 format %{ "MOVZX $dst,$mem\t# ushort/char -> int" %}
6047 ins_encode %{
6048 __ movzwl($dst$$Register, $mem$$Address);
6049 %}
6051 ins_pipe(ialu_reg_mem);
6052 %}
6054 // Load Unsigned Short/Char (16 bit UNsigned) to Byte (8 bit signed)
6055 instruct loadUS2B(eRegI dst, memory mem, immI_24 twentyfour) %{
6056 match(Set dst (RShiftI (LShiftI (LoadUS mem) twentyfour) twentyfour));
6058 ins_cost(125);
6059 format %{ "MOVSX $dst, $mem\t# ushort -> byte" %}
6060 ins_encode %{
6061 __ movsbl($dst$$Register, $mem$$Address);
6062 %}
6063 ins_pipe(ialu_reg_mem);
6064 %}
6066 // Load Unsigned Short/Char (16 bit UNsigned) into Long Register
6067 instruct loadUS2L(eRegL dst, memory mem, eFlagsReg cr) %{
6068 match(Set dst (ConvI2L (LoadUS mem)));
6069 effect(KILL cr);
6071 ins_cost(250);
6072 format %{ "MOVZX $dst.lo,$mem\t# ushort/char -> long\n\t"
6073 "XOR $dst.hi,$dst.hi" %}
6075 ins_encode %{
6076 __ movzwl($dst$$Register, $mem$$Address);
6077 __ xorl(HIGH_FROM_LOW($dst$$Register), HIGH_FROM_LOW($dst$$Register));
6078 %}
6080 ins_pipe(ialu_reg_mem);
6081 %}
6083 // Load Unsigned Short/Char (16 bit UNsigned) with mask 0xFF into Long Register
6084 instruct loadUS2L_immI_255(eRegL dst, memory mem, immI_255 mask, eFlagsReg cr) %{
6085 match(Set dst (ConvI2L (AndI (LoadUS mem) mask)));
6086 effect(KILL cr);
6088 format %{ "MOVZX8 $dst.lo,$mem\t# ushort/char & 0xFF -> long\n\t"
6089 "XOR $dst.hi,$dst.hi" %}
6090 ins_encode %{
6091 Register Rdst = $dst$$Register;
6092 __ movzbl(Rdst, $mem$$Address);
6093 __ xorl(HIGH_FROM_LOW(Rdst), HIGH_FROM_LOW(Rdst));
6094 %}
6095 ins_pipe(ialu_reg_mem);
6096 %}
6098 // Load Unsigned Short/Char (16 bit UNsigned) with a 16-bit mask into Long Register
6099 instruct loadUS2L_immI16(eRegL dst, memory mem, immI16 mask, eFlagsReg cr) %{
6100 match(Set dst (ConvI2L (AndI (LoadUS mem) mask)));
6101 effect(KILL cr);
6103 format %{ "MOVZX $dst.lo, $mem\t# ushort/char & 16-bit mask -> long\n\t"
6104 "XOR $dst.hi,$dst.hi\n\t"
6105 "AND $dst.lo,$mask" %}
6106 ins_encode %{
6107 Register Rdst = $dst$$Register;
6108 __ movzwl(Rdst, $mem$$Address);
6109 __ xorl(HIGH_FROM_LOW(Rdst), HIGH_FROM_LOW(Rdst));
6110 __ andl(Rdst, $mask$$constant);
6111 %}
6112 ins_pipe(ialu_reg_mem);
6113 %}
6115 // Load Integer
6116 instruct loadI(eRegI dst, memory mem) %{
6117 match(Set dst (LoadI mem));
6119 ins_cost(125);
6120 format %{ "MOV $dst,$mem\t# int" %}
6122 ins_encode %{
6123 __ movl($dst$$Register, $mem$$Address);
6124 %}
6126 ins_pipe(ialu_reg_mem);
6127 %}
6129 // Load Integer (32 bit signed) to Byte (8 bit signed)
6130 instruct loadI2B(eRegI dst, memory mem, immI_24 twentyfour) %{
6131 match(Set dst (RShiftI (LShiftI (LoadI mem) twentyfour) twentyfour));
6133 ins_cost(125);
6134 format %{ "MOVSX $dst, $mem\t# int -> byte" %}
6135 ins_encode %{
6136 __ movsbl($dst$$Register, $mem$$Address);
6137 %}
6138 ins_pipe(ialu_reg_mem);
6139 %}
6141 // Load Integer (32 bit signed) to Unsigned Byte (8 bit UNsigned)
6142 instruct loadI2UB(eRegI dst, memory mem, immI_255 mask) %{
6143 match(Set dst (AndI (LoadI mem) mask));
6145 ins_cost(125);
6146 format %{ "MOVZX $dst, $mem\t# int -> ubyte" %}
6147 ins_encode %{
6148 __ movzbl($dst$$Register, $mem$$Address);
6149 %}
6150 ins_pipe(ialu_reg_mem);
6151 %}
6153 // Load Integer (32 bit signed) to Short (16 bit signed)
6154 instruct loadI2S(eRegI dst, memory mem, immI_16 sixteen) %{
6155 match(Set dst (RShiftI (LShiftI (LoadI mem) sixteen) sixteen));
6157 ins_cost(125);
6158 format %{ "MOVSX $dst, $mem\t# int -> short" %}
6159 ins_encode %{
6160 __ movswl($dst$$Register, $mem$$Address);
6161 %}
6162 ins_pipe(ialu_reg_mem);
6163 %}
6165 // Load Integer (32 bit signed) to Unsigned Short/Char (16 bit UNsigned)
6166 instruct loadI2US(eRegI dst, memory mem, immI_65535 mask) %{
6167 match(Set dst (AndI (LoadI mem) mask));
6169 ins_cost(125);
6170 format %{ "MOVZX $dst, $mem\t# int -> ushort/char" %}
6171 ins_encode %{
6172 __ movzwl($dst$$Register, $mem$$Address);
6173 %}
6174 ins_pipe(ialu_reg_mem);
6175 %}
6177 // Load Integer into Long Register
6178 instruct loadI2L(eRegL dst, memory mem, eFlagsReg cr) %{
6179 match(Set dst (ConvI2L (LoadI mem)));
6180 effect(KILL cr);
6182 ins_cost(375);
6183 format %{ "MOV $dst.lo,$mem\t# int -> long\n\t"
6184 "MOV $dst.hi,$dst.lo\n\t"
6185 "SAR $dst.hi,31" %}
6187 ins_encode %{
6188 __ movl($dst$$Register, $mem$$Address);
6189 __ movl(HIGH_FROM_LOW($dst$$Register), $dst$$Register); // This is always a different register.
6190 __ sarl(HIGH_FROM_LOW($dst$$Register), 31);
6191 %}
6193 ins_pipe(ialu_reg_mem);
6194 %}
6196 // Load Integer with mask 0xFF into Long Register
6197 instruct loadI2L_immI_255(eRegL dst, memory mem, immI_255 mask, eFlagsReg cr) %{
6198 match(Set dst (ConvI2L (AndI (LoadI mem) mask)));
6199 effect(KILL cr);
6201 format %{ "MOVZX8 $dst.lo,$mem\t# int & 0xFF -> long\n\t"
6202 "XOR $dst.hi,$dst.hi" %}
6203 ins_encode %{
6204 Register Rdst = $dst$$Register;
6205 __ movzbl(Rdst, $mem$$Address);
6206 __ xorl(HIGH_FROM_LOW(Rdst), HIGH_FROM_LOW(Rdst));
6207 %}
6208 ins_pipe(ialu_reg_mem);
6209 %}
6211 // Load Integer with mask 0xFFFF into Long Register
6212 instruct loadI2L_immI_65535(eRegL dst, memory mem, immI_65535 mask, eFlagsReg cr) %{
6213 match(Set dst (ConvI2L (AndI (LoadI mem) mask)));
6214 effect(KILL cr);
6216 format %{ "MOVZX $dst.lo,$mem\t# int & 0xFFFF -> long\n\t"
6217 "XOR $dst.hi,$dst.hi" %}
6218 ins_encode %{
6219 Register Rdst = $dst$$Register;
6220 __ movzwl(Rdst, $mem$$Address);
6221 __ xorl(HIGH_FROM_LOW(Rdst), HIGH_FROM_LOW(Rdst));
6222 %}
6223 ins_pipe(ialu_reg_mem);
6224 %}
6226 // Load Integer with 32-bit mask into Long Register
6227 instruct loadI2L_immI(eRegL dst, memory mem, immI mask, eFlagsReg cr) %{
6228 match(Set dst (ConvI2L (AndI (LoadI mem) mask)));
6229 effect(KILL cr);
6231 format %{ "MOV $dst.lo,$mem\t# int & 32-bit mask -> long\n\t"
6232 "XOR $dst.hi,$dst.hi\n\t"
6233 "AND $dst.lo,$mask" %}
6234 ins_encode %{
6235 Register Rdst = $dst$$Register;
6236 __ movl(Rdst, $mem$$Address);
6237 __ xorl(HIGH_FROM_LOW(Rdst), HIGH_FROM_LOW(Rdst));
6238 __ andl(Rdst, $mask$$constant);
6239 %}
6240 ins_pipe(ialu_reg_mem);
6241 %}
6243 // Load Unsigned Integer into Long Register
6244 instruct loadUI2L(eRegL dst, memory mem, eFlagsReg cr) %{
6245 match(Set dst (LoadUI2L mem));
6246 effect(KILL cr);
6248 ins_cost(250);
6249 format %{ "MOV $dst.lo,$mem\t# uint -> long\n\t"
6250 "XOR $dst.hi,$dst.hi" %}
6252 ins_encode %{
6253 __ movl($dst$$Register, $mem$$Address);
6254 __ xorl(HIGH_FROM_LOW($dst$$Register), HIGH_FROM_LOW($dst$$Register));
6255 %}
6257 ins_pipe(ialu_reg_mem);
6258 %}
6260 // Load Long. Cannot clobber address while loading, so restrict address
6261 // register to ESI
6262 instruct loadL(eRegL dst, load_long_memory mem) %{
6263 predicate(!((LoadLNode*)n)->require_atomic_access());
6264 match(Set dst (LoadL mem));
6266 ins_cost(250);
6267 format %{ "MOV $dst.lo,$mem\t# long\n\t"
6268 "MOV $dst.hi,$mem+4" %}
6270 ins_encode %{
6271 Address Amemlo = Address::make_raw($mem$$base, $mem$$index, $mem$$scale, $mem$$disp, false);
6272 Address Amemhi = Address::make_raw($mem$$base, $mem$$index, $mem$$scale, $mem$$disp + 4, false);
6273 __ movl($dst$$Register, Amemlo);
6274 __ movl(HIGH_FROM_LOW($dst$$Register), Amemhi);
6275 %}
6277 ins_pipe(ialu_reg_long_mem);
6278 %}
6280 // Volatile Load Long. Must be atomic, so do 64-bit FILD
6281 // then store it down to the stack and reload on the int
6282 // side.
6283 instruct loadL_volatile(stackSlotL dst, memory mem) %{
6284 predicate(UseSSE<=1 && ((LoadLNode*)n)->require_atomic_access());
6285 match(Set dst (LoadL mem));
6287 ins_cost(200);
6288 format %{ "FILD $mem\t# Atomic volatile long load\n\t"
6289 "FISTp $dst" %}
6290 ins_encode(enc_loadL_volatile(mem,dst));
6291 ins_pipe( fpu_reg_mem );
6292 %}
6294 instruct loadLX_volatile(stackSlotL dst, memory mem, regD tmp) %{
6295 predicate(UseSSE>=2 && ((LoadLNode*)n)->require_atomic_access());
6296 match(Set dst (LoadL mem));
6297 effect(TEMP tmp);
6298 ins_cost(180);
6299 format %{ "MOVSD $tmp,$mem\t# Atomic volatile long load\n\t"
6300 "MOVSD $dst,$tmp" %}
6301 ins_encode %{
6302 __ movdbl($tmp$$XMMRegister, $mem$$Address);
6303 __ movdbl(Address(rsp, $dst$$disp), $tmp$$XMMRegister);
6304 %}
6305 ins_pipe( pipe_slow );
6306 %}
6308 instruct loadLX_reg_volatile(eRegL dst, memory mem, regD tmp) %{
6309 predicate(UseSSE>=2 && ((LoadLNode*)n)->require_atomic_access());
6310 match(Set dst (LoadL mem));
6311 effect(TEMP tmp);
6312 ins_cost(160);
6313 format %{ "MOVSD $tmp,$mem\t# Atomic volatile long load\n\t"
6314 "MOVD $dst.lo,$tmp\n\t"
6315 "PSRLQ $tmp,32\n\t"
6316 "MOVD $dst.hi,$tmp" %}
6317 ins_encode %{
6318 __ movdbl($tmp$$XMMRegister, $mem$$Address);
6319 __ movdl($dst$$Register, $tmp$$XMMRegister);
6320 __ psrlq($tmp$$XMMRegister, 32);
6321 __ movdl(HIGH_FROM_LOW($dst$$Register), $tmp$$XMMRegister);
6322 %}
6323 ins_pipe( pipe_slow );
6324 %}
6326 // Load Range
6327 instruct loadRange(eRegI dst, memory mem) %{
6328 match(Set dst (LoadRange mem));
6330 ins_cost(125);
6331 format %{ "MOV $dst,$mem" %}
6332 opcode(0x8B);
6333 ins_encode( OpcP, RegMem(dst,mem));
6334 ins_pipe( ialu_reg_mem );
6335 %}
6338 // Load Pointer
6339 instruct loadP(eRegP dst, memory mem) %{
6340 match(Set dst (LoadP mem));
6342 ins_cost(125);
6343 format %{ "MOV $dst,$mem" %}
6344 opcode(0x8B);
6345 ins_encode( OpcP, RegMem(dst,mem));
6346 ins_pipe( ialu_reg_mem );
6347 %}
6349 // Load Klass Pointer
6350 instruct loadKlass(eRegP dst, memory mem) %{
6351 match(Set dst (LoadKlass mem));
6353 ins_cost(125);
6354 format %{ "MOV $dst,$mem" %}
6355 opcode(0x8B);
6356 ins_encode( OpcP, RegMem(dst,mem));
6357 ins_pipe( ialu_reg_mem );
6358 %}
6360 // Load Double
6361 instruct loadDPR(regDPR dst, memory mem) %{
6362 predicate(UseSSE<=1);
6363 match(Set dst (LoadD mem));
6365 ins_cost(150);
6366 format %{ "FLD_D ST,$mem\n\t"
6367 "FSTP $dst" %}
6368 opcode(0xDD); /* DD /0 */
6369 ins_encode( OpcP, RMopc_Mem(0x00,mem),
6370 Pop_Reg_DPR(dst) );
6371 ins_pipe( fpu_reg_mem );
6372 %}
6374 // Load Double to XMM
6375 instruct loadD(regD dst, memory mem) %{
6376 predicate(UseSSE>=2 && UseXmmLoadAndClearUpper);
6377 match(Set dst (LoadD mem));
6378 ins_cost(145);
6379 format %{ "MOVSD $dst,$mem" %}
6380 ins_encode %{
6381 __ movdbl ($dst$$XMMRegister, $mem$$Address);
6382 %}
6383 ins_pipe( pipe_slow );
6384 %}
6386 instruct loadD_partial(regD dst, memory mem) %{
6387 predicate(UseSSE>=2 && !UseXmmLoadAndClearUpper);
6388 match(Set dst (LoadD mem));
6389 ins_cost(145);
6390 format %{ "MOVLPD $dst,$mem" %}
6391 ins_encode %{
6392 __ movdbl ($dst$$XMMRegister, $mem$$Address);
6393 %}
6394 ins_pipe( pipe_slow );
6395 %}
6397 // Load to XMM register (single-precision floating point)
6398 // MOVSS instruction
6399 instruct loadF(regF dst, memory mem) %{
6400 predicate(UseSSE>=1);
6401 match(Set dst (LoadF mem));
6402 ins_cost(145);
6403 format %{ "MOVSS $dst,$mem" %}
6404 ins_encode %{
6405 __ movflt ($dst$$XMMRegister, $mem$$Address);
6406 %}
6407 ins_pipe( pipe_slow );
6408 %}
6410 // Load Float
6411 instruct loadFPR(regFPR dst, memory mem) %{
6412 predicate(UseSSE==0);
6413 match(Set dst (LoadF mem));
6415 ins_cost(150);
6416 format %{ "FLD_S ST,$mem\n\t"
6417 "FSTP $dst" %}
6418 opcode(0xD9); /* D9 /0 */
6419 ins_encode( OpcP, RMopc_Mem(0x00,mem),
6420 Pop_Reg_FPR(dst) );
6421 ins_pipe( fpu_reg_mem );
6422 %}
6424 // Load Aligned Packed Byte to XMM register
6425 instruct loadA8B(regD dst, memory mem) %{
6426 predicate(UseSSE>=1);
6427 match(Set dst (Load8B mem));
6428 ins_cost(125);
6429 format %{ "MOVQ $dst,$mem\t! packed8B" %}
6430 ins_encode %{
6431 __ movq($dst$$XMMRegister, $mem$$Address);
6432 %}
6433 ins_pipe( pipe_slow );
6434 %}
6436 // Load Aligned Packed Short to XMM register
6437 instruct loadA4S(regD dst, memory mem) %{
6438 predicate(UseSSE>=1);
6439 match(Set dst (Load4S mem));
6440 ins_cost(125);
6441 format %{ "MOVQ $dst,$mem\t! packed4S" %}
6442 ins_encode %{
6443 __ movq($dst$$XMMRegister, $mem$$Address);
6444 %}
6445 ins_pipe( pipe_slow );
6446 %}
6448 // Load Aligned Packed Char to XMM register
6449 instruct loadA4C(regD dst, memory mem) %{
6450 predicate(UseSSE>=1);
6451 match(Set dst (Load4C mem));
6452 ins_cost(125);
6453 format %{ "MOVQ $dst,$mem\t! packed4C" %}
6454 ins_encode %{
6455 __ movq($dst$$XMMRegister, $mem$$Address);
6456 %}
6457 ins_pipe( pipe_slow );
6458 %}
6460 // Load Aligned Packed Integer to XMM register
6461 instruct load2IU(regD dst, memory mem) %{
6462 predicate(UseSSE>=1);
6463 match(Set dst (Load2I mem));
6464 ins_cost(125);
6465 format %{ "MOVQ $dst,$mem\t! packed2I" %}
6466 ins_encode %{
6467 __ movq($dst$$XMMRegister, $mem$$Address);
6468 %}
6469 ins_pipe( pipe_slow );
6470 %}
6472 // Load Aligned Packed Single to XMM
6473 instruct loadA2F(regD dst, memory mem) %{
6474 predicate(UseSSE>=1);
6475 match(Set dst (Load2F mem));
6476 ins_cost(145);
6477 format %{ "MOVQ $dst,$mem\t! packed2F" %}
6478 ins_encode %{
6479 __ movq($dst$$XMMRegister, $mem$$Address);
6480 %}
6481 ins_pipe( pipe_slow );
6482 %}
6484 // Load Effective Address
6485 instruct leaP8(eRegP dst, indOffset8 mem) %{
6486 match(Set dst mem);
6488 ins_cost(110);
6489 format %{ "LEA $dst,$mem" %}
6490 opcode(0x8D);
6491 ins_encode( OpcP, RegMem(dst,mem));
6492 ins_pipe( ialu_reg_reg_fat );
6493 %}
6495 instruct leaP32(eRegP dst, indOffset32 mem) %{
6496 match(Set dst mem);
6498 ins_cost(110);
6499 format %{ "LEA $dst,$mem" %}
6500 opcode(0x8D);
6501 ins_encode( OpcP, RegMem(dst,mem));
6502 ins_pipe( ialu_reg_reg_fat );
6503 %}
6505 instruct leaPIdxOff(eRegP dst, indIndexOffset mem) %{
6506 match(Set dst mem);
6508 ins_cost(110);
6509 format %{ "LEA $dst,$mem" %}
6510 opcode(0x8D);
6511 ins_encode( OpcP, RegMem(dst,mem));
6512 ins_pipe( ialu_reg_reg_fat );
6513 %}
6515 instruct leaPIdxScale(eRegP dst, indIndexScale mem) %{
6516 match(Set dst mem);
6518 ins_cost(110);
6519 format %{ "LEA $dst,$mem" %}
6520 opcode(0x8D);
6521 ins_encode( OpcP, RegMem(dst,mem));
6522 ins_pipe( ialu_reg_reg_fat );
6523 %}
6525 instruct leaPIdxScaleOff(eRegP dst, indIndexScaleOffset mem) %{
6526 match(Set dst mem);
6528 ins_cost(110);
6529 format %{ "LEA $dst,$mem" %}
6530 opcode(0x8D);
6531 ins_encode( OpcP, RegMem(dst,mem));
6532 ins_pipe( ialu_reg_reg_fat );
6533 %}
6535 // Load Constant
6536 instruct loadConI(eRegI dst, immI src) %{
6537 match(Set dst src);
6539 format %{ "MOV $dst,$src" %}
6540 ins_encode( LdImmI(dst, src) );
6541 ins_pipe( ialu_reg_fat );
6542 %}
6544 // Load Constant zero
6545 instruct loadConI0(eRegI dst, immI0 src, eFlagsReg cr) %{
6546 match(Set dst src);
6547 effect(KILL cr);
6549 ins_cost(50);
6550 format %{ "XOR $dst,$dst" %}
6551 opcode(0x33); /* + rd */
6552 ins_encode( OpcP, RegReg( dst, dst ) );
6553 ins_pipe( ialu_reg );
6554 %}
6556 instruct loadConP(eRegP dst, immP src) %{
6557 match(Set dst src);
6559 format %{ "MOV $dst,$src" %}
6560 opcode(0xB8); /* + rd */
6561 ins_encode( LdImmP(dst, src) );
6562 ins_pipe( ialu_reg_fat );
6563 %}
6565 instruct loadConL(eRegL dst, immL src, eFlagsReg cr) %{
6566 match(Set dst src);
6567 effect(KILL cr);
6568 ins_cost(200);
6569 format %{ "MOV $dst.lo,$src.lo\n\t"
6570 "MOV $dst.hi,$src.hi" %}
6571 opcode(0xB8);
6572 ins_encode( LdImmL_Lo(dst, src), LdImmL_Hi(dst, src) );
6573 ins_pipe( ialu_reg_long_fat );
6574 %}
6576 instruct loadConL0(eRegL dst, immL0 src, eFlagsReg cr) %{
6577 match(Set dst src);
6578 effect(KILL cr);
6579 ins_cost(150);
6580 format %{ "XOR $dst.lo,$dst.lo\n\t"
6581 "XOR $dst.hi,$dst.hi" %}
6582 opcode(0x33,0x33);
6583 ins_encode( RegReg_Lo(dst,dst), RegReg_Hi(dst, dst) );
6584 ins_pipe( ialu_reg_long );
6585 %}
6587 // The instruction usage is guarded by predicate in operand immFPR().
6588 instruct loadConFPR(regFPR dst, immFPR con) %{
6589 match(Set dst con);
6590 ins_cost(125);
6591 format %{ "FLD_S ST,[$constantaddress]\t# load from constant table: float=$con\n\t"
6592 "FSTP $dst" %}
6593 ins_encode %{
6594 __ fld_s($constantaddress($con));
6595 __ fstp_d($dst$$reg);
6596 %}
6597 ins_pipe(fpu_reg_con);
6598 %}
6600 // The instruction usage is guarded by predicate in operand immFPR0().
6601 instruct loadConFPR0(regFPR dst, immFPR0 con) %{
6602 match(Set dst con);
6603 ins_cost(125);
6604 format %{ "FLDZ ST\n\t"
6605 "FSTP $dst" %}
6606 ins_encode %{
6607 __ fldz();
6608 __ fstp_d($dst$$reg);
6609 %}
6610 ins_pipe(fpu_reg_con);
6611 %}
6613 // The instruction usage is guarded by predicate in operand immFPR1().
6614 instruct loadConFPR1(regFPR dst, immFPR1 con) %{
6615 match(Set dst con);
6616 ins_cost(125);
6617 format %{ "FLD1 ST\n\t"
6618 "FSTP $dst" %}
6619 ins_encode %{
6620 __ fld1();
6621 __ fstp_d($dst$$reg);
6622 %}
6623 ins_pipe(fpu_reg_con);
6624 %}
6626 // The instruction usage is guarded by predicate in operand immF().
6627 instruct loadConF(regF dst, immF con) %{
6628 match(Set dst con);
6629 ins_cost(125);
6630 format %{ "MOVSS $dst,[$constantaddress]\t# load from constant table: float=$con" %}
6631 ins_encode %{
6632 __ movflt($dst$$XMMRegister, $constantaddress($con));
6633 %}
6634 ins_pipe(pipe_slow);
6635 %}
6637 // The instruction usage is guarded by predicate in operand immF0().
6638 instruct loadConF0(regF dst, immF0 src) %{
6639 match(Set dst src);
6640 ins_cost(100);
6641 format %{ "XORPS $dst,$dst\t# float 0.0" %}
6642 ins_encode %{
6643 __ xorps($dst$$XMMRegister, $dst$$XMMRegister);
6644 %}
6645 ins_pipe(pipe_slow);
6646 %}
6648 // The instruction usage is guarded by predicate in operand immDPR().
6649 instruct loadConDPR(regDPR dst, immDPR con) %{
6650 match(Set dst con);
6651 ins_cost(125);
6653 format %{ "FLD_D ST,[$constantaddress]\t# load from constant table: double=$con\n\t"
6654 "FSTP $dst" %}
6655 ins_encode %{
6656 __ fld_d($constantaddress($con));
6657 __ fstp_d($dst$$reg);
6658 %}
6659 ins_pipe(fpu_reg_con);
6660 %}
6662 // The instruction usage is guarded by predicate in operand immDPR0().
6663 instruct loadConDPR0(regDPR dst, immDPR0 con) %{
6664 match(Set dst con);
6665 ins_cost(125);
6667 format %{ "FLDZ ST\n\t"
6668 "FSTP $dst" %}
6669 ins_encode %{
6670 __ fldz();
6671 __ fstp_d($dst$$reg);
6672 %}
6673 ins_pipe(fpu_reg_con);
6674 %}
6676 // The instruction usage is guarded by predicate in operand immDPR1().
6677 instruct loadConDPR1(regDPR dst, immDPR1 con) %{
6678 match(Set dst con);
6679 ins_cost(125);
6681 format %{ "FLD1 ST\n\t"
6682 "FSTP $dst" %}
6683 ins_encode %{
6684 __ fld1();
6685 __ fstp_d($dst$$reg);
6686 %}
6687 ins_pipe(fpu_reg_con);
6688 %}
6690 // The instruction usage is guarded by predicate in operand immD().
6691 instruct loadConD(regD dst, immD con) %{
6692 match(Set dst con);
6693 ins_cost(125);
6694 format %{ "MOVSD $dst,[$constantaddress]\t# load from constant table: double=$con" %}
6695 ins_encode %{
6696 __ movdbl($dst$$XMMRegister, $constantaddress($con));
6697 %}
6698 ins_pipe(pipe_slow);
6699 %}
6701 // The instruction usage is guarded by predicate in operand immD0().
6702 instruct loadConD0(regD dst, immD0 src) %{
6703 match(Set dst src);
6704 ins_cost(100);
6705 format %{ "XORPD $dst,$dst\t# double 0.0" %}
6706 ins_encode %{
6707 __ xorpd ($dst$$XMMRegister, $dst$$XMMRegister);
6708 %}
6709 ins_pipe( pipe_slow );
6710 %}
6712 // Load Stack Slot
6713 instruct loadSSI(eRegI dst, stackSlotI src) %{
6714 match(Set dst src);
6715 ins_cost(125);
6717 format %{ "MOV $dst,$src" %}
6718 opcode(0x8B);
6719 ins_encode( OpcP, RegMem(dst,src));
6720 ins_pipe( ialu_reg_mem );
6721 %}
6723 instruct loadSSL(eRegL dst, stackSlotL src) %{
6724 match(Set dst src);
6726 ins_cost(200);
6727 format %{ "MOV $dst,$src.lo\n\t"
6728 "MOV $dst+4,$src.hi" %}
6729 opcode(0x8B, 0x8B);
6730 ins_encode( OpcP, RegMem( dst, src ), OpcS, RegMem_Hi( dst, src ) );
6731 ins_pipe( ialu_mem_long_reg );
6732 %}
6734 // Load Stack Slot
6735 instruct loadSSP(eRegP dst, stackSlotP src) %{
6736 match(Set dst src);
6737 ins_cost(125);
6739 format %{ "MOV $dst,$src" %}
6740 opcode(0x8B);
6741 ins_encode( OpcP, RegMem(dst,src));
6742 ins_pipe( ialu_reg_mem );
6743 %}
6745 // Load Stack Slot
6746 instruct loadSSF(regFPR dst, stackSlotF src) %{
6747 match(Set dst src);
6748 ins_cost(125);
6750 format %{ "FLD_S $src\n\t"
6751 "FSTP $dst" %}
6752 opcode(0xD9); /* D9 /0, FLD m32real */
6753 ins_encode( OpcP, RMopc_Mem_no_oop(0x00,src),
6754 Pop_Reg_FPR(dst) );
6755 ins_pipe( fpu_reg_mem );
6756 %}
6758 // Load Stack Slot
6759 instruct loadSSD(regDPR dst, stackSlotD src) %{
6760 match(Set dst src);
6761 ins_cost(125);
6763 format %{ "FLD_D $src\n\t"
6764 "FSTP $dst" %}
6765 opcode(0xDD); /* DD /0, FLD m64real */
6766 ins_encode( OpcP, RMopc_Mem_no_oop(0x00,src),
6767 Pop_Reg_DPR(dst) );
6768 ins_pipe( fpu_reg_mem );
6769 %}
6771 // Prefetch instructions.
6772 // Must be safe to execute with invalid address (cannot fault).
6774 instruct prefetchr0( memory mem ) %{
6775 predicate(UseSSE==0 && !VM_Version::supports_3dnow_prefetch());
6776 match(PrefetchRead mem);
6777 ins_cost(0);
6778 size(0);
6779 format %{ "PREFETCHR (non-SSE is empty encoding)" %}
6780 ins_encode();
6781 ins_pipe(empty);
6782 %}
6784 instruct prefetchr( memory mem ) %{
6785 predicate(UseSSE==0 && VM_Version::supports_3dnow_prefetch() || ReadPrefetchInstr==3);
6786 match(PrefetchRead mem);
6787 ins_cost(100);
6789 format %{ "PREFETCHR $mem\t! Prefetch into level 1 cache for read" %}
6790 ins_encode %{
6791 __ prefetchr($mem$$Address);
6792 %}
6793 ins_pipe(ialu_mem);
6794 %}
6796 instruct prefetchrNTA( memory mem ) %{
6797 predicate(UseSSE>=1 && ReadPrefetchInstr==0);
6798 match(PrefetchRead mem);
6799 ins_cost(100);
6801 format %{ "PREFETCHNTA $mem\t! Prefetch into non-temporal cache for read" %}
6802 ins_encode %{
6803 __ prefetchnta($mem$$Address);
6804 %}
6805 ins_pipe(ialu_mem);
6806 %}
6808 instruct prefetchrT0( memory mem ) %{
6809 predicate(UseSSE>=1 && ReadPrefetchInstr==1);
6810 match(PrefetchRead mem);
6811 ins_cost(100);
6813 format %{ "PREFETCHT0 $mem\t! Prefetch into L1 and L2 caches for read" %}
6814 ins_encode %{
6815 __ prefetcht0($mem$$Address);
6816 %}
6817 ins_pipe(ialu_mem);
6818 %}
6820 instruct prefetchrT2( memory mem ) %{
6821 predicate(UseSSE>=1 && ReadPrefetchInstr==2);
6822 match(PrefetchRead mem);
6823 ins_cost(100);
6825 format %{ "PREFETCHT2 $mem\t! Prefetch into L2 cache for read" %}
6826 ins_encode %{
6827 __ prefetcht2($mem$$Address);
6828 %}
6829 ins_pipe(ialu_mem);
6830 %}
6832 instruct prefetchw0( memory mem ) %{
6833 predicate(UseSSE==0 && !VM_Version::supports_3dnow_prefetch());
6834 match(PrefetchWrite mem);
6835 ins_cost(0);
6836 size(0);
6837 format %{ "Prefetch (non-SSE is empty encoding)" %}
6838 ins_encode();
6839 ins_pipe(empty);
6840 %}
6842 instruct prefetchw( memory mem ) %{
6843 predicate(UseSSE==0 && VM_Version::supports_3dnow_prefetch());
6844 match( PrefetchWrite mem );
6845 ins_cost(100);
6847 format %{ "PREFETCHW $mem\t! Prefetch into L1 cache and mark modified" %}
6848 ins_encode %{
6849 __ prefetchw($mem$$Address);
6850 %}
6851 ins_pipe(ialu_mem);
6852 %}
6854 instruct prefetchwNTA( memory mem ) %{
6855 predicate(UseSSE>=1);
6856 match(PrefetchWrite mem);
6857 ins_cost(100);
6859 format %{ "PREFETCHNTA $mem\t! Prefetch into non-temporal cache for write" %}
6860 ins_encode %{
6861 __ prefetchnta($mem$$Address);
6862 %}
6863 ins_pipe(ialu_mem);
6864 %}
6866 // Prefetch instructions for allocation.
6868 instruct prefetchAlloc0( memory mem ) %{
6869 predicate(UseSSE==0 && AllocatePrefetchInstr!=3);
6870 match(PrefetchAllocation mem);
6871 ins_cost(0);
6872 size(0);
6873 format %{ "Prefetch allocation (non-SSE is empty encoding)" %}
6874 ins_encode();
6875 ins_pipe(empty);
6876 %}
6878 instruct prefetchAlloc( memory mem ) %{
6879 predicate(AllocatePrefetchInstr==3);
6880 match( PrefetchAllocation mem );
6881 ins_cost(100);
6883 format %{ "PREFETCHW $mem\t! Prefetch allocation into L1 cache and mark modified" %}
6884 ins_encode %{
6885 __ prefetchw($mem$$Address);
6886 %}
6887 ins_pipe(ialu_mem);
6888 %}
6890 instruct prefetchAllocNTA( memory mem ) %{
6891 predicate(UseSSE>=1 && AllocatePrefetchInstr==0);
6892 match(PrefetchAllocation mem);
6893 ins_cost(100);
6895 format %{ "PREFETCHNTA $mem\t! Prefetch allocation into non-temporal cache for write" %}
6896 ins_encode %{
6897 __ prefetchnta($mem$$Address);
6898 %}
6899 ins_pipe(ialu_mem);
6900 %}
6902 instruct prefetchAllocT0( memory mem ) %{
6903 predicate(UseSSE>=1 && AllocatePrefetchInstr==1);
6904 match(PrefetchAllocation mem);
6905 ins_cost(100);
6907 format %{ "PREFETCHT0 $mem\t! Prefetch allocation into L1 and L2 caches for write" %}
6908 ins_encode %{
6909 __ prefetcht0($mem$$Address);
6910 %}
6911 ins_pipe(ialu_mem);
6912 %}
6914 instruct prefetchAllocT2( memory mem ) %{
6915 predicate(UseSSE>=1 && AllocatePrefetchInstr==2);
6916 match(PrefetchAllocation mem);
6917 ins_cost(100);
6919 format %{ "PREFETCHT2 $mem\t! Prefetch allocation into L2 cache for write" %}
6920 ins_encode %{
6921 __ prefetcht2($mem$$Address);
6922 %}
6923 ins_pipe(ialu_mem);
6924 %}
6926 //----------Store Instructions-------------------------------------------------
6928 // Store Byte
6929 instruct storeB(memory mem, xRegI src) %{
6930 match(Set mem (StoreB mem src));
6932 ins_cost(125);
6933 format %{ "MOV8 $mem,$src" %}
6934 opcode(0x88);
6935 ins_encode( OpcP, RegMem( src, mem ) );
6936 ins_pipe( ialu_mem_reg );
6937 %}
6939 // Store Char/Short
6940 instruct storeC(memory mem, eRegI src) %{
6941 match(Set mem (StoreC mem src));
6943 ins_cost(125);
6944 format %{ "MOV16 $mem,$src" %}
6945 opcode(0x89, 0x66);
6946 ins_encode( OpcS, OpcP, RegMem( src, mem ) );
6947 ins_pipe( ialu_mem_reg );
6948 %}
6950 // Store Integer
6951 instruct storeI(memory mem, eRegI src) %{
6952 match(Set mem (StoreI mem src));
6954 ins_cost(125);
6955 format %{ "MOV $mem,$src" %}
6956 opcode(0x89);
6957 ins_encode( OpcP, RegMem( src, mem ) );
6958 ins_pipe( ialu_mem_reg );
6959 %}
6961 // Store Long
6962 instruct storeL(long_memory mem, eRegL src) %{
6963 predicate(!((StoreLNode*)n)->require_atomic_access());
6964 match(Set mem (StoreL mem src));
6966 ins_cost(200);
6967 format %{ "MOV $mem,$src.lo\n\t"
6968 "MOV $mem+4,$src.hi" %}
6969 opcode(0x89, 0x89);
6970 ins_encode( OpcP, RegMem( src, mem ), OpcS, RegMem_Hi( src, mem ) );
6971 ins_pipe( ialu_mem_long_reg );
6972 %}
6974 // Store Long to Integer
6975 instruct storeL2I(memory mem, eRegL src) %{
6976 match(Set mem (StoreI mem (ConvL2I src)));
6978 format %{ "MOV $mem,$src.lo\t# long -> int" %}
6979 ins_encode %{
6980 __ movl($mem$$Address, $src$$Register);
6981 %}
6982 ins_pipe(ialu_mem_reg);
6983 %}
6985 // Volatile Store Long. Must be atomic, so move it into
6986 // the FP TOS and then do a 64-bit FIST. Has to probe the
6987 // target address before the store (for null-ptr checks)
6988 // so the memory operand is used twice in the encoding.
6989 instruct storeL_volatile(memory mem, stackSlotL src, eFlagsReg cr ) %{
6990 predicate(UseSSE<=1 && ((StoreLNode*)n)->require_atomic_access());
6991 match(Set mem (StoreL mem src));
6992 effect( KILL cr );
6993 ins_cost(400);
6994 format %{ "CMP $mem,EAX\t# Probe address for implicit null check\n\t"
6995 "FILD $src\n\t"
6996 "FISTp $mem\t # 64-bit atomic volatile long store" %}
6997 opcode(0x3B);
6998 ins_encode( OpcP, RegMem( EAX, mem ), enc_storeL_volatile(mem,src));
6999 ins_pipe( fpu_reg_mem );
7000 %}
7002 instruct storeLX_volatile(memory mem, stackSlotL src, regD tmp, eFlagsReg cr) %{
7003 predicate(UseSSE>=2 && ((StoreLNode*)n)->require_atomic_access());
7004 match(Set mem (StoreL mem src));
7005 effect( TEMP tmp, KILL cr );
7006 ins_cost(380);
7007 format %{ "CMP $mem,EAX\t# Probe address for implicit null check\n\t"
7008 "MOVSD $tmp,$src\n\t"
7009 "MOVSD $mem,$tmp\t # 64-bit atomic volatile long store" %}
7010 ins_encode %{
7011 __ cmpl(rax, $mem$$Address);
7012 __ movdbl($tmp$$XMMRegister, Address(rsp, $src$$disp));
7013 __ movdbl($mem$$Address, $tmp$$XMMRegister);
7014 %}
7015 ins_pipe( pipe_slow );
7016 %}
7018 instruct storeLX_reg_volatile(memory mem, eRegL src, regD tmp2, regD tmp, eFlagsReg cr) %{
7019 predicate(UseSSE>=2 && ((StoreLNode*)n)->require_atomic_access());
7020 match(Set mem (StoreL mem src));
7021 effect( TEMP tmp2 , TEMP tmp, KILL cr );
7022 ins_cost(360);
7023 format %{ "CMP $mem,EAX\t# Probe address for implicit null check\n\t"
7024 "MOVD $tmp,$src.lo\n\t"
7025 "MOVD $tmp2,$src.hi\n\t"
7026 "PUNPCKLDQ $tmp,$tmp2\n\t"
7027 "MOVSD $mem,$tmp\t # 64-bit atomic volatile long store" %}
7028 ins_encode %{
7029 __ cmpl(rax, $mem$$Address);
7030 __ movdl($tmp$$XMMRegister, $src$$Register);
7031 __ movdl($tmp2$$XMMRegister, HIGH_FROM_LOW($src$$Register));
7032 __ punpckldq($tmp$$XMMRegister, $tmp2$$XMMRegister);
7033 __ movdbl($mem$$Address, $tmp$$XMMRegister);
7034 %}
7035 ins_pipe( pipe_slow );
7036 %}
7038 // Store Pointer; for storing unknown oops and raw pointers
7039 instruct storeP(memory mem, anyRegP src) %{
7040 match(Set mem (StoreP mem src));
7042 ins_cost(125);
7043 format %{ "MOV $mem,$src" %}
7044 opcode(0x89);
7045 ins_encode( OpcP, RegMem( src, mem ) );
7046 ins_pipe( ialu_mem_reg );
7047 %}
7049 // Store Integer Immediate
7050 instruct storeImmI(memory mem, immI src) %{
7051 match(Set mem (StoreI mem src));
7053 ins_cost(150);
7054 format %{ "MOV $mem,$src" %}
7055 opcode(0xC7); /* C7 /0 */
7056 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con32( src ));
7057 ins_pipe( ialu_mem_imm );
7058 %}
7060 // Store Short/Char Immediate
7061 instruct storeImmI16(memory mem, immI16 src) %{
7062 predicate(UseStoreImmI16);
7063 match(Set mem (StoreC mem src));
7065 ins_cost(150);
7066 format %{ "MOV16 $mem,$src" %}
7067 opcode(0xC7); /* C7 /0 Same as 32 store immediate with prefix */
7068 ins_encode( SizePrefix, OpcP, RMopc_Mem(0x00,mem), Con16( src ));
7069 ins_pipe( ialu_mem_imm );
7070 %}
7072 // Store Pointer Immediate; null pointers or constant oops that do not
7073 // need card-mark barriers.
7074 instruct storeImmP(memory mem, immP src) %{
7075 match(Set mem (StoreP mem src));
7077 ins_cost(150);
7078 format %{ "MOV $mem,$src" %}
7079 opcode(0xC7); /* C7 /0 */
7080 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con32( src ));
7081 ins_pipe( ialu_mem_imm );
7082 %}
7084 // Store Byte Immediate
7085 instruct storeImmB(memory mem, immI8 src) %{
7086 match(Set mem (StoreB mem src));
7088 ins_cost(150);
7089 format %{ "MOV8 $mem,$src" %}
7090 opcode(0xC6); /* C6 /0 */
7091 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con8or32( src ));
7092 ins_pipe( ialu_mem_imm );
7093 %}
7095 // Store Aligned Packed Byte XMM register to memory
7096 instruct storeA8B(memory mem, regD src) %{
7097 predicate(UseSSE>=1);
7098 match(Set mem (Store8B mem src));
7099 ins_cost(145);
7100 format %{ "MOVQ $mem,$src\t! packed8B" %}
7101 ins_encode %{
7102 __ movq($mem$$Address, $src$$XMMRegister);
7103 %}
7104 ins_pipe( pipe_slow );
7105 %}
7107 // Store Aligned Packed Char/Short XMM register to memory
7108 instruct storeA4C(memory mem, regD src) %{
7109 predicate(UseSSE>=1);
7110 match(Set mem (Store4C mem src));
7111 ins_cost(145);
7112 format %{ "MOVQ $mem,$src\t! packed4C" %}
7113 ins_encode %{
7114 __ movq($mem$$Address, $src$$XMMRegister);
7115 %}
7116 ins_pipe( pipe_slow );
7117 %}
7119 // Store Aligned Packed Integer XMM register to memory
7120 instruct storeA2I(memory mem, regD src) %{
7121 predicate(UseSSE>=1);
7122 match(Set mem (Store2I mem src));
7123 ins_cost(145);
7124 format %{ "MOVQ $mem,$src\t! packed2I" %}
7125 ins_encode %{
7126 __ movq($mem$$Address, $src$$XMMRegister);
7127 %}
7128 ins_pipe( pipe_slow );
7129 %}
7131 // Store CMS card-mark Immediate
7132 instruct storeImmCM(memory mem, immI8 src) %{
7133 match(Set mem (StoreCM mem src));
7135 ins_cost(150);
7136 format %{ "MOV8 $mem,$src\t! CMS card-mark imm0" %}
7137 opcode(0xC6); /* C6 /0 */
7138 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con8or32( src ));
7139 ins_pipe( ialu_mem_imm );
7140 %}
7142 // Store Double
7143 instruct storeDPR( memory mem, regDPR1 src) %{
7144 predicate(UseSSE<=1);
7145 match(Set mem (StoreD mem src));
7147 ins_cost(100);
7148 format %{ "FST_D $mem,$src" %}
7149 opcode(0xDD); /* DD /2 */
7150 ins_encode( enc_FPR_store(mem,src) );
7151 ins_pipe( fpu_mem_reg );
7152 %}
7154 // Store double does rounding on x86
7155 instruct storeDPR_rounded( memory mem, regDPR1 src) %{
7156 predicate(UseSSE<=1);
7157 match(Set mem (StoreD mem (RoundDouble src)));
7159 ins_cost(100);
7160 format %{ "FST_D $mem,$src\t# round" %}
7161 opcode(0xDD); /* DD /2 */
7162 ins_encode( enc_FPR_store(mem,src) );
7163 ins_pipe( fpu_mem_reg );
7164 %}
7166 // Store XMM register to memory (double-precision floating points)
7167 // MOVSD instruction
7168 instruct storeD(memory mem, regD src) %{
7169 predicate(UseSSE>=2);
7170 match(Set mem (StoreD mem src));
7171 ins_cost(95);
7172 format %{ "MOVSD $mem,$src" %}
7173 ins_encode %{
7174 __ movdbl($mem$$Address, $src$$XMMRegister);
7175 %}
7176 ins_pipe( pipe_slow );
7177 %}
7179 // Store XMM register to memory (single-precision floating point)
7180 // MOVSS instruction
7181 instruct storeF(memory mem, regF src) %{
7182 predicate(UseSSE>=1);
7183 match(Set mem (StoreF mem src));
7184 ins_cost(95);
7185 format %{ "MOVSS $mem,$src" %}
7186 ins_encode %{
7187 __ movflt($mem$$Address, $src$$XMMRegister);
7188 %}
7189 ins_pipe( pipe_slow );
7190 %}
7192 // Store Aligned Packed Single Float XMM register to memory
7193 instruct storeA2F(memory mem, regD src) %{
7194 predicate(UseSSE>=1);
7195 match(Set mem (Store2F mem src));
7196 ins_cost(145);
7197 format %{ "MOVQ $mem,$src\t! packed2F" %}
7198 ins_encode %{
7199 __ movq($mem$$Address, $src$$XMMRegister);
7200 %}
7201 ins_pipe( pipe_slow );
7202 %}
7204 // Store Float
7205 instruct storeFPR( memory mem, regFPR1 src) %{
7206 predicate(UseSSE==0);
7207 match(Set mem (StoreF mem src));
7209 ins_cost(100);
7210 format %{ "FST_S $mem,$src" %}
7211 opcode(0xD9); /* D9 /2 */
7212 ins_encode( enc_FPR_store(mem,src) );
7213 ins_pipe( fpu_mem_reg );
7214 %}
7216 // Store Float does rounding on x86
7217 instruct storeFPR_rounded( memory mem, regFPR1 src) %{
7218 predicate(UseSSE==0);
7219 match(Set mem (StoreF mem (RoundFloat src)));
7221 ins_cost(100);
7222 format %{ "FST_S $mem,$src\t# round" %}
7223 opcode(0xD9); /* D9 /2 */
7224 ins_encode( enc_FPR_store(mem,src) );
7225 ins_pipe( fpu_mem_reg );
7226 %}
7228 // Store Float does rounding on x86
7229 instruct storeFPR_Drounded( memory mem, regDPR1 src) %{
7230 predicate(UseSSE<=1);
7231 match(Set mem (StoreF mem (ConvD2F src)));
7233 ins_cost(100);
7234 format %{ "FST_S $mem,$src\t# D-round" %}
7235 opcode(0xD9); /* D9 /2 */
7236 ins_encode( enc_FPR_store(mem,src) );
7237 ins_pipe( fpu_mem_reg );
7238 %}
7240 // Store immediate Float value (it is faster than store from FPU register)
7241 // The instruction usage is guarded by predicate in operand immFPR().
7242 instruct storeFPR_imm( memory mem, immFPR src) %{
7243 match(Set mem (StoreF mem src));
7245 ins_cost(50);
7246 format %{ "MOV $mem,$src\t# store float" %}
7247 opcode(0xC7); /* C7 /0 */
7248 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con32FPR_as_bits( src ));
7249 ins_pipe( ialu_mem_imm );
7250 %}
7252 // Store immediate Float value (it is faster than store from XMM register)
7253 // The instruction usage is guarded by predicate in operand immF().
7254 instruct storeF_imm( memory mem, immF src) %{
7255 match(Set mem (StoreF mem src));
7257 ins_cost(50);
7258 format %{ "MOV $mem,$src\t# store float" %}
7259 opcode(0xC7); /* C7 /0 */
7260 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con32F_as_bits( src ));
7261 ins_pipe( ialu_mem_imm );
7262 %}
7264 // Store Integer to stack slot
7265 instruct storeSSI(stackSlotI dst, eRegI src) %{
7266 match(Set dst src);
7268 ins_cost(100);
7269 format %{ "MOV $dst,$src" %}
7270 opcode(0x89);
7271 ins_encode( OpcPRegSS( dst, src ) );
7272 ins_pipe( ialu_mem_reg );
7273 %}
7275 // Store Integer to stack slot
7276 instruct storeSSP(stackSlotP dst, eRegP src) %{
7277 match(Set dst src);
7279 ins_cost(100);
7280 format %{ "MOV $dst,$src" %}
7281 opcode(0x89);
7282 ins_encode( OpcPRegSS( dst, src ) );
7283 ins_pipe( ialu_mem_reg );
7284 %}
7286 // Store Long to stack slot
7287 instruct storeSSL(stackSlotL dst, eRegL src) %{
7288 match(Set dst src);
7290 ins_cost(200);
7291 format %{ "MOV $dst,$src.lo\n\t"
7292 "MOV $dst+4,$src.hi" %}
7293 opcode(0x89, 0x89);
7294 ins_encode( OpcP, RegMem( src, dst ), OpcS, RegMem_Hi( src, dst ) );
7295 ins_pipe( ialu_mem_long_reg );
7296 %}
7298 //----------MemBar Instructions-----------------------------------------------
7299 // Memory barrier flavors
7301 instruct membar_acquire() %{
7302 match(MemBarAcquire);
7303 ins_cost(400);
7305 size(0);
7306 format %{ "MEMBAR-acquire ! (empty encoding)" %}
7307 ins_encode();
7308 ins_pipe(empty);
7309 %}
7311 instruct membar_acquire_lock() %{
7312 match(MemBarAcquireLock);
7313 ins_cost(0);
7315 size(0);
7316 format %{ "MEMBAR-acquire (prior CMPXCHG in FastLock so empty encoding)" %}
7317 ins_encode( );
7318 ins_pipe(empty);
7319 %}
7321 instruct membar_release() %{
7322 match(MemBarRelease);
7323 ins_cost(400);
7325 size(0);
7326 format %{ "MEMBAR-release ! (empty encoding)" %}
7327 ins_encode( );
7328 ins_pipe(empty);
7329 %}
7331 instruct membar_release_lock() %{
7332 match(MemBarReleaseLock);
7333 ins_cost(0);
7335 size(0);
7336 format %{ "MEMBAR-release (a FastUnlock follows so empty encoding)" %}
7337 ins_encode( );
7338 ins_pipe(empty);
7339 %}
7341 instruct membar_volatile(eFlagsReg cr) %{
7342 match(MemBarVolatile);
7343 effect(KILL cr);
7344 ins_cost(400);
7346 format %{
7347 $$template
7348 if (os::is_MP()) {
7349 $$emit$$"LOCK ADDL [ESP + #0], 0\t! membar_volatile"
7350 } else {
7351 $$emit$$"MEMBAR-volatile ! (empty encoding)"
7352 }
7353 %}
7354 ins_encode %{
7355 __ membar(Assembler::StoreLoad);
7356 %}
7357 ins_pipe(pipe_slow);
7358 %}
7360 instruct unnecessary_membar_volatile() %{
7361 match(MemBarVolatile);
7362 predicate(Matcher::post_store_load_barrier(n));
7363 ins_cost(0);
7365 size(0);
7366 format %{ "MEMBAR-volatile (unnecessary so empty encoding)" %}
7367 ins_encode( );
7368 ins_pipe(empty);
7369 %}
7371 instruct membar_storestore() %{
7372 match(MemBarStoreStore);
7373 ins_cost(0);
7375 size(0);
7376 format %{ "MEMBAR-storestore (empty encoding)" %}
7377 ins_encode( );
7378 ins_pipe(empty);
7379 %}
7381 //----------Move Instructions--------------------------------------------------
7382 instruct castX2P(eAXRegP dst, eAXRegI src) %{
7383 match(Set dst (CastX2P src));
7384 format %{ "# X2P $dst, $src" %}
7385 ins_encode( /*empty encoding*/ );
7386 ins_cost(0);
7387 ins_pipe(empty);
7388 %}
7390 instruct castP2X(eRegI dst, eRegP src ) %{
7391 match(Set dst (CastP2X src));
7392 ins_cost(50);
7393 format %{ "MOV $dst, $src\t# CastP2X" %}
7394 ins_encode( enc_Copy( dst, src) );
7395 ins_pipe( ialu_reg_reg );
7396 %}
7398 //----------Conditional Move---------------------------------------------------
7399 // Conditional move
7400 instruct jmovI_reg(cmpOp cop, eFlagsReg cr, eRegI dst, eRegI src) %{
7401 predicate(!VM_Version::supports_cmov() );
7402 match(Set dst (CMoveI (Binary cop cr) (Binary dst src)));
7403 ins_cost(200);
7404 format %{ "J$cop,us skip\t# signed cmove\n\t"
7405 "MOV $dst,$src\n"
7406 "skip:" %}
7407 ins_encode %{
7408 Label Lskip;
7409 // Invert sense of branch from sense of CMOV
7410 __ jccb((Assembler::Condition)($cop$$cmpcode^1), Lskip);
7411 __ movl($dst$$Register, $src$$Register);
7412 __ bind(Lskip);
7413 %}
7414 ins_pipe( pipe_cmov_reg );
7415 %}
7417 instruct jmovI_regU(cmpOpU cop, eFlagsRegU cr, eRegI dst, eRegI src) %{
7418 predicate(!VM_Version::supports_cmov() );
7419 match(Set dst (CMoveI (Binary cop cr) (Binary dst src)));
7420 ins_cost(200);
7421 format %{ "J$cop,us skip\t# unsigned cmove\n\t"
7422 "MOV $dst,$src\n"
7423 "skip:" %}
7424 ins_encode %{
7425 Label Lskip;
7426 // Invert sense of branch from sense of CMOV
7427 __ jccb((Assembler::Condition)($cop$$cmpcode^1), Lskip);
7428 __ movl($dst$$Register, $src$$Register);
7429 __ bind(Lskip);
7430 %}
7431 ins_pipe( pipe_cmov_reg );
7432 %}
7434 instruct cmovI_reg(eRegI dst, eRegI src, eFlagsReg cr, cmpOp cop ) %{
7435 predicate(VM_Version::supports_cmov() );
7436 match(Set dst (CMoveI (Binary cop cr) (Binary dst src)));
7437 ins_cost(200);
7438 format %{ "CMOV$cop $dst,$src" %}
7439 opcode(0x0F,0x40);
7440 ins_encode( enc_cmov(cop), RegReg( dst, src ) );
7441 ins_pipe( pipe_cmov_reg );
7442 %}
7444 instruct cmovI_regU( cmpOpU cop, eFlagsRegU cr, eRegI dst, eRegI src ) %{
7445 predicate(VM_Version::supports_cmov() );
7446 match(Set dst (CMoveI (Binary cop cr) (Binary dst src)));
7447 ins_cost(200);
7448 format %{ "CMOV$cop $dst,$src" %}
7449 opcode(0x0F,0x40);
7450 ins_encode( enc_cmov(cop), RegReg( dst, src ) );
7451 ins_pipe( pipe_cmov_reg );
7452 %}
7454 instruct cmovI_regUCF( cmpOpUCF cop, eFlagsRegUCF cr, eRegI dst, eRegI src ) %{
7455 predicate(VM_Version::supports_cmov() );
7456 match(Set dst (CMoveI (Binary cop cr) (Binary dst src)));
7457 ins_cost(200);
7458 expand %{
7459 cmovI_regU(cop, cr, dst, src);
7460 %}
7461 %}
7463 // Conditional move
7464 instruct cmovI_mem(cmpOp cop, eFlagsReg cr, eRegI dst, memory src) %{
7465 predicate(VM_Version::supports_cmov() );
7466 match(Set dst (CMoveI (Binary cop cr) (Binary dst (LoadI src))));
7467 ins_cost(250);
7468 format %{ "CMOV$cop $dst,$src" %}
7469 opcode(0x0F,0x40);
7470 ins_encode( enc_cmov(cop), RegMem( dst, src ) );
7471 ins_pipe( pipe_cmov_mem );
7472 %}
7474 // Conditional move
7475 instruct cmovI_memU(cmpOpU cop, eFlagsRegU cr, eRegI dst, memory src) %{
7476 predicate(VM_Version::supports_cmov() );
7477 match(Set dst (CMoveI (Binary cop cr) (Binary dst (LoadI src))));
7478 ins_cost(250);
7479 format %{ "CMOV$cop $dst,$src" %}
7480 opcode(0x0F,0x40);
7481 ins_encode( enc_cmov(cop), RegMem( dst, src ) );
7482 ins_pipe( pipe_cmov_mem );
7483 %}
7485 instruct cmovI_memUCF(cmpOpUCF cop, eFlagsRegUCF cr, eRegI dst, memory src) %{
7486 predicate(VM_Version::supports_cmov() );
7487 match(Set dst (CMoveI (Binary cop cr) (Binary dst (LoadI src))));
7488 ins_cost(250);
7489 expand %{
7490 cmovI_memU(cop, cr, dst, src);
7491 %}
7492 %}
7494 // Conditional move
7495 instruct cmovP_reg(eRegP dst, eRegP src, eFlagsReg cr, cmpOp cop ) %{
7496 predicate(VM_Version::supports_cmov() );
7497 match(Set dst (CMoveP (Binary cop cr) (Binary dst src)));
7498 ins_cost(200);
7499 format %{ "CMOV$cop $dst,$src\t# ptr" %}
7500 opcode(0x0F,0x40);
7501 ins_encode( enc_cmov(cop), RegReg( dst, src ) );
7502 ins_pipe( pipe_cmov_reg );
7503 %}
7505 // Conditional move (non-P6 version)
7506 // Note: a CMoveP is generated for stubs and native wrappers
7507 // regardless of whether we are on a P6, so we
7508 // emulate a cmov here
7509 instruct cmovP_reg_nonP6(eRegP dst, eRegP src, eFlagsReg cr, cmpOp cop ) %{
7510 match(Set dst (CMoveP (Binary cop cr) (Binary dst src)));
7511 ins_cost(300);
7512 format %{ "Jn$cop skip\n\t"
7513 "MOV $dst,$src\t# pointer\n"
7514 "skip:" %}
7515 opcode(0x8b);
7516 ins_encode( enc_cmov_branch(cop, 0x2), OpcP, RegReg(dst, src));
7517 ins_pipe( pipe_cmov_reg );
7518 %}
7520 // Conditional move
7521 instruct cmovP_regU(cmpOpU cop, eFlagsRegU cr, eRegP dst, eRegP src ) %{
7522 predicate(VM_Version::supports_cmov() );
7523 match(Set dst (CMoveP (Binary cop cr) (Binary dst src)));
7524 ins_cost(200);
7525 format %{ "CMOV$cop $dst,$src\t# ptr" %}
7526 opcode(0x0F,0x40);
7527 ins_encode( enc_cmov(cop), RegReg( dst, src ) );
7528 ins_pipe( pipe_cmov_reg );
7529 %}
7531 instruct cmovP_regUCF(cmpOpUCF cop, eFlagsRegUCF cr, eRegP dst, eRegP src ) %{
7532 predicate(VM_Version::supports_cmov() );
7533 match(Set dst (CMoveP (Binary cop cr) (Binary dst src)));
7534 ins_cost(200);
7535 expand %{
7536 cmovP_regU(cop, cr, dst, src);
7537 %}
7538 %}
7540 // DISABLED: Requires the ADLC to emit a bottom_type call that
7541 // correctly meets the two pointer arguments; one is an incoming
7542 // register but the other is a memory operand. ALSO appears to
7543 // be buggy with implicit null checks.
7544 //
7545 //// Conditional move
7546 //instruct cmovP_mem(cmpOp cop, eFlagsReg cr, eRegP dst, memory src) %{
7547 // predicate(VM_Version::supports_cmov() );
7548 // match(Set dst (CMoveP (Binary cop cr) (Binary dst (LoadP src))));
7549 // ins_cost(250);
7550 // format %{ "CMOV$cop $dst,$src\t# ptr" %}
7551 // opcode(0x0F,0x40);
7552 // ins_encode( enc_cmov(cop), RegMem( dst, src ) );
7553 // ins_pipe( pipe_cmov_mem );
7554 //%}
7555 //
7556 //// Conditional move
7557 //instruct cmovP_memU(cmpOpU cop, eFlagsRegU cr, eRegP dst, memory src) %{
7558 // predicate(VM_Version::supports_cmov() );
7559 // match(Set dst (CMoveP (Binary cop cr) (Binary dst (LoadP src))));
7560 // ins_cost(250);
7561 // format %{ "CMOV$cop $dst,$src\t# ptr" %}
7562 // opcode(0x0F,0x40);
7563 // ins_encode( enc_cmov(cop), RegMem( dst, src ) );
7564 // ins_pipe( pipe_cmov_mem );
7565 //%}
7567 // Conditional move
7568 instruct fcmovDPR_regU(cmpOp_fcmov cop, eFlagsRegU cr, regDPR1 dst, regDPR src) %{
7569 predicate(UseSSE<=1);
7570 match(Set dst (CMoveD (Binary cop cr) (Binary dst src)));
7571 ins_cost(200);
7572 format %{ "FCMOV$cop $dst,$src\t# double" %}
7573 opcode(0xDA);
7574 ins_encode( enc_cmov_dpr(cop,src) );
7575 ins_pipe( pipe_cmovDPR_reg );
7576 %}
7578 // Conditional move
7579 instruct fcmovFPR_regU(cmpOp_fcmov cop, eFlagsRegU cr, regFPR1 dst, regFPR src) %{
7580 predicate(UseSSE==0);
7581 match(Set dst (CMoveF (Binary cop cr) (Binary dst src)));
7582 ins_cost(200);
7583 format %{ "FCMOV$cop $dst,$src\t# float" %}
7584 opcode(0xDA);
7585 ins_encode( enc_cmov_dpr(cop,src) );
7586 ins_pipe( pipe_cmovDPR_reg );
7587 %}
7589 // Float CMOV on Intel doesn't handle *signed* compares, only unsigned.
7590 instruct fcmovDPR_regS(cmpOp cop, eFlagsReg cr, regDPR dst, regDPR src) %{
7591 predicate(UseSSE<=1);
7592 match(Set dst (CMoveD (Binary cop cr) (Binary dst src)));
7593 ins_cost(200);
7594 format %{ "Jn$cop skip\n\t"
7595 "MOV $dst,$src\t# double\n"
7596 "skip:" %}
7597 opcode (0xdd, 0x3); /* DD D8+i or DD /3 */
7598 ins_encode( enc_cmov_branch( cop, 0x4 ), Push_Reg_DPR(src), OpcP, RegOpc(dst) );
7599 ins_pipe( pipe_cmovDPR_reg );
7600 %}
7602 // Float CMOV on Intel doesn't handle *signed* compares, only unsigned.
7603 instruct fcmovFPR_regS(cmpOp cop, eFlagsReg cr, regFPR dst, regFPR src) %{
7604 predicate(UseSSE==0);
7605 match(Set dst (CMoveF (Binary cop cr) (Binary dst src)));
7606 ins_cost(200);
7607 format %{ "Jn$cop skip\n\t"
7608 "MOV $dst,$src\t# float\n"
7609 "skip:" %}
7610 opcode (0xdd, 0x3); /* DD D8+i or DD /3 */
7611 ins_encode( enc_cmov_branch( cop, 0x4 ), Push_Reg_FPR(src), OpcP, RegOpc(dst) );
7612 ins_pipe( pipe_cmovDPR_reg );
7613 %}
7615 // No CMOVE with SSE/SSE2
7616 instruct fcmovF_regS(cmpOp cop, eFlagsReg cr, regF dst, regF src) %{
7617 predicate (UseSSE>=1);
7618 match(Set dst (CMoveF (Binary cop cr) (Binary dst src)));
7619 ins_cost(200);
7620 format %{ "Jn$cop skip\n\t"
7621 "MOVSS $dst,$src\t# float\n"
7622 "skip:" %}
7623 ins_encode %{
7624 Label skip;
7625 // Invert sense of branch from sense of CMOV
7626 __ jccb((Assembler::Condition)($cop$$cmpcode^1), skip);
7627 __ movflt($dst$$XMMRegister, $src$$XMMRegister);
7628 __ bind(skip);
7629 %}
7630 ins_pipe( pipe_slow );
7631 %}
7633 // No CMOVE with SSE/SSE2
7634 instruct fcmovD_regS(cmpOp cop, eFlagsReg cr, regD dst, regD src) %{
7635 predicate (UseSSE>=2);
7636 match(Set dst (CMoveD (Binary cop cr) (Binary dst src)));
7637 ins_cost(200);
7638 format %{ "Jn$cop skip\n\t"
7639 "MOVSD $dst,$src\t# float\n"
7640 "skip:" %}
7641 ins_encode %{
7642 Label skip;
7643 // Invert sense of branch from sense of CMOV
7644 __ jccb((Assembler::Condition)($cop$$cmpcode^1), skip);
7645 __ movdbl($dst$$XMMRegister, $src$$XMMRegister);
7646 __ bind(skip);
7647 %}
7648 ins_pipe( pipe_slow );
7649 %}
7651 // unsigned version
7652 instruct fcmovF_regU(cmpOpU cop, eFlagsRegU cr, regF dst, regF src) %{
7653 predicate (UseSSE>=1);
7654 match(Set dst (CMoveF (Binary cop cr) (Binary dst src)));
7655 ins_cost(200);
7656 format %{ "Jn$cop skip\n\t"
7657 "MOVSS $dst,$src\t# float\n"
7658 "skip:" %}
7659 ins_encode %{
7660 Label skip;
7661 // Invert sense of branch from sense of CMOV
7662 __ jccb((Assembler::Condition)($cop$$cmpcode^1), skip);
7663 __ movflt($dst$$XMMRegister, $src$$XMMRegister);
7664 __ bind(skip);
7665 %}
7666 ins_pipe( pipe_slow );
7667 %}
7669 instruct fcmovF_regUCF(cmpOpUCF cop, eFlagsRegUCF cr, regF dst, regF src) %{
7670 predicate (UseSSE>=1);
7671 match(Set dst (CMoveF (Binary cop cr) (Binary dst src)));
7672 ins_cost(200);
7673 expand %{
7674 fcmovF_regU(cop, cr, dst, src);
7675 %}
7676 %}
7678 // unsigned version
7679 instruct fcmovD_regU(cmpOpU cop, eFlagsRegU cr, regD dst, regD src) %{
7680 predicate (UseSSE>=2);
7681 match(Set dst (CMoveD (Binary cop cr) (Binary dst src)));
7682 ins_cost(200);
7683 format %{ "Jn$cop skip\n\t"
7684 "MOVSD $dst,$src\t# float\n"
7685 "skip:" %}
7686 ins_encode %{
7687 Label skip;
7688 // Invert sense of branch from sense of CMOV
7689 __ jccb((Assembler::Condition)($cop$$cmpcode^1), skip);
7690 __ movdbl($dst$$XMMRegister, $src$$XMMRegister);
7691 __ bind(skip);
7692 %}
7693 ins_pipe( pipe_slow );
7694 %}
7696 instruct fcmovD_regUCF(cmpOpUCF cop, eFlagsRegUCF cr, regD dst, regD src) %{
7697 predicate (UseSSE>=2);
7698 match(Set dst (CMoveD (Binary cop cr) (Binary dst src)));
7699 ins_cost(200);
7700 expand %{
7701 fcmovD_regU(cop, cr, dst, src);
7702 %}
7703 %}
7705 instruct cmovL_reg(cmpOp cop, eFlagsReg cr, eRegL dst, eRegL src) %{
7706 predicate(VM_Version::supports_cmov() );
7707 match(Set dst (CMoveL (Binary cop cr) (Binary dst src)));
7708 ins_cost(200);
7709 format %{ "CMOV$cop $dst.lo,$src.lo\n\t"
7710 "CMOV$cop $dst.hi,$src.hi" %}
7711 opcode(0x0F,0x40);
7712 ins_encode( enc_cmov(cop), RegReg_Lo2( dst, src ), enc_cmov(cop), RegReg_Hi2( dst, src ) );
7713 ins_pipe( pipe_cmov_reg_long );
7714 %}
7716 instruct cmovL_regU(cmpOpU cop, eFlagsRegU cr, eRegL dst, eRegL src) %{
7717 predicate(VM_Version::supports_cmov() );
7718 match(Set dst (CMoveL (Binary cop cr) (Binary dst src)));
7719 ins_cost(200);
7720 format %{ "CMOV$cop $dst.lo,$src.lo\n\t"
7721 "CMOV$cop $dst.hi,$src.hi" %}
7722 opcode(0x0F,0x40);
7723 ins_encode( enc_cmov(cop), RegReg_Lo2( dst, src ), enc_cmov(cop), RegReg_Hi2( dst, src ) );
7724 ins_pipe( pipe_cmov_reg_long );
7725 %}
7727 instruct cmovL_regUCF(cmpOpUCF cop, eFlagsRegUCF cr, eRegL dst, eRegL src) %{
7728 predicate(VM_Version::supports_cmov() );
7729 match(Set dst (CMoveL (Binary cop cr) (Binary dst src)));
7730 ins_cost(200);
7731 expand %{
7732 cmovL_regU(cop, cr, dst, src);
7733 %}
7734 %}
7736 //----------Arithmetic Instructions--------------------------------------------
7737 //----------Addition Instructions----------------------------------------------
7738 // Integer Addition Instructions
7739 instruct addI_eReg(eRegI dst, eRegI src, eFlagsReg cr) %{
7740 match(Set dst (AddI dst src));
7741 effect(KILL cr);
7743 size(2);
7744 format %{ "ADD $dst,$src" %}
7745 opcode(0x03);
7746 ins_encode( OpcP, RegReg( dst, src) );
7747 ins_pipe( ialu_reg_reg );
7748 %}
7750 instruct addI_eReg_imm(eRegI dst, immI src, eFlagsReg cr) %{
7751 match(Set dst (AddI dst src));
7752 effect(KILL cr);
7754 format %{ "ADD $dst,$src" %}
7755 opcode(0x81, 0x00); /* /0 id */
7756 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
7757 ins_pipe( ialu_reg );
7758 %}
7760 instruct incI_eReg(eRegI dst, immI1 src, eFlagsReg cr) %{
7761 predicate(UseIncDec);
7762 match(Set dst (AddI dst src));
7763 effect(KILL cr);
7765 size(1);
7766 format %{ "INC $dst" %}
7767 opcode(0x40); /* */
7768 ins_encode( Opc_plus( primary, dst ) );
7769 ins_pipe( ialu_reg );
7770 %}
7772 instruct leaI_eReg_immI(eRegI dst, eRegI src0, immI src1) %{
7773 match(Set dst (AddI src0 src1));
7774 ins_cost(110);
7776 format %{ "LEA $dst,[$src0 + $src1]" %}
7777 opcode(0x8D); /* 0x8D /r */
7778 ins_encode( OpcP, RegLea( dst, src0, src1 ) );
7779 ins_pipe( ialu_reg_reg );
7780 %}
7782 instruct leaP_eReg_immI(eRegP dst, eRegP src0, immI src1) %{
7783 match(Set dst (AddP src0 src1));
7784 ins_cost(110);
7786 format %{ "LEA $dst,[$src0 + $src1]\t# ptr" %}
7787 opcode(0x8D); /* 0x8D /r */
7788 ins_encode( OpcP, RegLea( dst, src0, src1 ) );
7789 ins_pipe( ialu_reg_reg );
7790 %}
7792 instruct decI_eReg(eRegI dst, immI_M1 src, eFlagsReg cr) %{
7793 predicate(UseIncDec);
7794 match(Set dst (AddI dst src));
7795 effect(KILL cr);
7797 size(1);
7798 format %{ "DEC $dst" %}
7799 opcode(0x48); /* */
7800 ins_encode( Opc_plus( primary, dst ) );
7801 ins_pipe( ialu_reg );
7802 %}
7804 instruct addP_eReg(eRegP dst, eRegI src, eFlagsReg cr) %{
7805 match(Set dst (AddP dst src));
7806 effect(KILL cr);
7808 size(2);
7809 format %{ "ADD $dst,$src" %}
7810 opcode(0x03);
7811 ins_encode( OpcP, RegReg( dst, src) );
7812 ins_pipe( ialu_reg_reg );
7813 %}
7815 instruct addP_eReg_imm(eRegP dst, immI src, eFlagsReg cr) %{
7816 match(Set dst (AddP dst src));
7817 effect(KILL cr);
7819 format %{ "ADD $dst,$src" %}
7820 opcode(0x81,0x00); /* Opcode 81 /0 id */
7821 // ins_encode( RegImm( dst, src) );
7822 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
7823 ins_pipe( ialu_reg );
7824 %}
7826 instruct addI_eReg_mem(eRegI dst, memory src, eFlagsReg cr) %{
7827 match(Set dst (AddI dst (LoadI src)));
7828 effect(KILL cr);
7830 ins_cost(125);
7831 format %{ "ADD $dst,$src" %}
7832 opcode(0x03);
7833 ins_encode( OpcP, RegMem( dst, src) );
7834 ins_pipe( ialu_reg_mem );
7835 %}
7837 instruct addI_mem_eReg(memory dst, eRegI src, eFlagsReg cr) %{
7838 match(Set dst (StoreI dst (AddI (LoadI dst) src)));
7839 effect(KILL cr);
7841 ins_cost(150);
7842 format %{ "ADD $dst,$src" %}
7843 opcode(0x01); /* Opcode 01 /r */
7844 ins_encode( OpcP, RegMem( src, dst ) );
7845 ins_pipe( ialu_mem_reg );
7846 %}
7848 // Add Memory with Immediate
7849 instruct addI_mem_imm(memory dst, immI src, eFlagsReg cr) %{
7850 match(Set dst (StoreI dst (AddI (LoadI dst) src)));
7851 effect(KILL cr);
7853 ins_cost(125);
7854 format %{ "ADD $dst,$src" %}
7855 opcode(0x81); /* Opcode 81 /0 id */
7856 ins_encode( OpcSE( src ), RMopc_Mem(0x00,dst), Con8or32( src ) );
7857 ins_pipe( ialu_mem_imm );
7858 %}
7860 instruct incI_mem(memory dst, immI1 src, eFlagsReg cr) %{
7861 match(Set dst (StoreI dst (AddI (LoadI dst) src)));
7862 effect(KILL cr);
7864 ins_cost(125);
7865 format %{ "INC $dst" %}
7866 opcode(0xFF); /* Opcode FF /0 */
7867 ins_encode( OpcP, RMopc_Mem(0x00,dst));
7868 ins_pipe( ialu_mem_imm );
7869 %}
7871 instruct decI_mem(memory dst, immI_M1 src, eFlagsReg cr) %{
7872 match(Set dst (StoreI dst (AddI (LoadI dst) src)));
7873 effect(KILL cr);
7875 ins_cost(125);
7876 format %{ "DEC $dst" %}
7877 opcode(0xFF); /* Opcode FF /1 */
7878 ins_encode( OpcP, RMopc_Mem(0x01,dst));
7879 ins_pipe( ialu_mem_imm );
7880 %}
7883 instruct checkCastPP( eRegP dst ) %{
7884 match(Set dst (CheckCastPP dst));
7886 size(0);
7887 format %{ "#checkcastPP of $dst" %}
7888 ins_encode( /*empty encoding*/ );
7889 ins_pipe( empty );
7890 %}
7892 instruct castPP( eRegP dst ) %{
7893 match(Set dst (CastPP dst));
7894 format %{ "#castPP of $dst" %}
7895 ins_encode( /*empty encoding*/ );
7896 ins_pipe( empty );
7897 %}
7899 instruct castII( eRegI dst ) %{
7900 match(Set dst (CastII dst));
7901 format %{ "#castII of $dst" %}
7902 ins_encode( /*empty encoding*/ );
7903 ins_cost(0);
7904 ins_pipe( empty );
7905 %}
7908 // Load-locked - same as a regular pointer load when used with compare-swap
7909 instruct loadPLocked(eRegP dst, memory mem) %{
7910 match(Set dst (LoadPLocked mem));
7912 ins_cost(125);
7913 format %{ "MOV $dst,$mem\t# Load ptr. locked" %}
7914 opcode(0x8B);
7915 ins_encode( OpcP, RegMem(dst,mem));
7916 ins_pipe( ialu_reg_mem );
7917 %}
7919 // LoadLong-locked - same as a volatile long load when used with compare-swap
7920 instruct loadLLocked(stackSlotL dst, memory mem) %{
7921 predicate(UseSSE<=1);
7922 match(Set dst (LoadLLocked mem));
7924 ins_cost(200);
7925 format %{ "FILD $mem\t# Atomic volatile long load\n\t"
7926 "FISTp $dst" %}
7927 ins_encode(enc_loadL_volatile(mem,dst));
7928 ins_pipe( fpu_reg_mem );
7929 %}
7931 instruct loadLX_Locked(stackSlotL dst, memory mem, regD tmp) %{
7932 predicate(UseSSE>=2);
7933 match(Set dst (LoadLLocked mem));
7934 effect(TEMP tmp);
7935 ins_cost(180);
7936 format %{ "MOVSD $tmp,$mem\t# Atomic volatile long load\n\t"
7937 "MOVSD $dst,$tmp" %}
7938 ins_encode %{
7939 __ movdbl($tmp$$XMMRegister, $mem$$Address);
7940 __ movdbl(Address(rsp, $dst$$disp), $tmp$$XMMRegister);
7941 %}
7942 ins_pipe( pipe_slow );
7943 %}
7945 instruct loadLX_reg_Locked(eRegL dst, memory mem, regD tmp) %{
7946 predicate(UseSSE>=2);
7947 match(Set dst (LoadLLocked mem));
7948 effect(TEMP tmp);
7949 ins_cost(160);
7950 format %{ "MOVSD $tmp,$mem\t# Atomic volatile long load\n\t"
7951 "MOVD $dst.lo,$tmp\n\t"
7952 "PSRLQ $tmp,32\n\t"
7953 "MOVD $dst.hi,$tmp" %}
7954 ins_encode %{
7955 __ movdbl($tmp$$XMMRegister, $mem$$Address);
7956 __ movdl($dst$$Register, $tmp$$XMMRegister);
7957 __ psrlq($tmp$$XMMRegister, 32);
7958 __ movdl(HIGH_FROM_LOW($dst$$Register), $tmp$$XMMRegister);
7959 %}
7960 ins_pipe( pipe_slow );
7961 %}
7963 // Conditional-store of the updated heap-top.
7964 // Used during allocation of the shared heap.
7965 // Sets flags (EQ) on success. Implemented with a CMPXCHG on Intel.
7966 instruct storePConditional( memory heap_top_ptr, eAXRegP oldval, eRegP newval, eFlagsReg cr ) %{
7967 match(Set cr (StorePConditional heap_top_ptr (Binary oldval newval)));
7968 // EAX is killed if there is contention, but then it's also unused.
7969 // In the common case of no contention, EAX holds the new oop address.
7970 format %{ "CMPXCHG $heap_top_ptr,$newval\t# If EAX==$heap_top_ptr Then store $newval into $heap_top_ptr" %}
7971 ins_encode( lock_prefix, Opcode(0x0F), Opcode(0xB1), RegMem(newval,heap_top_ptr) );
7972 ins_pipe( pipe_cmpxchg );
7973 %}
7975 // Conditional-store of an int value.
7976 // ZF flag is set on success, reset otherwise. Implemented with a CMPXCHG on Intel.
7977 instruct storeIConditional( memory mem, eAXRegI oldval, eRegI newval, eFlagsReg cr ) %{
7978 match(Set cr (StoreIConditional mem (Binary oldval newval)));
7979 effect(KILL oldval);
7980 format %{ "CMPXCHG $mem,$newval\t# If EAX==$mem Then store $newval into $mem" %}
7981 ins_encode( lock_prefix, Opcode(0x0F), Opcode(0xB1), RegMem(newval, mem) );
7982 ins_pipe( pipe_cmpxchg );
7983 %}
7985 // Conditional-store of a long value.
7986 // ZF flag is set on success, reset otherwise. Implemented with a CMPXCHG8 on Intel.
7987 instruct storeLConditional( memory mem, eADXRegL oldval, eBCXRegL newval, eFlagsReg cr ) %{
7988 match(Set cr (StoreLConditional mem (Binary oldval newval)));
7989 effect(KILL oldval);
7990 format %{ "XCHG EBX,ECX\t# correct order for CMPXCHG8 instruction\n\t"
7991 "CMPXCHG8 $mem,ECX:EBX\t# If EDX:EAX==$mem Then store ECX:EBX into $mem\n\t"
7992 "XCHG EBX,ECX"
7993 %}
7994 ins_encode %{
7995 // Note: we need to swap rbx, and rcx before and after the
7996 // cmpxchg8 instruction because the instruction uses
7997 // rcx as the high order word of the new value to store but
7998 // our register encoding uses rbx.
7999 __ xchgl(as_Register(EBX_enc), as_Register(ECX_enc));
8000 if( os::is_MP() )
8001 __ lock();
8002 __ cmpxchg8($mem$$Address);
8003 __ xchgl(as_Register(EBX_enc), as_Register(ECX_enc));
8004 %}
8005 ins_pipe( pipe_cmpxchg );
8006 %}
8008 // No flag versions for CompareAndSwap{P,I,L} because matcher can't match them
8010 instruct compareAndSwapL( eRegI res, eSIRegP mem_ptr, eADXRegL oldval, eBCXRegL newval, eFlagsReg cr ) %{
8011 match(Set res (CompareAndSwapL mem_ptr (Binary oldval newval)));
8012 effect(KILL cr, KILL oldval);
8013 format %{ "CMPXCHG8 [$mem_ptr],$newval\t# If EDX:EAX==[$mem_ptr] Then store $newval into [$mem_ptr]\n\t"
8014 "MOV $res,0\n\t"
8015 "JNE,s fail\n\t"
8016 "MOV $res,1\n"
8017 "fail:" %}
8018 ins_encode( enc_cmpxchg8(mem_ptr),
8019 enc_flags_ne_to_boolean(res) );
8020 ins_pipe( pipe_cmpxchg );
8021 %}
8023 instruct compareAndSwapP( eRegI res, pRegP mem_ptr, eAXRegP oldval, eCXRegP newval, eFlagsReg cr) %{
8024 match(Set res (CompareAndSwapP mem_ptr (Binary oldval newval)));
8025 effect(KILL cr, KILL oldval);
8026 format %{ "CMPXCHG [$mem_ptr],$newval\t# If EAX==[$mem_ptr] Then store $newval into [$mem_ptr]\n\t"
8027 "MOV $res,0\n\t"
8028 "JNE,s fail\n\t"
8029 "MOV $res,1\n"
8030 "fail:" %}
8031 ins_encode( enc_cmpxchg(mem_ptr), enc_flags_ne_to_boolean(res) );
8032 ins_pipe( pipe_cmpxchg );
8033 %}
8035 instruct compareAndSwapI( eRegI res, pRegP mem_ptr, eAXRegI oldval, eCXRegI newval, eFlagsReg cr) %{
8036 match(Set res (CompareAndSwapI mem_ptr (Binary oldval newval)));
8037 effect(KILL cr, KILL oldval);
8038 format %{ "CMPXCHG [$mem_ptr],$newval\t# If EAX==[$mem_ptr] Then store $newval into [$mem_ptr]\n\t"
8039 "MOV $res,0\n\t"
8040 "JNE,s fail\n\t"
8041 "MOV $res,1\n"
8042 "fail:" %}
8043 ins_encode( enc_cmpxchg(mem_ptr), enc_flags_ne_to_boolean(res) );
8044 ins_pipe( pipe_cmpxchg );
8045 %}
8047 //----------Subtraction Instructions-------------------------------------------
8048 // Integer Subtraction Instructions
8049 instruct subI_eReg(eRegI dst, eRegI src, eFlagsReg cr) %{
8050 match(Set dst (SubI dst src));
8051 effect(KILL cr);
8053 size(2);
8054 format %{ "SUB $dst,$src" %}
8055 opcode(0x2B);
8056 ins_encode( OpcP, RegReg( dst, src) );
8057 ins_pipe( ialu_reg_reg );
8058 %}
8060 instruct subI_eReg_imm(eRegI dst, immI src, eFlagsReg cr) %{
8061 match(Set dst (SubI dst src));
8062 effect(KILL cr);
8064 format %{ "SUB $dst,$src" %}
8065 opcode(0x81,0x05); /* Opcode 81 /5 */
8066 // ins_encode( RegImm( dst, src) );
8067 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
8068 ins_pipe( ialu_reg );
8069 %}
8071 instruct subI_eReg_mem(eRegI dst, memory src, eFlagsReg cr) %{
8072 match(Set dst (SubI dst (LoadI src)));
8073 effect(KILL cr);
8075 ins_cost(125);
8076 format %{ "SUB $dst,$src" %}
8077 opcode(0x2B);
8078 ins_encode( OpcP, RegMem( dst, src) );
8079 ins_pipe( ialu_reg_mem );
8080 %}
8082 instruct subI_mem_eReg(memory dst, eRegI src, eFlagsReg cr) %{
8083 match(Set dst (StoreI dst (SubI (LoadI dst) src)));
8084 effect(KILL cr);
8086 ins_cost(150);
8087 format %{ "SUB $dst,$src" %}
8088 opcode(0x29); /* Opcode 29 /r */
8089 ins_encode( OpcP, RegMem( src, dst ) );
8090 ins_pipe( ialu_mem_reg );
8091 %}
8093 // Subtract from a pointer
8094 instruct subP_eReg(eRegP dst, eRegI src, immI0 zero, eFlagsReg cr) %{
8095 match(Set dst (AddP dst (SubI zero src)));
8096 effect(KILL cr);
8098 size(2);
8099 format %{ "SUB $dst,$src" %}
8100 opcode(0x2B);
8101 ins_encode( OpcP, RegReg( dst, src) );
8102 ins_pipe( ialu_reg_reg );
8103 %}
8105 instruct negI_eReg(eRegI dst, immI0 zero, eFlagsReg cr) %{
8106 match(Set dst (SubI zero dst));
8107 effect(KILL cr);
8109 size(2);
8110 format %{ "NEG $dst" %}
8111 opcode(0xF7,0x03); // Opcode F7 /3
8112 ins_encode( OpcP, RegOpc( dst ) );
8113 ins_pipe( ialu_reg );
8114 %}
8117 //----------Multiplication/Division Instructions-------------------------------
8118 // Integer Multiplication Instructions
8119 // Multiply Register
8120 instruct mulI_eReg(eRegI dst, eRegI src, eFlagsReg cr) %{
8121 match(Set dst (MulI dst src));
8122 effect(KILL cr);
8124 size(3);
8125 ins_cost(300);
8126 format %{ "IMUL $dst,$src" %}
8127 opcode(0xAF, 0x0F);
8128 ins_encode( OpcS, OpcP, RegReg( dst, src) );
8129 ins_pipe( ialu_reg_reg_alu0 );
8130 %}
8132 // Multiply 32-bit Immediate
8133 instruct mulI_eReg_imm(eRegI dst, eRegI src, immI imm, eFlagsReg cr) %{
8134 match(Set dst (MulI src imm));
8135 effect(KILL cr);
8137 ins_cost(300);
8138 format %{ "IMUL $dst,$src,$imm" %}
8139 opcode(0x69); /* 69 /r id */
8140 ins_encode( OpcSE(imm), RegReg( dst, src ), Con8or32( imm ) );
8141 ins_pipe( ialu_reg_reg_alu0 );
8142 %}
8144 instruct loadConL_low_only(eADXRegL_low_only dst, immL32 src, eFlagsReg cr) %{
8145 match(Set dst src);
8146 effect(KILL cr);
8148 // Note that this is artificially increased to make it more expensive than loadConL
8149 ins_cost(250);
8150 format %{ "MOV EAX,$src\t// low word only" %}
8151 opcode(0xB8);
8152 ins_encode( LdImmL_Lo(dst, src) );
8153 ins_pipe( ialu_reg_fat );
8154 %}
8156 // Multiply by 32-bit Immediate, taking the shifted high order results
8157 // (special case for shift by 32)
8158 instruct mulI_imm_high(eDXRegI dst, nadxRegI src1, eADXRegL_low_only src2, immI_32 cnt, eFlagsReg cr) %{
8159 match(Set dst (ConvL2I (RShiftL (MulL (ConvI2L src1) src2) cnt)));
8160 predicate( _kids[0]->_kids[0]->_kids[1]->_leaf->Opcode() == Op_ConL &&
8161 _kids[0]->_kids[0]->_kids[1]->_leaf->as_Type()->type()->is_long()->get_con() >= min_jint &&
8162 _kids[0]->_kids[0]->_kids[1]->_leaf->as_Type()->type()->is_long()->get_con() <= max_jint );
8163 effect(USE src1, KILL cr);
8165 // Note that this is adjusted by 150 to compensate for the overcosting of loadConL_low_only
8166 ins_cost(0*100 + 1*400 - 150);
8167 format %{ "IMUL EDX:EAX,$src1" %}
8168 ins_encode( multiply_con_and_shift_high( dst, src1, src2, cnt, cr ) );
8169 ins_pipe( pipe_slow );
8170 %}
8172 // Multiply by 32-bit Immediate, taking the shifted high order results
8173 instruct mulI_imm_RShift_high(eDXRegI dst, nadxRegI src1, eADXRegL_low_only src2, immI_32_63 cnt, eFlagsReg cr) %{
8174 match(Set dst (ConvL2I (RShiftL (MulL (ConvI2L src1) src2) cnt)));
8175 predicate( _kids[0]->_kids[0]->_kids[1]->_leaf->Opcode() == Op_ConL &&
8176 _kids[0]->_kids[0]->_kids[1]->_leaf->as_Type()->type()->is_long()->get_con() >= min_jint &&
8177 _kids[0]->_kids[0]->_kids[1]->_leaf->as_Type()->type()->is_long()->get_con() <= max_jint );
8178 effect(USE src1, KILL cr);
8180 // Note that this is adjusted by 150 to compensate for the overcosting of loadConL_low_only
8181 ins_cost(1*100 + 1*400 - 150);
8182 format %{ "IMUL EDX:EAX,$src1\n\t"
8183 "SAR EDX,$cnt-32" %}
8184 ins_encode( multiply_con_and_shift_high( dst, src1, src2, cnt, cr ) );
8185 ins_pipe( pipe_slow );
8186 %}
8188 // Multiply Memory 32-bit Immediate
8189 instruct mulI_mem_imm(eRegI dst, memory src, immI imm, eFlagsReg cr) %{
8190 match(Set dst (MulI (LoadI src) imm));
8191 effect(KILL cr);
8193 ins_cost(300);
8194 format %{ "IMUL $dst,$src,$imm" %}
8195 opcode(0x69); /* 69 /r id */
8196 ins_encode( OpcSE(imm), RegMem( dst, src ), Con8or32( imm ) );
8197 ins_pipe( ialu_reg_mem_alu0 );
8198 %}
8200 // Multiply Memory
8201 instruct mulI(eRegI dst, memory src, eFlagsReg cr) %{
8202 match(Set dst (MulI dst (LoadI src)));
8203 effect(KILL cr);
8205 ins_cost(350);
8206 format %{ "IMUL $dst,$src" %}
8207 opcode(0xAF, 0x0F);
8208 ins_encode( OpcS, OpcP, RegMem( dst, src) );
8209 ins_pipe( ialu_reg_mem_alu0 );
8210 %}
8212 // Multiply Register Int to Long
8213 instruct mulI2L(eADXRegL dst, eAXRegI src, nadxRegI src1, eFlagsReg flags) %{
8214 // Basic Idea: long = (long)int * (long)int
8215 match(Set dst (MulL (ConvI2L src) (ConvI2L src1)));
8216 effect(DEF dst, USE src, USE src1, KILL flags);
8218 ins_cost(300);
8219 format %{ "IMUL $dst,$src1" %}
8221 ins_encode( long_int_multiply( dst, src1 ) );
8222 ins_pipe( ialu_reg_reg_alu0 );
8223 %}
8225 instruct mulIS_eReg(eADXRegL dst, immL_32bits mask, eFlagsReg flags, eAXRegI src, nadxRegI src1) %{
8226 // Basic Idea: long = (int & 0xffffffffL) * (int & 0xffffffffL)
8227 match(Set dst (MulL (AndL (ConvI2L src) mask) (AndL (ConvI2L src1) mask)));
8228 effect(KILL flags);
8230 ins_cost(300);
8231 format %{ "MUL $dst,$src1" %}
8233 ins_encode( long_uint_multiply(dst, src1) );
8234 ins_pipe( ialu_reg_reg_alu0 );
8235 %}
8237 // Multiply Register Long
8238 instruct mulL_eReg(eADXRegL dst, eRegL src, eRegI tmp, eFlagsReg cr) %{
8239 match(Set dst (MulL dst src));
8240 effect(KILL cr, TEMP tmp);
8241 ins_cost(4*100+3*400);
8242 // Basic idea: lo(result) = lo(x_lo * y_lo)
8243 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi)
8244 format %{ "MOV $tmp,$src.lo\n\t"
8245 "IMUL $tmp,EDX\n\t"
8246 "MOV EDX,$src.hi\n\t"
8247 "IMUL EDX,EAX\n\t"
8248 "ADD $tmp,EDX\n\t"
8249 "MUL EDX:EAX,$src.lo\n\t"
8250 "ADD EDX,$tmp" %}
8251 ins_encode( long_multiply( dst, src, tmp ) );
8252 ins_pipe( pipe_slow );
8253 %}
8255 // Multiply Register Long where the left operand's high 32 bits are zero
8256 instruct mulL_eReg_lhi0(eADXRegL dst, eRegL src, eRegI tmp, eFlagsReg cr) %{
8257 predicate(is_operand_hi32_zero(n->in(1)));
8258 match(Set dst (MulL dst src));
8259 effect(KILL cr, TEMP tmp);
8260 ins_cost(2*100+2*400);
8261 // Basic idea: lo(result) = lo(x_lo * y_lo)
8262 // hi(result) = hi(x_lo * y_lo) + lo(x_lo * y_hi) where lo(x_hi * y_lo) = 0 because x_hi = 0
8263 format %{ "MOV $tmp,$src.hi\n\t"
8264 "IMUL $tmp,EAX\n\t"
8265 "MUL EDX:EAX,$src.lo\n\t"
8266 "ADD EDX,$tmp" %}
8267 ins_encode %{
8268 __ movl($tmp$$Register, HIGH_FROM_LOW($src$$Register));
8269 __ imull($tmp$$Register, rax);
8270 __ mull($src$$Register);
8271 __ addl(rdx, $tmp$$Register);
8272 %}
8273 ins_pipe( pipe_slow );
8274 %}
8276 // Multiply Register Long where the right operand's high 32 bits are zero
8277 instruct mulL_eReg_rhi0(eADXRegL dst, eRegL src, eRegI tmp, eFlagsReg cr) %{
8278 predicate(is_operand_hi32_zero(n->in(2)));
8279 match(Set dst (MulL dst src));
8280 effect(KILL cr, TEMP tmp);
8281 ins_cost(2*100+2*400);
8282 // Basic idea: lo(result) = lo(x_lo * y_lo)
8283 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) where lo(x_lo * y_hi) = 0 because y_hi = 0
8284 format %{ "MOV $tmp,$src.lo\n\t"
8285 "IMUL $tmp,EDX\n\t"
8286 "MUL EDX:EAX,$src.lo\n\t"
8287 "ADD EDX,$tmp" %}
8288 ins_encode %{
8289 __ movl($tmp$$Register, $src$$Register);
8290 __ imull($tmp$$Register, rdx);
8291 __ mull($src$$Register);
8292 __ addl(rdx, $tmp$$Register);
8293 %}
8294 ins_pipe( pipe_slow );
8295 %}
8297 // Multiply Register Long where the left and the right operands' high 32 bits are zero
8298 instruct mulL_eReg_hi0(eADXRegL dst, eRegL src, eFlagsReg cr) %{
8299 predicate(is_operand_hi32_zero(n->in(1)) && is_operand_hi32_zero(n->in(2)));
8300 match(Set dst (MulL dst src));
8301 effect(KILL cr);
8302 ins_cost(1*400);
8303 // Basic idea: lo(result) = lo(x_lo * y_lo)
8304 // hi(result) = hi(x_lo * y_lo) where lo(x_hi * y_lo) = 0 and lo(x_lo * y_hi) = 0 because x_hi = 0 and y_hi = 0
8305 format %{ "MUL EDX:EAX,$src.lo\n\t" %}
8306 ins_encode %{
8307 __ mull($src$$Register);
8308 %}
8309 ins_pipe( pipe_slow );
8310 %}
8312 // Multiply Register Long by small constant
8313 instruct mulL_eReg_con(eADXRegL dst, immL_127 src, eRegI tmp, eFlagsReg cr) %{
8314 match(Set dst (MulL dst src));
8315 effect(KILL cr, TEMP tmp);
8316 ins_cost(2*100+2*400);
8317 size(12);
8318 // Basic idea: lo(result) = lo(src * EAX)
8319 // hi(result) = hi(src * EAX) + lo(src * EDX)
8320 format %{ "IMUL $tmp,EDX,$src\n\t"
8321 "MOV EDX,$src\n\t"
8322 "MUL EDX\t# EDX*EAX -> EDX:EAX\n\t"
8323 "ADD EDX,$tmp" %}
8324 ins_encode( long_multiply_con( dst, src, tmp ) );
8325 ins_pipe( pipe_slow );
8326 %}
8328 // Integer DIV with Register
8329 instruct divI_eReg(eAXRegI rax, eDXRegI rdx, eCXRegI div, eFlagsReg cr) %{
8330 match(Set rax (DivI rax div));
8331 effect(KILL rdx, KILL cr);
8332 size(26);
8333 ins_cost(30*100+10*100);
8334 format %{ "CMP EAX,0x80000000\n\t"
8335 "JNE,s normal\n\t"
8336 "XOR EDX,EDX\n\t"
8337 "CMP ECX,-1\n\t"
8338 "JE,s done\n"
8339 "normal: CDQ\n\t"
8340 "IDIV $div\n\t"
8341 "done:" %}
8342 opcode(0xF7, 0x7); /* Opcode F7 /7 */
8343 ins_encode( cdq_enc, OpcP, RegOpc(div) );
8344 ins_pipe( ialu_reg_reg_alu0 );
8345 %}
8347 // Divide Register Long
8348 instruct divL_eReg( eADXRegL dst, eRegL src1, eRegL src2, eFlagsReg cr, eCXRegI cx, eBXRegI bx ) %{
8349 match(Set dst (DivL src1 src2));
8350 effect( KILL cr, KILL cx, KILL bx );
8351 ins_cost(10000);
8352 format %{ "PUSH $src1.hi\n\t"
8353 "PUSH $src1.lo\n\t"
8354 "PUSH $src2.hi\n\t"
8355 "PUSH $src2.lo\n\t"
8356 "CALL SharedRuntime::ldiv\n\t"
8357 "ADD ESP,16" %}
8358 ins_encode( long_div(src1,src2) );
8359 ins_pipe( pipe_slow );
8360 %}
8362 // Integer DIVMOD with Register, both quotient and mod results
8363 instruct divModI_eReg_divmod(eAXRegI rax, eDXRegI rdx, eCXRegI div, eFlagsReg cr) %{
8364 match(DivModI rax div);
8365 effect(KILL cr);
8366 size(26);
8367 ins_cost(30*100+10*100);
8368 format %{ "CMP EAX,0x80000000\n\t"
8369 "JNE,s normal\n\t"
8370 "XOR EDX,EDX\n\t"
8371 "CMP ECX,-1\n\t"
8372 "JE,s done\n"
8373 "normal: CDQ\n\t"
8374 "IDIV $div\n\t"
8375 "done:" %}
8376 opcode(0xF7, 0x7); /* Opcode F7 /7 */
8377 ins_encode( cdq_enc, OpcP, RegOpc(div) );
8378 ins_pipe( pipe_slow );
8379 %}
8381 // Integer MOD with Register
8382 instruct modI_eReg(eDXRegI rdx, eAXRegI rax, eCXRegI div, eFlagsReg cr) %{
8383 match(Set rdx (ModI rax div));
8384 effect(KILL rax, KILL cr);
8386 size(26);
8387 ins_cost(300);
8388 format %{ "CDQ\n\t"
8389 "IDIV $div" %}
8390 opcode(0xF7, 0x7); /* Opcode F7 /7 */
8391 ins_encode( cdq_enc, OpcP, RegOpc(div) );
8392 ins_pipe( ialu_reg_reg_alu0 );
8393 %}
8395 // Remainder Register Long
8396 instruct modL_eReg( eADXRegL dst, eRegL src1, eRegL src2, eFlagsReg cr, eCXRegI cx, eBXRegI bx ) %{
8397 match(Set dst (ModL src1 src2));
8398 effect( KILL cr, KILL cx, KILL bx );
8399 ins_cost(10000);
8400 format %{ "PUSH $src1.hi\n\t"
8401 "PUSH $src1.lo\n\t"
8402 "PUSH $src2.hi\n\t"
8403 "PUSH $src2.lo\n\t"
8404 "CALL SharedRuntime::lrem\n\t"
8405 "ADD ESP,16" %}
8406 ins_encode( long_mod(src1,src2) );
8407 ins_pipe( pipe_slow );
8408 %}
8410 // Divide Register Long (no special case since divisor != -1)
8411 instruct divL_eReg_imm32( eADXRegL dst, immL32 imm, eRegI tmp, eRegI tmp2, eFlagsReg cr ) %{
8412 match(Set dst (DivL dst imm));
8413 effect( TEMP tmp, TEMP tmp2, KILL cr );
8414 ins_cost(1000);
8415 format %{ "MOV $tmp,abs($imm) # ldiv EDX:EAX,$imm\n\t"
8416 "XOR $tmp2,$tmp2\n\t"
8417 "CMP $tmp,EDX\n\t"
8418 "JA,s fast\n\t"
8419 "MOV $tmp2,EAX\n\t"
8420 "MOV EAX,EDX\n\t"
8421 "MOV EDX,0\n\t"
8422 "JLE,s pos\n\t"
8423 "LNEG EAX : $tmp2\n\t"
8424 "DIV $tmp # unsigned division\n\t"
8425 "XCHG EAX,$tmp2\n\t"
8426 "DIV $tmp\n\t"
8427 "LNEG $tmp2 : EAX\n\t"
8428 "JMP,s done\n"
8429 "pos:\n\t"
8430 "DIV $tmp\n\t"
8431 "XCHG EAX,$tmp2\n"
8432 "fast:\n\t"
8433 "DIV $tmp\n"
8434 "done:\n\t"
8435 "MOV EDX,$tmp2\n\t"
8436 "NEG EDX:EAX # if $imm < 0" %}
8437 ins_encode %{
8438 int con = (int)$imm$$constant;
8439 assert(con != 0 && con != -1 && con != min_jint, "wrong divisor");
8440 int pcon = (con > 0) ? con : -con;
8441 Label Lfast, Lpos, Ldone;
8443 __ movl($tmp$$Register, pcon);
8444 __ xorl($tmp2$$Register,$tmp2$$Register);
8445 __ cmpl($tmp$$Register, HIGH_FROM_LOW($dst$$Register));
8446 __ jccb(Assembler::above, Lfast); // result fits into 32 bit
8448 __ movl($tmp2$$Register, $dst$$Register); // save
8449 __ movl($dst$$Register, HIGH_FROM_LOW($dst$$Register));
8450 __ movl(HIGH_FROM_LOW($dst$$Register),0); // preserve flags
8451 __ jccb(Assembler::lessEqual, Lpos); // result is positive
8453 // Negative dividend.
8454 // convert value to positive to use unsigned division
8455 __ lneg($dst$$Register, $tmp2$$Register);
8456 __ divl($tmp$$Register);
8457 __ xchgl($dst$$Register, $tmp2$$Register);
8458 __ divl($tmp$$Register);
8459 // revert result back to negative
8460 __ lneg($tmp2$$Register, $dst$$Register);
8461 __ jmpb(Ldone);
8463 __ bind(Lpos);
8464 __ divl($tmp$$Register); // Use unsigned division
8465 __ xchgl($dst$$Register, $tmp2$$Register);
8466 // Fallthrow for final divide, tmp2 has 32 bit hi result
8468 __ bind(Lfast);
8469 // fast path: src is positive
8470 __ divl($tmp$$Register); // Use unsigned division
8472 __ bind(Ldone);
8473 __ movl(HIGH_FROM_LOW($dst$$Register),$tmp2$$Register);
8474 if (con < 0) {
8475 __ lneg(HIGH_FROM_LOW($dst$$Register), $dst$$Register);
8476 }
8477 %}
8478 ins_pipe( pipe_slow );
8479 %}
8481 // Remainder Register Long (remainder fit into 32 bits)
8482 instruct modL_eReg_imm32( eADXRegL dst, immL32 imm, eRegI tmp, eRegI tmp2, eFlagsReg cr ) %{
8483 match(Set dst (ModL dst imm));
8484 effect( TEMP tmp, TEMP tmp2, KILL cr );
8485 ins_cost(1000);
8486 format %{ "MOV $tmp,abs($imm) # lrem EDX:EAX,$imm\n\t"
8487 "CMP $tmp,EDX\n\t"
8488 "JA,s fast\n\t"
8489 "MOV $tmp2,EAX\n\t"
8490 "MOV EAX,EDX\n\t"
8491 "MOV EDX,0\n\t"
8492 "JLE,s pos\n\t"
8493 "LNEG EAX : $tmp2\n\t"
8494 "DIV $tmp # unsigned division\n\t"
8495 "MOV EAX,$tmp2\n\t"
8496 "DIV $tmp\n\t"
8497 "NEG EDX\n\t"
8498 "JMP,s done\n"
8499 "pos:\n\t"
8500 "DIV $tmp\n\t"
8501 "MOV EAX,$tmp2\n"
8502 "fast:\n\t"
8503 "DIV $tmp\n"
8504 "done:\n\t"
8505 "MOV EAX,EDX\n\t"
8506 "SAR EDX,31\n\t" %}
8507 ins_encode %{
8508 int con = (int)$imm$$constant;
8509 assert(con != 0 && con != -1 && con != min_jint, "wrong divisor");
8510 int pcon = (con > 0) ? con : -con;
8511 Label Lfast, Lpos, Ldone;
8513 __ movl($tmp$$Register, pcon);
8514 __ cmpl($tmp$$Register, HIGH_FROM_LOW($dst$$Register));
8515 __ jccb(Assembler::above, Lfast); // src is positive and result fits into 32 bit
8517 __ movl($tmp2$$Register, $dst$$Register); // save
8518 __ movl($dst$$Register, HIGH_FROM_LOW($dst$$Register));
8519 __ movl(HIGH_FROM_LOW($dst$$Register),0); // preserve flags
8520 __ jccb(Assembler::lessEqual, Lpos); // result is positive
8522 // Negative dividend.
8523 // convert value to positive to use unsigned division
8524 __ lneg($dst$$Register, $tmp2$$Register);
8525 __ divl($tmp$$Register);
8526 __ movl($dst$$Register, $tmp2$$Register);
8527 __ divl($tmp$$Register);
8528 // revert remainder back to negative
8529 __ negl(HIGH_FROM_LOW($dst$$Register));
8530 __ jmpb(Ldone);
8532 __ bind(Lpos);
8533 __ divl($tmp$$Register);
8534 __ movl($dst$$Register, $tmp2$$Register);
8536 __ bind(Lfast);
8537 // fast path: src is positive
8538 __ divl($tmp$$Register);
8540 __ bind(Ldone);
8541 __ movl($dst$$Register, HIGH_FROM_LOW($dst$$Register));
8542 __ sarl(HIGH_FROM_LOW($dst$$Register), 31); // result sign
8544 %}
8545 ins_pipe( pipe_slow );
8546 %}
8548 // Integer Shift Instructions
8549 // Shift Left by one
8550 instruct shlI_eReg_1(eRegI dst, immI1 shift, eFlagsReg cr) %{
8551 match(Set dst (LShiftI dst shift));
8552 effect(KILL cr);
8554 size(2);
8555 format %{ "SHL $dst,$shift" %}
8556 opcode(0xD1, 0x4); /* D1 /4 */
8557 ins_encode( OpcP, RegOpc( dst ) );
8558 ins_pipe( ialu_reg );
8559 %}
8561 // Shift Left by 8-bit immediate
8562 instruct salI_eReg_imm(eRegI dst, immI8 shift, eFlagsReg cr) %{
8563 match(Set dst (LShiftI dst shift));
8564 effect(KILL cr);
8566 size(3);
8567 format %{ "SHL $dst,$shift" %}
8568 opcode(0xC1, 0x4); /* C1 /4 ib */
8569 ins_encode( RegOpcImm( dst, shift) );
8570 ins_pipe( ialu_reg );
8571 %}
8573 // Shift Left by variable
8574 instruct salI_eReg_CL(eRegI dst, eCXRegI shift, eFlagsReg cr) %{
8575 match(Set dst (LShiftI dst shift));
8576 effect(KILL cr);
8578 size(2);
8579 format %{ "SHL $dst,$shift" %}
8580 opcode(0xD3, 0x4); /* D3 /4 */
8581 ins_encode( OpcP, RegOpc( dst ) );
8582 ins_pipe( ialu_reg_reg );
8583 %}
8585 // Arithmetic shift right by one
8586 instruct sarI_eReg_1(eRegI dst, immI1 shift, eFlagsReg cr) %{
8587 match(Set dst (RShiftI dst shift));
8588 effect(KILL cr);
8590 size(2);
8591 format %{ "SAR $dst,$shift" %}
8592 opcode(0xD1, 0x7); /* D1 /7 */
8593 ins_encode( OpcP, RegOpc( dst ) );
8594 ins_pipe( ialu_reg );
8595 %}
8597 // Arithmetic shift right by one
8598 instruct sarI_mem_1(memory dst, immI1 shift, eFlagsReg cr) %{
8599 match(Set dst (StoreI dst (RShiftI (LoadI dst) shift)));
8600 effect(KILL cr);
8601 format %{ "SAR $dst,$shift" %}
8602 opcode(0xD1, 0x7); /* D1 /7 */
8603 ins_encode( OpcP, RMopc_Mem(secondary,dst) );
8604 ins_pipe( ialu_mem_imm );
8605 %}
8607 // Arithmetic Shift Right by 8-bit immediate
8608 instruct sarI_eReg_imm(eRegI dst, immI8 shift, eFlagsReg cr) %{
8609 match(Set dst (RShiftI dst shift));
8610 effect(KILL cr);
8612 size(3);
8613 format %{ "SAR $dst,$shift" %}
8614 opcode(0xC1, 0x7); /* C1 /7 ib */
8615 ins_encode( RegOpcImm( dst, shift ) );
8616 ins_pipe( ialu_mem_imm );
8617 %}
8619 // Arithmetic Shift Right by 8-bit immediate
8620 instruct sarI_mem_imm(memory dst, immI8 shift, eFlagsReg cr) %{
8621 match(Set dst (StoreI dst (RShiftI (LoadI dst) shift)));
8622 effect(KILL cr);
8624 format %{ "SAR $dst,$shift" %}
8625 opcode(0xC1, 0x7); /* C1 /7 ib */
8626 ins_encode( OpcP, RMopc_Mem(secondary, dst ), Con8or32( shift ) );
8627 ins_pipe( ialu_mem_imm );
8628 %}
8630 // Arithmetic Shift Right by variable
8631 instruct sarI_eReg_CL(eRegI dst, eCXRegI shift, eFlagsReg cr) %{
8632 match(Set dst (RShiftI dst shift));
8633 effect(KILL cr);
8635 size(2);
8636 format %{ "SAR $dst,$shift" %}
8637 opcode(0xD3, 0x7); /* D3 /7 */
8638 ins_encode( OpcP, RegOpc( dst ) );
8639 ins_pipe( ialu_reg_reg );
8640 %}
8642 // Logical shift right by one
8643 instruct shrI_eReg_1(eRegI dst, immI1 shift, eFlagsReg cr) %{
8644 match(Set dst (URShiftI dst shift));
8645 effect(KILL cr);
8647 size(2);
8648 format %{ "SHR $dst,$shift" %}
8649 opcode(0xD1, 0x5); /* D1 /5 */
8650 ins_encode( OpcP, RegOpc( dst ) );
8651 ins_pipe( ialu_reg );
8652 %}
8654 // Logical Shift Right by 8-bit immediate
8655 instruct shrI_eReg_imm(eRegI dst, immI8 shift, eFlagsReg cr) %{
8656 match(Set dst (URShiftI dst shift));
8657 effect(KILL cr);
8659 size(3);
8660 format %{ "SHR $dst,$shift" %}
8661 opcode(0xC1, 0x5); /* C1 /5 ib */
8662 ins_encode( RegOpcImm( dst, shift) );
8663 ins_pipe( ialu_reg );
8664 %}
8667 // Logical Shift Right by 24, followed by Arithmetic Shift Left by 24.
8668 // This idiom is used by the compiler for the i2b bytecode.
8669 instruct i2b(eRegI dst, xRegI src, immI_24 twentyfour) %{
8670 match(Set dst (RShiftI (LShiftI src twentyfour) twentyfour));
8672 size(3);
8673 format %{ "MOVSX $dst,$src :8" %}
8674 ins_encode %{
8675 __ movsbl($dst$$Register, $src$$Register);
8676 %}
8677 ins_pipe(ialu_reg_reg);
8678 %}
8680 // Logical Shift Right by 16, followed by Arithmetic Shift Left by 16.
8681 // This idiom is used by the compiler the i2s bytecode.
8682 instruct i2s(eRegI dst, xRegI src, immI_16 sixteen) %{
8683 match(Set dst (RShiftI (LShiftI src sixteen) sixteen));
8685 size(3);
8686 format %{ "MOVSX $dst,$src :16" %}
8687 ins_encode %{
8688 __ movswl($dst$$Register, $src$$Register);
8689 %}
8690 ins_pipe(ialu_reg_reg);
8691 %}
8694 // Logical Shift Right by variable
8695 instruct shrI_eReg_CL(eRegI dst, eCXRegI shift, eFlagsReg cr) %{
8696 match(Set dst (URShiftI dst shift));
8697 effect(KILL cr);
8699 size(2);
8700 format %{ "SHR $dst,$shift" %}
8701 opcode(0xD3, 0x5); /* D3 /5 */
8702 ins_encode( OpcP, RegOpc( dst ) );
8703 ins_pipe( ialu_reg_reg );
8704 %}
8707 //----------Logical Instructions-----------------------------------------------
8708 //----------Integer Logical Instructions---------------------------------------
8709 // And Instructions
8710 // And Register with Register
8711 instruct andI_eReg(eRegI dst, eRegI src, eFlagsReg cr) %{
8712 match(Set dst (AndI dst src));
8713 effect(KILL cr);
8715 size(2);
8716 format %{ "AND $dst,$src" %}
8717 opcode(0x23);
8718 ins_encode( OpcP, RegReg( dst, src) );
8719 ins_pipe( ialu_reg_reg );
8720 %}
8722 // And Register with Immediate
8723 instruct andI_eReg_imm(eRegI dst, immI src, eFlagsReg cr) %{
8724 match(Set dst (AndI dst src));
8725 effect(KILL cr);
8727 format %{ "AND $dst,$src" %}
8728 opcode(0x81,0x04); /* Opcode 81 /4 */
8729 // ins_encode( RegImm( dst, src) );
8730 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
8731 ins_pipe( ialu_reg );
8732 %}
8734 // And Register with Memory
8735 instruct andI_eReg_mem(eRegI dst, memory src, eFlagsReg cr) %{
8736 match(Set dst (AndI dst (LoadI src)));
8737 effect(KILL cr);
8739 ins_cost(125);
8740 format %{ "AND $dst,$src" %}
8741 opcode(0x23);
8742 ins_encode( OpcP, RegMem( dst, src) );
8743 ins_pipe( ialu_reg_mem );
8744 %}
8746 // And Memory with Register
8747 instruct andI_mem_eReg(memory dst, eRegI src, eFlagsReg cr) %{
8748 match(Set dst (StoreI dst (AndI (LoadI dst) src)));
8749 effect(KILL cr);
8751 ins_cost(150);
8752 format %{ "AND $dst,$src" %}
8753 opcode(0x21); /* Opcode 21 /r */
8754 ins_encode( OpcP, RegMem( src, dst ) );
8755 ins_pipe( ialu_mem_reg );
8756 %}
8758 // And Memory with Immediate
8759 instruct andI_mem_imm(memory dst, immI src, eFlagsReg cr) %{
8760 match(Set dst (StoreI dst (AndI (LoadI dst) src)));
8761 effect(KILL cr);
8763 ins_cost(125);
8764 format %{ "AND $dst,$src" %}
8765 opcode(0x81, 0x4); /* Opcode 81 /4 id */
8766 // ins_encode( MemImm( dst, src) );
8767 ins_encode( OpcSE( src ), RMopc_Mem(secondary, dst ), Con8or32( src ) );
8768 ins_pipe( ialu_mem_imm );
8769 %}
8771 // Or Instructions
8772 // Or Register with Register
8773 instruct orI_eReg(eRegI dst, eRegI src, eFlagsReg cr) %{
8774 match(Set dst (OrI dst src));
8775 effect(KILL cr);
8777 size(2);
8778 format %{ "OR $dst,$src" %}
8779 opcode(0x0B);
8780 ins_encode( OpcP, RegReg( dst, src) );
8781 ins_pipe( ialu_reg_reg );
8782 %}
8784 instruct orI_eReg_castP2X(eRegI dst, eRegP src, eFlagsReg cr) %{
8785 match(Set dst (OrI dst (CastP2X src)));
8786 effect(KILL cr);
8788 size(2);
8789 format %{ "OR $dst,$src" %}
8790 opcode(0x0B);
8791 ins_encode( OpcP, RegReg( dst, src) );
8792 ins_pipe( ialu_reg_reg );
8793 %}
8796 // Or Register with Immediate
8797 instruct orI_eReg_imm(eRegI dst, immI src, eFlagsReg cr) %{
8798 match(Set dst (OrI dst src));
8799 effect(KILL cr);
8801 format %{ "OR $dst,$src" %}
8802 opcode(0x81,0x01); /* Opcode 81 /1 id */
8803 // ins_encode( RegImm( dst, src) );
8804 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
8805 ins_pipe( ialu_reg );
8806 %}
8808 // Or Register with Memory
8809 instruct orI_eReg_mem(eRegI dst, memory src, eFlagsReg cr) %{
8810 match(Set dst (OrI dst (LoadI src)));
8811 effect(KILL cr);
8813 ins_cost(125);
8814 format %{ "OR $dst,$src" %}
8815 opcode(0x0B);
8816 ins_encode( OpcP, RegMem( dst, src) );
8817 ins_pipe( ialu_reg_mem );
8818 %}
8820 // Or Memory with Register
8821 instruct orI_mem_eReg(memory dst, eRegI src, eFlagsReg cr) %{
8822 match(Set dst (StoreI dst (OrI (LoadI dst) src)));
8823 effect(KILL cr);
8825 ins_cost(150);
8826 format %{ "OR $dst,$src" %}
8827 opcode(0x09); /* Opcode 09 /r */
8828 ins_encode( OpcP, RegMem( src, dst ) );
8829 ins_pipe( ialu_mem_reg );
8830 %}
8832 // Or Memory with Immediate
8833 instruct orI_mem_imm(memory dst, immI src, eFlagsReg cr) %{
8834 match(Set dst (StoreI dst (OrI (LoadI dst) src)));
8835 effect(KILL cr);
8837 ins_cost(125);
8838 format %{ "OR $dst,$src" %}
8839 opcode(0x81,0x1); /* Opcode 81 /1 id */
8840 // ins_encode( MemImm( dst, src) );
8841 ins_encode( OpcSE( src ), RMopc_Mem(secondary, dst ), Con8or32( src ) );
8842 ins_pipe( ialu_mem_imm );
8843 %}
8845 // ROL/ROR
8846 // ROL expand
8847 instruct rolI_eReg_imm1(eRegI dst, immI1 shift, eFlagsReg cr) %{
8848 effect(USE_DEF dst, USE shift, KILL cr);
8850 format %{ "ROL $dst, $shift" %}
8851 opcode(0xD1, 0x0); /* Opcode D1 /0 */
8852 ins_encode( OpcP, RegOpc( dst ));
8853 ins_pipe( ialu_reg );
8854 %}
8856 instruct rolI_eReg_imm8(eRegI dst, immI8 shift, eFlagsReg cr) %{
8857 effect(USE_DEF dst, USE shift, KILL cr);
8859 format %{ "ROL $dst, $shift" %}
8860 opcode(0xC1, 0x0); /*Opcode /C1 /0 */
8861 ins_encode( RegOpcImm(dst, shift) );
8862 ins_pipe(ialu_reg);
8863 %}
8865 instruct rolI_eReg_CL(ncxRegI dst, eCXRegI shift, eFlagsReg cr) %{
8866 effect(USE_DEF dst, USE shift, KILL cr);
8868 format %{ "ROL $dst, $shift" %}
8869 opcode(0xD3, 0x0); /* Opcode D3 /0 */
8870 ins_encode(OpcP, RegOpc(dst));
8871 ins_pipe( ialu_reg_reg );
8872 %}
8873 // end of ROL expand
8875 // ROL 32bit by one once
8876 instruct rolI_eReg_i1(eRegI dst, immI1 lshift, immI_M1 rshift, eFlagsReg cr) %{
8877 match(Set dst ( OrI (LShiftI dst lshift) (URShiftI dst rshift)));
8879 expand %{
8880 rolI_eReg_imm1(dst, lshift, cr);
8881 %}
8882 %}
8884 // ROL 32bit var by imm8 once
8885 instruct rolI_eReg_i8(eRegI dst, immI8 lshift, immI8 rshift, eFlagsReg cr) %{
8886 predicate( 0 == ((n->in(1)->in(2)->get_int() + n->in(2)->in(2)->get_int()) & 0x1f));
8887 match(Set dst ( OrI (LShiftI dst lshift) (URShiftI dst rshift)));
8889 expand %{
8890 rolI_eReg_imm8(dst, lshift, cr);
8891 %}
8892 %}
8894 // ROL 32bit var by var once
8895 instruct rolI_eReg_Var_C0(ncxRegI dst, eCXRegI shift, immI0 zero, eFlagsReg cr) %{
8896 match(Set dst ( OrI (LShiftI dst shift) (URShiftI dst (SubI zero shift))));
8898 expand %{
8899 rolI_eReg_CL(dst, shift, cr);
8900 %}
8901 %}
8903 // ROL 32bit var by var once
8904 instruct rolI_eReg_Var_C32(ncxRegI dst, eCXRegI shift, immI_32 c32, eFlagsReg cr) %{
8905 match(Set dst ( OrI (LShiftI dst shift) (URShiftI dst (SubI c32 shift))));
8907 expand %{
8908 rolI_eReg_CL(dst, shift, cr);
8909 %}
8910 %}
8912 // ROR expand
8913 instruct rorI_eReg_imm1(eRegI dst, immI1 shift, eFlagsReg cr) %{
8914 effect(USE_DEF dst, USE shift, KILL cr);
8916 format %{ "ROR $dst, $shift" %}
8917 opcode(0xD1,0x1); /* Opcode D1 /1 */
8918 ins_encode( OpcP, RegOpc( dst ) );
8919 ins_pipe( ialu_reg );
8920 %}
8922 instruct rorI_eReg_imm8(eRegI dst, immI8 shift, eFlagsReg cr) %{
8923 effect (USE_DEF dst, USE shift, KILL cr);
8925 format %{ "ROR $dst, $shift" %}
8926 opcode(0xC1, 0x1); /* Opcode /C1 /1 ib */
8927 ins_encode( RegOpcImm(dst, shift) );
8928 ins_pipe( ialu_reg );
8929 %}
8931 instruct rorI_eReg_CL(ncxRegI dst, eCXRegI shift, eFlagsReg cr)%{
8932 effect(USE_DEF dst, USE shift, KILL cr);
8934 format %{ "ROR $dst, $shift" %}
8935 opcode(0xD3, 0x1); /* Opcode D3 /1 */
8936 ins_encode(OpcP, RegOpc(dst));
8937 ins_pipe( ialu_reg_reg );
8938 %}
8939 // end of ROR expand
8941 // ROR right once
8942 instruct rorI_eReg_i1(eRegI dst, immI1 rshift, immI_M1 lshift, eFlagsReg cr) %{
8943 match(Set dst ( OrI (URShiftI dst rshift) (LShiftI dst lshift)));
8945 expand %{
8946 rorI_eReg_imm1(dst, rshift, cr);
8947 %}
8948 %}
8950 // ROR 32bit by immI8 once
8951 instruct rorI_eReg_i8(eRegI dst, immI8 rshift, immI8 lshift, eFlagsReg cr) %{
8952 predicate( 0 == ((n->in(1)->in(2)->get_int() + n->in(2)->in(2)->get_int()) & 0x1f));
8953 match(Set dst ( OrI (URShiftI dst rshift) (LShiftI dst lshift)));
8955 expand %{
8956 rorI_eReg_imm8(dst, rshift, cr);
8957 %}
8958 %}
8960 // ROR 32bit var by var once
8961 instruct rorI_eReg_Var_C0(ncxRegI dst, eCXRegI shift, immI0 zero, eFlagsReg cr) %{
8962 match(Set dst ( OrI (URShiftI dst shift) (LShiftI dst (SubI zero shift))));
8964 expand %{
8965 rorI_eReg_CL(dst, shift, cr);
8966 %}
8967 %}
8969 // ROR 32bit var by var once
8970 instruct rorI_eReg_Var_C32(ncxRegI dst, eCXRegI shift, immI_32 c32, eFlagsReg cr) %{
8971 match(Set dst ( OrI (URShiftI dst shift) (LShiftI dst (SubI c32 shift))));
8973 expand %{
8974 rorI_eReg_CL(dst, shift, cr);
8975 %}
8976 %}
8978 // Xor Instructions
8979 // Xor Register with Register
8980 instruct xorI_eReg(eRegI dst, eRegI src, eFlagsReg cr) %{
8981 match(Set dst (XorI dst src));
8982 effect(KILL cr);
8984 size(2);
8985 format %{ "XOR $dst,$src" %}
8986 opcode(0x33);
8987 ins_encode( OpcP, RegReg( dst, src) );
8988 ins_pipe( ialu_reg_reg );
8989 %}
8991 // Xor Register with Immediate -1
8992 instruct xorI_eReg_im1(eRegI dst, immI_M1 imm) %{
8993 match(Set dst (XorI dst imm));
8995 size(2);
8996 format %{ "NOT $dst" %}
8997 ins_encode %{
8998 __ notl($dst$$Register);
8999 %}
9000 ins_pipe( ialu_reg );
9001 %}
9003 // Xor Register with Immediate
9004 instruct xorI_eReg_imm(eRegI dst, immI src, eFlagsReg cr) %{
9005 match(Set dst (XorI dst src));
9006 effect(KILL cr);
9008 format %{ "XOR $dst,$src" %}
9009 opcode(0x81,0x06); /* Opcode 81 /6 id */
9010 // ins_encode( RegImm( dst, src) );
9011 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
9012 ins_pipe( ialu_reg );
9013 %}
9015 // Xor Register with Memory
9016 instruct xorI_eReg_mem(eRegI dst, memory src, eFlagsReg cr) %{
9017 match(Set dst (XorI dst (LoadI src)));
9018 effect(KILL cr);
9020 ins_cost(125);
9021 format %{ "XOR $dst,$src" %}
9022 opcode(0x33);
9023 ins_encode( OpcP, RegMem(dst, src) );
9024 ins_pipe( ialu_reg_mem );
9025 %}
9027 // Xor Memory with Register
9028 instruct xorI_mem_eReg(memory dst, eRegI src, eFlagsReg cr) %{
9029 match(Set dst (StoreI dst (XorI (LoadI dst) src)));
9030 effect(KILL cr);
9032 ins_cost(150);
9033 format %{ "XOR $dst,$src" %}
9034 opcode(0x31); /* Opcode 31 /r */
9035 ins_encode( OpcP, RegMem( src, dst ) );
9036 ins_pipe( ialu_mem_reg );
9037 %}
9039 // Xor Memory with Immediate
9040 instruct xorI_mem_imm(memory dst, immI src, eFlagsReg cr) %{
9041 match(Set dst (StoreI dst (XorI (LoadI dst) src)));
9042 effect(KILL cr);
9044 ins_cost(125);
9045 format %{ "XOR $dst,$src" %}
9046 opcode(0x81,0x6); /* Opcode 81 /6 id */
9047 ins_encode( OpcSE( src ), RMopc_Mem(secondary, dst ), Con8or32( src ) );
9048 ins_pipe( ialu_mem_imm );
9049 %}
9051 //----------Convert Int to Boolean---------------------------------------------
9053 instruct movI_nocopy(eRegI dst, eRegI src) %{
9054 effect( DEF dst, USE src );
9055 format %{ "MOV $dst,$src" %}
9056 ins_encode( enc_Copy( dst, src) );
9057 ins_pipe( ialu_reg_reg );
9058 %}
9060 instruct ci2b( eRegI dst, eRegI src, eFlagsReg cr ) %{
9061 effect( USE_DEF dst, USE src, KILL cr );
9063 size(4);
9064 format %{ "NEG $dst\n\t"
9065 "ADC $dst,$src" %}
9066 ins_encode( neg_reg(dst),
9067 OpcRegReg(0x13,dst,src) );
9068 ins_pipe( ialu_reg_reg_long );
9069 %}
9071 instruct convI2B( eRegI dst, eRegI src, eFlagsReg cr ) %{
9072 match(Set dst (Conv2B src));
9074 expand %{
9075 movI_nocopy(dst,src);
9076 ci2b(dst,src,cr);
9077 %}
9078 %}
9080 instruct movP_nocopy(eRegI dst, eRegP src) %{
9081 effect( DEF dst, USE src );
9082 format %{ "MOV $dst,$src" %}
9083 ins_encode( enc_Copy( dst, src) );
9084 ins_pipe( ialu_reg_reg );
9085 %}
9087 instruct cp2b( eRegI dst, eRegP src, eFlagsReg cr ) %{
9088 effect( USE_DEF dst, USE src, KILL cr );
9089 format %{ "NEG $dst\n\t"
9090 "ADC $dst,$src" %}
9091 ins_encode( neg_reg(dst),
9092 OpcRegReg(0x13,dst,src) );
9093 ins_pipe( ialu_reg_reg_long );
9094 %}
9096 instruct convP2B( eRegI dst, eRegP src, eFlagsReg cr ) %{
9097 match(Set dst (Conv2B src));
9099 expand %{
9100 movP_nocopy(dst,src);
9101 cp2b(dst,src,cr);
9102 %}
9103 %}
9105 instruct cmpLTMask( eCXRegI dst, ncxRegI p, ncxRegI q, eFlagsReg cr ) %{
9106 match(Set dst (CmpLTMask p q));
9107 effect( KILL cr );
9108 ins_cost(400);
9110 // SETlt can only use low byte of EAX,EBX, ECX, or EDX as destination
9111 format %{ "XOR $dst,$dst\n\t"
9112 "CMP $p,$q\n\t"
9113 "SETlt $dst\n\t"
9114 "NEG $dst" %}
9115 ins_encode( OpcRegReg(0x33,dst,dst),
9116 OpcRegReg(0x3B,p,q),
9117 setLT_reg(dst), neg_reg(dst) );
9118 ins_pipe( pipe_slow );
9119 %}
9121 instruct cmpLTMask0( eRegI dst, immI0 zero, eFlagsReg cr ) %{
9122 match(Set dst (CmpLTMask dst zero));
9123 effect( DEF dst, KILL cr );
9124 ins_cost(100);
9126 format %{ "SAR $dst,31" %}
9127 opcode(0xC1, 0x7); /* C1 /7 ib */
9128 ins_encode( RegOpcImm( dst, 0x1F ) );
9129 ins_pipe( ialu_reg );
9130 %}
9133 instruct cadd_cmpLTMask( ncxRegI p, ncxRegI q, ncxRegI y, eCXRegI tmp, eFlagsReg cr ) %{
9134 match(Set p (AddI (AndI (CmpLTMask p q) y) (SubI p q)));
9135 effect( KILL tmp, KILL cr );
9136 ins_cost(400);
9137 // annoyingly, $tmp has no edges so you cant ask for it in
9138 // any format or encoding
9139 format %{ "SUB $p,$q\n\t"
9140 "SBB ECX,ECX\n\t"
9141 "AND ECX,$y\n\t"
9142 "ADD $p,ECX" %}
9143 ins_encode( enc_cmpLTP(p,q,y,tmp) );
9144 ins_pipe( pipe_cmplt );
9145 %}
9147 /* If I enable this, I encourage spilling in the inner loop of compress.
9148 instruct cadd_cmpLTMask_mem( ncxRegI p, ncxRegI q, memory y, eCXRegI tmp, eFlagsReg cr ) %{
9149 match(Set p (AddI (AndI (CmpLTMask p q) (LoadI y)) (SubI p q)));
9150 effect( USE_KILL tmp, KILL cr );
9151 ins_cost(400);
9153 format %{ "SUB $p,$q\n\t"
9154 "SBB ECX,ECX\n\t"
9155 "AND ECX,$y\n\t"
9156 "ADD $p,ECX" %}
9157 ins_encode( enc_cmpLTP_mem(p,q,y,tmp) );
9158 %}
9159 */
9161 //----------Long Instructions------------------------------------------------
9162 // Add Long Register with Register
9163 instruct addL_eReg(eRegL dst, eRegL src, eFlagsReg cr) %{
9164 match(Set dst (AddL dst src));
9165 effect(KILL cr);
9166 ins_cost(200);
9167 format %{ "ADD $dst.lo,$src.lo\n\t"
9168 "ADC $dst.hi,$src.hi" %}
9169 opcode(0x03, 0x13);
9170 ins_encode( RegReg_Lo(dst, src), RegReg_Hi(dst,src) );
9171 ins_pipe( ialu_reg_reg_long );
9172 %}
9174 // Add Long Register with Immediate
9175 instruct addL_eReg_imm(eRegL dst, immL src, eFlagsReg cr) %{
9176 match(Set dst (AddL dst src));
9177 effect(KILL cr);
9178 format %{ "ADD $dst.lo,$src.lo\n\t"
9179 "ADC $dst.hi,$src.hi" %}
9180 opcode(0x81,0x00,0x02); /* Opcode 81 /0, 81 /2 */
9181 ins_encode( Long_OpcSErm_Lo( dst, src ), Long_OpcSErm_Hi( dst, src ) );
9182 ins_pipe( ialu_reg_long );
9183 %}
9185 // Add Long Register with Memory
9186 instruct addL_eReg_mem(eRegL dst, load_long_memory mem, eFlagsReg cr) %{
9187 match(Set dst (AddL dst (LoadL mem)));
9188 effect(KILL cr);
9189 ins_cost(125);
9190 format %{ "ADD $dst.lo,$mem\n\t"
9191 "ADC $dst.hi,$mem+4" %}
9192 opcode(0x03, 0x13);
9193 ins_encode( OpcP, RegMem( dst, mem), OpcS, RegMem_Hi(dst,mem) );
9194 ins_pipe( ialu_reg_long_mem );
9195 %}
9197 // Subtract Long Register with Register.
9198 instruct subL_eReg(eRegL dst, eRegL src, eFlagsReg cr) %{
9199 match(Set dst (SubL dst src));
9200 effect(KILL cr);
9201 ins_cost(200);
9202 format %{ "SUB $dst.lo,$src.lo\n\t"
9203 "SBB $dst.hi,$src.hi" %}
9204 opcode(0x2B, 0x1B);
9205 ins_encode( RegReg_Lo(dst, src), RegReg_Hi(dst,src) );
9206 ins_pipe( ialu_reg_reg_long );
9207 %}
9209 // Subtract Long Register with Immediate
9210 instruct subL_eReg_imm(eRegL dst, immL src, eFlagsReg cr) %{
9211 match(Set dst (SubL dst src));
9212 effect(KILL cr);
9213 format %{ "SUB $dst.lo,$src.lo\n\t"
9214 "SBB $dst.hi,$src.hi" %}
9215 opcode(0x81,0x05,0x03); /* Opcode 81 /5, 81 /3 */
9216 ins_encode( Long_OpcSErm_Lo( dst, src ), Long_OpcSErm_Hi( dst, src ) );
9217 ins_pipe( ialu_reg_long );
9218 %}
9220 // Subtract Long Register with Memory
9221 instruct subL_eReg_mem(eRegL dst, load_long_memory mem, eFlagsReg cr) %{
9222 match(Set dst (SubL dst (LoadL mem)));
9223 effect(KILL cr);
9224 ins_cost(125);
9225 format %{ "SUB $dst.lo,$mem\n\t"
9226 "SBB $dst.hi,$mem+4" %}
9227 opcode(0x2B, 0x1B);
9228 ins_encode( OpcP, RegMem( dst, mem), OpcS, RegMem_Hi(dst,mem) );
9229 ins_pipe( ialu_reg_long_mem );
9230 %}
9232 instruct negL_eReg(eRegL dst, immL0 zero, eFlagsReg cr) %{
9233 match(Set dst (SubL zero dst));
9234 effect(KILL cr);
9235 ins_cost(300);
9236 format %{ "NEG $dst.hi\n\tNEG $dst.lo\n\tSBB $dst.hi,0" %}
9237 ins_encode( neg_long(dst) );
9238 ins_pipe( ialu_reg_reg_long );
9239 %}
9241 // And Long Register with Register
9242 instruct andL_eReg(eRegL dst, eRegL src, eFlagsReg cr) %{
9243 match(Set dst (AndL dst src));
9244 effect(KILL cr);
9245 format %{ "AND $dst.lo,$src.lo\n\t"
9246 "AND $dst.hi,$src.hi" %}
9247 opcode(0x23,0x23);
9248 ins_encode( RegReg_Lo( dst, src), RegReg_Hi( dst, src) );
9249 ins_pipe( ialu_reg_reg_long );
9250 %}
9252 // And Long Register with Immediate
9253 instruct andL_eReg_imm(eRegL dst, immL src, eFlagsReg cr) %{
9254 match(Set dst (AndL dst src));
9255 effect(KILL cr);
9256 format %{ "AND $dst.lo,$src.lo\n\t"
9257 "AND $dst.hi,$src.hi" %}
9258 opcode(0x81,0x04,0x04); /* Opcode 81 /4, 81 /4 */
9259 ins_encode( Long_OpcSErm_Lo( dst, src ), Long_OpcSErm_Hi( dst, src ) );
9260 ins_pipe( ialu_reg_long );
9261 %}
9263 // And Long Register with Memory
9264 instruct andL_eReg_mem(eRegL dst, load_long_memory mem, eFlagsReg cr) %{
9265 match(Set dst (AndL dst (LoadL mem)));
9266 effect(KILL cr);
9267 ins_cost(125);
9268 format %{ "AND $dst.lo,$mem\n\t"
9269 "AND $dst.hi,$mem+4" %}
9270 opcode(0x23, 0x23);
9271 ins_encode( OpcP, RegMem( dst, mem), OpcS, RegMem_Hi(dst,mem) );
9272 ins_pipe( ialu_reg_long_mem );
9273 %}
9275 // Or Long Register with Register
9276 instruct orl_eReg(eRegL dst, eRegL src, eFlagsReg cr) %{
9277 match(Set dst (OrL dst src));
9278 effect(KILL cr);
9279 format %{ "OR $dst.lo,$src.lo\n\t"
9280 "OR $dst.hi,$src.hi" %}
9281 opcode(0x0B,0x0B);
9282 ins_encode( RegReg_Lo( dst, src), RegReg_Hi( dst, src) );
9283 ins_pipe( ialu_reg_reg_long );
9284 %}
9286 // Or Long Register with Immediate
9287 instruct orl_eReg_imm(eRegL dst, immL src, eFlagsReg cr) %{
9288 match(Set dst (OrL dst src));
9289 effect(KILL cr);
9290 format %{ "OR $dst.lo,$src.lo\n\t"
9291 "OR $dst.hi,$src.hi" %}
9292 opcode(0x81,0x01,0x01); /* Opcode 81 /1, 81 /1 */
9293 ins_encode( Long_OpcSErm_Lo( dst, src ), Long_OpcSErm_Hi( dst, src ) );
9294 ins_pipe( ialu_reg_long );
9295 %}
9297 // Or Long Register with Memory
9298 instruct orl_eReg_mem(eRegL dst, load_long_memory mem, eFlagsReg cr) %{
9299 match(Set dst (OrL dst (LoadL mem)));
9300 effect(KILL cr);
9301 ins_cost(125);
9302 format %{ "OR $dst.lo,$mem\n\t"
9303 "OR $dst.hi,$mem+4" %}
9304 opcode(0x0B,0x0B);
9305 ins_encode( OpcP, RegMem( dst, mem), OpcS, RegMem_Hi(dst,mem) );
9306 ins_pipe( ialu_reg_long_mem );
9307 %}
9309 // Xor Long Register with Register
9310 instruct xorl_eReg(eRegL dst, eRegL src, eFlagsReg cr) %{
9311 match(Set dst (XorL dst src));
9312 effect(KILL cr);
9313 format %{ "XOR $dst.lo,$src.lo\n\t"
9314 "XOR $dst.hi,$src.hi" %}
9315 opcode(0x33,0x33);
9316 ins_encode( RegReg_Lo( dst, src), RegReg_Hi( dst, src) );
9317 ins_pipe( ialu_reg_reg_long );
9318 %}
9320 // Xor Long Register with Immediate -1
9321 instruct xorl_eReg_im1(eRegL dst, immL_M1 imm) %{
9322 match(Set dst (XorL dst imm));
9323 format %{ "NOT $dst.lo\n\t"
9324 "NOT $dst.hi" %}
9325 ins_encode %{
9326 __ notl($dst$$Register);
9327 __ notl(HIGH_FROM_LOW($dst$$Register));
9328 %}
9329 ins_pipe( ialu_reg_long );
9330 %}
9332 // Xor Long Register with Immediate
9333 instruct xorl_eReg_imm(eRegL dst, immL src, eFlagsReg cr) %{
9334 match(Set dst (XorL dst src));
9335 effect(KILL cr);
9336 format %{ "XOR $dst.lo,$src.lo\n\t"
9337 "XOR $dst.hi,$src.hi" %}
9338 opcode(0x81,0x06,0x06); /* Opcode 81 /6, 81 /6 */
9339 ins_encode( Long_OpcSErm_Lo( dst, src ), Long_OpcSErm_Hi( dst, src ) );
9340 ins_pipe( ialu_reg_long );
9341 %}
9343 // Xor Long Register with Memory
9344 instruct xorl_eReg_mem(eRegL dst, load_long_memory mem, eFlagsReg cr) %{
9345 match(Set dst (XorL dst (LoadL mem)));
9346 effect(KILL cr);
9347 ins_cost(125);
9348 format %{ "XOR $dst.lo,$mem\n\t"
9349 "XOR $dst.hi,$mem+4" %}
9350 opcode(0x33,0x33);
9351 ins_encode( OpcP, RegMem( dst, mem), OpcS, RegMem_Hi(dst,mem) );
9352 ins_pipe( ialu_reg_long_mem );
9353 %}
9355 // Shift Left Long by 1
9356 instruct shlL_eReg_1(eRegL dst, immI_1 cnt, eFlagsReg cr) %{
9357 predicate(UseNewLongLShift);
9358 match(Set dst (LShiftL dst cnt));
9359 effect(KILL cr);
9360 ins_cost(100);
9361 format %{ "ADD $dst.lo,$dst.lo\n\t"
9362 "ADC $dst.hi,$dst.hi" %}
9363 ins_encode %{
9364 __ addl($dst$$Register,$dst$$Register);
9365 __ adcl(HIGH_FROM_LOW($dst$$Register),HIGH_FROM_LOW($dst$$Register));
9366 %}
9367 ins_pipe( ialu_reg_long );
9368 %}
9370 // Shift Left Long by 2
9371 instruct shlL_eReg_2(eRegL dst, immI_2 cnt, eFlagsReg cr) %{
9372 predicate(UseNewLongLShift);
9373 match(Set dst (LShiftL dst cnt));
9374 effect(KILL cr);
9375 ins_cost(100);
9376 format %{ "ADD $dst.lo,$dst.lo\n\t"
9377 "ADC $dst.hi,$dst.hi\n\t"
9378 "ADD $dst.lo,$dst.lo\n\t"
9379 "ADC $dst.hi,$dst.hi" %}
9380 ins_encode %{
9381 __ addl($dst$$Register,$dst$$Register);
9382 __ adcl(HIGH_FROM_LOW($dst$$Register),HIGH_FROM_LOW($dst$$Register));
9383 __ addl($dst$$Register,$dst$$Register);
9384 __ adcl(HIGH_FROM_LOW($dst$$Register),HIGH_FROM_LOW($dst$$Register));
9385 %}
9386 ins_pipe( ialu_reg_long );
9387 %}
9389 // Shift Left Long by 3
9390 instruct shlL_eReg_3(eRegL dst, immI_3 cnt, eFlagsReg cr) %{
9391 predicate(UseNewLongLShift);
9392 match(Set dst (LShiftL dst cnt));
9393 effect(KILL cr);
9394 ins_cost(100);
9395 format %{ "ADD $dst.lo,$dst.lo\n\t"
9396 "ADC $dst.hi,$dst.hi\n\t"
9397 "ADD $dst.lo,$dst.lo\n\t"
9398 "ADC $dst.hi,$dst.hi\n\t"
9399 "ADD $dst.lo,$dst.lo\n\t"
9400 "ADC $dst.hi,$dst.hi" %}
9401 ins_encode %{
9402 __ addl($dst$$Register,$dst$$Register);
9403 __ adcl(HIGH_FROM_LOW($dst$$Register),HIGH_FROM_LOW($dst$$Register));
9404 __ addl($dst$$Register,$dst$$Register);
9405 __ adcl(HIGH_FROM_LOW($dst$$Register),HIGH_FROM_LOW($dst$$Register));
9406 __ addl($dst$$Register,$dst$$Register);
9407 __ adcl(HIGH_FROM_LOW($dst$$Register),HIGH_FROM_LOW($dst$$Register));
9408 %}
9409 ins_pipe( ialu_reg_long );
9410 %}
9412 // Shift Left Long by 1-31
9413 instruct shlL_eReg_1_31(eRegL dst, immI_1_31 cnt, eFlagsReg cr) %{
9414 match(Set dst (LShiftL dst cnt));
9415 effect(KILL cr);
9416 ins_cost(200);
9417 format %{ "SHLD $dst.hi,$dst.lo,$cnt\n\t"
9418 "SHL $dst.lo,$cnt" %}
9419 opcode(0xC1, 0x4, 0xA4); /* 0F/A4, then C1 /4 ib */
9420 ins_encode( move_long_small_shift(dst,cnt) );
9421 ins_pipe( ialu_reg_long );
9422 %}
9424 // Shift Left Long by 32-63
9425 instruct shlL_eReg_32_63(eRegL dst, immI_32_63 cnt, eFlagsReg cr) %{
9426 match(Set dst (LShiftL dst cnt));
9427 effect(KILL cr);
9428 ins_cost(300);
9429 format %{ "MOV $dst.hi,$dst.lo\n"
9430 "\tSHL $dst.hi,$cnt-32\n"
9431 "\tXOR $dst.lo,$dst.lo" %}
9432 opcode(0xC1, 0x4); /* C1 /4 ib */
9433 ins_encode( move_long_big_shift_clr(dst,cnt) );
9434 ins_pipe( ialu_reg_long );
9435 %}
9437 // Shift Left Long by variable
9438 instruct salL_eReg_CL(eRegL dst, eCXRegI shift, eFlagsReg cr) %{
9439 match(Set dst (LShiftL dst shift));
9440 effect(KILL cr);
9441 ins_cost(500+200);
9442 size(17);
9443 format %{ "TEST $shift,32\n\t"
9444 "JEQ,s small\n\t"
9445 "MOV $dst.hi,$dst.lo\n\t"
9446 "XOR $dst.lo,$dst.lo\n"
9447 "small:\tSHLD $dst.hi,$dst.lo,$shift\n\t"
9448 "SHL $dst.lo,$shift" %}
9449 ins_encode( shift_left_long( dst, shift ) );
9450 ins_pipe( pipe_slow );
9451 %}
9453 // Shift Right Long by 1-31
9454 instruct shrL_eReg_1_31(eRegL dst, immI_1_31 cnt, eFlagsReg cr) %{
9455 match(Set dst (URShiftL dst cnt));
9456 effect(KILL cr);
9457 ins_cost(200);
9458 format %{ "SHRD $dst.lo,$dst.hi,$cnt\n\t"
9459 "SHR $dst.hi,$cnt" %}
9460 opcode(0xC1, 0x5, 0xAC); /* 0F/AC, then C1 /5 ib */
9461 ins_encode( move_long_small_shift(dst,cnt) );
9462 ins_pipe( ialu_reg_long );
9463 %}
9465 // Shift Right Long by 32-63
9466 instruct shrL_eReg_32_63(eRegL dst, immI_32_63 cnt, eFlagsReg cr) %{
9467 match(Set dst (URShiftL dst cnt));
9468 effect(KILL cr);
9469 ins_cost(300);
9470 format %{ "MOV $dst.lo,$dst.hi\n"
9471 "\tSHR $dst.lo,$cnt-32\n"
9472 "\tXOR $dst.hi,$dst.hi" %}
9473 opcode(0xC1, 0x5); /* C1 /5 ib */
9474 ins_encode( move_long_big_shift_clr(dst,cnt) );
9475 ins_pipe( ialu_reg_long );
9476 %}
9478 // Shift Right Long by variable
9479 instruct shrL_eReg_CL(eRegL dst, eCXRegI shift, eFlagsReg cr) %{
9480 match(Set dst (URShiftL dst shift));
9481 effect(KILL cr);
9482 ins_cost(600);
9483 size(17);
9484 format %{ "TEST $shift,32\n\t"
9485 "JEQ,s small\n\t"
9486 "MOV $dst.lo,$dst.hi\n\t"
9487 "XOR $dst.hi,$dst.hi\n"
9488 "small:\tSHRD $dst.lo,$dst.hi,$shift\n\t"
9489 "SHR $dst.hi,$shift" %}
9490 ins_encode( shift_right_long( dst, shift ) );
9491 ins_pipe( pipe_slow );
9492 %}
9494 // Shift Right Long by 1-31
9495 instruct sarL_eReg_1_31(eRegL dst, immI_1_31 cnt, eFlagsReg cr) %{
9496 match(Set dst (RShiftL dst cnt));
9497 effect(KILL cr);
9498 ins_cost(200);
9499 format %{ "SHRD $dst.lo,$dst.hi,$cnt\n\t"
9500 "SAR $dst.hi,$cnt" %}
9501 opcode(0xC1, 0x7, 0xAC); /* 0F/AC, then C1 /7 ib */
9502 ins_encode( move_long_small_shift(dst,cnt) );
9503 ins_pipe( ialu_reg_long );
9504 %}
9506 // Shift Right Long by 32-63
9507 instruct sarL_eReg_32_63( eRegL dst, immI_32_63 cnt, eFlagsReg cr) %{
9508 match(Set dst (RShiftL dst cnt));
9509 effect(KILL cr);
9510 ins_cost(300);
9511 format %{ "MOV $dst.lo,$dst.hi\n"
9512 "\tSAR $dst.lo,$cnt-32\n"
9513 "\tSAR $dst.hi,31" %}
9514 opcode(0xC1, 0x7); /* C1 /7 ib */
9515 ins_encode( move_long_big_shift_sign(dst,cnt) );
9516 ins_pipe( ialu_reg_long );
9517 %}
9519 // Shift Right arithmetic Long by variable
9520 instruct sarL_eReg_CL(eRegL dst, eCXRegI shift, eFlagsReg cr) %{
9521 match(Set dst (RShiftL dst shift));
9522 effect(KILL cr);
9523 ins_cost(600);
9524 size(18);
9525 format %{ "TEST $shift,32\n\t"
9526 "JEQ,s small\n\t"
9527 "MOV $dst.lo,$dst.hi\n\t"
9528 "SAR $dst.hi,31\n"
9529 "small:\tSHRD $dst.lo,$dst.hi,$shift\n\t"
9530 "SAR $dst.hi,$shift" %}
9531 ins_encode( shift_right_arith_long( dst, shift ) );
9532 ins_pipe( pipe_slow );
9533 %}
9536 //----------Double Instructions------------------------------------------------
9537 // Double Math
9539 // Compare & branch
9541 // P6 version of float compare, sets condition codes in EFLAGS
9542 instruct cmpDPR_cc_P6(eFlagsRegU cr, regDPR src1, regDPR src2, eAXRegI rax) %{
9543 predicate(VM_Version::supports_cmov() && UseSSE <=1);
9544 match(Set cr (CmpD src1 src2));
9545 effect(KILL rax);
9546 ins_cost(150);
9547 format %{ "FLD $src1\n\t"
9548 "FUCOMIP ST,$src2 // P6 instruction\n\t"
9549 "JNP exit\n\t"
9550 "MOV ah,1 // saw a NaN, set CF\n\t"
9551 "SAHF\n"
9552 "exit:\tNOP // avoid branch to branch" %}
9553 opcode(0xDF, 0x05); /* DF E8+i or DF /5 */
9554 ins_encode( Push_Reg_DPR(src1),
9555 OpcP, RegOpc(src2),
9556 cmpF_P6_fixup );
9557 ins_pipe( pipe_slow );
9558 %}
9560 instruct cmpDPR_cc_P6CF(eFlagsRegUCF cr, regDPR src1, regDPR src2) %{
9561 predicate(VM_Version::supports_cmov() && UseSSE <=1);
9562 match(Set cr (CmpD src1 src2));
9563 ins_cost(150);
9564 format %{ "FLD $src1\n\t"
9565 "FUCOMIP ST,$src2 // P6 instruction" %}
9566 opcode(0xDF, 0x05); /* DF E8+i or DF /5 */
9567 ins_encode( Push_Reg_DPR(src1),
9568 OpcP, RegOpc(src2));
9569 ins_pipe( pipe_slow );
9570 %}
9572 // Compare & branch
9573 instruct cmpDPR_cc(eFlagsRegU cr, regDPR src1, regDPR src2, eAXRegI rax) %{
9574 predicate(UseSSE<=1);
9575 match(Set cr (CmpD src1 src2));
9576 effect(KILL rax);
9577 ins_cost(200);
9578 format %{ "FLD $src1\n\t"
9579 "FCOMp $src2\n\t"
9580 "FNSTSW AX\n\t"
9581 "TEST AX,0x400\n\t"
9582 "JZ,s flags\n\t"
9583 "MOV AH,1\t# unordered treat as LT\n"
9584 "flags:\tSAHF" %}
9585 opcode(0xD8, 0x3); /* D8 D8+i or D8 /3 */
9586 ins_encode( Push_Reg_DPR(src1),
9587 OpcP, RegOpc(src2),
9588 fpu_flags);
9589 ins_pipe( pipe_slow );
9590 %}
9592 // Compare vs zero into -1,0,1
9593 instruct cmpDPR_0(eRegI dst, regDPR src1, immDPR0 zero, eAXRegI rax, eFlagsReg cr) %{
9594 predicate(UseSSE<=1);
9595 match(Set dst (CmpD3 src1 zero));
9596 effect(KILL cr, KILL rax);
9597 ins_cost(280);
9598 format %{ "FTSTD $dst,$src1" %}
9599 opcode(0xE4, 0xD9);
9600 ins_encode( Push_Reg_DPR(src1),
9601 OpcS, OpcP, PopFPU,
9602 CmpF_Result(dst));
9603 ins_pipe( pipe_slow );
9604 %}
9606 // Compare into -1,0,1
9607 instruct cmpDPR_reg(eRegI dst, regDPR src1, regDPR src2, eAXRegI rax, eFlagsReg cr) %{
9608 predicate(UseSSE<=1);
9609 match(Set dst (CmpD3 src1 src2));
9610 effect(KILL cr, KILL rax);
9611 ins_cost(300);
9612 format %{ "FCMPD $dst,$src1,$src2" %}
9613 opcode(0xD8, 0x3); /* D8 D8+i or D8 /3 */
9614 ins_encode( Push_Reg_DPR(src1),
9615 OpcP, RegOpc(src2),
9616 CmpF_Result(dst));
9617 ins_pipe( pipe_slow );
9618 %}
9620 // float compare and set condition codes in EFLAGS by XMM regs
9621 instruct cmpD_cc(eFlagsRegU cr, regD src1, regD src2) %{
9622 predicate(UseSSE>=2);
9623 match(Set cr (CmpD src1 src2));
9624 ins_cost(145);
9625 format %{ "UCOMISD $src1,$src2\n\t"
9626 "JNP,s exit\n\t"
9627 "PUSHF\t# saw NaN, set CF\n\t"
9628 "AND [rsp], #0xffffff2b\n\t"
9629 "POPF\n"
9630 "exit:" %}
9631 ins_encode %{
9632 __ ucomisd($src1$$XMMRegister, $src2$$XMMRegister);
9633 emit_cmpfp_fixup(_masm);
9634 %}
9635 ins_pipe( pipe_slow );
9636 %}
9638 instruct cmpD_ccCF(eFlagsRegUCF cr, regD src1, regD src2) %{
9639 predicate(UseSSE>=2);
9640 match(Set cr (CmpD src1 src2));
9641 ins_cost(100);
9642 format %{ "UCOMISD $src1,$src2" %}
9643 ins_encode %{
9644 __ ucomisd($src1$$XMMRegister, $src2$$XMMRegister);
9645 %}
9646 ins_pipe( pipe_slow );
9647 %}
9649 // float compare and set condition codes in EFLAGS by XMM regs
9650 instruct cmpD_ccmem(eFlagsRegU cr, regD src1, memory src2) %{
9651 predicate(UseSSE>=2);
9652 match(Set cr (CmpD src1 (LoadD src2)));
9653 ins_cost(145);
9654 format %{ "UCOMISD $src1,$src2\n\t"
9655 "JNP,s exit\n\t"
9656 "PUSHF\t# saw NaN, set CF\n\t"
9657 "AND [rsp], #0xffffff2b\n\t"
9658 "POPF\n"
9659 "exit:" %}
9660 ins_encode %{
9661 __ ucomisd($src1$$XMMRegister, $src2$$Address);
9662 emit_cmpfp_fixup(_masm);
9663 %}
9664 ins_pipe( pipe_slow );
9665 %}
9667 instruct cmpD_ccmemCF(eFlagsRegUCF cr, regD src1, memory src2) %{
9668 predicate(UseSSE>=2);
9669 match(Set cr (CmpD src1 (LoadD src2)));
9670 ins_cost(100);
9671 format %{ "UCOMISD $src1,$src2" %}
9672 ins_encode %{
9673 __ ucomisd($src1$$XMMRegister, $src2$$Address);
9674 %}
9675 ins_pipe( pipe_slow );
9676 %}
9678 // Compare into -1,0,1 in XMM
9679 instruct cmpD_reg(xRegI dst, regD src1, regD src2, eFlagsReg cr) %{
9680 predicate(UseSSE>=2);
9681 match(Set dst (CmpD3 src1 src2));
9682 effect(KILL cr);
9683 ins_cost(255);
9684 format %{ "UCOMISD $src1, $src2\n\t"
9685 "MOV $dst, #-1\n\t"
9686 "JP,s done\n\t"
9687 "JB,s done\n\t"
9688 "SETNE $dst\n\t"
9689 "MOVZB $dst, $dst\n"
9690 "done:" %}
9691 ins_encode %{
9692 __ ucomisd($src1$$XMMRegister, $src2$$XMMRegister);
9693 emit_cmpfp3(_masm, $dst$$Register);
9694 %}
9695 ins_pipe( pipe_slow );
9696 %}
9698 // Compare into -1,0,1 in XMM and memory
9699 instruct cmpD_regmem(xRegI dst, regD src1, memory src2, eFlagsReg cr) %{
9700 predicate(UseSSE>=2);
9701 match(Set dst (CmpD3 src1 (LoadD src2)));
9702 effect(KILL cr);
9703 ins_cost(275);
9704 format %{ "UCOMISD $src1, $src2\n\t"
9705 "MOV $dst, #-1\n\t"
9706 "JP,s done\n\t"
9707 "JB,s done\n\t"
9708 "SETNE $dst\n\t"
9709 "MOVZB $dst, $dst\n"
9710 "done:" %}
9711 ins_encode %{
9712 __ ucomisd($src1$$XMMRegister, $src2$$Address);
9713 emit_cmpfp3(_masm, $dst$$Register);
9714 %}
9715 ins_pipe( pipe_slow );
9716 %}
9719 instruct subDPR_reg(regDPR dst, regDPR src) %{
9720 predicate (UseSSE <=1);
9721 match(Set dst (SubD dst src));
9723 format %{ "FLD $src\n\t"
9724 "DSUBp $dst,ST" %}
9725 opcode(0xDE, 0x5); /* DE E8+i or DE /5 */
9726 ins_cost(150);
9727 ins_encode( Push_Reg_DPR(src),
9728 OpcP, RegOpc(dst) );
9729 ins_pipe( fpu_reg_reg );
9730 %}
9732 instruct subDPR_reg_round(stackSlotD dst, regDPR src1, regDPR src2) %{
9733 predicate (UseSSE <=1);
9734 match(Set dst (RoundDouble (SubD src1 src2)));
9735 ins_cost(250);
9737 format %{ "FLD $src2\n\t"
9738 "DSUB ST,$src1\n\t"
9739 "FSTP_D $dst\t# D-round" %}
9740 opcode(0xD8, 0x5);
9741 ins_encode( Push_Reg_DPR(src2),
9742 OpcP, RegOpc(src1), Pop_Mem_DPR(dst) );
9743 ins_pipe( fpu_mem_reg_reg );
9744 %}
9747 instruct subDPR_reg_mem(regDPR dst, memory src) %{
9748 predicate (UseSSE <=1);
9749 match(Set dst (SubD dst (LoadD src)));
9750 ins_cost(150);
9752 format %{ "FLD $src\n\t"
9753 "DSUBp $dst,ST" %}
9754 opcode(0xDE, 0x5, 0xDD); /* DE C0+i */ /* LoadD DD /0 */
9755 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src),
9756 OpcP, RegOpc(dst) );
9757 ins_pipe( fpu_reg_mem );
9758 %}
9760 instruct absDPR_reg(regDPR1 dst, regDPR1 src) %{
9761 predicate (UseSSE<=1);
9762 match(Set dst (AbsD src));
9763 ins_cost(100);
9764 format %{ "FABS" %}
9765 opcode(0xE1, 0xD9);
9766 ins_encode( OpcS, OpcP );
9767 ins_pipe( fpu_reg_reg );
9768 %}
9770 instruct negDPR_reg(regDPR1 dst, regDPR1 src) %{
9771 predicate(UseSSE<=1);
9772 match(Set dst (NegD src));
9773 ins_cost(100);
9774 format %{ "FCHS" %}
9775 opcode(0xE0, 0xD9);
9776 ins_encode( OpcS, OpcP );
9777 ins_pipe( fpu_reg_reg );
9778 %}
9780 instruct addDPR_reg(regDPR dst, regDPR src) %{
9781 predicate(UseSSE<=1);
9782 match(Set dst (AddD dst src));
9783 format %{ "FLD $src\n\t"
9784 "DADD $dst,ST" %}
9785 size(4);
9786 ins_cost(150);
9787 opcode(0xDE, 0x0); /* DE C0+i or DE /0*/
9788 ins_encode( Push_Reg_DPR(src),
9789 OpcP, RegOpc(dst) );
9790 ins_pipe( fpu_reg_reg );
9791 %}
9794 instruct addDPR_reg_round(stackSlotD dst, regDPR src1, regDPR src2) %{
9795 predicate(UseSSE<=1);
9796 match(Set dst (RoundDouble (AddD src1 src2)));
9797 ins_cost(250);
9799 format %{ "FLD $src2\n\t"
9800 "DADD ST,$src1\n\t"
9801 "FSTP_D $dst\t# D-round" %}
9802 opcode(0xD8, 0x0); /* D8 C0+i or D8 /0*/
9803 ins_encode( Push_Reg_DPR(src2),
9804 OpcP, RegOpc(src1), Pop_Mem_DPR(dst) );
9805 ins_pipe( fpu_mem_reg_reg );
9806 %}
9809 instruct addDPR_reg_mem(regDPR dst, memory src) %{
9810 predicate(UseSSE<=1);
9811 match(Set dst (AddD dst (LoadD src)));
9812 ins_cost(150);
9814 format %{ "FLD $src\n\t"
9815 "DADDp $dst,ST" %}
9816 opcode(0xDE, 0x0, 0xDD); /* DE C0+i */ /* LoadD DD /0 */
9817 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src),
9818 OpcP, RegOpc(dst) );
9819 ins_pipe( fpu_reg_mem );
9820 %}
9822 // add-to-memory
9823 instruct addDPR_mem_reg(memory dst, regDPR src) %{
9824 predicate(UseSSE<=1);
9825 match(Set dst (StoreD dst (RoundDouble (AddD (LoadD dst) src))));
9826 ins_cost(150);
9828 format %{ "FLD_D $dst\n\t"
9829 "DADD ST,$src\n\t"
9830 "FST_D $dst" %}
9831 opcode(0xDD, 0x0);
9832 ins_encode( Opcode(0xDD), RMopc_Mem(0x00,dst),
9833 Opcode(0xD8), RegOpc(src),
9834 set_instruction_start,
9835 Opcode(0xDD), RMopc_Mem(0x03,dst) );
9836 ins_pipe( fpu_reg_mem );
9837 %}
9839 instruct addDPR_reg_imm1(regDPR dst, immDPR1 con) %{
9840 predicate(UseSSE<=1);
9841 match(Set dst (AddD dst con));
9842 ins_cost(125);
9843 format %{ "FLD1\n\t"
9844 "DADDp $dst,ST" %}
9845 ins_encode %{
9846 __ fld1();
9847 __ faddp($dst$$reg);
9848 %}
9849 ins_pipe(fpu_reg);
9850 %}
9852 instruct addDPR_reg_imm(regDPR dst, immDPR con) %{
9853 predicate(UseSSE<=1 && _kids[1]->_leaf->getd() != 0.0 && _kids[1]->_leaf->getd() != 1.0 );
9854 match(Set dst (AddD dst con));
9855 ins_cost(200);
9856 format %{ "FLD_D [$constantaddress]\t# load from constant table: double=$con\n\t"
9857 "DADDp $dst,ST" %}
9858 ins_encode %{
9859 __ fld_d($constantaddress($con));
9860 __ faddp($dst$$reg);
9861 %}
9862 ins_pipe(fpu_reg_mem);
9863 %}
9865 instruct addDPR_reg_imm_round(stackSlotD dst, regDPR src, immDPR con) %{
9866 predicate(UseSSE<=1 && _kids[0]->_kids[1]->_leaf->getd() != 0.0 && _kids[0]->_kids[1]->_leaf->getd() != 1.0 );
9867 match(Set dst (RoundDouble (AddD src con)));
9868 ins_cost(200);
9869 format %{ "FLD_D [$constantaddress]\t# load from constant table: double=$con\n\t"
9870 "DADD ST,$src\n\t"
9871 "FSTP_D $dst\t# D-round" %}
9872 ins_encode %{
9873 __ fld_d($constantaddress($con));
9874 __ fadd($src$$reg);
9875 __ fstp_d(Address(rsp, $dst$$disp));
9876 %}
9877 ins_pipe(fpu_mem_reg_con);
9878 %}
9880 instruct mulDPR_reg(regDPR dst, regDPR src) %{
9881 predicate(UseSSE<=1);
9882 match(Set dst (MulD dst src));
9883 format %{ "FLD $src\n\t"
9884 "DMULp $dst,ST" %}
9885 opcode(0xDE, 0x1); /* DE C8+i or DE /1*/
9886 ins_cost(150);
9887 ins_encode( Push_Reg_DPR(src),
9888 OpcP, RegOpc(dst) );
9889 ins_pipe( fpu_reg_reg );
9890 %}
9892 // Strict FP instruction biases argument before multiply then
9893 // biases result to avoid double rounding of subnormals.
9894 //
9895 // scale arg1 by multiplying arg1 by 2^(-15360)
9896 // load arg2
9897 // multiply scaled arg1 by arg2
9898 // rescale product by 2^(15360)
9899 //
9900 instruct strictfp_mulDPR_reg(regDPR1 dst, regnotDPR1 src) %{
9901 predicate( UseSSE<=1 && Compile::current()->has_method() && Compile::current()->method()->is_strict() );
9902 match(Set dst (MulD dst src));
9903 ins_cost(1); // Select this instruction for all strict FP double multiplies
9905 format %{ "FLD StubRoutines::_fpu_subnormal_bias1\n\t"
9906 "DMULp $dst,ST\n\t"
9907 "FLD $src\n\t"
9908 "DMULp $dst,ST\n\t"
9909 "FLD StubRoutines::_fpu_subnormal_bias2\n\t"
9910 "DMULp $dst,ST\n\t" %}
9911 opcode(0xDE, 0x1); /* DE C8+i or DE /1*/
9912 ins_encode( strictfp_bias1(dst),
9913 Push_Reg_DPR(src),
9914 OpcP, RegOpc(dst),
9915 strictfp_bias2(dst) );
9916 ins_pipe( fpu_reg_reg );
9917 %}
9919 instruct mulDPR_reg_imm(regDPR dst, immDPR con) %{
9920 predicate( UseSSE<=1 && _kids[1]->_leaf->getd() != 0.0 && _kids[1]->_leaf->getd() != 1.0 );
9921 match(Set dst (MulD dst con));
9922 ins_cost(200);
9923 format %{ "FLD_D [$constantaddress]\t# load from constant table: double=$con\n\t"
9924 "DMULp $dst,ST" %}
9925 ins_encode %{
9926 __ fld_d($constantaddress($con));
9927 __ fmulp($dst$$reg);
9928 %}
9929 ins_pipe(fpu_reg_mem);
9930 %}
9933 instruct mulDPR_reg_mem(regDPR dst, memory src) %{
9934 predicate( UseSSE<=1 );
9935 match(Set dst (MulD dst (LoadD src)));
9936 ins_cost(200);
9937 format %{ "FLD_D $src\n\t"
9938 "DMULp $dst,ST" %}
9939 opcode(0xDE, 0x1, 0xDD); /* DE C8+i or DE /1*/ /* LoadD DD /0 */
9940 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src),
9941 OpcP, RegOpc(dst) );
9942 ins_pipe( fpu_reg_mem );
9943 %}
9945 //
9946 // Cisc-alternate to reg-reg multiply
9947 instruct mulDPR_reg_mem_cisc(regDPR dst, regDPR src, memory mem) %{
9948 predicate( UseSSE<=1 );
9949 match(Set dst (MulD src (LoadD mem)));
9950 ins_cost(250);
9951 format %{ "FLD_D $mem\n\t"
9952 "DMUL ST,$src\n\t"
9953 "FSTP_D $dst" %}
9954 opcode(0xD8, 0x1, 0xD9); /* D8 C8+i */ /* LoadD D9 /0 */
9955 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,mem),
9956 OpcReg_FPR(src),
9957 Pop_Reg_DPR(dst) );
9958 ins_pipe( fpu_reg_reg_mem );
9959 %}
9962 // MACRO3 -- addDPR a mulDPR
9963 // This instruction is a '2-address' instruction in that the result goes
9964 // back to src2. This eliminates a move from the macro; possibly the
9965 // register allocator will have to add it back (and maybe not).
9966 instruct addDPR_mulDPR_reg(regDPR src2, regDPR src1, regDPR src0) %{
9967 predicate( UseSSE<=1 );
9968 match(Set src2 (AddD (MulD src0 src1) src2));
9969 format %{ "FLD $src0\t# ===MACRO3d===\n\t"
9970 "DMUL ST,$src1\n\t"
9971 "DADDp $src2,ST" %}
9972 ins_cost(250);
9973 opcode(0xDD); /* LoadD DD /0 */
9974 ins_encode( Push_Reg_FPR(src0),
9975 FMul_ST_reg(src1),
9976 FAddP_reg_ST(src2) );
9977 ins_pipe( fpu_reg_reg_reg );
9978 %}
9981 // MACRO3 -- subDPR a mulDPR
9982 instruct subDPR_mulDPR_reg(regDPR src2, regDPR src1, regDPR src0) %{
9983 predicate( UseSSE<=1 );
9984 match(Set src2 (SubD (MulD src0 src1) src2));
9985 format %{ "FLD $src0\t# ===MACRO3d===\n\t"
9986 "DMUL ST,$src1\n\t"
9987 "DSUBRp $src2,ST" %}
9988 ins_cost(250);
9989 ins_encode( Push_Reg_FPR(src0),
9990 FMul_ST_reg(src1),
9991 Opcode(0xDE), Opc_plus(0xE0,src2));
9992 ins_pipe( fpu_reg_reg_reg );
9993 %}
9996 instruct divDPR_reg(regDPR dst, regDPR src) %{
9997 predicate( UseSSE<=1 );
9998 match(Set dst (DivD dst src));
10000 format %{ "FLD $src\n\t"
10001 "FDIVp $dst,ST" %}
10002 opcode(0xDE, 0x7); /* DE F8+i or DE /7*/
10003 ins_cost(150);
10004 ins_encode( Push_Reg_DPR(src),
10005 OpcP, RegOpc(dst) );
10006 ins_pipe( fpu_reg_reg );
10007 %}
10009 // Strict FP instruction biases argument before division then
10010 // biases result, to avoid double rounding of subnormals.
10011 //
10012 // scale dividend by multiplying dividend by 2^(-15360)
10013 // load divisor
10014 // divide scaled dividend by divisor
10015 // rescale quotient by 2^(15360)
10016 //
10017 instruct strictfp_divDPR_reg(regDPR1 dst, regnotDPR1 src) %{
10018 predicate (UseSSE<=1);
10019 match(Set dst (DivD dst src));
10020 predicate( UseSSE<=1 && Compile::current()->has_method() && Compile::current()->method()->is_strict() );
10021 ins_cost(01);
10023 format %{ "FLD StubRoutines::_fpu_subnormal_bias1\n\t"
10024 "DMULp $dst,ST\n\t"
10025 "FLD $src\n\t"
10026 "FDIVp $dst,ST\n\t"
10027 "FLD StubRoutines::_fpu_subnormal_bias2\n\t"
10028 "DMULp $dst,ST\n\t" %}
10029 opcode(0xDE, 0x7); /* DE F8+i or DE /7*/
10030 ins_encode( strictfp_bias1(dst),
10031 Push_Reg_DPR(src),
10032 OpcP, RegOpc(dst),
10033 strictfp_bias2(dst) );
10034 ins_pipe( fpu_reg_reg );
10035 %}
10037 instruct divDPR_reg_round(stackSlotD dst, regDPR src1, regDPR src2) %{
10038 predicate( UseSSE<=1 && !(Compile::current()->has_method() && Compile::current()->method()->is_strict()) );
10039 match(Set dst (RoundDouble (DivD src1 src2)));
10041 format %{ "FLD $src1\n\t"
10042 "FDIV ST,$src2\n\t"
10043 "FSTP_D $dst\t# D-round" %}
10044 opcode(0xD8, 0x6); /* D8 F0+i or D8 /6 */
10045 ins_encode( Push_Reg_DPR(src1),
10046 OpcP, RegOpc(src2), Pop_Mem_DPR(dst) );
10047 ins_pipe( fpu_mem_reg_reg );
10048 %}
10051 instruct modDPR_reg(regDPR dst, regDPR src, eAXRegI rax, eFlagsReg cr) %{
10052 predicate(UseSSE<=1);
10053 match(Set dst (ModD dst src));
10054 effect(KILL rax, KILL cr); // emitModDPR() uses EAX and EFLAGS
10056 format %{ "DMOD $dst,$src" %}
10057 ins_cost(250);
10058 ins_encode(Push_Reg_Mod_DPR(dst, src),
10059 emitModDPR(),
10060 Push_Result_Mod_DPR(src),
10061 Pop_Reg_DPR(dst));
10062 ins_pipe( pipe_slow );
10063 %}
10065 instruct modD_reg(regD dst, regD src0, regD src1, eAXRegI rax, eFlagsReg cr) %{
10066 predicate(UseSSE>=2);
10067 match(Set dst (ModD src0 src1));
10068 effect(KILL rax, KILL cr);
10070 format %{ "SUB ESP,8\t # DMOD\n"
10071 "\tMOVSD [ESP+0],$src1\n"
10072 "\tFLD_D [ESP+0]\n"
10073 "\tMOVSD [ESP+0],$src0\n"
10074 "\tFLD_D [ESP+0]\n"
10075 "loop:\tFPREM\n"
10076 "\tFWAIT\n"
10077 "\tFNSTSW AX\n"
10078 "\tSAHF\n"
10079 "\tJP loop\n"
10080 "\tFSTP_D [ESP+0]\n"
10081 "\tMOVSD $dst,[ESP+0]\n"
10082 "\tADD ESP,8\n"
10083 "\tFSTP ST0\t # Restore FPU Stack"
10084 %}
10085 ins_cost(250);
10086 ins_encode( Push_ModD_encoding(src0, src1), emitModDPR(), Push_ResultD(dst), PopFPU);
10087 ins_pipe( pipe_slow );
10088 %}
10090 instruct sinDPR_reg(regDPR1 dst, regDPR1 src) %{
10091 predicate (UseSSE<=1);
10092 match(Set dst (SinD src));
10093 ins_cost(1800);
10094 format %{ "DSIN $dst" %}
10095 opcode(0xD9, 0xFE);
10096 ins_encode( OpcP, OpcS );
10097 ins_pipe( pipe_slow );
10098 %}
10100 instruct sinD_reg(regD dst, eFlagsReg cr) %{
10101 predicate (UseSSE>=2);
10102 match(Set dst (SinD dst));
10103 effect(KILL cr); // Push_{Src|Result}D() uses "{SUB|ADD} ESP,8"
10104 ins_cost(1800);
10105 format %{ "DSIN $dst" %}
10106 opcode(0xD9, 0xFE);
10107 ins_encode( Push_SrcD(dst), OpcP, OpcS, Push_ResultD(dst) );
10108 ins_pipe( pipe_slow );
10109 %}
10111 instruct cosDPR_reg(regDPR1 dst, regDPR1 src) %{
10112 predicate (UseSSE<=1);
10113 match(Set dst (CosD src));
10114 ins_cost(1800);
10115 format %{ "DCOS $dst" %}
10116 opcode(0xD9, 0xFF);
10117 ins_encode( OpcP, OpcS );
10118 ins_pipe( pipe_slow );
10119 %}
10121 instruct cosD_reg(regD dst, eFlagsReg cr) %{
10122 predicate (UseSSE>=2);
10123 match(Set dst (CosD dst));
10124 effect(KILL cr); // Push_{Src|Result}D() uses "{SUB|ADD} ESP,8"
10125 ins_cost(1800);
10126 format %{ "DCOS $dst" %}
10127 opcode(0xD9, 0xFF);
10128 ins_encode( Push_SrcD(dst), OpcP, OpcS, Push_ResultD(dst) );
10129 ins_pipe( pipe_slow );
10130 %}
10132 instruct tanDPR_reg(regDPR1 dst, regDPR1 src) %{
10133 predicate (UseSSE<=1);
10134 match(Set dst(TanD src));
10135 format %{ "DTAN $dst" %}
10136 ins_encode( Opcode(0xD9), Opcode(0xF2), // fptan
10137 Opcode(0xDD), Opcode(0xD8)); // fstp st
10138 ins_pipe( pipe_slow );
10139 %}
10141 instruct tanD_reg(regD dst, eFlagsReg cr) %{
10142 predicate (UseSSE>=2);
10143 match(Set dst(TanD dst));
10144 effect(KILL cr); // Push_{Src|Result}D() uses "{SUB|ADD} ESP,8"
10145 format %{ "DTAN $dst" %}
10146 ins_encode( Push_SrcD(dst),
10147 Opcode(0xD9), Opcode(0xF2), // fptan
10148 Opcode(0xDD), Opcode(0xD8), // fstp st
10149 Push_ResultD(dst) );
10150 ins_pipe( pipe_slow );
10151 %}
10153 instruct atanDPR_reg(regDPR dst, regDPR src) %{
10154 predicate (UseSSE<=1);
10155 match(Set dst(AtanD dst src));
10156 format %{ "DATA $dst,$src" %}
10157 opcode(0xD9, 0xF3);
10158 ins_encode( Push_Reg_DPR(src),
10159 OpcP, OpcS, RegOpc(dst) );
10160 ins_pipe( pipe_slow );
10161 %}
10163 instruct atanD_reg(regD dst, regD src, eFlagsReg cr) %{
10164 predicate (UseSSE>=2);
10165 match(Set dst(AtanD dst src));
10166 effect(KILL cr); // Push_{Src|Result}D() uses "{SUB|ADD} ESP,8"
10167 format %{ "DATA $dst,$src" %}
10168 opcode(0xD9, 0xF3);
10169 ins_encode( Push_SrcD(src),
10170 OpcP, OpcS, Push_ResultD(dst) );
10171 ins_pipe( pipe_slow );
10172 %}
10174 instruct sqrtDPR_reg(regDPR dst, regDPR src) %{
10175 predicate (UseSSE<=1);
10176 match(Set dst (SqrtD src));
10177 format %{ "DSQRT $dst,$src" %}
10178 opcode(0xFA, 0xD9);
10179 ins_encode( Push_Reg_DPR(src),
10180 OpcS, OpcP, Pop_Reg_DPR(dst) );
10181 ins_pipe( pipe_slow );
10182 %}
10184 instruct powDPR_reg(regDPR X, regDPR1 Y, eAXRegI rax, eBXRegI rbx, eCXRegI rcx) %{
10185 predicate (UseSSE<=1);
10186 match(Set Y (PowD X Y)); // Raise X to the Yth power
10187 effect(KILL rax, KILL rbx, KILL rcx);
10188 format %{ "SUB ESP,8\t\t# Fast-path POW encoding\n\t"
10189 "FLD_D $X\n\t"
10190 "FYL2X \t\t\t# Q=Y*ln2(X)\n\t"
10192 "FDUP \t\t\t# Q Q\n\t"
10193 "FRNDINT\t\t\t# int(Q) Q\n\t"
10194 "FSUB ST(1),ST(0)\t# int(Q) frac(Q)\n\t"
10195 "FISTP dword [ESP]\n\t"
10196 "F2XM1 \t\t\t# 2^frac(Q)-1 int(Q)\n\t"
10197 "FLD1 \t\t\t# 1 2^frac(Q)-1 int(Q)\n\t"
10198 "FADDP \t\t\t# 2^frac(Q) int(Q)\n\t" // could use FADD [1.000] instead
10199 "MOV EAX,[ESP]\t# Pick up int(Q)\n\t"
10200 "MOV ECX,0xFFFFF800\t# Overflow mask\n\t"
10201 "ADD EAX,1023\t\t# Double exponent bias\n\t"
10202 "MOV EBX,EAX\t\t# Preshifted biased expo\n\t"
10203 "SHL EAX,20\t\t# Shift exponent into place\n\t"
10204 "TEST EBX,ECX\t\t# Check for overflow\n\t"
10205 "CMOVne EAX,ECX\t\t# If overflow, stuff NaN into EAX\n\t"
10206 "MOV [ESP+4],EAX\t# Marshal 64-bit scaling double\n\t"
10207 "MOV [ESP+0],0\n\t"
10208 "FMUL ST(0),[ESP+0]\t# Scale\n\t"
10210 "ADD ESP,8"
10211 %}
10212 ins_encode( push_stack_temp_qword,
10213 Push_Reg_DPR(X),
10214 Opcode(0xD9), Opcode(0xF1), // fyl2x
10215 pow_exp_core_encoding,
10216 pop_stack_temp_qword);
10217 ins_pipe( pipe_slow );
10218 %}
10220 instruct powD_reg(regD dst, regD src0, regD src1, regDPR1 tmp1, eAXRegI rax, eBXRegI rbx, eCXRegI rcx ) %{
10221 predicate (UseSSE>=2);
10222 match(Set dst (PowD src0 src1)); // Raise src0 to the src1'th power
10223 effect(KILL tmp1, KILL rax, KILL rbx, KILL rcx );
10224 format %{ "SUB ESP,8\t\t# Fast-path POW encoding\n\t"
10225 "MOVSD [ESP],$src1\n\t"
10226 "FLD FPR1,$src1\n\t"
10227 "MOVSD [ESP],$src0\n\t"
10228 "FLD FPR1,$src0\n\t"
10229 "FYL2X \t\t\t# Q=Y*ln2(X)\n\t"
10231 "FDUP \t\t\t# Q Q\n\t"
10232 "FRNDINT\t\t\t# int(Q) Q\n\t"
10233 "FSUB ST(1),ST(0)\t# int(Q) frac(Q)\n\t"
10234 "FISTP dword [ESP]\n\t"
10235 "F2XM1 \t\t\t# 2^frac(Q)-1 int(Q)\n\t"
10236 "FLD1 \t\t\t# 1 2^frac(Q)-1 int(Q)\n\t"
10237 "FADDP \t\t\t# 2^frac(Q) int(Q)\n\t" // could use FADD [1.000] instead
10238 "MOV EAX,[ESP]\t# Pick up int(Q)\n\t"
10239 "MOV ECX,0xFFFFF800\t# Overflow mask\n\t"
10240 "ADD EAX,1023\t\t# Double exponent bias\n\t"
10241 "MOV EBX,EAX\t\t# Preshifted biased expo\n\t"
10242 "SHL EAX,20\t\t# Shift exponent into place\n\t"
10243 "TEST EBX,ECX\t\t# Check for overflow\n\t"
10244 "CMOVne EAX,ECX\t\t# If overflow, stuff NaN into EAX\n\t"
10245 "MOV [ESP+4],EAX\t# Marshal 64-bit scaling double\n\t"
10246 "MOV [ESP+0],0\n\t"
10247 "FMUL ST(0),[ESP+0]\t# Scale\n\t"
10249 "FST_D [ESP]\n\t"
10250 "MOVSD $dst,[ESP]\n\t"
10251 "ADD ESP,8"
10252 %}
10253 ins_encode( push_stack_temp_qword,
10254 push_xmm_to_fpr1(src1),
10255 push_xmm_to_fpr1(src0),
10256 Opcode(0xD9), Opcode(0xF1), // fyl2x
10257 pow_exp_core_encoding,
10258 Push_ResultD(dst) );
10259 ins_pipe( pipe_slow );
10260 %}
10263 instruct expDPR_reg(regDPR1 dpr1, eAXRegI rax, eBXRegI rbx, eCXRegI rcx) %{
10264 predicate (UseSSE<=1);
10265 match(Set dpr1 (ExpD dpr1));
10266 effect(KILL rax, KILL rbx, KILL rcx);
10267 format %{ "SUB ESP,8\t\t# Fast-path EXP encoding"
10268 "FLDL2E \t\t\t# Ld log2(e) X\n\t"
10269 "FMULP \t\t\t# Q=X*log2(e)\n\t"
10271 "FDUP \t\t\t# Q Q\n\t"
10272 "FRNDINT\t\t\t# int(Q) Q\n\t"
10273 "FSUB ST(1),ST(0)\t# int(Q) frac(Q)\n\t"
10274 "FISTP dword [ESP]\n\t"
10275 "F2XM1 \t\t\t# 2^frac(Q)-1 int(Q)\n\t"
10276 "FLD1 \t\t\t# 1 2^frac(Q)-1 int(Q)\n\t"
10277 "FADDP \t\t\t# 2^frac(Q) int(Q)\n\t" // could use FADD [1.000] instead
10278 "MOV EAX,[ESP]\t# Pick up int(Q)\n\t"
10279 "MOV ECX,0xFFFFF800\t# Overflow mask\n\t"
10280 "ADD EAX,1023\t\t# Double exponent bias\n\t"
10281 "MOV EBX,EAX\t\t# Preshifted biased expo\n\t"
10282 "SHL EAX,20\t\t# Shift exponent into place\n\t"
10283 "TEST EBX,ECX\t\t# Check for overflow\n\t"
10284 "CMOVne EAX,ECX\t\t# If overflow, stuff NaN into EAX\n\t"
10285 "MOV [ESP+4],EAX\t# Marshal 64-bit scaling double\n\t"
10286 "MOV [ESP+0],0\n\t"
10287 "FMUL ST(0),[ESP+0]\t# Scale\n\t"
10289 "ADD ESP,8"
10290 %}
10291 ins_encode( push_stack_temp_qword,
10292 Opcode(0xD9), Opcode(0xEA), // fldl2e
10293 Opcode(0xDE), Opcode(0xC9), // fmulp
10294 pow_exp_core_encoding,
10295 pop_stack_temp_qword);
10296 ins_pipe( pipe_slow );
10297 %}
10299 instruct expD_reg(regD dst, regD src, regDPR1 tmp1, eAXRegI rax, eBXRegI rbx, eCXRegI rcx) %{
10300 predicate (UseSSE>=2);
10301 match(Set dst (ExpD src));
10302 effect(KILL tmp1, KILL rax, KILL rbx, KILL rcx);
10303 format %{ "SUB ESP,8\t\t# Fast-path EXP encoding\n\t"
10304 "MOVSD [ESP],$src\n\t"
10305 "FLDL2E \t\t\t# Ld log2(e) X\n\t"
10306 "FMULP \t\t\t# Q=X*log2(e) X\n\t"
10308 "FDUP \t\t\t# Q Q\n\t"
10309 "FRNDINT\t\t\t# int(Q) Q\n\t"
10310 "FSUB ST(1),ST(0)\t# int(Q) frac(Q)\n\t"
10311 "FISTP dword [ESP]\n\t"
10312 "F2XM1 \t\t\t# 2^frac(Q)-1 int(Q)\n\t"
10313 "FLD1 \t\t\t# 1 2^frac(Q)-1 int(Q)\n\t"
10314 "FADDP \t\t\t# 2^frac(Q) int(Q)\n\t" // could use FADD [1.000] instead
10315 "MOV EAX,[ESP]\t# Pick up int(Q)\n\t"
10316 "MOV ECX,0xFFFFF800\t# Overflow mask\n\t"
10317 "ADD EAX,1023\t\t# Double exponent bias\n\t"
10318 "MOV EBX,EAX\t\t# Preshifted biased expo\n\t"
10319 "SHL EAX,20\t\t# Shift exponent into place\n\t"
10320 "TEST EBX,ECX\t\t# Check for overflow\n\t"
10321 "CMOVne EAX,ECX\t\t# If overflow, stuff NaN into EAX\n\t"
10322 "MOV [ESP+4],EAX\t# Marshal 64-bit scaling double\n\t"
10323 "MOV [ESP+0],0\n\t"
10324 "FMUL ST(0),[ESP+0]\t# Scale\n\t"
10326 "FST_D [ESP]\n\t"
10327 "MOVSD $dst,[ESP]\n\t"
10328 "ADD ESP,8"
10329 %}
10330 ins_encode( Push_SrcD(src),
10331 Opcode(0xD9), Opcode(0xEA), // fldl2e
10332 Opcode(0xDE), Opcode(0xC9), // fmulp
10333 pow_exp_core_encoding,
10334 Push_ResultD(dst) );
10335 ins_pipe( pipe_slow );
10336 %}
10340 instruct log10DPR_reg(regDPR1 dst, regDPR1 src) %{
10341 predicate (UseSSE<=1);
10342 // The source Double operand on FPU stack
10343 match(Set dst (Log10D src));
10344 // fldlg2 ; push log_10(2) on the FPU stack; full 80-bit number
10345 // fxch ; swap ST(0) with ST(1)
10346 // fyl2x ; compute log_10(2) * log_2(x)
10347 format %{ "FLDLG2 \t\t\t#Log10\n\t"
10348 "FXCH \n\t"
10349 "FYL2X \t\t\t# Q=Log10*Log_2(x)"
10350 %}
10351 ins_encode( Opcode(0xD9), Opcode(0xEC), // fldlg2
10352 Opcode(0xD9), Opcode(0xC9), // fxch
10353 Opcode(0xD9), Opcode(0xF1)); // fyl2x
10355 ins_pipe( pipe_slow );
10356 %}
10358 instruct log10D_reg(regD dst, regD src, eFlagsReg cr) %{
10359 predicate (UseSSE>=2);
10360 effect(KILL cr);
10361 match(Set dst (Log10D src));
10362 // fldlg2 ; push log_10(2) on the FPU stack; full 80-bit number
10363 // fyl2x ; compute log_10(2) * log_2(x)
10364 format %{ "FLDLG2 \t\t\t#Log10\n\t"
10365 "FYL2X \t\t\t# Q=Log10*Log_2(x)"
10366 %}
10367 ins_encode( Opcode(0xD9), Opcode(0xEC), // fldlg2
10368 Push_SrcD(src),
10369 Opcode(0xD9), Opcode(0xF1), // fyl2x
10370 Push_ResultD(dst));
10372 ins_pipe( pipe_slow );
10373 %}
10375 instruct logDPR_reg(regDPR1 dst, regDPR1 src) %{
10376 predicate (UseSSE<=1);
10377 // The source Double operand on FPU stack
10378 match(Set dst (LogD src));
10379 // fldln2 ; push log_e(2) on the FPU stack; full 80-bit number
10380 // fxch ; swap ST(0) with ST(1)
10381 // fyl2x ; compute log_e(2) * log_2(x)
10382 format %{ "FLDLN2 \t\t\t#Log_e\n\t"
10383 "FXCH \n\t"
10384 "FYL2X \t\t\t# Q=Log_e*Log_2(x)"
10385 %}
10386 ins_encode( Opcode(0xD9), Opcode(0xED), // fldln2
10387 Opcode(0xD9), Opcode(0xC9), // fxch
10388 Opcode(0xD9), Opcode(0xF1)); // fyl2x
10390 ins_pipe( pipe_slow );
10391 %}
10393 instruct logD_reg(regD dst, regD src, eFlagsReg cr) %{
10394 predicate (UseSSE>=2);
10395 effect(KILL cr);
10396 // The source and result Double operands in XMM registers
10397 match(Set dst (LogD src));
10398 // fldln2 ; push log_e(2) on the FPU stack; full 80-bit number
10399 // fyl2x ; compute log_e(2) * log_2(x)
10400 format %{ "FLDLN2 \t\t\t#Log_e\n\t"
10401 "FYL2X \t\t\t# Q=Log_e*Log_2(x)"
10402 %}
10403 ins_encode( Opcode(0xD9), Opcode(0xED), // fldln2
10404 Push_SrcD(src),
10405 Opcode(0xD9), Opcode(0xF1), // fyl2x
10406 Push_ResultD(dst));
10407 ins_pipe( pipe_slow );
10408 %}
10410 //-------------Float Instructions-------------------------------
10411 // Float Math
10413 // Code for float compare:
10414 // fcompp();
10415 // fwait(); fnstsw_ax();
10416 // sahf();
10417 // movl(dst, unordered_result);
10418 // jcc(Assembler::parity, exit);
10419 // movl(dst, less_result);
10420 // jcc(Assembler::below, exit);
10421 // movl(dst, equal_result);
10422 // jcc(Assembler::equal, exit);
10423 // movl(dst, greater_result);
10424 // exit:
10426 // P6 version of float compare, sets condition codes in EFLAGS
10427 instruct cmpFPR_cc_P6(eFlagsRegU cr, regFPR src1, regFPR src2, eAXRegI rax) %{
10428 predicate(VM_Version::supports_cmov() && UseSSE == 0);
10429 match(Set cr (CmpF src1 src2));
10430 effect(KILL rax);
10431 ins_cost(150);
10432 format %{ "FLD $src1\n\t"
10433 "FUCOMIP ST,$src2 // P6 instruction\n\t"
10434 "JNP exit\n\t"
10435 "MOV ah,1 // saw a NaN, set CF (treat as LT)\n\t"
10436 "SAHF\n"
10437 "exit:\tNOP // avoid branch to branch" %}
10438 opcode(0xDF, 0x05); /* DF E8+i or DF /5 */
10439 ins_encode( Push_Reg_DPR(src1),
10440 OpcP, RegOpc(src2),
10441 cmpF_P6_fixup );
10442 ins_pipe( pipe_slow );
10443 %}
10445 instruct cmpFPR_cc_P6CF(eFlagsRegUCF cr, regFPR src1, regFPR src2) %{
10446 predicate(VM_Version::supports_cmov() && UseSSE == 0);
10447 match(Set cr (CmpF src1 src2));
10448 ins_cost(100);
10449 format %{ "FLD $src1\n\t"
10450 "FUCOMIP ST,$src2 // P6 instruction" %}
10451 opcode(0xDF, 0x05); /* DF E8+i or DF /5 */
10452 ins_encode( Push_Reg_DPR(src1),
10453 OpcP, RegOpc(src2));
10454 ins_pipe( pipe_slow );
10455 %}
10458 // Compare & branch
10459 instruct cmpFPR_cc(eFlagsRegU cr, regFPR src1, regFPR src2, eAXRegI rax) %{
10460 predicate(UseSSE == 0);
10461 match(Set cr (CmpF src1 src2));
10462 effect(KILL rax);
10463 ins_cost(200);
10464 format %{ "FLD $src1\n\t"
10465 "FCOMp $src2\n\t"
10466 "FNSTSW AX\n\t"
10467 "TEST AX,0x400\n\t"
10468 "JZ,s flags\n\t"
10469 "MOV AH,1\t# unordered treat as LT\n"
10470 "flags:\tSAHF" %}
10471 opcode(0xD8, 0x3); /* D8 D8+i or D8 /3 */
10472 ins_encode( Push_Reg_DPR(src1),
10473 OpcP, RegOpc(src2),
10474 fpu_flags);
10475 ins_pipe( pipe_slow );
10476 %}
10478 // Compare vs zero into -1,0,1
10479 instruct cmpFPR_0(eRegI dst, regFPR src1, immFPR0 zero, eAXRegI rax, eFlagsReg cr) %{
10480 predicate(UseSSE == 0);
10481 match(Set dst (CmpF3 src1 zero));
10482 effect(KILL cr, KILL rax);
10483 ins_cost(280);
10484 format %{ "FTSTF $dst,$src1" %}
10485 opcode(0xE4, 0xD9);
10486 ins_encode( Push_Reg_DPR(src1),
10487 OpcS, OpcP, PopFPU,
10488 CmpF_Result(dst));
10489 ins_pipe( pipe_slow );
10490 %}
10492 // Compare into -1,0,1
10493 instruct cmpFPR_reg(eRegI dst, regFPR src1, regFPR src2, eAXRegI rax, eFlagsReg cr) %{
10494 predicate(UseSSE == 0);
10495 match(Set dst (CmpF3 src1 src2));
10496 effect(KILL cr, KILL rax);
10497 ins_cost(300);
10498 format %{ "FCMPF $dst,$src1,$src2" %}
10499 opcode(0xD8, 0x3); /* D8 D8+i or D8 /3 */
10500 ins_encode( Push_Reg_DPR(src1),
10501 OpcP, RegOpc(src2),
10502 CmpF_Result(dst));
10503 ins_pipe( pipe_slow );
10504 %}
10506 // float compare and set condition codes in EFLAGS by XMM regs
10507 instruct cmpF_cc(eFlagsRegU cr, regF src1, regF src2) %{
10508 predicate(UseSSE>=1);
10509 match(Set cr (CmpF src1 src2));
10510 ins_cost(145);
10511 format %{ "UCOMISS $src1,$src2\n\t"
10512 "JNP,s exit\n\t"
10513 "PUSHF\t# saw NaN, set CF\n\t"
10514 "AND [rsp], #0xffffff2b\n\t"
10515 "POPF\n"
10516 "exit:" %}
10517 ins_encode %{
10518 __ ucomiss($src1$$XMMRegister, $src2$$XMMRegister);
10519 emit_cmpfp_fixup(_masm);
10520 %}
10521 ins_pipe( pipe_slow );
10522 %}
10524 instruct cmpF_ccCF(eFlagsRegUCF cr, regF src1, regF src2) %{
10525 predicate(UseSSE>=1);
10526 match(Set cr (CmpF src1 src2));
10527 ins_cost(100);
10528 format %{ "UCOMISS $src1,$src2" %}
10529 ins_encode %{
10530 __ ucomiss($src1$$XMMRegister, $src2$$XMMRegister);
10531 %}
10532 ins_pipe( pipe_slow );
10533 %}
10535 // float compare and set condition codes in EFLAGS by XMM regs
10536 instruct cmpF_ccmem(eFlagsRegU cr, regF src1, memory src2) %{
10537 predicate(UseSSE>=1);
10538 match(Set cr (CmpF src1 (LoadF src2)));
10539 ins_cost(165);
10540 format %{ "UCOMISS $src1,$src2\n\t"
10541 "JNP,s exit\n\t"
10542 "PUSHF\t# saw NaN, set CF\n\t"
10543 "AND [rsp], #0xffffff2b\n\t"
10544 "POPF\n"
10545 "exit:" %}
10546 ins_encode %{
10547 __ ucomiss($src1$$XMMRegister, $src2$$Address);
10548 emit_cmpfp_fixup(_masm);
10549 %}
10550 ins_pipe( pipe_slow );
10551 %}
10553 instruct cmpF_ccmemCF(eFlagsRegUCF cr, regF src1, memory src2) %{
10554 predicate(UseSSE>=1);
10555 match(Set cr (CmpF src1 (LoadF src2)));
10556 ins_cost(100);
10557 format %{ "UCOMISS $src1,$src2" %}
10558 ins_encode %{
10559 __ ucomiss($src1$$XMMRegister, $src2$$Address);
10560 %}
10561 ins_pipe( pipe_slow );
10562 %}
10564 // Compare into -1,0,1 in XMM
10565 instruct cmpF_reg(xRegI dst, regF src1, regF src2, eFlagsReg cr) %{
10566 predicate(UseSSE>=1);
10567 match(Set dst (CmpF3 src1 src2));
10568 effect(KILL cr);
10569 ins_cost(255);
10570 format %{ "UCOMISS $src1, $src2\n\t"
10571 "MOV $dst, #-1\n\t"
10572 "JP,s done\n\t"
10573 "JB,s done\n\t"
10574 "SETNE $dst\n\t"
10575 "MOVZB $dst, $dst\n"
10576 "done:" %}
10577 ins_encode %{
10578 __ ucomiss($src1$$XMMRegister, $src2$$XMMRegister);
10579 emit_cmpfp3(_masm, $dst$$Register);
10580 %}
10581 ins_pipe( pipe_slow );
10582 %}
10584 // Compare into -1,0,1 in XMM and memory
10585 instruct cmpF_regmem(xRegI dst, regF src1, memory src2, eFlagsReg cr) %{
10586 predicate(UseSSE>=1);
10587 match(Set dst (CmpF3 src1 (LoadF src2)));
10588 effect(KILL cr);
10589 ins_cost(275);
10590 format %{ "UCOMISS $src1, $src2\n\t"
10591 "MOV $dst, #-1\n\t"
10592 "JP,s done\n\t"
10593 "JB,s done\n\t"
10594 "SETNE $dst\n\t"
10595 "MOVZB $dst, $dst\n"
10596 "done:" %}
10597 ins_encode %{
10598 __ ucomiss($src1$$XMMRegister, $src2$$Address);
10599 emit_cmpfp3(_masm, $dst$$Register);
10600 %}
10601 ins_pipe( pipe_slow );
10602 %}
10604 // Spill to obtain 24-bit precision
10605 instruct subFPR24_reg(stackSlotF dst, regFPR src1, regFPR src2) %{
10606 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10607 match(Set dst (SubF src1 src2));
10609 format %{ "FSUB $dst,$src1 - $src2" %}
10610 opcode(0xD8, 0x4); /* D8 E0+i or D8 /4 mod==0x3 ;; result in TOS */
10611 ins_encode( Push_Reg_FPR(src1),
10612 OpcReg_FPR(src2),
10613 Pop_Mem_FPR(dst) );
10614 ins_pipe( fpu_mem_reg_reg );
10615 %}
10616 //
10617 // This instruction does not round to 24-bits
10618 instruct subFPR_reg(regFPR dst, regFPR src) %{
10619 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10620 match(Set dst (SubF dst src));
10622 format %{ "FSUB $dst,$src" %}
10623 opcode(0xDE, 0x5); /* DE E8+i or DE /5 */
10624 ins_encode( Push_Reg_FPR(src),
10625 OpcP, RegOpc(dst) );
10626 ins_pipe( fpu_reg_reg );
10627 %}
10629 // Spill to obtain 24-bit precision
10630 instruct addFPR24_reg(stackSlotF dst, regFPR src1, regFPR src2) %{
10631 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10632 match(Set dst (AddF src1 src2));
10634 format %{ "FADD $dst,$src1,$src2" %}
10635 opcode(0xD8, 0x0); /* D8 C0+i */
10636 ins_encode( Push_Reg_FPR(src2),
10637 OpcReg_FPR(src1),
10638 Pop_Mem_FPR(dst) );
10639 ins_pipe( fpu_mem_reg_reg );
10640 %}
10641 //
10642 // This instruction does not round to 24-bits
10643 instruct addFPR_reg(regFPR dst, regFPR src) %{
10644 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10645 match(Set dst (AddF dst src));
10647 format %{ "FLD $src\n\t"
10648 "FADDp $dst,ST" %}
10649 opcode(0xDE, 0x0); /* DE C0+i or DE /0*/
10650 ins_encode( Push_Reg_FPR(src),
10651 OpcP, RegOpc(dst) );
10652 ins_pipe( fpu_reg_reg );
10653 %}
10655 instruct absFPR_reg(regFPR1 dst, regFPR1 src) %{
10656 predicate(UseSSE==0);
10657 match(Set dst (AbsF src));
10658 ins_cost(100);
10659 format %{ "FABS" %}
10660 opcode(0xE1, 0xD9);
10661 ins_encode( OpcS, OpcP );
10662 ins_pipe( fpu_reg_reg );
10663 %}
10665 instruct negFPR_reg(regFPR1 dst, regFPR1 src) %{
10666 predicate(UseSSE==0);
10667 match(Set dst (NegF src));
10668 ins_cost(100);
10669 format %{ "FCHS" %}
10670 opcode(0xE0, 0xD9);
10671 ins_encode( OpcS, OpcP );
10672 ins_pipe( fpu_reg_reg );
10673 %}
10675 // Cisc-alternate to addFPR_reg
10676 // Spill to obtain 24-bit precision
10677 instruct addFPR24_reg_mem(stackSlotF dst, regFPR src1, memory src2) %{
10678 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10679 match(Set dst (AddF src1 (LoadF src2)));
10681 format %{ "FLD $src2\n\t"
10682 "FADD ST,$src1\n\t"
10683 "FSTP_S $dst" %}
10684 opcode(0xD8, 0x0, 0xD9); /* D8 C0+i */ /* LoadF D9 /0 */
10685 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10686 OpcReg_FPR(src1),
10687 Pop_Mem_FPR(dst) );
10688 ins_pipe( fpu_mem_reg_mem );
10689 %}
10690 //
10691 // Cisc-alternate to addFPR_reg
10692 // This instruction does not round to 24-bits
10693 instruct addFPR_reg_mem(regFPR dst, memory src) %{
10694 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10695 match(Set dst (AddF dst (LoadF src)));
10697 format %{ "FADD $dst,$src" %}
10698 opcode(0xDE, 0x0, 0xD9); /* DE C0+i or DE /0*/ /* LoadF D9 /0 */
10699 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src),
10700 OpcP, RegOpc(dst) );
10701 ins_pipe( fpu_reg_mem );
10702 %}
10704 // // Following two instructions for _222_mpegaudio
10705 // Spill to obtain 24-bit precision
10706 instruct addFPR24_mem_reg(stackSlotF dst, regFPR src2, memory src1 ) %{
10707 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10708 match(Set dst (AddF src1 src2));
10710 format %{ "FADD $dst,$src1,$src2" %}
10711 opcode(0xD8, 0x0, 0xD9); /* D8 C0+i */ /* LoadF D9 /0 */
10712 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src1),
10713 OpcReg_FPR(src2),
10714 Pop_Mem_FPR(dst) );
10715 ins_pipe( fpu_mem_reg_mem );
10716 %}
10718 // Cisc-spill variant
10719 // Spill to obtain 24-bit precision
10720 instruct addFPR24_mem_cisc(stackSlotF dst, memory src1, memory src2) %{
10721 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10722 match(Set dst (AddF src1 (LoadF src2)));
10724 format %{ "FADD $dst,$src1,$src2 cisc" %}
10725 opcode(0xD8, 0x0, 0xD9); /* D8 C0+i */ /* LoadF D9 /0 */
10726 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10727 set_instruction_start,
10728 OpcP, RMopc_Mem(secondary,src1),
10729 Pop_Mem_FPR(dst) );
10730 ins_pipe( fpu_mem_mem_mem );
10731 %}
10733 // Spill to obtain 24-bit precision
10734 instruct addFPR24_mem_mem(stackSlotF dst, memory src1, memory src2) %{
10735 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10736 match(Set dst (AddF src1 src2));
10738 format %{ "FADD $dst,$src1,$src2" %}
10739 opcode(0xD8, 0x0, 0xD9); /* D8 /0 */ /* LoadF D9 /0 */
10740 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10741 set_instruction_start,
10742 OpcP, RMopc_Mem(secondary,src1),
10743 Pop_Mem_FPR(dst) );
10744 ins_pipe( fpu_mem_mem_mem );
10745 %}
10748 // Spill to obtain 24-bit precision
10749 instruct addFPR24_reg_imm(stackSlotF dst, regFPR src, immFPR con) %{
10750 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10751 match(Set dst (AddF src con));
10752 format %{ "FLD $src\n\t"
10753 "FADD_S [$constantaddress]\t# load from constant table: float=$con\n\t"
10754 "FSTP_S $dst" %}
10755 ins_encode %{
10756 __ fld_s($src$$reg - 1); // FLD ST(i-1)
10757 __ fadd_s($constantaddress($con));
10758 __ fstp_s(Address(rsp, $dst$$disp));
10759 %}
10760 ins_pipe(fpu_mem_reg_con);
10761 %}
10762 //
10763 // This instruction does not round to 24-bits
10764 instruct addFPR_reg_imm(regFPR dst, regFPR src, immFPR con) %{
10765 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10766 match(Set dst (AddF src con));
10767 format %{ "FLD $src\n\t"
10768 "FADD_S [$constantaddress]\t# load from constant table: float=$con\n\t"
10769 "FSTP $dst" %}
10770 ins_encode %{
10771 __ fld_s($src$$reg - 1); // FLD ST(i-1)
10772 __ fadd_s($constantaddress($con));
10773 __ fstp_d($dst$$reg);
10774 %}
10775 ins_pipe(fpu_reg_reg_con);
10776 %}
10778 // Spill to obtain 24-bit precision
10779 instruct mulFPR24_reg(stackSlotF dst, regFPR src1, regFPR src2) %{
10780 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10781 match(Set dst (MulF src1 src2));
10783 format %{ "FLD $src1\n\t"
10784 "FMUL $src2\n\t"
10785 "FSTP_S $dst" %}
10786 opcode(0xD8, 0x1); /* D8 C8+i or D8 /1 ;; result in TOS */
10787 ins_encode( Push_Reg_FPR(src1),
10788 OpcReg_FPR(src2),
10789 Pop_Mem_FPR(dst) );
10790 ins_pipe( fpu_mem_reg_reg );
10791 %}
10792 //
10793 // This instruction does not round to 24-bits
10794 instruct mulFPR_reg(regFPR dst, regFPR src1, regFPR src2) %{
10795 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10796 match(Set dst (MulF src1 src2));
10798 format %{ "FLD $src1\n\t"
10799 "FMUL $src2\n\t"
10800 "FSTP_S $dst" %}
10801 opcode(0xD8, 0x1); /* D8 C8+i */
10802 ins_encode( Push_Reg_FPR(src2),
10803 OpcReg_FPR(src1),
10804 Pop_Reg_FPR(dst) );
10805 ins_pipe( fpu_reg_reg_reg );
10806 %}
10809 // Spill to obtain 24-bit precision
10810 // Cisc-alternate to reg-reg multiply
10811 instruct mulFPR24_reg_mem(stackSlotF dst, regFPR src1, memory src2) %{
10812 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10813 match(Set dst (MulF src1 (LoadF src2)));
10815 format %{ "FLD_S $src2\n\t"
10816 "FMUL $src1\n\t"
10817 "FSTP_S $dst" %}
10818 opcode(0xD8, 0x1, 0xD9); /* D8 C8+i or DE /1*/ /* LoadF D9 /0 */
10819 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10820 OpcReg_FPR(src1),
10821 Pop_Mem_FPR(dst) );
10822 ins_pipe( fpu_mem_reg_mem );
10823 %}
10824 //
10825 // This instruction does not round to 24-bits
10826 // Cisc-alternate to reg-reg multiply
10827 instruct mulFPR_reg_mem(regFPR dst, regFPR src1, memory src2) %{
10828 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10829 match(Set dst (MulF src1 (LoadF src2)));
10831 format %{ "FMUL $dst,$src1,$src2" %}
10832 opcode(0xD8, 0x1, 0xD9); /* D8 C8+i */ /* LoadF D9 /0 */
10833 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10834 OpcReg_FPR(src1),
10835 Pop_Reg_FPR(dst) );
10836 ins_pipe( fpu_reg_reg_mem );
10837 %}
10839 // Spill to obtain 24-bit precision
10840 instruct mulFPR24_mem_mem(stackSlotF dst, memory src1, memory src2) %{
10841 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10842 match(Set dst (MulF src1 src2));
10844 format %{ "FMUL $dst,$src1,$src2" %}
10845 opcode(0xD8, 0x1, 0xD9); /* D8 /1 */ /* LoadF D9 /0 */
10846 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10847 set_instruction_start,
10848 OpcP, RMopc_Mem(secondary,src1),
10849 Pop_Mem_FPR(dst) );
10850 ins_pipe( fpu_mem_mem_mem );
10851 %}
10853 // Spill to obtain 24-bit precision
10854 instruct mulFPR24_reg_imm(stackSlotF dst, regFPR src, immFPR con) %{
10855 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10856 match(Set dst (MulF src con));
10858 format %{ "FLD $src\n\t"
10859 "FMUL_S [$constantaddress]\t# load from constant table: float=$con\n\t"
10860 "FSTP_S $dst" %}
10861 ins_encode %{
10862 __ fld_s($src$$reg - 1); // FLD ST(i-1)
10863 __ fmul_s($constantaddress($con));
10864 __ fstp_s(Address(rsp, $dst$$disp));
10865 %}
10866 ins_pipe(fpu_mem_reg_con);
10867 %}
10868 //
10869 // This instruction does not round to 24-bits
10870 instruct mulFPR_reg_imm(regFPR dst, regFPR src, immFPR con) %{
10871 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10872 match(Set dst (MulF src con));
10874 format %{ "FLD $src\n\t"
10875 "FMUL_S [$constantaddress]\t# load from constant table: float=$con\n\t"
10876 "FSTP $dst" %}
10877 ins_encode %{
10878 __ fld_s($src$$reg - 1); // FLD ST(i-1)
10879 __ fmul_s($constantaddress($con));
10880 __ fstp_d($dst$$reg);
10881 %}
10882 ins_pipe(fpu_reg_reg_con);
10883 %}
10886 //
10887 // MACRO1 -- subsume unshared load into mulFPR
10888 // This instruction does not round to 24-bits
10889 instruct mulFPR_reg_load1(regFPR dst, regFPR src, memory mem1 ) %{
10890 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10891 match(Set dst (MulF (LoadF mem1) src));
10893 format %{ "FLD $mem1 ===MACRO1===\n\t"
10894 "FMUL ST,$src\n\t"
10895 "FSTP $dst" %}
10896 opcode(0xD8, 0x1, 0xD9); /* D8 C8+i or D8 /1 */ /* LoadF D9 /0 */
10897 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,mem1),
10898 OpcReg_FPR(src),
10899 Pop_Reg_FPR(dst) );
10900 ins_pipe( fpu_reg_reg_mem );
10901 %}
10902 //
10903 // MACRO2 -- addFPR a mulFPR which subsumed an unshared load
10904 // This instruction does not round to 24-bits
10905 instruct addFPR_mulFPR_reg_load1(regFPR dst, memory mem1, regFPR src1, regFPR src2) %{
10906 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10907 match(Set dst (AddF (MulF (LoadF mem1) src1) src2));
10908 ins_cost(95);
10910 format %{ "FLD $mem1 ===MACRO2===\n\t"
10911 "FMUL ST,$src1 subsume mulFPR left load\n\t"
10912 "FADD ST,$src2\n\t"
10913 "FSTP $dst" %}
10914 opcode(0xD9); /* LoadF D9 /0 */
10915 ins_encode( OpcP, RMopc_Mem(0x00,mem1),
10916 FMul_ST_reg(src1),
10917 FAdd_ST_reg(src2),
10918 Pop_Reg_FPR(dst) );
10919 ins_pipe( fpu_reg_mem_reg_reg );
10920 %}
10922 // MACRO3 -- addFPR a mulFPR
10923 // This instruction does not round to 24-bits. It is a '2-address'
10924 // instruction in that the result goes back to src2. This eliminates
10925 // a move from the macro; possibly the register allocator will have
10926 // to add it back (and maybe not).
10927 instruct addFPR_mulFPR_reg(regFPR src2, regFPR src1, regFPR src0) %{
10928 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10929 match(Set src2 (AddF (MulF src0 src1) src2));
10931 format %{ "FLD $src0 ===MACRO3===\n\t"
10932 "FMUL ST,$src1\n\t"
10933 "FADDP $src2,ST" %}
10934 opcode(0xD9); /* LoadF D9 /0 */
10935 ins_encode( Push_Reg_FPR(src0),
10936 FMul_ST_reg(src1),
10937 FAddP_reg_ST(src2) );
10938 ins_pipe( fpu_reg_reg_reg );
10939 %}
10941 // MACRO4 -- divFPR subFPR
10942 // This instruction does not round to 24-bits
10943 instruct subFPR_divFPR_reg(regFPR dst, regFPR src1, regFPR src2, regFPR src3) %{
10944 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10945 match(Set dst (DivF (SubF src2 src1) src3));
10947 format %{ "FLD $src2 ===MACRO4===\n\t"
10948 "FSUB ST,$src1\n\t"
10949 "FDIV ST,$src3\n\t"
10950 "FSTP $dst" %}
10951 opcode(0xDE, 0x7); /* DE F8+i or DE /7*/
10952 ins_encode( Push_Reg_FPR(src2),
10953 subFPR_divFPR_encode(src1,src3),
10954 Pop_Reg_FPR(dst) );
10955 ins_pipe( fpu_reg_reg_reg_reg );
10956 %}
10958 // Spill to obtain 24-bit precision
10959 instruct divFPR24_reg(stackSlotF dst, regFPR src1, regFPR src2) %{
10960 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10961 match(Set dst (DivF src1 src2));
10963 format %{ "FDIV $dst,$src1,$src2" %}
10964 opcode(0xD8, 0x6); /* D8 F0+i or DE /6*/
10965 ins_encode( Push_Reg_FPR(src1),
10966 OpcReg_FPR(src2),
10967 Pop_Mem_FPR(dst) );
10968 ins_pipe( fpu_mem_reg_reg );
10969 %}
10970 //
10971 // This instruction does not round to 24-bits
10972 instruct divFPR_reg(regFPR dst, regFPR src) %{
10973 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10974 match(Set dst (DivF dst src));
10976 format %{ "FDIV $dst,$src" %}
10977 opcode(0xDE, 0x7); /* DE F8+i or DE /7*/
10978 ins_encode( Push_Reg_FPR(src),
10979 OpcP, RegOpc(dst) );
10980 ins_pipe( fpu_reg_reg );
10981 %}
10984 // Spill to obtain 24-bit precision
10985 instruct modFPR24_reg(stackSlotF dst, regFPR src1, regFPR src2, eAXRegI rax, eFlagsReg cr) %{
10986 predicate( UseSSE==0 && Compile::current()->select_24_bit_instr());
10987 match(Set dst (ModF src1 src2));
10988 effect(KILL rax, KILL cr); // emitModDPR() uses EAX and EFLAGS
10990 format %{ "FMOD $dst,$src1,$src2" %}
10991 ins_encode( Push_Reg_Mod_DPR(src1, src2),
10992 emitModDPR(),
10993 Push_Result_Mod_DPR(src2),
10994 Pop_Mem_FPR(dst));
10995 ins_pipe( pipe_slow );
10996 %}
10997 //
10998 // This instruction does not round to 24-bits
10999 instruct modFPR_reg(regFPR dst, regFPR src, eAXRegI rax, eFlagsReg cr) %{
11000 predicate( UseSSE==0 && !Compile::current()->select_24_bit_instr());
11001 match(Set dst (ModF dst src));
11002 effect(KILL rax, KILL cr); // emitModDPR() uses EAX and EFLAGS
11004 format %{ "FMOD $dst,$src" %}
11005 ins_encode(Push_Reg_Mod_DPR(dst, src),
11006 emitModDPR(),
11007 Push_Result_Mod_DPR(src),
11008 Pop_Reg_FPR(dst));
11009 ins_pipe( pipe_slow );
11010 %}
11012 instruct modF_reg(regF dst, regF src0, regF src1, eAXRegI rax, eFlagsReg cr) %{
11013 predicate(UseSSE>=1);
11014 match(Set dst (ModF src0 src1));
11015 effect(KILL rax, KILL cr);
11016 format %{ "SUB ESP,4\t # FMOD\n"
11017 "\tMOVSS [ESP+0],$src1\n"
11018 "\tFLD_S [ESP+0]\n"
11019 "\tMOVSS [ESP+0],$src0\n"
11020 "\tFLD_S [ESP+0]\n"
11021 "loop:\tFPREM\n"
11022 "\tFWAIT\n"
11023 "\tFNSTSW AX\n"
11024 "\tSAHF\n"
11025 "\tJP loop\n"
11026 "\tFSTP_S [ESP+0]\n"
11027 "\tMOVSS $dst,[ESP+0]\n"
11028 "\tADD ESP,4\n"
11029 "\tFSTP ST0\t # Restore FPU Stack"
11030 %}
11031 ins_cost(250);
11032 ins_encode( Push_ModF_encoding(src0, src1), emitModDPR(), Push_ResultF(dst,0x4), PopFPU);
11033 ins_pipe( pipe_slow );
11034 %}
11037 //----------Arithmetic Conversion Instructions---------------------------------
11038 // The conversions operations are all Alpha sorted. Please keep it that way!
11040 instruct roundFloat_mem_reg(stackSlotF dst, regFPR src) %{
11041 predicate(UseSSE==0);
11042 match(Set dst (RoundFloat src));
11043 ins_cost(125);
11044 format %{ "FST_S $dst,$src\t# F-round" %}
11045 ins_encode( Pop_Mem_Reg_FPR(dst, src) );
11046 ins_pipe( fpu_mem_reg );
11047 %}
11049 instruct roundDouble_mem_reg(stackSlotD dst, regDPR src) %{
11050 predicate(UseSSE<=1);
11051 match(Set dst (RoundDouble src));
11052 ins_cost(125);
11053 format %{ "FST_D $dst,$src\t# D-round" %}
11054 ins_encode( Pop_Mem_Reg_DPR(dst, src) );
11055 ins_pipe( fpu_mem_reg );
11056 %}
11058 // Force rounding to 24-bit precision and 6-bit exponent
11059 instruct convDPR2FPR_reg(stackSlotF dst, regDPR src) %{
11060 predicate(UseSSE==0);
11061 match(Set dst (ConvD2F src));
11062 format %{ "FST_S $dst,$src\t# F-round" %}
11063 expand %{
11064 roundFloat_mem_reg(dst,src);
11065 %}
11066 %}
11068 // Force rounding to 24-bit precision and 6-bit exponent
11069 instruct convDPR2F_reg(regF dst, regDPR src, eFlagsReg cr) %{
11070 predicate(UseSSE==1);
11071 match(Set dst (ConvD2F src));
11072 effect( KILL cr );
11073 format %{ "SUB ESP,4\n\t"
11074 "FST_S [ESP],$src\t# F-round\n\t"
11075 "MOVSS $dst,[ESP]\n\t"
11076 "ADD ESP,4" %}
11077 ins_encode %{
11078 __ subptr(rsp, 4);
11079 if ($src$$reg != FPR1L_enc) {
11080 __ fld_s($src$$reg-1);
11081 __ fstp_s(Address(rsp, 0));
11082 } else {
11083 __ fst_s(Address(rsp, 0));
11084 }
11085 __ movflt($dst$$XMMRegister, Address(rsp, 0));
11086 __ addptr(rsp, 4);
11087 %}
11088 ins_pipe( pipe_slow );
11089 %}
11091 // Force rounding double precision to single precision
11092 instruct convD2F_reg(regF dst, regD src) %{
11093 predicate(UseSSE>=2);
11094 match(Set dst (ConvD2F src));
11095 format %{ "CVTSD2SS $dst,$src\t# F-round" %}
11096 ins_encode %{
11097 __ cvtsd2ss ($dst$$XMMRegister, $src$$XMMRegister);
11098 %}
11099 ins_pipe( pipe_slow );
11100 %}
11102 instruct convFPR2DPR_reg_reg(regDPR dst, regFPR src) %{
11103 predicate(UseSSE==0);
11104 match(Set dst (ConvF2D src));
11105 format %{ "FST_S $dst,$src\t# D-round" %}
11106 ins_encode( Pop_Reg_Reg_DPR(dst, src));
11107 ins_pipe( fpu_reg_reg );
11108 %}
11110 instruct convFPR2D_reg(stackSlotD dst, regFPR src) %{
11111 predicate(UseSSE==1);
11112 match(Set dst (ConvF2D src));
11113 format %{ "FST_D $dst,$src\t# D-round" %}
11114 expand %{
11115 roundDouble_mem_reg(dst,src);
11116 %}
11117 %}
11119 instruct convF2DPR_reg(regDPR dst, regF src, eFlagsReg cr) %{
11120 predicate(UseSSE==1);
11121 match(Set dst (ConvF2D src));
11122 effect( KILL cr );
11123 format %{ "SUB ESP,4\n\t"
11124 "MOVSS [ESP] $src\n\t"
11125 "FLD_S [ESP]\n\t"
11126 "ADD ESP,4\n\t"
11127 "FSTP $dst\t# D-round" %}
11128 ins_encode %{
11129 __ subptr(rsp, 4);
11130 __ movflt(Address(rsp, 0), $src$$XMMRegister);
11131 __ fld_s(Address(rsp, 0));
11132 __ addptr(rsp, 4);
11133 __ fstp_d($dst$$reg);
11134 %}
11135 ins_pipe( pipe_slow );
11136 %}
11138 instruct convF2D_reg(regD dst, regF src) %{
11139 predicate(UseSSE>=2);
11140 match(Set dst (ConvF2D src));
11141 format %{ "CVTSS2SD $dst,$src\t# D-round" %}
11142 ins_encode %{
11143 __ cvtss2sd ($dst$$XMMRegister, $src$$XMMRegister);
11144 %}
11145 ins_pipe( pipe_slow );
11146 %}
11148 // Convert a double to an int. If the double is a NAN, stuff a zero in instead.
11149 instruct convDPR2I_reg_reg( eAXRegI dst, eDXRegI tmp, regDPR src, eFlagsReg cr ) %{
11150 predicate(UseSSE<=1);
11151 match(Set dst (ConvD2I src));
11152 effect( KILL tmp, KILL cr );
11153 format %{ "FLD $src\t# Convert double to int \n\t"
11154 "FLDCW trunc mode\n\t"
11155 "SUB ESP,4\n\t"
11156 "FISTp [ESP + #0]\n\t"
11157 "FLDCW std/24-bit mode\n\t"
11158 "POP EAX\n\t"
11159 "CMP EAX,0x80000000\n\t"
11160 "JNE,s fast\n\t"
11161 "FLD_D $src\n\t"
11162 "CALL d2i_wrapper\n"
11163 "fast:" %}
11164 ins_encode( Push_Reg_DPR(src), DPR2I_encoding(src) );
11165 ins_pipe( pipe_slow );
11166 %}
11168 // Convert a double to an int. If the double is a NAN, stuff a zero in instead.
11169 instruct convD2I_reg_reg( eAXRegI dst, eDXRegI tmp, regD src, eFlagsReg cr ) %{
11170 predicate(UseSSE>=2);
11171 match(Set dst (ConvD2I src));
11172 effect( KILL tmp, KILL cr );
11173 format %{ "CVTTSD2SI $dst, $src\n\t"
11174 "CMP $dst,0x80000000\n\t"
11175 "JNE,s fast\n\t"
11176 "SUB ESP, 8\n\t"
11177 "MOVSD [ESP], $src\n\t"
11178 "FLD_D [ESP]\n\t"
11179 "ADD ESP, 8\n\t"
11180 "CALL d2i_wrapper\n"
11181 "fast:" %}
11182 ins_encode %{
11183 Label fast;
11184 __ cvttsd2sil($dst$$Register, $src$$XMMRegister);
11185 __ cmpl($dst$$Register, 0x80000000);
11186 __ jccb(Assembler::notEqual, fast);
11187 __ subptr(rsp, 8);
11188 __ movdbl(Address(rsp, 0), $src$$XMMRegister);
11189 __ fld_d(Address(rsp, 0));
11190 __ addptr(rsp, 8);
11191 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, StubRoutines::d2i_wrapper())));
11192 __ bind(fast);
11193 %}
11194 ins_pipe( pipe_slow );
11195 %}
11197 instruct convDPR2L_reg_reg( eADXRegL dst, regDPR src, eFlagsReg cr ) %{
11198 predicate(UseSSE<=1);
11199 match(Set dst (ConvD2L src));
11200 effect( KILL cr );
11201 format %{ "FLD $src\t# Convert double to long\n\t"
11202 "FLDCW trunc mode\n\t"
11203 "SUB ESP,8\n\t"
11204 "FISTp [ESP + #0]\n\t"
11205 "FLDCW std/24-bit mode\n\t"
11206 "POP EAX\n\t"
11207 "POP EDX\n\t"
11208 "CMP EDX,0x80000000\n\t"
11209 "JNE,s fast\n\t"
11210 "TEST EAX,EAX\n\t"
11211 "JNE,s fast\n\t"
11212 "FLD $src\n\t"
11213 "CALL d2l_wrapper\n"
11214 "fast:" %}
11215 ins_encode( Push_Reg_DPR(src), DPR2L_encoding(src) );
11216 ins_pipe( pipe_slow );
11217 %}
11219 // XMM lacks a float/double->long conversion, so use the old FPU stack.
11220 instruct convD2L_reg_reg( eADXRegL dst, regD src, eFlagsReg cr ) %{
11221 predicate (UseSSE>=2);
11222 match(Set dst (ConvD2L src));
11223 effect( KILL cr );
11224 format %{ "SUB ESP,8\t# Convert double to long\n\t"
11225 "MOVSD [ESP],$src\n\t"
11226 "FLD_D [ESP]\n\t"
11227 "FLDCW trunc mode\n\t"
11228 "FISTp [ESP + #0]\n\t"
11229 "FLDCW std/24-bit mode\n\t"
11230 "POP EAX\n\t"
11231 "POP EDX\n\t"
11232 "CMP EDX,0x80000000\n\t"
11233 "JNE,s fast\n\t"
11234 "TEST EAX,EAX\n\t"
11235 "JNE,s fast\n\t"
11236 "SUB ESP,8\n\t"
11237 "MOVSD [ESP],$src\n\t"
11238 "FLD_D [ESP]\n\t"
11239 "ADD ESP,8\n\t"
11240 "CALL d2l_wrapper\n"
11241 "fast:" %}
11242 ins_encode %{
11243 Label fast;
11244 __ subptr(rsp, 8);
11245 __ movdbl(Address(rsp, 0), $src$$XMMRegister);
11246 __ fld_d(Address(rsp, 0));
11247 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_trunc()));
11248 __ fistp_d(Address(rsp, 0));
11249 // Restore the rounding mode, mask the exception
11250 if (Compile::current()->in_24_bit_fp_mode()) {
11251 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
11252 } else {
11253 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
11254 }
11255 // Load the converted long, adjust CPU stack
11256 __ pop(rax);
11257 __ pop(rdx);
11258 __ cmpl(rdx, 0x80000000);
11259 __ jccb(Assembler::notEqual, fast);
11260 __ testl(rax, rax);
11261 __ jccb(Assembler::notEqual, fast);
11262 __ subptr(rsp, 8);
11263 __ movdbl(Address(rsp, 0), $src$$XMMRegister);
11264 __ fld_d(Address(rsp, 0));
11265 __ addptr(rsp, 8);
11266 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, StubRoutines::d2l_wrapper())));
11267 __ bind(fast);
11268 %}
11269 ins_pipe( pipe_slow );
11270 %}
11272 // Convert a double to an int. Java semantics require we do complex
11273 // manglations in the corner cases. So we set the rounding mode to
11274 // 'zero', store the darned double down as an int, and reset the
11275 // rounding mode to 'nearest'. The hardware stores a flag value down
11276 // if we would overflow or converted a NAN; we check for this and
11277 // and go the slow path if needed.
11278 instruct convFPR2I_reg_reg(eAXRegI dst, eDXRegI tmp, regFPR src, eFlagsReg cr ) %{
11279 predicate(UseSSE==0);
11280 match(Set dst (ConvF2I src));
11281 effect( KILL tmp, KILL cr );
11282 format %{ "FLD $src\t# Convert float to int \n\t"
11283 "FLDCW trunc mode\n\t"
11284 "SUB ESP,4\n\t"
11285 "FISTp [ESP + #0]\n\t"
11286 "FLDCW std/24-bit mode\n\t"
11287 "POP EAX\n\t"
11288 "CMP EAX,0x80000000\n\t"
11289 "JNE,s fast\n\t"
11290 "FLD $src\n\t"
11291 "CALL d2i_wrapper\n"
11292 "fast:" %}
11293 // DPR2I_encoding works for FPR2I
11294 ins_encode( Push_Reg_FPR(src), DPR2I_encoding(src) );
11295 ins_pipe( pipe_slow );
11296 %}
11298 // Convert a float in xmm to an int reg.
11299 instruct convF2I_reg(eAXRegI dst, eDXRegI tmp, regF src, eFlagsReg cr ) %{
11300 predicate(UseSSE>=1);
11301 match(Set dst (ConvF2I src));
11302 effect( KILL tmp, KILL cr );
11303 format %{ "CVTTSS2SI $dst, $src\n\t"
11304 "CMP $dst,0x80000000\n\t"
11305 "JNE,s fast\n\t"
11306 "SUB ESP, 4\n\t"
11307 "MOVSS [ESP], $src\n\t"
11308 "FLD [ESP]\n\t"
11309 "ADD ESP, 4\n\t"
11310 "CALL d2i_wrapper\n"
11311 "fast:" %}
11312 ins_encode %{
11313 Label fast;
11314 __ cvttss2sil($dst$$Register, $src$$XMMRegister);
11315 __ cmpl($dst$$Register, 0x80000000);
11316 __ jccb(Assembler::notEqual, fast);
11317 __ subptr(rsp, 4);
11318 __ movflt(Address(rsp, 0), $src$$XMMRegister);
11319 __ fld_s(Address(rsp, 0));
11320 __ addptr(rsp, 4);
11321 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, StubRoutines::d2i_wrapper())));
11322 __ bind(fast);
11323 %}
11324 ins_pipe( pipe_slow );
11325 %}
11327 instruct convFPR2L_reg_reg( eADXRegL dst, regFPR src, eFlagsReg cr ) %{
11328 predicate(UseSSE==0);
11329 match(Set dst (ConvF2L src));
11330 effect( KILL cr );
11331 format %{ "FLD $src\t# Convert float to long\n\t"
11332 "FLDCW trunc mode\n\t"
11333 "SUB ESP,8\n\t"
11334 "FISTp [ESP + #0]\n\t"
11335 "FLDCW std/24-bit mode\n\t"
11336 "POP EAX\n\t"
11337 "POP EDX\n\t"
11338 "CMP EDX,0x80000000\n\t"
11339 "JNE,s fast\n\t"
11340 "TEST EAX,EAX\n\t"
11341 "JNE,s fast\n\t"
11342 "FLD $src\n\t"
11343 "CALL d2l_wrapper\n"
11344 "fast:" %}
11345 // DPR2L_encoding works for FPR2L
11346 ins_encode( Push_Reg_FPR(src), DPR2L_encoding(src) );
11347 ins_pipe( pipe_slow );
11348 %}
11350 // XMM lacks a float/double->long conversion, so use the old FPU stack.
11351 instruct convF2L_reg_reg( eADXRegL dst, regF src, eFlagsReg cr ) %{
11352 predicate (UseSSE>=1);
11353 match(Set dst (ConvF2L src));
11354 effect( KILL cr );
11355 format %{ "SUB ESP,8\t# Convert float to long\n\t"
11356 "MOVSS [ESP],$src\n\t"
11357 "FLD_S [ESP]\n\t"
11358 "FLDCW trunc mode\n\t"
11359 "FISTp [ESP + #0]\n\t"
11360 "FLDCW std/24-bit mode\n\t"
11361 "POP EAX\n\t"
11362 "POP EDX\n\t"
11363 "CMP EDX,0x80000000\n\t"
11364 "JNE,s fast\n\t"
11365 "TEST EAX,EAX\n\t"
11366 "JNE,s fast\n\t"
11367 "SUB ESP,4\t# Convert float to long\n\t"
11368 "MOVSS [ESP],$src\n\t"
11369 "FLD_S [ESP]\n\t"
11370 "ADD ESP,4\n\t"
11371 "CALL d2l_wrapper\n"
11372 "fast:" %}
11373 ins_encode %{
11374 Label fast;
11375 __ subptr(rsp, 8);
11376 __ movflt(Address(rsp, 0), $src$$XMMRegister);
11377 __ fld_s(Address(rsp, 0));
11378 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_trunc()));
11379 __ fistp_d(Address(rsp, 0));
11380 // Restore the rounding mode, mask the exception
11381 if (Compile::current()->in_24_bit_fp_mode()) {
11382 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
11383 } else {
11384 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
11385 }
11386 // Load the converted long, adjust CPU stack
11387 __ pop(rax);
11388 __ pop(rdx);
11389 __ cmpl(rdx, 0x80000000);
11390 __ jccb(Assembler::notEqual, fast);
11391 __ testl(rax, rax);
11392 __ jccb(Assembler::notEqual, fast);
11393 __ subptr(rsp, 4);
11394 __ movflt(Address(rsp, 0), $src$$XMMRegister);
11395 __ fld_s(Address(rsp, 0));
11396 __ addptr(rsp, 4);
11397 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, StubRoutines::d2l_wrapper())));
11398 __ bind(fast);
11399 %}
11400 ins_pipe( pipe_slow );
11401 %}
11403 instruct convI2DPR_reg(regDPR dst, stackSlotI src) %{
11404 predicate( UseSSE<=1 );
11405 match(Set dst (ConvI2D src));
11406 format %{ "FILD $src\n\t"
11407 "FSTP $dst" %}
11408 opcode(0xDB, 0x0); /* DB /0 */
11409 ins_encode(Push_Mem_I(src), Pop_Reg_DPR(dst));
11410 ins_pipe( fpu_reg_mem );
11411 %}
11413 instruct convI2D_reg(regD dst, eRegI src) %{
11414 predicate( UseSSE>=2 && !UseXmmI2D );
11415 match(Set dst (ConvI2D src));
11416 format %{ "CVTSI2SD $dst,$src" %}
11417 ins_encode %{
11418 __ cvtsi2sdl ($dst$$XMMRegister, $src$$Register);
11419 %}
11420 ins_pipe( pipe_slow );
11421 %}
11423 instruct convI2D_mem(regD dst, memory mem) %{
11424 predicate( UseSSE>=2 );
11425 match(Set dst (ConvI2D (LoadI mem)));
11426 format %{ "CVTSI2SD $dst,$mem" %}
11427 ins_encode %{
11428 __ cvtsi2sdl ($dst$$XMMRegister, $mem$$Address);
11429 %}
11430 ins_pipe( pipe_slow );
11431 %}
11433 instruct convXI2D_reg(regD dst, eRegI src)
11434 %{
11435 predicate( UseSSE>=2 && UseXmmI2D );
11436 match(Set dst (ConvI2D src));
11438 format %{ "MOVD $dst,$src\n\t"
11439 "CVTDQ2PD $dst,$dst\t# i2d" %}
11440 ins_encode %{
11441 __ movdl($dst$$XMMRegister, $src$$Register);
11442 __ cvtdq2pd($dst$$XMMRegister, $dst$$XMMRegister);
11443 %}
11444 ins_pipe(pipe_slow); // XXX
11445 %}
11447 instruct convI2DPR_mem(regDPR dst, memory mem) %{
11448 predicate( UseSSE<=1 && !Compile::current()->select_24_bit_instr());
11449 match(Set dst (ConvI2D (LoadI mem)));
11450 format %{ "FILD $mem\n\t"
11451 "FSTP $dst" %}
11452 opcode(0xDB); /* DB /0 */
11453 ins_encode( OpcP, RMopc_Mem(0x00,mem),
11454 Pop_Reg_DPR(dst));
11455 ins_pipe( fpu_reg_mem );
11456 %}
11458 // Convert a byte to a float; no rounding step needed.
11459 instruct conv24I2FPR_reg(regFPR dst, stackSlotI src) %{
11460 predicate( UseSSE==0 && n->in(1)->Opcode() == Op_AndI && n->in(1)->in(2)->is_Con() && n->in(1)->in(2)->get_int() == 255 );
11461 match(Set dst (ConvI2F src));
11462 format %{ "FILD $src\n\t"
11463 "FSTP $dst" %}
11465 opcode(0xDB, 0x0); /* DB /0 */
11466 ins_encode(Push_Mem_I(src), Pop_Reg_FPR(dst));
11467 ins_pipe( fpu_reg_mem );
11468 %}
11470 // In 24-bit mode, force exponent rounding by storing back out
11471 instruct convI2FPR_SSF(stackSlotF dst, stackSlotI src) %{
11472 predicate( UseSSE==0 && Compile::current()->select_24_bit_instr());
11473 match(Set dst (ConvI2F src));
11474 ins_cost(200);
11475 format %{ "FILD $src\n\t"
11476 "FSTP_S $dst" %}
11477 opcode(0xDB, 0x0); /* DB /0 */
11478 ins_encode( Push_Mem_I(src),
11479 Pop_Mem_FPR(dst));
11480 ins_pipe( fpu_mem_mem );
11481 %}
11483 // In 24-bit mode, force exponent rounding by storing back out
11484 instruct convI2FPR_SSF_mem(stackSlotF dst, memory mem) %{
11485 predicate( UseSSE==0 && Compile::current()->select_24_bit_instr());
11486 match(Set dst (ConvI2F (LoadI mem)));
11487 ins_cost(200);
11488 format %{ "FILD $mem\n\t"
11489 "FSTP_S $dst" %}
11490 opcode(0xDB); /* DB /0 */
11491 ins_encode( OpcP, RMopc_Mem(0x00,mem),
11492 Pop_Mem_FPR(dst));
11493 ins_pipe( fpu_mem_mem );
11494 %}
11496 // This instruction does not round to 24-bits
11497 instruct convI2FPR_reg(regFPR dst, stackSlotI src) %{
11498 predicate( UseSSE==0 && !Compile::current()->select_24_bit_instr());
11499 match(Set dst (ConvI2F src));
11500 format %{ "FILD $src\n\t"
11501 "FSTP $dst" %}
11502 opcode(0xDB, 0x0); /* DB /0 */
11503 ins_encode( Push_Mem_I(src),
11504 Pop_Reg_FPR(dst));
11505 ins_pipe( fpu_reg_mem );
11506 %}
11508 // This instruction does not round to 24-bits
11509 instruct convI2FPR_mem(regFPR dst, memory mem) %{
11510 predicate( UseSSE==0 && !Compile::current()->select_24_bit_instr());
11511 match(Set dst (ConvI2F (LoadI mem)));
11512 format %{ "FILD $mem\n\t"
11513 "FSTP $dst" %}
11514 opcode(0xDB); /* DB /0 */
11515 ins_encode( OpcP, RMopc_Mem(0x00,mem),
11516 Pop_Reg_FPR(dst));
11517 ins_pipe( fpu_reg_mem );
11518 %}
11520 // Convert an int to a float in xmm; no rounding step needed.
11521 instruct convI2F_reg(regF dst, eRegI src) %{
11522 predicate( UseSSE==1 || UseSSE>=2 && !UseXmmI2F );
11523 match(Set dst (ConvI2F src));
11524 format %{ "CVTSI2SS $dst, $src" %}
11525 ins_encode %{
11526 __ cvtsi2ssl ($dst$$XMMRegister, $src$$Register);
11527 %}
11528 ins_pipe( pipe_slow );
11529 %}
11531 instruct convXI2F_reg(regF dst, eRegI src)
11532 %{
11533 predicate( UseSSE>=2 && UseXmmI2F );
11534 match(Set dst (ConvI2F src));
11536 format %{ "MOVD $dst,$src\n\t"
11537 "CVTDQ2PS $dst,$dst\t# i2f" %}
11538 ins_encode %{
11539 __ movdl($dst$$XMMRegister, $src$$Register);
11540 __ cvtdq2ps($dst$$XMMRegister, $dst$$XMMRegister);
11541 %}
11542 ins_pipe(pipe_slow); // XXX
11543 %}
11545 instruct convI2L_reg( eRegL dst, eRegI src, eFlagsReg cr) %{
11546 match(Set dst (ConvI2L src));
11547 effect(KILL cr);
11548 ins_cost(375);
11549 format %{ "MOV $dst.lo,$src\n\t"
11550 "MOV $dst.hi,$src\n\t"
11551 "SAR $dst.hi,31" %}
11552 ins_encode(convert_int_long(dst,src));
11553 ins_pipe( ialu_reg_reg_long );
11554 %}
11556 // Zero-extend convert int to long
11557 instruct convI2L_reg_zex(eRegL dst, eRegI src, immL_32bits mask, eFlagsReg flags ) %{
11558 match(Set dst (AndL (ConvI2L src) mask) );
11559 effect( KILL flags );
11560 ins_cost(250);
11561 format %{ "MOV $dst.lo,$src\n\t"
11562 "XOR $dst.hi,$dst.hi" %}
11563 opcode(0x33); // XOR
11564 ins_encode(enc_Copy(dst,src), OpcP, RegReg_Hi2(dst,dst) );
11565 ins_pipe( ialu_reg_reg_long );
11566 %}
11568 // Zero-extend long
11569 instruct zerox_long(eRegL dst, eRegL src, immL_32bits mask, eFlagsReg flags ) %{
11570 match(Set dst (AndL src mask) );
11571 effect( KILL flags );
11572 ins_cost(250);
11573 format %{ "MOV $dst.lo,$src.lo\n\t"
11574 "XOR $dst.hi,$dst.hi\n\t" %}
11575 opcode(0x33); // XOR
11576 ins_encode(enc_Copy(dst,src), OpcP, RegReg_Hi2(dst,dst) );
11577 ins_pipe( ialu_reg_reg_long );
11578 %}
11580 instruct convL2DPR_reg( stackSlotD dst, eRegL src, eFlagsReg cr) %{
11581 predicate (UseSSE<=1);
11582 match(Set dst (ConvL2D src));
11583 effect( KILL cr );
11584 format %{ "PUSH $src.hi\t# Convert long to double\n\t"
11585 "PUSH $src.lo\n\t"
11586 "FILD ST,[ESP + #0]\n\t"
11587 "ADD ESP,8\n\t"
11588 "FSTP_D $dst\t# D-round" %}
11589 opcode(0xDF, 0x5); /* DF /5 */
11590 ins_encode(convert_long_double(src), Pop_Mem_DPR(dst));
11591 ins_pipe( pipe_slow );
11592 %}
11594 instruct convL2D_reg( regD dst, eRegL src, eFlagsReg cr) %{
11595 predicate (UseSSE>=2);
11596 match(Set dst (ConvL2D src));
11597 effect( KILL cr );
11598 format %{ "PUSH $src.hi\t# Convert long to double\n\t"
11599 "PUSH $src.lo\n\t"
11600 "FILD_D [ESP]\n\t"
11601 "FSTP_D [ESP]\n\t"
11602 "MOVSD $dst,[ESP]\n\t"
11603 "ADD ESP,8" %}
11604 opcode(0xDF, 0x5); /* DF /5 */
11605 ins_encode(convert_long_double2(src), Push_ResultD(dst));
11606 ins_pipe( pipe_slow );
11607 %}
11609 instruct convL2F_reg( regF dst, eRegL src, eFlagsReg cr) %{
11610 predicate (UseSSE>=1);
11611 match(Set dst (ConvL2F src));
11612 effect( KILL cr );
11613 format %{ "PUSH $src.hi\t# Convert long to single float\n\t"
11614 "PUSH $src.lo\n\t"
11615 "FILD_D [ESP]\n\t"
11616 "FSTP_S [ESP]\n\t"
11617 "MOVSS $dst,[ESP]\n\t"
11618 "ADD ESP,8" %}
11619 opcode(0xDF, 0x5); /* DF /5 */
11620 ins_encode(convert_long_double2(src), Push_ResultF(dst,0x8));
11621 ins_pipe( pipe_slow );
11622 %}
11624 instruct convL2FPR_reg( stackSlotF dst, eRegL src, eFlagsReg cr) %{
11625 match(Set dst (ConvL2F src));
11626 effect( KILL cr );
11627 format %{ "PUSH $src.hi\t# Convert long to single float\n\t"
11628 "PUSH $src.lo\n\t"
11629 "FILD ST,[ESP + #0]\n\t"
11630 "ADD ESP,8\n\t"
11631 "FSTP_S $dst\t# F-round" %}
11632 opcode(0xDF, 0x5); /* DF /5 */
11633 ins_encode(convert_long_double(src), Pop_Mem_FPR(dst));
11634 ins_pipe( pipe_slow );
11635 %}
11637 instruct convL2I_reg( eRegI dst, eRegL src ) %{
11638 match(Set dst (ConvL2I src));
11639 effect( DEF dst, USE src );
11640 format %{ "MOV $dst,$src.lo" %}
11641 ins_encode(enc_CopyL_Lo(dst,src));
11642 ins_pipe( ialu_reg_reg );
11643 %}
11646 instruct MoveF2I_stack_reg(eRegI dst, stackSlotF src) %{
11647 match(Set dst (MoveF2I src));
11648 effect( DEF dst, USE src );
11649 ins_cost(100);
11650 format %{ "MOV $dst,$src\t# MoveF2I_stack_reg" %}
11651 ins_encode %{
11652 __ movl($dst$$Register, Address(rsp, $src$$disp));
11653 %}
11654 ins_pipe( ialu_reg_mem );
11655 %}
11657 instruct MoveFPR2I_reg_stack(stackSlotI dst, regFPR src) %{
11658 predicate(UseSSE==0);
11659 match(Set dst (MoveF2I src));
11660 effect( DEF dst, USE src );
11662 ins_cost(125);
11663 format %{ "FST_S $dst,$src\t# MoveF2I_reg_stack" %}
11664 ins_encode( Pop_Mem_Reg_FPR(dst, src) );
11665 ins_pipe( fpu_mem_reg );
11666 %}
11668 instruct MoveF2I_reg_stack_sse(stackSlotI dst, regF src) %{
11669 predicate(UseSSE>=1);
11670 match(Set dst (MoveF2I src));
11671 effect( DEF dst, USE src );
11673 ins_cost(95);
11674 format %{ "MOVSS $dst,$src\t# MoveF2I_reg_stack_sse" %}
11675 ins_encode %{
11676 __ movflt(Address(rsp, $dst$$disp), $src$$XMMRegister);
11677 %}
11678 ins_pipe( pipe_slow );
11679 %}
11681 instruct MoveF2I_reg_reg_sse(eRegI dst, regF src) %{
11682 predicate(UseSSE>=2);
11683 match(Set dst (MoveF2I src));
11684 effect( DEF dst, USE src );
11685 ins_cost(85);
11686 format %{ "MOVD $dst,$src\t# MoveF2I_reg_reg_sse" %}
11687 ins_encode %{
11688 __ movdl($dst$$Register, $src$$XMMRegister);
11689 %}
11690 ins_pipe( pipe_slow );
11691 %}
11693 instruct MoveI2F_reg_stack(stackSlotF dst, eRegI src) %{
11694 match(Set dst (MoveI2F src));
11695 effect( DEF dst, USE src );
11697 ins_cost(100);
11698 format %{ "MOV $dst,$src\t# MoveI2F_reg_stack" %}
11699 ins_encode %{
11700 __ movl(Address(rsp, $dst$$disp), $src$$Register);
11701 %}
11702 ins_pipe( ialu_mem_reg );
11703 %}
11706 instruct MoveI2FPR_stack_reg(regFPR dst, stackSlotI src) %{
11707 predicate(UseSSE==0);
11708 match(Set dst (MoveI2F src));
11709 effect(DEF dst, USE src);
11711 ins_cost(125);
11712 format %{ "FLD_S $src\n\t"
11713 "FSTP $dst\t# MoveI2F_stack_reg" %}
11714 opcode(0xD9); /* D9 /0, FLD m32real */
11715 ins_encode( OpcP, RMopc_Mem_no_oop(0x00,src),
11716 Pop_Reg_FPR(dst) );
11717 ins_pipe( fpu_reg_mem );
11718 %}
11720 instruct MoveI2F_stack_reg_sse(regF dst, stackSlotI src) %{
11721 predicate(UseSSE>=1);
11722 match(Set dst (MoveI2F src));
11723 effect( DEF dst, USE src );
11725 ins_cost(95);
11726 format %{ "MOVSS $dst,$src\t# MoveI2F_stack_reg_sse" %}
11727 ins_encode %{
11728 __ movflt($dst$$XMMRegister, Address(rsp, $src$$disp));
11729 %}
11730 ins_pipe( pipe_slow );
11731 %}
11733 instruct MoveI2F_reg_reg_sse(regF dst, eRegI src) %{
11734 predicate(UseSSE>=2);
11735 match(Set dst (MoveI2F src));
11736 effect( DEF dst, USE src );
11738 ins_cost(85);
11739 format %{ "MOVD $dst,$src\t# MoveI2F_reg_reg_sse" %}
11740 ins_encode %{
11741 __ movdl($dst$$XMMRegister, $src$$Register);
11742 %}
11743 ins_pipe( pipe_slow );
11744 %}
11746 instruct MoveD2L_stack_reg(eRegL dst, stackSlotD src) %{
11747 match(Set dst (MoveD2L src));
11748 effect(DEF dst, USE src);
11750 ins_cost(250);
11751 format %{ "MOV $dst.lo,$src\n\t"
11752 "MOV $dst.hi,$src+4\t# MoveD2L_stack_reg" %}
11753 opcode(0x8B, 0x8B);
11754 ins_encode( OpcP, RegMem(dst,src), OpcS, RegMem_Hi(dst,src));
11755 ins_pipe( ialu_mem_long_reg );
11756 %}
11758 instruct MoveDPR2L_reg_stack(stackSlotL dst, regDPR src) %{
11759 predicate(UseSSE<=1);
11760 match(Set dst (MoveD2L src));
11761 effect(DEF dst, USE src);
11763 ins_cost(125);
11764 format %{ "FST_D $dst,$src\t# MoveD2L_reg_stack" %}
11765 ins_encode( Pop_Mem_Reg_DPR(dst, src) );
11766 ins_pipe( fpu_mem_reg );
11767 %}
11769 instruct MoveD2L_reg_stack_sse(stackSlotL dst, regD src) %{
11770 predicate(UseSSE>=2);
11771 match(Set dst (MoveD2L src));
11772 effect(DEF dst, USE src);
11773 ins_cost(95);
11774 format %{ "MOVSD $dst,$src\t# MoveD2L_reg_stack_sse" %}
11775 ins_encode %{
11776 __ movdbl(Address(rsp, $dst$$disp), $src$$XMMRegister);
11777 %}
11778 ins_pipe( pipe_slow );
11779 %}
11781 instruct MoveD2L_reg_reg_sse(eRegL dst, regD src, regD tmp) %{
11782 predicate(UseSSE>=2);
11783 match(Set dst (MoveD2L src));
11784 effect(DEF dst, USE src, TEMP tmp);
11785 ins_cost(85);
11786 format %{ "MOVD $dst.lo,$src\n\t"
11787 "PSHUFLW $tmp,$src,0x4E\n\t"
11788 "MOVD $dst.hi,$tmp\t# MoveD2L_reg_reg_sse" %}
11789 ins_encode %{
11790 __ movdl($dst$$Register, $src$$XMMRegister);
11791 __ pshuflw($tmp$$XMMRegister, $src$$XMMRegister, 0x4e);
11792 __ movdl(HIGH_FROM_LOW($dst$$Register), $tmp$$XMMRegister);
11793 %}
11794 ins_pipe( pipe_slow );
11795 %}
11797 instruct MoveL2D_reg_stack(stackSlotD dst, eRegL src) %{
11798 match(Set dst (MoveL2D src));
11799 effect(DEF dst, USE src);
11801 ins_cost(200);
11802 format %{ "MOV $dst,$src.lo\n\t"
11803 "MOV $dst+4,$src.hi\t# MoveL2D_reg_stack" %}
11804 opcode(0x89, 0x89);
11805 ins_encode( OpcP, RegMem( src, dst ), OpcS, RegMem_Hi( src, dst ) );
11806 ins_pipe( ialu_mem_long_reg );
11807 %}
11810 instruct MoveL2DPR_stack_reg(regDPR dst, stackSlotL src) %{
11811 predicate(UseSSE<=1);
11812 match(Set dst (MoveL2D src));
11813 effect(DEF dst, USE src);
11814 ins_cost(125);
11816 format %{ "FLD_D $src\n\t"
11817 "FSTP $dst\t# MoveL2D_stack_reg" %}
11818 opcode(0xDD); /* DD /0, FLD m64real */
11819 ins_encode( OpcP, RMopc_Mem_no_oop(0x00,src),
11820 Pop_Reg_DPR(dst) );
11821 ins_pipe( fpu_reg_mem );
11822 %}
11825 instruct MoveL2D_stack_reg_sse(regD dst, stackSlotL src) %{
11826 predicate(UseSSE>=2 && UseXmmLoadAndClearUpper);
11827 match(Set dst (MoveL2D src));
11828 effect(DEF dst, USE src);
11830 ins_cost(95);
11831 format %{ "MOVSD $dst,$src\t# MoveL2D_stack_reg_sse" %}
11832 ins_encode %{
11833 __ movdbl($dst$$XMMRegister, Address(rsp, $src$$disp));
11834 %}
11835 ins_pipe( pipe_slow );
11836 %}
11838 instruct MoveL2D_stack_reg_sse_partial(regD dst, stackSlotL src) %{
11839 predicate(UseSSE>=2 && !UseXmmLoadAndClearUpper);
11840 match(Set dst (MoveL2D src));
11841 effect(DEF dst, USE src);
11843 ins_cost(95);
11844 format %{ "MOVLPD $dst,$src\t# MoveL2D_stack_reg_sse" %}
11845 ins_encode %{
11846 __ movdbl($dst$$XMMRegister, Address(rsp, $src$$disp));
11847 %}
11848 ins_pipe( pipe_slow );
11849 %}
11851 instruct MoveL2D_reg_reg_sse(regD dst, eRegL src, regD tmp) %{
11852 predicate(UseSSE>=2);
11853 match(Set dst (MoveL2D src));
11854 effect(TEMP dst, USE src, TEMP tmp);
11855 ins_cost(85);
11856 format %{ "MOVD $dst,$src.lo\n\t"
11857 "MOVD $tmp,$src.hi\n\t"
11858 "PUNPCKLDQ $dst,$tmp\t# MoveL2D_reg_reg_sse" %}
11859 ins_encode %{
11860 __ movdl($dst$$XMMRegister, $src$$Register);
11861 __ movdl($tmp$$XMMRegister, HIGH_FROM_LOW($src$$Register));
11862 __ punpckldq($dst$$XMMRegister, $tmp$$XMMRegister);
11863 %}
11864 ins_pipe( pipe_slow );
11865 %}
11867 // Replicate scalar to packed byte (1 byte) values in xmm
11868 instruct Repl8B_reg(regD dst, regD src) %{
11869 predicate(UseSSE>=2);
11870 match(Set dst (Replicate8B src));
11871 format %{ "MOVDQA $dst,$src\n\t"
11872 "PUNPCKLBW $dst,$dst\n\t"
11873 "PSHUFLW $dst,$dst,0x00\t! replicate8B" %}
11874 ins_encode %{
11875 if ($dst$$reg != $src$$reg) {
11876 __ movdqa($dst$$XMMRegister, $src$$XMMRegister);
11877 }
11878 __ punpcklbw($dst$$XMMRegister, $dst$$XMMRegister);
11879 __ pshuflw($dst$$XMMRegister, $dst$$XMMRegister, 0x00);
11880 %}
11881 ins_pipe( pipe_slow );
11882 %}
11884 // Replicate scalar to packed byte (1 byte) values in xmm
11885 instruct Repl8B_eRegI(regD dst, eRegI src) %{
11886 predicate(UseSSE>=2);
11887 match(Set dst (Replicate8B src));
11888 format %{ "MOVD $dst,$src\n\t"
11889 "PUNPCKLBW $dst,$dst\n\t"
11890 "PSHUFLW $dst,$dst,0x00\t! replicate8B" %}
11891 ins_encode %{
11892 __ movdl($dst$$XMMRegister, $src$$Register);
11893 __ punpcklbw($dst$$XMMRegister, $dst$$XMMRegister);
11894 __ pshuflw($dst$$XMMRegister, $dst$$XMMRegister, 0x00);
11895 %}
11896 ins_pipe( pipe_slow );
11897 %}
11899 // Replicate scalar zero to packed byte (1 byte) values in xmm
11900 instruct Repl8B_immI0(regD dst, immI0 zero) %{
11901 predicate(UseSSE>=2);
11902 match(Set dst (Replicate8B zero));
11903 format %{ "PXOR $dst,$dst\t! replicate8B" %}
11904 ins_encode %{
11905 __ pxor($dst$$XMMRegister, $dst$$XMMRegister);
11906 %}
11907 ins_pipe( fpu_reg_reg );
11908 %}
11910 // Replicate scalar to packed shore (2 byte) values in xmm
11911 instruct Repl4S_reg(regD dst, regD src) %{
11912 predicate(UseSSE>=2);
11913 match(Set dst (Replicate4S src));
11914 format %{ "PSHUFLW $dst,$src,0x00\t! replicate4S" %}
11915 ins_encode %{
11916 __ pshuflw($dst$$XMMRegister, $src$$XMMRegister, 0x00);
11917 %}
11918 ins_pipe( fpu_reg_reg );
11919 %}
11921 // Replicate scalar to packed shore (2 byte) values in xmm
11922 instruct Repl4S_eRegI(regD dst, eRegI src) %{
11923 predicate(UseSSE>=2);
11924 match(Set dst (Replicate4S src));
11925 format %{ "MOVD $dst,$src\n\t"
11926 "PSHUFLW $dst,$dst,0x00\t! replicate4S" %}
11927 ins_encode %{
11928 __ movdl($dst$$XMMRegister, $src$$Register);
11929 __ pshuflw($dst$$XMMRegister, $dst$$XMMRegister, 0x00);
11930 %}
11931 ins_pipe( fpu_reg_reg );
11932 %}
11934 // Replicate scalar zero to packed short (2 byte) values in xmm
11935 instruct Repl4S_immI0(regD dst, immI0 zero) %{
11936 predicate(UseSSE>=2);
11937 match(Set dst (Replicate4S zero));
11938 format %{ "PXOR $dst,$dst\t! replicate4S" %}
11939 ins_encode %{
11940 __ pxor($dst$$XMMRegister, $dst$$XMMRegister);
11941 %}
11942 ins_pipe( fpu_reg_reg );
11943 %}
11945 // Replicate scalar to packed char (2 byte) values in xmm
11946 instruct Repl4C_reg(regD dst, regD src) %{
11947 predicate(UseSSE>=2);
11948 match(Set dst (Replicate4C src));
11949 format %{ "PSHUFLW $dst,$src,0x00\t! replicate4C" %}
11950 ins_encode %{
11951 __ pshuflw($dst$$XMMRegister, $src$$XMMRegister, 0x00);
11952 %}
11953 ins_pipe( fpu_reg_reg );
11954 %}
11956 // Replicate scalar to packed char (2 byte) values in xmm
11957 instruct Repl4C_eRegI(regD dst, eRegI src) %{
11958 predicate(UseSSE>=2);
11959 match(Set dst (Replicate4C src));
11960 format %{ "MOVD $dst,$src\n\t"
11961 "PSHUFLW $dst,$dst,0x00\t! replicate4C" %}
11962 ins_encode %{
11963 __ movdl($dst$$XMMRegister, $src$$Register);
11964 __ pshuflw($dst$$XMMRegister, $dst$$XMMRegister, 0x00);
11965 %}
11966 ins_pipe( fpu_reg_reg );
11967 %}
11969 // Replicate scalar zero to packed char (2 byte) values in xmm
11970 instruct Repl4C_immI0(regD dst, immI0 zero) %{
11971 predicate(UseSSE>=2);
11972 match(Set dst (Replicate4C zero));
11973 format %{ "PXOR $dst,$dst\t! replicate4C" %}
11974 ins_encode %{
11975 __ pxor($dst$$XMMRegister, $dst$$XMMRegister);
11976 %}
11977 ins_pipe( fpu_reg_reg );
11978 %}
11980 // Replicate scalar to packed integer (4 byte) values in xmm
11981 instruct Repl2I_reg(regD dst, regD src) %{
11982 predicate(UseSSE>=2);
11983 match(Set dst (Replicate2I src));
11984 format %{ "PSHUFD $dst,$src,0x00\t! replicate2I" %}
11985 ins_encode %{
11986 __ pshufd($dst$$XMMRegister, $src$$XMMRegister, 0x00);
11987 %}
11988 ins_pipe( fpu_reg_reg );
11989 %}
11991 // Replicate scalar to packed integer (4 byte) values in xmm
11992 instruct Repl2I_eRegI(regD dst, eRegI src) %{
11993 predicate(UseSSE>=2);
11994 match(Set dst (Replicate2I src));
11995 format %{ "MOVD $dst,$src\n\t"
11996 "PSHUFD $dst,$dst,0x00\t! replicate2I" %}
11997 ins_encode %{
11998 __ movdl($dst$$XMMRegister, $src$$Register);
11999 __ pshufd($dst$$XMMRegister, $dst$$XMMRegister, 0x00);
12000 %}
12001 ins_pipe( fpu_reg_reg );
12002 %}
12004 // Replicate scalar zero to packed integer (2 byte) values in xmm
12005 instruct Repl2I_immI0(regD dst, immI0 zero) %{
12006 predicate(UseSSE>=2);
12007 match(Set dst (Replicate2I zero));
12008 format %{ "PXOR $dst,$dst\t! replicate2I" %}
12009 ins_encode %{
12010 __ pxor($dst$$XMMRegister, $dst$$XMMRegister);
12011 %}
12012 ins_pipe( fpu_reg_reg );
12013 %}
12015 // Replicate scalar to packed single precision floating point values in xmm
12016 instruct Repl2F_reg(regD dst, regD src) %{
12017 predicate(UseSSE>=2);
12018 match(Set dst (Replicate2F src));
12019 format %{ "PSHUFD $dst,$src,0xe0\t! replicate2F" %}
12020 ins_encode %{
12021 __ pshufd($dst$$XMMRegister, $src$$XMMRegister, 0xe0);
12022 %}
12023 ins_pipe( fpu_reg_reg );
12024 %}
12026 // Replicate scalar to packed single precision floating point values in xmm
12027 instruct Repl2F_regF(regD dst, regF src) %{
12028 predicate(UseSSE>=2);
12029 match(Set dst (Replicate2F src));
12030 format %{ "PSHUFD $dst,$src,0xe0\t! replicate2F" %}
12031 ins_encode %{
12032 __ pshufd($dst$$XMMRegister, $src$$XMMRegister, 0xe0);
12033 %}
12034 ins_pipe( fpu_reg_reg );
12035 %}
12037 // Replicate scalar to packed single precision floating point values in xmm
12038 instruct Repl2F_immF0(regD dst, immF0 zero) %{
12039 predicate(UseSSE>=2);
12040 match(Set dst (Replicate2F zero));
12041 format %{ "PXOR $dst,$dst\t! replicate2F" %}
12042 ins_encode %{
12043 __ pxor($dst$$XMMRegister, $dst$$XMMRegister);
12044 %}
12045 ins_pipe( fpu_reg_reg );
12046 %}
12048 // =======================================================================
12049 // fast clearing of an array
12050 instruct rep_stos(eCXRegI cnt, eDIRegP base, eAXRegI zero, Universe dummy, eFlagsReg cr) %{
12051 match(Set dummy (ClearArray cnt base));
12052 effect(USE_KILL cnt, USE_KILL base, KILL zero, KILL cr);
12053 format %{ "SHL ECX,1\t# Convert doublewords to words\n\t"
12054 "XOR EAX,EAX\n\t"
12055 "REP STOS\t# store EAX into [EDI++] while ECX--" %}
12056 opcode(0,0x4);
12057 ins_encode( Opcode(0xD1), RegOpc(ECX),
12058 OpcRegReg(0x33,EAX,EAX),
12059 Opcode(0xF3), Opcode(0xAB) );
12060 ins_pipe( pipe_slow );
12061 %}
12063 instruct string_compare(eDIRegP str1, eCXRegI cnt1, eSIRegP str2, eDXRegI cnt2,
12064 eAXRegI result, regD tmp1, eFlagsReg cr) %{
12065 match(Set result (StrComp (Binary str1 cnt1) (Binary str2 cnt2)));
12066 effect(TEMP tmp1, USE_KILL str1, USE_KILL str2, USE_KILL cnt1, USE_KILL cnt2, KILL cr);
12068 format %{ "String Compare $str1,$cnt1,$str2,$cnt2 -> $result // KILL $tmp1" %}
12069 ins_encode %{
12070 __ string_compare($str1$$Register, $str2$$Register,
12071 $cnt1$$Register, $cnt2$$Register, $result$$Register,
12072 $tmp1$$XMMRegister);
12073 %}
12074 ins_pipe( pipe_slow );
12075 %}
12077 // fast string equals
12078 instruct string_equals(eDIRegP str1, eSIRegP str2, eCXRegI cnt, eAXRegI result,
12079 regD tmp1, regD tmp2, eBXRegI tmp3, eFlagsReg cr) %{
12080 match(Set result (StrEquals (Binary str1 str2) cnt));
12081 effect(TEMP tmp1, TEMP tmp2, USE_KILL str1, USE_KILL str2, USE_KILL cnt, KILL tmp3, KILL cr);
12083 format %{ "String Equals $str1,$str2,$cnt -> $result // KILL $tmp1, $tmp2, $tmp3" %}
12084 ins_encode %{
12085 __ char_arrays_equals(false, $str1$$Register, $str2$$Register,
12086 $cnt$$Register, $result$$Register, $tmp3$$Register,
12087 $tmp1$$XMMRegister, $tmp2$$XMMRegister);
12088 %}
12089 ins_pipe( pipe_slow );
12090 %}
12092 // fast search of substring with known size.
12093 instruct string_indexof_con(eDIRegP str1, eDXRegI cnt1, eSIRegP str2, immI int_cnt2,
12094 eBXRegI result, regD vec, eAXRegI cnt2, eCXRegI tmp, eFlagsReg cr) %{
12095 predicate(UseSSE42Intrinsics);
12096 match(Set result (StrIndexOf (Binary str1 cnt1) (Binary str2 int_cnt2)));
12097 effect(TEMP vec, USE_KILL str1, USE_KILL str2, USE_KILL cnt1, KILL cnt2, KILL tmp, KILL cr);
12099 format %{ "String IndexOf $str1,$cnt1,$str2,$int_cnt2 -> $result // KILL $vec, $cnt1, $cnt2, $tmp" %}
12100 ins_encode %{
12101 int icnt2 = (int)$int_cnt2$$constant;
12102 if (icnt2 >= 8) {
12103 // IndexOf for constant substrings with size >= 8 elements
12104 // which don't need to be loaded through stack.
12105 __ string_indexofC8($str1$$Register, $str2$$Register,
12106 $cnt1$$Register, $cnt2$$Register,
12107 icnt2, $result$$Register,
12108 $vec$$XMMRegister, $tmp$$Register);
12109 } else {
12110 // Small strings are loaded through stack if they cross page boundary.
12111 __ string_indexof($str1$$Register, $str2$$Register,
12112 $cnt1$$Register, $cnt2$$Register,
12113 icnt2, $result$$Register,
12114 $vec$$XMMRegister, $tmp$$Register);
12115 }
12116 %}
12117 ins_pipe( pipe_slow );
12118 %}
12120 instruct string_indexof(eDIRegP str1, eDXRegI cnt1, eSIRegP str2, eAXRegI cnt2,
12121 eBXRegI result, regD vec, eCXRegI tmp, eFlagsReg cr) %{
12122 predicate(UseSSE42Intrinsics);
12123 match(Set result (StrIndexOf (Binary str1 cnt1) (Binary str2 cnt2)));
12124 effect(TEMP vec, USE_KILL str1, USE_KILL str2, USE_KILL cnt1, USE_KILL cnt2, KILL tmp, KILL cr);
12126 format %{ "String IndexOf $str1,$cnt1,$str2,$cnt2 -> $result // KILL all" %}
12127 ins_encode %{
12128 __ string_indexof($str1$$Register, $str2$$Register,
12129 $cnt1$$Register, $cnt2$$Register,
12130 (-1), $result$$Register,
12131 $vec$$XMMRegister, $tmp$$Register);
12132 %}
12133 ins_pipe( pipe_slow );
12134 %}
12136 // fast array equals
12137 instruct array_equals(eDIRegP ary1, eSIRegP ary2, eAXRegI result,
12138 regD tmp1, regD tmp2, eCXRegI tmp3, eBXRegI tmp4, eFlagsReg cr)
12139 %{
12140 match(Set result (AryEq ary1 ary2));
12141 effect(TEMP tmp1, TEMP tmp2, USE_KILL ary1, USE_KILL ary2, KILL tmp3, KILL tmp4, KILL cr);
12142 //ins_cost(300);
12144 format %{ "Array Equals $ary1,$ary2 -> $result // KILL $tmp1, $tmp2, $tmp3, $tmp4" %}
12145 ins_encode %{
12146 __ char_arrays_equals(true, $ary1$$Register, $ary2$$Register,
12147 $tmp3$$Register, $result$$Register, $tmp4$$Register,
12148 $tmp1$$XMMRegister, $tmp2$$XMMRegister);
12149 %}
12150 ins_pipe( pipe_slow );
12151 %}
12153 //----------Control Flow Instructions------------------------------------------
12154 // Signed compare Instructions
12155 instruct compI_eReg(eFlagsReg cr, eRegI op1, eRegI op2) %{
12156 match(Set cr (CmpI op1 op2));
12157 effect( DEF cr, USE op1, USE op2 );
12158 format %{ "CMP $op1,$op2" %}
12159 opcode(0x3B); /* Opcode 3B /r */
12160 ins_encode( OpcP, RegReg( op1, op2) );
12161 ins_pipe( ialu_cr_reg_reg );
12162 %}
12164 instruct compI_eReg_imm(eFlagsReg cr, eRegI op1, immI op2) %{
12165 match(Set cr (CmpI op1 op2));
12166 effect( DEF cr, USE op1 );
12167 format %{ "CMP $op1,$op2" %}
12168 opcode(0x81,0x07); /* Opcode 81 /7 */
12169 // ins_encode( RegImm( op1, op2) ); /* Was CmpImm */
12170 ins_encode( OpcSErm( op1, op2 ), Con8or32( op2 ) );
12171 ins_pipe( ialu_cr_reg_imm );
12172 %}
12174 // Cisc-spilled version of cmpI_eReg
12175 instruct compI_eReg_mem(eFlagsReg cr, eRegI op1, memory op2) %{
12176 match(Set cr (CmpI op1 (LoadI op2)));
12178 format %{ "CMP $op1,$op2" %}
12179 ins_cost(500);
12180 opcode(0x3B); /* Opcode 3B /r */
12181 ins_encode( OpcP, RegMem( op1, op2) );
12182 ins_pipe( ialu_cr_reg_mem );
12183 %}
12185 instruct testI_reg( eFlagsReg cr, eRegI src, immI0 zero ) %{
12186 match(Set cr (CmpI src zero));
12187 effect( DEF cr, USE src );
12189 format %{ "TEST $src,$src" %}
12190 opcode(0x85);
12191 ins_encode( OpcP, RegReg( src, src ) );
12192 ins_pipe( ialu_cr_reg_imm );
12193 %}
12195 instruct testI_reg_imm( eFlagsReg cr, eRegI src, immI con, immI0 zero ) %{
12196 match(Set cr (CmpI (AndI src con) zero));
12198 format %{ "TEST $src,$con" %}
12199 opcode(0xF7,0x00);
12200 ins_encode( OpcP, RegOpc(src), Con32(con) );
12201 ins_pipe( ialu_cr_reg_imm );
12202 %}
12204 instruct testI_reg_mem( eFlagsReg cr, eRegI src, memory mem, immI0 zero ) %{
12205 match(Set cr (CmpI (AndI src mem) zero));
12207 format %{ "TEST $src,$mem" %}
12208 opcode(0x85);
12209 ins_encode( OpcP, RegMem( src, mem ) );
12210 ins_pipe( ialu_cr_reg_mem );
12211 %}
12213 // Unsigned compare Instructions; really, same as signed except they
12214 // produce an eFlagsRegU instead of eFlagsReg.
12215 instruct compU_eReg(eFlagsRegU cr, eRegI op1, eRegI op2) %{
12216 match(Set cr (CmpU op1 op2));
12218 format %{ "CMPu $op1,$op2" %}
12219 opcode(0x3B); /* Opcode 3B /r */
12220 ins_encode( OpcP, RegReg( op1, op2) );
12221 ins_pipe( ialu_cr_reg_reg );
12222 %}
12224 instruct compU_eReg_imm(eFlagsRegU cr, eRegI op1, immI op2) %{
12225 match(Set cr (CmpU op1 op2));
12227 format %{ "CMPu $op1,$op2" %}
12228 opcode(0x81,0x07); /* Opcode 81 /7 */
12229 ins_encode( OpcSErm( op1, op2 ), Con8or32( op2 ) );
12230 ins_pipe( ialu_cr_reg_imm );
12231 %}
12233 // // Cisc-spilled version of cmpU_eReg
12234 instruct compU_eReg_mem(eFlagsRegU cr, eRegI op1, memory op2) %{
12235 match(Set cr (CmpU op1 (LoadI op2)));
12237 format %{ "CMPu $op1,$op2" %}
12238 ins_cost(500);
12239 opcode(0x3B); /* Opcode 3B /r */
12240 ins_encode( OpcP, RegMem( op1, op2) );
12241 ins_pipe( ialu_cr_reg_mem );
12242 %}
12244 // // Cisc-spilled version of cmpU_eReg
12245 //instruct compU_mem_eReg(eFlagsRegU cr, memory op1, eRegI op2) %{
12246 // match(Set cr (CmpU (LoadI op1) op2));
12247 //
12248 // format %{ "CMPu $op1,$op2" %}
12249 // ins_cost(500);
12250 // opcode(0x39); /* Opcode 39 /r */
12251 // ins_encode( OpcP, RegMem( op1, op2) );
12252 //%}
12254 instruct testU_reg( eFlagsRegU cr, eRegI src, immI0 zero ) %{
12255 match(Set cr (CmpU src zero));
12257 format %{ "TESTu $src,$src" %}
12258 opcode(0x85);
12259 ins_encode( OpcP, RegReg( src, src ) );
12260 ins_pipe( ialu_cr_reg_imm );
12261 %}
12263 // Unsigned pointer compare Instructions
12264 instruct compP_eReg(eFlagsRegU cr, eRegP op1, eRegP op2) %{
12265 match(Set cr (CmpP op1 op2));
12267 format %{ "CMPu $op1,$op2" %}
12268 opcode(0x3B); /* Opcode 3B /r */
12269 ins_encode( OpcP, RegReg( op1, op2) );
12270 ins_pipe( ialu_cr_reg_reg );
12271 %}
12273 instruct compP_eReg_imm(eFlagsRegU cr, eRegP op1, immP op2) %{
12274 match(Set cr (CmpP op1 op2));
12276 format %{ "CMPu $op1,$op2" %}
12277 opcode(0x81,0x07); /* Opcode 81 /7 */
12278 ins_encode( OpcSErm( op1, op2 ), Con8or32( op2 ) );
12279 ins_pipe( ialu_cr_reg_imm );
12280 %}
12282 // // Cisc-spilled version of cmpP_eReg
12283 instruct compP_eReg_mem(eFlagsRegU cr, eRegP op1, memory op2) %{
12284 match(Set cr (CmpP op1 (LoadP op2)));
12286 format %{ "CMPu $op1,$op2" %}
12287 ins_cost(500);
12288 opcode(0x3B); /* Opcode 3B /r */
12289 ins_encode( OpcP, RegMem( op1, op2) );
12290 ins_pipe( ialu_cr_reg_mem );
12291 %}
12293 // // Cisc-spilled version of cmpP_eReg
12294 //instruct compP_mem_eReg(eFlagsRegU cr, memory op1, eRegP op2) %{
12295 // match(Set cr (CmpP (LoadP op1) op2));
12296 //
12297 // format %{ "CMPu $op1,$op2" %}
12298 // ins_cost(500);
12299 // opcode(0x39); /* Opcode 39 /r */
12300 // ins_encode( OpcP, RegMem( op1, op2) );
12301 //%}
12303 // Compare raw pointer (used in out-of-heap check).
12304 // Only works because non-oop pointers must be raw pointers
12305 // and raw pointers have no anti-dependencies.
12306 instruct compP_mem_eReg( eFlagsRegU cr, eRegP op1, memory op2 ) %{
12307 predicate( !n->in(2)->in(2)->bottom_type()->isa_oop_ptr() );
12308 match(Set cr (CmpP op1 (LoadP op2)));
12310 format %{ "CMPu $op1,$op2" %}
12311 opcode(0x3B); /* Opcode 3B /r */
12312 ins_encode( OpcP, RegMem( op1, op2) );
12313 ins_pipe( ialu_cr_reg_mem );
12314 %}
12316 //
12317 // This will generate a signed flags result. This should be ok
12318 // since any compare to a zero should be eq/neq.
12319 instruct testP_reg( eFlagsReg cr, eRegP src, immP0 zero ) %{
12320 match(Set cr (CmpP src zero));
12322 format %{ "TEST $src,$src" %}
12323 opcode(0x85);
12324 ins_encode( OpcP, RegReg( src, src ) );
12325 ins_pipe( ialu_cr_reg_imm );
12326 %}
12328 // Cisc-spilled version of testP_reg
12329 // This will generate a signed flags result. This should be ok
12330 // since any compare to a zero should be eq/neq.
12331 instruct testP_Reg_mem( eFlagsReg cr, memory op, immI0 zero ) %{
12332 match(Set cr (CmpP (LoadP op) zero));
12334 format %{ "TEST $op,0xFFFFFFFF" %}
12335 ins_cost(500);
12336 opcode(0xF7); /* Opcode F7 /0 */
12337 ins_encode( OpcP, RMopc_Mem(0x00,op), Con_d32(0xFFFFFFFF) );
12338 ins_pipe( ialu_cr_reg_imm );
12339 %}
12341 // Yanked all unsigned pointer compare operations.
12342 // Pointer compares are done with CmpP which is already unsigned.
12344 //----------Max and Min--------------------------------------------------------
12345 // Min Instructions
12346 ////
12347 // *** Min and Max using the conditional move are slower than the
12348 // *** branch version on a Pentium III.
12349 // // Conditional move for min
12350 //instruct cmovI_reg_lt( eRegI op2, eRegI op1, eFlagsReg cr ) %{
12351 // effect( USE_DEF op2, USE op1, USE cr );
12352 // format %{ "CMOVlt $op2,$op1\t! min" %}
12353 // opcode(0x4C,0x0F);
12354 // ins_encode( OpcS, OpcP, RegReg( op2, op1 ) );
12355 // ins_pipe( pipe_cmov_reg );
12356 //%}
12357 //
12358 //// Min Register with Register (P6 version)
12359 //instruct minI_eReg_p6( eRegI op1, eRegI op2 ) %{
12360 // predicate(VM_Version::supports_cmov() );
12361 // match(Set op2 (MinI op1 op2));
12362 // ins_cost(200);
12363 // expand %{
12364 // eFlagsReg cr;
12365 // compI_eReg(cr,op1,op2);
12366 // cmovI_reg_lt(op2,op1,cr);
12367 // %}
12368 //%}
12370 // Min Register with Register (generic version)
12371 instruct minI_eReg(eRegI dst, eRegI src, eFlagsReg flags) %{
12372 match(Set dst (MinI dst src));
12373 effect(KILL flags);
12374 ins_cost(300);
12376 format %{ "MIN $dst,$src" %}
12377 opcode(0xCC);
12378 ins_encode( min_enc(dst,src) );
12379 ins_pipe( pipe_slow );
12380 %}
12382 // Max Register with Register
12383 // *** Min and Max using the conditional move are slower than the
12384 // *** branch version on a Pentium III.
12385 // // Conditional move for max
12386 //instruct cmovI_reg_gt( eRegI op2, eRegI op1, eFlagsReg cr ) %{
12387 // effect( USE_DEF op2, USE op1, USE cr );
12388 // format %{ "CMOVgt $op2,$op1\t! max" %}
12389 // opcode(0x4F,0x0F);
12390 // ins_encode( OpcS, OpcP, RegReg( op2, op1 ) );
12391 // ins_pipe( pipe_cmov_reg );
12392 //%}
12393 //
12394 // // Max Register with Register (P6 version)
12395 //instruct maxI_eReg_p6( eRegI op1, eRegI op2 ) %{
12396 // predicate(VM_Version::supports_cmov() );
12397 // match(Set op2 (MaxI op1 op2));
12398 // ins_cost(200);
12399 // expand %{
12400 // eFlagsReg cr;
12401 // compI_eReg(cr,op1,op2);
12402 // cmovI_reg_gt(op2,op1,cr);
12403 // %}
12404 //%}
12406 // Max Register with Register (generic version)
12407 instruct maxI_eReg(eRegI dst, eRegI src, eFlagsReg flags) %{
12408 match(Set dst (MaxI dst src));
12409 effect(KILL flags);
12410 ins_cost(300);
12412 format %{ "MAX $dst,$src" %}
12413 opcode(0xCC);
12414 ins_encode( max_enc(dst,src) );
12415 ins_pipe( pipe_slow );
12416 %}
12418 // ============================================================================
12419 // Counted Loop limit node which represents exact final iterator value.
12420 // Note: the resulting value should fit into integer range since
12421 // counted loops have limit check on overflow.
12422 instruct loopLimit_eReg(eAXRegI limit, nadxRegI init, immI stride, eDXRegI limit_hi, nadxRegI tmp, eFlagsReg flags) %{
12423 match(Set limit (LoopLimit (Binary init limit) stride));
12424 effect(TEMP limit_hi, TEMP tmp, KILL flags);
12425 ins_cost(300);
12427 format %{ "loopLimit $init,$limit,$stride # $limit = $init + $stride *( $limit - $init + $stride -1)/ $stride, kills $limit_hi" %}
12428 ins_encode %{
12429 int strd = (int)$stride$$constant;
12430 assert(strd != 1 && strd != -1, "sanity");
12431 int m1 = (strd > 0) ? 1 : -1;
12432 // Convert limit to long (EAX:EDX)
12433 __ cdql();
12434 // Convert init to long (init:tmp)
12435 __ movl($tmp$$Register, $init$$Register);
12436 __ sarl($tmp$$Register, 31);
12437 // $limit - $init
12438 __ subl($limit$$Register, $init$$Register);
12439 __ sbbl($limit_hi$$Register, $tmp$$Register);
12440 // + ($stride - 1)
12441 if (strd > 0) {
12442 __ addl($limit$$Register, (strd - 1));
12443 __ adcl($limit_hi$$Register, 0);
12444 __ movl($tmp$$Register, strd);
12445 } else {
12446 __ addl($limit$$Register, (strd + 1));
12447 __ adcl($limit_hi$$Register, -1);
12448 __ lneg($limit_hi$$Register, $limit$$Register);
12449 __ movl($tmp$$Register, -strd);
12450 }
12451 // signed devision: (EAX:EDX) / pos_stride
12452 __ idivl($tmp$$Register);
12453 if (strd < 0) {
12454 // restore sign
12455 __ negl($tmp$$Register);
12456 }
12457 // (EAX) * stride
12458 __ mull($tmp$$Register);
12459 // + init (ignore upper bits)
12460 __ addl($limit$$Register, $init$$Register);
12461 %}
12462 ins_pipe( pipe_slow );
12463 %}
12465 // ============================================================================
12466 // Branch Instructions
12467 // Jump Table
12468 instruct jumpXtnd(eRegI switch_val) %{
12469 match(Jump switch_val);
12470 ins_cost(350);
12471 format %{ "JMP [$constantaddress](,$switch_val,1)\n\t" %}
12472 ins_encode %{
12473 // Jump to Address(table_base + switch_reg)
12474 Address index(noreg, $switch_val$$Register, Address::times_1);
12475 __ jump(ArrayAddress($constantaddress, index));
12476 %}
12477 ins_pipe(pipe_jmp);
12478 %}
12480 // Jump Direct - Label defines a relative address from JMP+1
12481 instruct jmpDir(label labl) %{
12482 match(Goto);
12483 effect(USE labl);
12485 ins_cost(300);
12486 format %{ "JMP $labl" %}
12487 size(5);
12488 ins_encode %{
12489 Label* L = $labl$$label;
12490 __ jmp(*L, false); // Always long jump
12491 %}
12492 ins_pipe( pipe_jmp );
12493 %}
12495 // Jump Direct Conditional - Label defines a relative address from Jcc+1
12496 instruct jmpCon(cmpOp cop, eFlagsReg cr, label labl) %{
12497 match(If cop cr);
12498 effect(USE labl);
12500 ins_cost(300);
12501 format %{ "J$cop $labl" %}
12502 size(6);
12503 ins_encode %{
12504 Label* L = $labl$$label;
12505 __ jcc((Assembler::Condition)($cop$$cmpcode), *L, false); // Always long jump
12506 %}
12507 ins_pipe( pipe_jcc );
12508 %}
12510 // Jump Direct Conditional - Label defines a relative address from Jcc+1
12511 instruct jmpLoopEnd(cmpOp cop, eFlagsReg cr, label labl) %{
12512 match(CountedLoopEnd cop cr);
12513 effect(USE labl);
12515 ins_cost(300);
12516 format %{ "J$cop $labl\t# Loop end" %}
12517 size(6);
12518 ins_encode %{
12519 Label* L = $labl$$label;
12520 __ jcc((Assembler::Condition)($cop$$cmpcode), *L, false); // Always long jump
12521 %}
12522 ins_pipe( pipe_jcc );
12523 %}
12525 // Jump Direct Conditional - Label defines a relative address from Jcc+1
12526 instruct jmpLoopEndU(cmpOpU cop, eFlagsRegU cmp, label labl) %{
12527 match(CountedLoopEnd cop cmp);
12528 effect(USE labl);
12530 ins_cost(300);
12531 format %{ "J$cop,u $labl\t# Loop end" %}
12532 size(6);
12533 ins_encode %{
12534 Label* L = $labl$$label;
12535 __ jcc((Assembler::Condition)($cop$$cmpcode), *L, false); // Always long jump
12536 %}
12537 ins_pipe( pipe_jcc );
12538 %}
12540 instruct jmpLoopEndUCF(cmpOpUCF cop, eFlagsRegUCF cmp, label labl) %{
12541 match(CountedLoopEnd cop cmp);
12542 effect(USE labl);
12544 ins_cost(200);
12545 format %{ "J$cop,u $labl\t# Loop end" %}
12546 size(6);
12547 ins_encode %{
12548 Label* L = $labl$$label;
12549 __ jcc((Assembler::Condition)($cop$$cmpcode), *L, false); // Always long jump
12550 %}
12551 ins_pipe( pipe_jcc );
12552 %}
12554 // Jump Direct Conditional - using unsigned comparison
12555 instruct jmpConU(cmpOpU cop, eFlagsRegU cmp, label labl) %{
12556 match(If cop cmp);
12557 effect(USE labl);
12559 ins_cost(300);
12560 format %{ "J$cop,u $labl" %}
12561 size(6);
12562 ins_encode %{
12563 Label* L = $labl$$label;
12564 __ jcc((Assembler::Condition)($cop$$cmpcode), *L, false); // Always long jump
12565 %}
12566 ins_pipe(pipe_jcc);
12567 %}
12569 instruct jmpConUCF(cmpOpUCF cop, eFlagsRegUCF cmp, label labl) %{
12570 match(If cop cmp);
12571 effect(USE labl);
12573 ins_cost(200);
12574 format %{ "J$cop,u $labl" %}
12575 size(6);
12576 ins_encode %{
12577 Label* L = $labl$$label;
12578 __ jcc((Assembler::Condition)($cop$$cmpcode), *L, false); // Always long jump
12579 %}
12580 ins_pipe(pipe_jcc);
12581 %}
12583 instruct jmpConUCF2(cmpOpUCF2 cop, eFlagsRegUCF cmp, label labl) %{
12584 match(If cop cmp);
12585 effect(USE labl);
12587 ins_cost(200);
12588 format %{ $$template
12589 if ($cop$$cmpcode == Assembler::notEqual) {
12590 $$emit$$"JP,u $labl\n\t"
12591 $$emit$$"J$cop,u $labl"
12592 } else {
12593 $$emit$$"JP,u done\n\t"
12594 $$emit$$"J$cop,u $labl\n\t"
12595 $$emit$$"done:"
12596 }
12597 %}
12598 ins_encode %{
12599 Label* l = $labl$$label;
12600 if ($cop$$cmpcode == Assembler::notEqual) {
12601 __ jcc(Assembler::parity, *l, false);
12602 __ jcc(Assembler::notEqual, *l, false);
12603 } else if ($cop$$cmpcode == Assembler::equal) {
12604 Label done;
12605 __ jccb(Assembler::parity, done);
12606 __ jcc(Assembler::equal, *l, false);
12607 __ bind(done);
12608 } else {
12609 ShouldNotReachHere();
12610 }
12611 %}
12612 ins_pipe(pipe_jcc);
12613 %}
12615 // ============================================================================
12616 // The 2nd slow-half of a subtype check. Scan the subklass's 2ndary superklass
12617 // array for an instance of the superklass. Set a hidden internal cache on a
12618 // hit (cache is checked with exposed code in gen_subtype_check()). Return
12619 // NZ for a miss or zero for a hit. The encoding ALSO sets flags.
12620 instruct partialSubtypeCheck( eDIRegP result, eSIRegP sub, eAXRegP super, eCXRegI rcx, eFlagsReg cr ) %{
12621 match(Set result (PartialSubtypeCheck sub super));
12622 effect( KILL rcx, KILL cr );
12624 ins_cost(1100); // slightly larger than the next version
12625 format %{ "MOV EDI,[$sub+Klass::secondary_supers]\n\t"
12626 "MOV ECX,[EDI+arrayKlass::length]\t# length to scan\n\t"
12627 "ADD EDI,arrayKlass::base_offset\t# Skip to start of data; set NZ in case count is zero\n\t"
12628 "REPNE SCASD\t# Scan *EDI++ for a match with EAX while CX-- != 0\n\t"
12629 "JNE,s miss\t\t# Missed: EDI not-zero\n\t"
12630 "MOV [$sub+Klass::secondary_super_cache],$super\t# Hit: update cache\n\t"
12631 "XOR $result,$result\t\t Hit: EDI zero\n\t"
12632 "miss:\t" %}
12634 opcode(0x1); // Force a XOR of EDI
12635 ins_encode( enc_PartialSubtypeCheck() );
12636 ins_pipe( pipe_slow );
12637 %}
12639 instruct partialSubtypeCheck_vs_Zero( eFlagsReg cr, eSIRegP sub, eAXRegP super, eCXRegI rcx, eDIRegP result, immP0 zero ) %{
12640 match(Set cr (CmpP (PartialSubtypeCheck sub super) zero));
12641 effect( KILL rcx, KILL result );
12643 ins_cost(1000);
12644 format %{ "MOV EDI,[$sub+Klass::secondary_supers]\n\t"
12645 "MOV ECX,[EDI+arrayKlass::length]\t# length to scan\n\t"
12646 "ADD EDI,arrayKlass::base_offset\t# Skip to start of data; set NZ in case count is zero\n\t"
12647 "REPNE SCASD\t# Scan *EDI++ for a match with EAX while CX-- != 0\n\t"
12648 "JNE,s miss\t\t# Missed: flags NZ\n\t"
12649 "MOV [$sub+Klass::secondary_super_cache],$super\t# Hit: update cache, flags Z\n\t"
12650 "miss:\t" %}
12652 opcode(0x0); // No need to XOR EDI
12653 ins_encode( enc_PartialSubtypeCheck() );
12654 ins_pipe( pipe_slow );
12655 %}
12657 // ============================================================================
12658 // Branch Instructions -- short offset versions
12659 //
12660 // These instructions are used to replace jumps of a long offset (the default
12661 // match) with jumps of a shorter offset. These instructions are all tagged
12662 // with the ins_short_branch attribute, which causes the ADLC to suppress the
12663 // match rules in general matching. Instead, the ADLC generates a conversion
12664 // method in the MachNode which can be used to do in-place replacement of the
12665 // long variant with the shorter variant. The compiler will determine if a
12666 // branch can be taken by the is_short_branch_offset() predicate in the machine
12667 // specific code section of the file.
12669 // Jump Direct - Label defines a relative address from JMP+1
12670 instruct jmpDir_short(label labl) %{
12671 match(Goto);
12672 effect(USE labl);
12674 ins_cost(300);
12675 format %{ "JMP,s $labl" %}
12676 size(2);
12677 ins_encode %{
12678 Label* L = $labl$$label;
12679 __ jmpb(*L);
12680 %}
12681 ins_pipe( pipe_jmp );
12682 ins_short_branch(1);
12683 %}
12685 // Jump Direct Conditional - Label defines a relative address from Jcc+1
12686 instruct jmpCon_short(cmpOp cop, eFlagsReg cr, label labl) %{
12687 match(If cop cr);
12688 effect(USE labl);
12690 ins_cost(300);
12691 format %{ "J$cop,s $labl" %}
12692 size(2);
12693 ins_encode %{
12694 Label* L = $labl$$label;
12695 __ jccb((Assembler::Condition)($cop$$cmpcode), *L);
12696 %}
12697 ins_pipe( pipe_jcc );
12698 ins_short_branch(1);
12699 %}
12701 // Jump Direct Conditional - Label defines a relative address from Jcc+1
12702 instruct jmpLoopEnd_short(cmpOp cop, eFlagsReg cr, label labl) %{
12703 match(CountedLoopEnd cop cr);
12704 effect(USE labl);
12706 ins_cost(300);
12707 format %{ "J$cop,s $labl\t# Loop end" %}
12708 size(2);
12709 ins_encode %{
12710 Label* L = $labl$$label;
12711 __ jccb((Assembler::Condition)($cop$$cmpcode), *L);
12712 %}
12713 ins_pipe( pipe_jcc );
12714 ins_short_branch(1);
12715 %}
12717 // Jump Direct Conditional - Label defines a relative address from Jcc+1
12718 instruct jmpLoopEndU_short(cmpOpU cop, eFlagsRegU cmp, label labl) %{
12719 match(CountedLoopEnd cop cmp);
12720 effect(USE labl);
12722 ins_cost(300);
12723 format %{ "J$cop,us $labl\t# Loop end" %}
12724 size(2);
12725 ins_encode %{
12726 Label* L = $labl$$label;
12727 __ jccb((Assembler::Condition)($cop$$cmpcode), *L);
12728 %}
12729 ins_pipe( pipe_jcc );
12730 ins_short_branch(1);
12731 %}
12733 instruct jmpLoopEndUCF_short(cmpOpUCF cop, eFlagsRegUCF cmp, label labl) %{
12734 match(CountedLoopEnd cop cmp);
12735 effect(USE labl);
12737 ins_cost(300);
12738 format %{ "J$cop,us $labl\t# Loop end" %}
12739 size(2);
12740 ins_encode %{
12741 Label* L = $labl$$label;
12742 __ jccb((Assembler::Condition)($cop$$cmpcode), *L);
12743 %}
12744 ins_pipe( pipe_jcc );
12745 ins_short_branch(1);
12746 %}
12748 // Jump Direct Conditional - using unsigned comparison
12749 instruct jmpConU_short(cmpOpU cop, eFlagsRegU cmp, label labl) %{
12750 match(If cop cmp);
12751 effect(USE labl);
12753 ins_cost(300);
12754 format %{ "J$cop,us $labl" %}
12755 size(2);
12756 ins_encode %{
12757 Label* L = $labl$$label;
12758 __ jccb((Assembler::Condition)($cop$$cmpcode), *L);
12759 %}
12760 ins_pipe( pipe_jcc );
12761 ins_short_branch(1);
12762 %}
12764 instruct jmpConUCF_short(cmpOpUCF cop, eFlagsRegUCF cmp, label labl) %{
12765 match(If cop cmp);
12766 effect(USE labl);
12768 ins_cost(300);
12769 format %{ "J$cop,us $labl" %}
12770 size(2);
12771 ins_encode %{
12772 Label* L = $labl$$label;
12773 __ jccb((Assembler::Condition)($cop$$cmpcode), *L);
12774 %}
12775 ins_pipe( pipe_jcc );
12776 ins_short_branch(1);
12777 %}
12779 instruct jmpConUCF2_short(cmpOpUCF2 cop, eFlagsRegUCF cmp, label labl) %{
12780 match(If cop cmp);
12781 effect(USE labl);
12783 ins_cost(300);
12784 format %{ $$template
12785 if ($cop$$cmpcode == Assembler::notEqual) {
12786 $$emit$$"JP,u,s $labl\n\t"
12787 $$emit$$"J$cop,u,s $labl"
12788 } else {
12789 $$emit$$"JP,u,s done\n\t"
12790 $$emit$$"J$cop,u,s $labl\n\t"
12791 $$emit$$"done:"
12792 }
12793 %}
12794 size(4);
12795 ins_encode %{
12796 Label* l = $labl$$label;
12797 if ($cop$$cmpcode == Assembler::notEqual) {
12798 __ jccb(Assembler::parity, *l);
12799 __ jccb(Assembler::notEqual, *l);
12800 } else if ($cop$$cmpcode == Assembler::equal) {
12801 Label done;
12802 __ jccb(Assembler::parity, done);
12803 __ jccb(Assembler::equal, *l);
12804 __ bind(done);
12805 } else {
12806 ShouldNotReachHere();
12807 }
12808 %}
12809 ins_pipe(pipe_jcc);
12810 ins_short_branch(1);
12811 %}
12813 // ============================================================================
12814 // Long Compare
12815 //
12816 // Currently we hold longs in 2 registers. Comparing such values efficiently
12817 // is tricky. The flavor of compare used depends on whether we are testing
12818 // for LT, LE, or EQ. For a simple LT test we can check just the sign bit.
12819 // The GE test is the negated LT test. The LE test can be had by commuting
12820 // the operands (yielding a GE test) and then negating; negate again for the
12821 // GT test. The EQ test is done by ORcc'ing the high and low halves, and the
12822 // NE test is negated from that.
12824 // Due to a shortcoming in the ADLC, it mixes up expressions like:
12825 // (foo (CmpI (CmpL X Y) 0)) and (bar (CmpI (CmpL X 0L) 0)). Note the
12826 // difference between 'Y' and '0L'. The tree-matches for the CmpI sections
12827 // are collapsed internally in the ADLC's dfa-gen code. The match for
12828 // (CmpI (CmpL X Y) 0) is silently replaced with (CmpI (CmpL X 0L) 0) and the
12829 // foo match ends up with the wrong leaf. One fix is to not match both
12830 // reg-reg and reg-zero forms of long-compare. This is unfortunate because
12831 // both forms beat the trinary form of long-compare and both are very useful
12832 // on Intel which has so few registers.
12834 // Manifest a CmpL result in an integer register. Very painful.
12835 // This is the test to avoid.
12836 instruct cmpL3_reg_reg(eSIRegI dst, eRegL src1, eRegL src2, eFlagsReg flags ) %{
12837 match(Set dst (CmpL3 src1 src2));
12838 effect( KILL flags );
12839 ins_cost(1000);
12840 format %{ "XOR $dst,$dst\n\t"
12841 "CMP $src1.hi,$src2.hi\n\t"
12842 "JLT,s m_one\n\t"
12843 "JGT,s p_one\n\t"
12844 "CMP $src1.lo,$src2.lo\n\t"
12845 "JB,s m_one\n\t"
12846 "JEQ,s done\n"
12847 "p_one:\tINC $dst\n\t"
12848 "JMP,s done\n"
12849 "m_one:\tDEC $dst\n"
12850 "done:" %}
12851 ins_encode %{
12852 Label p_one, m_one, done;
12853 __ xorptr($dst$$Register, $dst$$Register);
12854 __ cmpl(HIGH_FROM_LOW($src1$$Register), HIGH_FROM_LOW($src2$$Register));
12855 __ jccb(Assembler::less, m_one);
12856 __ jccb(Assembler::greater, p_one);
12857 __ cmpl($src1$$Register, $src2$$Register);
12858 __ jccb(Assembler::below, m_one);
12859 __ jccb(Assembler::equal, done);
12860 __ bind(p_one);
12861 __ incrementl($dst$$Register);
12862 __ jmpb(done);
12863 __ bind(m_one);
12864 __ decrementl($dst$$Register);
12865 __ bind(done);
12866 %}
12867 ins_pipe( pipe_slow );
12868 %}
12870 //======
12871 // Manifest a CmpL result in the normal flags. Only good for LT or GE
12872 // compares. Can be used for LE or GT compares by reversing arguments.
12873 // NOT GOOD FOR EQ/NE tests.
12874 instruct cmpL_zero_flags_LTGE( flagsReg_long_LTGE flags, eRegL src, immL0 zero ) %{
12875 match( Set flags (CmpL src zero ));
12876 ins_cost(100);
12877 format %{ "TEST $src.hi,$src.hi" %}
12878 opcode(0x85);
12879 ins_encode( OpcP, RegReg_Hi2( src, src ) );
12880 ins_pipe( ialu_cr_reg_reg );
12881 %}
12883 // Manifest a CmpL result in the normal flags. Only good for LT or GE
12884 // compares. Can be used for LE or GT compares by reversing arguments.
12885 // NOT GOOD FOR EQ/NE tests.
12886 instruct cmpL_reg_flags_LTGE( flagsReg_long_LTGE flags, eRegL src1, eRegL src2, eRegI tmp ) %{
12887 match( Set flags (CmpL src1 src2 ));
12888 effect( TEMP tmp );
12889 ins_cost(300);
12890 format %{ "CMP $src1.lo,$src2.lo\t! Long compare; set flags for low bits\n\t"
12891 "MOV $tmp,$src1.hi\n\t"
12892 "SBB $tmp,$src2.hi\t! Compute flags for long compare" %}
12893 ins_encode( long_cmp_flags2( src1, src2, tmp ) );
12894 ins_pipe( ialu_cr_reg_reg );
12895 %}
12897 // Long compares reg < zero/req OR reg >= zero/req.
12898 // Just a wrapper for a normal branch, plus the predicate test.
12899 instruct cmpL_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, label labl) %{
12900 match(If cmp flags);
12901 effect(USE labl);
12902 predicate( _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::lt || _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ge );
12903 expand %{
12904 jmpCon(cmp,flags,labl); // JLT or JGE...
12905 %}
12906 %}
12908 // Compare 2 longs and CMOVE longs.
12909 instruct cmovLL_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, eRegL dst, eRegL src) %{
12910 match(Set dst (CMoveL (Binary cmp flags) (Binary dst src)));
12911 predicate(VM_Version::supports_cmov() && ( _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::lt || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ge ));
12912 ins_cost(400);
12913 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
12914 "CMOV$cmp $dst.hi,$src.hi" %}
12915 opcode(0x0F,0x40);
12916 ins_encode( enc_cmov(cmp), RegReg_Lo2( dst, src ), enc_cmov(cmp), RegReg_Hi2( dst, src ) );
12917 ins_pipe( pipe_cmov_reg_long );
12918 %}
12920 instruct cmovLL_mem_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, eRegL dst, load_long_memory src) %{
12921 match(Set dst (CMoveL (Binary cmp flags) (Binary dst (LoadL src))));
12922 predicate(VM_Version::supports_cmov() && ( _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::lt || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ge ));
12923 ins_cost(500);
12924 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
12925 "CMOV$cmp $dst.hi,$src.hi" %}
12926 opcode(0x0F,0x40);
12927 ins_encode( enc_cmov(cmp), RegMem(dst, src), enc_cmov(cmp), RegMem_Hi(dst, src) );
12928 ins_pipe( pipe_cmov_reg_long );
12929 %}
12931 // Compare 2 longs and CMOVE ints.
12932 instruct cmovII_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, eRegI dst, eRegI src) %{
12933 predicate(VM_Version::supports_cmov() && ( _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::lt || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ge ));
12934 match(Set dst (CMoveI (Binary cmp flags) (Binary dst src)));
12935 ins_cost(200);
12936 format %{ "CMOV$cmp $dst,$src" %}
12937 opcode(0x0F,0x40);
12938 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
12939 ins_pipe( pipe_cmov_reg );
12940 %}
12942 instruct cmovII_mem_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, eRegI dst, memory src) %{
12943 predicate(VM_Version::supports_cmov() && ( _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::lt || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ge ));
12944 match(Set dst (CMoveI (Binary cmp flags) (Binary dst (LoadI src))));
12945 ins_cost(250);
12946 format %{ "CMOV$cmp $dst,$src" %}
12947 opcode(0x0F,0x40);
12948 ins_encode( enc_cmov(cmp), RegMem( dst, src ) );
12949 ins_pipe( pipe_cmov_mem );
12950 %}
12952 // Compare 2 longs and CMOVE ints.
12953 instruct cmovPP_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, eRegP dst, eRegP src) %{
12954 predicate(VM_Version::supports_cmov() && ( _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::lt || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ge ));
12955 match(Set dst (CMoveP (Binary cmp flags) (Binary dst src)));
12956 ins_cost(200);
12957 format %{ "CMOV$cmp $dst,$src" %}
12958 opcode(0x0F,0x40);
12959 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
12960 ins_pipe( pipe_cmov_reg );
12961 %}
12963 // Compare 2 longs and CMOVE doubles
12964 instruct cmovDDPR_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, regDPR dst, regDPR src) %{
12965 predicate( UseSSE<=1 && _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::lt || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ge );
12966 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
12967 ins_cost(200);
12968 expand %{
12969 fcmovDPR_regS(cmp,flags,dst,src);
12970 %}
12971 %}
12973 // Compare 2 longs and CMOVE doubles
12974 instruct cmovDD_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, regD dst, regD src) %{
12975 predicate( UseSSE>=2 && _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::lt || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ge );
12976 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
12977 ins_cost(200);
12978 expand %{
12979 fcmovD_regS(cmp,flags,dst,src);
12980 %}
12981 %}
12983 instruct cmovFFPR_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, regFPR dst, regFPR src) %{
12984 predicate( UseSSE==0 && _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::lt || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ge );
12985 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
12986 ins_cost(200);
12987 expand %{
12988 fcmovFPR_regS(cmp,flags,dst,src);
12989 %}
12990 %}
12992 instruct cmovFF_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, regF dst, regF src) %{
12993 predicate( UseSSE>=1 && _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::lt || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ge );
12994 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
12995 ins_cost(200);
12996 expand %{
12997 fcmovF_regS(cmp,flags,dst,src);
12998 %}
12999 %}
13001 //======
13002 // Manifest a CmpL result in the normal flags. Only good for EQ/NE compares.
13003 instruct cmpL_zero_flags_EQNE( flagsReg_long_EQNE flags, eRegL src, immL0 zero, eRegI tmp ) %{
13004 match( Set flags (CmpL src zero ));
13005 effect(TEMP tmp);
13006 ins_cost(200);
13007 format %{ "MOV $tmp,$src.lo\n\t"
13008 "OR $tmp,$src.hi\t! Long is EQ/NE 0?" %}
13009 ins_encode( long_cmp_flags0( src, tmp ) );
13010 ins_pipe( ialu_reg_reg_long );
13011 %}
13013 // Manifest a CmpL result in the normal flags. Only good for EQ/NE compares.
13014 instruct cmpL_reg_flags_EQNE( flagsReg_long_EQNE flags, eRegL src1, eRegL src2 ) %{
13015 match( Set flags (CmpL src1 src2 ));
13016 ins_cost(200+300);
13017 format %{ "CMP $src1.lo,$src2.lo\t! Long compare; set flags for low bits\n\t"
13018 "JNE,s skip\n\t"
13019 "CMP $src1.hi,$src2.hi\n\t"
13020 "skip:\t" %}
13021 ins_encode( long_cmp_flags1( src1, src2 ) );
13022 ins_pipe( ialu_cr_reg_reg );
13023 %}
13025 // Long compare reg == zero/reg OR reg != zero/reg
13026 // Just a wrapper for a normal branch, plus the predicate test.
13027 instruct cmpL_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, label labl) %{
13028 match(If cmp flags);
13029 effect(USE labl);
13030 predicate( _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::eq || _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ne );
13031 expand %{
13032 jmpCon(cmp,flags,labl); // JEQ or JNE...
13033 %}
13034 %}
13036 // Compare 2 longs and CMOVE longs.
13037 instruct cmovLL_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, eRegL dst, eRegL src) %{
13038 match(Set dst (CMoveL (Binary cmp flags) (Binary dst src)));
13039 predicate(VM_Version::supports_cmov() && ( _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::eq || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ne ));
13040 ins_cost(400);
13041 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
13042 "CMOV$cmp $dst.hi,$src.hi" %}
13043 opcode(0x0F,0x40);
13044 ins_encode( enc_cmov(cmp), RegReg_Lo2( dst, src ), enc_cmov(cmp), RegReg_Hi2( dst, src ) );
13045 ins_pipe( pipe_cmov_reg_long );
13046 %}
13048 instruct cmovLL_mem_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, eRegL dst, load_long_memory src) %{
13049 match(Set dst (CMoveL (Binary cmp flags) (Binary dst (LoadL src))));
13050 predicate(VM_Version::supports_cmov() && ( _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::eq || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ne ));
13051 ins_cost(500);
13052 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
13053 "CMOV$cmp $dst.hi,$src.hi" %}
13054 opcode(0x0F,0x40);
13055 ins_encode( enc_cmov(cmp), RegMem(dst, src), enc_cmov(cmp), RegMem_Hi(dst, src) );
13056 ins_pipe( pipe_cmov_reg_long );
13057 %}
13059 // Compare 2 longs and CMOVE ints.
13060 instruct cmovII_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, eRegI dst, eRegI src) %{
13061 predicate(VM_Version::supports_cmov() && ( _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::eq || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ne ));
13062 match(Set dst (CMoveI (Binary cmp flags) (Binary dst src)));
13063 ins_cost(200);
13064 format %{ "CMOV$cmp $dst,$src" %}
13065 opcode(0x0F,0x40);
13066 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
13067 ins_pipe( pipe_cmov_reg );
13068 %}
13070 instruct cmovII_mem_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, eRegI dst, memory src) %{
13071 predicate(VM_Version::supports_cmov() && ( _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::eq || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ne ));
13072 match(Set dst (CMoveI (Binary cmp flags) (Binary dst (LoadI src))));
13073 ins_cost(250);
13074 format %{ "CMOV$cmp $dst,$src" %}
13075 opcode(0x0F,0x40);
13076 ins_encode( enc_cmov(cmp), RegMem( dst, src ) );
13077 ins_pipe( pipe_cmov_mem );
13078 %}
13080 // Compare 2 longs and CMOVE ints.
13081 instruct cmovPP_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, eRegP dst, eRegP src) %{
13082 predicate(VM_Version::supports_cmov() && ( _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::eq || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ne ));
13083 match(Set dst (CMoveP (Binary cmp flags) (Binary dst src)));
13084 ins_cost(200);
13085 format %{ "CMOV$cmp $dst,$src" %}
13086 opcode(0x0F,0x40);
13087 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
13088 ins_pipe( pipe_cmov_reg );
13089 %}
13091 // Compare 2 longs and CMOVE doubles
13092 instruct cmovDDPR_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, regDPR dst, regDPR src) %{
13093 predicate( UseSSE<=1 && _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::eq || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ne );
13094 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
13095 ins_cost(200);
13096 expand %{
13097 fcmovDPR_regS(cmp,flags,dst,src);
13098 %}
13099 %}
13101 // Compare 2 longs and CMOVE doubles
13102 instruct cmovDD_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, regD dst, regD src) %{
13103 predicate( UseSSE>=2 && _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::eq || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ne );
13104 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
13105 ins_cost(200);
13106 expand %{
13107 fcmovD_regS(cmp,flags,dst,src);
13108 %}
13109 %}
13111 instruct cmovFFPR_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, regFPR dst, regFPR src) %{
13112 predicate( UseSSE==0 && _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::eq || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ne );
13113 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
13114 ins_cost(200);
13115 expand %{
13116 fcmovFPR_regS(cmp,flags,dst,src);
13117 %}
13118 %}
13120 instruct cmovFF_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, regF dst, regF src) %{
13121 predicate( UseSSE>=1 && _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::eq || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ne );
13122 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
13123 ins_cost(200);
13124 expand %{
13125 fcmovF_regS(cmp,flags,dst,src);
13126 %}
13127 %}
13129 //======
13130 // Manifest a CmpL result in the normal flags. Only good for LE or GT compares.
13131 // Same as cmpL_reg_flags_LEGT except must negate src
13132 instruct cmpL_zero_flags_LEGT( flagsReg_long_LEGT flags, eRegL src, immL0 zero, eRegI tmp ) %{
13133 match( Set flags (CmpL src zero ));
13134 effect( TEMP tmp );
13135 ins_cost(300);
13136 format %{ "XOR $tmp,$tmp\t# Long compare for -$src < 0, use commuted test\n\t"
13137 "CMP $tmp,$src.lo\n\t"
13138 "SBB $tmp,$src.hi\n\t" %}
13139 ins_encode( long_cmp_flags3(src, tmp) );
13140 ins_pipe( ialu_reg_reg_long );
13141 %}
13143 // Manifest a CmpL result in the normal flags. Only good for LE or GT compares.
13144 // Same as cmpL_reg_flags_LTGE except operands swapped. Swapping operands
13145 // requires a commuted test to get the same result.
13146 instruct cmpL_reg_flags_LEGT( flagsReg_long_LEGT flags, eRegL src1, eRegL src2, eRegI tmp ) %{
13147 match( Set flags (CmpL src1 src2 ));
13148 effect( TEMP tmp );
13149 ins_cost(300);
13150 format %{ "CMP $src2.lo,$src1.lo\t! Long compare, swapped operands, use with commuted test\n\t"
13151 "MOV $tmp,$src2.hi\n\t"
13152 "SBB $tmp,$src1.hi\t! Compute flags for long compare" %}
13153 ins_encode( long_cmp_flags2( src2, src1, tmp ) );
13154 ins_pipe( ialu_cr_reg_reg );
13155 %}
13157 // Long compares reg < zero/req OR reg >= zero/req.
13158 // Just a wrapper for a normal branch, plus the predicate test
13159 instruct cmpL_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, label labl) %{
13160 match(If cmp flags);
13161 effect(USE labl);
13162 predicate( _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::gt || _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::le );
13163 ins_cost(300);
13164 expand %{
13165 jmpCon(cmp,flags,labl); // JGT or JLE...
13166 %}
13167 %}
13169 // Compare 2 longs and CMOVE longs.
13170 instruct cmovLL_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, eRegL dst, eRegL src) %{
13171 match(Set dst (CMoveL (Binary cmp flags) (Binary dst src)));
13172 predicate(VM_Version::supports_cmov() && ( _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::le || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::gt ));
13173 ins_cost(400);
13174 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
13175 "CMOV$cmp $dst.hi,$src.hi" %}
13176 opcode(0x0F,0x40);
13177 ins_encode( enc_cmov(cmp), RegReg_Lo2( dst, src ), enc_cmov(cmp), RegReg_Hi2( dst, src ) );
13178 ins_pipe( pipe_cmov_reg_long );
13179 %}
13181 instruct cmovLL_mem_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, eRegL dst, load_long_memory src) %{
13182 match(Set dst (CMoveL (Binary cmp flags) (Binary dst (LoadL src))));
13183 predicate(VM_Version::supports_cmov() && ( _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::le || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::gt ));
13184 ins_cost(500);
13185 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
13186 "CMOV$cmp $dst.hi,$src.hi+4" %}
13187 opcode(0x0F,0x40);
13188 ins_encode( enc_cmov(cmp), RegMem(dst, src), enc_cmov(cmp), RegMem_Hi(dst, src) );
13189 ins_pipe( pipe_cmov_reg_long );
13190 %}
13192 // Compare 2 longs and CMOVE ints.
13193 instruct cmovII_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, eRegI dst, eRegI src) %{
13194 predicate(VM_Version::supports_cmov() && ( _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::le || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::gt ));
13195 match(Set dst (CMoveI (Binary cmp flags) (Binary dst src)));
13196 ins_cost(200);
13197 format %{ "CMOV$cmp $dst,$src" %}
13198 opcode(0x0F,0x40);
13199 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
13200 ins_pipe( pipe_cmov_reg );
13201 %}
13203 instruct cmovII_mem_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, eRegI dst, memory src) %{
13204 predicate(VM_Version::supports_cmov() && ( _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::le || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::gt ));
13205 match(Set dst (CMoveI (Binary cmp flags) (Binary dst (LoadI src))));
13206 ins_cost(250);
13207 format %{ "CMOV$cmp $dst,$src" %}
13208 opcode(0x0F,0x40);
13209 ins_encode( enc_cmov(cmp), RegMem( dst, src ) );
13210 ins_pipe( pipe_cmov_mem );
13211 %}
13213 // Compare 2 longs and CMOVE ptrs.
13214 instruct cmovPP_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, eRegP dst, eRegP src) %{
13215 predicate(VM_Version::supports_cmov() && ( _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::le || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::gt ));
13216 match(Set dst (CMoveP (Binary cmp flags) (Binary dst src)));
13217 ins_cost(200);
13218 format %{ "CMOV$cmp $dst,$src" %}
13219 opcode(0x0F,0x40);
13220 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
13221 ins_pipe( pipe_cmov_reg );
13222 %}
13224 // Compare 2 longs and CMOVE doubles
13225 instruct cmovDDPR_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, regDPR dst, regDPR src) %{
13226 predicate( UseSSE<=1 && _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::le || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::gt );
13227 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
13228 ins_cost(200);
13229 expand %{
13230 fcmovDPR_regS(cmp,flags,dst,src);
13231 %}
13232 %}
13234 // Compare 2 longs and CMOVE doubles
13235 instruct cmovDD_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, regD dst, regD src) %{
13236 predicate( UseSSE>=2 && _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::le || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::gt );
13237 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
13238 ins_cost(200);
13239 expand %{
13240 fcmovD_regS(cmp,flags,dst,src);
13241 %}
13242 %}
13244 instruct cmovFFPR_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, regFPR dst, regFPR src) %{
13245 predicate( UseSSE==0 && _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::le || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::gt );
13246 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
13247 ins_cost(200);
13248 expand %{
13249 fcmovFPR_regS(cmp,flags,dst,src);
13250 %}
13251 %}
13254 instruct cmovFF_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, regF dst, regF src) %{
13255 predicate( UseSSE>=1 && _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::le || _kids[0]->_kids[0]->_leaf->as_Bool()->_test._test == BoolTest::gt );
13256 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
13257 ins_cost(200);
13258 expand %{
13259 fcmovF_regS(cmp,flags,dst,src);
13260 %}
13261 %}
13264 // ============================================================================
13265 // Procedure Call/Return Instructions
13266 // Call Java Static Instruction
13267 // Note: If this code changes, the corresponding ret_addr_offset() and
13268 // compute_padding() functions will have to be adjusted.
13269 instruct CallStaticJavaDirect(method meth) %{
13270 match(CallStaticJava);
13271 predicate(! ((CallStaticJavaNode*)n)->is_method_handle_invoke());
13272 effect(USE meth);
13274 ins_cost(300);
13275 format %{ "CALL,static " %}
13276 opcode(0xE8); /* E8 cd */
13277 ins_encode( pre_call_FPU,
13278 Java_Static_Call( meth ),
13279 call_epilog,
13280 post_call_FPU );
13281 ins_pipe( pipe_slow );
13282 ins_alignment(4);
13283 %}
13285 // Call Java Static Instruction (method handle version)
13286 // Note: If this code changes, the corresponding ret_addr_offset() and
13287 // compute_padding() functions will have to be adjusted.
13288 instruct CallStaticJavaHandle(method meth, eBPRegP ebp_mh_SP_save) %{
13289 match(CallStaticJava);
13290 predicate(((CallStaticJavaNode*)n)->is_method_handle_invoke());
13291 effect(USE meth);
13292 // EBP is saved by all callees (for interpreter stack correction).
13293 // We use it here for a similar purpose, in {preserve,restore}_SP.
13295 ins_cost(300);
13296 format %{ "CALL,static/MethodHandle " %}
13297 opcode(0xE8); /* E8 cd */
13298 ins_encode( pre_call_FPU,
13299 preserve_SP,
13300 Java_Static_Call( meth ),
13301 restore_SP,
13302 call_epilog,
13303 post_call_FPU );
13304 ins_pipe( pipe_slow );
13305 ins_alignment(4);
13306 %}
13308 // Call Java Dynamic Instruction
13309 // Note: If this code changes, the corresponding ret_addr_offset() and
13310 // compute_padding() functions will have to be adjusted.
13311 instruct CallDynamicJavaDirect(method meth) %{
13312 match(CallDynamicJava);
13313 effect(USE meth);
13315 ins_cost(300);
13316 format %{ "MOV EAX,(oop)-1\n\t"
13317 "CALL,dynamic" %}
13318 opcode(0xE8); /* E8 cd */
13319 ins_encode( pre_call_FPU,
13320 Java_Dynamic_Call( meth ),
13321 call_epilog,
13322 post_call_FPU );
13323 ins_pipe( pipe_slow );
13324 ins_alignment(4);
13325 %}
13327 // Call Runtime Instruction
13328 instruct CallRuntimeDirect(method meth) %{
13329 match(CallRuntime );
13330 effect(USE meth);
13332 ins_cost(300);
13333 format %{ "CALL,runtime " %}
13334 opcode(0xE8); /* E8 cd */
13335 // Use FFREEs to clear entries in float stack
13336 ins_encode( pre_call_FPU,
13337 FFree_Float_Stack_All,
13338 Java_To_Runtime( meth ),
13339 post_call_FPU );
13340 ins_pipe( pipe_slow );
13341 %}
13343 // Call runtime without safepoint
13344 instruct CallLeafDirect(method meth) %{
13345 match(CallLeaf);
13346 effect(USE meth);
13348 ins_cost(300);
13349 format %{ "CALL_LEAF,runtime " %}
13350 opcode(0xE8); /* E8 cd */
13351 ins_encode( pre_call_FPU,
13352 FFree_Float_Stack_All,
13353 Java_To_Runtime( meth ),
13354 Verify_FPU_For_Leaf, post_call_FPU );
13355 ins_pipe( pipe_slow );
13356 %}
13358 instruct CallLeafNoFPDirect(method meth) %{
13359 match(CallLeafNoFP);
13360 effect(USE meth);
13362 ins_cost(300);
13363 format %{ "CALL_LEAF_NOFP,runtime " %}
13364 opcode(0xE8); /* E8 cd */
13365 ins_encode(Java_To_Runtime(meth));
13366 ins_pipe( pipe_slow );
13367 %}
13370 // Return Instruction
13371 // Remove the return address & jump to it.
13372 instruct Ret() %{
13373 match(Return);
13374 format %{ "RET" %}
13375 opcode(0xC3);
13376 ins_encode(OpcP);
13377 ins_pipe( pipe_jmp );
13378 %}
13380 // Tail Call; Jump from runtime stub to Java code.
13381 // Also known as an 'interprocedural jump'.
13382 // Target of jump will eventually return to caller.
13383 // TailJump below removes the return address.
13384 instruct TailCalljmpInd(eRegP_no_EBP jump_target, eBXRegP method_oop) %{
13385 match(TailCall jump_target method_oop );
13386 ins_cost(300);
13387 format %{ "JMP $jump_target \t# EBX holds method oop" %}
13388 opcode(0xFF, 0x4); /* Opcode FF /4 */
13389 ins_encode( OpcP, RegOpc(jump_target) );
13390 ins_pipe( pipe_jmp );
13391 %}
13394 // Tail Jump; remove the return address; jump to target.
13395 // TailCall above leaves the return address around.
13396 instruct tailjmpInd(eRegP_no_EBP jump_target, eAXRegP ex_oop) %{
13397 match( TailJump jump_target ex_oop );
13398 ins_cost(300);
13399 format %{ "POP EDX\t# pop return address into dummy\n\t"
13400 "JMP $jump_target " %}
13401 opcode(0xFF, 0x4); /* Opcode FF /4 */
13402 ins_encode( enc_pop_rdx,
13403 OpcP, RegOpc(jump_target) );
13404 ins_pipe( pipe_jmp );
13405 %}
13407 // Create exception oop: created by stack-crawling runtime code.
13408 // Created exception is now available to this handler, and is setup
13409 // just prior to jumping to this handler. No code emitted.
13410 instruct CreateException( eAXRegP ex_oop )
13411 %{
13412 match(Set ex_oop (CreateEx));
13414 size(0);
13415 // use the following format syntax
13416 format %{ "# exception oop is in EAX; no code emitted" %}
13417 ins_encode();
13418 ins_pipe( empty );
13419 %}
13422 // Rethrow exception:
13423 // The exception oop will come in the first argument position.
13424 // Then JUMP (not call) to the rethrow stub code.
13425 instruct RethrowException()
13426 %{
13427 match(Rethrow);
13429 // use the following format syntax
13430 format %{ "JMP rethrow_stub" %}
13431 ins_encode(enc_rethrow);
13432 ins_pipe( pipe_jmp );
13433 %}
13435 // inlined locking and unlocking
13438 instruct cmpFastLock( eFlagsReg cr, eRegP object, eRegP box, eAXRegI tmp, eRegP scr) %{
13439 match( Set cr (FastLock object box) );
13440 effect( TEMP tmp, TEMP scr );
13441 ins_cost(300);
13442 format %{ "FASTLOCK $object, $box KILLS $tmp,$scr" %}
13443 ins_encode( Fast_Lock(object,box,tmp,scr) );
13444 ins_pipe( pipe_slow );
13445 %}
13447 instruct cmpFastUnlock( eFlagsReg cr, eRegP object, eAXRegP box, eRegP tmp ) %{
13448 match( Set cr (FastUnlock object box) );
13449 effect( TEMP tmp );
13450 ins_cost(300);
13451 format %{ "FASTUNLOCK $object, $box, $tmp" %}
13452 ins_encode( Fast_Unlock(object,box,tmp) );
13453 ins_pipe( pipe_slow );
13454 %}
13458 // ============================================================================
13459 // Safepoint Instruction
13460 instruct safePoint_poll(eFlagsReg cr) %{
13461 match(SafePoint);
13462 effect(KILL cr);
13464 // TODO-FIXME: we currently poll at offset 0 of the safepoint polling page.
13465 // On SPARC that might be acceptable as we can generate the address with
13466 // just a sethi, saving an or. By polling at offset 0 we can end up
13467 // putting additional pressure on the index-0 in the D$. Because of
13468 // alignment (just like the situation at hand) the lower indices tend
13469 // to see more traffic. It'd be better to change the polling address
13470 // to offset 0 of the last $line in the polling page.
13472 format %{ "TSTL #polladdr,EAX\t! Safepoint: poll for GC" %}
13473 ins_cost(125);
13474 size(6) ;
13475 ins_encode( Safepoint_Poll() );
13476 ins_pipe( ialu_reg_mem );
13477 %}
13479 //----------PEEPHOLE RULES-----------------------------------------------------
13480 // These must follow all instruction definitions as they use the names
13481 // defined in the instructions definitions.
13482 //
13483 // peepmatch ( root_instr_name [preceding_instruction]* );
13484 //
13485 // peepconstraint %{
13486 // (instruction_number.operand_name relational_op instruction_number.operand_name
13487 // [, ...] );
13488 // // instruction numbers are zero-based using left to right order in peepmatch
13489 //
13490 // peepreplace ( instr_name ( [instruction_number.operand_name]* ) );
13491 // // provide an instruction_number.operand_name for each operand that appears
13492 // // in the replacement instruction's match rule
13493 //
13494 // ---------VM FLAGS---------------------------------------------------------
13495 //
13496 // All peephole optimizations can be turned off using -XX:-OptoPeephole
13497 //
13498 // Each peephole rule is given an identifying number starting with zero and
13499 // increasing by one in the order seen by the parser. An individual peephole
13500 // can be enabled, and all others disabled, by using -XX:OptoPeepholeAt=#
13501 // on the command-line.
13502 //
13503 // ---------CURRENT LIMITATIONS----------------------------------------------
13504 //
13505 // Only match adjacent instructions in same basic block
13506 // Only equality constraints
13507 // Only constraints between operands, not (0.dest_reg == EAX_enc)
13508 // Only one replacement instruction
13509 //
13510 // ---------EXAMPLE----------------------------------------------------------
13511 //
13512 // // pertinent parts of existing instructions in architecture description
13513 // instruct movI(eRegI dst, eRegI src) %{
13514 // match(Set dst (CopyI src));
13515 // %}
13516 //
13517 // instruct incI_eReg(eRegI dst, immI1 src, eFlagsReg cr) %{
13518 // match(Set dst (AddI dst src));
13519 // effect(KILL cr);
13520 // %}
13521 //
13522 // // Change (inc mov) to lea
13523 // peephole %{
13524 // // increment preceeded by register-register move
13525 // peepmatch ( incI_eReg movI );
13526 // // require that the destination register of the increment
13527 // // match the destination register of the move
13528 // peepconstraint ( 0.dst == 1.dst );
13529 // // construct a replacement instruction that sets
13530 // // the destination to ( move's source register + one )
13531 // peepreplace ( leaI_eReg_immI( 0.dst 1.src 0.src ) );
13532 // %}
13533 //
13534 // Implementation no longer uses movX instructions since
13535 // machine-independent system no longer uses CopyX nodes.
13536 //
13537 // peephole %{
13538 // peepmatch ( incI_eReg movI );
13539 // peepconstraint ( 0.dst == 1.dst );
13540 // peepreplace ( leaI_eReg_immI( 0.dst 1.src 0.src ) );
13541 // %}
13542 //
13543 // peephole %{
13544 // peepmatch ( decI_eReg movI );
13545 // peepconstraint ( 0.dst == 1.dst );
13546 // peepreplace ( leaI_eReg_immI( 0.dst 1.src 0.src ) );
13547 // %}
13548 //
13549 // peephole %{
13550 // peepmatch ( addI_eReg_imm movI );
13551 // peepconstraint ( 0.dst == 1.dst );
13552 // peepreplace ( leaI_eReg_immI( 0.dst 1.src 0.src ) );
13553 // %}
13554 //
13555 // peephole %{
13556 // peepmatch ( addP_eReg_imm movP );
13557 // peepconstraint ( 0.dst == 1.dst );
13558 // peepreplace ( leaP_eReg_immI( 0.dst 1.src 0.src ) );
13559 // %}
13561 // // Change load of spilled value to only a spill
13562 // instruct storeI(memory mem, eRegI src) %{
13563 // match(Set mem (StoreI mem src));
13564 // %}
13565 //
13566 // instruct loadI(eRegI dst, memory mem) %{
13567 // match(Set dst (LoadI mem));
13568 // %}
13569 //
13570 peephole %{
13571 peepmatch ( loadI storeI );
13572 peepconstraint ( 1.src == 0.dst, 1.mem == 0.mem );
13573 peepreplace ( storeI( 1.mem 1.mem 1.src ) );
13574 %}
13576 //----------SMARTSPILL RULES---------------------------------------------------
13577 // These must follow all instruction definitions as they use the names
13578 // defined in the instructions definitions.