Wed, 19 Mar 2008 15:33:25 -0700
6662967: Optimize I2D conversion on new x86
Summary: Use CVTDQ2PS and CVTDQ2PD for integer values conversions to float and double values on new AMD cpu.
Reviewed-by: sgoldman, never
1 //
2 // Copyright 1997-2007 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 // CA 95054 USA or visit www.sun.com if you need additional information or
21 // have any 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 boundry (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 %{
239 #define RELOC_IMM32 Assembler::imm32_operand
240 #define RELOC_DISP32 Assembler::disp32_operand
242 #define __ _masm.
244 // How to find the high register of a Long pair, given the low register
245 #define HIGH_FROM_LOW(x) ((x)+2)
247 // These masks are used to provide 128-bit aligned bitmasks to the XMM
248 // instructions, to allow sign-masking or sign-bit flipping. They allow
249 // fast versions of NegF/NegD and AbsF/AbsD.
251 // Note: 'double' and 'long long' have 32-bits alignment on x86.
252 static jlong* double_quadword(jlong *adr, jlong lo, jlong hi) {
253 // Use the expression (adr)&(~0xF) to provide 128-bits aligned address
254 // of 128-bits operands for SSE instructions.
255 jlong *operand = (jlong*)(((uintptr_t)adr)&((uintptr_t)(~0xF)));
256 // Store the value to a 128-bits operand.
257 operand[0] = lo;
258 operand[1] = hi;
259 return operand;
260 }
262 // Buffer for 128-bits masks used by SSE instructions.
263 static jlong fp_signmask_pool[(4+1)*2]; // 4*128bits(data) + 128bits(alignment)
265 // Static initialization during VM startup.
266 static jlong *float_signmask_pool = double_quadword(&fp_signmask_pool[1*2], CONST64(0x7FFFFFFF7FFFFFFF), CONST64(0x7FFFFFFF7FFFFFFF));
267 static jlong *double_signmask_pool = double_quadword(&fp_signmask_pool[2*2], CONST64(0x7FFFFFFFFFFFFFFF), CONST64(0x7FFFFFFFFFFFFFFF));
268 static jlong *float_signflip_pool = double_quadword(&fp_signmask_pool[3*2], CONST64(0x8000000080000000), CONST64(0x8000000080000000));
269 static jlong *double_signflip_pool = double_quadword(&fp_signmask_pool[4*2], CONST64(0x8000000000000000), CONST64(0x8000000000000000));
271 // !!!!! Special hack to get all type of calls to specify the byte offset
272 // from the start of the call to the point where the return address
273 // will point.
274 int MachCallStaticJavaNode::ret_addr_offset() {
275 return 5 + (Compile::current()->in_24_bit_fp_mode() ? 6 : 0); // 5 bytes from start of call to where return address points
276 }
278 int MachCallDynamicJavaNode::ret_addr_offset() {
279 return 10 + (Compile::current()->in_24_bit_fp_mode() ? 6 : 0); // 10 bytes from start of call to where return address points
280 }
282 static int sizeof_FFree_Float_Stack_All = -1;
284 int MachCallRuntimeNode::ret_addr_offset() {
285 assert(sizeof_FFree_Float_Stack_All != -1, "must have been emitted already");
286 return sizeof_FFree_Float_Stack_All + 5 + (Compile::current()->in_24_bit_fp_mode() ? 6 : 0);
287 }
289 // Indicate if the safepoint node needs the polling page as an input.
290 // Since x86 does have absolute addressing, it doesn't.
291 bool SafePointNode::needs_polling_address_input() {
292 return false;
293 }
295 //
296 // Compute padding required for nodes which need alignment
297 //
299 // The address of the call instruction needs to be 4-byte aligned to
300 // ensure that it does not span a cache line so that it can be patched.
301 int CallStaticJavaDirectNode::compute_padding(int current_offset) const {
302 if (Compile::current()->in_24_bit_fp_mode())
303 current_offset += 6; // skip fldcw in pre_call_FPU, if any
304 current_offset += 1; // skip call opcode byte
305 return round_to(current_offset, alignment_required()) - current_offset;
306 }
308 // The address of the call instruction needs to be 4-byte aligned to
309 // ensure that it does not span a cache line so that it can be patched.
310 int CallDynamicJavaDirectNode::compute_padding(int current_offset) const {
311 if (Compile::current()->in_24_bit_fp_mode())
312 current_offset += 6; // skip fldcw in pre_call_FPU, if any
313 current_offset += 5; // skip MOV instruction
314 current_offset += 1; // skip call opcode byte
315 return round_to(current_offset, alignment_required()) - current_offset;
316 }
318 #ifndef PRODUCT
319 void MachBreakpointNode::format( PhaseRegAlloc *, outputStream* st ) const {
320 st->print("INT3");
321 }
322 #endif
324 // EMIT_RM()
325 void emit_rm(CodeBuffer &cbuf, int f1, int f2, int f3) {
326 unsigned char c = (unsigned char)((f1 << 6) | (f2 << 3) | f3);
327 *(cbuf.code_end()) = c;
328 cbuf.set_code_end(cbuf.code_end() + 1);
329 }
331 // EMIT_CC()
332 void emit_cc(CodeBuffer &cbuf, int f1, int f2) {
333 unsigned char c = (unsigned char)( f1 | f2 );
334 *(cbuf.code_end()) = c;
335 cbuf.set_code_end(cbuf.code_end() + 1);
336 }
338 // EMIT_OPCODE()
339 void emit_opcode(CodeBuffer &cbuf, int code) {
340 *(cbuf.code_end()) = (unsigned char)code;
341 cbuf.set_code_end(cbuf.code_end() + 1);
342 }
344 // EMIT_OPCODE() w/ relocation information
345 void emit_opcode(CodeBuffer &cbuf, int code, relocInfo::relocType reloc, int offset = 0) {
346 cbuf.relocate(cbuf.inst_mark() + offset, reloc);
347 emit_opcode(cbuf, code);
348 }
350 // EMIT_D8()
351 void emit_d8(CodeBuffer &cbuf, int d8) {
352 *(cbuf.code_end()) = (unsigned char)d8;
353 cbuf.set_code_end(cbuf.code_end() + 1);
354 }
356 // EMIT_D16()
357 void emit_d16(CodeBuffer &cbuf, int d16) {
358 *((short *)(cbuf.code_end())) = d16;
359 cbuf.set_code_end(cbuf.code_end() + 2);
360 }
362 // EMIT_D32()
363 void emit_d32(CodeBuffer &cbuf, int d32) {
364 *((int *)(cbuf.code_end())) = d32;
365 cbuf.set_code_end(cbuf.code_end() + 4);
366 }
368 // emit 32 bit value and construct relocation entry from relocInfo::relocType
369 void emit_d32_reloc(CodeBuffer &cbuf, int d32, relocInfo::relocType reloc,
370 int format) {
371 cbuf.relocate(cbuf.inst_mark(), reloc, format);
373 *((int *)(cbuf.code_end())) = d32;
374 cbuf.set_code_end(cbuf.code_end() + 4);
375 }
377 // emit 32 bit value and construct relocation entry from RelocationHolder
378 void emit_d32_reloc(CodeBuffer &cbuf, int d32, RelocationHolder const& rspec,
379 int format) {
380 #ifdef ASSERT
381 if (rspec.reloc()->type() == relocInfo::oop_type && d32 != 0 && d32 != (int)Universe::non_oop_word()) {
382 assert(oop(d32)->is_oop() && oop(d32)->is_perm(), "cannot embed non-perm oops in code");
383 }
384 #endif
385 cbuf.relocate(cbuf.inst_mark(), rspec, format);
387 *((int *)(cbuf.code_end())) = d32;
388 cbuf.set_code_end(cbuf.code_end() + 4);
389 }
391 // Access stack slot for load or store
392 void store_to_stackslot(CodeBuffer &cbuf, int opcode, int rm_field, int disp) {
393 emit_opcode( cbuf, opcode ); // (e.g., FILD [ESP+src])
394 if( -128 <= disp && disp <= 127 ) {
395 emit_rm( cbuf, 0x01, rm_field, ESP_enc ); // R/M byte
396 emit_rm( cbuf, 0x00, ESP_enc, ESP_enc); // SIB byte
397 emit_d8 (cbuf, disp); // Displacement // R/M byte
398 } else {
399 emit_rm( cbuf, 0x02, rm_field, ESP_enc ); // R/M byte
400 emit_rm( cbuf, 0x00, ESP_enc, ESP_enc); // SIB byte
401 emit_d32(cbuf, disp); // Displacement // R/M byte
402 }
403 }
405 // eRegI ereg, memory mem) %{ // emit_reg_mem
406 void encode_RegMem( CodeBuffer &cbuf, int reg_encoding, int base, int index, int scale, int displace, bool displace_is_oop ) {
407 // There is no index & no scale, use form without SIB byte
408 if ((index == 0x4) &&
409 (scale == 0) && (base != ESP_enc)) {
410 // If no displacement, mode is 0x0; unless base is [EBP]
411 if ( (displace == 0) && (base != EBP_enc) ) {
412 emit_rm(cbuf, 0x0, reg_encoding, base);
413 }
414 else { // If 8-bit displacement, mode 0x1
415 if ((displace >= -128) && (displace <= 127)
416 && !(displace_is_oop) ) {
417 emit_rm(cbuf, 0x1, reg_encoding, base);
418 emit_d8(cbuf, displace);
419 }
420 else { // If 32-bit displacement
421 if (base == -1) { // Special flag for absolute address
422 emit_rm(cbuf, 0x0, reg_encoding, 0x5);
423 // (manual lies; no SIB needed here)
424 if ( displace_is_oop ) {
425 emit_d32_reloc(cbuf, displace, relocInfo::oop_type, 1);
426 } else {
427 emit_d32 (cbuf, displace);
428 }
429 }
430 else { // Normal base + offset
431 emit_rm(cbuf, 0x2, reg_encoding, base);
432 if ( displace_is_oop ) {
433 emit_d32_reloc(cbuf, displace, relocInfo::oop_type, 1);
434 } else {
435 emit_d32 (cbuf, displace);
436 }
437 }
438 }
439 }
440 }
441 else { // Else, encode with the SIB byte
442 // If no displacement, mode is 0x0; unless base is [EBP]
443 if (displace == 0 && (base != EBP_enc)) { // If no displacement
444 emit_rm(cbuf, 0x0, reg_encoding, 0x4);
445 emit_rm(cbuf, scale, index, base);
446 }
447 else { // If 8-bit displacement, mode 0x1
448 if ((displace >= -128) && (displace <= 127)
449 && !(displace_is_oop) ) {
450 emit_rm(cbuf, 0x1, reg_encoding, 0x4);
451 emit_rm(cbuf, scale, index, base);
452 emit_d8(cbuf, displace);
453 }
454 else { // If 32-bit displacement
455 if (base == 0x04 ) {
456 emit_rm(cbuf, 0x2, reg_encoding, 0x4);
457 emit_rm(cbuf, scale, index, 0x04);
458 } else {
459 emit_rm(cbuf, 0x2, reg_encoding, 0x4);
460 emit_rm(cbuf, scale, index, base);
461 }
462 if ( displace_is_oop ) {
463 emit_d32_reloc(cbuf, displace, relocInfo::oop_type, 1);
464 } else {
465 emit_d32 (cbuf, displace);
466 }
467 }
468 }
469 }
470 }
473 void encode_Copy( CodeBuffer &cbuf, int dst_encoding, int src_encoding ) {
474 if( dst_encoding == src_encoding ) {
475 // reg-reg copy, use an empty encoding
476 } else {
477 emit_opcode( cbuf, 0x8B );
478 emit_rm(cbuf, 0x3, dst_encoding, src_encoding );
479 }
480 }
482 void encode_CopyXD( CodeBuffer &cbuf, int dst_encoding, int src_encoding ) {
483 if( dst_encoding == src_encoding ) {
484 // reg-reg copy, use an empty encoding
485 } else {
486 MacroAssembler _masm(&cbuf);
488 __ movdqa(as_XMMRegister(dst_encoding), as_XMMRegister(src_encoding));
489 }
490 }
493 //=============================================================================
494 #ifndef PRODUCT
495 void MachPrologNode::format( PhaseRegAlloc *ra_, outputStream* st ) const {
496 Compile* C = ra_->C;
497 if( C->in_24_bit_fp_mode() ) {
498 tty->print("FLDCW 24 bit fpu control word");
499 tty->print_cr(""); tty->print("\t");
500 }
502 int framesize = C->frame_slots() << LogBytesPerInt;
503 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
504 // Remove two words for return addr and rbp,
505 framesize -= 2*wordSize;
507 // Calls to C2R adapters often do not accept exceptional returns.
508 // We require that their callers must bang for them. But be careful, because
509 // some VM calls (such as call site linkage) can use several kilobytes of
510 // stack. But the stack safety zone should account for that.
511 // See bugs 4446381, 4468289, 4497237.
512 if (C->need_stack_bang(framesize)) {
513 tty->print_cr("# stack bang"); tty->print("\t");
514 }
515 tty->print_cr("PUSHL EBP"); tty->print("\t");
517 if( VerifyStackAtCalls ) { // Majik cookie to verify stack depth
518 tty->print("PUSH 0xBADB100D\t# Majik cookie for stack depth check");
519 tty->print_cr(""); tty->print("\t");
520 framesize -= wordSize;
521 }
523 if ((C->in_24_bit_fp_mode() || VerifyStackAtCalls ) && framesize < 128 ) {
524 if (framesize) {
525 tty->print("SUB ESP,%d\t# Create frame",framesize);
526 }
527 } else {
528 tty->print("SUB ESP,%d\t# Create frame",framesize);
529 }
530 }
531 #endif
534 void MachPrologNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
535 Compile* C = ra_->C;
537 if (UseSSE >= 2 && VerifyFPU) {
538 MacroAssembler masm(&cbuf);
539 masm.verify_FPU(0, "FPU stack must be clean on entry");
540 }
542 // WARNING: Initial instruction MUST be 5 bytes or longer so that
543 // NativeJump::patch_verified_entry will be able to patch out the entry
544 // code safely. The fldcw is ok at 6 bytes, the push to verify stack
545 // depth is ok at 5 bytes, the frame allocation can be either 3 or
546 // 6 bytes. So if we don't do the fldcw or the push then we must
547 // use the 6 byte frame allocation even if we have no frame. :-(
548 // If method sets FPU control word do it now
549 if( C->in_24_bit_fp_mode() ) {
550 MacroAssembler masm(&cbuf);
551 masm.fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
552 }
554 int framesize = C->frame_slots() << LogBytesPerInt;
555 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
556 // Remove two words for return addr and rbp,
557 framesize -= 2*wordSize;
559 // Calls to C2R adapters often do not accept exceptional returns.
560 // We require that their callers must bang for them. But be careful, because
561 // some VM calls (such as call site linkage) can use several kilobytes of
562 // stack. But the stack safety zone should account for that.
563 // See bugs 4446381, 4468289, 4497237.
564 if (C->need_stack_bang(framesize)) {
565 MacroAssembler masm(&cbuf);
566 masm.generate_stack_overflow_check(framesize);
567 }
569 // We always push rbp, so that on return to interpreter rbp, will be
570 // restored correctly and we can correct the stack.
571 emit_opcode(cbuf, 0x50 | EBP_enc);
573 if( VerifyStackAtCalls ) { // Majik cookie to verify stack depth
574 emit_opcode(cbuf, 0x68); // push 0xbadb100d
575 emit_d32(cbuf, 0xbadb100d);
576 framesize -= wordSize;
577 }
579 if ((C->in_24_bit_fp_mode() || VerifyStackAtCalls ) && framesize < 128 ) {
580 if (framesize) {
581 emit_opcode(cbuf, 0x83); // sub SP,#framesize
582 emit_rm(cbuf, 0x3, 0x05, ESP_enc);
583 emit_d8(cbuf, framesize);
584 }
585 } else {
586 emit_opcode(cbuf, 0x81); // sub SP,#framesize
587 emit_rm(cbuf, 0x3, 0x05, ESP_enc);
588 emit_d32(cbuf, framesize);
589 }
590 C->set_frame_complete(cbuf.code_end() - cbuf.code_begin());
592 #ifdef ASSERT
593 if (VerifyStackAtCalls) {
594 Label L;
595 MacroAssembler masm(&cbuf);
596 masm.pushl(rax);
597 masm.movl(rax, rsp);
598 masm.andl(rax, StackAlignmentInBytes-1);
599 masm.cmpl(rax, StackAlignmentInBytes-wordSize);
600 masm.popl(rax);
601 masm.jcc(Assembler::equal, L);
602 masm.stop("Stack is not properly aligned!");
603 masm.bind(L);
604 }
605 #endif
607 }
609 uint MachPrologNode::size(PhaseRegAlloc *ra_) const {
610 return MachNode::size(ra_); // too many variables; just compute it the hard way
611 }
613 int MachPrologNode::reloc() const {
614 return 0; // a large enough number
615 }
617 //=============================================================================
618 #ifndef PRODUCT
619 void MachEpilogNode::format( PhaseRegAlloc *ra_, outputStream* st ) const {
620 Compile *C = ra_->C;
621 int framesize = C->frame_slots() << LogBytesPerInt;
622 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
623 // Remove two words for return addr and rbp,
624 framesize -= 2*wordSize;
626 if( C->in_24_bit_fp_mode() ) {
627 st->print("FLDCW standard control word");
628 st->cr(); st->print("\t");
629 }
630 if( framesize ) {
631 st->print("ADD ESP,%d\t# Destroy frame",framesize);
632 st->cr(); st->print("\t");
633 }
634 st->print_cr("POPL EBP"); st->print("\t");
635 if( do_polling() && C->is_method_compilation() ) {
636 st->print("TEST PollPage,EAX\t! Poll Safepoint");
637 st->cr(); st->print("\t");
638 }
639 }
640 #endif
642 void MachEpilogNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
643 Compile *C = ra_->C;
645 // If method set FPU control word, restore to standard control word
646 if( C->in_24_bit_fp_mode() ) {
647 MacroAssembler masm(&cbuf);
648 masm.fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
649 }
651 int framesize = C->frame_slots() << LogBytesPerInt;
652 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
653 // Remove two words for return addr and rbp,
654 framesize -= 2*wordSize;
656 // Note that VerifyStackAtCalls' Majik cookie does not change the frame size popped here
658 if( framesize >= 128 ) {
659 emit_opcode(cbuf, 0x81); // add SP, #framesize
660 emit_rm(cbuf, 0x3, 0x00, ESP_enc);
661 emit_d32(cbuf, framesize);
662 }
663 else if( framesize ) {
664 emit_opcode(cbuf, 0x83); // add SP, #framesize
665 emit_rm(cbuf, 0x3, 0x00, ESP_enc);
666 emit_d8(cbuf, framesize);
667 }
669 emit_opcode(cbuf, 0x58 | EBP_enc);
671 if( do_polling() && C->is_method_compilation() ) {
672 cbuf.relocate(cbuf.code_end(), relocInfo::poll_return_type, 0);
673 emit_opcode(cbuf,0x85);
674 emit_rm(cbuf, 0x0, EAX_enc, 0x5); // EAX
675 emit_d32(cbuf, (intptr_t)os::get_polling_page());
676 }
677 }
679 uint MachEpilogNode::size(PhaseRegAlloc *ra_) const {
680 Compile *C = ra_->C;
681 // If method set FPU control word, restore to standard control word
682 int size = C->in_24_bit_fp_mode() ? 6 : 0;
683 if( do_polling() && C->is_method_compilation() ) size += 6;
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 size++; // popl rbp,
692 if( framesize >= 128 ) {
693 size += 6;
694 } else {
695 size += framesize ? 3 : 0;
696 }
697 return size;
698 }
700 int MachEpilogNode::reloc() const {
701 return 0; // a large enough number
702 }
704 const Pipeline * MachEpilogNode::pipeline() const {
705 return MachNode::pipeline_class();
706 }
708 int MachEpilogNode::safepoint_offset() const { return 0; }
710 //=============================================================================
712 enum RC { rc_bad, rc_int, rc_float, rc_xmm, rc_stack };
713 static enum RC rc_class( OptoReg::Name reg ) {
715 if( !OptoReg::is_valid(reg) ) return rc_bad;
716 if (OptoReg::is_stack(reg)) return rc_stack;
718 VMReg r = OptoReg::as_VMReg(reg);
719 if (r->is_Register()) return rc_int;
720 if (r->is_FloatRegister()) {
721 assert(UseSSE < 2, "shouldn't be used in SSE2+ mode");
722 return rc_float;
723 }
724 assert(r->is_XMMRegister(), "must be");
725 return rc_xmm;
726 }
728 static int impl_helper( CodeBuffer *cbuf, bool do_size, bool is_load, int offset, int reg, int opcode, const char *op_str, int size ) {
729 if( cbuf ) {
730 emit_opcode (*cbuf, opcode );
731 encode_RegMem(*cbuf, Matcher::_regEncode[reg], ESP_enc, 0x4, 0, offset, false);
732 #ifndef PRODUCT
733 } else if( !do_size ) {
734 if( size != 0 ) tty->print("\n\t");
735 if( opcode == 0x8B || opcode == 0x89 ) { // MOV
736 if( is_load ) tty->print("%s %s,[ESP + #%d]",op_str,Matcher::regName[reg],offset);
737 else tty->print("%s [ESP + #%d],%s",op_str,offset,Matcher::regName[reg]);
738 } else { // FLD, FST, PUSH, POP
739 tty->print("%s [ESP + #%d]",op_str,offset);
740 }
741 #endif
742 }
743 int offset_size = (offset == 0) ? 0 : ((offset <= 127) ? 1 : 4);
744 return size+3+offset_size;
745 }
747 // Helper for XMM registers. Extra opcode bits, limited syntax.
748 static int impl_x_helper( CodeBuffer *cbuf, bool do_size, bool is_load,
749 int offset, int reg_lo, int reg_hi, int size ) {
750 if( cbuf ) {
751 if( reg_lo+1 == reg_hi ) { // double move?
752 if( is_load && !UseXmmLoadAndClearUpper )
753 emit_opcode(*cbuf, 0x66 ); // use 'movlpd' for load
754 else
755 emit_opcode(*cbuf, 0xF2 ); // use 'movsd' otherwise
756 } else {
757 emit_opcode(*cbuf, 0xF3 );
758 }
759 emit_opcode(*cbuf, 0x0F );
760 if( reg_lo+1 == reg_hi && is_load && !UseXmmLoadAndClearUpper )
761 emit_opcode(*cbuf, 0x12 ); // use 'movlpd' for load
762 else
763 emit_opcode(*cbuf, is_load ? 0x10 : 0x11 );
764 encode_RegMem(*cbuf, Matcher::_regEncode[reg_lo], ESP_enc, 0x4, 0, offset, false);
765 #ifndef PRODUCT
766 } else if( !do_size ) {
767 if( size != 0 ) tty->print("\n\t");
768 if( reg_lo+1 == reg_hi ) { // double move?
769 if( is_load ) tty->print("%s %s,[ESP + #%d]",
770 UseXmmLoadAndClearUpper ? "MOVSD " : "MOVLPD",
771 Matcher::regName[reg_lo], offset);
772 else tty->print("MOVSD [ESP + #%d],%s",
773 offset, Matcher::regName[reg_lo]);
774 } else {
775 if( is_load ) tty->print("MOVSS %s,[ESP + #%d]",
776 Matcher::regName[reg_lo], offset);
777 else tty->print("MOVSS [ESP + #%d],%s",
778 offset, Matcher::regName[reg_lo]);
779 }
780 #endif
781 }
782 int offset_size = (offset == 0) ? 0 : ((offset <= 127) ? 1 : 4);
783 return size+5+offset_size;
784 }
787 static int impl_movx_helper( CodeBuffer *cbuf, bool do_size, int src_lo, int dst_lo,
788 int src_hi, int dst_hi, int size ) {
789 if( UseXmmRegToRegMoveAll ) {//Use movaps,movapd to move between xmm registers
790 if( cbuf ) {
791 if( (src_lo+1 == src_hi && dst_lo+1 == dst_hi) ) {
792 emit_opcode(*cbuf, 0x66 );
793 }
794 emit_opcode(*cbuf, 0x0F );
795 emit_opcode(*cbuf, 0x28 );
796 emit_rm (*cbuf, 0x3, Matcher::_regEncode[dst_lo], Matcher::_regEncode[src_lo] );
797 #ifndef PRODUCT
798 } else if( !do_size ) {
799 if( size != 0 ) tty->print("\n\t");
800 if( src_lo+1 == src_hi && dst_lo+1 == dst_hi ) { // double move?
801 tty->print("MOVAPD %s,%s",Matcher::regName[dst_lo],Matcher::regName[src_lo]);
802 } else {
803 tty->print("MOVAPS %s,%s",Matcher::regName[dst_lo],Matcher::regName[src_lo]);
804 }
805 #endif
806 }
807 return size + ((src_lo+1 == src_hi && dst_lo+1 == dst_hi) ? 4 : 3);
808 } else {
809 if( cbuf ) {
810 emit_opcode(*cbuf, (src_lo+1 == src_hi && dst_lo+1 == dst_hi) ? 0xF2 : 0xF3 );
811 emit_opcode(*cbuf, 0x0F );
812 emit_opcode(*cbuf, 0x10 );
813 emit_rm (*cbuf, 0x3, Matcher::_regEncode[dst_lo], Matcher::_regEncode[src_lo] );
814 #ifndef PRODUCT
815 } else if( !do_size ) {
816 if( size != 0 ) tty->print("\n\t");
817 if( src_lo+1 == src_hi && dst_lo+1 == dst_hi ) { // double move?
818 tty->print("MOVSD %s,%s",Matcher::regName[dst_lo],Matcher::regName[src_lo]);
819 } else {
820 tty->print("MOVSS %s,%s",Matcher::regName[dst_lo],Matcher::regName[src_lo]);
821 }
822 #endif
823 }
824 return size+4;
825 }
826 }
828 static int impl_mov_helper( CodeBuffer *cbuf, bool do_size, int src, int dst, int size ) {
829 if( cbuf ) {
830 emit_opcode(*cbuf, 0x8B );
831 emit_rm (*cbuf, 0x3, Matcher::_regEncode[dst], Matcher::_regEncode[src] );
832 #ifndef PRODUCT
833 } else if( !do_size ) {
834 if( size != 0 ) tty->print("\n\t");
835 tty->print("MOV %s,%s",Matcher::regName[dst],Matcher::regName[src]);
836 #endif
837 }
838 return size+2;
839 }
841 static int impl_fp_store_helper( CodeBuffer *cbuf, bool do_size, int src_lo, int src_hi, int dst_lo, int dst_hi, int offset, int size ) {
842 if( src_lo != FPR1L_num ) { // Move value to top of FP stack, if not already there
843 if( cbuf ) {
844 emit_opcode( *cbuf, 0xD9 ); // FLD (i.e., push it)
845 emit_d8( *cbuf, 0xC0-1+Matcher::_regEncode[src_lo] );
846 #ifndef PRODUCT
847 } else if( !do_size ) {
848 if( size != 0 ) tty->print("\n\t");
849 tty->print("FLD %s",Matcher::regName[src_lo]);
850 #endif
851 }
852 size += 2;
853 }
855 int st_op = (src_lo != FPR1L_num) ? EBX_num /*store & pop*/ : EDX_num /*store no pop*/;
856 const char *op_str;
857 int op;
858 if( src_lo+1 == src_hi && dst_lo+1 == dst_hi ) { // double store?
859 op_str = (src_lo != FPR1L_num) ? "FSTP_D" : "FST_D ";
860 op = 0xDD;
861 } else { // 32-bit store
862 op_str = (src_lo != FPR1L_num) ? "FSTP_S" : "FST_S ";
863 op = 0xD9;
864 assert( !OptoReg::is_valid(src_hi) && !OptoReg::is_valid(dst_hi), "no non-adjacent float-stores" );
865 }
867 return impl_helper(cbuf,do_size,false,offset,st_op,op,op_str,size);
868 }
870 uint MachSpillCopyNode::implementation( CodeBuffer *cbuf, PhaseRegAlloc *ra_, bool do_size, outputStream* st ) const {
871 // Get registers to move
872 OptoReg::Name src_second = ra_->get_reg_second(in(1));
873 OptoReg::Name src_first = ra_->get_reg_first(in(1));
874 OptoReg::Name dst_second = ra_->get_reg_second(this );
875 OptoReg::Name dst_first = ra_->get_reg_first(this );
877 enum RC src_second_rc = rc_class(src_second);
878 enum RC src_first_rc = rc_class(src_first);
879 enum RC dst_second_rc = rc_class(dst_second);
880 enum RC dst_first_rc = rc_class(dst_first);
882 assert( OptoReg::is_valid(src_first) && OptoReg::is_valid(dst_first), "must move at least 1 register" );
884 // Generate spill code!
885 int size = 0;
887 if( src_first == dst_first && src_second == dst_second )
888 return size; // Self copy, no move
890 // --------------------------------------
891 // Check for mem-mem move. push/pop to move.
892 if( src_first_rc == rc_stack && dst_first_rc == rc_stack ) {
893 if( src_second == dst_first ) { // overlapping stack copy ranges
894 assert( src_second_rc == rc_stack && dst_second_rc == rc_stack, "we only expect a stk-stk copy here" );
895 size = impl_helper(cbuf,do_size,true ,ra_->reg2offset(src_second),ESI_num,0xFF,"PUSH ",size);
896 size = impl_helper(cbuf,do_size,false,ra_->reg2offset(dst_second),EAX_num,0x8F,"POP ",size);
897 src_second_rc = dst_second_rc = rc_bad; // flag as already moved the second bits
898 }
899 // move low bits
900 size = impl_helper(cbuf,do_size,true ,ra_->reg2offset(src_first),ESI_num,0xFF,"PUSH ",size);
901 size = impl_helper(cbuf,do_size,false,ra_->reg2offset(dst_first),EAX_num,0x8F,"POP ",size);
902 if( src_second_rc == rc_stack && dst_second_rc == rc_stack ) { // mov second bits
903 size = impl_helper(cbuf,do_size,true ,ra_->reg2offset(src_second),ESI_num,0xFF,"PUSH ",size);
904 size = impl_helper(cbuf,do_size,false,ra_->reg2offset(dst_second),EAX_num,0x8F,"POP ",size);
905 }
906 return size;
907 }
909 // --------------------------------------
910 // Check for integer reg-reg copy
911 if( src_first_rc == rc_int && dst_first_rc == rc_int )
912 size = impl_mov_helper(cbuf,do_size,src_first,dst_first,size);
914 // Check for integer store
915 if( src_first_rc == rc_int && dst_first_rc == rc_stack )
916 size = impl_helper(cbuf,do_size,false,ra_->reg2offset(dst_first),src_first,0x89,"MOV ",size);
918 // Check for integer load
919 if( dst_first_rc == rc_int && src_first_rc == rc_stack )
920 size = impl_helper(cbuf,do_size,true ,ra_->reg2offset(src_first),dst_first,0x8B,"MOV ",size);
922 // --------------------------------------
923 // Check for float reg-reg copy
924 if( src_first_rc == rc_float && dst_first_rc == rc_float ) {
925 assert( (src_second_rc == rc_bad && dst_second_rc == rc_bad) ||
926 (src_first+1 == src_second && dst_first+1 == dst_second), "no non-adjacent float-moves" );
927 if( cbuf ) {
929 // Note the mucking with the register encode to compensate for the 0/1
930 // indexing issue mentioned in a comment in the reg_def sections
931 // for FPR registers many lines above here.
933 if( src_first != FPR1L_num ) {
934 emit_opcode (*cbuf, 0xD9 ); // FLD ST(i)
935 emit_d8 (*cbuf, 0xC0+Matcher::_regEncode[src_first]-1 );
936 emit_opcode (*cbuf, 0xDD ); // FSTP ST(i)
937 emit_d8 (*cbuf, 0xD8+Matcher::_regEncode[dst_first] );
938 } else {
939 emit_opcode (*cbuf, 0xDD ); // FST ST(i)
940 emit_d8 (*cbuf, 0xD0+Matcher::_regEncode[dst_first]-1 );
941 }
942 #ifndef PRODUCT
943 } else if( !do_size ) {
944 if( size != 0 ) st->print("\n\t");
945 if( src_first != FPR1L_num ) st->print("FLD %s\n\tFSTP %s",Matcher::regName[src_first],Matcher::regName[dst_first]);
946 else st->print( "FST %s", Matcher::regName[dst_first]);
947 #endif
948 }
949 return size + ((src_first != FPR1L_num) ? 2+2 : 2);
950 }
952 // Check for float store
953 if( src_first_rc == rc_float && dst_first_rc == rc_stack ) {
954 return impl_fp_store_helper(cbuf,do_size,src_first,src_second,dst_first,dst_second,ra_->reg2offset(dst_first),size);
955 }
957 // Check for float load
958 if( dst_first_rc == rc_float && src_first_rc == rc_stack ) {
959 int offset = ra_->reg2offset(src_first);
960 const char *op_str;
961 int op;
962 if( src_first+1 == src_second && dst_first+1 == dst_second ) { // double load?
963 op_str = "FLD_D";
964 op = 0xDD;
965 } else { // 32-bit load
966 op_str = "FLD_S";
967 op = 0xD9;
968 assert( src_second_rc == rc_bad && dst_second_rc == rc_bad, "no non-adjacent float-loads" );
969 }
970 if( cbuf ) {
971 emit_opcode (*cbuf, op );
972 encode_RegMem(*cbuf, 0x0, ESP_enc, 0x4, 0, offset, false);
973 emit_opcode (*cbuf, 0xDD ); // FSTP ST(i)
974 emit_d8 (*cbuf, 0xD8+Matcher::_regEncode[dst_first] );
975 #ifndef PRODUCT
976 } else if( !do_size ) {
977 if( size != 0 ) st->print("\n\t");
978 st->print("%s ST,[ESP + #%d]\n\tFSTP %s",op_str, offset,Matcher::regName[dst_first]);
979 #endif
980 }
981 int offset_size = (offset == 0) ? 0 : ((offset <= 127) ? 1 : 4);
982 return size + 3+offset_size+2;
983 }
985 // Check for xmm reg-reg copy
986 if( src_first_rc == rc_xmm && dst_first_rc == rc_xmm ) {
987 assert( (src_second_rc == rc_bad && dst_second_rc == rc_bad) ||
988 (src_first+1 == src_second && dst_first+1 == dst_second),
989 "no non-adjacent float-moves" );
990 return impl_movx_helper(cbuf,do_size,src_first,dst_first,src_second, dst_second, size);
991 }
993 // Check for xmm store
994 if( src_first_rc == rc_xmm && dst_first_rc == rc_stack ) {
995 return impl_x_helper(cbuf,do_size,false,ra_->reg2offset(dst_first),src_first, src_second, size);
996 }
998 // Check for float xmm load
999 if( dst_first_rc == rc_xmm && src_first_rc == rc_stack ) {
1000 return impl_x_helper(cbuf,do_size,true ,ra_->reg2offset(src_first),dst_first, dst_second, size);
1001 }
1003 // Copy from float reg to xmm reg
1004 if( dst_first_rc == rc_xmm && src_first_rc == rc_float ) {
1005 // copy to the top of stack from floating point reg
1006 // and use LEA to preserve flags
1007 if( cbuf ) {
1008 emit_opcode(*cbuf,0x8D); // LEA ESP,[ESP-8]
1009 emit_rm(*cbuf, 0x1, ESP_enc, 0x04);
1010 emit_rm(*cbuf, 0x0, 0x04, ESP_enc);
1011 emit_d8(*cbuf,0xF8);
1012 #ifndef PRODUCT
1013 } else if( !do_size ) {
1014 if( size != 0 ) st->print("\n\t");
1015 st->print("LEA ESP,[ESP-8]");
1016 #endif
1017 }
1018 size += 4;
1020 size = impl_fp_store_helper(cbuf,do_size,src_first,src_second,dst_first,dst_second,0,size);
1022 // Copy from the temp memory to the xmm reg.
1023 size = impl_x_helper(cbuf,do_size,true ,0,dst_first, dst_second, size);
1025 if( cbuf ) {
1026 emit_opcode(*cbuf,0x8D); // LEA ESP,[ESP+8]
1027 emit_rm(*cbuf, 0x1, ESP_enc, 0x04);
1028 emit_rm(*cbuf, 0x0, 0x04, ESP_enc);
1029 emit_d8(*cbuf,0x08);
1030 #ifndef PRODUCT
1031 } else if( !do_size ) {
1032 if( size != 0 ) st->print("\n\t");
1033 st->print("LEA ESP,[ESP+8]");
1034 #endif
1035 }
1036 size += 4;
1037 return size;
1038 }
1040 assert( size > 0, "missed a case" );
1042 // --------------------------------------------------------------------
1043 // Check for second bits still needing moving.
1044 if( src_second == dst_second )
1045 return size; // Self copy; no move
1046 assert( src_second_rc != rc_bad && dst_second_rc != rc_bad, "src_second & dst_second cannot be Bad" );
1048 // Check for second word int-int move
1049 if( src_second_rc == rc_int && dst_second_rc == rc_int )
1050 return impl_mov_helper(cbuf,do_size,src_second,dst_second,size);
1052 // Check for second word integer store
1053 if( src_second_rc == rc_int && dst_second_rc == rc_stack )
1054 return impl_helper(cbuf,do_size,false,ra_->reg2offset(dst_second),src_second,0x89,"MOV ",size);
1056 // Check for second word integer load
1057 if( dst_second_rc == rc_int && src_second_rc == rc_stack )
1058 return impl_helper(cbuf,do_size,true ,ra_->reg2offset(src_second),dst_second,0x8B,"MOV ",size);
1061 Unimplemented();
1062 }
1064 #ifndef PRODUCT
1065 void MachSpillCopyNode::format( PhaseRegAlloc *ra_, outputStream* st ) const {
1066 implementation( NULL, ra_, false, st );
1067 }
1068 #endif
1070 void MachSpillCopyNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
1071 implementation( &cbuf, ra_, false, NULL );
1072 }
1074 uint MachSpillCopyNode::size(PhaseRegAlloc *ra_) const {
1075 return implementation( NULL, ra_, true, NULL );
1076 }
1078 //=============================================================================
1079 #ifndef PRODUCT
1080 void MachNopNode::format( PhaseRegAlloc *, outputStream* st ) const {
1081 st->print("NOP \t# %d bytes pad for loops and calls", _count);
1082 }
1083 #endif
1085 void MachNopNode::emit(CodeBuffer &cbuf, PhaseRegAlloc * ) const {
1086 MacroAssembler _masm(&cbuf);
1087 __ nop(_count);
1088 }
1090 uint MachNopNode::size(PhaseRegAlloc *) const {
1091 return _count;
1092 }
1095 //=============================================================================
1096 #ifndef PRODUCT
1097 void BoxLockNode::format( PhaseRegAlloc *ra_, outputStream* st ) const {
1098 int offset = ra_->reg2offset(in_RegMask(0).find_first_elem());
1099 int reg = ra_->get_reg_first(this);
1100 st->print("LEA %s,[ESP + #%d]",Matcher::regName[reg],offset);
1101 }
1102 #endif
1104 void BoxLockNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
1105 int offset = ra_->reg2offset(in_RegMask(0).find_first_elem());
1106 int reg = ra_->get_encode(this);
1107 if( offset >= 128 ) {
1108 emit_opcode(cbuf, 0x8D); // LEA reg,[SP+offset]
1109 emit_rm(cbuf, 0x2, reg, 0x04);
1110 emit_rm(cbuf, 0x0, 0x04, ESP_enc);
1111 emit_d32(cbuf, offset);
1112 }
1113 else {
1114 emit_opcode(cbuf, 0x8D); // LEA reg,[SP+offset]
1115 emit_rm(cbuf, 0x1, reg, 0x04);
1116 emit_rm(cbuf, 0x0, 0x04, ESP_enc);
1117 emit_d8(cbuf, offset);
1118 }
1119 }
1121 uint BoxLockNode::size(PhaseRegAlloc *ra_) const {
1122 int offset = ra_->reg2offset(in_RegMask(0).find_first_elem());
1123 if( offset >= 128 ) {
1124 return 7;
1125 }
1126 else {
1127 return 4;
1128 }
1129 }
1131 //=============================================================================
1133 // emit call stub, compiled java to interpreter
1134 void emit_java_to_interp(CodeBuffer &cbuf ) {
1135 // Stub is fixed up when the corresponding call is converted from calling
1136 // compiled code to calling interpreted code.
1137 // mov rbx,0
1138 // jmp -1
1140 address mark = cbuf.inst_mark(); // get mark within main instrs section
1142 // Note that the code buffer's inst_mark is always relative to insts.
1143 // That's why we must use the macroassembler to generate a stub.
1144 MacroAssembler _masm(&cbuf);
1146 address base =
1147 __ start_a_stub(Compile::MAX_stubs_size);
1148 if (base == NULL) return; // CodeBuffer::expand failed
1149 // static stub relocation stores the instruction address of the call
1150 __ relocate(static_stub_Relocation::spec(mark), RELOC_IMM32);
1151 // static stub relocation also tags the methodOop in the code-stream.
1152 __ movoop(rbx, (jobject)NULL); // method is zapped till fixup time
1153 __ jump(RuntimeAddress((address)-1));
1155 __ end_a_stub();
1156 // Update current stubs pointer and restore code_end.
1157 }
1158 // size of call stub, compiled java to interpretor
1159 uint size_java_to_interp() {
1160 return 10; // movl; jmp
1161 }
1162 // relocation entries for call stub, compiled java to interpretor
1163 uint reloc_java_to_interp() {
1164 return 4; // 3 in emit_java_to_interp + 1 in Java_Static_Call
1165 }
1167 //=============================================================================
1168 #ifndef PRODUCT
1169 void MachUEPNode::format( PhaseRegAlloc *ra_, outputStream* st ) const {
1170 st->print_cr( "CMP EAX,[ECX+4]\t# Inline cache check");
1171 st->print_cr("\tJNE SharedRuntime::handle_ic_miss_stub");
1172 st->print_cr("\tNOP");
1173 st->print_cr("\tNOP");
1174 if( !OptoBreakpoint )
1175 st->print_cr("\tNOP");
1176 }
1177 #endif
1179 void MachUEPNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
1180 MacroAssembler masm(&cbuf);
1181 #ifdef ASSERT
1182 uint code_size = cbuf.code_size();
1183 #endif
1184 masm.cmpl(rax, Address(rcx, oopDesc::klass_offset_in_bytes()));
1185 masm.jump_cc(Assembler::notEqual,
1186 RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1187 /* WARNING these NOPs are critical so that verified entry point is properly
1188 aligned for patching by NativeJump::patch_verified_entry() */
1189 int nops_cnt = 2;
1190 if( !OptoBreakpoint ) // Leave space for int3
1191 nops_cnt += 1;
1192 masm.nop(nops_cnt);
1194 assert(cbuf.code_size() - code_size == size(ra_), "checking code size of inline cache node");
1195 }
1197 uint MachUEPNode::size(PhaseRegAlloc *ra_) const {
1198 return OptoBreakpoint ? 11 : 12;
1199 }
1202 //=============================================================================
1203 uint size_exception_handler() {
1204 // NativeCall instruction size is the same as NativeJump.
1205 // exception handler starts out as jump and can be patched to
1206 // a call be deoptimization. (4932387)
1207 // Note that this value is also credited (in output.cpp) to
1208 // the size of the code section.
1209 return NativeJump::instruction_size;
1210 }
1212 // Emit exception handler code. Stuff framesize into a register
1213 // and call a VM stub routine.
1214 int emit_exception_handler(CodeBuffer& cbuf) {
1216 // Note that the code buffer's inst_mark is always relative to insts.
1217 // That's why we must use the macroassembler to generate a handler.
1218 MacroAssembler _masm(&cbuf);
1219 address base =
1220 __ start_a_stub(size_exception_handler());
1221 if (base == NULL) return 0; // CodeBuffer::expand failed
1222 int offset = __ offset();
1223 __ jump(RuntimeAddress(OptoRuntime::exception_blob()->instructions_begin()));
1224 assert(__ offset() - offset <= (int) size_exception_handler(), "overflow");
1225 __ end_a_stub();
1226 return offset;
1227 }
1229 uint size_deopt_handler() {
1230 // NativeCall instruction size is the same as NativeJump.
1231 // exception handler starts out as jump and can be patched to
1232 // a call be deoptimization. (4932387)
1233 // Note that this value is also credited (in output.cpp) to
1234 // the size of the code section.
1235 return 5 + NativeJump::instruction_size; // pushl(); jmp;
1236 }
1238 // Emit deopt handler code.
1239 int emit_deopt_handler(CodeBuffer& cbuf) {
1241 // Note that the code buffer's inst_mark is always relative to insts.
1242 // That's why we must use the macroassembler to generate a handler.
1243 MacroAssembler _masm(&cbuf);
1244 address base =
1245 __ start_a_stub(size_exception_handler());
1246 if (base == NULL) return 0; // CodeBuffer::expand failed
1247 int offset = __ offset();
1248 InternalAddress here(__ pc());
1249 __ pushptr(here.addr());
1251 __ jump(RuntimeAddress(SharedRuntime::deopt_blob()->unpack()));
1252 assert(__ offset() - offset <= (int) size_deopt_handler(), "overflow");
1253 __ end_a_stub();
1254 return offset;
1255 }
1258 static void emit_double_constant(CodeBuffer& cbuf, double x) {
1259 int mark = cbuf.insts()->mark_off();
1260 MacroAssembler _masm(&cbuf);
1261 address double_address = __ double_constant(x);
1262 cbuf.insts()->set_mark_off(mark); // preserve mark across masm shift
1263 emit_d32_reloc(cbuf,
1264 (int)double_address,
1265 internal_word_Relocation::spec(double_address),
1266 RELOC_DISP32);
1267 }
1269 static void emit_float_constant(CodeBuffer& cbuf, float x) {
1270 int mark = cbuf.insts()->mark_off();
1271 MacroAssembler _masm(&cbuf);
1272 address float_address = __ float_constant(x);
1273 cbuf.insts()->set_mark_off(mark); // preserve mark across masm shift
1274 emit_d32_reloc(cbuf,
1275 (int)float_address,
1276 internal_word_Relocation::spec(float_address),
1277 RELOC_DISP32);
1278 }
1281 int Matcher::regnum_to_fpu_offset(int regnum) {
1282 return regnum - 32; // The FP registers are in the second chunk
1283 }
1285 bool is_positive_zero_float(jfloat f) {
1286 return jint_cast(f) == jint_cast(0.0F);
1287 }
1289 bool is_positive_one_float(jfloat f) {
1290 return jint_cast(f) == jint_cast(1.0F);
1291 }
1293 bool is_positive_zero_double(jdouble d) {
1294 return jlong_cast(d) == jlong_cast(0.0);
1295 }
1297 bool is_positive_one_double(jdouble d) {
1298 return jlong_cast(d) == jlong_cast(1.0);
1299 }
1301 // This is UltraSparc specific, true just means we have fast l2f conversion
1302 const bool Matcher::convL2FSupported(void) {
1303 return true;
1304 }
1306 // Vector width in bytes
1307 const uint Matcher::vector_width_in_bytes(void) {
1308 return UseSSE >= 2 ? 8 : 0;
1309 }
1311 // Vector ideal reg
1312 const uint Matcher::vector_ideal_reg(void) {
1313 return Op_RegD;
1314 }
1316 // Is this branch offset short enough that a short branch can be used?
1317 //
1318 // NOTE: If the platform does not provide any short branch variants, then
1319 // this method should return false for offset 0.
1320 bool Matcher::is_short_branch_offset(int offset) {
1321 return (-128 <= offset && offset <= 127);
1322 }
1324 const bool Matcher::isSimpleConstant64(jlong value) {
1325 // Will one (StoreL ConL) be cheaper than two (StoreI ConI)?.
1326 return false;
1327 }
1329 // The ecx parameter to rep stos for the ClearArray node is in dwords.
1330 const bool Matcher::init_array_count_is_in_bytes = false;
1332 // Threshold size for cleararray.
1333 const int Matcher::init_array_short_size = 8 * BytesPerLong;
1335 // Should the Matcher clone shifts on addressing modes, expecting them to
1336 // be subsumed into complex addressing expressions or compute them into
1337 // registers? True for Intel but false for most RISCs
1338 const bool Matcher::clone_shift_expressions = true;
1340 // Is it better to copy float constants, or load them directly from memory?
1341 // Intel can load a float constant from a direct address, requiring no
1342 // extra registers. Most RISCs will have to materialize an address into a
1343 // register first, so they would do better to copy the constant from stack.
1344 const bool Matcher::rematerialize_float_constants = true;
1346 // If CPU can load and store mis-aligned doubles directly then no fixup is
1347 // needed. Else we split the double into 2 integer pieces and move it
1348 // piece-by-piece. Only happens when passing doubles into C code as the
1349 // Java calling convention forces doubles to be aligned.
1350 const bool Matcher::misaligned_doubles_ok = true;
1353 void Matcher::pd_implicit_null_fixup(MachNode *node, uint idx) {
1354 // Get the memory operand from the node
1355 uint numopnds = node->num_opnds(); // Virtual call for number of operands
1356 uint skipped = node->oper_input_base(); // Sum of leaves skipped so far
1357 assert( idx >= skipped, "idx too low in pd_implicit_null_fixup" );
1358 uint opcnt = 1; // First operand
1359 uint num_edges = node->_opnds[1]->num_edges(); // leaves for first operand
1360 while( idx >= skipped+num_edges ) {
1361 skipped += num_edges;
1362 opcnt++; // Bump operand count
1363 assert( opcnt < numopnds, "Accessing non-existent operand" );
1364 num_edges = node->_opnds[opcnt]->num_edges(); // leaves for next operand
1365 }
1367 MachOper *memory = node->_opnds[opcnt];
1368 MachOper *new_memory = NULL;
1369 switch (memory->opcode()) {
1370 case DIRECT:
1371 case INDOFFSET32X:
1372 // No transformation necessary.
1373 return;
1374 case INDIRECT:
1375 new_memory = new (C) indirect_win95_safeOper( );
1376 break;
1377 case INDOFFSET8:
1378 new_memory = new (C) indOffset8_win95_safeOper(memory->disp(NULL, NULL, 0));
1379 break;
1380 case INDOFFSET32:
1381 new_memory = new (C) indOffset32_win95_safeOper(memory->disp(NULL, NULL, 0));
1382 break;
1383 case INDINDEXOFFSET:
1384 new_memory = new (C) indIndexOffset_win95_safeOper(memory->disp(NULL, NULL, 0));
1385 break;
1386 case INDINDEXSCALE:
1387 new_memory = new (C) indIndexScale_win95_safeOper(memory->scale());
1388 break;
1389 case INDINDEXSCALEOFFSET:
1390 new_memory = new (C) indIndexScaleOffset_win95_safeOper(memory->scale(), memory->disp(NULL, NULL, 0));
1391 break;
1392 case LOAD_LONG_INDIRECT:
1393 case LOAD_LONG_INDOFFSET32:
1394 // Does not use EBP as address register, use { EDX, EBX, EDI, ESI}
1395 return;
1396 default:
1397 assert(false, "unexpected memory operand in pd_implicit_null_fixup()");
1398 return;
1399 }
1400 node->_opnds[opcnt] = new_memory;
1401 }
1403 // Advertise here if the CPU requires explicit rounding operations
1404 // to implement the UseStrictFP mode.
1405 const bool Matcher::strict_fp_requires_explicit_rounding = true;
1407 // Do floats take an entire double register or just half?
1408 const bool Matcher::float_in_double = true;
1409 // Do ints take an entire long register or just half?
1410 const bool Matcher::int_in_long = false;
1412 // Return whether or not this register is ever used as an argument. This
1413 // function is used on startup to build the trampoline stubs in generateOptoStub.
1414 // Registers not mentioned will be killed by the VM call in the trampoline, and
1415 // arguments in those registers not be available to the callee.
1416 bool Matcher::can_be_java_arg( int reg ) {
1417 if( reg == ECX_num || reg == EDX_num ) return true;
1418 if( (reg == XMM0a_num || reg == XMM1a_num) && UseSSE>=1 ) return true;
1419 if( (reg == XMM0b_num || reg == XMM1b_num) && UseSSE>=2 ) return true;
1420 return false;
1421 }
1423 bool Matcher::is_spillable_arg( int reg ) {
1424 return can_be_java_arg(reg);
1425 }
1427 // Register for DIVI projection of divmodI
1428 RegMask Matcher::divI_proj_mask() {
1429 return EAX_REG_mask;
1430 }
1432 // Register for MODI projection of divmodI
1433 RegMask Matcher::modI_proj_mask() {
1434 return EDX_REG_mask;
1435 }
1437 // Register for DIVL projection of divmodL
1438 RegMask Matcher::divL_proj_mask() {
1439 ShouldNotReachHere();
1440 return RegMask();
1441 }
1443 // Register for MODL projection of divmodL
1444 RegMask Matcher::modL_proj_mask() {
1445 ShouldNotReachHere();
1446 return RegMask();
1447 }
1449 %}
1451 //----------ENCODING BLOCK-----------------------------------------------------
1452 // This block specifies the encoding classes used by the compiler to output
1453 // byte streams. Encoding classes generate functions which are called by
1454 // Machine Instruction Nodes in order to generate the bit encoding of the
1455 // instruction. Operands specify their base encoding interface with the
1456 // interface keyword. There are currently supported four interfaces,
1457 // REG_INTER, CONST_INTER, MEMORY_INTER, & COND_INTER. REG_INTER causes an
1458 // operand to generate a function which returns its register number when
1459 // queried. CONST_INTER causes an operand to generate a function which
1460 // returns the value of the constant when queried. MEMORY_INTER causes an
1461 // operand to generate four functions which return the Base Register, the
1462 // Index Register, the Scale Value, and the Offset Value of the operand when
1463 // queried. COND_INTER causes an operand to generate six functions which
1464 // return the encoding code (ie - encoding bits for the instruction)
1465 // associated with each basic boolean condition for a conditional instruction.
1466 // Instructions specify two basic values for encoding. They use the
1467 // ins_encode keyword to specify their encoding class (which must be one of
1468 // the class names specified in the encoding block), and they use the
1469 // opcode keyword to specify, in order, their primary, secondary, and
1470 // tertiary opcode. Only the opcode sections which a particular instruction
1471 // needs for encoding need to be specified.
1472 encode %{
1473 // Build emit functions for each basic byte or larger field in the intel
1474 // encoding scheme (opcode, rm, sib, immediate), and call them from C++
1475 // code in the enc_class source block. Emit functions will live in the
1476 // main source block for now. In future, we can generalize this by
1477 // adding a syntax that specifies the sizes of fields in an order,
1478 // so that the adlc can build the emit functions automagically
1479 enc_class OpcP %{ // Emit opcode
1480 emit_opcode(cbuf,$primary);
1481 %}
1483 enc_class OpcS %{ // Emit opcode
1484 emit_opcode(cbuf,$secondary);
1485 %}
1487 enc_class Opcode(immI d8 ) %{ // Emit opcode
1488 emit_opcode(cbuf,$d8$$constant);
1489 %}
1491 enc_class SizePrefix %{
1492 emit_opcode(cbuf,0x66);
1493 %}
1495 enc_class RegReg (eRegI dst, eRegI src) %{ // RegReg(Many)
1496 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
1497 %}
1499 enc_class OpcRegReg (immI opcode, eRegI dst, eRegI src) %{ // OpcRegReg(Many)
1500 emit_opcode(cbuf,$opcode$$constant);
1501 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
1502 %}
1504 enc_class mov_r32_imm0( eRegI dst ) %{
1505 emit_opcode( cbuf, 0xB8 + $dst$$reg ); // 0xB8+ rd -- MOV r32 ,imm32
1506 emit_d32 ( cbuf, 0x0 ); // imm32==0x0
1507 %}
1509 enc_class cdq_enc %{
1510 // Full implementation of Java idiv and irem; checks for
1511 // special case as described in JVM spec., p.243 & p.271.
1512 //
1513 // normal case special case
1514 //
1515 // input : rax,: dividend min_int
1516 // reg: divisor -1
1517 //
1518 // output: rax,: quotient (= rax, idiv reg) min_int
1519 // rdx: remainder (= rax, irem reg) 0
1520 //
1521 // Code sequnce:
1522 //
1523 // 81 F8 00 00 00 80 cmp rax,80000000h
1524 // 0F 85 0B 00 00 00 jne normal_case
1525 // 33 D2 xor rdx,edx
1526 // 83 F9 FF cmp rcx,0FFh
1527 // 0F 84 03 00 00 00 je done
1528 // normal_case:
1529 // 99 cdq
1530 // F7 F9 idiv rax,ecx
1531 // done:
1532 //
1533 emit_opcode(cbuf,0x81); emit_d8(cbuf,0xF8);
1534 emit_opcode(cbuf,0x00); emit_d8(cbuf,0x00);
1535 emit_opcode(cbuf,0x00); emit_d8(cbuf,0x80); // cmp rax,80000000h
1536 emit_opcode(cbuf,0x0F); emit_d8(cbuf,0x85);
1537 emit_opcode(cbuf,0x0B); emit_d8(cbuf,0x00);
1538 emit_opcode(cbuf,0x00); emit_d8(cbuf,0x00); // jne normal_case
1539 emit_opcode(cbuf,0x33); emit_d8(cbuf,0xD2); // xor rdx,edx
1540 emit_opcode(cbuf,0x83); emit_d8(cbuf,0xF9); emit_d8(cbuf,0xFF); // cmp rcx,0FFh
1541 emit_opcode(cbuf,0x0F); emit_d8(cbuf,0x84);
1542 emit_opcode(cbuf,0x03); emit_d8(cbuf,0x00);
1543 emit_opcode(cbuf,0x00); emit_d8(cbuf,0x00); // je done
1544 // normal_case:
1545 emit_opcode(cbuf,0x99); // cdq
1546 // idiv (note: must be emitted by the user of this rule)
1547 // normal:
1548 %}
1550 // Dense encoding for older common ops
1551 enc_class Opc_plus(immI opcode, eRegI reg) %{
1552 emit_opcode(cbuf, $opcode$$constant + $reg$$reg);
1553 %}
1556 // Opcde enc_class for 8/32 bit immediate instructions with sign-extension
1557 enc_class OpcSE (immI imm) %{ // Emit primary opcode and set sign-extend bit
1558 // Check for 8-bit immediate, and set sign extend bit in opcode
1559 if (($imm$$constant >= -128) && ($imm$$constant <= 127)) {
1560 emit_opcode(cbuf, $primary | 0x02);
1561 }
1562 else { // If 32-bit immediate
1563 emit_opcode(cbuf, $primary);
1564 }
1565 %}
1567 enc_class OpcSErm (eRegI dst, immI imm) %{ // OpcSEr/m
1568 // Emit primary opcode and set sign-extend bit
1569 // Check for 8-bit immediate, and set sign extend bit in opcode
1570 if (($imm$$constant >= -128) && ($imm$$constant <= 127)) {
1571 emit_opcode(cbuf, $primary | 0x02); }
1572 else { // If 32-bit immediate
1573 emit_opcode(cbuf, $primary);
1574 }
1575 // Emit r/m byte with secondary opcode, after primary opcode.
1576 emit_rm(cbuf, 0x3, $secondary, $dst$$reg);
1577 %}
1579 enc_class Con8or32 (immI imm) %{ // Con8or32(storeImmI), 8 or 32 bits
1580 // Check for 8-bit immediate, and set sign extend bit in opcode
1581 if (($imm$$constant >= -128) && ($imm$$constant <= 127)) {
1582 $$$emit8$imm$$constant;
1583 }
1584 else { // If 32-bit immediate
1585 // Output immediate
1586 $$$emit32$imm$$constant;
1587 }
1588 %}
1590 enc_class Long_OpcSErm_Lo(eRegL dst, immL imm) %{
1591 // Emit primary opcode and set sign-extend bit
1592 // Check for 8-bit immediate, and set sign extend bit in opcode
1593 int con = (int)$imm$$constant; // Throw away top bits
1594 emit_opcode(cbuf, ((con >= -128) && (con <= 127)) ? ($primary | 0x02) : $primary);
1595 // Emit r/m byte with secondary opcode, after primary opcode.
1596 emit_rm(cbuf, 0x3, $secondary, $dst$$reg);
1597 if ((con >= -128) && (con <= 127)) emit_d8 (cbuf,con);
1598 else emit_d32(cbuf,con);
1599 %}
1601 enc_class Long_OpcSErm_Hi(eRegL dst, immL imm) %{
1602 // Emit primary opcode and set sign-extend bit
1603 // Check for 8-bit immediate, and set sign extend bit in opcode
1604 int con = (int)($imm$$constant >> 32); // Throw away bottom bits
1605 emit_opcode(cbuf, ((con >= -128) && (con <= 127)) ? ($primary | 0x02) : $primary);
1606 // Emit r/m byte with tertiary opcode, after primary opcode.
1607 emit_rm(cbuf, 0x3, $tertiary, HIGH_FROM_LOW($dst$$reg));
1608 if ((con >= -128) && (con <= 127)) emit_d8 (cbuf,con);
1609 else emit_d32(cbuf,con);
1610 %}
1612 enc_class Lbl (label labl) %{ // JMP, CALL
1613 Label *l = $labl$$label;
1614 emit_d32(cbuf, l ? (l->loc_pos() - (cbuf.code_size()+4)) : 0);
1615 %}
1617 enc_class LblShort (label labl) %{ // JMP, CALL
1618 Label *l = $labl$$label;
1619 int disp = l ? (l->loc_pos() - (cbuf.code_size()+1)) : 0;
1620 assert(-128 <= disp && disp <= 127, "Displacement too large for short jmp");
1621 emit_d8(cbuf, disp);
1622 %}
1624 enc_class OpcSReg (eRegI dst) %{ // BSWAP
1625 emit_cc(cbuf, $secondary, $dst$$reg );
1626 %}
1628 enc_class bswap_long_bytes(eRegL dst) %{ // BSWAP
1629 int destlo = $dst$$reg;
1630 int desthi = HIGH_FROM_LOW(destlo);
1631 // bswap lo
1632 emit_opcode(cbuf, 0x0F);
1633 emit_cc(cbuf, 0xC8, destlo);
1634 // bswap hi
1635 emit_opcode(cbuf, 0x0F);
1636 emit_cc(cbuf, 0xC8, desthi);
1637 // xchg lo and hi
1638 emit_opcode(cbuf, 0x87);
1639 emit_rm(cbuf, 0x3, destlo, desthi);
1640 %}
1642 enc_class RegOpc (eRegI div) %{ // IDIV, IMOD, JMP indirect, ...
1643 emit_rm(cbuf, 0x3, $secondary, $div$$reg );
1644 %}
1646 enc_class Jcc (cmpOp cop, label labl) %{ // JCC
1647 Label *l = $labl$$label;
1648 $$$emit8$primary;
1649 emit_cc(cbuf, $secondary, $cop$$cmpcode);
1650 emit_d32(cbuf, l ? (l->loc_pos() - (cbuf.code_size()+4)) : 0);
1651 %}
1653 enc_class JccShort (cmpOp cop, label labl) %{ // JCC
1654 Label *l = $labl$$label;
1655 emit_cc(cbuf, $primary, $cop$$cmpcode);
1656 int disp = l ? (l->loc_pos() - (cbuf.code_size()+1)) : 0;
1657 assert(-128 <= disp && disp <= 127, "Displacement too large for short jmp");
1658 emit_d8(cbuf, disp);
1659 %}
1661 enc_class enc_cmov(cmpOp cop ) %{ // CMOV
1662 $$$emit8$primary;
1663 emit_cc(cbuf, $secondary, $cop$$cmpcode);
1664 %}
1666 enc_class enc_cmov_d(cmpOp cop, regD src ) %{ // CMOV
1667 int op = 0xDA00 + $cop$$cmpcode + ($src$$reg-1);
1668 emit_d8(cbuf, op >> 8 );
1669 emit_d8(cbuf, op & 255);
1670 %}
1672 // emulate a CMOV with a conditional branch around a MOV
1673 enc_class enc_cmov_branch( cmpOp cop, immI brOffs ) %{ // CMOV
1674 // Invert sense of branch from sense of CMOV
1675 emit_cc( cbuf, 0x70, ($cop$$cmpcode^1) );
1676 emit_d8( cbuf, $brOffs$$constant );
1677 %}
1679 enc_class enc_PartialSubtypeCheck( ) %{
1680 Register Redi = as_Register(EDI_enc); // result register
1681 Register Reax = as_Register(EAX_enc); // super class
1682 Register Recx = as_Register(ECX_enc); // killed
1683 Register Resi = as_Register(ESI_enc); // sub class
1684 Label hit, miss;
1686 MacroAssembler _masm(&cbuf);
1687 // Compare super with sub directly, since super is not in its own SSA.
1688 // The compiler used to emit this test, but we fold it in here,
1689 // to allow platform-specific tweaking on sparc.
1690 __ cmpl(Reax, Resi);
1691 __ jcc(Assembler::equal, hit);
1692 #ifndef PRODUCT
1693 __ increment(ExternalAddress((address)&SharedRuntime::_partial_subtype_ctr));
1694 #endif //PRODUCT
1695 __ movl(Redi,Address(Resi,sizeof(oopDesc) + Klass::secondary_supers_offset_in_bytes()));
1696 __ movl(Recx,Address(Redi,arrayOopDesc::length_offset_in_bytes()));
1697 __ addl(Redi,arrayOopDesc::base_offset_in_bytes(T_OBJECT));
1698 __ repne_scan();
1699 __ jcc(Assembler::notEqual, miss);
1700 __ movl(Address(Resi,sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes()),Reax);
1701 __ bind(hit);
1702 if( $primary )
1703 __ xorl(Redi,Redi);
1704 __ bind(miss);
1705 %}
1707 enc_class FFree_Float_Stack_All %{ // Free_Float_Stack_All
1708 MacroAssembler masm(&cbuf);
1709 int start = masm.offset();
1710 if (UseSSE >= 2) {
1711 if (VerifyFPU) {
1712 masm.verify_FPU(0, "must be empty in SSE2+ mode");
1713 }
1714 } else {
1715 // External c_calling_convention expects the FPU stack to be 'clean'.
1716 // Compiled code leaves it dirty. Do cleanup now.
1717 masm.empty_FPU_stack();
1718 }
1719 if (sizeof_FFree_Float_Stack_All == -1) {
1720 sizeof_FFree_Float_Stack_All = masm.offset() - start;
1721 } else {
1722 assert(masm.offset() - start == sizeof_FFree_Float_Stack_All, "wrong size");
1723 }
1724 %}
1726 enc_class Verify_FPU_For_Leaf %{
1727 if( VerifyFPU ) {
1728 MacroAssembler masm(&cbuf);
1729 masm.verify_FPU( -3, "Returning from Runtime Leaf call");
1730 }
1731 %}
1733 enc_class Java_To_Runtime (method meth) %{ // CALL Java_To_Runtime, Java_To_Runtime_Leaf
1734 // This is the instruction starting address for relocation info.
1735 cbuf.set_inst_mark();
1736 $$$emit8$primary;
1737 // CALL directly to the runtime
1738 emit_d32_reloc(cbuf, ($meth$$method - (int)(cbuf.code_end()) - 4),
1739 runtime_call_Relocation::spec(), RELOC_IMM32 );
1741 if (UseSSE >= 2) {
1742 MacroAssembler _masm(&cbuf);
1743 BasicType rt = tf()->return_type();
1745 if ((rt == T_FLOAT || rt == T_DOUBLE) && !return_value_is_used()) {
1746 // A C runtime call where the return value is unused. In SSE2+
1747 // mode the result needs to be removed from the FPU stack. It's
1748 // likely that this function call could be removed by the
1749 // optimizer if the C function is a pure function.
1750 __ ffree(0);
1751 } else if (rt == T_FLOAT) {
1752 __ leal(rsp, Address(rsp, -4));
1753 __ fstp_s(Address(rsp, 0));
1754 __ movflt(xmm0, Address(rsp, 0));
1755 __ leal(rsp, Address(rsp, 4));
1756 } else if (rt == T_DOUBLE) {
1757 __ leal(rsp, Address(rsp, -8));
1758 __ fstp_d(Address(rsp, 0));
1759 __ movdbl(xmm0, Address(rsp, 0));
1760 __ leal(rsp, Address(rsp, 8));
1761 }
1762 }
1763 %}
1766 enc_class pre_call_FPU %{
1767 // If method sets FPU control word restore it here
1768 if( Compile::current()->in_24_bit_fp_mode() ) {
1769 MacroAssembler masm(&cbuf);
1770 masm.fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
1771 }
1772 %}
1774 enc_class post_call_FPU %{
1775 // If method sets FPU control word do it here also
1776 if( Compile::current()->in_24_bit_fp_mode() ) {
1777 MacroAssembler masm(&cbuf);
1778 masm.fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
1779 }
1780 %}
1782 enc_class Java_Static_Call (method meth) %{ // JAVA STATIC CALL
1783 // CALL to fixup routine. Fixup routine uses ScopeDesc info to determine
1784 // who we intended to call.
1785 cbuf.set_inst_mark();
1786 $$$emit8$primary;
1787 if ( !_method ) {
1788 emit_d32_reloc(cbuf, ($meth$$method - (int)(cbuf.code_end()) - 4),
1789 runtime_call_Relocation::spec(), RELOC_IMM32 );
1790 } else if(_optimized_virtual) {
1791 emit_d32_reloc(cbuf, ($meth$$method - (int)(cbuf.code_end()) - 4),
1792 opt_virtual_call_Relocation::spec(), RELOC_IMM32 );
1793 } else {
1794 emit_d32_reloc(cbuf, ($meth$$method - (int)(cbuf.code_end()) - 4),
1795 static_call_Relocation::spec(), RELOC_IMM32 );
1796 }
1797 if( _method ) { // Emit stub for static call
1798 emit_java_to_interp(cbuf);
1799 }
1800 %}
1802 enc_class Java_Dynamic_Call (method meth) %{ // JAVA DYNAMIC CALL
1803 // !!!!!
1804 // Generate "Mov EAX,0x00", placeholder instruction to load oop-info
1805 // emit_call_dynamic_prologue( cbuf );
1806 cbuf.set_inst_mark();
1807 emit_opcode(cbuf, 0xB8 + EAX_enc); // mov EAX,-1
1808 emit_d32_reloc(cbuf, (int)Universe::non_oop_word(), oop_Relocation::spec_for_immediate(), RELOC_IMM32);
1809 address virtual_call_oop_addr = cbuf.inst_mark();
1810 // CALL to fixup routine. Fixup routine uses ScopeDesc info to determine
1811 // who we intended to call.
1812 cbuf.set_inst_mark();
1813 $$$emit8$primary;
1814 emit_d32_reloc(cbuf, ($meth$$method - (int)(cbuf.code_end()) - 4),
1815 virtual_call_Relocation::spec(virtual_call_oop_addr), RELOC_IMM32 );
1816 %}
1818 enc_class Java_Compiled_Call (method meth) %{ // JAVA COMPILED CALL
1819 int disp = in_bytes(methodOopDesc::from_compiled_offset());
1820 assert( -128 <= disp && disp <= 127, "compiled_code_offset isn't small");
1822 // CALL *[EAX+in_bytes(methodOopDesc::from_compiled_code_entry_point_offset())]
1823 cbuf.set_inst_mark();
1824 $$$emit8$primary;
1825 emit_rm(cbuf, 0x01, $secondary, EAX_enc ); // R/M byte
1826 emit_d8(cbuf, disp); // Displacement
1828 %}
1830 enc_class Xor_Reg (eRegI dst) %{
1831 emit_opcode(cbuf, 0x33);
1832 emit_rm(cbuf, 0x3, $dst$$reg, $dst$$reg);
1833 %}
1835 // Following encoding is no longer used, but may be restored if calling
1836 // convention changes significantly.
1837 // Became: Xor_Reg(EBP), Java_To_Runtime( labl )
1838 //
1839 // enc_class Java_Interpreter_Call (label labl) %{ // JAVA INTERPRETER CALL
1840 // // int ic_reg = Matcher::inline_cache_reg();
1841 // // int ic_encode = Matcher::_regEncode[ic_reg];
1842 // // int imo_reg = Matcher::interpreter_method_oop_reg();
1843 // // int imo_encode = Matcher::_regEncode[imo_reg];
1844 //
1845 // // // Interpreter expects method_oop in EBX, currently a callee-saved register,
1846 // // // so we load it immediately before the call
1847 // // emit_opcode(cbuf, 0x8B); // MOV imo_reg,ic_reg # method_oop
1848 // // emit_rm(cbuf, 0x03, imo_encode, ic_encode ); // R/M byte
1849 //
1850 // // xor rbp,ebp
1851 // emit_opcode(cbuf, 0x33);
1852 // emit_rm(cbuf, 0x3, EBP_enc, EBP_enc);
1853 //
1854 // // CALL to interpreter.
1855 // cbuf.set_inst_mark();
1856 // $$$emit8$primary;
1857 // emit_d32_reloc(cbuf, ($labl$$label - (int)(cbuf.code_end()) - 4),
1858 // runtime_call_Relocation::spec(), RELOC_IMM32 );
1859 // %}
1861 enc_class RegOpcImm (eRegI dst, immI8 shift) %{ // SHL, SAR, SHR
1862 $$$emit8$primary;
1863 emit_rm(cbuf, 0x3, $secondary, $dst$$reg);
1864 $$$emit8$shift$$constant;
1865 %}
1867 enc_class LdImmI (eRegI dst, immI src) %{ // Load Immediate
1868 // Load immediate does not have a zero or sign extended version
1869 // for 8-bit immediates
1870 emit_opcode(cbuf, 0xB8 + $dst$$reg);
1871 $$$emit32$src$$constant;
1872 %}
1874 enc_class LdImmP (eRegI dst, immI src) %{ // Load Immediate
1875 // Load immediate does not have a zero or sign extended version
1876 // for 8-bit immediates
1877 emit_opcode(cbuf, $primary + $dst$$reg);
1878 $$$emit32$src$$constant;
1879 %}
1881 enc_class LdImmL_Lo( eRegL dst, immL src) %{ // Load Immediate
1882 // Load immediate does not have a zero or sign extended version
1883 // for 8-bit immediates
1884 int dst_enc = $dst$$reg;
1885 int src_con = $src$$constant & 0x0FFFFFFFFL;
1886 if (src_con == 0) {
1887 // xor dst, dst
1888 emit_opcode(cbuf, 0x33);
1889 emit_rm(cbuf, 0x3, dst_enc, dst_enc);
1890 } else {
1891 emit_opcode(cbuf, $primary + dst_enc);
1892 emit_d32(cbuf, src_con);
1893 }
1894 %}
1896 enc_class LdImmL_Hi( eRegL dst, immL src) %{ // Load Immediate
1897 // Load immediate does not have a zero or sign extended version
1898 // for 8-bit immediates
1899 int dst_enc = $dst$$reg + 2;
1900 int src_con = ((julong)($src$$constant)) >> 32;
1901 if (src_con == 0) {
1902 // xor dst, dst
1903 emit_opcode(cbuf, 0x33);
1904 emit_rm(cbuf, 0x3, dst_enc, dst_enc);
1905 } else {
1906 emit_opcode(cbuf, $primary + dst_enc);
1907 emit_d32(cbuf, src_con);
1908 }
1909 %}
1912 enc_class LdImmD (immD src) %{ // Load Immediate
1913 if( is_positive_zero_double($src$$constant)) {
1914 // FLDZ
1915 emit_opcode(cbuf,0xD9);
1916 emit_opcode(cbuf,0xEE);
1917 } else if( is_positive_one_double($src$$constant)) {
1918 // FLD1
1919 emit_opcode(cbuf,0xD9);
1920 emit_opcode(cbuf,0xE8);
1921 } else {
1922 emit_opcode(cbuf,0xDD);
1923 emit_rm(cbuf, 0x0, 0x0, 0x5);
1924 emit_double_constant(cbuf, $src$$constant);
1925 }
1926 %}
1929 enc_class LdImmF (immF src) %{ // Load Immediate
1930 if( is_positive_zero_float($src$$constant)) {
1931 emit_opcode(cbuf,0xD9);
1932 emit_opcode(cbuf,0xEE);
1933 } else if( is_positive_one_float($src$$constant)) {
1934 emit_opcode(cbuf,0xD9);
1935 emit_opcode(cbuf,0xE8);
1936 } else {
1937 $$$emit8$primary;
1938 // Load immediate does not have a zero or sign extended version
1939 // for 8-bit immediates
1940 // First load to TOS, then move to dst
1941 emit_rm(cbuf, 0x0, 0x0, 0x5);
1942 emit_float_constant(cbuf, $src$$constant);
1943 }
1944 %}
1946 enc_class LdImmX (regX dst, immXF con) %{ // Load Immediate
1947 emit_rm(cbuf, 0x0, $dst$$reg, 0x5);
1948 emit_float_constant(cbuf, $con$$constant);
1949 %}
1951 enc_class LdImmXD (regXD dst, immXD con) %{ // Load Immediate
1952 emit_rm(cbuf, 0x0, $dst$$reg, 0x5);
1953 emit_double_constant(cbuf, $con$$constant);
1954 %}
1956 enc_class load_conXD (regXD dst, immXD con) %{ // Load double constant
1957 // UseXmmLoadAndClearUpper ? movsd(dst, con) : movlpd(dst, con)
1958 emit_opcode(cbuf, UseXmmLoadAndClearUpper ? 0xF2 : 0x66);
1959 emit_opcode(cbuf, 0x0F);
1960 emit_opcode(cbuf, UseXmmLoadAndClearUpper ? 0x10 : 0x12);
1961 emit_rm(cbuf, 0x0, $dst$$reg, 0x5);
1962 emit_double_constant(cbuf, $con$$constant);
1963 %}
1965 enc_class Opc_MemImm_F(immF src) %{
1966 cbuf.set_inst_mark();
1967 $$$emit8$primary;
1968 emit_rm(cbuf, 0x0, $secondary, 0x5);
1969 emit_float_constant(cbuf, $src$$constant);
1970 %}
1973 enc_class MovI2X_reg(regX dst, eRegI src) %{
1974 emit_opcode(cbuf, 0x66 ); // MOVD dst,src
1975 emit_opcode(cbuf, 0x0F );
1976 emit_opcode(cbuf, 0x6E );
1977 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
1978 %}
1980 enc_class MovX2I_reg(eRegI dst, regX src) %{
1981 emit_opcode(cbuf, 0x66 ); // MOVD dst,src
1982 emit_opcode(cbuf, 0x0F );
1983 emit_opcode(cbuf, 0x7E );
1984 emit_rm(cbuf, 0x3, $src$$reg, $dst$$reg);
1985 %}
1987 enc_class MovL2XD_reg(regXD dst, eRegL src, regXD tmp) %{
1988 { // MOVD $dst,$src.lo
1989 emit_opcode(cbuf,0x66);
1990 emit_opcode(cbuf,0x0F);
1991 emit_opcode(cbuf,0x6E);
1992 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
1993 }
1994 { // MOVD $tmp,$src.hi
1995 emit_opcode(cbuf,0x66);
1996 emit_opcode(cbuf,0x0F);
1997 emit_opcode(cbuf,0x6E);
1998 emit_rm(cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($src$$reg));
1999 }
2000 { // PUNPCKLDQ $dst,$tmp
2001 emit_opcode(cbuf,0x66);
2002 emit_opcode(cbuf,0x0F);
2003 emit_opcode(cbuf,0x62);
2004 emit_rm(cbuf, 0x3, $dst$$reg, $tmp$$reg);
2005 }
2006 %}
2008 enc_class MovXD2L_reg(eRegL dst, regXD src, regXD tmp) %{
2009 { // MOVD $dst.lo,$src
2010 emit_opcode(cbuf,0x66);
2011 emit_opcode(cbuf,0x0F);
2012 emit_opcode(cbuf,0x7E);
2013 emit_rm(cbuf, 0x3, $src$$reg, $dst$$reg);
2014 }
2015 { // PSHUFLW $tmp,$src,0x4E (01001110b)
2016 emit_opcode(cbuf,0xF2);
2017 emit_opcode(cbuf,0x0F);
2018 emit_opcode(cbuf,0x70);
2019 emit_rm(cbuf, 0x3, $tmp$$reg, $src$$reg);
2020 emit_d8(cbuf, 0x4E);
2021 }
2022 { // MOVD $dst.hi,$tmp
2023 emit_opcode(cbuf,0x66);
2024 emit_opcode(cbuf,0x0F);
2025 emit_opcode(cbuf,0x7E);
2026 emit_rm(cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($dst$$reg));
2027 }
2028 %}
2031 // Encode a reg-reg copy. If it is useless, then empty encoding.
2032 enc_class enc_Copy( eRegI dst, eRegI src ) %{
2033 encode_Copy( cbuf, $dst$$reg, $src$$reg );
2034 %}
2036 enc_class enc_CopyL_Lo( eRegI dst, eRegL src ) %{
2037 encode_Copy( cbuf, $dst$$reg, $src$$reg );
2038 %}
2040 // Encode xmm reg-reg copy. If it is useless, then empty encoding.
2041 enc_class enc_CopyXD( RegXD dst, RegXD src ) %{
2042 encode_CopyXD( cbuf, $dst$$reg, $src$$reg );
2043 %}
2045 enc_class RegReg (eRegI dst, eRegI src) %{ // RegReg(Many)
2046 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2047 %}
2049 enc_class RegReg_Lo(eRegL dst, eRegL src) %{ // RegReg(Many)
2050 $$$emit8$primary;
2051 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2052 %}
2054 enc_class RegReg_Hi(eRegL dst, eRegL src) %{ // RegReg(Many)
2055 $$$emit8$secondary;
2056 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), HIGH_FROM_LOW($src$$reg));
2057 %}
2059 enc_class RegReg_Lo2(eRegL dst, eRegL src) %{ // RegReg(Many)
2060 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2061 %}
2063 enc_class RegReg_Hi2(eRegL dst, eRegL src) %{ // RegReg(Many)
2064 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), HIGH_FROM_LOW($src$$reg));
2065 %}
2067 enc_class RegReg_HiLo( eRegL src, eRegI dst ) %{
2068 emit_rm(cbuf, 0x3, $dst$$reg, HIGH_FROM_LOW($src$$reg));
2069 %}
2071 enc_class Con32 (immI src) %{ // Con32(storeImmI)
2072 // Output immediate
2073 $$$emit32$src$$constant;
2074 %}
2076 enc_class Con32F_as_bits(immF src) %{ // storeF_imm
2077 // Output Float immediate bits
2078 jfloat jf = $src$$constant;
2079 int jf_as_bits = jint_cast( jf );
2080 emit_d32(cbuf, jf_as_bits);
2081 %}
2083 enc_class Con32XF_as_bits(immXF src) %{ // storeX_imm
2084 // Output Float immediate bits
2085 jfloat jf = $src$$constant;
2086 int jf_as_bits = jint_cast( jf );
2087 emit_d32(cbuf, jf_as_bits);
2088 %}
2090 enc_class Con16 (immI src) %{ // Con16(storeImmI)
2091 // Output immediate
2092 $$$emit16$src$$constant;
2093 %}
2095 enc_class Con_d32(immI src) %{
2096 emit_d32(cbuf,$src$$constant);
2097 %}
2099 enc_class conmemref (eRegP t1) %{ // Con32(storeImmI)
2100 // Output immediate memory reference
2101 emit_rm(cbuf, 0x00, $t1$$reg, 0x05 );
2102 emit_d32(cbuf, 0x00);
2103 %}
2105 enc_class lock_prefix( ) %{
2106 if( os::is_MP() )
2107 emit_opcode(cbuf,0xF0); // [Lock]
2108 %}
2110 // Cmp-xchg long value.
2111 // Note: we need to swap rbx, and rcx before and after the
2112 // cmpxchg8 instruction because the instruction uses
2113 // rcx as the high order word of the new value to store but
2114 // our register encoding uses rbx,.
2115 enc_class enc_cmpxchg8(eSIRegP mem_ptr) %{
2117 // XCHG rbx,ecx
2118 emit_opcode(cbuf,0x87);
2119 emit_opcode(cbuf,0xD9);
2120 // [Lock]
2121 if( os::is_MP() )
2122 emit_opcode(cbuf,0xF0);
2123 // CMPXCHG8 [Eptr]
2124 emit_opcode(cbuf,0x0F);
2125 emit_opcode(cbuf,0xC7);
2126 emit_rm( cbuf, 0x0, 1, $mem_ptr$$reg );
2127 // XCHG rbx,ecx
2128 emit_opcode(cbuf,0x87);
2129 emit_opcode(cbuf,0xD9);
2130 %}
2132 enc_class enc_cmpxchg(eSIRegP mem_ptr) %{
2133 // [Lock]
2134 if( os::is_MP() )
2135 emit_opcode(cbuf,0xF0);
2137 // CMPXCHG [Eptr]
2138 emit_opcode(cbuf,0x0F);
2139 emit_opcode(cbuf,0xB1);
2140 emit_rm( cbuf, 0x0, 1, $mem_ptr$$reg );
2141 %}
2143 enc_class enc_flags_ne_to_boolean( iRegI res ) %{
2144 int res_encoding = $res$$reg;
2146 // MOV res,0
2147 emit_opcode( cbuf, 0xB8 + res_encoding);
2148 emit_d32( cbuf, 0 );
2149 // JNE,s fail
2150 emit_opcode(cbuf,0x75);
2151 emit_d8(cbuf, 5 );
2152 // MOV res,1
2153 emit_opcode( cbuf, 0xB8 + res_encoding);
2154 emit_d32( cbuf, 1 );
2155 // fail:
2156 %}
2158 enc_class set_instruction_start( ) %{
2159 cbuf.set_inst_mark(); // Mark start of opcode for reloc info in mem operand
2160 %}
2162 enc_class RegMem (eRegI ereg, memory mem) %{ // emit_reg_mem
2163 int reg_encoding = $ereg$$reg;
2164 int base = $mem$$base;
2165 int index = $mem$$index;
2166 int scale = $mem$$scale;
2167 int displace = $mem$$disp;
2168 bool disp_is_oop = $mem->disp_is_oop();
2169 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, disp_is_oop);
2170 %}
2172 enc_class RegMem_Hi(eRegL ereg, memory mem) %{ // emit_reg_mem
2173 int reg_encoding = HIGH_FROM_LOW($ereg$$reg); // Hi register of pair, computed from lo
2174 int base = $mem$$base;
2175 int index = $mem$$index;
2176 int scale = $mem$$scale;
2177 int displace = $mem$$disp + 4; // Offset is 4 further in memory
2178 assert( !$mem->disp_is_oop(), "Cannot add 4 to oop" );
2179 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, false/*disp_is_oop*/);
2180 %}
2182 enc_class move_long_small_shift( eRegL dst, immI_1_31 cnt ) %{
2183 int r1, r2;
2184 if( $tertiary == 0xA4 ) { r1 = $dst$$reg; r2 = HIGH_FROM_LOW($dst$$reg); }
2185 else { r2 = $dst$$reg; r1 = HIGH_FROM_LOW($dst$$reg); }
2186 emit_opcode(cbuf,0x0F);
2187 emit_opcode(cbuf,$tertiary);
2188 emit_rm(cbuf, 0x3, r1, r2);
2189 emit_d8(cbuf,$cnt$$constant);
2190 emit_d8(cbuf,$primary);
2191 emit_rm(cbuf, 0x3, $secondary, r1);
2192 emit_d8(cbuf,$cnt$$constant);
2193 %}
2195 enc_class move_long_big_shift_sign( eRegL dst, immI_32_63 cnt ) %{
2196 emit_opcode( cbuf, 0x8B ); // Move
2197 emit_rm(cbuf, 0x3, $dst$$reg, HIGH_FROM_LOW($dst$$reg));
2198 emit_d8(cbuf,$primary);
2199 emit_rm(cbuf, 0x3, $secondary, $dst$$reg);
2200 emit_d8(cbuf,$cnt$$constant-32);
2201 emit_d8(cbuf,$primary);
2202 emit_rm(cbuf, 0x3, $secondary, HIGH_FROM_LOW($dst$$reg));
2203 emit_d8(cbuf,31);
2204 %}
2206 enc_class move_long_big_shift_clr( eRegL dst, immI_32_63 cnt ) %{
2207 int r1, r2;
2208 if( $secondary == 0x5 ) { r1 = $dst$$reg; r2 = HIGH_FROM_LOW($dst$$reg); }
2209 else { r2 = $dst$$reg; r1 = HIGH_FROM_LOW($dst$$reg); }
2211 emit_opcode( cbuf, 0x8B ); // Move r1,r2
2212 emit_rm(cbuf, 0x3, r1, r2);
2213 if( $cnt$$constant > 32 ) { // Shift, if not by zero
2214 emit_opcode(cbuf,$primary);
2215 emit_rm(cbuf, 0x3, $secondary, r1);
2216 emit_d8(cbuf,$cnt$$constant-32);
2217 }
2218 emit_opcode(cbuf,0x33); // XOR r2,r2
2219 emit_rm(cbuf, 0x3, r2, r2);
2220 %}
2222 // Clone of RegMem but accepts an extra parameter to access each
2223 // half of a double in memory; it never needs relocation info.
2224 enc_class Mov_MemD_half_to_Reg (immI opcode, memory mem, immI disp_for_half, eRegI rm_reg) %{
2225 emit_opcode(cbuf,$opcode$$constant);
2226 int reg_encoding = $rm_reg$$reg;
2227 int base = $mem$$base;
2228 int index = $mem$$index;
2229 int scale = $mem$$scale;
2230 int displace = $mem$$disp + $disp_for_half$$constant;
2231 bool disp_is_oop = false;
2232 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, disp_is_oop);
2233 %}
2235 // !!!!! Special Custom Code used by MemMove, and stack access instructions !!!!!
2236 //
2237 // Clone of RegMem except the RM-byte's reg/opcode field is an ADLC-time constant
2238 // and it never needs relocation information.
2239 // Frequently used to move data between FPU's Stack Top and memory.
2240 enc_class RMopc_Mem_no_oop (immI rm_opcode, memory mem) %{
2241 int rm_byte_opcode = $rm_opcode$$constant;
2242 int base = $mem$$base;
2243 int index = $mem$$index;
2244 int scale = $mem$$scale;
2245 int displace = $mem$$disp;
2246 assert( !$mem->disp_is_oop(), "No oops here because no relo info allowed" );
2247 encode_RegMem(cbuf, rm_byte_opcode, base, index, scale, displace, false);
2248 %}
2250 enc_class RMopc_Mem (immI rm_opcode, memory mem) %{
2251 int rm_byte_opcode = $rm_opcode$$constant;
2252 int base = $mem$$base;
2253 int index = $mem$$index;
2254 int scale = $mem$$scale;
2255 int displace = $mem$$disp;
2256 bool disp_is_oop = $mem->disp_is_oop(); // disp-as-oop when working with static globals
2257 encode_RegMem(cbuf, rm_byte_opcode, base, index, scale, displace, disp_is_oop);
2258 %}
2260 enc_class RegLea (eRegI dst, eRegI src0, immI src1 ) %{ // emit_reg_lea
2261 int reg_encoding = $dst$$reg;
2262 int base = $src0$$reg; // 0xFFFFFFFF indicates no base
2263 int index = 0x04; // 0x04 indicates no index
2264 int scale = 0x00; // 0x00 indicates no scale
2265 int displace = $src1$$constant; // 0x00 indicates no displacement
2266 bool disp_is_oop = false;
2267 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, disp_is_oop);
2268 %}
2270 enc_class min_enc (eRegI dst, eRegI src) %{ // MIN
2271 // Compare dst,src
2272 emit_opcode(cbuf,0x3B);
2273 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2274 // jmp dst < src around move
2275 emit_opcode(cbuf,0x7C);
2276 emit_d8(cbuf,2);
2277 // move dst,src
2278 emit_opcode(cbuf,0x8B);
2279 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2280 %}
2282 enc_class max_enc (eRegI dst, eRegI src) %{ // MAX
2283 // Compare dst,src
2284 emit_opcode(cbuf,0x3B);
2285 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2286 // jmp dst > src around move
2287 emit_opcode(cbuf,0x7F);
2288 emit_d8(cbuf,2);
2289 // move dst,src
2290 emit_opcode(cbuf,0x8B);
2291 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2292 %}
2294 enc_class enc_FP_store(memory mem, regD src) %{
2295 // If src is FPR1, we can just FST to store it.
2296 // Else we need to FLD it to FPR1, then FSTP to store/pop it.
2297 int reg_encoding = 0x2; // Just store
2298 int base = $mem$$base;
2299 int index = $mem$$index;
2300 int scale = $mem$$scale;
2301 int displace = $mem$$disp;
2302 bool disp_is_oop = $mem->disp_is_oop(); // disp-as-oop when working with static globals
2303 if( $src$$reg != FPR1L_enc ) {
2304 reg_encoding = 0x3; // Store & pop
2305 emit_opcode( cbuf, 0xD9 ); // FLD (i.e., push it)
2306 emit_d8( cbuf, 0xC0-1+$src$$reg );
2307 }
2308 cbuf.set_inst_mark(); // Mark start of opcode for reloc info in mem operand
2309 emit_opcode(cbuf,$primary);
2310 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, disp_is_oop);
2311 %}
2313 enc_class neg_reg(eRegI dst) %{
2314 // NEG $dst
2315 emit_opcode(cbuf,0xF7);
2316 emit_rm(cbuf, 0x3, 0x03, $dst$$reg );
2317 %}
2319 enc_class setLT_reg(eCXRegI dst) %{
2320 // SETLT $dst
2321 emit_opcode(cbuf,0x0F);
2322 emit_opcode(cbuf,0x9C);
2323 emit_rm( cbuf, 0x3, 0x4, $dst$$reg );
2324 %}
2326 enc_class enc_cmpLTP(ncxRegI p, ncxRegI q, ncxRegI y, eCXRegI tmp) %{ // cadd_cmpLT
2327 int tmpReg = $tmp$$reg;
2329 // SUB $p,$q
2330 emit_opcode(cbuf,0x2B);
2331 emit_rm(cbuf, 0x3, $p$$reg, $q$$reg);
2332 // SBB $tmp,$tmp
2333 emit_opcode(cbuf,0x1B);
2334 emit_rm(cbuf, 0x3, tmpReg, tmpReg);
2335 // AND $tmp,$y
2336 emit_opcode(cbuf,0x23);
2337 emit_rm(cbuf, 0x3, tmpReg, $y$$reg);
2338 // ADD $p,$tmp
2339 emit_opcode(cbuf,0x03);
2340 emit_rm(cbuf, 0x3, $p$$reg, tmpReg);
2341 %}
2343 enc_class enc_cmpLTP_mem(eRegI p, eRegI q, memory mem, eCXRegI tmp) %{ // cadd_cmpLT
2344 int tmpReg = $tmp$$reg;
2346 // SUB $p,$q
2347 emit_opcode(cbuf,0x2B);
2348 emit_rm(cbuf, 0x3, $p$$reg, $q$$reg);
2349 // SBB $tmp,$tmp
2350 emit_opcode(cbuf,0x1B);
2351 emit_rm(cbuf, 0x3, tmpReg, tmpReg);
2352 // AND $tmp,$y
2353 cbuf.set_inst_mark(); // Mark start of opcode for reloc info in mem operand
2354 emit_opcode(cbuf,0x23);
2355 int reg_encoding = tmpReg;
2356 int base = $mem$$base;
2357 int index = $mem$$index;
2358 int scale = $mem$$scale;
2359 int displace = $mem$$disp;
2360 bool disp_is_oop = $mem->disp_is_oop();
2361 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, disp_is_oop);
2362 // ADD $p,$tmp
2363 emit_opcode(cbuf,0x03);
2364 emit_rm(cbuf, 0x3, $p$$reg, tmpReg);
2365 %}
2367 enc_class shift_left_long( eRegL dst, eCXRegI shift ) %{
2368 // TEST shift,32
2369 emit_opcode(cbuf,0xF7);
2370 emit_rm(cbuf, 0x3, 0, ECX_enc);
2371 emit_d32(cbuf,0x20);
2372 // JEQ,s small
2373 emit_opcode(cbuf, 0x74);
2374 emit_d8(cbuf, 0x04);
2375 // MOV $dst.hi,$dst.lo
2376 emit_opcode( cbuf, 0x8B );
2377 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), $dst$$reg );
2378 // CLR $dst.lo
2379 emit_opcode(cbuf, 0x33);
2380 emit_rm(cbuf, 0x3, $dst$$reg, $dst$$reg);
2381 // small:
2382 // SHLD $dst.hi,$dst.lo,$shift
2383 emit_opcode(cbuf,0x0F);
2384 emit_opcode(cbuf,0xA5);
2385 emit_rm(cbuf, 0x3, $dst$$reg, HIGH_FROM_LOW($dst$$reg));
2386 // SHL $dst.lo,$shift"
2387 emit_opcode(cbuf,0xD3);
2388 emit_rm(cbuf, 0x3, 0x4, $dst$$reg );
2389 %}
2391 enc_class shift_right_long( eRegL dst, eCXRegI shift ) %{
2392 // TEST shift,32
2393 emit_opcode(cbuf,0xF7);
2394 emit_rm(cbuf, 0x3, 0, ECX_enc);
2395 emit_d32(cbuf,0x20);
2396 // JEQ,s small
2397 emit_opcode(cbuf, 0x74);
2398 emit_d8(cbuf, 0x04);
2399 // MOV $dst.lo,$dst.hi
2400 emit_opcode( cbuf, 0x8B );
2401 emit_rm(cbuf, 0x3, $dst$$reg, HIGH_FROM_LOW($dst$$reg) );
2402 // CLR $dst.hi
2403 emit_opcode(cbuf, 0x33);
2404 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), HIGH_FROM_LOW($dst$$reg));
2405 // small:
2406 // SHRD $dst.lo,$dst.hi,$shift
2407 emit_opcode(cbuf,0x0F);
2408 emit_opcode(cbuf,0xAD);
2409 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), $dst$$reg);
2410 // SHR $dst.hi,$shift"
2411 emit_opcode(cbuf,0xD3);
2412 emit_rm(cbuf, 0x3, 0x5, HIGH_FROM_LOW($dst$$reg) );
2413 %}
2415 enc_class shift_right_arith_long( eRegL dst, eCXRegI shift ) %{
2416 // TEST shift,32
2417 emit_opcode(cbuf,0xF7);
2418 emit_rm(cbuf, 0x3, 0, ECX_enc);
2419 emit_d32(cbuf,0x20);
2420 // JEQ,s small
2421 emit_opcode(cbuf, 0x74);
2422 emit_d8(cbuf, 0x05);
2423 // MOV $dst.lo,$dst.hi
2424 emit_opcode( cbuf, 0x8B );
2425 emit_rm(cbuf, 0x3, $dst$$reg, HIGH_FROM_LOW($dst$$reg) );
2426 // SAR $dst.hi,31
2427 emit_opcode(cbuf, 0xC1);
2428 emit_rm(cbuf, 0x3, 7, HIGH_FROM_LOW($dst$$reg) );
2429 emit_d8(cbuf, 0x1F );
2430 // small:
2431 // SHRD $dst.lo,$dst.hi,$shift
2432 emit_opcode(cbuf,0x0F);
2433 emit_opcode(cbuf,0xAD);
2434 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), $dst$$reg);
2435 // SAR $dst.hi,$shift"
2436 emit_opcode(cbuf,0xD3);
2437 emit_rm(cbuf, 0x3, 0x7, HIGH_FROM_LOW($dst$$reg) );
2438 %}
2441 // ----------------- Encodings for floating point unit -----------------
2442 // May leave result in FPU-TOS or FPU reg depending on opcodes
2443 enc_class OpcReg_F (regF src) %{ // FMUL, FDIV
2444 $$$emit8$primary;
2445 emit_rm(cbuf, 0x3, $secondary, $src$$reg );
2446 %}
2448 // Pop argument in FPR0 with FSTP ST(0)
2449 enc_class PopFPU() %{
2450 emit_opcode( cbuf, 0xDD );
2451 emit_d8( cbuf, 0xD8 );
2452 %}
2454 // !!!!! equivalent to Pop_Reg_F
2455 enc_class Pop_Reg_D( regD dst ) %{
2456 emit_opcode( cbuf, 0xDD ); // FSTP ST(i)
2457 emit_d8( cbuf, 0xD8+$dst$$reg );
2458 %}
2460 enc_class Push_Reg_D( regD dst ) %{
2461 emit_opcode( cbuf, 0xD9 );
2462 emit_d8( cbuf, 0xC0-1+$dst$$reg ); // FLD ST(i-1)
2463 %}
2465 enc_class strictfp_bias1( regD dst ) %{
2466 emit_opcode( cbuf, 0xDB ); // FLD m80real
2467 emit_opcode( cbuf, 0x2D );
2468 emit_d32( cbuf, (int)StubRoutines::addr_fpu_subnormal_bias1() );
2469 emit_opcode( cbuf, 0xDE ); // FMULP ST(dst), ST0
2470 emit_opcode( cbuf, 0xC8+$dst$$reg );
2471 %}
2473 enc_class strictfp_bias2( regD dst ) %{
2474 emit_opcode( cbuf, 0xDB ); // FLD m80real
2475 emit_opcode( cbuf, 0x2D );
2476 emit_d32( cbuf, (int)StubRoutines::addr_fpu_subnormal_bias2() );
2477 emit_opcode( cbuf, 0xDE ); // FMULP ST(dst), ST0
2478 emit_opcode( cbuf, 0xC8+$dst$$reg );
2479 %}
2481 // Special case for moving an integer register to a stack slot.
2482 enc_class OpcPRegSS( stackSlotI dst, eRegI src ) %{ // RegSS
2483 store_to_stackslot( cbuf, $primary, $src$$reg, $dst$$disp );
2484 %}
2486 // Special case for moving a register to a stack slot.
2487 enc_class RegSS( stackSlotI dst, eRegI src ) %{ // RegSS
2488 // Opcode already emitted
2489 emit_rm( cbuf, 0x02, $src$$reg, ESP_enc ); // R/M byte
2490 emit_rm( cbuf, 0x00, ESP_enc, ESP_enc); // SIB byte
2491 emit_d32(cbuf, $dst$$disp); // Displacement
2492 %}
2494 // Push the integer in stackSlot 'src' onto FP-stack
2495 enc_class Push_Mem_I( memory src ) %{ // FILD [ESP+src]
2496 store_to_stackslot( cbuf, $primary, $secondary, $src$$disp );
2497 %}
2499 // Push the float in stackSlot 'src' onto FP-stack
2500 enc_class Push_Mem_F( memory src ) %{ // FLD_S [ESP+src]
2501 store_to_stackslot( cbuf, 0xD9, 0x00, $src$$disp );
2502 %}
2504 // Push the double in stackSlot 'src' onto FP-stack
2505 enc_class Push_Mem_D( memory src ) %{ // FLD_D [ESP+src]
2506 store_to_stackslot( cbuf, 0xDD, 0x00, $src$$disp );
2507 %}
2509 // Push FPU's TOS float to a stack-slot, and pop FPU-stack
2510 enc_class Pop_Mem_F( stackSlotF dst ) %{ // FSTP_S [ESP+dst]
2511 store_to_stackslot( cbuf, 0xD9, 0x03, $dst$$disp );
2512 %}
2514 // Same as Pop_Mem_F except for opcode
2515 // Push FPU's TOS double to a stack-slot, and pop FPU-stack
2516 enc_class Pop_Mem_D( stackSlotD dst ) %{ // FSTP_D [ESP+dst]
2517 store_to_stackslot( cbuf, 0xDD, 0x03, $dst$$disp );
2518 %}
2520 enc_class Pop_Reg_F( regF dst ) %{
2521 emit_opcode( cbuf, 0xDD ); // FSTP ST(i)
2522 emit_d8( cbuf, 0xD8+$dst$$reg );
2523 %}
2525 enc_class Push_Reg_F( regF dst ) %{
2526 emit_opcode( cbuf, 0xD9 ); // FLD ST(i-1)
2527 emit_d8( cbuf, 0xC0-1+$dst$$reg );
2528 %}
2530 // Push FPU's float to a stack-slot, and pop FPU-stack
2531 enc_class Pop_Mem_Reg_F( stackSlotF dst, regF src ) %{
2532 int pop = 0x02;
2533 if ($src$$reg != FPR1L_enc) {
2534 emit_opcode( cbuf, 0xD9 ); // FLD ST(i-1)
2535 emit_d8( cbuf, 0xC0-1+$src$$reg );
2536 pop = 0x03;
2537 }
2538 store_to_stackslot( cbuf, 0xD9, pop, $dst$$disp ); // FST<P>_S [ESP+dst]
2539 %}
2541 // Push FPU's double to a stack-slot, and pop FPU-stack
2542 enc_class Pop_Mem_Reg_D( stackSlotD dst, regD src ) %{
2543 int pop = 0x02;
2544 if ($src$$reg != FPR1L_enc) {
2545 emit_opcode( cbuf, 0xD9 ); // FLD ST(i-1)
2546 emit_d8( cbuf, 0xC0-1+$src$$reg );
2547 pop = 0x03;
2548 }
2549 store_to_stackslot( cbuf, 0xDD, pop, $dst$$disp ); // FST<P>_D [ESP+dst]
2550 %}
2552 // Push FPU's double to a FPU-stack-slot, and pop FPU-stack
2553 enc_class Pop_Reg_Reg_D( regD dst, regF src ) %{
2554 int pop = 0xD0 - 1; // -1 since we skip FLD
2555 if ($src$$reg != FPR1L_enc) {
2556 emit_opcode( cbuf, 0xD9 ); // FLD ST(src-1)
2557 emit_d8( cbuf, 0xC0-1+$src$$reg );
2558 pop = 0xD8;
2559 }
2560 emit_opcode( cbuf, 0xDD );
2561 emit_d8( cbuf, pop+$dst$$reg ); // FST<P> ST(i)
2562 %}
2565 enc_class Mul_Add_F( regF dst, regF src, regF src1, regF src2 ) %{
2566 MacroAssembler masm(&cbuf);
2567 masm.fld_s( $src1$$reg-1); // nothing at TOS, load TOS from src1.reg
2568 masm.fmul( $src2$$reg+0); // value at TOS
2569 masm.fadd( $src$$reg+0); // value at TOS
2570 masm.fstp_d( $dst$$reg+0); // value at TOS, popped off after store
2571 %}
2574 enc_class Push_Reg_Mod_D( regD dst, regD src) %{
2575 // load dst in FPR0
2576 emit_opcode( cbuf, 0xD9 );
2577 emit_d8( cbuf, 0xC0-1+$dst$$reg );
2578 if ($src$$reg != FPR1L_enc) {
2579 // fincstp
2580 emit_opcode (cbuf, 0xD9);
2581 emit_opcode (cbuf, 0xF7);
2582 // swap src with FPR1:
2583 // FXCH FPR1 with src
2584 emit_opcode(cbuf, 0xD9);
2585 emit_d8(cbuf, 0xC8-1+$src$$reg );
2586 // fdecstp
2587 emit_opcode (cbuf, 0xD9);
2588 emit_opcode (cbuf, 0xF6);
2589 }
2590 %}
2592 enc_class Push_ModD_encoding( regXD src0, regXD src1) %{
2593 // Allocate a word
2594 emit_opcode(cbuf,0x83); // SUB ESP,8
2595 emit_opcode(cbuf,0xEC);
2596 emit_d8(cbuf,0x08);
2598 emit_opcode (cbuf, 0xF2 ); // MOVSD [ESP], src1
2599 emit_opcode (cbuf, 0x0F );
2600 emit_opcode (cbuf, 0x11 );
2601 encode_RegMem(cbuf, $src1$$reg, ESP_enc, 0x4, 0, 0, false);
2603 emit_opcode(cbuf,0xDD ); // FLD_D [ESP]
2604 encode_RegMem(cbuf, 0x0, ESP_enc, 0x4, 0, 0, false);
2606 emit_opcode (cbuf, 0xF2 ); // MOVSD [ESP], src0
2607 emit_opcode (cbuf, 0x0F );
2608 emit_opcode (cbuf, 0x11 );
2609 encode_RegMem(cbuf, $src0$$reg, ESP_enc, 0x4, 0, 0, false);
2611 emit_opcode(cbuf,0xDD ); // FLD_D [ESP]
2612 encode_RegMem(cbuf, 0x0, ESP_enc, 0x4, 0, 0, false);
2614 %}
2616 enc_class Push_ModX_encoding( regX src0, regX src1) %{
2617 // Allocate a word
2618 emit_opcode(cbuf,0x83); // SUB ESP,4
2619 emit_opcode(cbuf,0xEC);
2620 emit_d8(cbuf,0x04);
2622 emit_opcode (cbuf, 0xF3 ); // MOVSS [ESP], src1
2623 emit_opcode (cbuf, 0x0F );
2624 emit_opcode (cbuf, 0x11 );
2625 encode_RegMem(cbuf, $src1$$reg, ESP_enc, 0x4, 0, 0, false);
2627 emit_opcode(cbuf,0xD9 ); // FLD [ESP]
2628 encode_RegMem(cbuf, 0x0, ESP_enc, 0x4, 0, 0, false);
2630 emit_opcode (cbuf, 0xF3 ); // MOVSS [ESP], src0
2631 emit_opcode (cbuf, 0x0F );
2632 emit_opcode (cbuf, 0x11 );
2633 encode_RegMem(cbuf, $src0$$reg, ESP_enc, 0x4, 0, 0, false);
2635 emit_opcode(cbuf,0xD9 ); // FLD [ESP]
2636 encode_RegMem(cbuf, 0x0, ESP_enc, 0x4, 0, 0, false);
2638 %}
2640 enc_class Push_ResultXD(regXD dst) %{
2641 store_to_stackslot( cbuf, 0xDD, 0x03, 0 ); //FSTP [ESP]
2643 // UseXmmLoadAndClearUpper ? movsd dst,[esp] : movlpd dst,[esp]
2644 emit_opcode (cbuf, UseXmmLoadAndClearUpper ? 0xF2 : 0x66);
2645 emit_opcode (cbuf, 0x0F );
2646 emit_opcode (cbuf, UseXmmLoadAndClearUpper ? 0x10 : 0x12);
2647 encode_RegMem(cbuf, $dst$$reg, ESP_enc, 0x4, 0, 0, false);
2649 emit_opcode(cbuf,0x83); // ADD ESP,8
2650 emit_opcode(cbuf,0xC4);
2651 emit_d8(cbuf,0x08);
2652 %}
2654 enc_class Push_ResultX(regX dst, immI d8) %{
2655 store_to_stackslot( cbuf, 0xD9, 0x03, 0 ); //FSTP_S [ESP]
2657 emit_opcode (cbuf, 0xF3 ); // MOVSS dst(xmm), [ESP]
2658 emit_opcode (cbuf, 0x0F );
2659 emit_opcode (cbuf, 0x10 );
2660 encode_RegMem(cbuf, $dst$$reg, ESP_enc, 0x4, 0, 0, false);
2662 emit_opcode(cbuf,0x83); // ADD ESP,d8 (4 or 8)
2663 emit_opcode(cbuf,0xC4);
2664 emit_d8(cbuf,$d8$$constant);
2665 %}
2667 enc_class Push_SrcXD(regXD src) %{
2668 // Allocate a word
2669 emit_opcode(cbuf,0x83); // SUB ESP,8
2670 emit_opcode(cbuf,0xEC);
2671 emit_d8(cbuf,0x08);
2673 emit_opcode (cbuf, 0xF2 ); // MOVSD [ESP], src
2674 emit_opcode (cbuf, 0x0F );
2675 emit_opcode (cbuf, 0x11 );
2676 encode_RegMem(cbuf, $src$$reg, ESP_enc, 0x4, 0, 0, false);
2678 emit_opcode(cbuf,0xDD ); // FLD_D [ESP]
2679 encode_RegMem(cbuf, 0x0, ESP_enc, 0x4, 0, 0, false);
2680 %}
2682 enc_class push_stack_temp_qword() %{
2683 emit_opcode(cbuf,0x83); // SUB ESP,8
2684 emit_opcode(cbuf,0xEC);
2685 emit_d8 (cbuf,0x08);
2686 %}
2688 enc_class pop_stack_temp_qword() %{
2689 emit_opcode(cbuf,0x83); // ADD ESP,8
2690 emit_opcode(cbuf,0xC4);
2691 emit_d8 (cbuf,0x08);
2692 %}
2694 enc_class push_xmm_to_fpr1( regXD xmm_src ) %{
2695 emit_opcode (cbuf, 0xF2 ); // MOVSD [ESP], xmm_src
2696 emit_opcode (cbuf, 0x0F );
2697 emit_opcode (cbuf, 0x11 );
2698 encode_RegMem(cbuf, $xmm_src$$reg, ESP_enc, 0x4, 0, 0, false);
2700 emit_opcode(cbuf,0xDD ); // FLD_D [ESP]
2701 encode_RegMem(cbuf, 0x0, ESP_enc, 0x4, 0, 0, false);
2702 %}
2704 // Compute X^Y using Intel's fast hardware instructions, if possible.
2705 // Otherwise return a NaN.
2706 enc_class pow_exp_core_encoding %{
2707 // FPR1 holds Y*ln2(X). Compute FPR1 = 2^(Y*ln2(X))
2708 emit_opcode(cbuf,0xD9); emit_opcode(cbuf,0xC0); // fdup = fld st(0) Q Q
2709 emit_opcode(cbuf,0xD9); emit_opcode(cbuf,0xFC); // frndint int(Q) Q
2710 emit_opcode(cbuf,0xDC); emit_opcode(cbuf,0xE9); // fsub st(1) -= st(0); int(Q) frac(Q)
2711 emit_opcode(cbuf,0xDB); // FISTP [ESP] frac(Q)
2712 emit_opcode(cbuf,0x1C);
2713 emit_d8(cbuf,0x24);
2714 emit_opcode(cbuf,0xD9); emit_opcode(cbuf,0xF0); // f2xm1 2^frac(Q)-1
2715 emit_opcode(cbuf,0xD9); emit_opcode(cbuf,0xE8); // fld1 1 2^frac(Q)-1
2716 emit_opcode(cbuf,0xDE); emit_opcode(cbuf,0xC1); // faddp 2^frac(Q)
2717 emit_opcode(cbuf,0x8B); // mov rax,[esp+0]=int(Q)
2718 encode_RegMem(cbuf, EAX_enc, ESP_enc, 0x4, 0, 0, false);
2719 emit_opcode(cbuf,0xC7); // mov rcx,0xFFFFF800 - overflow mask
2720 emit_rm(cbuf, 0x3, 0x0, ECX_enc);
2721 emit_d32(cbuf,0xFFFFF800);
2722 emit_opcode(cbuf,0x81); // add rax,1023 - the double exponent bias
2723 emit_rm(cbuf, 0x3, 0x0, EAX_enc);
2724 emit_d32(cbuf,1023);
2725 emit_opcode(cbuf,0x8B); // mov rbx,eax
2726 emit_rm(cbuf, 0x3, EBX_enc, EAX_enc);
2727 emit_opcode(cbuf,0xC1); // shl rax,20 - Slide to exponent position
2728 emit_rm(cbuf,0x3,0x4,EAX_enc);
2729 emit_d8(cbuf,20);
2730 emit_opcode(cbuf,0x85); // test rbx,ecx - check for overflow
2731 emit_rm(cbuf, 0x3, EBX_enc, ECX_enc);
2732 emit_opcode(cbuf,0x0F); emit_opcode(cbuf,0x45); // CMOVne rax,ecx - overflow; stuff NAN into EAX
2733 emit_rm(cbuf, 0x3, EAX_enc, ECX_enc);
2734 emit_opcode(cbuf,0x89); // mov [esp+4],eax - Store as part of double word
2735 encode_RegMem(cbuf, EAX_enc, ESP_enc, 0x4, 0, 4, false);
2736 emit_opcode(cbuf,0xC7); // mov [esp+0],0 - [ESP] = (double)(1<<int(Q)) = 2^int(Q)
2737 encode_RegMem(cbuf, 0x0, ESP_enc, 0x4, 0, 0, false);
2738 emit_d32(cbuf,0);
2739 emit_opcode(cbuf,0xDC); // fmul dword st(0),[esp+0]; FPR1 = 2^int(Q)*2^frac(Q) = 2^Q
2740 encode_RegMem(cbuf, 0x1, ESP_enc, 0x4, 0, 0, false);
2741 %}
2743 // enc_class Pop_Reg_Mod_D( regD dst, regD src)
2744 // was replaced by Push_Result_Mod_D followed by Pop_Reg_X() or Pop_Mem_X()
2746 enc_class Push_Result_Mod_D( regD src) %{
2747 if ($src$$reg != FPR1L_enc) {
2748 // fincstp
2749 emit_opcode (cbuf, 0xD9);
2750 emit_opcode (cbuf, 0xF7);
2751 // FXCH FPR1 with src
2752 emit_opcode(cbuf, 0xD9);
2753 emit_d8(cbuf, 0xC8-1+$src$$reg );
2754 // fdecstp
2755 emit_opcode (cbuf, 0xD9);
2756 emit_opcode (cbuf, 0xF6);
2757 }
2758 // // following asm replaced with Pop_Reg_F or Pop_Mem_F
2759 // // FSTP FPR$dst$$reg
2760 // emit_opcode( cbuf, 0xDD );
2761 // emit_d8( cbuf, 0xD8+$dst$$reg );
2762 %}
2764 enc_class fnstsw_sahf_skip_parity() %{
2765 // fnstsw ax
2766 emit_opcode( cbuf, 0xDF );
2767 emit_opcode( cbuf, 0xE0 );
2768 // sahf
2769 emit_opcode( cbuf, 0x9E );
2770 // jnp ::skip
2771 emit_opcode( cbuf, 0x7B );
2772 emit_opcode( cbuf, 0x05 );
2773 %}
2775 enc_class emitModD() %{
2776 // fprem must be iterative
2777 // :: loop
2778 // fprem
2779 emit_opcode( cbuf, 0xD9 );
2780 emit_opcode( cbuf, 0xF8 );
2781 // wait
2782 emit_opcode( cbuf, 0x9b );
2783 // fnstsw ax
2784 emit_opcode( cbuf, 0xDF );
2785 emit_opcode( cbuf, 0xE0 );
2786 // sahf
2787 emit_opcode( cbuf, 0x9E );
2788 // jp ::loop
2789 emit_opcode( cbuf, 0x0F );
2790 emit_opcode( cbuf, 0x8A );
2791 emit_opcode( cbuf, 0xF4 );
2792 emit_opcode( cbuf, 0xFF );
2793 emit_opcode( cbuf, 0xFF );
2794 emit_opcode( cbuf, 0xFF );
2795 %}
2797 enc_class fpu_flags() %{
2798 // fnstsw_ax
2799 emit_opcode( cbuf, 0xDF);
2800 emit_opcode( cbuf, 0xE0);
2801 // test ax,0x0400
2802 emit_opcode( cbuf, 0x66 ); // operand-size prefix for 16-bit immediate
2803 emit_opcode( cbuf, 0xA9 );
2804 emit_d16 ( cbuf, 0x0400 );
2805 // // // This sequence works, but stalls for 12-16 cycles on PPro
2806 // // test rax,0x0400
2807 // emit_opcode( cbuf, 0xA9 );
2808 // emit_d32 ( cbuf, 0x00000400 );
2809 //
2810 // jz exit (no unordered comparison)
2811 emit_opcode( cbuf, 0x74 );
2812 emit_d8 ( cbuf, 0x02 );
2813 // mov ah,1 - treat as LT case (set carry flag)
2814 emit_opcode( cbuf, 0xB4 );
2815 emit_d8 ( cbuf, 0x01 );
2816 // sahf
2817 emit_opcode( cbuf, 0x9E);
2818 %}
2820 enc_class cmpF_P6_fixup() %{
2821 // Fixup the integer flags in case comparison involved a NaN
2822 //
2823 // JNP exit (no unordered comparison, P-flag is set by NaN)
2824 emit_opcode( cbuf, 0x7B );
2825 emit_d8 ( cbuf, 0x03 );
2826 // MOV AH,1 - treat as LT case (set carry flag)
2827 emit_opcode( cbuf, 0xB4 );
2828 emit_d8 ( cbuf, 0x01 );
2829 // SAHF
2830 emit_opcode( cbuf, 0x9E);
2831 // NOP // target for branch to avoid branch to branch
2832 emit_opcode( cbuf, 0x90);
2833 %}
2835 // fnstsw_ax();
2836 // sahf();
2837 // movl(dst, nan_result);
2838 // jcc(Assembler::parity, exit);
2839 // movl(dst, less_result);
2840 // jcc(Assembler::below, exit);
2841 // movl(dst, equal_result);
2842 // jcc(Assembler::equal, exit);
2843 // movl(dst, greater_result);
2845 // less_result = 1;
2846 // greater_result = -1;
2847 // equal_result = 0;
2848 // nan_result = -1;
2850 enc_class CmpF_Result(eRegI dst) %{
2851 // fnstsw_ax();
2852 emit_opcode( cbuf, 0xDF);
2853 emit_opcode( cbuf, 0xE0);
2854 // sahf
2855 emit_opcode( cbuf, 0x9E);
2856 // movl(dst, nan_result);
2857 emit_opcode( cbuf, 0xB8 + $dst$$reg);
2858 emit_d32( cbuf, -1 );
2859 // jcc(Assembler::parity, exit);
2860 emit_opcode( cbuf, 0x7A );
2861 emit_d8 ( cbuf, 0x13 );
2862 // movl(dst, less_result);
2863 emit_opcode( cbuf, 0xB8 + $dst$$reg);
2864 emit_d32( cbuf, -1 );
2865 // jcc(Assembler::below, exit);
2866 emit_opcode( cbuf, 0x72 );
2867 emit_d8 ( cbuf, 0x0C );
2868 // movl(dst, equal_result);
2869 emit_opcode( cbuf, 0xB8 + $dst$$reg);
2870 emit_d32( cbuf, 0 );
2871 // jcc(Assembler::equal, exit);
2872 emit_opcode( cbuf, 0x74 );
2873 emit_d8 ( cbuf, 0x05 );
2874 // movl(dst, greater_result);
2875 emit_opcode( cbuf, 0xB8 + $dst$$reg);
2876 emit_d32( cbuf, 1 );
2877 %}
2880 // XMM version of CmpF_Result. Because the XMM compare
2881 // instructions set the EFLAGS directly. It becomes simpler than
2882 // the float version above.
2883 enc_class CmpX_Result(eRegI dst) %{
2884 MacroAssembler _masm(&cbuf);
2885 Label nan, inc, done;
2887 __ jccb(Assembler::parity, nan);
2888 __ jccb(Assembler::equal, done);
2889 __ jccb(Assembler::above, inc);
2890 __ bind(nan);
2891 __ decrement(as_Register($dst$$reg));
2892 __ jmpb(done);
2893 __ bind(inc);
2894 __ increment(as_Register($dst$$reg));
2895 __ bind(done);
2896 %}
2898 // Compare the longs and set flags
2899 // BROKEN! Do Not use as-is
2900 enc_class cmpl_test( eRegL src1, eRegL src2 ) %{
2901 // CMP $src1.hi,$src2.hi
2902 emit_opcode( cbuf, 0x3B );
2903 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($src1$$reg), HIGH_FROM_LOW($src2$$reg) );
2904 // JNE,s done
2905 emit_opcode(cbuf,0x75);
2906 emit_d8(cbuf, 2 );
2907 // CMP $src1.lo,$src2.lo
2908 emit_opcode( cbuf, 0x3B );
2909 emit_rm(cbuf, 0x3, $src1$$reg, $src2$$reg );
2910 // done:
2911 %}
2913 enc_class convert_int_long( regL dst, eRegI src ) %{
2914 // mov $dst.lo,$src
2915 int dst_encoding = $dst$$reg;
2916 int src_encoding = $src$$reg;
2917 encode_Copy( cbuf, dst_encoding , src_encoding );
2918 // mov $dst.hi,$src
2919 encode_Copy( cbuf, HIGH_FROM_LOW(dst_encoding), src_encoding );
2920 // sar $dst.hi,31
2921 emit_opcode( cbuf, 0xC1 );
2922 emit_rm(cbuf, 0x3, 7, HIGH_FROM_LOW(dst_encoding) );
2923 emit_d8(cbuf, 0x1F );
2924 %}
2926 enc_class convert_long_double( eRegL src ) %{
2927 // push $src.hi
2928 emit_opcode(cbuf, 0x50+HIGH_FROM_LOW($src$$reg));
2929 // push $src.lo
2930 emit_opcode(cbuf, 0x50+$src$$reg );
2931 // fild 64-bits at [SP]
2932 emit_opcode(cbuf,0xdf);
2933 emit_d8(cbuf, 0x6C);
2934 emit_d8(cbuf, 0x24);
2935 emit_d8(cbuf, 0x00);
2936 // pop stack
2937 emit_opcode(cbuf, 0x83); // add SP, #8
2938 emit_rm(cbuf, 0x3, 0x00, ESP_enc);
2939 emit_d8(cbuf, 0x8);
2940 %}
2942 enc_class multiply_con_and_shift_high( eDXRegI dst, nadxRegI src1, eADXRegL_low_only src2, immI_32_63 cnt, eFlagsReg cr ) %{
2943 // IMUL EDX:EAX,$src1
2944 emit_opcode( cbuf, 0xF7 );
2945 emit_rm( cbuf, 0x3, 0x5, $src1$$reg );
2946 // SAR EDX,$cnt-32
2947 int shift_count = ((int)$cnt$$constant) - 32;
2948 if (shift_count > 0) {
2949 emit_opcode(cbuf, 0xC1);
2950 emit_rm(cbuf, 0x3, 7, $dst$$reg );
2951 emit_d8(cbuf, shift_count);
2952 }
2953 %}
2955 // this version doesn't have add sp, 8
2956 enc_class convert_long_double2( eRegL src ) %{
2957 // push $src.hi
2958 emit_opcode(cbuf, 0x50+HIGH_FROM_LOW($src$$reg));
2959 // push $src.lo
2960 emit_opcode(cbuf, 0x50+$src$$reg );
2961 // fild 64-bits at [SP]
2962 emit_opcode(cbuf,0xdf);
2963 emit_d8(cbuf, 0x6C);
2964 emit_d8(cbuf, 0x24);
2965 emit_d8(cbuf, 0x00);
2966 %}
2968 enc_class long_int_multiply( eADXRegL dst, nadxRegI src) %{
2969 // Basic idea: long = (long)int * (long)int
2970 // IMUL EDX:EAX, src
2971 emit_opcode( cbuf, 0xF7 );
2972 emit_rm( cbuf, 0x3, 0x5, $src$$reg);
2973 %}
2975 enc_class long_uint_multiply( eADXRegL dst, nadxRegI src) %{
2976 // Basic Idea: long = (int & 0xffffffffL) * (int & 0xffffffffL)
2977 // MUL EDX:EAX, src
2978 emit_opcode( cbuf, 0xF7 );
2979 emit_rm( cbuf, 0x3, 0x4, $src$$reg);
2980 %}
2982 enc_class long_multiply( eADXRegL dst, eRegL src, eRegI tmp ) %{
2983 // Basic idea: lo(result) = lo(x_lo * y_lo)
2984 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi)
2985 // MOV $tmp,$src.lo
2986 encode_Copy( cbuf, $tmp$$reg, $src$$reg );
2987 // IMUL $tmp,EDX
2988 emit_opcode( cbuf, 0x0F );
2989 emit_opcode( cbuf, 0xAF );
2990 emit_rm( cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($dst$$reg) );
2991 // MOV EDX,$src.hi
2992 encode_Copy( cbuf, HIGH_FROM_LOW($dst$$reg), HIGH_FROM_LOW($src$$reg) );
2993 // IMUL EDX,EAX
2994 emit_opcode( cbuf, 0x0F );
2995 emit_opcode( cbuf, 0xAF );
2996 emit_rm( cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), $dst$$reg );
2997 // ADD $tmp,EDX
2998 emit_opcode( cbuf, 0x03 );
2999 emit_rm( cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($dst$$reg) );
3000 // MUL EDX:EAX,$src.lo
3001 emit_opcode( cbuf, 0xF7 );
3002 emit_rm( cbuf, 0x3, 0x4, $src$$reg );
3003 // ADD EDX,ESI
3004 emit_opcode( cbuf, 0x03 );
3005 emit_rm( cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), $tmp$$reg );
3006 %}
3008 enc_class long_multiply_con( eADXRegL dst, immL_127 src, eRegI tmp ) %{
3009 // Basic idea: lo(result) = lo(src * y_lo)
3010 // hi(result) = hi(src * y_lo) + lo(src * y_hi)
3011 // IMUL $tmp,EDX,$src
3012 emit_opcode( cbuf, 0x6B );
3013 emit_rm( cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($dst$$reg) );
3014 emit_d8( cbuf, (int)$src$$constant );
3015 // MOV EDX,$src
3016 emit_opcode(cbuf, 0xB8 + EDX_enc);
3017 emit_d32( cbuf, (int)$src$$constant );
3018 // MUL EDX:EAX,EDX
3019 emit_opcode( cbuf, 0xF7 );
3020 emit_rm( cbuf, 0x3, 0x4, EDX_enc );
3021 // ADD EDX,ESI
3022 emit_opcode( cbuf, 0x03 );
3023 emit_rm( cbuf, 0x3, EDX_enc, $tmp$$reg );
3024 %}
3026 enc_class long_div( eRegL src1, eRegL src2 ) %{
3027 // PUSH src1.hi
3028 emit_opcode(cbuf, HIGH_FROM_LOW(0x50+$src1$$reg) );
3029 // PUSH src1.lo
3030 emit_opcode(cbuf, 0x50+$src1$$reg );
3031 // PUSH src2.hi
3032 emit_opcode(cbuf, HIGH_FROM_LOW(0x50+$src2$$reg) );
3033 // PUSH src2.lo
3034 emit_opcode(cbuf, 0x50+$src2$$reg );
3035 // CALL directly to the runtime
3036 cbuf.set_inst_mark();
3037 emit_opcode(cbuf,0xE8); // Call into runtime
3038 emit_d32_reloc(cbuf, (CAST_FROM_FN_PTR(address, SharedRuntime::ldiv) - cbuf.code_end()) - 4, runtime_call_Relocation::spec(), RELOC_IMM32 );
3039 // Restore stack
3040 emit_opcode(cbuf, 0x83); // add SP, #framesize
3041 emit_rm(cbuf, 0x3, 0x00, ESP_enc);
3042 emit_d8(cbuf, 4*4);
3043 %}
3045 enc_class long_mod( eRegL src1, eRegL src2 ) %{
3046 // PUSH src1.hi
3047 emit_opcode(cbuf, HIGH_FROM_LOW(0x50+$src1$$reg) );
3048 // PUSH src1.lo
3049 emit_opcode(cbuf, 0x50+$src1$$reg );
3050 // PUSH src2.hi
3051 emit_opcode(cbuf, HIGH_FROM_LOW(0x50+$src2$$reg) );
3052 // PUSH src2.lo
3053 emit_opcode(cbuf, 0x50+$src2$$reg );
3054 // CALL directly to the runtime
3055 cbuf.set_inst_mark();
3056 emit_opcode(cbuf,0xE8); // Call into runtime
3057 emit_d32_reloc(cbuf, (CAST_FROM_FN_PTR(address, SharedRuntime::lrem ) - cbuf.code_end()) - 4, runtime_call_Relocation::spec(), RELOC_IMM32 );
3058 // Restore stack
3059 emit_opcode(cbuf, 0x83); // add SP, #framesize
3060 emit_rm(cbuf, 0x3, 0x00, ESP_enc);
3061 emit_d8(cbuf, 4*4);
3062 %}
3064 enc_class long_cmp_flags0( eRegL src, eRegI tmp ) %{
3065 // MOV $tmp,$src.lo
3066 emit_opcode(cbuf, 0x8B);
3067 emit_rm(cbuf, 0x3, $tmp$$reg, $src$$reg);
3068 // OR $tmp,$src.hi
3069 emit_opcode(cbuf, 0x0B);
3070 emit_rm(cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($src$$reg));
3071 %}
3073 enc_class long_cmp_flags1( eRegL src1, eRegL src2 ) %{
3074 // CMP $src1.lo,$src2.lo
3075 emit_opcode( cbuf, 0x3B );
3076 emit_rm(cbuf, 0x3, $src1$$reg, $src2$$reg );
3077 // JNE,s skip
3078 emit_cc(cbuf, 0x70, 0x5);
3079 emit_d8(cbuf,2);
3080 // CMP $src1.hi,$src2.hi
3081 emit_opcode( cbuf, 0x3B );
3082 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($src1$$reg), HIGH_FROM_LOW($src2$$reg) );
3083 %}
3085 enc_class long_cmp_flags2( eRegL src1, eRegL src2, eRegI tmp ) %{
3086 // CMP $src1.lo,$src2.lo\t! Long compare; set flags for low bits
3087 emit_opcode( cbuf, 0x3B );
3088 emit_rm(cbuf, 0x3, $src1$$reg, $src2$$reg );
3089 // MOV $tmp,$src1.hi
3090 emit_opcode( cbuf, 0x8B );
3091 emit_rm(cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($src1$$reg) );
3092 // SBB $tmp,$src2.hi\t! Compute flags for long compare
3093 emit_opcode( cbuf, 0x1B );
3094 emit_rm(cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($src2$$reg) );
3095 %}
3097 enc_class long_cmp_flags3( eRegL src, eRegI tmp ) %{
3098 // XOR $tmp,$tmp
3099 emit_opcode(cbuf,0x33); // XOR
3100 emit_rm(cbuf,0x3, $tmp$$reg, $tmp$$reg);
3101 // CMP $tmp,$src.lo
3102 emit_opcode( cbuf, 0x3B );
3103 emit_rm(cbuf, 0x3, $tmp$$reg, $src$$reg );
3104 // SBB $tmp,$src.hi
3105 emit_opcode( cbuf, 0x1B );
3106 emit_rm(cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($src$$reg) );
3107 %}
3109 // Sniff, sniff... smells like Gnu Superoptimizer
3110 enc_class neg_long( eRegL dst ) %{
3111 emit_opcode(cbuf,0xF7); // NEG hi
3112 emit_rm (cbuf,0x3, 0x3, HIGH_FROM_LOW($dst$$reg));
3113 emit_opcode(cbuf,0xF7); // NEG lo
3114 emit_rm (cbuf,0x3, 0x3, $dst$$reg );
3115 emit_opcode(cbuf,0x83); // SBB hi,0
3116 emit_rm (cbuf,0x3, 0x3, HIGH_FROM_LOW($dst$$reg));
3117 emit_d8 (cbuf,0 );
3118 %}
3120 enc_class movq_ld(regXD dst, memory mem) %{
3121 MacroAssembler _masm(&cbuf);
3122 Address madr = Address::make_raw($mem$$base, $mem$$index, $mem$$scale, $mem$$disp);
3123 __ movq(as_XMMRegister($dst$$reg), madr);
3124 %}
3126 enc_class movq_st(memory mem, regXD src) %{
3127 MacroAssembler _masm(&cbuf);
3128 Address madr = Address::make_raw($mem$$base, $mem$$index, $mem$$scale, $mem$$disp);
3129 __ movq(madr, as_XMMRegister($src$$reg));
3130 %}
3132 enc_class pshufd_8x8(regX dst, regX src) %{
3133 MacroAssembler _masm(&cbuf);
3135 encode_CopyXD(cbuf, $dst$$reg, $src$$reg);
3136 __ punpcklbw(as_XMMRegister($dst$$reg), as_XMMRegister($dst$$reg));
3137 __ pshuflw(as_XMMRegister($dst$$reg), as_XMMRegister($dst$$reg), 0x00);
3138 %}
3140 enc_class pshufd_4x16(regX dst, regX src) %{
3141 MacroAssembler _masm(&cbuf);
3143 __ pshuflw(as_XMMRegister($dst$$reg), as_XMMRegister($src$$reg), 0x00);
3144 %}
3146 enc_class pshufd(regXD dst, regXD src, int mode) %{
3147 MacroAssembler _masm(&cbuf);
3149 __ pshufd(as_XMMRegister($dst$$reg), as_XMMRegister($src$$reg), $mode);
3150 %}
3152 enc_class pxor(regXD dst, regXD src) %{
3153 MacroAssembler _masm(&cbuf);
3155 __ pxor(as_XMMRegister($dst$$reg), as_XMMRegister($src$$reg));
3156 %}
3158 enc_class mov_i2x(regXD dst, eRegI src) %{
3159 MacroAssembler _masm(&cbuf);
3161 __ movd(as_XMMRegister($dst$$reg), as_Register($src$$reg));
3162 %}
3165 // Because the transitions from emitted code to the runtime
3166 // monitorenter/exit helper stubs are so slow it's critical that
3167 // we inline both the stack-locking fast-path and the inflated fast path.
3168 //
3169 // See also: cmpFastLock and cmpFastUnlock.
3170 //
3171 // What follows is a specialized inline transliteration of the code
3172 // in slow_enter() and slow_exit(). If we're concerned about I$ bloat
3173 // another option would be to emit TrySlowEnter and TrySlowExit methods
3174 // at startup-time. These methods would accept arguments as
3175 // (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure
3176 // indications in the icc.ZFlag. Fast_Lock and Fast_Unlock would simply
3177 // marshal the arguments and emit calls to TrySlowEnter and TrySlowExit.
3178 // In practice, however, the # of lock sites is bounded and is usually small.
3179 // Besides the call overhead, TrySlowEnter and TrySlowExit might suffer
3180 // if the processor uses simple bimodal branch predictors keyed by EIP
3181 // Since the helper routines would be called from multiple synchronization
3182 // sites.
3183 //
3184 // An even better approach would be write "MonitorEnter()" and "MonitorExit()"
3185 // in java - using j.u.c and unsafe - and just bind the lock and unlock sites
3186 // to those specialized methods. That'd give us a mostly platform-independent
3187 // implementation that the JITs could optimize and inline at their pleasure.
3188 // Done correctly, the only time we'd need to cross to native could would be
3189 // to park() or unpark() threads. We'd also need a few more unsafe operators
3190 // to (a) prevent compiler-JIT reordering of non-volatile accesses, and
3191 // (b) explicit barriers or fence operations.
3192 //
3193 // TODO:
3194 //
3195 // * Arrange for C2 to pass "Self" into Fast_Lock and Fast_Unlock in one of the registers (scr).
3196 // This avoids manifesting the Self pointer in the Fast_Lock and Fast_Unlock terminals.
3197 // Given TLAB allocation, Self is usually manifested in a register, so passing it into
3198 // the lock operators would typically be faster than reifying Self.
3199 //
3200 // * Ideally I'd define the primitives as:
3201 // fast_lock (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED.
3202 // fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED
3203 // Unfortunately ADLC bugs prevent us from expressing the ideal form.
3204 // Instead, we're stuck with a rather awkward and brittle register assignments below.
3205 // Furthermore the register assignments are overconstrained, possibly resulting in
3206 // sub-optimal code near the synchronization site.
3207 //
3208 // * Eliminate the sp-proximity tests and just use "== Self" tests instead.
3209 // Alternately, use a better sp-proximity test.
3210 //
3211 // * Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value.
3212 // Either one is sufficient to uniquely identify a thread.
3213 // TODO: eliminate use of sp in _owner and use get_thread(tr) instead.
3214 //
3215 // * Intrinsify notify() and notifyAll() for the common cases where the
3216 // object is locked by the calling thread but the waitlist is empty.
3217 // avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll().
3218 //
3219 // * use jccb and jmpb instead of jcc and jmp to improve code density.
3220 // But beware of excessive branch density on AMD Opterons.
3221 //
3222 // * Both Fast_Lock and Fast_Unlock set the ICC.ZF to indicate success
3223 // or failure of the fast-path. If the fast-path fails then we pass
3224 // control to the slow-path, typically in C. In Fast_Lock and
3225 // Fast_Unlock we often branch to DONE_LABEL, just to find that C2
3226 // will emit a conditional branch immediately after the node.
3227 // So we have branches to branches and lots of ICC.ZF games.
3228 // Instead, it might be better to have C2 pass a "FailureLabel"
3229 // into Fast_Lock and Fast_Unlock. In the case of success, control
3230 // will drop through the node. ICC.ZF is undefined at exit.
3231 // In the case of failure, the node will branch directly to the
3232 // FailureLabel
3235 // obj: object to lock
3236 // box: on-stack box address (displaced header location) - KILLED
3237 // rax,: tmp -- KILLED
3238 // scr: tmp -- KILLED
3239 enc_class Fast_Lock( eRegP obj, eRegP box, eAXRegI tmp, eRegP scr ) %{
3241 Register objReg = as_Register($obj$$reg);
3242 Register boxReg = as_Register($box$$reg);
3243 Register tmpReg = as_Register($tmp$$reg);
3244 Register scrReg = as_Register($scr$$reg);
3246 // Ensure the register assignents are disjoint
3247 guarantee (objReg != boxReg, "") ;
3248 guarantee (objReg != tmpReg, "") ;
3249 guarantee (objReg != scrReg, "") ;
3250 guarantee (boxReg != tmpReg, "") ;
3251 guarantee (boxReg != scrReg, "") ;
3252 guarantee (tmpReg == as_Register(EAX_enc), "") ;
3254 MacroAssembler masm(&cbuf);
3256 if (_counters != NULL) {
3257 masm.atomic_incl(ExternalAddress((address) _counters->total_entry_count_addr()));
3258 }
3259 if (EmitSync & 1) {
3260 // set box->dhw = unused_mark (3)
3261 // Force all sync thru slow-path: slow_enter() and slow_exit()
3262 masm.movl (Address(boxReg, 0), intptr_t(markOopDesc::unused_mark())) ;
3263 masm.cmpl (rsp, 0) ;
3264 } else
3265 if (EmitSync & 2) {
3266 Label DONE_LABEL ;
3267 if (UseBiasedLocking) {
3268 // Note: tmpReg maps to the swap_reg argument and scrReg to the tmp_reg argument.
3269 masm.biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, _counters);
3270 }
3272 masm.movl (tmpReg, Address(objReg, 0)) ; // fetch markword
3273 masm.orl (tmpReg, 0x1);
3274 masm.movl (Address(boxReg, 0), tmpReg); // Anticipate successful CAS
3275 if (os::is_MP()) { masm.lock(); }
3276 masm.cmpxchg(boxReg, Address(objReg, 0)); // Updates tmpReg
3277 masm.jcc(Assembler::equal, DONE_LABEL);
3278 // Recursive locking
3279 masm.subl(tmpReg, rsp);
3280 masm.andl(tmpReg, 0xFFFFF003 );
3281 masm.movl(Address(boxReg, 0), tmpReg);
3282 masm.bind(DONE_LABEL) ;
3283 } else {
3284 // Possible cases that we'll encounter in fast_lock
3285 // ------------------------------------------------
3286 // * Inflated
3287 // -- unlocked
3288 // -- Locked
3289 // = by self
3290 // = by other
3291 // * biased
3292 // -- by Self
3293 // -- by other
3294 // * neutral
3295 // * stack-locked
3296 // -- by self
3297 // = sp-proximity test hits
3298 // = sp-proximity test generates false-negative
3299 // -- by other
3300 //
3302 Label IsInflated, DONE_LABEL, PopDone ;
3304 // TODO: optimize away redundant LDs of obj->mark and improve the markword triage
3305 // order to reduce the number of conditional branches in the most common cases.
3306 // Beware -- there's a subtle invariant that fetch of the markword
3307 // at [FETCH], below, will never observe a biased encoding (*101b).
3308 // If this invariant is not held we risk exclusion (safety) failure.
3309 if (UseBiasedLocking) {
3310 masm.biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, _counters);
3311 }
3313 masm.movl (tmpReg, Address(objReg, 0)) ; // [FETCH]
3314 masm.testl (tmpReg, 0x02) ; // Inflated v (Stack-locked or neutral)
3315 masm.jccb (Assembler::notZero, IsInflated) ;
3317 // Attempt stack-locking ...
3318 masm.orl (tmpReg, 0x1);
3319 masm.movl (Address(boxReg, 0), tmpReg); // Anticipate successful CAS
3320 if (os::is_MP()) { masm.lock(); }
3321 masm.cmpxchg(boxReg, Address(objReg, 0)); // Updates tmpReg
3322 if (_counters != NULL) {
3323 masm.cond_inc32(Assembler::equal,
3324 ExternalAddress((address)_counters->fast_path_entry_count_addr()));
3325 }
3326 masm.jccb (Assembler::equal, DONE_LABEL);
3328 // Recursive locking
3329 masm.subl(tmpReg, rsp);
3330 masm.andl(tmpReg, 0xFFFFF003 );
3331 masm.movl(Address(boxReg, 0), tmpReg);
3332 if (_counters != NULL) {
3333 masm.cond_inc32(Assembler::equal,
3334 ExternalAddress((address)_counters->fast_path_entry_count_addr()));
3335 }
3336 masm.jmp (DONE_LABEL) ;
3338 masm.bind (IsInflated) ;
3340 // The object is inflated.
3341 //
3342 // TODO-FIXME: eliminate the ugly use of manifest constants:
3343 // Use markOopDesc::monitor_value instead of "2".
3344 // use markOop::unused_mark() instead of "3".
3345 // The tmpReg value is an objectMonitor reference ORed with
3346 // markOopDesc::monitor_value (2). We can either convert tmpReg to an
3347 // objectmonitor pointer by masking off the "2" bit or we can just
3348 // use tmpReg as an objectmonitor pointer but bias the objectmonitor
3349 // field offsets with "-2" to compensate for and annul the low-order tag bit.
3350 //
3351 // I use the latter as it avoids AGI stalls.
3352 // As such, we write "mov r, [tmpReg+OFFSETOF(Owner)-2]"
3353 // instead of "mov r, [tmpReg+OFFSETOF(Owner)]".
3354 //
3355 #define OFFSET_SKEWED(f) ((ObjectMonitor::f ## _offset_in_bytes())-2)
3357 // boxReg refers to the on-stack BasicLock in the current frame.
3358 // We'd like to write:
3359 // set box->_displaced_header = markOop::unused_mark(). Any non-0 value suffices.
3360 // This is convenient but results a ST-before-CAS penalty. The following CAS suffers
3361 // additional latency as we have another ST in the store buffer that must drain.
3363 if (EmitSync & 8192) {
3364 masm.movl (Address(boxReg, 0), 3) ; // results in ST-before-CAS penalty
3365 masm.get_thread (scrReg) ;
3366 masm.movl (boxReg, tmpReg); // consider: LEA box, [tmp-2]
3367 masm.movl (tmpReg, 0); // consider: xor vs mov
3368 if (os::is_MP()) { masm.lock(); }
3369 masm.cmpxchg (scrReg, Address(boxReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
3370 } else
3371 if ((EmitSync & 128) == 0) { // avoid ST-before-CAS
3372 masm.movl (scrReg, boxReg) ;
3373 masm.movl (boxReg, tmpReg); // consider: LEA box, [tmp-2]
3375 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
3376 if ((EmitSync & 2048) && VM_Version::supports_3dnow() && os::is_MP()) {
3377 // prefetchw [eax + Offset(_owner)-2]
3378 masm.emit_raw (0x0F) ;
3379 masm.emit_raw (0x0D) ;
3380 masm.emit_raw (0x48) ;
3381 masm.emit_raw (ObjectMonitor::owner_offset_in_bytes()-2) ;
3382 }
3384 if ((EmitSync & 64) == 0) {
3385 // Optimistic form: consider XORL tmpReg,tmpReg
3386 masm.movl (tmpReg, 0 ) ;
3387 } else {
3388 // Can suffer RTS->RTO upgrades on shared or cold $ lines
3389 // Test-And-CAS instead of CAS
3390 masm.movl (tmpReg, Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ; // rax, = m->_owner
3391 masm.testl (tmpReg, tmpReg) ; // Locked ?
3392 masm.jccb (Assembler::notZero, DONE_LABEL) ;
3393 }
3395 // Appears unlocked - try to swing _owner from null to non-null.
3396 // Ideally, I'd manifest "Self" with get_thread and then attempt
3397 // to CAS the register containing Self into m->Owner.
3398 // But we don't have enough registers, so instead we can either try to CAS
3399 // rsp or the address of the box (in scr) into &m->owner. If the CAS succeeds
3400 // we later store "Self" into m->Owner. Transiently storing a stack address
3401 // (rsp or the address of the box) into m->owner is harmless.
3402 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
3403 if (os::is_MP()) { masm.lock(); }
3404 masm.cmpxchg (scrReg, Address(boxReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
3405 masm.movl (Address(scrReg, 0), 3) ; // box->_displaced_header = 3
3406 masm.jccb (Assembler::notZero, DONE_LABEL) ;
3407 masm.get_thread (scrReg) ; // beware: clobbers ICCs
3408 masm.movl (Address(boxReg, ObjectMonitor::owner_offset_in_bytes()-2), scrReg) ;
3409 masm.xorl (boxReg, boxReg) ; // set icc.ZFlag = 1 to indicate success
3411 // If the CAS fails we can either retry or pass control to the slow-path.
3412 // We use the latter tactic.
3413 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
3414 // If the CAS was successful ...
3415 // Self has acquired the lock
3416 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
3417 // Intentional fall-through into DONE_LABEL ...
3418 } else {
3419 masm.movl (Address(boxReg, 0), 3) ; // results in ST-before-CAS penalty
3420 masm.movl (boxReg, tmpReg) ;
3422 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
3423 if ((EmitSync & 2048) && VM_Version::supports_3dnow() && os::is_MP()) {
3424 // prefetchw [eax + Offset(_owner)-2]
3425 masm.emit_raw (0x0F) ;
3426 masm.emit_raw (0x0D) ;
3427 masm.emit_raw (0x48) ;
3428 masm.emit_raw (ObjectMonitor::owner_offset_in_bytes()-2) ;
3429 }
3431 if ((EmitSync & 64) == 0) {
3432 // Optimistic form
3433 masm.xorl (tmpReg, tmpReg) ;
3434 } else {
3435 // Can suffer RTS->RTO upgrades on shared or cold $ lines
3436 masm.movl (tmpReg, Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ; // rax, = m->_owner
3437 masm.testl (tmpReg, tmpReg) ; // Locked ?
3438 masm.jccb (Assembler::notZero, DONE_LABEL) ;
3439 }
3441 // Appears unlocked - try to swing _owner from null to non-null.
3442 // Use either "Self" (in scr) or rsp as thread identity in _owner.
3443 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
3444 masm.get_thread (scrReg) ;
3445 if (os::is_MP()) { masm.lock(); }
3446 masm.cmpxchg (scrReg, Address(boxReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
3448 // If the CAS fails we can either retry or pass control to the slow-path.
3449 // We use the latter tactic.
3450 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
3451 // If the CAS was successful ...
3452 // Self has acquired the lock
3453 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
3454 // Intentional fall-through into DONE_LABEL ...
3455 }
3457 // DONE_LABEL is a hot target - we'd really like to place it at the
3458 // start of cache line by padding with NOPs.
3459 // See the AMD and Intel software optimization manuals for the
3460 // most efficient "long" NOP encodings.
3461 // Unfortunately none of our alignment mechanisms suffice.
3462 masm.bind(DONE_LABEL);
3464 // Avoid branch-to-branch on AMD processors
3465 // This appears to be superstition.
3466 if (EmitSync & 32) masm.nop() ;
3469 // At DONE_LABEL the icc ZFlag is set as follows ...
3470 // Fast_Unlock uses the same protocol.
3471 // ZFlag == 1 -> Success
3472 // ZFlag == 0 -> Failure - force control through the slow-path
3473 }
3474 %}
3476 // obj: object to unlock
3477 // box: box address (displaced header location), killed. Must be EAX.
3478 // rbx,: killed tmp; cannot be obj nor box.
3479 //
3480 // Some commentary on balanced locking:
3481 //
3482 // Fast_Lock and Fast_Unlock are emitted only for provably balanced lock sites.
3483 // Methods that don't have provably balanced locking are forced to run in the
3484 // interpreter - such methods won't be compiled to use fast_lock and fast_unlock.
3485 // The interpreter provides two properties:
3486 // I1: At return-time the interpreter automatically and quietly unlocks any
3487 // objects acquired the current activation (frame). Recall that the
3488 // interpreter maintains an on-stack list of locks currently held by
3489 // a frame.
3490 // I2: If a method attempts to unlock an object that is not held by the
3491 // the frame the interpreter throws IMSX.
3492 //
3493 // Lets say A(), which has provably balanced locking, acquires O and then calls B().
3494 // B() doesn't have provably balanced locking so it runs in the interpreter.
3495 // Control returns to A() and A() unlocks O. By I1 and I2, above, we know that O
3496 // is still locked by A().
3497 //
3498 // The only other source of unbalanced locking would be JNI. The "Java Native Interface:
3499 // Programmer's Guide and Specification" claims that an object locked by jni_monitorenter
3500 // should not be unlocked by "normal" java-level locking and vice-versa. The specification
3501 // doesn't specify what will occur if a program engages in such mixed-mode locking, however.
3503 enc_class Fast_Unlock( nabxRegP obj, eAXRegP box, eRegP tmp) %{
3505 Register objReg = as_Register($obj$$reg);
3506 Register boxReg = as_Register($box$$reg);
3507 Register tmpReg = as_Register($tmp$$reg);
3509 guarantee (objReg != boxReg, "") ;
3510 guarantee (objReg != tmpReg, "") ;
3511 guarantee (boxReg != tmpReg, "") ;
3512 guarantee (boxReg == as_Register(EAX_enc), "") ;
3513 MacroAssembler masm(&cbuf);
3515 if (EmitSync & 4) {
3516 // Disable - inhibit all inlining. Force control through the slow-path
3517 masm.cmpl (rsp, 0) ;
3518 } else
3519 if (EmitSync & 8) {
3520 Label DONE_LABEL ;
3521 if (UseBiasedLocking) {
3522 masm.biased_locking_exit(objReg, tmpReg, DONE_LABEL);
3523 }
3524 // classic stack-locking code ...
3525 masm.movl (tmpReg, Address(boxReg, 0)) ;
3526 masm.testl (tmpReg, tmpReg) ;
3527 masm.jcc (Assembler::zero, DONE_LABEL) ;
3528 if (os::is_MP()) { masm.lock(); }
3529 masm.cmpxchg(tmpReg, Address(objReg, 0)); // Uses EAX which is box
3530 masm.bind(DONE_LABEL);
3531 } else {
3532 Label DONE_LABEL, Stacked, CheckSucc, Inflated ;
3534 // Critically, the biased locking test must have precedence over
3535 // and appear before the (box->dhw == 0) recursive stack-lock test.
3536 if (UseBiasedLocking) {
3537 masm.biased_locking_exit(objReg, tmpReg, DONE_LABEL);
3538 }
3540 masm.cmpl (Address(boxReg, 0), 0) ; // Examine the displaced header
3541 masm.movl (tmpReg, Address(objReg, 0)) ; // Examine the object's markword
3542 masm.jccb (Assembler::zero, DONE_LABEL) ; // 0 indicates recursive stack-lock
3544 masm.testl (tmpReg, 0x02) ; // Inflated?
3545 masm.jccb (Assembler::zero, Stacked) ;
3547 masm.bind (Inflated) ;
3548 // It's inflated.
3549 // Despite our balanced locking property we still check that m->_owner == Self
3550 // as java routines or native JNI code called by this thread might
3551 // have released the lock.
3552 // Refer to the comments in synchronizer.cpp for how we might encode extra
3553 // state in _succ so we can avoid fetching EntryList|cxq.
3554 //
3555 // I'd like to add more cases in fast_lock() and fast_unlock() --
3556 // such as recursive enter and exit -- but we have to be wary of
3557 // I$ bloat, T$ effects and BP$ effects.
3558 //
3559 // If there's no contention try a 1-0 exit. That is, exit without
3560 // a costly MEMBAR or CAS. See synchronizer.cpp for details on how
3561 // we detect and recover from the race that the 1-0 exit admits.
3562 //
3563 // Conceptually Fast_Unlock() must execute a STST|LDST "release" barrier
3564 // before it STs null into _owner, releasing the lock. Updates
3565 // to data protected by the critical section must be visible before
3566 // we drop the lock (and thus before any other thread could acquire
3567 // the lock and observe the fields protected by the lock).
3568 // IA32's memory-model is SPO, so STs are ordered with respect to
3569 // each other and there's no need for an explicit barrier (fence).
3570 // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
3572 masm.get_thread (boxReg) ;
3573 if ((EmitSync & 4096) && VM_Version::supports_3dnow() && os::is_MP()) {
3574 // prefetchw [ebx + Offset(_owner)-2]
3575 masm.emit_raw (0x0F) ;
3576 masm.emit_raw (0x0D) ;
3577 masm.emit_raw (0x4B) ;
3578 masm.emit_raw (ObjectMonitor::owner_offset_in_bytes()-2) ;
3579 }
3581 // Note that we could employ various encoding schemes to reduce
3582 // the number of loads below (currently 4) to just 2 or 3.
3583 // Refer to the comments in synchronizer.cpp.
3584 // In practice the chain of fetches doesn't seem to impact performance, however.
3585 if ((EmitSync & 65536) == 0 && (EmitSync & 256)) {
3586 // Attempt to reduce branch density - AMD's branch predictor.
3587 masm.xorl (boxReg, Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
3588 masm.orl (boxReg, Address (tmpReg, ObjectMonitor::recursions_offset_in_bytes()-2)) ;
3589 masm.orl (boxReg, Address (tmpReg, ObjectMonitor::EntryList_offset_in_bytes()-2)) ;
3590 masm.orl (boxReg, Address (tmpReg, ObjectMonitor::cxq_offset_in_bytes()-2)) ;
3591 masm.jccb (Assembler::notZero, DONE_LABEL) ;
3592 masm.movl (Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2), 0) ;
3593 masm.jmpb (DONE_LABEL) ;
3594 } else {
3595 masm.xorl (boxReg, Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
3596 masm.orl (boxReg, Address (tmpReg, ObjectMonitor::recursions_offset_in_bytes()-2)) ;
3597 masm.jccb (Assembler::notZero, DONE_LABEL) ;
3598 masm.movl (boxReg, Address (tmpReg, ObjectMonitor::EntryList_offset_in_bytes()-2)) ;
3599 masm.orl (boxReg, Address (tmpReg, ObjectMonitor::cxq_offset_in_bytes()-2)) ;
3600 masm.jccb (Assembler::notZero, CheckSucc) ;
3601 masm.movl (Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2), 0) ;
3602 masm.jmpb (DONE_LABEL) ;
3603 }
3605 // The Following code fragment (EmitSync & 65536) improves the performance of
3606 // contended applications and contended synchronization microbenchmarks.
3607 // Unfortunately the emission of the code - even though not executed - causes regressions
3608 // in scimark and jetstream, evidently because of $ effects. Replacing the code
3609 // with an equal number of never-executed NOPs results in the same regression.
3610 // We leave it off by default.
3612 if ((EmitSync & 65536) != 0) {
3613 Label LSuccess, LGoSlowPath ;
3615 masm.bind (CheckSucc) ;
3617 // Optional pre-test ... it's safe to elide this
3618 if ((EmitSync & 16) == 0) {
3619 masm.cmpl (Address (tmpReg, ObjectMonitor::succ_offset_in_bytes()-2), 0) ;
3620 masm.jccb (Assembler::zero, LGoSlowPath) ;
3621 }
3623 // We have a classic Dekker-style idiom:
3624 // ST m->_owner = 0 ; MEMBAR; LD m->_succ
3625 // There are a number of ways to implement the barrier:
3626 // (1) lock:andl &m->_owner, 0
3627 // is fast, but mask doesn't currently support the "ANDL M,IMM32" form.
3628 // LOCK: ANDL [ebx+Offset(_Owner)-2], 0
3629 // Encodes as 81 31 OFF32 IMM32 or 83 63 OFF8 IMM8
3630 // (2) If supported, an explicit MFENCE is appealing.
3631 // In older IA32 processors MFENCE is slower than lock:add or xchg
3632 // particularly if the write-buffer is full as might be the case if
3633 // if stores closely precede the fence or fence-equivalent instruction.
3634 // In more modern implementations MFENCE appears faster, however.
3635 // (3) In lieu of an explicit fence, use lock:addl to the top-of-stack
3636 // The $lines underlying the top-of-stack should be in M-state.
3637 // The locked add instruction is serializing, of course.
3638 // (4) Use xchg, which is serializing
3639 // mov boxReg, 0; xchgl boxReg, [tmpReg + Offset(_owner)-2] also works
3640 // (5) ST m->_owner = 0 and then execute lock:orl &m->_succ, 0.
3641 // The integer condition codes will tell us if succ was 0.
3642 // Since _succ and _owner should reside in the same $line and
3643 // we just stored into _owner, it's likely that the $line
3644 // remains in M-state for the lock:orl.
3645 //
3646 // We currently use (3), although it's likely that switching to (2)
3647 // is correct for the future.
3649 masm.movl (Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2), 0) ;
3650 if (os::is_MP()) {
3651 if (VM_Version::supports_sse2() && 1 == FenceInstruction) {
3652 masm.emit_raw (0x0F) ; // MFENCE ...
3653 masm.emit_raw (0xAE) ;
3654 masm.emit_raw (0xF0) ;
3655 } else {
3656 masm.lock () ; masm.addl (Address(rsp, 0), 0) ;
3657 }
3658 }
3659 // Ratify _succ remains non-null
3660 masm.cmpl (Address (tmpReg, ObjectMonitor::succ_offset_in_bytes()-2), 0) ;
3661 masm.jccb (Assembler::notZero, LSuccess) ;
3663 masm.xorl (boxReg, boxReg) ; // box is really EAX
3664 if (os::is_MP()) { masm.lock(); }
3665 masm.cmpxchg(rsp, Address(tmpReg, ObjectMonitor::owner_offset_in_bytes()-2));
3666 masm.jccb (Assembler::notEqual, LSuccess) ;
3667 // Since we're low on registers we installed rsp as a placeholding in _owner.
3668 // Now install Self over rsp. This is safe as we're transitioning from
3669 // non-null to non=null
3670 masm.get_thread (boxReg) ;
3671 masm.movl (Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2), boxReg) ;
3672 // Intentional fall-through into LGoSlowPath ...
3674 masm.bind (LGoSlowPath) ;
3675 masm.orl (boxReg, 1) ; // set ICC.ZF=0 to indicate failure
3676 masm.jmpb (DONE_LABEL) ;
3678 masm.bind (LSuccess) ;
3679 masm.xorl (boxReg, boxReg) ; // set ICC.ZF=1 to indicate success
3680 masm.jmpb (DONE_LABEL) ;
3681 }
3683 masm.bind (Stacked) ;
3684 // It's not inflated and it's not recursively stack-locked and it's not biased.
3685 // It must be stack-locked.
3686 // Try to reset the header to displaced header.
3687 // The "box" value on the stack is stable, so we can reload
3688 // and be assured we observe the same value as above.
3689 masm.movl (tmpReg, Address(boxReg, 0)) ;
3690 if (os::is_MP()) { masm.lock(); }
3691 masm.cmpxchg(tmpReg, Address(objReg, 0)); // Uses EAX which is box
3692 // Intention fall-thru into DONE_LABEL
3695 // DONE_LABEL is a hot target - we'd really like to place it at the
3696 // start of cache line by padding with NOPs.
3697 // See the AMD and Intel software optimization manuals for the
3698 // most efficient "long" NOP encodings.
3699 // Unfortunately none of our alignment mechanisms suffice.
3700 if ((EmitSync & 65536) == 0) {
3701 masm.bind (CheckSucc) ;
3702 }
3703 masm.bind(DONE_LABEL);
3705 // Avoid branch to branch on AMD processors
3706 if (EmitSync & 32768) { masm.nop() ; }
3707 }
3708 %}
3710 enc_class enc_String_Compare() %{
3711 Label ECX_GOOD_LABEL, LENGTH_DIFF_LABEL,
3712 POP_LABEL, DONE_LABEL, CONT_LABEL,
3713 WHILE_HEAD_LABEL;
3714 MacroAssembler masm(&cbuf);
3716 // Get the first character position in both strings
3717 // [8] char array, [12] offset, [16] count
3718 int value_offset = java_lang_String::value_offset_in_bytes();
3719 int offset_offset = java_lang_String::offset_offset_in_bytes();
3720 int count_offset = java_lang_String::count_offset_in_bytes();
3721 int base_offset = arrayOopDesc::base_offset_in_bytes(T_CHAR);
3723 masm.movl(rax, Address(rsi, value_offset));
3724 masm.movl(rcx, Address(rsi, offset_offset));
3725 masm.leal(rax, Address(rax, rcx, Address::times_2, base_offset));
3726 masm.movl(rbx, Address(rdi, value_offset));
3727 masm.movl(rcx, Address(rdi, offset_offset));
3728 masm.leal(rbx, Address(rbx, rcx, Address::times_2, base_offset));
3730 // Compute the minimum of the string lengths(rsi) and the
3731 // difference of the string lengths (stack)
3734 if (VM_Version::supports_cmov()) {
3735 masm.movl(rdi, Address(rdi, count_offset));
3736 masm.movl(rsi, Address(rsi, count_offset));
3737 masm.movl(rcx, rdi);
3738 masm.subl(rdi, rsi);
3739 masm.pushl(rdi);
3740 masm.cmovl(Assembler::lessEqual, rsi, rcx);
3741 } else {
3742 masm.movl(rdi, Address(rdi, count_offset));
3743 masm.movl(rcx, Address(rsi, count_offset));
3744 masm.movl(rsi, rdi);
3745 masm.subl(rdi, rcx);
3746 masm.pushl(rdi);
3747 masm.jcc(Assembler::lessEqual, ECX_GOOD_LABEL);
3748 masm.movl(rsi, rcx);
3749 // rsi holds min, rcx is unused
3750 }
3752 // Is the minimum length zero?
3753 masm.bind(ECX_GOOD_LABEL);
3754 masm.testl(rsi, rsi);
3755 masm.jcc(Assembler::zero, LENGTH_DIFF_LABEL);
3757 // Load first characters
3758 masm.load_unsigned_word(rcx, Address(rbx, 0));
3759 masm.load_unsigned_word(rdi, Address(rax, 0));
3761 // Compare first characters
3762 masm.subl(rcx, rdi);
3763 masm.jcc(Assembler::notZero, POP_LABEL);
3764 masm.decrement(rsi);
3765 masm.jcc(Assembler::zero, LENGTH_DIFF_LABEL);
3767 {
3768 // Check after comparing first character to see if strings are equivalent
3769 Label LSkip2;
3770 // Check if the strings start at same location
3771 masm.cmpl(rbx,rax);
3772 masm.jcc(Assembler::notEqual, LSkip2);
3774 // Check if the length difference is zero (from stack)
3775 masm.cmpl(Address(rsp, 0), 0x0);
3776 masm.jcc(Assembler::equal, LENGTH_DIFF_LABEL);
3778 // Strings might not be equivalent
3779 masm.bind(LSkip2);
3780 }
3782 // Shift rax, and rbx, to the end of the arrays, negate min
3783 masm.leal(rax, Address(rax, rsi, Address::times_2, 2));
3784 masm.leal(rbx, Address(rbx, rsi, Address::times_2, 2));
3785 masm.negl(rsi);
3787 // Compare the rest of the characters
3788 masm.bind(WHILE_HEAD_LABEL);
3789 masm.load_unsigned_word(rcx, Address(rbx, rsi, Address::times_2, 0));
3790 masm.load_unsigned_word(rdi, Address(rax, rsi, Address::times_2, 0));
3791 masm.subl(rcx, rdi);
3792 masm.jcc(Assembler::notZero, POP_LABEL);
3793 masm.increment(rsi);
3794 masm.jcc(Assembler::notZero, WHILE_HEAD_LABEL);
3796 // Strings are equal up to min length. Return the length difference.
3797 masm.bind(LENGTH_DIFF_LABEL);
3798 masm.popl(rcx);
3799 masm.jmp(DONE_LABEL);
3801 // Discard the stored length difference
3802 masm.bind(POP_LABEL);
3803 masm.addl(rsp, 4);
3805 // That's it
3806 masm.bind(DONE_LABEL);
3807 %}
3809 enc_class enc_pop_rdx() %{
3810 emit_opcode(cbuf,0x5A);
3811 %}
3813 enc_class enc_rethrow() %{
3814 cbuf.set_inst_mark();
3815 emit_opcode(cbuf, 0xE9); // jmp entry
3816 emit_d32_reloc(cbuf, (int)OptoRuntime::rethrow_stub() - ((int)cbuf.code_end())-4,
3817 runtime_call_Relocation::spec(), RELOC_IMM32 );
3818 %}
3821 // Convert a double to an int. Java semantics require we do complex
3822 // manglelations in the corner cases. So we set the rounding mode to
3823 // 'zero', store the darned double down as an int, and reset the
3824 // rounding mode to 'nearest'. The hardware throws an exception which
3825 // patches up the correct value directly to the stack.
3826 enc_class D2I_encoding( regD src ) %{
3827 // Flip to round-to-zero mode. We attempted to allow invalid-op
3828 // exceptions here, so that a NAN or other corner-case value will
3829 // thrown an exception (but normal values get converted at full speed).
3830 // However, I2C adapters and other float-stack manglers leave pending
3831 // invalid-op exceptions hanging. We would have to clear them before
3832 // enabling them and that is more expensive than just testing for the
3833 // invalid value Intel stores down in the corner cases.
3834 emit_opcode(cbuf,0xD9); // FLDCW trunc
3835 emit_opcode(cbuf,0x2D);
3836 emit_d32(cbuf,(int)StubRoutines::addr_fpu_cntrl_wrd_trunc());
3837 // Allocate a word
3838 emit_opcode(cbuf,0x83); // SUB ESP,4
3839 emit_opcode(cbuf,0xEC);
3840 emit_d8(cbuf,0x04);
3841 // Encoding assumes a double has been pushed into FPR0.
3842 // Store down the double as an int, popping the FPU stack
3843 emit_opcode(cbuf,0xDB); // FISTP [ESP]
3844 emit_opcode(cbuf,0x1C);
3845 emit_d8(cbuf,0x24);
3846 // Restore the rounding mode; mask the exception
3847 emit_opcode(cbuf,0xD9); // FLDCW std/24-bit mode
3848 emit_opcode(cbuf,0x2D);
3849 emit_d32( cbuf, Compile::current()->in_24_bit_fp_mode()
3850 ? (int)StubRoutines::addr_fpu_cntrl_wrd_24()
3851 : (int)StubRoutines::addr_fpu_cntrl_wrd_std());
3853 // Load the converted int; adjust CPU stack
3854 emit_opcode(cbuf,0x58); // POP EAX
3855 emit_opcode(cbuf,0x3D); // CMP EAX,imm
3856 emit_d32 (cbuf,0x80000000); // 0x80000000
3857 emit_opcode(cbuf,0x75); // JNE around_slow_call
3858 emit_d8 (cbuf,0x07); // Size of slow_call
3859 // Push src onto stack slow-path
3860 emit_opcode(cbuf,0xD9 ); // FLD ST(i)
3861 emit_d8 (cbuf,0xC0-1+$src$$reg );
3862 // CALL directly to the runtime
3863 cbuf.set_inst_mark();
3864 emit_opcode(cbuf,0xE8); // Call into runtime
3865 emit_d32_reloc(cbuf, (StubRoutines::d2i_wrapper() - cbuf.code_end()) - 4, runtime_call_Relocation::spec(), RELOC_IMM32 );
3866 // Carry on here...
3867 %}
3869 enc_class D2L_encoding( regD src ) %{
3870 emit_opcode(cbuf,0xD9); // FLDCW trunc
3871 emit_opcode(cbuf,0x2D);
3872 emit_d32(cbuf,(int)StubRoutines::addr_fpu_cntrl_wrd_trunc());
3873 // Allocate a word
3874 emit_opcode(cbuf,0x83); // SUB ESP,8
3875 emit_opcode(cbuf,0xEC);
3876 emit_d8(cbuf,0x08);
3877 // Encoding assumes a double has been pushed into FPR0.
3878 // Store down the double as a long, popping the FPU stack
3879 emit_opcode(cbuf,0xDF); // FISTP [ESP]
3880 emit_opcode(cbuf,0x3C);
3881 emit_d8(cbuf,0x24);
3882 // Restore the rounding mode; mask the exception
3883 emit_opcode(cbuf,0xD9); // FLDCW std/24-bit mode
3884 emit_opcode(cbuf,0x2D);
3885 emit_d32( cbuf, Compile::current()->in_24_bit_fp_mode()
3886 ? (int)StubRoutines::addr_fpu_cntrl_wrd_24()
3887 : (int)StubRoutines::addr_fpu_cntrl_wrd_std());
3889 // Load the converted int; adjust CPU stack
3890 emit_opcode(cbuf,0x58); // POP EAX
3891 emit_opcode(cbuf,0x5A); // POP EDX
3892 emit_opcode(cbuf,0x81); // CMP EDX,imm
3893 emit_d8 (cbuf,0xFA); // rdx
3894 emit_d32 (cbuf,0x80000000); // 0x80000000
3895 emit_opcode(cbuf,0x75); // JNE around_slow_call
3896 emit_d8 (cbuf,0x07+4); // Size of slow_call
3897 emit_opcode(cbuf,0x85); // TEST EAX,EAX
3898 emit_opcode(cbuf,0xC0); // 2/rax,/rax,
3899 emit_opcode(cbuf,0x75); // JNE around_slow_call
3900 emit_d8 (cbuf,0x07); // Size of slow_call
3901 // Push src onto stack slow-path
3902 emit_opcode(cbuf,0xD9 ); // FLD ST(i)
3903 emit_d8 (cbuf,0xC0-1+$src$$reg );
3904 // CALL directly to the runtime
3905 cbuf.set_inst_mark();
3906 emit_opcode(cbuf,0xE8); // Call into runtime
3907 emit_d32_reloc(cbuf, (StubRoutines::d2l_wrapper() - cbuf.code_end()) - 4, runtime_call_Relocation::spec(), RELOC_IMM32 );
3908 // Carry on here...
3909 %}
3911 enc_class X2L_encoding( regX src ) %{
3912 // Allocate a word
3913 emit_opcode(cbuf,0x83); // SUB ESP,8
3914 emit_opcode(cbuf,0xEC);
3915 emit_d8(cbuf,0x08);
3917 emit_opcode (cbuf, 0xF3 ); // MOVSS [ESP], src
3918 emit_opcode (cbuf, 0x0F );
3919 emit_opcode (cbuf, 0x11 );
3920 encode_RegMem(cbuf, $src$$reg, ESP_enc, 0x4, 0, 0, false);
3922 emit_opcode(cbuf,0xD9 ); // FLD_S [ESP]
3923 encode_RegMem(cbuf, 0x0, ESP_enc, 0x4, 0, 0, false);
3925 emit_opcode(cbuf,0xD9); // FLDCW trunc
3926 emit_opcode(cbuf,0x2D);
3927 emit_d32(cbuf,(int)StubRoutines::addr_fpu_cntrl_wrd_trunc());
3929 // Encoding assumes a double has been pushed into FPR0.
3930 // Store down the double as a long, popping the FPU stack
3931 emit_opcode(cbuf,0xDF); // FISTP [ESP]
3932 emit_opcode(cbuf,0x3C);
3933 emit_d8(cbuf,0x24);
3935 // Restore the rounding mode; mask the exception
3936 emit_opcode(cbuf,0xD9); // FLDCW std/24-bit mode
3937 emit_opcode(cbuf,0x2D);
3938 emit_d32( cbuf, Compile::current()->in_24_bit_fp_mode()
3939 ? (int)StubRoutines::addr_fpu_cntrl_wrd_24()
3940 : (int)StubRoutines::addr_fpu_cntrl_wrd_std());
3942 // Load the converted int; adjust CPU stack
3943 emit_opcode(cbuf,0x58); // POP EAX
3945 emit_opcode(cbuf,0x5A); // POP EDX
3947 emit_opcode(cbuf,0x81); // CMP EDX,imm
3948 emit_d8 (cbuf,0xFA); // rdx
3949 emit_d32 (cbuf,0x80000000);// 0x80000000
3951 emit_opcode(cbuf,0x75); // JNE around_slow_call
3952 emit_d8 (cbuf,0x13+4); // Size of slow_call
3954 emit_opcode(cbuf,0x85); // TEST EAX,EAX
3955 emit_opcode(cbuf,0xC0); // 2/rax,/rax,
3957 emit_opcode(cbuf,0x75); // JNE around_slow_call
3958 emit_d8 (cbuf,0x13); // Size of slow_call
3960 // Allocate a word
3961 emit_opcode(cbuf,0x83); // SUB ESP,4
3962 emit_opcode(cbuf,0xEC);
3963 emit_d8(cbuf,0x04);
3965 emit_opcode (cbuf, 0xF3 ); // MOVSS [ESP], src
3966 emit_opcode (cbuf, 0x0F );
3967 emit_opcode (cbuf, 0x11 );
3968 encode_RegMem(cbuf, $src$$reg, ESP_enc, 0x4, 0, 0, false);
3970 emit_opcode(cbuf,0xD9 ); // FLD_S [ESP]
3971 encode_RegMem(cbuf, 0x0, ESP_enc, 0x4, 0, 0, false);
3973 emit_opcode(cbuf,0x83); // ADD ESP,4
3974 emit_opcode(cbuf,0xC4);
3975 emit_d8(cbuf,0x04);
3977 // CALL directly to the runtime
3978 cbuf.set_inst_mark();
3979 emit_opcode(cbuf,0xE8); // Call into runtime
3980 emit_d32_reloc(cbuf, (StubRoutines::d2l_wrapper() - cbuf.code_end()) - 4, runtime_call_Relocation::spec(), RELOC_IMM32 );
3981 // Carry on here...
3982 %}
3984 enc_class XD2L_encoding( regXD src ) %{
3985 // Allocate a word
3986 emit_opcode(cbuf,0x83); // SUB ESP,8
3987 emit_opcode(cbuf,0xEC);
3988 emit_d8(cbuf,0x08);
3990 emit_opcode (cbuf, 0xF2 ); // MOVSD [ESP], src
3991 emit_opcode (cbuf, 0x0F );
3992 emit_opcode (cbuf, 0x11 );
3993 encode_RegMem(cbuf, $src$$reg, ESP_enc, 0x4, 0, 0, false);
3995 emit_opcode(cbuf,0xDD ); // FLD_D [ESP]
3996 encode_RegMem(cbuf, 0x0, ESP_enc, 0x4, 0, 0, false);
3998 emit_opcode(cbuf,0xD9); // FLDCW trunc
3999 emit_opcode(cbuf,0x2D);
4000 emit_d32(cbuf,(int)StubRoutines::addr_fpu_cntrl_wrd_trunc());
4002 // Encoding assumes a double has been pushed into FPR0.
4003 // Store down the double as a long, popping the FPU stack
4004 emit_opcode(cbuf,0xDF); // FISTP [ESP]
4005 emit_opcode(cbuf,0x3C);
4006 emit_d8(cbuf,0x24);
4008 // Restore the rounding mode; mask the exception
4009 emit_opcode(cbuf,0xD9); // FLDCW std/24-bit mode
4010 emit_opcode(cbuf,0x2D);
4011 emit_d32( cbuf, Compile::current()->in_24_bit_fp_mode()
4012 ? (int)StubRoutines::addr_fpu_cntrl_wrd_24()
4013 : (int)StubRoutines::addr_fpu_cntrl_wrd_std());
4015 // Load the converted int; adjust CPU stack
4016 emit_opcode(cbuf,0x58); // POP EAX
4018 emit_opcode(cbuf,0x5A); // POP EDX
4020 emit_opcode(cbuf,0x81); // CMP EDX,imm
4021 emit_d8 (cbuf,0xFA); // rdx
4022 emit_d32 (cbuf,0x80000000); // 0x80000000
4024 emit_opcode(cbuf,0x75); // JNE around_slow_call
4025 emit_d8 (cbuf,0x13+4); // Size of slow_call
4027 emit_opcode(cbuf,0x85); // TEST EAX,EAX
4028 emit_opcode(cbuf,0xC0); // 2/rax,/rax,
4030 emit_opcode(cbuf,0x75); // JNE around_slow_call
4031 emit_d8 (cbuf,0x13); // Size of slow_call
4033 // Push src onto stack slow-path
4034 // Allocate a word
4035 emit_opcode(cbuf,0x83); // SUB ESP,8
4036 emit_opcode(cbuf,0xEC);
4037 emit_d8(cbuf,0x08);
4039 emit_opcode (cbuf, 0xF2 ); // MOVSD [ESP], src
4040 emit_opcode (cbuf, 0x0F );
4041 emit_opcode (cbuf, 0x11 );
4042 encode_RegMem(cbuf, $src$$reg, ESP_enc, 0x4, 0, 0, false);
4044 emit_opcode(cbuf,0xDD ); // FLD_D [ESP]
4045 encode_RegMem(cbuf, 0x0, ESP_enc, 0x4, 0, 0, false);
4047 emit_opcode(cbuf,0x83); // ADD ESP,8
4048 emit_opcode(cbuf,0xC4);
4049 emit_d8(cbuf,0x08);
4051 // CALL directly to the runtime
4052 cbuf.set_inst_mark();
4053 emit_opcode(cbuf,0xE8); // Call into runtime
4054 emit_d32_reloc(cbuf, (StubRoutines::d2l_wrapper() - cbuf.code_end()) - 4, runtime_call_Relocation::spec(), RELOC_IMM32 );
4055 // Carry on here...
4056 %}
4058 enc_class D2X_encoding( regX dst, regD src ) %{
4059 // Allocate a word
4060 emit_opcode(cbuf,0x83); // SUB ESP,4
4061 emit_opcode(cbuf,0xEC);
4062 emit_d8(cbuf,0x04);
4063 int pop = 0x02;
4064 if ($src$$reg != FPR1L_enc) {
4065 emit_opcode( cbuf, 0xD9 ); // FLD ST(i-1)
4066 emit_d8( cbuf, 0xC0-1+$src$$reg );
4067 pop = 0x03;
4068 }
4069 store_to_stackslot( cbuf, 0xD9, pop, 0 ); // FST<P>_S [ESP]
4071 emit_opcode (cbuf, 0xF3 ); // MOVSS dst(xmm), [ESP]
4072 emit_opcode (cbuf, 0x0F );
4073 emit_opcode (cbuf, 0x10 );
4074 encode_RegMem(cbuf, $dst$$reg, ESP_enc, 0x4, 0, 0, false);
4076 emit_opcode(cbuf,0x83); // ADD ESP,4
4077 emit_opcode(cbuf,0xC4);
4078 emit_d8(cbuf,0x04);
4079 // Carry on here...
4080 %}
4082 enc_class FX2I_encoding( regX src, eRegI dst ) %{
4083 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
4085 // Compare the result to see if we need to go to the slow path
4086 emit_opcode(cbuf,0x81); // CMP dst,imm
4087 emit_rm (cbuf,0x3,0x7,$dst$$reg);
4088 emit_d32 (cbuf,0x80000000); // 0x80000000
4090 emit_opcode(cbuf,0x75); // JNE around_slow_call
4091 emit_d8 (cbuf,0x13); // Size of slow_call
4092 // Store xmm to a temp memory
4093 // location and push it onto stack.
4095 emit_opcode(cbuf,0x83); // SUB ESP,4
4096 emit_opcode(cbuf,0xEC);
4097 emit_d8(cbuf, $primary ? 0x8 : 0x4);
4099 emit_opcode (cbuf, $primary ? 0xF2 : 0xF3 ); // MOVSS [ESP], xmm
4100 emit_opcode (cbuf, 0x0F );
4101 emit_opcode (cbuf, 0x11 );
4102 encode_RegMem(cbuf, $src$$reg, ESP_enc, 0x4, 0, 0, false);
4104 emit_opcode(cbuf, $primary ? 0xDD : 0xD9 ); // FLD [ESP]
4105 encode_RegMem(cbuf, 0x0, ESP_enc, 0x4, 0, 0, false);
4107 emit_opcode(cbuf,0x83); // ADD ESP,4
4108 emit_opcode(cbuf,0xC4);
4109 emit_d8(cbuf, $primary ? 0x8 : 0x4);
4111 // CALL directly to the runtime
4112 cbuf.set_inst_mark();
4113 emit_opcode(cbuf,0xE8); // Call into runtime
4114 emit_d32_reloc(cbuf, (StubRoutines::d2i_wrapper() - cbuf.code_end()) - 4, runtime_call_Relocation::spec(), RELOC_IMM32 );
4116 // Carry on here...
4117 %}
4119 enc_class X2D_encoding( regD dst, regX src ) %{
4120 // Allocate a word
4121 emit_opcode(cbuf,0x83); // SUB ESP,4
4122 emit_opcode(cbuf,0xEC);
4123 emit_d8(cbuf,0x04);
4125 emit_opcode (cbuf, 0xF3 ); // MOVSS [ESP], xmm
4126 emit_opcode (cbuf, 0x0F );
4127 emit_opcode (cbuf, 0x11 );
4128 encode_RegMem(cbuf, $src$$reg, ESP_enc, 0x4, 0, 0, false);
4130 emit_opcode(cbuf,0xD9 ); // FLD_S [ESP]
4131 encode_RegMem(cbuf, 0x0, ESP_enc, 0x4, 0, 0, false);
4133 emit_opcode(cbuf,0x83); // ADD ESP,4
4134 emit_opcode(cbuf,0xC4);
4135 emit_d8(cbuf,0x04);
4137 // Carry on here...
4138 %}
4140 enc_class AbsXF_encoding(regX dst) %{
4141 address signmask_address=(address)float_signmask_pool;
4142 // andpd:\tANDPS $dst,[signconst]
4143 emit_opcode(cbuf, 0x0F);
4144 emit_opcode(cbuf, 0x54);
4145 emit_rm(cbuf, 0x0, $dst$$reg, 0x5);
4146 emit_d32(cbuf, (int)signmask_address);
4147 %}
4149 enc_class AbsXD_encoding(regXD dst) %{
4150 address signmask_address=(address)double_signmask_pool;
4151 // andpd:\tANDPD $dst,[signconst]
4152 emit_opcode(cbuf, 0x66);
4153 emit_opcode(cbuf, 0x0F);
4154 emit_opcode(cbuf, 0x54);
4155 emit_rm(cbuf, 0x0, $dst$$reg, 0x5);
4156 emit_d32(cbuf, (int)signmask_address);
4157 %}
4159 enc_class NegXF_encoding(regX dst) %{
4160 address signmask_address=(address)float_signflip_pool;
4161 // andpd:\tXORPS $dst,[signconst]
4162 emit_opcode(cbuf, 0x0F);
4163 emit_opcode(cbuf, 0x57);
4164 emit_rm(cbuf, 0x0, $dst$$reg, 0x5);
4165 emit_d32(cbuf, (int)signmask_address);
4166 %}
4168 enc_class NegXD_encoding(regXD dst) %{
4169 address signmask_address=(address)double_signflip_pool;
4170 // andpd:\tXORPD $dst,[signconst]
4171 emit_opcode(cbuf, 0x66);
4172 emit_opcode(cbuf, 0x0F);
4173 emit_opcode(cbuf, 0x57);
4174 emit_rm(cbuf, 0x0, $dst$$reg, 0x5);
4175 emit_d32(cbuf, (int)signmask_address);
4176 %}
4178 enc_class FMul_ST_reg( eRegF src1 ) %{
4179 // Operand was loaded from memory into fp ST (stack top)
4180 // FMUL ST,$src /* D8 C8+i */
4181 emit_opcode(cbuf, 0xD8);
4182 emit_opcode(cbuf, 0xC8 + $src1$$reg);
4183 %}
4185 enc_class FAdd_ST_reg( eRegF src2 ) %{
4186 // FADDP ST,src2 /* D8 C0+i */
4187 emit_opcode(cbuf, 0xD8);
4188 emit_opcode(cbuf, 0xC0 + $src2$$reg);
4189 //could use FADDP src2,fpST /* DE C0+i */
4190 %}
4192 enc_class FAddP_reg_ST( eRegF src2 ) %{
4193 // FADDP src2,ST /* DE C0+i */
4194 emit_opcode(cbuf, 0xDE);
4195 emit_opcode(cbuf, 0xC0 + $src2$$reg);
4196 %}
4198 enc_class subF_divF_encode( eRegF src1, eRegF src2) %{
4199 // Operand has been loaded into fp ST (stack top)
4200 // FSUB ST,$src1
4201 emit_opcode(cbuf, 0xD8);
4202 emit_opcode(cbuf, 0xE0 + $src1$$reg);
4204 // FDIV
4205 emit_opcode(cbuf, 0xD8);
4206 emit_opcode(cbuf, 0xF0 + $src2$$reg);
4207 %}
4209 enc_class MulFAddF (eRegF src1, eRegF src2) %{
4210 // Operand was loaded from memory into fp ST (stack top)
4211 // FADD ST,$src /* D8 C0+i */
4212 emit_opcode(cbuf, 0xD8);
4213 emit_opcode(cbuf, 0xC0 + $src1$$reg);
4215 // FMUL ST,src2 /* D8 C*+i */
4216 emit_opcode(cbuf, 0xD8);
4217 emit_opcode(cbuf, 0xC8 + $src2$$reg);
4218 %}
4221 enc_class MulFAddFreverse (eRegF src1, eRegF src2) %{
4222 // Operand was loaded from memory into fp ST (stack top)
4223 // FADD ST,$src /* D8 C0+i */
4224 emit_opcode(cbuf, 0xD8);
4225 emit_opcode(cbuf, 0xC0 + $src1$$reg);
4227 // FMULP src2,ST /* DE C8+i */
4228 emit_opcode(cbuf, 0xDE);
4229 emit_opcode(cbuf, 0xC8 + $src2$$reg);
4230 %}
4232 enc_class enc_membar_acquire %{
4233 // Doug Lea believes this is not needed with current Sparcs and TSO.
4234 // MacroAssembler masm(&cbuf);
4235 // masm.membar();
4236 %}
4238 enc_class enc_membar_release %{
4239 // Doug Lea believes this is not needed with current Sparcs and TSO.
4240 // MacroAssembler masm(&cbuf);
4241 // masm.membar();
4242 %}
4244 enc_class enc_membar_volatile %{
4245 MacroAssembler masm(&cbuf);
4246 masm.membar();
4247 %}
4249 // Atomically load the volatile long
4250 enc_class enc_loadL_volatile( memory mem, stackSlotL dst ) %{
4251 emit_opcode(cbuf,0xDF);
4252 int rm_byte_opcode = 0x05;
4253 int base = $mem$$base;
4254 int index = $mem$$index;
4255 int scale = $mem$$scale;
4256 int displace = $mem$$disp;
4257 bool disp_is_oop = $mem->disp_is_oop(); // disp-as-oop when working with static globals
4258 encode_RegMem(cbuf, rm_byte_opcode, base, index, scale, displace, disp_is_oop);
4259 store_to_stackslot( cbuf, 0x0DF, 0x07, $dst$$disp );
4260 %}
4262 enc_class enc_loadLX_volatile( memory mem, stackSlotL dst, regXD tmp ) %{
4263 { // Atomic long load
4264 // UseXmmLoadAndClearUpper ? movsd $tmp,$mem : movlpd $tmp,$mem
4265 emit_opcode(cbuf,UseXmmLoadAndClearUpper ? 0xF2 : 0x66);
4266 emit_opcode(cbuf,0x0F);
4267 emit_opcode(cbuf,UseXmmLoadAndClearUpper ? 0x10 : 0x12);
4268 int base = $mem$$base;
4269 int index = $mem$$index;
4270 int scale = $mem$$scale;
4271 int displace = $mem$$disp;
4272 bool disp_is_oop = $mem->disp_is_oop(); // disp-as-oop when working with static globals
4273 encode_RegMem(cbuf, $tmp$$reg, base, index, scale, displace, disp_is_oop);
4274 }
4275 { // MOVSD $dst,$tmp ! atomic long store
4276 emit_opcode(cbuf,0xF2);
4277 emit_opcode(cbuf,0x0F);
4278 emit_opcode(cbuf,0x11);
4279 int base = $dst$$base;
4280 int index = $dst$$index;
4281 int scale = $dst$$scale;
4282 int displace = $dst$$disp;
4283 bool disp_is_oop = $dst->disp_is_oop(); // disp-as-oop when working with static globals
4284 encode_RegMem(cbuf, $tmp$$reg, base, index, scale, displace, disp_is_oop);
4285 }
4286 %}
4288 enc_class enc_loadLX_reg_volatile( memory mem, eRegL dst, regXD tmp ) %{
4289 { // Atomic long load
4290 // UseXmmLoadAndClearUpper ? movsd $tmp,$mem : movlpd $tmp,$mem
4291 emit_opcode(cbuf,UseXmmLoadAndClearUpper ? 0xF2 : 0x66);
4292 emit_opcode(cbuf,0x0F);
4293 emit_opcode(cbuf,UseXmmLoadAndClearUpper ? 0x10 : 0x12);
4294 int base = $mem$$base;
4295 int index = $mem$$index;
4296 int scale = $mem$$scale;
4297 int displace = $mem$$disp;
4298 bool disp_is_oop = $mem->disp_is_oop(); // disp-as-oop when working with static globals
4299 encode_RegMem(cbuf, $tmp$$reg, base, index, scale, displace, disp_is_oop);
4300 }
4301 { // MOVD $dst.lo,$tmp
4302 emit_opcode(cbuf,0x66);
4303 emit_opcode(cbuf,0x0F);
4304 emit_opcode(cbuf,0x7E);
4305 emit_rm(cbuf, 0x3, $tmp$$reg, $dst$$reg);
4306 }
4307 { // PSRLQ $tmp,32
4308 emit_opcode(cbuf,0x66);
4309 emit_opcode(cbuf,0x0F);
4310 emit_opcode(cbuf,0x73);
4311 emit_rm(cbuf, 0x3, 0x02, $tmp$$reg);
4312 emit_d8(cbuf, 0x20);
4313 }
4314 { // MOVD $dst.hi,$tmp
4315 emit_opcode(cbuf,0x66);
4316 emit_opcode(cbuf,0x0F);
4317 emit_opcode(cbuf,0x7E);
4318 emit_rm(cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($dst$$reg));
4319 }
4320 %}
4322 // Volatile Store Long. Must be atomic, so move it into
4323 // the FP TOS and then do a 64-bit FIST. Has to probe the
4324 // target address before the store (for null-ptr checks)
4325 // so the memory operand is used twice in the encoding.
4326 enc_class enc_storeL_volatile( memory mem, stackSlotL src ) %{
4327 store_to_stackslot( cbuf, 0x0DF, 0x05, $src$$disp );
4328 cbuf.set_inst_mark(); // Mark start of FIST in case $mem has an oop
4329 emit_opcode(cbuf,0xDF);
4330 int rm_byte_opcode = 0x07;
4331 int base = $mem$$base;
4332 int index = $mem$$index;
4333 int scale = $mem$$scale;
4334 int displace = $mem$$disp;
4335 bool disp_is_oop = $mem->disp_is_oop(); // disp-as-oop when working with static globals
4336 encode_RegMem(cbuf, rm_byte_opcode, base, index, scale, displace, disp_is_oop);
4337 %}
4339 enc_class enc_storeLX_volatile( memory mem, stackSlotL src, regXD tmp) %{
4340 { // Atomic long load
4341 // UseXmmLoadAndClearUpper ? movsd $tmp,[$src] : movlpd $tmp,[$src]
4342 emit_opcode(cbuf,UseXmmLoadAndClearUpper ? 0xF2 : 0x66);
4343 emit_opcode(cbuf,0x0F);
4344 emit_opcode(cbuf,UseXmmLoadAndClearUpper ? 0x10 : 0x12);
4345 int base = $src$$base;
4346 int index = $src$$index;
4347 int scale = $src$$scale;
4348 int displace = $src$$disp;
4349 bool disp_is_oop = $src->disp_is_oop(); // disp-as-oop when working with static globals
4350 encode_RegMem(cbuf, $tmp$$reg, base, index, scale, displace, disp_is_oop);
4351 }
4352 cbuf.set_inst_mark(); // Mark start of MOVSD in case $mem has an oop
4353 { // MOVSD $mem,$tmp ! atomic long store
4354 emit_opcode(cbuf,0xF2);
4355 emit_opcode(cbuf,0x0F);
4356 emit_opcode(cbuf,0x11);
4357 int base = $mem$$base;
4358 int index = $mem$$index;
4359 int scale = $mem$$scale;
4360 int displace = $mem$$disp;
4361 bool disp_is_oop = $mem->disp_is_oop(); // disp-as-oop when working with static globals
4362 encode_RegMem(cbuf, $tmp$$reg, base, index, scale, displace, disp_is_oop);
4363 }
4364 %}
4366 enc_class enc_storeLX_reg_volatile( memory mem, eRegL src, regXD tmp, regXD tmp2) %{
4367 { // MOVD $tmp,$src.lo
4368 emit_opcode(cbuf,0x66);
4369 emit_opcode(cbuf,0x0F);
4370 emit_opcode(cbuf,0x6E);
4371 emit_rm(cbuf, 0x3, $tmp$$reg, $src$$reg);
4372 }
4373 { // MOVD $tmp2,$src.hi
4374 emit_opcode(cbuf,0x66);
4375 emit_opcode(cbuf,0x0F);
4376 emit_opcode(cbuf,0x6E);
4377 emit_rm(cbuf, 0x3, $tmp2$$reg, HIGH_FROM_LOW($src$$reg));
4378 }
4379 { // PUNPCKLDQ $tmp,$tmp2
4380 emit_opcode(cbuf,0x66);
4381 emit_opcode(cbuf,0x0F);
4382 emit_opcode(cbuf,0x62);
4383 emit_rm(cbuf, 0x3, $tmp$$reg, $tmp2$$reg);
4384 }
4385 cbuf.set_inst_mark(); // Mark start of MOVSD in case $mem has an oop
4386 { // MOVSD $mem,$tmp ! atomic long store
4387 emit_opcode(cbuf,0xF2);
4388 emit_opcode(cbuf,0x0F);
4389 emit_opcode(cbuf,0x11);
4390 int base = $mem$$base;
4391 int index = $mem$$index;
4392 int scale = $mem$$scale;
4393 int displace = $mem$$disp;
4394 bool disp_is_oop = $mem->disp_is_oop(); // disp-as-oop when working with static globals
4395 encode_RegMem(cbuf, $tmp$$reg, base, index, scale, displace, disp_is_oop);
4396 }
4397 %}
4399 // Safepoint Poll. This polls the safepoint page, and causes an
4400 // exception if it is not readable. Unfortunately, it kills the condition code
4401 // in the process
4402 // We current use TESTL [spp],EDI
4403 // A better choice might be TESTB [spp + pagesize() - CacheLineSize()],0
4405 enc_class Safepoint_Poll() %{
4406 cbuf.relocate(cbuf.inst_mark(), relocInfo::poll_type, 0);
4407 emit_opcode(cbuf,0x85);
4408 emit_rm (cbuf, 0x0, 0x7, 0x5);
4409 emit_d32(cbuf, (intptr_t)os::get_polling_page());
4410 %}
4411 %}
4414 //----------FRAME--------------------------------------------------------------
4415 // Definition of frame structure and management information.
4416 //
4417 // S T A C K L A Y O U T Allocators stack-slot number
4418 // | (to get allocators register number
4419 // G Owned by | | v add OptoReg::stack0())
4420 // r CALLER | |
4421 // o | +--------+ pad to even-align allocators stack-slot
4422 // w V | pad0 | numbers; owned by CALLER
4423 // t -----------+--------+----> Matcher::_in_arg_limit, unaligned
4424 // h ^ | in | 5
4425 // | | args | 4 Holes in incoming args owned by SELF
4426 // | | | | 3
4427 // | | +--------+
4428 // V | | old out| Empty on Intel, window on Sparc
4429 // | old |preserve| Must be even aligned.
4430 // | SP-+--------+----> Matcher::_old_SP, even aligned
4431 // | | in | 3 area for Intel ret address
4432 // Owned by |preserve| Empty on Sparc.
4433 // SELF +--------+
4434 // | | pad2 | 2 pad to align old SP
4435 // | +--------+ 1
4436 // | | locks | 0
4437 // | +--------+----> OptoReg::stack0(), even aligned
4438 // | | pad1 | 11 pad to align new SP
4439 // | +--------+
4440 // | | | 10
4441 // | | spills | 9 spills
4442 // V | | 8 (pad0 slot for callee)
4443 // -----------+--------+----> Matcher::_out_arg_limit, unaligned
4444 // ^ | out | 7
4445 // | | args | 6 Holes in outgoing args owned by CALLEE
4446 // Owned by +--------+
4447 // CALLEE | new out| 6 Empty on Intel, window on Sparc
4448 // | new |preserve| Must be even-aligned.
4449 // | SP-+--------+----> Matcher::_new_SP, even aligned
4450 // | | |
4451 //
4452 // Note 1: Only region 8-11 is determined by the allocator. Region 0-5 is
4453 // known from SELF's arguments and the Java calling convention.
4454 // Region 6-7 is determined per call site.
4455 // Note 2: If the calling convention leaves holes in the incoming argument
4456 // area, those holes are owned by SELF. Holes in the outgoing area
4457 // are owned by the CALLEE. Holes should not be nessecary in the
4458 // incoming area, as the Java calling convention is completely under
4459 // the control of the AD file. Doubles can be sorted and packed to
4460 // avoid holes. Holes in the outgoing arguments may be nessecary for
4461 // varargs C calling conventions.
4462 // Note 3: Region 0-3 is even aligned, with pad2 as needed. Region 3-5 is
4463 // even aligned with pad0 as needed.
4464 // Region 6 is even aligned. Region 6-7 is NOT even aligned;
4465 // region 6-11 is even aligned; it may be padded out more so that
4466 // the region from SP to FP meets the minimum stack alignment.
4468 frame %{
4469 // What direction does stack grow in (assumed to be same for C & Java)
4470 stack_direction(TOWARDS_LOW);
4472 // These three registers define part of the calling convention
4473 // between compiled code and the interpreter.
4474 inline_cache_reg(EAX); // Inline Cache Register
4475 interpreter_method_oop_reg(EBX); // Method Oop Register when calling interpreter
4477 // Optional: name the operand used by cisc-spilling to access [stack_pointer + offset]
4478 cisc_spilling_operand_name(indOffset32);
4480 // Number of stack slots consumed by locking an object
4481 sync_stack_slots(1);
4483 // Compiled code's Frame Pointer
4484 frame_pointer(ESP);
4485 // Interpreter stores its frame pointer in a register which is
4486 // stored to the stack by I2CAdaptors.
4487 // I2CAdaptors convert from interpreted java to compiled java.
4488 interpreter_frame_pointer(EBP);
4490 // Stack alignment requirement
4491 // Alignment size in bytes (128-bit -> 16 bytes)
4492 stack_alignment(StackAlignmentInBytes);
4494 // Number of stack slots between incoming argument block and the start of
4495 // a new frame. The PROLOG must add this many slots to the stack. The
4496 // EPILOG must remove this many slots. Intel needs one slot for
4497 // return address and one for rbp, (must save rbp)
4498 in_preserve_stack_slots(2+VerifyStackAtCalls);
4500 // Number of outgoing stack slots killed above the out_preserve_stack_slots
4501 // for calls to C. Supports the var-args backing area for register parms.
4502 varargs_C_out_slots_killed(0);
4504 // The after-PROLOG location of the return address. Location of
4505 // return address specifies a type (REG or STACK) and a number
4506 // representing the register number (i.e. - use a register name) or
4507 // stack slot.
4508 // Ret Addr is on stack in slot 0 if no locks or verification or alignment.
4509 // Otherwise, it is above the locks and verification slot and alignment word
4510 return_addr(STACK - 1 +
4511 round_to(1+VerifyStackAtCalls+
4512 Compile::current()->fixed_slots(),
4513 (StackAlignmentInBytes/wordSize)));
4515 // Body of function which returns an integer array locating
4516 // arguments either in registers or in stack slots. Passed an array
4517 // of ideal registers called "sig" and a "length" count. Stack-slot
4518 // offsets are based on outgoing arguments, i.e. a CALLER setting up
4519 // arguments for a CALLEE. Incoming stack arguments are
4520 // automatically biased by the preserve_stack_slots field above.
4521 calling_convention %{
4522 // No difference between ingoing/outgoing just pass false
4523 SharedRuntime::java_calling_convention(sig_bt, regs, length, false);
4524 %}
4527 // Body of function which returns an integer array locating
4528 // arguments either in registers or in stack slots. Passed an array
4529 // of ideal registers called "sig" and a "length" count. Stack-slot
4530 // offsets are based on outgoing arguments, i.e. a CALLER setting up
4531 // arguments for a CALLEE. Incoming stack arguments are
4532 // automatically biased by the preserve_stack_slots field above.
4533 c_calling_convention %{
4534 // This is obviously always outgoing
4535 (void) SharedRuntime::c_calling_convention(sig_bt, regs, length);
4536 %}
4538 // Location of C & interpreter return values
4539 c_return_value %{
4540 assert( ideal_reg >= Op_RegI && ideal_reg <= Op_RegL, "only return normal values" );
4541 static int lo[Op_RegL+1] = { 0, 0, EAX_num, EAX_num, FPR1L_num, FPR1L_num, EAX_num };
4542 static int hi[Op_RegL+1] = { 0, 0, OptoReg::Bad, OptoReg::Bad, OptoReg::Bad, FPR1H_num, EDX_num };
4544 // in SSE2+ mode we want to keep the FPU stack clean so pretend
4545 // that C functions return float and double results in XMM0.
4546 if( ideal_reg == Op_RegD && UseSSE>=2 )
4547 return OptoRegPair(XMM0b_num,XMM0a_num);
4548 if( ideal_reg == Op_RegF && UseSSE>=2 )
4549 return OptoRegPair(OptoReg::Bad,XMM0a_num);
4551 return OptoRegPair(hi[ideal_reg],lo[ideal_reg]);
4552 %}
4554 // Location of return values
4555 return_value %{
4556 assert( ideal_reg >= Op_RegI && ideal_reg <= Op_RegL, "only return normal values" );
4557 static int lo[Op_RegL+1] = { 0, 0, EAX_num, EAX_num, FPR1L_num, FPR1L_num, EAX_num };
4558 static int hi[Op_RegL+1] = { 0, 0, OptoReg::Bad, OptoReg::Bad, OptoReg::Bad, FPR1H_num, EDX_num };
4559 if( ideal_reg == Op_RegD && UseSSE>=2 )
4560 return OptoRegPair(XMM0b_num,XMM0a_num);
4561 if( ideal_reg == Op_RegF && UseSSE>=1 )
4562 return OptoRegPair(OptoReg::Bad,XMM0a_num);
4563 return OptoRegPair(hi[ideal_reg],lo[ideal_reg]);
4564 %}
4566 %}
4568 //----------ATTRIBUTES---------------------------------------------------------
4569 //----------Operand Attributes-------------------------------------------------
4570 op_attrib op_cost(0); // Required cost attribute
4572 //----------Instruction Attributes---------------------------------------------
4573 ins_attrib ins_cost(100); // Required cost attribute
4574 ins_attrib ins_size(8); // Required size attribute (in bits)
4575 ins_attrib ins_pc_relative(0); // Required PC Relative flag
4576 ins_attrib ins_short_branch(0); // Required flag: is this instruction a
4577 // non-matching short branch variant of some
4578 // long branch?
4579 ins_attrib ins_alignment(1); // Required alignment attribute (must be a power of 2)
4580 // specifies the alignment that some part of the instruction (not
4581 // necessarily the start) requires. If > 1, a compute_padding()
4582 // function must be provided for the instruction
4584 //----------OPERANDS-----------------------------------------------------------
4585 // Operand definitions must precede instruction definitions for correct parsing
4586 // in the ADLC because operands constitute user defined types which are used in
4587 // instruction definitions.
4589 //----------Simple Operands----------------------------------------------------
4590 // Immediate Operands
4591 // Integer Immediate
4592 operand immI() %{
4593 match(ConI);
4595 op_cost(10);
4596 format %{ %}
4597 interface(CONST_INTER);
4598 %}
4600 // Constant for test vs zero
4601 operand immI0() %{
4602 predicate(n->get_int() == 0);
4603 match(ConI);
4605 op_cost(0);
4606 format %{ %}
4607 interface(CONST_INTER);
4608 %}
4610 // Constant for increment
4611 operand immI1() %{
4612 predicate(n->get_int() == 1);
4613 match(ConI);
4615 op_cost(0);
4616 format %{ %}
4617 interface(CONST_INTER);
4618 %}
4620 // Constant for decrement
4621 operand immI_M1() %{
4622 predicate(n->get_int() == -1);
4623 match(ConI);
4625 op_cost(0);
4626 format %{ %}
4627 interface(CONST_INTER);
4628 %}
4630 // Valid scale values for addressing modes
4631 operand immI2() %{
4632 predicate(0 <= n->get_int() && (n->get_int() <= 3));
4633 match(ConI);
4635 format %{ %}
4636 interface(CONST_INTER);
4637 %}
4639 operand immI8() %{
4640 predicate((-128 <= n->get_int()) && (n->get_int() <= 127));
4641 match(ConI);
4643 op_cost(5);
4644 format %{ %}
4645 interface(CONST_INTER);
4646 %}
4648 operand immI16() %{
4649 predicate((-32768 <= n->get_int()) && (n->get_int() <= 32767));
4650 match(ConI);
4652 op_cost(10);
4653 format %{ %}
4654 interface(CONST_INTER);
4655 %}
4657 // Constant for long shifts
4658 operand immI_32() %{
4659 predicate( n->get_int() == 32 );
4660 match(ConI);
4662 op_cost(0);
4663 format %{ %}
4664 interface(CONST_INTER);
4665 %}
4667 operand immI_1_31() %{
4668 predicate( n->get_int() >= 1 && n->get_int() <= 31 );
4669 match(ConI);
4671 op_cost(0);
4672 format %{ %}
4673 interface(CONST_INTER);
4674 %}
4676 operand immI_32_63() %{
4677 predicate( n->get_int() >= 32 && n->get_int() <= 63 );
4678 match(ConI);
4679 op_cost(0);
4681 format %{ %}
4682 interface(CONST_INTER);
4683 %}
4685 // Pointer Immediate
4686 operand immP() %{
4687 match(ConP);
4689 op_cost(10);
4690 format %{ %}
4691 interface(CONST_INTER);
4692 %}
4694 // NULL Pointer Immediate
4695 operand immP0() %{
4696 predicate( n->get_ptr() == 0 );
4697 match(ConP);
4698 op_cost(0);
4700 format %{ %}
4701 interface(CONST_INTER);
4702 %}
4704 // Long Immediate
4705 operand immL() %{
4706 match(ConL);
4708 op_cost(20);
4709 format %{ %}
4710 interface(CONST_INTER);
4711 %}
4713 // Long Immediate zero
4714 operand immL0() %{
4715 predicate( n->get_long() == 0L );
4716 match(ConL);
4717 op_cost(0);
4719 format %{ %}
4720 interface(CONST_INTER);
4721 %}
4723 // Long immediate from 0 to 127.
4724 // Used for a shorter form of long mul by 10.
4725 operand immL_127() %{
4726 predicate((0 <= n->get_long()) && (n->get_long() <= 127));
4727 match(ConL);
4728 op_cost(0);
4730 format %{ %}
4731 interface(CONST_INTER);
4732 %}
4734 // Long Immediate: low 32-bit mask
4735 operand immL_32bits() %{
4736 predicate(n->get_long() == 0xFFFFFFFFL);
4737 match(ConL);
4738 op_cost(0);
4740 format %{ %}
4741 interface(CONST_INTER);
4742 %}
4744 // Long Immediate: low 32-bit mask
4745 operand immL32() %{
4746 predicate(n->get_long() == (int)(n->get_long()));
4747 match(ConL);
4748 op_cost(20);
4750 format %{ %}
4751 interface(CONST_INTER);
4752 %}
4754 //Double Immediate zero
4755 operand immD0() %{
4756 // Do additional (and counter-intuitive) test against NaN to work around VC++
4757 // bug that generates code such that NaNs compare equal to 0.0
4758 predicate( UseSSE<=1 && n->getd() == 0.0 && !g_isnan(n->getd()) );
4759 match(ConD);
4761 op_cost(5);
4762 format %{ %}
4763 interface(CONST_INTER);
4764 %}
4766 // Double Immediate
4767 operand immD1() %{
4768 predicate( UseSSE<=1 && n->getd() == 1.0 );
4769 match(ConD);
4771 op_cost(5);
4772 format %{ %}
4773 interface(CONST_INTER);
4774 %}
4776 // Double Immediate
4777 operand immD() %{
4778 predicate(UseSSE<=1);
4779 match(ConD);
4781 op_cost(5);
4782 format %{ %}
4783 interface(CONST_INTER);
4784 %}
4786 operand immXD() %{
4787 predicate(UseSSE>=2);
4788 match(ConD);
4790 op_cost(5);
4791 format %{ %}
4792 interface(CONST_INTER);
4793 %}
4795 // Double Immediate zero
4796 operand immXD0() %{
4797 // Do additional (and counter-intuitive) test against NaN to work around VC++
4798 // bug that generates code such that NaNs compare equal to 0.0 AND do not
4799 // compare equal to -0.0.
4800 predicate( UseSSE>=2 && jlong_cast(n->getd()) == 0 );
4801 match(ConD);
4803 format %{ %}
4804 interface(CONST_INTER);
4805 %}
4807 // Float Immediate zero
4808 operand immF0() %{
4809 predicate( UseSSE == 0 && n->getf() == 0.0 );
4810 match(ConF);
4812 op_cost(5);
4813 format %{ %}
4814 interface(CONST_INTER);
4815 %}
4817 // Float Immediate
4818 operand immF() %{
4819 predicate( UseSSE == 0 );
4820 match(ConF);
4822 op_cost(5);
4823 format %{ %}
4824 interface(CONST_INTER);
4825 %}
4827 // Float Immediate
4828 operand immXF() %{
4829 predicate(UseSSE >= 1);
4830 match(ConF);
4832 op_cost(5);
4833 format %{ %}
4834 interface(CONST_INTER);
4835 %}
4837 // Float Immediate zero. Zero and not -0.0
4838 operand immXF0() %{
4839 predicate( UseSSE >= 1 && jint_cast(n->getf()) == 0 );
4840 match(ConF);
4842 op_cost(5);
4843 format %{ %}
4844 interface(CONST_INTER);
4845 %}
4847 // Immediates for special shifts (sign extend)
4849 // Constants for increment
4850 operand immI_16() %{
4851 predicate( n->get_int() == 16 );
4852 match(ConI);
4854 format %{ %}
4855 interface(CONST_INTER);
4856 %}
4858 operand immI_24() %{
4859 predicate( n->get_int() == 24 );
4860 match(ConI);
4862 format %{ %}
4863 interface(CONST_INTER);
4864 %}
4866 // Constant for byte-wide masking
4867 operand immI_255() %{
4868 predicate( n->get_int() == 255 );
4869 match(ConI);
4871 format %{ %}
4872 interface(CONST_INTER);
4873 %}
4875 // Register Operands
4876 // Integer Register
4877 operand eRegI() %{
4878 constraint(ALLOC_IN_RC(e_reg));
4879 match(RegI);
4880 match(xRegI);
4881 match(eAXRegI);
4882 match(eBXRegI);
4883 match(eCXRegI);
4884 match(eDXRegI);
4885 match(eDIRegI);
4886 match(eSIRegI);
4888 format %{ %}
4889 interface(REG_INTER);
4890 %}
4892 // Subset of Integer Register
4893 operand xRegI(eRegI reg) %{
4894 constraint(ALLOC_IN_RC(x_reg));
4895 match(reg);
4896 match(eAXRegI);
4897 match(eBXRegI);
4898 match(eCXRegI);
4899 match(eDXRegI);
4901 format %{ %}
4902 interface(REG_INTER);
4903 %}
4905 // Special Registers
4906 operand eAXRegI(xRegI reg) %{
4907 constraint(ALLOC_IN_RC(eax_reg));
4908 match(reg);
4909 match(eRegI);
4911 format %{ "EAX" %}
4912 interface(REG_INTER);
4913 %}
4915 // Special Registers
4916 operand eBXRegI(xRegI reg) %{
4917 constraint(ALLOC_IN_RC(ebx_reg));
4918 match(reg);
4919 match(eRegI);
4921 format %{ "EBX" %}
4922 interface(REG_INTER);
4923 %}
4925 operand eCXRegI(xRegI reg) %{
4926 constraint(ALLOC_IN_RC(ecx_reg));
4927 match(reg);
4928 match(eRegI);
4930 format %{ "ECX" %}
4931 interface(REG_INTER);
4932 %}
4934 operand eDXRegI(xRegI reg) %{
4935 constraint(ALLOC_IN_RC(edx_reg));
4936 match(reg);
4937 match(eRegI);
4939 format %{ "EDX" %}
4940 interface(REG_INTER);
4941 %}
4943 operand eDIRegI(xRegI reg) %{
4944 constraint(ALLOC_IN_RC(edi_reg));
4945 match(reg);
4946 match(eRegI);
4948 format %{ "EDI" %}
4949 interface(REG_INTER);
4950 %}
4952 operand naxRegI() %{
4953 constraint(ALLOC_IN_RC(nax_reg));
4954 match(RegI);
4955 match(eCXRegI);
4956 match(eDXRegI);
4957 match(eSIRegI);
4958 match(eDIRegI);
4960 format %{ %}
4961 interface(REG_INTER);
4962 %}
4964 operand nadxRegI() %{
4965 constraint(ALLOC_IN_RC(nadx_reg));
4966 match(RegI);
4967 match(eBXRegI);
4968 match(eCXRegI);
4969 match(eSIRegI);
4970 match(eDIRegI);
4972 format %{ %}
4973 interface(REG_INTER);
4974 %}
4976 operand ncxRegI() %{
4977 constraint(ALLOC_IN_RC(ncx_reg));
4978 match(RegI);
4979 match(eAXRegI);
4980 match(eDXRegI);
4981 match(eSIRegI);
4982 match(eDIRegI);
4984 format %{ %}
4985 interface(REG_INTER);
4986 %}
4988 // // This operand was used by cmpFastUnlock, but conflicted with 'object' reg
4989 // //
4990 operand eSIRegI(xRegI reg) %{
4991 constraint(ALLOC_IN_RC(esi_reg));
4992 match(reg);
4993 match(eRegI);
4995 format %{ "ESI" %}
4996 interface(REG_INTER);
4997 %}
4999 // Pointer Register
5000 operand anyRegP() %{
5001 constraint(ALLOC_IN_RC(any_reg));
5002 match(RegP);
5003 match(eAXRegP);
5004 match(eBXRegP);
5005 match(eCXRegP);
5006 match(eDIRegP);
5007 match(eRegP);
5009 format %{ %}
5010 interface(REG_INTER);
5011 %}
5013 operand eRegP() %{
5014 constraint(ALLOC_IN_RC(e_reg));
5015 match(RegP);
5016 match(eAXRegP);
5017 match(eBXRegP);
5018 match(eCXRegP);
5019 match(eDIRegP);
5021 format %{ %}
5022 interface(REG_INTER);
5023 %}
5025 // On windows95, EBP is not safe to use for implicit null tests.
5026 operand eRegP_no_EBP() %{
5027 constraint(ALLOC_IN_RC(e_reg_no_rbp));
5028 match(RegP);
5029 match(eAXRegP);
5030 match(eBXRegP);
5031 match(eCXRegP);
5032 match(eDIRegP);
5034 op_cost(100);
5035 format %{ %}
5036 interface(REG_INTER);
5037 %}
5039 operand naxRegP() %{
5040 constraint(ALLOC_IN_RC(nax_reg));
5041 match(RegP);
5042 match(eBXRegP);
5043 match(eDXRegP);
5044 match(eCXRegP);
5045 match(eSIRegP);
5046 match(eDIRegP);
5048 format %{ %}
5049 interface(REG_INTER);
5050 %}
5052 operand nabxRegP() %{
5053 constraint(ALLOC_IN_RC(nabx_reg));
5054 match(RegP);
5055 match(eCXRegP);
5056 match(eDXRegP);
5057 match(eSIRegP);
5058 match(eDIRegP);
5060 format %{ %}
5061 interface(REG_INTER);
5062 %}
5064 operand pRegP() %{
5065 constraint(ALLOC_IN_RC(p_reg));
5066 match(RegP);
5067 match(eBXRegP);
5068 match(eDXRegP);
5069 match(eSIRegP);
5070 match(eDIRegP);
5072 format %{ %}
5073 interface(REG_INTER);
5074 %}
5076 // Special Registers
5077 // Return a pointer value
5078 operand eAXRegP(eRegP reg) %{
5079 constraint(ALLOC_IN_RC(eax_reg));
5080 match(reg);
5081 format %{ "EAX" %}
5082 interface(REG_INTER);
5083 %}
5085 // Used in AtomicAdd
5086 operand eBXRegP(eRegP reg) %{
5087 constraint(ALLOC_IN_RC(ebx_reg));
5088 match(reg);
5089 format %{ "EBX" %}
5090 interface(REG_INTER);
5091 %}
5093 // Tail-call (interprocedural jump) to interpreter
5094 operand eCXRegP(eRegP reg) %{
5095 constraint(ALLOC_IN_RC(ecx_reg));
5096 match(reg);
5097 format %{ "ECX" %}
5098 interface(REG_INTER);
5099 %}
5101 operand eSIRegP(eRegP reg) %{
5102 constraint(ALLOC_IN_RC(esi_reg));
5103 match(reg);
5104 format %{ "ESI" %}
5105 interface(REG_INTER);
5106 %}
5108 // Used in rep stosw
5109 operand eDIRegP(eRegP reg) %{
5110 constraint(ALLOC_IN_RC(edi_reg));
5111 match(reg);
5112 format %{ "EDI" %}
5113 interface(REG_INTER);
5114 %}
5116 operand eBPRegP() %{
5117 constraint(ALLOC_IN_RC(ebp_reg));
5118 match(RegP);
5119 format %{ "EBP" %}
5120 interface(REG_INTER);
5121 %}
5123 operand eRegL() %{
5124 constraint(ALLOC_IN_RC(long_reg));
5125 match(RegL);
5126 match(eADXRegL);
5128 format %{ %}
5129 interface(REG_INTER);
5130 %}
5132 operand eADXRegL( eRegL reg ) %{
5133 constraint(ALLOC_IN_RC(eadx_reg));
5134 match(reg);
5136 format %{ "EDX:EAX" %}
5137 interface(REG_INTER);
5138 %}
5140 operand eBCXRegL( eRegL reg ) %{
5141 constraint(ALLOC_IN_RC(ebcx_reg));
5142 match(reg);
5144 format %{ "EBX:ECX" %}
5145 interface(REG_INTER);
5146 %}
5148 // Special case for integer high multiply
5149 operand eADXRegL_low_only() %{
5150 constraint(ALLOC_IN_RC(eadx_reg));
5151 match(RegL);
5153 format %{ "EAX" %}
5154 interface(REG_INTER);
5155 %}
5157 // Flags register, used as output of compare instructions
5158 operand eFlagsReg() %{
5159 constraint(ALLOC_IN_RC(int_flags));
5160 match(RegFlags);
5162 format %{ "EFLAGS" %}
5163 interface(REG_INTER);
5164 %}
5166 // Flags register, used as output of FLOATING POINT compare instructions
5167 operand eFlagsRegU() %{
5168 constraint(ALLOC_IN_RC(int_flags));
5169 match(RegFlags);
5171 format %{ "EFLAGS_U" %}
5172 interface(REG_INTER);
5173 %}
5175 // Condition Code Register used by long compare
5176 operand flagsReg_long_LTGE() %{
5177 constraint(ALLOC_IN_RC(int_flags));
5178 match(RegFlags);
5179 format %{ "FLAGS_LTGE" %}
5180 interface(REG_INTER);
5181 %}
5182 operand flagsReg_long_EQNE() %{
5183 constraint(ALLOC_IN_RC(int_flags));
5184 match(RegFlags);
5185 format %{ "FLAGS_EQNE" %}
5186 interface(REG_INTER);
5187 %}
5188 operand flagsReg_long_LEGT() %{
5189 constraint(ALLOC_IN_RC(int_flags));
5190 match(RegFlags);
5191 format %{ "FLAGS_LEGT" %}
5192 interface(REG_INTER);
5193 %}
5195 // Float register operands
5196 operand regD() %{
5197 predicate( UseSSE < 2 );
5198 constraint(ALLOC_IN_RC(dbl_reg));
5199 match(RegD);
5200 match(regDPR1);
5201 match(regDPR2);
5202 format %{ %}
5203 interface(REG_INTER);
5204 %}
5206 operand regDPR1(regD reg) %{
5207 predicate( UseSSE < 2 );
5208 constraint(ALLOC_IN_RC(dbl_reg0));
5209 match(reg);
5210 format %{ "FPR1" %}
5211 interface(REG_INTER);
5212 %}
5214 operand regDPR2(regD reg) %{
5215 predicate( UseSSE < 2 );
5216 constraint(ALLOC_IN_RC(dbl_reg1));
5217 match(reg);
5218 format %{ "FPR2" %}
5219 interface(REG_INTER);
5220 %}
5222 operand regnotDPR1(regD reg) %{
5223 predicate( UseSSE < 2 );
5224 constraint(ALLOC_IN_RC(dbl_notreg0));
5225 match(reg);
5226 format %{ %}
5227 interface(REG_INTER);
5228 %}
5230 // XMM Double register operands
5231 operand regXD() %{
5232 predicate( UseSSE>=2 );
5233 constraint(ALLOC_IN_RC(xdb_reg));
5234 match(RegD);
5235 match(regXD6);
5236 match(regXD7);
5237 format %{ %}
5238 interface(REG_INTER);
5239 %}
5241 // XMM6 double register operands
5242 operand regXD6(regXD reg) %{
5243 predicate( UseSSE>=2 );
5244 constraint(ALLOC_IN_RC(xdb_reg6));
5245 match(reg);
5246 format %{ "XMM6" %}
5247 interface(REG_INTER);
5248 %}
5250 // XMM7 double register operands
5251 operand regXD7(regXD reg) %{
5252 predicate( UseSSE>=2 );
5253 constraint(ALLOC_IN_RC(xdb_reg7));
5254 match(reg);
5255 format %{ "XMM7" %}
5256 interface(REG_INTER);
5257 %}
5259 // Float register operands
5260 operand regF() %{
5261 predicate( UseSSE < 2 );
5262 constraint(ALLOC_IN_RC(flt_reg));
5263 match(RegF);
5264 match(regFPR1);
5265 format %{ %}
5266 interface(REG_INTER);
5267 %}
5269 // Float register operands
5270 operand regFPR1(regF reg) %{
5271 predicate( UseSSE < 2 );
5272 constraint(ALLOC_IN_RC(flt_reg0));
5273 match(reg);
5274 format %{ "FPR1" %}
5275 interface(REG_INTER);
5276 %}
5278 // XMM register operands
5279 operand regX() %{
5280 predicate( UseSSE>=1 );
5281 constraint(ALLOC_IN_RC(xmm_reg));
5282 match(RegF);
5283 format %{ %}
5284 interface(REG_INTER);
5285 %}
5288 //----------Memory Operands----------------------------------------------------
5289 // Direct Memory Operand
5290 operand direct(immP addr) %{
5291 match(addr);
5293 format %{ "[$addr]" %}
5294 interface(MEMORY_INTER) %{
5295 base(0xFFFFFFFF);
5296 index(0x4);
5297 scale(0x0);
5298 disp($addr);
5299 %}
5300 %}
5302 // Indirect Memory Operand
5303 operand indirect(eRegP reg) %{
5304 constraint(ALLOC_IN_RC(e_reg));
5305 match(reg);
5307 format %{ "[$reg]" %}
5308 interface(MEMORY_INTER) %{
5309 base($reg);
5310 index(0x4);
5311 scale(0x0);
5312 disp(0x0);
5313 %}
5314 %}
5316 // Indirect Memory Plus Short Offset Operand
5317 operand indOffset8(eRegP reg, immI8 off) %{
5318 match(AddP reg off);
5320 format %{ "[$reg + $off]" %}
5321 interface(MEMORY_INTER) %{
5322 base($reg);
5323 index(0x4);
5324 scale(0x0);
5325 disp($off);
5326 %}
5327 %}
5329 // Indirect Memory Plus Long Offset Operand
5330 operand indOffset32(eRegP reg, immI off) %{
5331 match(AddP reg off);
5333 format %{ "[$reg + $off]" %}
5334 interface(MEMORY_INTER) %{
5335 base($reg);
5336 index(0x4);
5337 scale(0x0);
5338 disp($off);
5339 %}
5340 %}
5342 // Indirect Memory Plus Long Offset Operand
5343 operand indOffset32X(eRegI reg, immP off) %{
5344 match(AddP off reg);
5346 format %{ "[$reg + $off]" %}
5347 interface(MEMORY_INTER) %{
5348 base($reg);
5349 index(0x4);
5350 scale(0x0);
5351 disp($off);
5352 %}
5353 %}
5355 // Indirect Memory Plus Index Register Plus Offset Operand
5356 operand indIndexOffset(eRegP reg, eRegI ireg, immI off) %{
5357 match(AddP (AddP reg ireg) off);
5359 op_cost(10);
5360 format %{"[$reg + $off + $ireg]" %}
5361 interface(MEMORY_INTER) %{
5362 base($reg);
5363 index($ireg);
5364 scale(0x0);
5365 disp($off);
5366 %}
5367 %}
5369 // Indirect Memory Plus Index Register Plus Offset Operand
5370 operand indIndex(eRegP reg, eRegI ireg) %{
5371 match(AddP reg ireg);
5373 op_cost(10);
5374 format %{"[$reg + $ireg]" %}
5375 interface(MEMORY_INTER) %{
5376 base($reg);
5377 index($ireg);
5378 scale(0x0);
5379 disp(0x0);
5380 %}
5381 %}
5383 // // -------------------------------------------------------------------------
5384 // // 486 architecture doesn't support "scale * index + offset" with out a base
5385 // // -------------------------------------------------------------------------
5386 // // Scaled Memory Operands
5387 // // Indirect Memory Times Scale Plus Offset Operand
5388 // operand indScaleOffset(immP off, eRegI ireg, immI2 scale) %{
5389 // match(AddP off (LShiftI ireg scale));
5390 //
5391 // op_cost(10);
5392 // format %{"[$off + $ireg << $scale]" %}
5393 // interface(MEMORY_INTER) %{
5394 // base(0x4);
5395 // index($ireg);
5396 // scale($scale);
5397 // disp($off);
5398 // %}
5399 // %}
5401 // Indirect Memory Times Scale Plus Index Register
5402 operand indIndexScale(eRegP reg, eRegI ireg, immI2 scale) %{
5403 match(AddP reg (LShiftI ireg scale));
5405 op_cost(10);
5406 format %{"[$reg + $ireg << $scale]" %}
5407 interface(MEMORY_INTER) %{
5408 base($reg);
5409 index($ireg);
5410 scale($scale);
5411 disp(0x0);
5412 %}
5413 %}
5415 // Indirect Memory Times Scale Plus Index Register Plus Offset Operand
5416 operand indIndexScaleOffset(eRegP reg, immI off, eRegI ireg, immI2 scale) %{
5417 match(AddP (AddP reg (LShiftI ireg scale)) off);
5419 op_cost(10);
5420 format %{"[$reg + $off + $ireg << $scale]" %}
5421 interface(MEMORY_INTER) %{
5422 base($reg);
5423 index($ireg);
5424 scale($scale);
5425 disp($off);
5426 %}
5427 %}
5429 //----------Load Long Memory Operands------------------------------------------
5430 // The load-long idiom will use it's address expression again after loading
5431 // the first word of the long. If the load-long destination overlaps with
5432 // registers used in the addressing expression, the 2nd half will be loaded
5433 // from a clobbered address. Fix this by requiring that load-long use
5434 // address registers that do not overlap with the load-long target.
5436 // load-long support
5437 operand load_long_RegP() %{
5438 constraint(ALLOC_IN_RC(esi_reg));
5439 match(RegP);
5440 match(eSIRegP);
5441 op_cost(100);
5442 format %{ %}
5443 interface(REG_INTER);
5444 %}
5446 // Indirect Memory Operand Long
5447 operand load_long_indirect(load_long_RegP reg) %{
5448 constraint(ALLOC_IN_RC(esi_reg));
5449 match(reg);
5451 format %{ "[$reg]" %}
5452 interface(MEMORY_INTER) %{
5453 base($reg);
5454 index(0x4);
5455 scale(0x0);
5456 disp(0x0);
5457 %}
5458 %}
5460 // Indirect Memory Plus Long Offset Operand
5461 operand load_long_indOffset32(load_long_RegP reg, immI off) %{
5462 match(AddP reg off);
5464 format %{ "[$reg + $off]" %}
5465 interface(MEMORY_INTER) %{
5466 base($reg);
5467 index(0x4);
5468 scale(0x0);
5469 disp($off);
5470 %}
5471 %}
5473 opclass load_long_memory(load_long_indirect, load_long_indOffset32);
5476 //----------Special Memory Operands--------------------------------------------
5477 // Stack Slot Operand - This operand is used for loading and storing temporary
5478 // values on the stack where a match requires a value to
5479 // flow through memory.
5480 operand stackSlotP(sRegP reg) %{
5481 constraint(ALLOC_IN_RC(stack_slots));
5482 // No match rule because this operand is only generated in matching
5483 format %{ "[$reg]" %}
5484 interface(MEMORY_INTER) %{
5485 base(0x4); // ESP
5486 index(0x4); // No Index
5487 scale(0x0); // No Scale
5488 disp($reg); // Stack Offset
5489 %}
5490 %}
5492 operand stackSlotI(sRegI reg) %{
5493 constraint(ALLOC_IN_RC(stack_slots));
5494 // No match rule because this operand is only generated in matching
5495 format %{ "[$reg]" %}
5496 interface(MEMORY_INTER) %{
5497 base(0x4); // ESP
5498 index(0x4); // No Index
5499 scale(0x0); // No Scale
5500 disp($reg); // Stack Offset
5501 %}
5502 %}
5504 operand stackSlotF(sRegF reg) %{
5505 constraint(ALLOC_IN_RC(stack_slots));
5506 // No match rule because this operand is only generated in matching
5507 format %{ "[$reg]" %}
5508 interface(MEMORY_INTER) %{
5509 base(0x4); // ESP
5510 index(0x4); // No Index
5511 scale(0x0); // No Scale
5512 disp($reg); // Stack Offset
5513 %}
5514 %}
5516 operand stackSlotD(sRegD reg) %{
5517 constraint(ALLOC_IN_RC(stack_slots));
5518 // No match rule because this operand is only generated in matching
5519 format %{ "[$reg]" %}
5520 interface(MEMORY_INTER) %{
5521 base(0x4); // ESP
5522 index(0x4); // No Index
5523 scale(0x0); // No Scale
5524 disp($reg); // Stack Offset
5525 %}
5526 %}
5528 operand stackSlotL(sRegL reg) %{
5529 constraint(ALLOC_IN_RC(stack_slots));
5530 // No match rule because this operand is only generated in matching
5531 format %{ "[$reg]" %}
5532 interface(MEMORY_INTER) %{
5533 base(0x4); // ESP
5534 index(0x4); // No Index
5535 scale(0x0); // No Scale
5536 disp($reg); // Stack Offset
5537 %}
5538 %}
5540 //----------Memory Operands - Win95 Implicit Null Variants----------------
5541 // Indirect Memory Operand
5542 operand indirect_win95_safe(eRegP_no_EBP reg)
5543 %{
5544 constraint(ALLOC_IN_RC(e_reg));
5545 match(reg);
5547 op_cost(100);
5548 format %{ "[$reg]" %}
5549 interface(MEMORY_INTER) %{
5550 base($reg);
5551 index(0x4);
5552 scale(0x0);
5553 disp(0x0);
5554 %}
5555 %}
5557 // Indirect Memory Plus Short Offset Operand
5558 operand indOffset8_win95_safe(eRegP_no_EBP reg, immI8 off)
5559 %{
5560 match(AddP reg off);
5562 op_cost(100);
5563 format %{ "[$reg + $off]" %}
5564 interface(MEMORY_INTER) %{
5565 base($reg);
5566 index(0x4);
5567 scale(0x0);
5568 disp($off);
5569 %}
5570 %}
5572 // Indirect Memory Plus Long Offset Operand
5573 operand indOffset32_win95_safe(eRegP_no_EBP reg, immI off)
5574 %{
5575 match(AddP reg off);
5577 op_cost(100);
5578 format %{ "[$reg + $off]" %}
5579 interface(MEMORY_INTER) %{
5580 base($reg);
5581 index(0x4);
5582 scale(0x0);
5583 disp($off);
5584 %}
5585 %}
5587 // Indirect Memory Plus Index Register Plus Offset Operand
5588 operand indIndexOffset_win95_safe(eRegP_no_EBP reg, eRegI ireg, immI off)
5589 %{
5590 match(AddP (AddP reg ireg) off);
5592 op_cost(100);
5593 format %{"[$reg + $off + $ireg]" %}
5594 interface(MEMORY_INTER) %{
5595 base($reg);
5596 index($ireg);
5597 scale(0x0);
5598 disp($off);
5599 %}
5600 %}
5602 // Indirect Memory Times Scale Plus Index Register
5603 operand indIndexScale_win95_safe(eRegP_no_EBP reg, eRegI ireg, immI2 scale)
5604 %{
5605 match(AddP reg (LShiftI ireg scale));
5607 op_cost(100);
5608 format %{"[$reg + $ireg << $scale]" %}
5609 interface(MEMORY_INTER) %{
5610 base($reg);
5611 index($ireg);
5612 scale($scale);
5613 disp(0x0);
5614 %}
5615 %}
5617 // Indirect Memory Times Scale Plus Index Register Plus Offset Operand
5618 operand indIndexScaleOffset_win95_safe(eRegP_no_EBP reg, immI off, eRegI ireg, immI2 scale)
5619 %{
5620 match(AddP (AddP reg (LShiftI ireg scale)) off);
5622 op_cost(100);
5623 format %{"[$reg + $off + $ireg << $scale]" %}
5624 interface(MEMORY_INTER) %{
5625 base($reg);
5626 index($ireg);
5627 scale($scale);
5628 disp($off);
5629 %}
5630 %}
5632 //----------Conditional Branch Operands----------------------------------------
5633 // Comparison Op - This is the operation of the comparison, and is limited to
5634 // the following set of codes:
5635 // L (<), LE (<=), G (>), GE (>=), E (==), NE (!=)
5636 //
5637 // Other attributes of the comparison, such as unsignedness, are specified
5638 // by the comparison instruction that sets a condition code flags register.
5639 // That result is represented by a flags operand whose subtype is appropriate
5640 // to the unsignedness (etc.) of the comparison.
5641 //
5642 // Later, the instruction which matches both the Comparison Op (a Bool) and
5643 // the flags (produced by the Cmp) specifies the coding of the comparison op
5644 // by matching a specific subtype of Bool operand below, such as cmpOpU.
5646 // Comparision Code
5647 operand cmpOp() %{
5648 match(Bool);
5650 format %{ "" %}
5651 interface(COND_INTER) %{
5652 equal(0x4);
5653 not_equal(0x5);
5654 less(0xC);
5655 greater_equal(0xD);
5656 less_equal(0xE);
5657 greater(0xF);
5658 %}
5659 %}
5661 // Comparison Code, unsigned compare. Used by FP also, with
5662 // C2 (unordered) turned into GT or LT already. The other bits
5663 // C0 and C3 are turned into Carry & Zero flags.
5664 operand cmpOpU() %{
5665 match(Bool);
5667 format %{ "" %}
5668 interface(COND_INTER) %{
5669 equal(0x4);
5670 not_equal(0x5);
5671 less(0x2);
5672 greater_equal(0x3);
5673 less_equal(0x6);
5674 greater(0x7);
5675 %}
5676 %}
5678 // Comparison Code for FP conditional move
5679 operand cmpOp_fcmov() %{
5680 match(Bool);
5682 format %{ "" %}
5683 interface(COND_INTER) %{
5684 equal (0x0C8);
5685 not_equal (0x1C8);
5686 less (0x0C0);
5687 greater_equal(0x1C0);
5688 less_equal (0x0D0);
5689 greater (0x1D0);
5690 %}
5691 %}
5693 // Comparision Code used in long compares
5694 operand cmpOp_commute() %{
5695 match(Bool);
5697 format %{ "" %}
5698 interface(COND_INTER) %{
5699 equal(0x4);
5700 not_equal(0x5);
5701 less(0xF);
5702 greater_equal(0xE);
5703 less_equal(0xD);
5704 greater(0xC);
5705 %}
5706 %}
5708 //----------OPERAND CLASSES----------------------------------------------------
5709 // Operand Classes are groups of operands that are used as to simplify
5710 // instruction definitions by not requiring the AD writer to specify seperate
5711 // instructions for every form of operand when the instruction accepts
5712 // multiple operand types with the same basic encoding and format. The classic
5713 // case of this is memory operands.
5715 opclass memory(direct, indirect, indOffset8, indOffset32, indOffset32X, indIndexOffset,
5716 indIndex, indIndexScale, indIndexScaleOffset);
5718 // Long memory operations are encoded in 2 instructions and a +4 offset.
5719 // This means some kind of offset is always required and you cannot use
5720 // an oop as the offset (done when working on static globals).
5721 opclass long_memory(direct, indirect, indOffset8, indOffset32, indIndexOffset,
5722 indIndex, indIndexScale, indIndexScaleOffset);
5725 //----------PIPELINE-----------------------------------------------------------
5726 // Rules which define the behavior of the target architectures pipeline.
5727 pipeline %{
5729 //----------ATTRIBUTES---------------------------------------------------------
5730 attributes %{
5731 variable_size_instructions; // Fixed size instructions
5732 max_instructions_per_bundle = 3; // Up to 3 instructions per bundle
5733 instruction_unit_size = 1; // An instruction is 1 bytes long
5734 instruction_fetch_unit_size = 16; // The processor fetches one line
5735 instruction_fetch_units = 1; // of 16 bytes
5737 // List of nop instructions
5738 nops( MachNop );
5739 %}
5741 //----------RESOURCES----------------------------------------------------------
5742 // Resources are the functional units available to the machine
5744 // Generic P2/P3 pipeline
5745 // 3 decoders, only D0 handles big operands; a "bundle" is the limit of
5746 // 3 instructions decoded per cycle.
5747 // 2 load/store ops per cycle, 1 branch, 1 FPU,
5748 // 2 ALU op, only ALU0 handles mul/div instructions.
5749 resources( D0, D1, D2, DECODE = D0 | D1 | D2,
5750 MS0, MS1, MEM = MS0 | MS1,
5751 BR, FPU,
5752 ALU0, ALU1, ALU = ALU0 | ALU1 );
5754 //----------PIPELINE DESCRIPTION-----------------------------------------------
5755 // Pipeline Description specifies the stages in the machine's pipeline
5757 // Generic P2/P3 pipeline
5758 pipe_desc(S0, S1, S2, S3, S4, S5);
5760 //----------PIPELINE CLASSES---------------------------------------------------
5761 // Pipeline Classes describe the stages in which input and output are
5762 // referenced by the hardware pipeline.
5764 // Naming convention: ialu or fpu
5765 // Then: _reg
5766 // Then: _reg if there is a 2nd register
5767 // Then: _long if it's a pair of instructions implementing a long
5768 // Then: _fat if it requires the big decoder
5769 // Or: _mem if it requires the big decoder and a memory unit.
5771 // Integer ALU reg operation
5772 pipe_class ialu_reg(eRegI dst) %{
5773 single_instruction;
5774 dst : S4(write);
5775 dst : S3(read);
5776 DECODE : S0; // any decoder
5777 ALU : S3; // any alu
5778 %}
5780 // Long ALU reg operation
5781 pipe_class ialu_reg_long(eRegL dst) %{
5782 instruction_count(2);
5783 dst : S4(write);
5784 dst : S3(read);
5785 DECODE : S0(2); // any 2 decoders
5786 ALU : S3(2); // both alus
5787 %}
5789 // Integer ALU reg operation using big decoder
5790 pipe_class ialu_reg_fat(eRegI dst) %{
5791 single_instruction;
5792 dst : S4(write);
5793 dst : S3(read);
5794 D0 : S0; // big decoder only
5795 ALU : S3; // any alu
5796 %}
5798 // Long ALU reg operation using big decoder
5799 pipe_class ialu_reg_long_fat(eRegL dst) %{
5800 instruction_count(2);
5801 dst : S4(write);
5802 dst : S3(read);
5803 D0 : S0(2); // big decoder only; twice
5804 ALU : S3(2); // any 2 alus
5805 %}
5807 // Integer ALU reg-reg operation
5808 pipe_class ialu_reg_reg(eRegI dst, eRegI src) %{
5809 single_instruction;
5810 dst : S4(write);
5811 src : S3(read);
5812 DECODE : S0; // any decoder
5813 ALU : S3; // any alu
5814 %}
5816 // Long ALU reg-reg operation
5817 pipe_class ialu_reg_reg_long(eRegL dst, eRegL src) %{
5818 instruction_count(2);
5819 dst : S4(write);
5820 src : S3(read);
5821 DECODE : S0(2); // any 2 decoders
5822 ALU : S3(2); // both alus
5823 %}
5825 // Integer ALU reg-reg operation
5826 pipe_class ialu_reg_reg_fat(eRegI dst, memory src) %{
5827 single_instruction;
5828 dst : S4(write);
5829 src : S3(read);
5830 D0 : S0; // big decoder only
5831 ALU : S3; // any alu
5832 %}
5834 // Long ALU reg-reg operation
5835 pipe_class ialu_reg_reg_long_fat(eRegL dst, eRegL src) %{
5836 instruction_count(2);
5837 dst : S4(write);
5838 src : S3(read);
5839 D0 : S0(2); // big decoder only; twice
5840 ALU : S3(2); // both alus
5841 %}
5843 // Integer ALU reg-mem operation
5844 pipe_class ialu_reg_mem(eRegI dst, memory mem) %{
5845 single_instruction;
5846 dst : S5(write);
5847 mem : S3(read);
5848 D0 : S0; // big decoder only
5849 ALU : S4; // any alu
5850 MEM : S3; // any mem
5851 %}
5853 // Long ALU reg-mem operation
5854 pipe_class ialu_reg_long_mem(eRegL dst, load_long_memory mem) %{
5855 instruction_count(2);
5856 dst : S5(write);
5857 mem : S3(read);
5858 D0 : S0(2); // big decoder only; twice
5859 ALU : S4(2); // any 2 alus
5860 MEM : S3(2); // both mems
5861 %}
5863 // Integer mem operation (prefetch)
5864 pipe_class ialu_mem(memory mem)
5865 %{
5866 single_instruction;
5867 mem : S3(read);
5868 D0 : S0; // big decoder only
5869 MEM : S3; // any mem
5870 %}
5872 // Integer Store to Memory
5873 pipe_class ialu_mem_reg(memory mem, eRegI src) %{
5874 single_instruction;
5875 mem : S3(read);
5876 src : S5(read);
5877 D0 : S0; // big decoder only
5878 ALU : S4; // any alu
5879 MEM : S3;
5880 %}
5882 // Long Store to Memory
5883 pipe_class ialu_mem_long_reg(memory mem, eRegL src) %{
5884 instruction_count(2);
5885 mem : S3(read);
5886 src : S5(read);
5887 D0 : S0(2); // big decoder only; twice
5888 ALU : S4(2); // any 2 alus
5889 MEM : S3(2); // Both mems
5890 %}
5892 // Integer Store to Memory
5893 pipe_class ialu_mem_imm(memory mem) %{
5894 single_instruction;
5895 mem : S3(read);
5896 D0 : S0; // big decoder only
5897 ALU : S4; // any alu
5898 MEM : S3;
5899 %}
5901 // Integer ALU0 reg-reg operation
5902 pipe_class ialu_reg_reg_alu0(eRegI dst, eRegI src) %{
5903 single_instruction;
5904 dst : S4(write);
5905 src : S3(read);
5906 D0 : S0; // Big decoder only
5907 ALU0 : S3; // only alu0
5908 %}
5910 // Integer ALU0 reg-mem operation
5911 pipe_class ialu_reg_mem_alu0(eRegI dst, memory mem) %{
5912 single_instruction;
5913 dst : S5(write);
5914 mem : S3(read);
5915 D0 : S0; // big decoder only
5916 ALU0 : S4; // ALU0 only
5917 MEM : S3; // any mem
5918 %}
5920 // Integer ALU reg-reg operation
5921 pipe_class ialu_cr_reg_reg(eFlagsReg cr, eRegI src1, eRegI src2) %{
5922 single_instruction;
5923 cr : S4(write);
5924 src1 : S3(read);
5925 src2 : S3(read);
5926 DECODE : S0; // any decoder
5927 ALU : S3; // any alu
5928 %}
5930 // Integer ALU reg-imm operation
5931 pipe_class ialu_cr_reg_imm(eFlagsReg cr, eRegI src1) %{
5932 single_instruction;
5933 cr : S4(write);
5934 src1 : S3(read);
5935 DECODE : S0; // any decoder
5936 ALU : S3; // any alu
5937 %}
5939 // Integer ALU reg-mem operation
5940 pipe_class ialu_cr_reg_mem(eFlagsReg cr, eRegI src1, memory src2) %{
5941 single_instruction;
5942 cr : S4(write);
5943 src1 : S3(read);
5944 src2 : S3(read);
5945 D0 : S0; // big decoder only
5946 ALU : S4; // any alu
5947 MEM : S3;
5948 %}
5950 // Conditional move reg-reg
5951 pipe_class pipe_cmplt( eRegI p, eRegI q, eRegI y ) %{
5952 instruction_count(4);
5953 y : S4(read);
5954 q : S3(read);
5955 p : S3(read);
5956 DECODE : S0(4); // any decoder
5957 %}
5959 // Conditional move reg-reg
5960 pipe_class pipe_cmov_reg( eRegI dst, eRegI src, eFlagsReg cr ) %{
5961 single_instruction;
5962 dst : S4(write);
5963 src : S3(read);
5964 cr : S3(read);
5965 DECODE : S0; // any decoder
5966 %}
5968 // Conditional move reg-mem
5969 pipe_class pipe_cmov_mem( eFlagsReg cr, eRegI dst, memory src) %{
5970 single_instruction;
5971 dst : S4(write);
5972 src : S3(read);
5973 cr : S3(read);
5974 DECODE : S0; // any decoder
5975 MEM : S3;
5976 %}
5978 // Conditional move reg-reg long
5979 pipe_class pipe_cmov_reg_long( eFlagsReg cr, eRegL dst, eRegL src) %{
5980 single_instruction;
5981 dst : S4(write);
5982 src : S3(read);
5983 cr : S3(read);
5984 DECODE : S0(2); // any 2 decoders
5985 %}
5987 // Conditional move double reg-reg
5988 pipe_class pipe_cmovD_reg( eFlagsReg cr, regDPR1 dst, regD src) %{
5989 single_instruction;
5990 dst : S4(write);
5991 src : S3(read);
5992 cr : S3(read);
5993 DECODE : S0; // any decoder
5994 %}
5996 // Float reg-reg operation
5997 pipe_class fpu_reg(regD dst) %{
5998 instruction_count(2);
5999 dst : S3(read);
6000 DECODE : S0(2); // any 2 decoders
6001 FPU : S3;
6002 %}
6004 // Float reg-reg operation
6005 pipe_class fpu_reg_reg(regD dst, regD src) %{
6006 instruction_count(2);
6007 dst : S4(write);
6008 src : S3(read);
6009 DECODE : S0(2); // any 2 decoders
6010 FPU : S3;
6011 %}
6013 // Float reg-reg operation
6014 pipe_class fpu_reg_reg_reg(regD dst, regD src1, regD src2) %{
6015 instruction_count(3);
6016 dst : S4(write);
6017 src1 : S3(read);
6018 src2 : S3(read);
6019 DECODE : S0(3); // any 3 decoders
6020 FPU : S3(2);
6021 %}
6023 // Float reg-reg operation
6024 pipe_class fpu_reg_reg_reg_reg(regD dst, regD src1, regD src2, regD src3) %{
6025 instruction_count(4);
6026 dst : S4(write);
6027 src1 : S3(read);
6028 src2 : S3(read);
6029 src3 : S3(read);
6030 DECODE : S0(4); // any 3 decoders
6031 FPU : S3(2);
6032 %}
6034 // Float reg-reg operation
6035 pipe_class fpu_reg_mem_reg_reg(regD dst, memory src1, regD src2, regD src3) %{
6036 instruction_count(4);
6037 dst : S4(write);
6038 src1 : S3(read);
6039 src2 : S3(read);
6040 src3 : S3(read);
6041 DECODE : S1(3); // any 3 decoders
6042 D0 : S0; // Big decoder only
6043 FPU : S3(2);
6044 MEM : S3;
6045 %}
6047 // Float reg-mem operation
6048 pipe_class fpu_reg_mem(regD dst, memory mem) %{
6049 instruction_count(2);
6050 dst : S5(write);
6051 mem : S3(read);
6052 D0 : S0; // big decoder only
6053 DECODE : S1; // any decoder for FPU POP
6054 FPU : S4;
6055 MEM : S3; // any mem
6056 %}
6058 // Float reg-mem operation
6059 pipe_class fpu_reg_reg_mem(regD dst, regD src1, memory mem) %{
6060 instruction_count(3);
6061 dst : S5(write);
6062 src1 : S3(read);
6063 mem : S3(read);
6064 D0 : S0; // big decoder only
6065 DECODE : S1(2); // any decoder for FPU POP
6066 FPU : S4;
6067 MEM : S3; // any mem
6068 %}
6070 // Float mem-reg operation
6071 pipe_class fpu_mem_reg(memory mem, regD src) %{
6072 instruction_count(2);
6073 src : S5(read);
6074 mem : S3(read);
6075 DECODE : S0; // any decoder for FPU PUSH
6076 D0 : S1; // big decoder only
6077 FPU : S4;
6078 MEM : S3; // any mem
6079 %}
6081 pipe_class fpu_mem_reg_reg(memory mem, regD src1, regD src2) %{
6082 instruction_count(3);
6083 src1 : S3(read);
6084 src2 : S3(read);
6085 mem : S3(read);
6086 DECODE : S0(2); // any decoder for FPU PUSH
6087 D0 : S1; // big decoder only
6088 FPU : S4;
6089 MEM : S3; // any mem
6090 %}
6092 pipe_class fpu_mem_reg_mem(memory mem, regD src1, memory src2) %{
6093 instruction_count(3);
6094 src1 : S3(read);
6095 src2 : S3(read);
6096 mem : S4(read);
6097 DECODE : S0; // any decoder for FPU PUSH
6098 D0 : S0(2); // big decoder only
6099 FPU : S4;
6100 MEM : S3(2); // any mem
6101 %}
6103 pipe_class fpu_mem_mem(memory dst, memory src1) %{
6104 instruction_count(2);
6105 src1 : S3(read);
6106 dst : S4(read);
6107 D0 : S0(2); // big decoder only
6108 MEM : S3(2); // any mem
6109 %}
6111 pipe_class fpu_mem_mem_mem(memory dst, memory src1, memory src2) %{
6112 instruction_count(3);
6113 src1 : S3(read);
6114 src2 : S3(read);
6115 dst : S4(read);
6116 D0 : S0(3); // big decoder only
6117 FPU : S4;
6118 MEM : S3(3); // any mem
6119 %}
6121 pipe_class fpu_mem_reg_con(memory mem, regD src1) %{
6122 instruction_count(3);
6123 src1 : S4(read);
6124 mem : S4(read);
6125 DECODE : S0; // any decoder for FPU PUSH
6126 D0 : S0(2); // big decoder only
6127 FPU : S4;
6128 MEM : S3(2); // any mem
6129 %}
6131 // Float load constant
6132 pipe_class fpu_reg_con(regD dst) %{
6133 instruction_count(2);
6134 dst : S5(write);
6135 D0 : S0; // big decoder only for the load
6136 DECODE : S1; // any decoder for FPU POP
6137 FPU : S4;
6138 MEM : S3; // any mem
6139 %}
6141 // Float load constant
6142 pipe_class fpu_reg_reg_con(regD dst, regD src) %{
6143 instruction_count(3);
6144 dst : S5(write);
6145 src : S3(read);
6146 D0 : S0; // big decoder only for the load
6147 DECODE : S1(2); // any decoder for FPU POP
6148 FPU : S4;
6149 MEM : S3; // any mem
6150 %}
6152 // UnConditional branch
6153 pipe_class pipe_jmp( label labl ) %{
6154 single_instruction;
6155 BR : S3;
6156 %}
6158 // Conditional branch
6159 pipe_class pipe_jcc( cmpOp cmp, eFlagsReg cr, label labl ) %{
6160 single_instruction;
6161 cr : S1(read);
6162 BR : S3;
6163 %}
6165 // Allocation idiom
6166 pipe_class pipe_cmpxchg( eRegP dst, eRegP heap_ptr ) %{
6167 instruction_count(1); force_serialization;
6168 fixed_latency(6);
6169 heap_ptr : S3(read);
6170 DECODE : S0(3);
6171 D0 : S2;
6172 MEM : S3;
6173 ALU : S3(2);
6174 dst : S5(write);
6175 BR : S5;
6176 %}
6178 // Generic big/slow expanded idiom
6179 pipe_class pipe_slow( ) %{
6180 instruction_count(10); multiple_bundles; force_serialization;
6181 fixed_latency(100);
6182 D0 : S0(2);
6183 MEM : S3(2);
6184 %}
6186 // The real do-nothing guy
6187 pipe_class empty( ) %{
6188 instruction_count(0);
6189 %}
6191 // Define the class for the Nop node
6192 define %{
6193 MachNop = empty;
6194 %}
6196 %}
6198 //----------INSTRUCTIONS-------------------------------------------------------
6199 //
6200 // match -- States which machine-independent subtree may be replaced
6201 // by this instruction.
6202 // ins_cost -- The estimated cost of this instruction is used by instruction
6203 // selection to identify a minimum cost tree of machine
6204 // instructions that matches a tree of machine-independent
6205 // instructions.
6206 // format -- A string providing the disassembly for this instruction.
6207 // The value of an instruction's operand may be inserted
6208 // by referring to it with a '$' prefix.
6209 // opcode -- Three instruction opcodes may be provided. These are referred
6210 // to within an encode class as $primary, $secondary, and $tertiary
6211 // respectively. The primary opcode is commonly used to
6212 // indicate the type of machine instruction, while secondary
6213 // and tertiary are often used for prefix options or addressing
6214 // modes.
6215 // ins_encode -- A list of encode classes with parameters. The encode class
6216 // name must have been defined in an 'enc_class' specification
6217 // in the encode section of the architecture description.
6219 //----------BSWAP-Instruction--------------------------------------------------
6220 instruct bytes_reverse_int(eRegI dst) %{
6221 match(Set dst (ReverseBytesI dst));
6223 format %{ "BSWAP $dst" %}
6224 opcode(0x0F, 0xC8);
6225 ins_encode( OpcP, OpcSReg(dst) );
6226 ins_pipe( ialu_reg );
6227 %}
6229 instruct bytes_reverse_long(eRegL dst) %{
6230 match(Set dst (ReverseBytesL dst));
6232 format %{ "BSWAP $dst.lo\n\t"
6233 "BSWAP $dst.hi\n\t"
6234 "XCHG $dst.lo $dst.hi" %}
6236 ins_cost(125);
6237 ins_encode( bswap_long_bytes(dst) );
6238 ins_pipe( ialu_reg_reg);
6239 %}
6242 //----------Load/Store/Move Instructions---------------------------------------
6243 //----------Load Instructions--------------------------------------------------
6244 // Load Byte (8bit signed)
6245 instruct loadB(xRegI dst, memory mem) %{
6246 match(Set dst (LoadB mem));
6248 ins_cost(125);
6249 format %{ "MOVSX8 $dst,$mem" %}
6250 opcode(0xBE, 0x0F);
6251 ins_encode( OpcS, OpcP, RegMem(dst,mem));
6252 ins_pipe( ialu_reg_mem );
6253 %}
6255 // Load Byte (8bit UNsigned)
6256 instruct loadUB(xRegI dst, memory mem, immI_255 bytemask) %{
6257 match(Set dst (AndI (LoadB mem) bytemask));
6259 ins_cost(125);
6260 format %{ "MOVZX8 $dst,$mem" %}
6261 opcode(0xB6, 0x0F);
6262 ins_encode( OpcS, OpcP, RegMem(dst,mem));
6263 ins_pipe( ialu_reg_mem );
6264 %}
6266 // Load Char (16bit unsigned)
6267 instruct loadC(eRegI dst, memory mem) %{
6268 match(Set dst (LoadC mem));
6270 ins_cost(125);
6271 format %{ "MOVZX $dst,$mem" %}
6272 opcode(0xB7, 0x0F);
6273 ins_encode( OpcS, OpcP, RegMem(dst,mem));
6274 ins_pipe( ialu_reg_mem );
6275 %}
6277 // Load Integer
6278 instruct loadI(eRegI dst, memory mem) %{
6279 match(Set dst (LoadI mem));
6281 ins_cost(125);
6282 format %{ "MOV $dst,$mem" %}
6283 opcode(0x8B);
6284 ins_encode( OpcP, RegMem(dst,mem));
6285 ins_pipe( ialu_reg_mem );
6286 %}
6288 // Load Long. Cannot clobber address while loading, so restrict address
6289 // register to ESI
6290 instruct loadL(eRegL dst, load_long_memory mem) %{
6291 predicate(!((LoadLNode*)n)->require_atomic_access());
6292 match(Set dst (LoadL mem));
6294 ins_cost(250);
6295 format %{ "MOV $dst.lo,$mem\n\t"
6296 "MOV $dst.hi,$mem+4" %}
6297 opcode(0x8B, 0x8B);
6298 ins_encode( OpcP, RegMem(dst,mem), OpcS, RegMem_Hi(dst,mem));
6299 ins_pipe( ialu_reg_long_mem );
6300 %}
6302 // Volatile Load Long. Must be atomic, so do 64-bit FILD
6303 // then store it down to the stack and reload on the int
6304 // side.
6305 instruct loadL_volatile(stackSlotL dst, memory mem) %{
6306 predicate(UseSSE<=1 && ((LoadLNode*)n)->require_atomic_access());
6307 match(Set dst (LoadL mem));
6309 ins_cost(200);
6310 format %{ "FILD $mem\t# Atomic volatile long load\n\t"
6311 "FISTp $dst" %}
6312 ins_encode(enc_loadL_volatile(mem,dst));
6313 ins_pipe( fpu_reg_mem );
6314 %}
6316 instruct loadLX_volatile(stackSlotL dst, memory mem, regXD tmp) %{
6317 predicate(UseSSE>=2 && ((LoadLNode*)n)->require_atomic_access());
6318 match(Set dst (LoadL mem));
6319 effect(TEMP tmp);
6320 ins_cost(180);
6321 format %{ "MOVSD $tmp,$mem\t# Atomic volatile long load\n\t"
6322 "MOVSD $dst,$tmp" %}
6323 ins_encode(enc_loadLX_volatile(mem, dst, tmp));
6324 ins_pipe( pipe_slow );
6325 %}
6327 instruct loadLX_reg_volatile(eRegL dst, memory mem, regXD tmp) %{
6328 predicate(UseSSE>=2 && ((LoadLNode*)n)->require_atomic_access());
6329 match(Set dst (LoadL mem));
6330 effect(TEMP tmp);
6331 ins_cost(160);
6332 format %{ "MOVSD $tmp,$mem\t# Atomic volatile long load\n\t"
6333 "MOVD $dst.lo,$tmp\n\t"
6334 "PSRLQ $tmp,32\n\t"
6335 "MOVD $dst.hi,$tmp" %}
6336 ins_encode(enc_loadLX_reg_volatile(mem, dst, tmp));
6337 ins_pipe( pipe_slow );
6338 %}
6340 // Load Range
6341 instruct loadRange(eRegI dst, memory mem) %{
6342 match(Set dst (LoadRange mem));
6344 ins_cost(125);
6345 format %{ "MOV $dst,$mem" %}
6346 opcode(0x8B);
6347 ins_encode( OpcP, RegMem(dst,mem));
6348 ins_pipe( ialu_reg_mem );
6349 %}
6352 // Load Pointer
6353 instruct loadP(eRegP dst, memory mem) %{
6354 match(Set dst (LoadP mem));
6356 ins_cost(125);
6357 format %{ "MOV $dst,$mem" %}
6358 opcode(0x8B);
6359 ins_encode( OpcP, RegMem(dst,mem));
6360 ins_pipe( ialu_reg_mem );
6361 %}
6363 // Load Klass Pointer
6364 instruct loadKlass(eRegP dst, memory mem) %{
6365 match(Set dst (LoadKlass mem));
6367 ins_cost(125);
6368 format %{ "MOV $dst,$mem" %}
6369 opcode(0x8B);
6370 ins_encode( OpcP, RegMem(dst,mem));
6371 ins_pipe( ialu_reg_mem );
6372 %}
6374 // Load Short (16bit signed)
6375 instruct loadS(eRegI dst, memory mem) %{
6376 match(Set dst (LoadS mem));
6378 ins_cost(125);
6379 format %{ "MOVSX $dst,$mem" %}
6380 opcode(0xBF, 0x0F);
6381 ins_encode( OpcS, OpcP, RegMem(dst,mem));
6382 ins_pipe( ialu_reg_mem );
6383 %}
6385 // Load Double
6386 instruct loadD(regD dst, memory mem) %{
6387 predicate(UseSSE<=1);
6388 match(Set dst (LoadD mem));
6390 ins_cost(150);
6391 format %{ "FLD_D ST,$mem\n\t"
6392 "FSTP $dst" %}
6393 opcode(0xDD); /* DD /0 */
6394 ins_encode( OpcP, RMopc_Mem(0x00,mem),
6395 Pop_Reg_D(dst) );
6396 ins_pipe( fpu_reg_mem );
6397 %}
6399 // Load Double to XMM
6400 instruct loadXD(regXD dst, memory mem) %{
6401 predicate(UseSSE>=2 && UseXmmLoadAndClearUpper);
6402 match(Set dst (LoadD mem));
6403 ins_cost(145);
6404 format %{ "MOVSD $dst,$mem" %}
6405 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x10), RegMem(dst,mem));
6406 ins_pipe( pipe_slow );
6407 %}
6409 instruct loadXD_partial(regXD dst, memory mem) %{
6410 predicate(UseSSE>=2 && !UseXmmLoadAndClearUpper);
6411 match(Set dst (LoadD mem));
6412 ins_cost(145);
6413 format %{ "MOVLPD $dst,$mem" %}
6414 ins_encode( Opcode(0x66), Opcode(0x0F), Opcode(0x12), RegMem(dst,mem));
6415 ins_pipe( pipe_slow );
6416 %}
6418 // Load to XMM register (single-precision floating point)
6419 // MOVSS instruction
6420 instruct loadX(regX dst, memory mem) %{
6421 predicate(UseSSE>=1);
6422 match(Set dst (LoadF mem));
6423 ins_cost(145);
6424 format %{ "MOVSS $dst,$mem" %}
6425 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x10), RegMem(dst,mem));
6426 ins_pipe( pipe_slow );
6427 %}
6429 // Load Float
6430 instruct loadF(regF dst, memory mem) %{
6431 predicate(UseSSE==0);
6432 match(Set dst (LoadF mem));
6434 ins_cost(150);
6435 format %{ "FLD_S ST,$mem\n\t"
6436 "FSTP $dst" %}
6437 opcode(0xD9); /* D9 /0 */
6438 ins_encode( OpcP, RMopc_Mem(0x00,mem),
6439 Pop_Reg_F(dst) );
6440 ins_pipe( fpu_reg_mem );
6441 %}
6443 // Load Aligned Packed Byte to XMM register
6444 instruct loadA8B(regXD dst, memory mem) %{
6445 predicate(UseSSE>=1);
6446 match(Set dst (Load8B mem));
6447 ins_cost(125);
6448 format %{ "MOVQ $dst,$mem\t! packed8B" %}
6449 ins_encode( movq_ld(dst, mem));
6450 ins_pipe( pipe_slow );
6451 %}
6453 // Load Aligned Packed Short to XMM register
6454 instruct loadA4S(regXD dst, memory mem) %{
6455 predicate(UseSSE>=1);
6456 match(Set dst (Load4S mem));
6457 ins_cost(125);
6458 format %{ "MOVQ $dst,$mem\t! packed4S" %}
6459 ins_encode( movq_ld(dst, mem));
6460 ins_pipe( pipe_slow );
6461 %}
6463 // Load Aligned Packed Char to XMM register
6464 instruct loadA4C(regXD dst, memory mem) %{
6465 predicate(UseSSE>=1);
6466 match(Set dst (Load4C mem));
6467 ins_cost(125);
6468 format %{ "MOVQ $dst,$mem\t! packed4C" %}
6469 ins_encode( movq_ld(dst, mem));
6470 ins_pipe( pipe_slow );
6471 %}
6473 // Load Aligned Packed Integer to XMM register
6474 instruct load2IU(regXD dst, memory mem) %{
6475 predicate(UseSSE>=1);
6476 match(Set dst (Load2I mem));
6477 ins_cost(125);
6478 format %{ "MOVQ $dst,$mem\t! packed2I" %}
6479 ins_encode( movq_ld(dst, mem));
6480 ins_pipe( pipe_slow );
6481 %}
6483 // Load Aligned Packed Single to XMM
6484 instruct loadA2F(regXD dst, memory mem) %{
6485 predicate(UseSSE>=1);
6486 match(Set dst (Load2F mem));
6487 ins_cost(145);
6488 format %{ "MOVQ $dst,$mem\t! packed2F" %}
6489 ins_encode( movq_ld(dst, mem));
6490 ins_pipe( pipe_slow );
6491 %}
6493 // Load Effective Address
6494 instruct leaP8(eRegP dst, indOffset8 mem) %{
6495 match(Set dst mem);
6497 ins_cost(110);
6498 format %{ "LEA $dst,$mem" %}
6499 opcode(0x8D);
6500 ins_encode( OpcP, RegMem(dst,mem));
6501 ins_pipe( ialu_reg_reg_fat );
6502 %}
6504 instruct leaP32(eRegP dst, indOffset32 mem) %{
6505 match(Set dst mem);
6507 ins_cost(110);
6508 format %{ "LEA $dst,$mem" %}
6509 opcode(0x8D);
6510 ins_encode( OpcP, RegMem(dst,mem));
6511 ins_pipe( ialu_reg_reg_fat );
6512 %}
6514 instruct leaPIdxOff(eRegP dst, indIndexOffset mem) %{
6515 match(Set dst mem);
6517 ins_cost(110);
6518 format %{ "LEA $dst,$mem" %}
6519 opcode(0x8D);
6520 ins_encode( OpcP, RegMem(dst,mem));
6521 ins_pipe( ialu_reg_reg_fat );
6522 %}
6524 instruct leaPIdxScale(eRegP dst, indIndexScale mem) %{
6525 match(Set dst mem);
6527 ins_cost(110);
6528 format %{ "LEA $dst,$mem" %}
6529 opcode(0x8D);
6530 ins_encode( OpcP, RegMem(dst,mem));
6531 ins_pipe( ialu_reg_reg_fat );
6532 %}
6534 instruct leaPIdxScaleOff(eRegP dst, indIndexScaleOffset mem) %{
6535 match(Set dst mem);
6537 ins_cost(110);
6538 format %{ "LEA $dst,$mem" %}
6539 opcode(0x8D);
6540 ins_encode( OpcP, RegMem(dst,mem));
6541 ins_pipe( ialu_reg_reg_fat );
6542 %}
6544 // Load Constant
6545 instruct loadConI(eRegI dst, immI src) %{
6546 match(Set dst src);
6548 format %{ "MOV $dst,$src" %}
6549 ins_encode( LdImmI(dst, src) );
6550 ins_pipe( ialu_reg_fat );
6551 %}
6553 // Load Constant zero
6554 instruct loadConI0(eRegI dst, immI0 src, eFlagsReg cr) %{
6555 match(Set dst src);
6556 effect(KILL cr);
6558 ins_cost(50);
6559 format %{ "XOR $dst,$dst" %}
6560 opcode(0x33); /* + rd */
6561 ins_encode( OpcP, RegReg( dst, dst ) );
6562 ins_pipe( ialu_reg );
6563 %}
6565 instruct loadConP(eRegP dst, immP src) %{
6566 match(Set dst src);
6568 format %{ "MOV $dst,$src" %}
6569 opcode(0xB8); /* + rd */
6570 ins_encode( LdImmP(dst, src) );
6571 ins_pipe( ialu_reg_fat );
6572 %}
6574 instruct loadConL(eRegL dst, immL src, eFlagsReg cr) %{
6575 match(Set dst src);
6576 effect(KILL cr);
6577 ins_cost(200);
6578 format %{ "MOV $dst.lo,$src.lo\n\t"
6579 "MOV $dst.hi,$src.hi" %}
6580 opcode(0xB8);
6581 ins_encode( LdImmL_Lo(dst, src), LdImmL_Hi(dst, src) );
6582 ins_pipe( ialu_reg_long_fat );
6583 %}
6585 instruct loadConL0(eRegL dst, immL0 src, eFlagsReg cr) %{
6586 match(Set dst src);
6587 effect(KILL cr);
6588 ins_cost(150);
6589 format %{ "XOR $dst.lo,$dst.lo\n\t"
6590 "XOR $dst.hi,$dst.hi" %}
6591 opcode(0x33,0x33);
6592 ins_encode( RegReg_Lo(dst,dst), RegReg_Hi(dst, dst) );
6593 ins_pipe( ialu_reg_long );
6594 %}
6596 // The instruction usage is guarded by predicate in operand immF().
6597 instruct loadConF(regF dst, immF src) %{
6598 match(Set dst src);
6599 ins_cost(125);
6601 format %{ "FLD_S ST,$src\n\t"
6602 "FSTP $dst" %}
6603 opcode(0xD9, 0x00); /* D9 /0 */
6604 ins_encode(LdImmF(src), Pop_Reg_F(dst) );
6605 ins_pipe( fpu_reg_con );
6606 %}
6608 // The instruction usage is guarded by predicate in operand immXF().
6609 instruct loadConX(regX dst, immXF con) %{
6610 match(Set dst con);
6611 ins_cost(125);
6612 format %{ "MOVSS $dst,[$con]" %}
6613 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x10), LdImmX(dst, con));
6614 ins_pipe( pipe_slow );
6615 %}
6617 // The instruction usage is guarded by predicate in operand immXF0().
6618 instruct loadConX0(regX dst, immXF0 src) %{
6619 match(Set dst src);
6620 ins_cost(100);
6621 format %{ "XORPS $dst,$dst\t# float 0.0" %}
6622 ins_encode( Opcode(0x0F), Opcode(0x57), RegReg(dst,dst));
6623 ins_pipe( pipe_slow );
6624 %}
6626 // The instruction usage is guarded by predicate in operand immD().
6627 instruct loadConD(regD dst, immD src) %{
6628 match(Set dst src);
6629 ins_cost(125);
6631 format %{ "FLD_D ST,$src\n\t"
6632 "FSTP $dst" %}
6633 ins_encode(LdImmD(src), Pop_Reg_D(dst) );
6634 ins_pipe( fpu_reg_con );
6635 %}
6637 // The instruction usage is guarded by predicate in operand immXD().
6638 instruct loadConXD(regXD dst, immXD con) %{
6639 match(Set dst con);
6640 ins_cost(125);
6641 format %{ "MOVSD $dst,[$con]" %}
6642 ins_encode(load_conXD(dst, con));
6643 ins_pipe( pipe_slow );
6644 %}
6646 // The instruction usage is guarded by predicate in operand immXD0().
6647 instruct loadConXD0(regXD dst, immXD0 src) %{
6648 match(Set dst src);
6649 ins_cost(100);
6650 format %{ "XORPD $dst,$dst\t# double 0.0" %}
6651 ins_encode( Opcode(0x66), Opcode(0x0F), Opcode(0x57), RegReg(dst,dst));
6652 ins_pipe( pipe_slow );
6653 %}
6655 // Load Stack Slot
6656 instruct loadSSI(eRegI dst, stackSlotI src) %{
6657 match(Set dst src);
6658 ins_cost(125);
6660 format %{ "MOV $dst,$src" %}
6661 opcode(0x8B);
6662 ins_encode( OpcP, RegMem(dst,src));
6663 ins_pipe( ialu_reg_mem );
6664 %}
6666 instruct loadSSL(eRegL dst, stackSlotL src) %{
6667 match(Set dst src);
6669 ins_cost(200);
6670 format %{ "MOV $dst,$src.lo\n\t"
6671 "MOV $dst+4,$src.hi" %}
6672 opcode(0x8B, 0x8B);
6673 ins_encode( OpcP, RegMem( dst, src ), OpcS, RegMem_Hi( dst, src ) );
6674 ins_pipe( ialu_mem_long_reg );
6675 %}
6677 // Load Stack Slot
6678 instruct loadSSP(eRegP dst, stackSlotP src) %{
6679 match(Set dst src);
6680 ins_cost(125);
6682 format %{ "MOV $dst,$src" %}
6683 opcode(0x8B);
6684 ins_encode( OpcP, RegMem(dst,src));
6685 ins_pipe( ialu_reg_mem );
6686 %}
6688 // Load Stack Slot
6689 instruct loadSSF(regF dst, stackSlotF src) %{
6690 match(Set dst src);
6691 ins_cost(125);
6693 format %{ "FLD_S $src\n\t"
6694 "FSTP $dst" %}
6695 opcode(0xD9); /* D9 /0, FLD m32real */
6696 ins_encode( OpcP, RMopc_Mem_no_oop(0x00,src),
6697 Pop_Reg_F(dst) );
6698 ins_pipe( fpu_reg_mem );
6699 %}
6701 // Load Stack Slot
6702 instruct loadSSD(regD dst, stackSlotD src) %{
6703 match(Set dst src);
6704 ins_cost(125);
6706 format %{ "FLD_D $src\n\t"
6707 "FSTP $dst" %}
6708 opcode(0xDD); /* DD /0, FLD m64real */
6709 ins_encode( OpcP, RMopc_Mem_no_oop(0x00,src),
6710 Pop_Reg_D(dst) );
6711 ins_pipe( fpu_reg_mem );
6712 %}
6714 // Prefetch instructions.
6715 // Must be safe to execute with invalid address (cannot fault).
6717 instruct prefetchr0( memory mem ) %{
6718 predicate(UseSSE==0 && !VM_Version::supports_3dnow());
6719 match(PrefetchRead mem);
6720 ins_cost(0);
6721 size(0);
6722 format %{ "PREFETCHR (non-SSE is empty encoding)" %}
6723 ins_encode();
6724 ins_pipe(empty);
6725 %}
6727 instruct prefetchr( memory mem ) %{
6728 predicate(UseSSE==0 && VM_Version::supports_3dnow() || ReadPrefetchInstr==3);
6729 match(PrefetchRead mem);
6730 ins_cost(100);
6732 format %{ "PREFETCHR $mem\t! Prefetch into level 1 cache for read" %}
6733 opcode(0x0F, 0x0d); /* Opcode 0F 0d /0 */
6734 ins_encode(OpcP, OpcS, RMopc_Mem(0x00,mem));
6735 ins_pipe(ialu_mem);
6736 %}
6738 instruct prefetchrNTA( memory mem ) %{
6739 predicate(UseSSE>=1 && ReadPrefetchInstr==0);
6740 match(PrefetchRead mem);
6741 ins_cost(100);
6743 format %{ "PREFETCHNTA $mem\t! Prefetch into non-temporal cache for read" %}
6744 opcode(0x0F, 0x18); /* Opcode 0F 18 /0 */
6745 ins_encode(OpcP, OpcS, RMopc_Mem(0x00,mem));
6746 ins_pipe(ialu_mem);
6747 %}
6749 instruct prefetchrT0( memory mem ) %{
6750 predicate(UseSSE>=1 && ReadPrefetchInstr==1);
6751 match(PrefetchRead mem);
6752 ins_cost(100);
6754 format %{ "PREFETCHT0 $mem\t! Prefetch into L1 and L2 caches for read" %}
6755 opcode(0x0F, 0x18); /* Opcode 0F 18 /1 */
6756 ins_encode(OpcP, OpcS, RMopc_Mem(0x01,mem));
6757 ins_pipe(ialu_mem);
6758 %}
6760 instruct prefetchrT2( memory mem ) %{
6761 predicate(UseSSE>=1 && ReadPrefetchInstr==2);
6762 match(PrefetchRead mem);
6763 ins_cost(100);
6765 format %{ "PREFETCHT2 $mem\t! Prefetch into L2 cache for read" %}
6766 opcode(0x0F, 0x18); /* Opcode 0F 18 /3 */
6767 ins_encode(OpcP, OpcS, RMopc_Mem(0x03,mem));
6768 ins_pipe(ialu_mem);
6769 %}
6771 instruct prefetchw0( memory mem ) %{
6772 predicate(UseSSE==0 && !VM_Version::supports_3dnow());
6773 match(PrefetchWrite mem);
6774 ins_cost(0);
6775 size(0);
6776 format %{ "Prefetch (non-SSE is empty encoding)" %}
6777 ins_encode();
6778 ins_pipe(empty);
6779 %}
6781 instruct prefetchw( memory mem ) %{
6782 predicate(UseSSE==0 && VM_Version::supports_3dnow() || AllocatePrefetchInstr==3);
6783 match( PrefetchWrite mem );
6784 ins_cost(100);
6786 format %{ "PREFETCHW $mem\t! Prefetch into L1 cache and mark modified" %}
6787 opcode(0x0F, 0x0D); /* Opcode 0F 0D /1 */
6788 ins_encode(OpcP, OpcS, RMopc_Mem(0x01,mem));
6789 ins_pipe(ialu_mem);
6790 %}
6792 instruct prefetchwNTA( memory mem ) %{
6793 predicate(UseSSE>=1 && AllocatePrefetchInstr==0);
6794 match(PrefetchWrite mem);
6795 ins_cost(100);
6797 format %{ "PREFETCHNTA $mem\t! Prefetch into non-temporal cache for write" %}
6798 opcode(0x0F, 0x18); /* Opcode 0F 18 /0 */
6799 ins_encode(OpcP, OpcS, RMopc_Mem(0x00,mem));
6800 ins_pipe(ialu_mem);
6801 %}
6803 instruct prefetchwT0( memory mem ) %{
6804 predicate(UseSSE>=1 && AllocatePrefetchInstr==1);
6805 match(PrefetchWrite mem);
6806 ins_cost(100);
6808 format %{ "PREFETCHT0 $mem\t! Prefetch into L1 and L2 caches for write" %}
6809 opcode(0x0F, 0x18); /* Opcode 0F 18 /1 */
6810 ins_encode(OpcP, OpcS, RMopc_Mem(0x01,mem));
6811 ins_pipe(ialu_mem);
6812 %}
6814 instruct prefetchwT2( memory mem ) %{
6815 predicate(UseSSE>=1 && AllocatePrefetchInstr==2);
6816 match(PrefetchWrite mem);
6817 ins_cost(100);
6819 format %{ "PREFETCHT2 $mem\t! Prefetch into L2 cache for write" %}
6820 opcode(0x0F, 0x18); /* Opcode 0F 18 /3 */
6821 ins_encode(OpcP, OpcS, RMopc_Mem(0x03,mem));
6822 ins_pipe(ialu_mem);
6823 %}
6825 //----------Store Instructions-------------------------------------------------
6827 // Store Byte
6828 instruct storeB(memory mem, xRegI src) %{
6829 match(Set mem (StoreB mem src));
6831 ins_cost(125);
6832 format %{ "MOV8 $mem,$src" %}
6833 opcode(0x88);
6834 ins_encode( OpcP, RegMem( src, mem ) );
6835 ins_pipe( ialu_mem_reg );
6836 %}
6838 // Store Char/Short
6839 instruct storeC(memory mem, eRegI src) %{
6840 match(Set mem (StoreC mem src));
6842 ins_cost(125);
6843 format %{ "MOV16 $mem,$src" %}
6844 opcode(0x89, 0x66);
6845 ins_encode( OpcS, OpcP, RegMem( src, mem ) );
6846 ins_pipe( ialu_mem_reg );
6847 %}
6849 // Store Integer
6850 instruct storeI(memory mem, eRegI src) %{
6851 match(Set mem (StoreI mem src));
6853 ins_cost(125);
6854 format %{ "MOV $mem,$src" %}
6855 opcode(0x89);
6856 ins_encode( OpcP, RegMem( src, mem ) );
6857 ins_pipe( ialu_mem_reg );
6858 %}
6860 // Store Long
6861 instruct storeL(long_memory mem, eRegL src) %{
6862 predicate(!((StoreLNode*)n)->require_atomic_access());
6863 match(Set mem (StoreL mem src));
6865 ins_cost(200);
6866 format %{ "MOV $mem,$src.lo\n\t"
6867 "MOV $mem+4,$src.hi" %}
6868 opcode(0x89, 0x89);
6869 ins_encode( OpcP, RegMem( src, mem ), OpcS, RegMem_Hi( src, mem ) );
6870 ins_pipe( ialu_mem_long_reg );
6871 %}
6873 // Volatile Store Long. Must be atomic, so move it into
6874 // the FP TOS and then do a 64-bit FIST. Has to probe the
6875 // target address before the store (for null-ptr checks)
6876 // so the memory operand is used twice in the encoding.
6877 instruct storeL_volatile(memory mem, stackSlotL src, eFlagsReg cr ) %{
6878 predicate(UseSSE<=1 && ((StoreLNode*)n)->require_atomic_access());
6879 match(Set mem (StoreL mem src));
6880 effect( KILL cr );
6881 ins_cost(400);
6882 format %{ "CMP $mem,EAX\t# Probe address for implicit null check\n\t"
6883 "FILD $src\n\t"
6884 "FISTp $mem\t # 64-bit atomic volatile long store" %}
6885 opcode(0x3B);
6886 ins_encode( OpcP, RegMem( EAX, mem ), enc_storeL_volatile(mem,src));
6887 ins_pipe( fpu_reg_mem );
6888 %}
6890 instruct storeLX_volatile(memory mem, stackSlotL src, regXD tmp, eFlagsReg cr) %{
6891 predicate(UseSSE>=2 && ((StoreLNode*)n)->require_atomic_access());
6892 match(Set mem (StoreL mem src));
6893 effect( TEMP tmp, KILL cr );
6894 ins_cost(380);
6895 format %{ "CMP $mem,EAX\t# Probe address for implicit null check\n\t"
6896 "MOVSD $tmp,$src\n\t"
6897 "MOVSD $mem,$tmp\t # 64-bit atomic volatile long store" %}
6898 opcode(0x3B);
6899 ins_encode( OpcP, RegMem( EAX, mem ), enc_storeLX_volatile(mem, src, tmp));
6900 ins_pipe( pipe_slow );
6901 %}
6903 instruct storeLX_reg_volatile(memory mem, eRegL src, regXD tmp2, regXD tmp, eFlagsReg cr) %{
6904 predicate(UseSSE>=2 && ((StoreLNode*)n)->require_atomic_access());
6905 match(Set mem (StoreL mem src));
6906 effect( TEMP tmp2 , TEMP tmp, KILL cr );
6907 ins_cost(360);
6908 format %{ "CMP $mem,EAX\t# Probe address for implicit null check\n\t"
6909 "MOVD $tmp,$src.lo\n\t"
6910 "MOVD $tmp2,$src.hi\n\t"
6911 "PUNPCKLDQ $tmp,$tmp2\n\t"
6912 "MOVSD $mem,$tmp\t # 64-bit atomic volatile long store" %}
6913 opcode(0x3B);
6914 ins_encode( OpcP, RegMem( EAX, mem ), enc_storeLX_reg_volatile(mem, src, tmp, tmp2));
6915 ins_pipe( pipe_slow );
6916 %}
6918 // Store Pointer; for storing unknown oops and raw pointers
6919 instruct storeP(memory mem, anyRegP src) %{
6920 match(Set mem (StoreP mem src));
6922 ins_cost(125);
6923 format %{ "MOV $mem,$src" %}
6924 opcode(0x89);
6925 ins_encode( OpcP, RegMem( src, mem ) );
6926 ins_pipe( ialu_mem_reg );
6927 %}
6929 // Store Integer Immediate
6930 instruct storeImmI(memory mem, immI src) %{
6931 match(Set mem (StoreI mem src));
6933 ins_cost(150);
6934 format %{ "MOV $mem,$src" %}
6935 opcode(0xC7); /* C7 /0 */
6936 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con32( src ));
6937 ins_pipe( ialu_mem_imm );
6938 %}
6940 // Store Short/Char Immediate
6941 instruct storeImmI16(memory mem, immI16 src) %{
6942 predicate(UseStoreImmI16);
6943 match(Set mem (StoreC mem src));
6945 ins_cost(150);
6946 format %{ "MOV16 $mem,$src" %}
6947 opcode(0xC7); /* C7 /0 Same as 32 store immediate with prefix */
6948 ins_encode( SizePrefix, OpcP, RMopc_Mem(0x00,mem), Con16( src ));
6949 ins_pipe( ialu_mem_imm );
6950 %}
6952 // Store Pointer Immediate; null pointers or constant oops that do not
6953 // need card-mark barriers.
6954 instruct storeImmP(memory mem, immP src) %{
6955 match(Set mem (StoreP mem src));
6957 ins_cost(150);
6958 format %{ "MOV $mem,$src" %}
6959 opcode(0xC7); /* C7 /0 */
6960 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con32( src ));
6961 ins_pipe( ialu_mem_imm );
6962 %}
6964 // Store Byte Immediate
6965 instruct storeImmB(memory mem, immI8 src) %{
6966 match(Set mem (StoreB mem src));
6968 ins_cost(150);
6969 format %{ "MOV8 $mem,$src" %}
6970 opcode(0xC6); /* C6 /0 */
6971 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con8or32( src ));
6972 ins_pipe( ialu_mem_imm );
6973 %}
6975 // Store Aligned Packed Byte XMM register to memory
6976 instruct storeA8B(memory mem, regXD src) %{
6977 predicate(UseSSE>=1);
6978 match(Set mem (Store8B mem src));
6979 ins_cost(145);
6980 format %{ "MOVQ $mem,$src\t! packed8B" %}
6981 ins_encode( movq_st(mem, src));
6982 ins_pipe( pipe_slow );
6983 %}
6985 // Store Aligned Packed Char/Short XMM register to memory
6986 instruct storeA4C(memory mem, regXD src) %{
6987 predicate(UseSSE>=1);
6988 match(Set mem (Store4C mem src));
6989 ins_cost(145);
6990 format %{ "MOVQ $mem,$src\t! packed4C" %}
6991 ins_encode( movq_st(mem, src));
6992 ins_pipe( pipe_slow );
6993 %}
6995 // Store Aligned Packed Integer XMM register to memory
6996 instruct storeA2I(memory mem, regXD src) %{
6997 predicate(UseSSE>=1);
6998 match(Set mem (Store2I mem src));
6999 ins_cost(145);
7000 format %{ "MOVQ $mem,$src\t! packed2I" %}
7001 ins_encode( movq_st(mem, src));
7002 ins_pipe( pipe_slow );
7003 %}
7005 // Store CMS card-mark Immediate
7006 instruct storeImmCM(memory mem, immI8 src) %{
7007 match(Set mem (StoreCM mem src));
7009 ins_cost(150);
7010 format %{ "MOV8 $mem,$src\t! CMS card-mark imm0" %}
7011 opcode(0xC6); /* C6 /0 */
7012 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con8or32( src ));
7013 ins_pipe( ialu_mem_imm );
7014 %}
7016 // Store Double
7017 instruct storeD( memory mem, regDPR1 src) %{
7018 predicate(UseSSE<=1);
7019 match(Set mem (StoreD mem src));
7021 ins_cost(100);
7022 format %{ "FST_D $mem,$src" %}
7023 opcode(0xDD); /* DD /2 */
7024 ins_encode( enc_FP_store(mem,src) );
7025 ins_pipe( fpu_mem_reg );
7026 %}
7028 // Store double does rounding on x86
7029 instruct storeD_rounded( memory mem, regDPR1 src) %{
7030 predicate(UseSSE<=1);
7031 match(Set mem (StoreD mem (RoundDouble src)));
7033 ins_cost(100);
7034 format %{ "FST_D $mem,$src\t# round" %}
7035 opcode(0xDD); /* DD /2 */
7036 ins_encode( enc_FP_store(mem,src) );
7037 ins_pipe( fpu_mem_reg );
7038 %}
7040 // Store XMM register to memory (double-precision floating points)
7041 // MOVSD instruction
7042 instruct storeXD(memory mem, regXD src) %{
7043 predicate(UseSSE>=2);
7044 match(Set mem (StoreD mem src));
7045 ins_cost(95);
7046 format %{ "MOVSD $mem,$src" %}
7047 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x11), RegMem(src, mem));
7048 ins_pipe( pipe_slow );
7049 %}
7051 // Store XMM register to memory (single-precision floating point)
7052 // MOVSS instruction
7053 instruct storeX(memory mem, regX src) %{
7054 predicate(UseSSE>=1);
7055 match(Set mem (StoreF mem src));
7056 ins_cost(95);
7057 format %{ "MOVSS $mem,$src" %}
7058 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x11), RegMem(src, mem));
7059 ins_pipe( pipe_slow );
7060 %}
7062 // Store Aligned Packed Single Float XMM register to memory
7063 instruct storeA2F(memory mem, regXD src) %{
7064 predicate(UseSSE>=1);
7065 match(Set mem (Store2F mem src));
7066 ins_cost(145);
7067 format %{ "MOVQ $mem,$src\t! packed2F" %}
7068 ins_encode( movq_st(mem, src));
7069 ins_pipe( pipe_slow );
7070 %}
7072 // Store Float
7073 instruct storeF( memory mem, regFPR1 src) %{
7074 predicate(UseSSE==0);
7075 match(Set mem (StoreF mem src));
7077 ins_cost(100);
7078 format %{ "FST_S $mem,$src" %}
7079 opcode(0xD9); /* D9 /2 */
7080 ins_encode( enc_FP_store(mem,src) );
7081 ins_pipe( fpu_mem_reg );
7082 %}
7084 // Store Float does rounding on x86
7085 instruct storeF_rounded( memory mem, regFPR1 src) %{
7086 predicate(UseSSE==0);
7087 match(Set mem (StoreF mem (RoundFloat src)));
7089 ins_cost(100);
7090 format %{ "FST_S $mem,$src\t# round" %}
7091 opcode(0xD9); /* D9 /2 */
7092 ins_encode( enc_FP_store(mem,src) );
7093 ins_pipe( fpu_mem_reg );
7094 %}
7096 // Store Float does rounding on x86
7097 instruct storeF_Drounded( memory mem, regDPR1 src) %{
7098 predicate(UseSSE<=1);
7099 match(Set mem (StoreF mem (ConvD2F src)));
7101 ins_cost(100);
7102 format %{ "FST_S $mem,$src\t# D-round" %}
7103 opcode(0xD9); /* D9 /2 */
7104 ins_encode( enc_FP_store(mem,src) );
7105 ins_pipe( fpu_mem_reg );
7106 %}
7108 // Store immediate Float value (it is faster than store from FPU register)
7109 // The instruction usage is guarded by predicate in operand immF().
7110 instruct storeF_imm( memory mem, immF src) %{
7111 match(Set mem (StoreF mem src));
7113 ins_cost(50);
7114 format %{ "MOV $mem,$src\t# store float" %}
7115 opcode(0xC7); /* C7 /0 */
7116 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con32F_as_bits( src ));
7117 ins_pipe( ialu_mem_imm );
7118 %}
7120 // Store immediate Float value (it is faster than store from XMM register)
7121 // The instruction usage is guarded by predicate in operand immXF().
7122 instruct storeX_imm( memory mem, immXF src) %{
7123 match(Set mem (StoreF mem src));
7125 ins_cost(50);
7126 format %{ "MOV $mem,$src\t# store float" %}
7127 opcode(0xC7); /* C7 /0 */
7128 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con32XF_as_bits( src ));
7129 ins_pipe( ialu_mem_imm );
7130 %}
7132 // Store Integer to stack slot
7133 instruct storeSSI(stackSlotI dst, eRegI src) %{
7134 match(Set dst src);
7136 ins_cost(100);
7137 format %{ "MOV $dst,$src" %}
7138 opcode(0x89);
7139 ins_encode( OpcPRegSS( dst, src ) );
7140 ins_pipe( ialu_mem_reg );
7141 %}
7143 // Store Integer to stack slot
7144 instruct storeSSP(stackSlotP dst, eRegP src) %{
7145 match(Set dst src);
7147 ins_cost(100);
7148 format %{ "MOV $dst,$src" %}
7149 opcode(0x89);
7150 ins_encode( OpcPRegSS( dst, src ) );
7151 ins_pipe( ialu_mem_reg );
7152 %}
7154 // Store Long to stack slot
7155 instruct storeSSL(stackSlotL dst, eRegL src) %{
7156 match(Set dst src);
7158 ins_cost(200);
7159 format %{ "MOV $dst,$src.lo\n\t"
7160 "MOV $dst+4,$src.hi" %}
7161 opcode(0x89, 0x89);
7162 ins_encode( OpcP, RegMem( src, dst ), OpcS, RegMem_Hi( src, dst ) );
7163 ins_pipe( ialu_mem_long_reg );
7164 %}
7166 //----------MemBar Instructions-----------------------------------------------
7167 // Memory barrier flavors
7169 instruct membar_acquire() %{
7170 match(MemBarAcquire);
7171 ins_cost(400);
7173 size(0);
7174 format %{ "MEMBAR-acquire" %}
7175 ins_encode( enc_membar_acquire );
7176 ins_pipe(pipe_slow);
7177 %}
7179 instruct membar_acquire_lock() %{
7180 match(MemBarAcquire);
7181 predicate(Matcher::prior_fast_lock(n));
7182 ins_cost(0);
7184 size(0);
7185 format %{ "MEMBAR-acquire (prior CMPXCHG in FastLock so empty encoding)" %}
7186 ins_encode( );
7187 ins_pipe(empty);
7188 %}
7190 instruct membar_release() %{
7191 match(MemBarRelease);
7192 ins_cost(400);
7194 size(0);
7195 format %{ "MEMBAR-release" %}
7196 ins_encode( enc_membar_release );
7197 ins_pipe(pipe_slow);
7198 %}
7200 instruct membar_release_lock() %{
7201 match(MemBarRelease);
7202 predicate(Matcher::post_fast_unlock(n));
7203 ins_cost(0);
7205 size(0);
7206 format %{ "MEMBAR-release (a FastUnlock follows so empty encoding)" %}
7207 ins_encode( );
7208 ins_pipe(empty);
7209 %}
7211 instruct membar_volatile() %{
7212 match(MemBarVolatile);
7213 ins_cost(400);
7215 format %{ "MEMBAR-volatile" %}
7216 ins_encode( enc_membar_volatile );
7217 ins_pipe(pipe_slow);
7218 %}
7220 instruct unnecessary_membar_volatile() %{
7221 match(MemBarVolatile);
7222 predicate(Matcher::post_store_load_barrier(n));
7223 ins_cost(0);
7225 size(0);
7226 format %{ "MEMBAR-volatile (unnecessary so empty encoding)" %}
7227 ins_encode( );
7228 ins_pipe(empty);
7229 %}
7231 //----------Move Instructions--------------------------------------------------
7232 instruct castX2P(eAXRegP dst, eAXRegI src) %{
7233 match(Set dst (CastX2P src));
7234 format %{ "# X2P $dst, $src" %}
7235 ins_encode( /*empty encoding*/ );
7236 ins_cost(0);
7237 ins_pipe(empty);
7238 %}
7240 instruct castP2X(eRegI dst, eRegP src ) %{
7241 match(Set dst (CastP2X src));
7242 ins_cost(50);
7243 format %{ "MOV $dst, $src\t# CastP2X" %}
7244 ins_encode( enc_Copy( dst, src) );
7245 ins_pipe( ialu_reg_reg );
7246 %}
7248 //----------Conditional Move---------------------------------------------------
7249 // Conditional move
7250 instruct cmovI_reg(eRegI dst, eRegI src, eFlagsReg cr, cmpOp cop ) %{
7251 predicate(VM_Version::supports_cmov() );
7252 match(Set dst (CMoveI (Binary cop cr) (Binary dst src)));
7253 ins_cost(200);
7254 format %{ "CMOV$cop $dst,$src" %}
7255 opcode(0x0F,0x40);
7256 ins_encode( enc_cmov(cop), RegReg( dst, src ) );
7257 ins_pipe( pipe_cmov_reg );
7258 %}
7260 instruct cmovI_regU( eRegI dst, eRegI src, eFlagsRegU cr, cmpOpU cop ) %{
7261 predicate(VM_Version::supports_cmov() );
7262 match(Set dst (CMoveI (Binary cop cr) (Binary dst src)));
7263 ins_cost(200);
7264 format %{ "CMOV$cop $dst,$src" %}
7265 opcode(0x0F,0x40);
7266 ins_encode( enc_cmov(cop), RegReg( dst, src ) );
7267 ins_pipe( pipe_cmov_reg );
7268 %}
7270 // Conditional move
7271 instruct cmovI_mem(cmpOp cop, eFlagsReg cr, eRegI dst, memory src) %{
7272 predicate(VM_Version::supports_cmov() );
7273 match(Set dst (CMoveI (Binary cop cr) (Binary dst (LoadI src))));
7274 ins_cost(250);
7275 format %{ "CMOV$cop $dst,$src" %}
7276 opcode(0x0F,0x40);
7277 ins_encode( enc_cmov(cop), RegMem( dst, src ) );
7278 ins_pipe( pipe_cmov_mem );
7279 %}
7281 // Conditional move
7282 instruct cmovI_memu(cmpOpU cop, eFlagsRegU cr, eRegI dst, memory src) %{
7283 predicate(VM_Version::supports_cmov() );
7284 match(Set dst (CMoveI (Binary cop cr) (Binary dst (LoadI src))));
7285 ins_cost(250);
7286 format %{ "CMOV$cop $dst,$src" %}
7287 opcode(0x0F,0x40);
7288 ins_encode( enc_cmov(cop), RegMem( dst, src ) );
7289 ins_pipe( pipe_cmov_mem );
7290 %}
7292 // Conditional move
7293 instruct cmovP_reg(eRegP dst, eRegP src, eFlagsReg cr, cmpOp cop ) %{
7294 predicate(VM_Version::supports_cmov() );
7295 match(Set dst (CMoveP (Binary cop cr) (Binary dst src)));
7296 ins_cost(200);
7297 format %{ "CMOV$cop $dst,$src\t# ptr" %}
7298 opcode(0x0F,0x40);
7299 ins_encode( enc_cmov(cop), RegReg( dst, src ) );
7300 ins_pipe( pipe_cmov_reg );
7301 %}
7303 // Conditional move (non-P6 version)
7304 // Note: a CMoveP is generated for stubs and native wrappers
7305 // regardless of whether we are on a P6, so we
7306 // emulate a cmov here
7307 instruct cmovP_reg_nonP6(eRegP dst, eRegP src, eFlagsReg cr, cmpOp cop ) %{
7308 match(Set dst (CMoveP (Binary cop cr) (Binary dst src)));
7309 ins_cost(300);
7310 format %{ "Jn$cop skip\n\t"
7311 "MOV $dst,$src\t# pointer\n"
7312 "skip:" %}
7313 opcode(0x8b);
7314 ins_encode( enc_cmov_branch(cop, 0x2), OpcP, RegReg(dst, src));
7315 ins_pipe( pipe_cmov_reg );
7316 %}
7318 // Conditional move
7319 instruct cmovP_regU(eRegP dst, eRegP src, eFlagsRegU cr, cmpOpU cop ) %{
7320 predicate(VM_Version::supports_cmov() );
7321 match(Set dst (CMoveP (Binary cop cr) (Binary dst src)));
7322 ins_cost(200);
7323 format %{ "CMOV$cop $dst,$src\t# ptr" %}
7324 opcode(0x0F,0x40);
7325 ins_encode( enc_cmov(cop), RegReg( dst, src ) );
7326 ins_pipe( pipe_cmov_reg );
7327 %}
7329 // DISABLED: Requires the ADLC to emit a bottom_type call that
7330 // correctly meets the two pointer arguments; one is an incoming
7331 // register but the other is a memory operand. ALSO appears to
7332 // be buggy with implicit null checks.
7333 //
7334 //// Conditional move
7335 //instruct cmovP_mem(cmpOp cop, eFlagsReg cr, eRegP dst, memory src) %{
7336 // predicate(VM_Version::supports_cmov() );
7337 // match(Set dst (CMoveP (Binary cop cr) (Binary dst (LoadP src))));
7338 // ins_cost(250);
7339 // format %{ "CMOV$cop $dst,$src\t# ptr" %}
7340 // opcode(0x0F,0x40);
7341 // ins_encode( enc_cmov(cop), RegMem( dst, src ) );
7342 // ins_pipe( pipe_cmov_mem );
7343 //%}
7344 //
7345 //// Conditional move
7346 //instruct cmovP_memU(cmpOpU cop, eFlagsRegU cr, eRegP dst, memory src) %{
7347 // predicate(VM_Version::supports_cmov() );
7348 // match(Set dst (CMoveP (Binary cop cr) (Binary dst (LoadP src))));
7349 // ins_cost(250);
7350 // format %{ "CMOV$cop $dst,$src\t# ptr" %}
7351 // opcode(0x0F,0x40);
7352 // ins_encode( enc_cmov(cop), RegMem( dst, src ) );
7353 // ins_pipe( pipe_cmov_mem );
7354 //%}
7356 // Conditional move
7357 instruct fcmovD_regU(cmpOp_fcmov cop, eFlagsRegU cr, regDPR1 dst, regD src) %{
7358 predicate(UseSSE<=1);
7359 match(Set dst (CMoveD (Binary cop cr) (Binary dst src)));
7360 ins_cost(200);
7361 format %{ "FCMOV$cop $dst,$src\t# double" %}
7362 opcode(0xDA);
7363 ins_encode( enc_cmov_d(cop,src) );
7364 ins_pipe( pipe_cmovD_reg );
7365 %}
7367 // Conditional move
7368 instruct fcmovF_regU(cmpOp_fcmov cop, eFlagsRegU cr, regFPR1 dst, regF src) %{
7369 predicate(UseSSE==0);
7370 match(Set dst (CMoveF (Binary cop cr) (Binary dst src)));
7371 ins_cost(200);
7372 format %{ "FCMOV$cop $dst,$src\t# float" %}
7373 opcode(0xDA);
7374 ins_encode( enc_cmov_d(cop,src) );
7375 ins_pipe( pipe_cmovD_reg );
7376 %}
7378 // Float CMOV on Intel doesn't handle *signed* compares, only unsigned.
7379 instruct fcmovD_regS(cmpOp cop, eFlagsReg cr, regD dst, regD src) %{
7380 predicate(UseSSE<=1);
7381 match(Set dst (CMoveD (Binary cop cr) (Binary dst src)));
7382 ins_cost(200);
7383 format %{ "Jn$cop skip\n\t"
7384 "MOV $dst,$src\t# double\n"
7385 "skip:" %}
7386 opcode (0xdd, 0x3); /* DD D8+i or DD /3 */
7387 ins_encode( enc_cmov_branch( cop, 0x4 ), Push_Reg_D(src), OpcP, RegOpc(dst) );
7388 ins_pipe( pipe_cmovD_reg );
7389 %}
7391 // Float CMOV on Intel doesn't handle *signed* compares, only unsigned.
7392 instruct fcmovF_regS(cmpOp cop, eFlagsReg cr, regF dst, regF src) %{
7393 predicate(UseSSE==0);
7394 match(Set dst (CMoveF (Binary cop cr) (Binary dst src)));
7395 ins_cost(200);
7396 format %{ "Jn$cop skip\n\t"
7397 "MOV $dst,$src\t# float\n"
7398 "skip:" %}
7399 opcode (0xdd, 0x3); /* DD D8+i or DD /3 */
7400 ins_encode( enc_cmov_branch( cop, 0x4 ), Push_Reg_F(src), OpcP, RegOpc(dst) );
7401 ins_pipe( pipe_cmovD_reg );
7402 %}
7404 // No CMOVE with SSE/SSE2
7405 instruct fcmovX_regS(cmpOp cop, eFlagsReg cr, regX dst, regX src) %{
7406 predicate (UseSSE>=1);
7407 match(Set dst (CMoveF (Binary cop cr) (Binary dst src)));
7408 ins_cost(200);
7409 format %{ "Jn$cop skip\n\t"
7410 "MOVSS $dst,$src\t# float\n"
7411 "skip:" %}
7412 ins_encode %{
7413 Label skip;
7414 // Invert sense of branch from sense of CMOV
7415 __ jccb((Assembler::Condition)($cop$$cmpcode^1), skip);
7416 __ movflt($dst$$XMMRegister, $src$$XMMRegister);
7417 __ bind(skip);
7418 %}
7419 ins_pipe( pipe_slow );
7420 %}
7422 // No CMOVE with SSE/SSE2
7423 instruct fcmovXD_regS(cmpOp cop, eFlagsReg cr, regXD dst, regXD src) %{
7424 predicate (UseSSE>=2);
7425 match(Set dst (CMoveD (Binary cop cr) (Binary dst src)));
7426 ins_cost(200);
7427 format %{ "Jn$cop skip\n\t"
7428 "MOVSD $dst,$src\t# float\n"
7429 "skip:" %}
7430 ins_encode %{
7431 Label skip;
7432 // Invert sense of branch from sense of CMOV
7433 __ jccb((Assembler::Condition)($cop$$cmpcode^1), skip);
7434 __ movdbl($dst$$XMMRegister, $src$$XMMRegister);
7435 __ bind(skip);
7436 %}
7437 ins_pipe( pipe_slow );
7438 %}
7440 // unsigned version
7441 instruct fcmovX_regU(cmpOpU cop, eFlagsRegU cr, regX dst, regX src) %{
7442 predicate (UseSSE>=1);
7443 match(Set dst (CMoveF (Binary cop cr) (Binary dst src)));
7444 ins_cost(200);
7445 format %{ "Jn$cop skip\n\t"
7446 "MOVSS $dst,$src\t# float\n"
7447 "skip:" %}
7448 ins_encode %{
7449 Label skip;
7450 // Invert sense of branch from sense of CMOV
7451 __ jccb((Assembler::Condition)($cop$$cmpcode^1), skip);
7452 __ movflt($dst$$XMMRegister, $src$$XMMRegister);
7453 __ bind(skip);
7454 %}
7455 ins_pipe( pipe_slow );
7456 %}
7458 // unsigned version
7459 instruct fcmovXD_regU(cmpOpU cop, eFlagsRegU cr, regXD dst, regXD src) %{
7460 predicate (UseSSE>=2);
7461 match(Set dst (CMoveD (Binary cop cr) (Binary dst src)));
7462 ins_cost(200);
7463 format %{ "Jn$cop skip\n\t"
7464 "MOVSD $dst,$src\t# float\n"
7465 "skip:" %}
7466 ins_encode %{
7467 Label skip;
7468 // Invert sense of branch from sense of CMOV
7469 __ jccb((Assembler::Condition)($cop$$cmpcode^1), skip);
7470 __ movdbl($dst$$XMMRegister, $src$$XMMRegister);
7471 __ bind(skip);
7472 %}
7473 ins_pipe( pipe_slow );
7474 %}
7476 instruct cmovL_reg(cmpOp cop, eFlagsReg cr, eRegL dst, eRegL src) %{
7477 predicate(VM_Version::supports_cmov() );
7478 match(Set dst (CMoveL (Binary cop cr) (Binary dst src)));
7479 ins_cost(200);
7480 format %{ "CMOV$cop $dst.lo,$src.lo\n\t"
7481 "CMOV$cop $dst.hi,$src.hi" %}
7482 opcode(0x0F,0x40);
7483 ins_encode( enc_cmov(cop), RegReg_Lo2( dst, src ), enc_cmov(cop), RegReg_Hi2( dst, src ) );
7484 ins_pipe( pipe_cmov_reg_long );
7485 %}
7487 instruct cmovL_regU(cmpOpU cop, eFlagsRegU cr, eRegL dst, eRegL src) %{
7488 predicate(VM_Version::supports_cmov() );
7489 match(Set dst (CMoveL (Binary cop cr) (Binary dst src)));
7490 ins_cost(200);
7491 format %{ "CMOV$cop $dst.lo,$src.lo\n\t"
7492 "CMOV$cop $dst.hi,$src.hi" %}
7493 opcode(0x0F,0x40);
7494 ins_encode( enc_cmov(cop), RegReg_Lo2( dst, src ), enc_cmov(cop), RegReg_Hi2( dst, src ) );
7495 ins_pipe( pipe_cmov_reg_long );
7496 %}
7498 //----------Arithmetic Instructions--------------------------------------------
7499 //----------Addition Instructions----------------------------------------------
7500 // Integer Addition Instructions
7501 instruct addI_eReg(eRegI dst, eRegI src, eFlagsReg cr) %{
7502 match(Set dst (AddI dst src));
7503 effect(KILL cr);
7505 size(2);
7506 format %{ "ADD $dst,$src" %}
7507 opcode(0x03);
7508 ins_encode( OpcP, RegReg( dst, src) );
7509 ins_pipe( ialu_reg_reg );
7510 %}
7512 instruct addI_eReg_imm(eRegI dst, immI src, eFlagsReg cr) %{
7513 match(Set dst (AddI dst src));
7514 effect(KILL cr);
7516 format %{ "ADD $dst,$src" %}
7517 opcode(0x81, 0x00); /* /0 id */
7518 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
7519 ins_pipe( ialu_reg );
7520 %}
7522 instruct incI_eReg(eRegI dst, immI1 src, eFlagsReg cr) %{
7523 predicate(UseIncDec);
7524 match(Set dst (AddI dst src));
7525 effect(KILL cr);
7527 size(1);
7528 format %{ "INC $dst" %}
7529 opcode(0x40); /* */
7530 ins_encode( Opc_plus( primary, dst ) );
7531 ins_pipe( ialu_reg );
7532 %}
7534 instruct leaI_eReg_immI(eRegI dst, eRegI src0, immI src1) %{
7535 match(Set dst (AddI src0 src1));
7536 ins_cost(110);
7538 format %{ "LEA $dst,[$src0 + $src1]" %}
7539 opcode(0x8D); /* 0x8D /r */
7540 ins_encode( OpcP, RegLea( dst, src0, src1 ) );
7541 ins_pipe( ialu_reg_reg );
7542 %}
7544 instruct leaP_eReg_immI(eRegP dst, eRegP src0, immI src1) %{
7545 match(Set dst (AddP src0 src1));
7546 ins_cost(110);
7548 format %{ "LEA $dst,[$src0 + $src1]\t# ptr" %}
7549 opcode(0x8D); /* 0x8D /r */
7550 ins_encode( OpcP, RegLea( dst, src0, src1 ) );
7551 ins_pipe( ialu_reg_reg );
7552 %}
7554 instruct decI_eReg(eRegI dst, immI_M1 src, eFlagsReg cr) %{
7555 predicate(UseIncDec);
7556 match(Set dst (AddI dst src));
7557 effect(KILL cr);
7559 size(1);
7560 format %{ "DEC $dst" %}
7561 opcode(0x48); /* */
7562 ins_encode( Opc_plus( primary, dst ) );
7563 ins_pipe( ialu_reg );
7564 %}
7566 instruct addP_eReg(eRegP dst, eRegI src, eFlagsReg cr) %{
7567 match(Set dst (AddP dst src));
7568 effect(KILL cr);
7570 size(2);
7571 format %{ "ADD $dst,$src" %}
7572 opcode(0x03);
7573 ins_encode( OpcP, RegReg( dst, src) );
7574 ins_pipe( ialu_reg_reg );
7575 %}
7577 instruct addP_eReg_imm(eRegP dst, immI src, eFlagsReg cr) %{
7578 match(Set dst (AddP dst src));
7579 effect(KILL cr);
7581 format %{ "ADD $dst,$src" %}
7582 opcode(0x81,0x00); /* Opcode 81 /0 id */
7583 // ins_encode( RegImm( dst, src) );
7584 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
7585 ins_pipe( ialu_reg );
7586 %}
7588 instruct addI_eReg_mem(eRegI dst, memory src, eFlagsReg cr) %{
7589 match(Set dst (AddI dst (LoadI src)));
7590 effect(KILL cr);
7592 ins_cost(125);
7593 format %{ "ADD $dst,$src" %}
7594 opcode(0x03);
7595 ins_encode( OpcP, RegMem( dst, src) );
7596 ins_pipe( ialu_reg_mem );
7597 %}
7599 instruct addI_mem_eReg(memory dst, eRegI src, eFlagsReg cr) %{
7600 match(Set dst (StoreI dst (AddI (LoadI dst) src)));
7601 effect(KILL cr);
7603 ins_cost(150);
7604 format %{ "ADD $dst,$src" %}
7605 opcode(0x01); /* Opcode 01 /r */
7606 ins_encode( OpcP, RegMem( src, dst ) );
7607 ins_pipe( ialu_mem_reg );
7608 %}
7610 // Add Memory with Immediate
7611 instruct addI_mem_imm(memory dst, immI src, eFlagsReg cr) %{
7612 match(Set dst (StoreI dst (AddI (LoadI dst) src)));
7613 effect(KILL cr);
7615 ins_cost(125);
7616 format %{ "ADD $dst,$src" %}
7617 opcode(0x81); /* Opcode 81 /0 id */
7618 ins_encode( OpcSE( src ), RMopc_Mem(0x00,dst), Con8or32( src ) );
7619 ins_pipe( ialu_mem_imm );
7620 %}
7622 instruct incI_mem(memory dst, immI1 src, eFlagsReg cr) %{
7623 match(Set dst (StoreI dst (AddI (LoadI dst) src)));
7624 effect(KILL cr);
7626 ins_cost(125);
7627 format %{ "INC $dst" %}
7628 opcode(0xFF); /* Opcode FF /0 */
7629 ins_encode( OpcP, RMopc_Mem(0x00,dst));
7630 ins_pipe( ialu_mem_imm );
7631 %}
7633 instruct decI_mem(memory dst, immI_M1 src, eFlagsReg cr) %{
7634 match(Set dst (StoreI dst (AddI (LoadI dst) src)));
7635 effect(KILL cr);
7637 ins_cost(125);
7638 format %{ "DEC $dst" %}
7639 opcode(0xFF); /* Opcode FF /1 */
7640 ins_encode( OpcP, RMopc_Mem(0x01,dst));
7641 ins_pipe( ialu_mem_imm );
7642 %}
7645 instruct checkCastPP( eRegP dst ) %{
7646 match(Set dst (CheckCastPP dst));
7648 size(0);
7649 format %{ "#checkcastPP of $dst" %}
7650 ins_encode( /*empty encoding*/ );
7651 ins_pipe( empty );
7652 %}
7654 instruct castPP( eRegP dst ) %{
7655 match(Set dst (CastPP dst));
7656 format %{ "#castPP of $dst" %}
7657 ins_encode( /*empty encoding*/ );
7658 ins_pipe( empty );
7659 %}
7661 instruct castII( eRegI dst ) %{
7662 match(Set dst (CastII dst));
7663 format %{ "#castII of $dst" %}
7664 ins_encode( /*empty encoding*/ );
7665 ins_cost(0);
7666 ins_pipe( empty );
7667 %}
7670 // Load-locked - same as a regular pointer load when used with compare-swap
7671 instruct loadPLocked(eRegP dst, memory mem) %{
7672 match(Set dst (LoadPLocked mem));
7674 ins_cost(125);
7675 format %{ "MOV $dst,$mem\t# Load ptr. locked" %}
7676 opcode(0x8B);
7677 ins_encode( OpcP, RegMem(dst,mem));
7678 ins_pipe( ialu_reg_mem );
7679 %}
7681 // LoadLong-locked - same as a volatile long load when used with compare-swap
7682 instruct loadLLocked(stackSlotL dst, load_long_memory mem) %{
7683 predicate(UseSSE<=1);
7684 match(Set dst (LoadLLocked mem));
7686 ins_cost(200);
7687 format %{ "FILD $mem\t# Atomic volatile long load\n\t"
7688 "FISTp $dst" %}
7689 ins_encode(enc_loadL_volatile(mem,dst));
7690 ins_pipe( fpu_reg_mem );
7691 %}
7693 instruct loadLX_Locked(stackSlotL dst, load_long_memory mem, regXD tmp) %{
7694 predicate(UseSSE>=2);
7695 match(Set dst (LoadLLocked mem));
7696 effect(TEMP tmp);
7697 ins_cost(180);
7698 format %{ "MOVSD $tmp,$mem\t# Atomic volatile long load\n\t"
7699 "MOVSD $dst,$tmp" %}
7700 ins_encode(enc_loadLX_volatile(mem, dst, tmp));
7701 ins_pipe( pipe_slow );
7702 %}
7704 instruct loadLX_reg_Locked(eRegL dst, load_long_memory mem, regXD tmp) %{
7705 predicate(UseSSE>=2);
7706 match(Set dst (LoadLLocked mem));
7707 effect(TEMP tmp);
7708 ins_cost(160);
7709 format %{ "MOVSD $tmp,$mem\t# Atomic volatile long load\n\t"
7710 "MOVD $dst.lo,$tmp\n\t"
7711 "PSRLQ $tmp,32\n\t"
7712 "MOVD $dst.hi,$tmp" %}
7713 ins_encode(enc_loadLX_reg_volatile(mem, dst, tmp));
7714 ins_pipe( pipe_slow );
7715 %}
7717 // Conditional-store of the updated heap-top.
7718 // Used during allocation of the shared heap.
7719 // Sets flags (EQ) on success. Implemented with a CMPXCHG on Intel.
7720 instruct storePConditional( memory heap_top_ptr, eAXRegP oldval, eRegP newval, eFlagsReg cr ) %{
7721 match(Set cr (StorePConditional heap_top_ptr (Binary oldval newval)));
7722 // EAX is killed if there is contention, but then it's also unused.
7723 // In the common case of no contention, EAX holds the new oop address.
7724 format %{ "CMPXCHG $heap_top_ptr,$newval\t# If EAX==$heap_top_ptr Then store $newval into $heap_top_ptr" %}
7725 ins_encode( lock_prefix, Opcode(0x0F), Opcode(0xB1), RegMem(newval,heap_top_ptr) );
7726 ins_pipe( pipe_cmpxchg );
7727 %}
7729 // Conditional-store of a long value
7730 // Returns a boolean value (0/1) on success. Implemented with a CMPXCHG8 on Intel.
7731 // mem_ptr can actually be in either ESI or EDI
7732 instruct storeLConditional( eRegI res, eSIRegP mem_ptr, eADXRegL oldval, eBCXRegL newval, eFlagsReg cr ) %{
7733 match(Set res (StoreLConditional mem_ptr (Binary oldval newval)));
7734 effect(KILL cr);
7735 // EDX:EAX is killed if there is contention, but then it's also unused.
7736 // In the common case of no contention, EDX:EAX holds the new oop address.
7737 format %{ "CMPXCHG8 [$mem_ptr],$newval\t# If EDX:EAX==[$mem_ptr] Then store $newval into [$mem_ptr]\n\t"
7738 "MOV $res,0\n\t"
7739 "JNE,s fail\n\t"
7740 "MOV $res,1\n"
7741 "fail:" %}
7742 ins_encode( enc_cmpxchg8(mem_ptr),
7743 enc_flags_ne_to_boolean(res) );
7744 ins_pipe( pipe_cmpxchg );
7745 %}
7747 // Conditional-store of a long value
7748 // ZF flag is set on success, reset otherwise. Implemented with a CMPXCHG8 on Intel.
7749 // mem_ptr can actually be in either ESI or EDI
7750 instruct storeLConditional_flags( eSIRegP mem_ptr, eADXRegL oldval, eBCXRegL newval, eFlagsReg cr, immI0 zero ) %{
7751 match(Set cr (CmpI (StoreLConditional mem_ptr (Binary oldval newval)) zero));
7752 // EDX:EAX is killed if there is contention, but then it's also unused.
7753 // In the common case of no contention, EDX:EAX holds the new oop address.
7754 format %{ "CMPXCHG8 [$mem_ptr],$newval\t# If EAX==[$mem_ptr] Then store $newval into [$mem_ptr]\n\t" %}
7755 ins_encode( enc_cmpxchg8(mem_ptr) );
7756 ins_pipe( pipe_cmpxchg );
7757 %}
7759 // No flag versions for CompareAndSwap{P,I,L} because matcher can't match them
7761 instruct compareAndSwapL( eRegI res, eSIRegP mem_ptr, eADXRegL oldval, eBCXRegL newval, eFlagsReg cr ) %{
7762 match(Set res (CompareAndSwapL mem_ptr (Binary oldval newval)));
7763 effect(KILL cr, KILL oldval);
7764 format %{ "CMPXCHG8 [$mem_ptr],$newval\t# If EDX:EAX==[$mem_ptr] Then store $newval into [$mem_ptr]\n\t"
7765 "MOV $res,0\n\t"
7766 "JNE,s fail\n\t"
7767 "MOV $res,1\n"
7768 "fail:" %}
7769 ins_encode( enc_cmpxchg8(mem_ptr),
7770 enc_flags_ne_to_boolean(res) );
7771 ins_pipe( pipe_cmpxchg );
7772 %}
7774 instruct compareAndSwapP( eRegI res, pRegP mem_ptr, eAXRegP oldval, eCXRegP newval, eFlagsReg cr) %{
7775 match(Set res (CompareAndSwapP mem_ptr (Binary oldval newval)));
7776 effect(KILL cr, KILL oldval);
7777 format %{ "CMPXCHG [$mem_ptr],$newval\t# If EAX==[$mem_ptr] Then store $newval into [$mem_ptr]\n\t"
7778 "MOV $res,0\n\t"
7779 "JNE,s fail\n\t"
7780 "MOV $res,1\n"
7781 "fail:" %}
7782 ins_encode( enc_cmpxchg(mem_ptr), enc_flags_ne_to_boolean(res) );
7783 ins_pipe( pipe_cmpxchg );
7784 %}
7786 instruct compareAndSwapI( eRegI res, pRegP mem_ptr, eAXRegI oldval, eCXRegI newval, eFlagsReg cr) %{
7787 match(Set res (CompareAndSwapI mem_ptr (Binary oldval newval)));
7788 effect(KILL cr, KILL oldval);
7789 format %{ "CMPXCHG [$mem_ptr],$newval\t# If EAX==[$mem_ptr] Then store $newval into [$mem_ptr]\n\t"
7790 "MOV $res,0\n\t"
7791 "JNE,s fail\n\t"
7792 "MOV $res,1\n"
7793 "fail:" %}
7794 ins_encode( enc_cmpxchg(mem_ptr), enc_flags_ne_to_boolean(res) );
7795 ins_pipe( pipe_cmpxchg );
7796 %}
7798 //----------Subtraction Instructions-------------------------------------------
7799 // Integer Subtraction Instructions
7800 instruct subI_eReg(eRegI dst, eRegI src, eFlagsReg cr) %{
7801 match(Set dst (SubI dst src));
7802 effect(KILL cr);
7804 size(2);
7805 format %{ "SUB $dst,$src" %}
7806 opcode(0x2B);
7807 ins_encode( OpcP, RegReg( dst, src) );
7808 ins_pipe( ialu_reg_reg );
7809 %}
7811 instruct subI_eReg_imm(eRegI dst, immI src, eFlagsReg cr) %{
7812 match(Set dst (SubI dst src));
7813 effect(KILL cr);
7815 format %{ "SUB $dst,$src" %}
7816 opcode(0x81,0x05); /* Opcode 81 /5 */
7817 // ins_encode( RegImm( dst, src) );
7818 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
7819 ins_pipe( ialu_reg );
7820 %}
7822 instruct subI_eReg_mem(eRegI dst, memory src, eFlagsReg cr) %{
7823 match(Set dst (SubI dst (LoadI src)));
7824 effect(KILL cr);
7826 ins_cost(125);
7827 format %{ "SUB $dst,$src" %}
7828 opcode(0x2B);
7829 ins_encode( OpcP, RegMem( dst, src) );
7830 ins_pipe( ialu_reg_mem );
7831 %}
7833 instruct subI_mem_eReg(memory dst, eRegI src, eFlagsReg cr) %{
7834 match(Set dst (StoreI dst (SubI (LoadI dst) src)));
7835 effect(KILL cr);
7837 ins_cost(150);
7838 format %{ "SUB $dst,$src" %}
7839 opcode(0x29); /* Opcode 29 /r */
7840 ins_encode( OpcP, RegMem( src, dst ) );
7841 ins_pipe( ialu_mem_reg );
7842 %}
7844 // Subtract from a pointer
7845 instruct subP_eReg(eRegP dst, eRegI src, immI0 zero, eFlagsReg cr) %{
7846 match(Set dst (AddP dst (SubI zero src)));
7847 effect(KILL cr);
7849 size(2);
7850 format %{ "SUB $dst,$src" %}
7851 opcode(0x2B);
7852 ins_encode( OpcP, RegReg( dst, src) );
7853 ins_pipe( ialu_reg_reg );
7854 %}
7856 instruct negI_eReg(eRegI dst, immI0 zero, eFlagsReg cr) %{
7857 match(Set dst (SubI zero dst));
7858 effect(KILL cr);
7860 size(2);
7861 format %{ "NEG $dst" %}
7862 opcode(0xF7,0x03); // Opcode F7 /3
7863 ins_encode( OpcP, RegOpc( dst ) );
7864 ins_pipe( ialu_reg );
7865 %}
7868 //----------Multiplication/Division Instructions-------------------------------
7869 // Integer Multiplication Instructions
7870 // Multiply Register
7871 instruct mulI_eReg(eRegI dst, eRegI src, eFlagsReg cr) %{
7872 match(Set dst (MulI dst src));
7873 effect(KILL cr);
7875 size(3);
7876 ins_cost(300);
7877 format %{ "IMUL $dst,$src" %}
7878 opcode(0xAF, 0x0F);
7879 ins_encode( OpcS, OpcP, RegReg( dst, src) );
7880 ins_pipe( ialu_reg_reg_alu0 );
7881 %}
7883 // Multiply 32-bit Immediate
7884 instruct mulI_eReg_imm(eRegI dst, eRegI src, immI imm, eFlagsReg cr) %{
7885 match(Set dst (MulI src imm));
7886 effect(KILL cr);
7888 ins_cost(300);
7889 format %{ "IMUL $dst,$src,$imm" %}
7890 opcode(0x69); /* 69 /r id */
7891 ins_encode( OpcSE(imm), RegReg( dst, src ), Con8or32( imm ) );
7892 ins_pipe( ialu_reg_reg_alu0 );
7893 %}
7895 instruct loadConL_low_only(eADXRegL_low_only dst, immL32 src, eFlagsReg cr) %{
7896 match(Set dst src);
7897 effect(KILL cr);
7899 // Note that this is artificially increased to make it more expensive than loadConL
7900 ins_cost(250);
7901 format %{ "MOV EAX,$src\t// low word only" %}
7902 opcode(0xB8);
7903 ins_encode( LdImmL_Lo(dst, src) );
7904 ins_pipe( ialu_reg_fat );
7905 %}
7907 // Multiply by 32-bit Immediate, taking the shifted high order results
7908 // (special case for shift by 32)
7909 instruct mulI_imm_high(eDXRegI dst, nadxRegI src1, eADXRegL_low_only src2, immI_32 cnt, eFlagsReg cr) %{
7910 match(Set dst (ConvL2I (RShiftL (MulL (ConvI2L src1) src2) cnt)));
7911 predicate( _kids[0]->_kids[0]->_kids[1]->_leaf->Opcode() == Op_ConL &&
7912 _kids[0]->_kids[0]->_kids[1]->_leaf->as_Type()->type()->is_long()->get_con() >= min_jint &&
7913 _kids[0]->_kids[0]->_kids[1]->_leaf->as_Type()->type()->is_long()->get_con() <= max_jint );
7914 effect(USE src1, KILL cr);
7916 // Note that this is adjusted by 150 to compensate for the overcosting of loadConL_low_only
7917 ins_cost(0*100 + 1*400 - 150);
7918 format %{ "IMUL EDX:EAX,$src1" %}
7919 ins_encode( multiply_con_and_shift_high( dst, src1, src2, cnt, cr ) );
7920 ins_pipe( pipe_slow );
7921 %}
7923 // Multiply by 32-bit Immediate, taking the shifted high order results
7924 instruct mulI_imm_RShift_high(eDXRegI dst, nadxRegI src1, eADXRegL_low_only src2, immI_32_63 cnt, eFlagsReg cr) %{
7925 match(Set dst (ConvL2I (RShiftL (MulL (ConvI2L src1) src2) cnt)));
7926 predicate( _kids[0]->_kids[0]->_kids[1]->_leaf->Opcode() == Op_ConL &&
7927 _kids[0]->_kids[0]->_kids[1]->_leaf->as_Type()->type()->is_long()->get_con() >= min_jint &&
7928 _kids[0]->_kids[0]->_kids[1]->_leaf->as_Type()->type()->is_long()->get_con() <= max_jint );
7929 effect(USE src1, KILL cr);
7931 // Note that this is adjusted by 150 to compensate for the overcosting of loadConL_low_only
7932 ins_cost(1*100 + 1*400 - 150);
7933 format %{ "IMUL EDX:EAX,$src1\n\t"
7934 "SAR EDX,$cnt-32" %}
7935 ins_encode( multiply_con_and_shift_high( dst, src1, src2, cnt, cr ) );
7936 ins_pipe( pipe_slow );
7937 %}
7939 // Multiply Memory 32-bit Immediate
7940 instruct mulI_mem_imm(eRegI dst, memory src, immI imm, eFlagsReg cr) %{
7941 match(Set dst (MulI (LoadI src) imm));
7942 effect(KILL cr);
7944 ins_cost(300);
7945 format %{ "IMUL $dst,$src,$imm" %}
7946 opcode(0x69); /* 69 /r id */
7947 ins_encode( OpcSE(imm), RegMem( dst, src ), Con8or32( imm ) );
7948 ins_pipe( ialu_reg_mem_alu0 );
7949 %}
7951 // Multiply Memory
7952 instruct mulI(eRegI dst, memory src, eFlagsReg cr) %{
7953 match(Set dst (MulI dst (LoadI src)));
7954 effect(KILL cr);
7956 ins_cost(350);
7957 format %{ "IMUL $dst,$src" %}
7958 opcode(0xAF, 0x0F);
7959 ins_encode( OpcS, OpcP, RegMem( dst, src) );
7960 ins_pipe( ialu_reg_mem_alu0 );
7961 %}
7963 // Multiply Register Int to Long
7964 instruct mulI2L(eADXRegL dst, eAXRegI src, nadxRegI src1, eFlagsReg flags) %{
7965 // Basic Idea: long = (long)int * (long)int
7966 match(Set dst (MulL (ConvI2L src) (ConvI2L src1)));
7967 effect(DEF dst, USE src, USE src1, KILL flags);
7969 ins_cost(300);
7970 format %{ "IMUL $dst,$src1" %}
7972 ins_encode( long_int_multiply( dst, src1 ) );
7973 ins_pipe( ialu_reg_reg_alu0 );
7974 %}
7976 instruct mulIS_eReg(eADXRegL dst, immL_32bits mask, eFlagsReg flags, eAXRegI src, nadxRegI src1) %{
7977 // Basic Idea: long = (int & 0xffffffffL) * (int & 0xffffffffL)
7978 match(Set dst (MulL (AndL (ConvI2L src) mask) (AndL (ConvI2L src1) mask)));
7979 effect(KILL flags);
7981 ins_cost(300);
7982 format %{ "MUL $dst,$src1" %}
7984 ins_encode( long_uint_multiply(dst, src1) );
7985 ins_pipe( ialu_reg_reg_alu0 );
7986 %}
7988 // Multiply Register Long
7989 instruct mulL_eReg(eADXRegL dst, eRegL src, eRegI tmp, eFlagsReg cr) %{
7990 match(Set dst (MulL dst src));
7991 effect(KILL cr, TEMP tmp);
7992 ins_cost(4*100+3*400);
7993 // Basic idea: lo(result) = lo(x_lo * y_lo)
7994 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi)
7995 format %{ "MOV $tmp,$src.lo\n\t"
7996 "IMUL $tmp,EDX\n\t"
7997 "MOV EDX,$src.hi\n\t"
7998 "IMUL EDX,EAX\n\t"
7999 "ADD $tmp,EDX\n\t"
8000 "MUL EDX:EAX,$src.lo\n\t"
8001 "ADD EDX,$tmp" %}
8002 ins_encode( long_multiply( dst, src, tmp ) );
8003 ins_pipe( pipe_slow );
8004 %}
8006 // Multiply Register Long by small constant
8007 instruct mulL_eReg_con(eADXRegL dst, immL_127 src, eRegI tmp, eFlagsReg cr) %{
8008 match(Set dst (MulL dst src));
8009 effect(KILL cr, TEMP tmp);
8010 ins_cost(2*100+2*400);
8011 size(12);
8012 // Basic idea: lo(result) = lo(src * EAX)
8013 // hi(result) = hi(src * EAX) + lo(src * EDX)
8014 format %{ "IMUL $tmp,EDX,$src\n\t"
8015 "MOV EDX,$src\n\t"
8016 "MUL EDX\t# EDX*EAX -> EDX:EAX\n\t"
8017 "ADD EDX,$tmp" %}
8018 ins_encode( long_multiply_con( dst, src, tmp ) );
8019 ins_pipe( pipe_slow );
8020 %}
8022 // Integer DIV with Register
8023 instruct divI_eReg(eAXRegI rax, eDXRegI rdx, eCXRegI div, eFlagsReg cr) %{
8024 match(Set rax (DivI rax div));
8025 effect(KILL rdx, KILL cr);
8026 size(26);
8027 ins_cost(30*100+10*100);
8028 format %{ "CMP EAX,0x80000000\n\t"
8029 "JNE,s normal\n\t"
8030 "XOR EDX,EDX\n\t"
8031 "CMP ECX,-1\n\t"
8032 "JE,s done\n"
8033 "normal: CDQ\n\t"
8034 "IDIV $div\n\t"
8035 "done:" %}
8036 opcode(0xF7, 0x7); /* Opcode F7 /7 */
8037 ins_encode( cdq_enc, OpcP, RegOpc(div) );
8038 ins_pipe( ialu_reg_reg_alu0 );
8039 %}
8041 // Divide Register Long
8042 instruct divL_eReg( eADXRegL dst, eRegL src1, eRegL src2, eFlagsReg cr, eCXRegI cx, eBXRegI bx ) %{
8043 match(Set dst (DivL src1 src2));
8044 effect( KILL cr, KILL cx, KILL bx );
8045 ins_cost(10000);
8046 format %{ "PUSH $src1.hi\n\t"
8047 "PUSH $src1.lo\n\t"
8048 "PUSH $src2.hi\n\t"
8049 "PUSH $src2.lo\n\t"
8050 "CALL SharedRuntime::ldiv\n\t"
8051 "ADD ESP,16" %}
8052 ins_encode( long_div(src1,src2) );
8053 ins_pipe( pipe_slow );
8054 %}
8056 // Integer DIVMOD with Register, both quotient and mod results
8057 instruct divModI_eReg_divmod(eAXRegI rax, eDXRegI rdx, eCXRegI div, eFlagsReg cr) %{
8058 match(DivModI rax div);
8059 effect(KILL cr);
8060 size(26);
8061 ins_cost(30*100+10*100);
8062 format %{ "CMP EAX,0x80000000\n\t"
8063 "JNE,s normal\n\t"
8064 "XOR EDX,EDX\n\t"
8065 "CMP ECX,-1\n\t"
8066 "JE,s done\n"
8067 "normal: CDQ\n\t"
8068 "IDIV $div\n\t"
8069 "done:" %}
8070 opcode(0xF7, 0x7); /* Opcode F7 /7 */
8071 ins_encode( cdq_enc, OpcP, RegOpc(div) );
8072 ins_pipe( pipe_slow );
8073 %}
8075 // Integer MOD with Register
8076 instruct modI_eReg(eDXRegI rdx, eAXRegI rax, eCXRegI div, eFlagsReg cr) %{
8077 match(Set rdx (ModI rax div));
8078 effect(KILL rax, KILL cr);
8080 size(26);
8081 ins_cost(300);
8082 format %{ "CDQ\n\t"
8083 "IDIV $div" %}
8084 opcode(0xF7, 0x7); /* Opcode F7 /7 */
8085 ins_encode( cdq_enc, OpcP, RegOpc(div) );
8086 ins_pipe( ialu_reg_reg_alu0 );
8087 %}
8089 // Remainder Register Long
8090 instruct modL_eReg( eADXRegL dst, eRegL src1, eRegL src2, eFlagsReg cr, eCXRegI cx, eBXRegI bx ) %{
8091 match(Set dst (ModL src1 src2));
8092 effect( KILL cr, KILL cx, KILL bx );
8093 ins_cost(10000);
8094 format %{ "PUSH $src1.hi\n\t"
8095 "PUSH $src1.lo\n\t"
8096 "PUSH $src2.hi\n\t"
8097 "PUSH $src2.lo\n\t"
8098 "CALL SharedRuntime::lrem\n\t"
8099 "ADD ESP,16" %}
8100 ins_encode( long_mod(src1,src2) );
8101 ins_pipe( pipe_slow );
8102 %}
8104 // Integer Shift Instructions
8105 // Shift Left by one
8106 instruct shlI_eReg_1(eRegI dst, immI1 shift, eFlagsReg cr) %{
8107 match(Set dst (LShiftI dst shift));
8108 effect(KILL cr);
8110 size(2);
8111 format %{ "SHL $dst,$shift" %}
8112 opcode(0xD1, 0x4); /* D1 /4 */
8113 ins_encode( OpcP, RegOpc( dst ) );
8114 ins_pipe( ialu_reg );
8115 %}
8117 // Shift Left by 8-bit immediate
8118 instruct salI_eReg_imm(eRegI dst, immI8 shift, eFlagsReg cr) %{
8119 match(Set dst (LShiftI dst shift));
8120 effect(KILL cr);
8122 size(3);
8123 format %{ "SHL $dst,$shift" %}
8124 opcode(0xC1, 0x4); /* C1 /4 ib */
8125 ins_encode( RegOpcImm( dst, shift) );
8126 ins_pipe( ialu_reg );
8127 %}
8129 // Shift Left by variable
8130 instruct salI_eReg_CL(eRegI dst, eCXRegI shift, eFlagsReg cr) %{
8131 match(Set dst (LShiftI dst shift));
8132 effect(KILL cr);
8134 size(2);
8135 format %{ "SHL $dst,$shift" %}
8136 opcode(0xD3, 0x4); /* D3 /4 */
8137 ins_encode( OpcP, RegOpc( dst ) );
8138 ins_pipe( ialu_reg_reg );
8139 %}
8141 // Arithmetic shift right by one
8142 instruct sarI_eReg_1(eRegI dst, immI1 shift, eFlagsReg cr) %{
8143 match(Set dst (RShiftI dst shift));
8144 effect(KILL cr);
8146 size(2);
8147 format %{ "SAR $dst,$shift" %}
8148 opcode(0xD1, 0x7); /* D1 /7 */
8149 ins_encode( OpcP, RegOpc( dst ) );
8150 ins_pipe( ialu_reg );
8151 %}
8153 // Arithmetic shift right by one
8154 instruct sarI_mem_1(memory dst, immI1 shift, eFlagsReg cr) %{
8155 match(Set dst (StoreI dst (RShiftI (LoadI dst) shift)));
8156 effect(KILL cr);
8157 format %{ "SAR $dst,$shift" %}
8158 opcode(0xD1, 0x7); /* D1 /7 */
8159 ins_encode( OpcP, RMopc_Mem(secondary,dst) );
8160 ins_pipe( ialu_mem_imm );
8161 %}
8163 // Arithmetic Shift Right by 8-bit immediate
8164 instruct sarI_eReg_imm(eRegI dst, immI8 shift, eFlagsReg cr) %{
8165 match(Set dst (RShiftI dst shift));
8166 effect(KILL cr);
8168 size(3);
8169 format %{ "SAR $dst,$shift" %}
8170 opcode(0xC1, 0x7); /* C1 /7 ib */
8171 ins_encode( RegOpcImm( dst, shift ) );
8172 ins_pipe( ialu_mem_imm );
8173 %}
8175 // Arithmetic Shift Right by 8-bit immediate
8176 instruct sarI_mem_imm(memory dst, immI8 shift, eFlagsReg cr) %{
8177 match(Set dst (StoreI dst (RShiftI (LoadI dst) shift)));
8178 effect(KILL cr);
8180 format %{ "SAR $dst,$shift" %}
8181 opcode(0xC1, 0x7); /* C1 /7 ib */
8182 ins_encode( OpcP, RMopc_Mem(secondary, dst ), Con8or32( shift ) );
8183 ins_pipe( ialu_mem_imm );
8184 %}
8186 // Arithmetic Shift Right by variable
8187 instruct sarI_eReg_CL(eRegI dst, eCXRegI shift, eFlagsReg cr) %{
8188 match(Set dst (RShiftI dst shift));
8189 effect(KILL cr);
8191 size(2);
8192 format %{ "SAR $dst,$shift" %}
8193 opcode(0xD3, 0x7); /* D3 /7 */
8194 ins_encode( OpcP, RegOpc( dst ) );
8195 ins_pipe( ialu_reg_reg );
8196 %}
8198 // Logical shift right by one
8199 instruct shrI_eReg_1(eRegI dst, immI1 shift, eFlagsReg cr) %{
8200 match(Set dst (URShiftI dst shift));
8201 effect(KILL cr);
8203 size(2);
8204 format %{ "SHR $dst,$shift" %}
8205 opcode(0xD1, 0x5); /* D1 /5 */
8206 ins_encode( OpcP, RegOpc( dst ) );
8207 ins_pipe( ialu_reg );
8208 %}
8210 // Logical Shift Right by 8-bit immediate
8211 instruct shrI_eReg_imm(eRegI dst, immI8 shift, eFlagsReg cr) %{
8212 match(Set dst (URShiftI dst shift));
8213 effect(KILL cr);
8215 size(3);
8216 format %{ "SHR $dst,$shift" %}
8217 opcode(0xC1, 0x5); /* C1 /5 ib */
8218 ins_encode( RegOpcImm( dst, shift) );
8219 ins_pipe( ialu_reg );
8220 %}
8222 // Logical Shift Right by 24, followed by Arithmetic Shift Left by 24.
8223 // This idiom is used by the compiler for the i2b bytecode.
8224 instruct i2b(eRegI dst, xRegI src, immI_24 twentyfour, eFlagsReg cr) %{
8225 match(Set dst (RShiftI (LShiftI src twentyfour) twentyfour));
8226 effect(KILL cr);
8228 size(3);
8229 format %{ "MOVSX $dst,$src :8" %}
8230 opcode(0xBE, 0x0F);
8231 ins_encode( OpcS, OpcP, RegReg( dst, src));
8232 ins_pipe( ialu_reg_reg );
8233 %}
8235 // Logical Shift Right by 16, followed by Arithmetic Shift Left by 16.
8236 // This idiom is used by the compiler the i2s bytecode.
8237 instruct i2s(eRegI dst, xRegI src, immI_16 sixteen, eFlagsReg cr) %{
8238 match(Set dst (RShiftI (LShiftI src sixteen) sixteen));
8239 effect(KILL cr);
8241 size(3);
8242 format %{ "MOVSX $dst,$src :16" %}
8243 opcode(0xBF, 0x0F);
8244 ins_encode( OpcS, OpcP, RegReg( dst, src));
8245 ins_pipe( ialu_reg_reg );
8246 %}
8249 // Logical Shift Right by variable
8250 instruct shrI_eReg_CL(eRegI dst, eCXRegI shift, eFlagsReg cr) %{
8251 match(Set dst (URShiftI dst shift));
8252 effect(KILL cr);
8254 size(2);
8255 format %{ "SHR $dst,$shift" %}
8256 opcode(0xD3, 0x5); /* D3 /5 */
8257 ins_encode( OpcP, RegOpc( dst ) );
8258 ins_pipe( ialu_reg_reg );
8259 %}
8262 //----------Logical Instructions-----------------------------------------------
8263 //----------Integer Logical Instructions---------------------------------------
8264 // And Instructions
8265 // And Register with Register
8266 instruct andI_eReg(eRegI dst, eRegI src, eFlagsReg cr) %{
8267 match(Set dst (AndI dst src));
8268 effect(KILL cr);
8270 size(2);
8271 format %{ "AND $dst,$src" %}
8272 opcode(0x23);
8273 ins_encode( OpcP, RegReg( dst, src) );
8274 ins_pipe( ialu_reg_reg );
8275 %}
8277 // And Register with Immediate
8278 instruct andI_eReg_imm(eRegI dst, immI src, eFlagsReg cr) %{
8279 match(Set dst (AndI dst src));
8280 effect(KILL cr);
8282 format %{ "AND $dst,$src" %}
8283 opcode(0x81,0x04); /* Opcode 81 /4 */
8284 // ins_encode( RegImm( dst, src) );
8285 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
8286 ins_pipe( ialu_reg );
8287 %}
8289 // And Register with Memory
8290 instruct andI_eReg_mem(eRegI dst, memory src, eFlagsReg cr) %{
8291 match(Set dst (AndI dst (LoadI src)));
8292 effect(KILL cr);
8294 ins_cost(125);
8295 format %{ "AND $dst,$src" %}
8296 opcode(0x23);
8297 ins_encode( OpcP, RegMem( dst, src) );
8298 ins_pipe( ialu_reg_mem );
8299 %}
8301 // And Memory with Register
8302 instruct andI_mem_eReg(memory dst, eRegI src, eFlagsReg cr) %{
8303 match(Set dst (StoreI dst (AndI (LoadI dst) src)));
8304 effect(KILL cr);
8306 ins_cost(150);
8307 format %{ "AND $dst,$src" %}
8308 opcode(0x21); /* Opcode 21 /r */
8309 ins_encode( OpcP, RegMem( src, dst ) );
8310 ins_pipe( ialu_mem_reg );
8311 %}
8313 // And Memory with Immediate
8314 instruct andI_mem_imm(memory dst, immI src, eFlagsReg cr) %{
8315 match(Set dst (StoreI dst (AndI (LoadI dst) src)));
8316 effect(KILL cr);
8318 ins_cost(125);
8319 format %{ "AND $dst,$src" %}
8320 opcode(0x81, 0x4); /* Opcode 81 /4 id */
8321 // ins_encode( MemImm( dst, src) );
8322 ins_encode( OpcSE( src ), RMopc_Mem(secondary, dst ), Con8or32( src ) );
8323 ins_pipe( ialu_mem_imm );
8324 %}
8326 // Or Instructions
8327 // Or Register with Register
8328 instruct orI_eReg(eRegI dst, eRegI src, eFlagsReg cr) %{
8329 match(Set dst (OrI dst src));
8330 effect(KILL cr);
8332 size(2);
8333 format %{ "OR $dst,$src" %}
8334 opcode(0x0B);
8335 ins_encode( OpcP, RegReg( dst, src) );
8336 ins_pipe( ialu_reg_reg );
8337 %}
8339 // Or Register with Immediate
8340 instruct orI_eReg_imm(eRegI dst, immI src, eFlagsReg cr) %{
8341 match(Set dst (OrI dst src));
8342 effect(KILL cr);
8344 format %{ "OR $dst,$src" %}
8345 opcode(0x81,0x01); /* Opcode 81 /1 id */
8346 // ins_encode( RegImm( dst, src) );
8347 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
8348 ins_pipe( ialu_reg );
8349 %}
8351 // Or Register with Memory
8352 instruct orI_eReg_mem(eRegI dst, memory src, eFlagsReg cr) %{
8353 match(Set dst (OrI dst (LoadI src)));
8354 effect(KILL cr);
8356 ins_cost(125);
8357 format %{ "OR $dst,$src" %}
8358 opcode(0x0B);
8359 ins_encode( OpcP, RegMem( dst, src) );
8360 ins_pipe( ialu_reg_mem );
8361 %}
8363 // Or Memory with Register
8364 instruct orI_mem_eReg(memory dst, eRegI src, eFlagsReg cr) %{
8365 match(Set dst (StoreI dst (OrI (LoadI dst) src)));
8366 effect(KILL cr);
8368 ins_cost(150);
8369 format %{ "OR $dst,$src" %}
8370 opcode(0x09); /* Opcode 09 /r */
8371 ins_encode( OpcP, RegMem( src, dst ) );
8372 ins_pipe( ialu_mem_reg );
8373 %}
8375 // Or Memory with Immediate
8376 instruct orI_mem_imm(memory dst, immI src, eFlagsReg cr) %{
8377 match(Set dst (StoreI dst (OrI (LoadI dst) src)));
8378 effect(KILL cr);
8380 ins_cost(125);
8381 format %{ "OR $dst,$src" %}
8382 opcode(0x81,0x1); /* Opcode 81 /1 id */
8383 // ins_encode( MemImm( dst, src) );
8384 ins_encode( OpcSE( src ), RMopc_Mem(secondary, dst ), Con8or32( src ) );
8385 ins_pipe( ialu_mem_imm );
8386 %}
8388 // ROL/ROR
8389 // ROL expand
8390 instruct rolI_eReg_imm1(eRegI dst, immI1 shift, eFlagsReg cr) %{
8391 effect(USE_DEF dst, USE shift, KILL cr);
8393 format %{ "ROL $dst, $shift" %}
8394 opcode(0xD1, 0x0); /* Opcode D1 /0 */
8395 ins_encode( OpcP, RegOpc( dst ));
8396 ins_pipe( ialu_reg );
8397 %}
8399 instruct rolI_eReg_imm8(eRegI dst, immI8 shift, eFlagsReg cr) %{
8400 effect(USE_DEF dst, USE shift, KILL cr);
8402 format %{ "ROL $dst, $shift" %}
8403 opcode(0xC1, 0x0); /*Opcode /C1 /0 */
8404 ins_encode( RegOpcImm(dst, shift) );
8405 ins_pipe(ialu_reg);
8406 %}
8408 instruct rolI_eReg_CL(ncxRegI dst, eCXRegI shift, eFlagsReg cr) %{
8409 effect(USE_DEF dst, USE shift, KILL cr);
8411 format %{ "ROL $dst, $shift" %}
8412 opcode(0xD3, 0x0); /* Opcode D3 /0 */
8413 ins_encode(OpcP, RegOpc(dst));
8414 ins_pipe( ialu_reg_reg );
8415 %}
8416 // end of ROL expand
8418 // ROL 32bit by one once
8419 instruct rolI_eReg_i1(eRegI dst, immI1 lshift, immI_M1 rshift, eFlagsReg cr) %{
8420 match(Set dst ( OrI (LShiftI dst lshift) (URShiftI dst rshift)));
8422 expand %{
8423 rolI_eReg_imm1(dst, lshift, cr);
8424 %}
8425 %}
8427 // ROL 32bit var by imm8 once
8428 instruct rolI_eReg_i8(eRegI dst, immI8 lshift, immI8 rshift, eFlagsReg cr) %{
8429 predicate( 0 == ((n->in(1)->in(2)->get_int() + n->in(2)->in(2)->get_int()) & 0x1f));
8430 match(Set dst ( OrI (LShiftI dst lshift) (URShiftI dst rshift)));
8432 expand %{
8433 rolI_eReg_imm8(dst, lshift, cr);
8434 %}
8435 %}
8437 // ROL 32bit var by var once
8438 instruct rolI_eReg_Var_C0(ncxRegI dst, eCXRegI shift, immI0 zero, eFlagsReg cr) %{
8439 match(Set dst ( OrI (LShiftI dst shift) (URShiftI dst (SubI zero shift))));
8441 expand %{
8442 rolI_eReg_CL(dst, shift, cr);
8443 %}
8444 %}
8446 // ROL 32bit var by var once
8447 instruct rolI_eReg_Var_C32(ncxRegI dst, eCXRegI shift, immI_32 c32, eFlagsReg cr) %{
8448 match(Set dst ( OrI (LShiftI dst shift) (URShiftI dst (SubI c32 shift))));
8450 expand %{
8451 rolI_eReg_CL(dst, shift, cr);
8452 %}
8453 %}
8455 // ROR expand
8456 instruct rorI_eReg_imm1(eRegI dst, immI1 shift, eFlagsReg cr) %{
8457 effect(USE_DEF dst, USE shift, KILL cr);
8459 format %{ "ROR $dst, $shift" %}
8460 opcode(0xD1,0x1); /* Opcode D1 /1 */
8461 ins_encode( OpcP, RegOpc( dst ) );
8462 ins_pipe( ialu_reg );
8463 %}
8465 instruct rorI_eReg_imm8(eRegI dst, immI8 shift, eFlagsReg cr) %{
8466 effect (USE_DEF dst, USE shift, KILL cr);
8468 format %{ "ROR $dst, $shift" %}
8469 opcode(0xC1, 0x1); /* Opcode /C1 /1 ib */
8470 ins_encode( RegOpcImm(dst, shift) );
8471 ins_pipe( ialu_reg );
8472 %}
8474 instruct rorI_eReg_CL(ncxRegI dst, eCXRegI shift, eFlagsReg cr)%{
8475 effect(USE_DEF dst, USE shift, KILL cr);
8477 format %{ "ROR $dst, $shift" %}
8478 opcode(0xD3, 0x1); /* Opcode D3 /1 */
8479 ins_encode(OpcP, RegOpc(dst));
8480 ins_pipe( ialu_reg_reg );
8481 %}
8482 // end of ROR expand
8484 // ROR right once
8485 instruct rorI_eReg_i1(eRegI dst, immI1 rshift, immI_M1 lshift, eFlagsReg cr) %{
8486 match(Set dst ( OrI (URShiftI dst rshift) (LShiftI dst lshift)));
8488 expand %{
8489 rorI_eReg_imm1(dst, rshift, cr);
8490 %}
8491 %}
8493 // ROR 32bit by immI8 once
8494 instruct rorI_eReg_i8(eRegI dst, immI8 rshift, immI8 lshift, eFlagsReg cr) %{
8495 predicate( 0 == ((n->in(1)->in(2)->get_int() + n->in(2)->in(2)->get_int()) & 0x1f));
8496 match(Set dst ( OrI (URShiftI dst rshift) (LShiftI dst lshift)));
8498 expand %{
8499 rorI_eReg_imm8(dst, rshift, cr);
8500 %}
8501 %}
8503 // ROR 32bit var by var once
8504 instruct rorI_eReg_Var_C0(ncxRegI dst, eCXRegI shift, immI0 zero, eFlagsReg cr) %{
8505 match(Set dst ( OrI (URShiftI dst shift) (LShiftI dst (SubI zero shift))));
8507 expand %{
8508 rorI_eReg_CL(dst, shift, cr);
8509 %}
8510 %}
8512 // ROR 32bit var by var once
8513 instruct rorI_eReg_Var_C32(ncxRegI dst, eCXRegI shift, immI_32 c32, eFlagsReg cr) %{
8514 match(Set dst ( OrI (URShiftI dst shift) (LShiftI dst (SubI c32 shift))));
8516 expand %{
8517 rorI_eReg_CL(dst, shift, cr);
8518 %}
8519 %}
8521 // Xor Instructions
8522 // Xor Register with Register
8523 instruct xorI_eReg(eRegI dst, eRegI src, eFlagsReg cr) %{
8524 match(Set dst (XorI dst src));
8525 effect(KILL cr);
8527 size(2);
8528 format %{ "XOR $dst,$src" %}
8529 opcode(0x33);
8530 ins_encode( OpcP, RegReg( dst, src) );
8531 ins_pipe( ialu_reg_reg );
8532 %}
8534 // Xor Register with Immediate
8535 instruct xorI_eReg_imm(eRegI dst, immI src, eFlagsReg cr) %{
8536 match(Set dst (XorI dst src));
8537 effect(KILL cr);
8539 format %{ "XOR $dst,$src" %}
8540 opcode(0x81,0x06); /* Opcode 81 /6 id */
8541 // ins_encode( RegImm( dst, src) );
8542 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
8543 ins_pipe( ialu_reg );
8544 %}
8546 // Xor Register with Memory
8547 instruct xorI_eReg_mem(eRegI dst, memory src, eFlagsReg cr) %{
8548 match(Set dst (XorI dst (LoadI src)));
8549 effect(KILL cr);
8551 ins_cost(125);
8552 format %{ "XOR $dst,$src" %}
8553 opcode(0x33);
8554 ins_encode( OpcP, RegMem(dst, src) );
8555 ins_pipe( ialu_reg_mem );
8556 %}
8558 // Xor Memory with Register
8559 instruct xorI_mem_eReg(memory dst, eRegI src, eFlagsReg cr) %{
8560 match(Set dst (StoreI dst (XorI (LoadI dst) src)));
8561 effect(KILL cr);
8563 ins_cost(150);
8564 format %{ "XOR $dst,$src" %}
8565 opcode(0x31); /* Opcode 31 /r */
8566 ins_encode( OpcP, RegMem( src, dst ) );
8567 ins_pipe( ialu_mem_reg );
8568 %}
8570 // Xor Memory with Immediate
8571 instruct xorI_mem_imm(memory dst, immI src, eFlagsReg cr) %{
8572 match(Set dst (StoreI dst (XorI (LoadI dst) src)));
8573 effect(KILL cr);
8575 ins_cost(125);
8576 format %{ "XOR $dst,$src" %}
8577 opcode(0x81,0x6); /* Opcode 81 /6 id */
8578 ins_encode( OpcSE( src ), RMopc_Mem(secondary, dst ), Con8or32( src ) );
8579 ins_pipe( ialu_mem_imm );
8580 %}
8582 //----------Convert Int to Boolean---------------------------------------------
8584 instruct movI_nocopy(eRegI dst, eRegI src) %{
8585 effect( DEF dst, USE src );
8586 format %{ "MOV $dst,$src" %}
8587 ins_encode( enc_Copy( dst, src) );
8588 ins_pipe( ialu_reg_reg );
8589 %}
8591 instruct ci2b( eRegI dst, eRegI src, eFlagsReg cr ) %{
8592 effect( USE_DEF dst, USE src, KILL cr );
8594 size(4);
8595 format %{ "NEG $dst\n\t"
8596 "ADC $dst,$src" %}
8597 ins_encode( neg_reg(dst),
8598 OpcRegReg(0x13,dst,src) );
8599 ins_pipe( ialu_reg_reg_long );
8600 %}
8602 instruct convI2B( eRegI dst, eRegI src, eFlagsReg cr ) %{
8603 match(Set dst (Conv2B src));
8605 expand %{
8606 movI_nocopy(dst,src);
8607 ci2b(dst,src,cr);
8608 %}
8609 %}
8611 instruct movP_nocopy(eRegI dst, eRegP src) %{
8612 effect( DEF dst, USE src );
8613 format %{ "MOV $dst,$src" %}
8614 ins_encode( enc_Copy( dst, src) );
8615 ins_pipe( ialu_reg_reg );
8616 %}
8618 instruct cp2b( eRegI dst, eRegP src, eFlagsReg cr ) %{
8619 effect( USE_DEF dst, USE src, KILL cr );
8620 format %{ "NEG $dst\n\t"
8621 "ADC $dst,$src" %}
8622 ins_encode( neg_reg(dst),
8623 OpcRegReg(0x13,dst,src) );
8624 ins_pipe( ialu_reg_reg_long );
8625 %}
8627 instruct convP2B( eRegI dst, eRegP src, eFlagsReg cr ) %{
8628 match(Set dst (Conv2B src));
8630 expand %{
8631 movP_nocopy(dst,src);
8632 cp2b(dst,src,cr);
8633 %}
8634 %}
8636 instruct cmpLTMask( eCXRegI dst, ncxRegI p, ncxRegI q, eFlagsReg cr ) %{
8637 match(Set dst (CmpLTMask p q));
8638 effect( KILL cr );
8639 ins_cost(400);
8641 // SETlt can only use low byte of EAX,EBX, ECX, or EDX as destination
8642 format %{ "XOR $dst,$dst\n\t"
8643 "CMP $p,$q\n\t"
8644 "SETlt $dst\n\t"
8645 "NEG $dst" %}
8646 ins_encode( OpcRegReg(0x33,dst,dst),
8647 OpcRegReg(0x3B,p,q),
8648 setLT_reg(dst), neg_reg(dst) );
8649 ins_pipe( pipe_slow );
8650 %}
8652 instruct cmpLTMask0( eRegI dst, immI0 zero, eFlagsReg cr ) %{
8653 match(Set dst (CmpLTMask dst zero));
8654 effect( DEF dst, KILL cr );
8655 ins_cost(100);
8657 format %{ "SAR $dst,31" %}
8658 opcode(0xC1, 0x7); /* C1 /7 ib */
8659 ins_encode( RegOpcImm( dst, 0x1F ) );
8660 ins_pipe( ialu_reg );
8661 %}
8664 instruct cadd_cmpLTMask( ncxRegI p, ncxRegI q, ncxRegI y, eCXRegI tmp, eFlagsReg cr ) %{
8665 match(Set p (AddI (AndI (CmpLTMask p q) y) (SubI p q)));
8666 effect( KILL tmp, KILL cr );
8667 ins_cost(400);
8668 // annoyingly, $tmp has no edges so you cant ask for it in
8669 // any format or encoding
8670 format %{ "SUB $p,$q\n\t"
8671 "SBB ECX,ECX\n\t"
8672 "AND ECX,$y\n\t"
8673 "ADD $p,ECX" %}
8674 ins_encode( enc_cmpLTP(p,q,y,tmp) );
8675 ins_pipe( pipe_cmplt );
8676 %}
8678 /* If I enable this, I encourage spilling in the inner loop of compress.
8679 instruct cadd_cmpLTMask_mem( ncxRegI p, ncxRegI q, memory y, eCXRegI tmp, eFlagsReg cr ) %{
8680 match(Set p (AddI (AndI (CmpLTMask p q) (LoadI y)) (SubI p q)));
8681 effect( USE_KILL tmp, KILL cr );
8682 ins_cost(400);
8684 format %{ "SUB $p,$q\n\t"
8685 "SBB ECX,ECX\n\t"
8686 "AND ECX,$y\n\t"
8687 "ADD $p,ECX" %}
8688 ins_encode( enc_cmpLTP_mem(p,q,y,tmp) );
8689 %}
8690 */
8692 //----------Long Instructions------------------------------------------------
8693 // Add Long Register with Register
8694 instruct addL_eReg(eRegL dst, eRegL src, eFlagsReg cr) %{
8695 match(Set dst (AddL dst src));
8696 effect(KILL cr);
8697 ins_cost(200);
8698 format %{ "ADD $dst.lo,$src.lo\n\t"
8699 "ADC $dst.hi,$src.hi" %}
8700 opcode(0x03, 0x13);
8701 ins_encode( RegReg_Lo(dst, src), RegReg_Hi(dst,src) );
8702 ins_pipe( ialu_reg_reg_long );
8703 %}
8705 // Add Long Register with Immediate
8706 instruct addL_eReg_imm(eRegL dst, immL src, eFlagsReg cr) %{
8707 match(Set dst (AddL dst src));
8708 effect(KILL cr);
8709 format %{ "ADD $dst.lo,$src.lo\n\t"
8710 "ADC $dst.hi,$src.hi" %}
8711 opcode(0x81,0x00,0x02); /* Opcode 81 /0, 81 /2 */
8712 ins_encode( Long_OpcSErm_Lo( dst, src ), Long_OpcSErm_Hi( dst, src ) );
8713 ins_pipe( ialu_reg_long );
8714 %}
8716 // Add Long Register with Memory
8717 instruct addL_eReg_mem(eRegL dst, load_long_memory mem, eFlagsReg cr) %{
8718 match(Set dst (AddL dst (LoadL mem)));
8719 effect(KILL cr);
8720 ins_cost(125);
8721 format %{ "ADD $dst.lo,$mem\n\t"
8722 "ADC $dst.hi,$mem+4" %}
8723 opcode(0x03, 0x13);
8724 ins_encode( OpcP, RegMem( dst, mem), OpcS, RegMem_Hi(dst,mem) );
8725 ins_pipe( ialu_reg_long_mem );
8726 %}
8728 // Subtract Long Register with Register.
8729 instruct subL_eReg(eRegL dst, eRegL src, eFlagsReg cr) %{
8730 match(Set dst (SubL dst src));
8731 effect(KILL cr);
8732 ins_cost(200);
8733 format %{ "SUB $dst.lo,$src.lo\n\t"
8734 "SBB $dst.hi,$src.hi" %}
8735 opcode(0x2B, 0x1B);
8736 ins_encode( RegReg_Lo(dst, src), RegReg_Hi(dst,src) );
8737 ins_pipe( ialu_reg_reg_long );
8738 %}
8740 // Subtract Long Register with Immediate
8741 instruct subL_eReg_imm(eRegL dst, immL src, eFlagsReg cr) %{
8742 match(Set dst (SubL dst src));
8743 effect(KILL cr);
8744 format %{ "SUB $dst.lo,$src.lo\n\t"
8745 "SBB $dst.hi,$src.hi" %}
8746 opcode(0x81,0x05,0x03); /* Opcode 81 /5, 81 /3 */
8747 ins_encode( Long_OpcSErm_Lo( dst, src ), Long_OpcSErm_Hi( dst, src ) );
8748 ins_pipe( ialu_reg_long );
8749 %}
8751 // Subtract Long Register with Memory
8752 instruct subL_eReg_mem(eRegL dst, load_long_memory mem, eFlagsReg cr) %{
8753 match(Set dst (SubL dst (LoadL mem)));
8754 effect(KILL cr);
8755 ins_cost(125);
8756 format %{ "SUB $dst.lo,$mem\n\t"
8757 "SBB $dst.hi,$mem+4" %}
8758 opcode(0x2B, 0x1B);
8759 ins_encode( OpcP, RegMem( dst, mem), OpcS, RegMem_Hi(dst,mem) );
8760 ins_pipe( ialu_reg_long_mem );
8761 %}
8763 instruct negL_eReg(eRegL dst, immL0 zero, eFlagsReg cr) %{
8764 match(Set dst (SubL zero dst));
8765 effect(KILL cr);
8766 ins_cost(300);
8767 format %{ "NEG $dst.hi\n\tNEG $dst.lo\n\tSBB $dst.hi,0" %}
8768 ins_encode( neg_long(dst) );
8769 ins_pipe( ialu_reg_reg_long );
8770 %}
8772 // And Long Register with Register
8773 instruct andL_eReg(eRegL dst, eRegL src, eFlagsReg cr) %{
8774 match(Set dst (AndL dst src));
8775 effect(KILL cr);
8776 format %{ "AND $dst.lo,$src.lo\n\t"
8777 "AND $dst.hi,$src.hi" %}
8778 opcode(0x23,0x23);
8779 ins_encode( RegReg_Lo( dst, src), RegReg_Hi( dst, src) );
8780 ins_pipe( ialu_reg_reg_long );
8781 %}
8783 // And Long Register with Immediate
8784 instruct andL_eReg_imm(eRegL dst, immL src, eFlagsReg cr) %{
8785 match(Set dst (AndL dst src));
8786 effect(KILL cr);
8787 format %{ "AND $dst.lo,$src.lo\n\t"
8788 "AND $dst.hi,$src.hi" %}
8789 opcode(0x81,0x04,0x04); /* Opcode 81 /4, 81 /4 */
8790 ins_encode( Long_OpcSErm_Lo( dst, src ), Long_OpcSErm_Hi( dst, src ) );
8791 ins_pipe( ialu_reg_long );
8792 %}
8794 // And Long Register with Memory
8795 instruct andL_eReg_mem(eRegL dst, load_long_memory mem, eFlagsReg cr) %{
8796 match(Set dst (AndL dst (LoadL mem)));
8797 effect(KILL cr);
8798 ins_cost(125);
8799 format %{ "AND $dst.lo,$mem\n\t"
8800 "AND $dst.hi,$mem+4" %}
8801 opcode(0x23, 0x23);
8802 ins_encode( OpcP, RegMem( dst, mem), OpcS, RegMem_Hi(dst,mem) );
8803 ins_pipe( ialu_reg_long_mem );
8804 %}
8806 // Or Long Register with Register
8807 instruct orl_eReg(eRegL dst, eRegL src, eFlagsReg cr) %{
8808 match(Set dst (OrL dst src));
8809 effect(KILL cr);
8810 format %{ "OR $dst.lo,$src.lo\n\t"
8811 "OR $dst.hi,$src.hi" %}
8812 opcode(0x0B,0x0B);
8813 ins_encode( RegReg_Lo( dst, src), RegReg_Hi( dst, src) );
8814 ins_pipe( ialu_reg_reg_long );
8815 %}
8817 // Or Long Register with Immediate
8818 instruct orl_eReg_imm(eRegL dst, immL src, eFlagsReg cr) %{
8819 match(Set dst (OrL dst src));
8820 effect(KILL cr);
8821 format %{ "OR $dst.lo,$src.lo\n\t"
8822 "OR $dst.hi,$src.hi" %}
8823 opcode(0x81,0x01,0x01); /* Opcode 81 /1, 81 /1 */
8824 ins_encode( Long_OpcSErm_Lo( dst, src ), Long_OpcSErm_Hi( dst, src ) );
8825 ins_pipe( ialu_reg_long );
8826 %}
8828 // Or Long Register with Memory
8829 instruct orl_eReg_mem(eRegL dst, load_long_memory mem, eFlagsReg cr) %{
8830 match(Set dst (OrL dst (LoadL mem)));
8831 effect(KILL cr);
8832 ins_cost(125);
8833 format %{ "OR $dst.lo,$mem\n\t"
8834 "OR $dst.hi,$mem+4" %}
8835 opcode(0x0B,0x0B);
8836 ins_encode( OpcP, RegMem( dst, mem), OpcS, RegMem_Hi(dst,mem) );
8837 ins_pipe( ialu_reg_long_mem );
8838 %}
8840 // Xor Long Register with Register
8841 instruct xorl_eReg(eRegL dst, eRegL src, eFlagsReg cr) %{
8842 match(Set dst (XorL dst src));
8843 effect(KILL cr);
8844 format %{ "XOR $dst.lo,$src.lo\n\t"
8845 "XOR $dst.hi,$src.hi" %}
8846 opcode(0x33,0x33);
8847 ins_encode( RegReg_Lo( dst, src), RegReg_Hi( dst, src) );
8848 ins_pipe( ialu_reg_reg_long );
8849 %}
8851 // Xor Long Register with Immediate
8852 instruct xorl_eReg_imm(eRegL dst, immL src, eFlagsReg cr) %{
8853 match(Set dst (XorL dst src));
8854 effect(KILL cr);
8855 format %{ "XOR $dst.lo,$src.lo\n\t"
8856 "XOR $dst.hi,$src.hi" %}
8857 opcode(0x81,0x06,0x06); /* Opcode 81 /6, 81 /6 */
8858 ins_encode( Long_OpcSErm_Lo( dst, src ), Long_OpcSErm_Hi( dst, src ) );
8859 ins_pipe( ialu_reg_long );
8860 %}
8862 // Xor Long Register with Memory
8863 instruct xorl_eReg_mem(eRegL dst, load_long_memory mem, eFlagsReg cr) %{
8864 match(Set dst (XorL dst (LoadL mem)));
8865 effect(KILL cr);
8866 ins_cost(125);
8867 format %{ "XOR $dst.lo,$mem\n\t"
8868 "XOR $dst.hi,$mem+4" %}
8869 opcode(0x33,0x33);
8870 ins_encode( OpcP, RegMem( dst, mem), OpcS, RegMem_Hi(dst,mem) );
8871 ins_pipe( ialu_reg_long_mem );
8872 %}
8874 // Shift Left Long by 1-31
8875 instruct shlL_eReg_1_31(eRegL dst, immI_1_31 cnt, eFlagsReg cr) %{
8876 match(Set dst (LShiftL dst cnt));
8877 effect(KILL cr);
8878 ins_cost(200);
8879 format %{ "SHLD $dst.hi,$dst.lo,$cnt\n\t"
8880 "SHL $dst.lo,$cnt" %}
8881 opcode(0xC1, 0x4, 0xA4); /* 0F/A4, then C1 /4 ib */
8882 ins_encode( move_long_small_shift(dst,cnt) );
8883 ins_pipe( ialu_reg_long );
8884 %}
8886 // Shift Left Long by 32-63
8887 instruct shlL_eReg_32_63(eRegL dst, immI_32_63 cnt, eFlagsReg cr) %{
8888 match(Set dst (LShiftL dst cnt));
8889 effect(KILL cr);
8890 ins_cost(300);
8891 format %{ "MOV $dst.hi,$dst.lo\n"
8892 "\tSHL $dst.hi,$cnt-32\n"
8893 "\tXOR $dst.lo,$dst.lo" %}
8894 opcode(0xC1, 0x4); /* C1 /4 ib */
8895 ins_encode( move_long_big_shift_clr(dst,cnt) );
8896 ins_pipe( ialu_reg_long );
8897 %}
8899 // Shift Left Long by variable
8900 instruct salL_eReg_CL(eRegL dst, eCXRegI shift, eFlagsReg cr) %{
8901 match(Set dst (LShiftL dst shift));
8902 effect(KILL cr);
8903 ins_cost(500+200);
8904 size(17);
8905 format %{ "TEST $shift,32\n\t"
8906 "JEQ,s small\n\t"
8907 "MOV $dst.hi,$dst.lo\n\t"
8908 "XOR $dst.lo,$dst.lo\n"
8909 "small:\tSHLD $dst.hi,$dst.lo,$shift\n\t"
8910 "SHL $dst.lo,$shift" %}
8911 ins_encode( shift_left_long( dst, shift ) );
8912 ins_pipe( pipe_slow );
8913 %}
8915 // Shift Right Long by 1-31
8916 instruct shrL_eReg_1_31(eRegL dst, immI_1_31 cnt, eFlagsReg cr) %{
8917 match(Set dst (URShiftL dst cnt));
8918 effect(KILL cr);
8919 ins_cost(200);
8920 format %{ "SHRD $dst.lo,$dst.hi,$cnt\n\t"
8921 "SHR $dst.hi,$cnt" %}
8922 opcode(0xC1, 0x5, 0xAC); /* 0F/AC, then C1 /5 ib */
8923 ins_encode( move_long_small_shift(dst,cnt) );
8924 ins_pipe( ialu_reg_long );
8925 %}
8927 // Shift Right Long by 32-63
8928 instruct shrL_eReg_32_63(eRegL dst, immI_32_63 cnt, eFlagsReg cr) %{
8929 match(Set dst (URShiftL dst cnt));
8930 effect(KILL cr);
8931 ins_cost(300);
8932 format %{ "MOV $dst.lo,$dst.hi\n"
8933 "\tSHR $dst.lo,$cnt-32\n"
8934 "\tXOR $dst.hi,$dst.hi" %}
8935 opcode(0xC1, 0x5); /* C1 /5 ib */
8936 ins_encode( move_long_big_shift_clr(dst,cnt) );
8937 ins_pipe( ialu_reg_long );
8938 %}
8940 // Shift Right Long by variable
8941 instruct shrL_eReg_CL(eRegL dst, eCXRegI shift, eFlagsReg cr) %{
8942 match(Set dst (URShiftL dst shift));
8943 effect(KILL cr);
8944 ins_cost(600);
8945 size(17);
8946 format %{ "TEST $shift,32\n\t"
8947 "JEQ,s small\n\t"
8948 "MOV $dst.lo,$dst.hi\n\t"
8949 "XOR $dst.hi,$dst.hi\n"
8950 "small:\tSHRD $dst.lo,$dst.hi,$shift\n\t"
8951 "SHR $dst.hi,$shift" %}
8952 ins_encode( shift_right_long( dst, shift ) );
8953 ins_pipe( pipe_slow );
8954 %}
8956 // Shift Right Long by 1-31
8957 instruct sarL_eReg_1_31(eRegL dst, immI_1_31 cnt, eFlagsReg cr) %{
8958 match(Set dst (RShiftL dst cnt));
8959 effect(KILL cr);
8960 ins_cost(200);
8961 format %{ "SHRD $dst.lo,$dst.hi,$cnt\n\t"
8962 "SAR $dst.hi,$cnt" %}
8963 opcode(0xC1, 0x7, 0xAC); /* 0F/AC, then C1 /7 ib */
8964 ins_encode( move_long_small_shift(dst,cnt) );
8965 ins_pipe( ialu_reg_long );
8966 %}
8968 // Shift Right Long by 32-63
8969 instruct sarL_eReg_32_63( eRegL dst, immI_32_63 cnt, eFlagsReg cr) %{
8970 match(Set dst (RShiftL dst cnt));
8971 effect(KILL cr);
8972 ins_cost(300);
8973 format %{ "MOV $dst.lo,$dst.hi\n"
8974 "\tSAR $dst.lo,$cnt-32\n"
8975 "\tSAR $dst.hi,31" %}
8976 opcode(0xC1, 0x7); /* C1 /7 ib */
8977 ins_encode( move_long_big_shift_sign(dst,cnt) );
8978 ins_pipe( ialu_reg_long );
8979 %}
8981 // Shift Right arithmetic Long by variable
8982 instruct sarL_eReg_CL(eRegL dst, eCXRegI shift, eFlagsReg cr) %{
8983 match(Set dst (RShiftL dst shift));
8984 effect(KILL cr);
8985 ins_cost(600);
8986 size(18);
8987 format %{ "TEST $shift,32\n\t"
8988 "JEQ,s small\n\t"
8989 "MOV $dst.lo,$dst.hi\n\t"
8990 "SAR $dst.hi,31\n"
8991 "small:\tSHRD $dst.lo,$dst.hi,$shift\n\t"
8992 "SAR $dst.hi,$shift" %}
8993 ins_encode( shift_right_arith_long( dst, shift ) );
8994 ins_pipe( pipe_slow );
8995 %}
8998 //----------Double Instructions------------------------------------------------
8999 // Double Math
9001 // Compare & branch
9003 // P6 version of float compare, sets condition codes in EFLAGS
9004 instruct cmpD_cc_P6(eFlagsRegU cr, regD src1, regD src2, eAXRegI rax) %{
9005 predicate(VM_Version::supports_cmov() && UseSSE <=1);
9006 match(Set cr (CmpD src1 src2));
9007 effect(KILL rax);
9008 ins_cost(150);
9009 format %{ "FLD $src1\n\t"
9010 "FUCOMIP ST,$src2 // P6 instruction\n\t"
9011 "JNP exit\n\t"
9012 "MOV ah,1 // saw a NaN, set CF\n\t"
9013 "SAHF\n"
9014 "exit:\tNOP // avoid branch to branch" %}
9015 opcode(0xDF, 0x05); /* DF E8+i or DF /5 */
9016 ins_encode( Push_Reg_D(src1),
9017 OpcP, RegOpc(src2),
9018 cmpF_P6_fixup );
9019 ins_pipe( pipe_slow );
9020 %}
9022 // Compare & branch
9023 instruct cmpD_cc(eFlagsRegU cr, regD src1, regD src2, eAXRegI rax) %{
9024 predicate(UseSSE<=1);
9025 match(Set cr (CmpD src1 src2));
9026 effect(KILL rax);
9027 ins_cost(200);
9028 format %{ "FLD $src1\n\t"
9029 "FCOMp $src2\n\t"
9030 "FNSTSW AX\n\t"
9031 "TEST AX,0x400\n\t"
9032 "JZ,s flags\n\t"
9033 "MOV AH,1\t# unordered treat as LT\n"
9034 "flags:\tSAHF" %}
9035 opcode(0xD8, 0x3); /* D8 D8+i or D8 /3 */
9036 ins_encode( Push_Reg_D(src1),
9037 OpcP, RegOpc(src2),
9038 fpu_flags);
9039 ins_pipe( pipe_slow );
9040 %}
9042 // Compare vs zero into -1,0,1
9043 instruct cmpD_0(eRegI dst, regD src1, immD0 zero, eAXRegI rax, eFlagsReg cr) %{
9044 predicate(UseSSE<=1);
9045 match(Set dst (CmpD3 src1 zero));
9046 effect(KILL cr, KILL rax);
9047 ins_cost(280);
9048 format %{ "FTSTD $dst,$src1" %}
9049 opcode(0xE4, 0xD9);
9050 ins_encode( Push_Reg_D(src1),
9051 OpcS, OpcP, PopFPU,
9052 CmpF_Result(dst));
9053 ins_pipe( pipe_slow );
9054 %}
9056 // Compare into -1,0,1
9057 instruct cmpD_reg(eRegI dst, regD src1, regD src2, eAXRegI rax, eFlagsReg cr) %{
9058 predicate(UseSSE<=1);
9059 match(Set dst (CmpD3 src1 src2));
9060 effect(KILL cr, KILL rax);
9061 ins_cost(300);
9062 format %{ "FCMPD $dst,$src1,$src2" %}
9063 opcode(0xD8, 0x3); /* D8 D8+i or D8 /3 */
9064 ins_encode( Push_Reg_D(src1),
9065 OpcP, RegOpc(src2),
9066 CmpF_Result(dst));
9067 ins_pipe( pipe_slow );
9068 %}
9070 // float compare and set condition codes in EFLAGS by XMM regs
9071 instruct cmpXD_cc(eFlagsRegU cr, regXD dst, regXD src, eAXRegI rax) %{
9072 predicate(UseSSE>=2);
9073 match(Set cr (CmpD dst src));
9074 effect(KILL rax);
9075 ins_cost(125);
9076 format %{ "COMISD $dst,$src\n"
9077 "\tJNP exit\n"
9078 "\tMOV ah,1 // saw a NaN, set CF\n"
9079 "\tSAHF\n"
9080 "exit:\tNOP // avoid branch to branch" %}
9081 opcode(0x66, 0x0F, 0x2F);
9082 ins_encode(OpcP, OpcS, Opcode(tertiary), RegReg(dst, src), cmpF_P6_fixup);
9083 ins_pipe( pipe_slow );
9084 %}
9086 // float compare and set condition codes in EFLAGS by XMM regs
9087 instruct cmpXD_ccmem(eFlagsRegU cr, regXD dst, memory src, eAXRegI rax) %{
9088 predicate(UseSSE>=2);
9089 match(Set cr (CmpD dst (LoadD src)));
9090 effect(KILL rax);
9091 ins_cost(145);
9092 format %{ "COMISD $dst,$src\n"
9093 "\tJNP exit\n"
9094 "\tMOV ah,1 // saw a NaN, set CF\n"
9095 "\tSAHF\n"
9096 "exit:\tNOP // avoid branch to branch" %}
9097 opcode(0x66, 0x0F, 0x2F);
9098 ins_encode(OpcP, OpcS, Opcode(tertiary), RegMem(dst, src), cmpF_P6_fixup);
9099 ins_pipe( pipe_slow );
9100 %}
9102 // Compare into -1,0,1 in XMM
9103 instruct cmpXD_reg(eRegI dst, regXD src1, regXD src2, eFlagsReg cr) %{
9104 predicate(UseSSE>=2);
9105 match(Set dst (CmpD3 src1 src2));
9106 effect(KILL cr);
9107 ins_cost(255);
9108 format %{ "XOR $dst,$dst\n"
9109 "\tCOMISD $src1,$src2\n"
9110 "\tJP,s nan\n"
9111 "\tJEQ,s exit\n"
9112 "\tJA,s inc\n"
9113 "nan:\tDEC $dst\n"
9114 "\tJMP,s exit\n"
9115 "inc:\tINC $dst\n"
9116 "exit:"
9117 %}
9118 opcode(0x66, 0x0F, 0x2F);
9119 ins_encode(Xor_Reg(dst), OpcP, OpcS, Opcode(tertiary), RegReg(src1, src2),
9120 CmpX_Result(dst));
9121 ins_pipe( pipe_slow );
9122 %}
9124 // Compare into -1,0,1 in XMM and memory
9125 instruct cmpXD_regmem(eRegI dst, regXD src1, memory mem, eFlagsReg cr) %{
9126 predicate(UseSSE>=2);
9127 match(Set dst (CmpD3 src1 (LoadD mem)));
9128 effect(KILL cr);
9129 ins_cost(275);
9130 format %{ "COMISD $src1,$mem\n"
9131 "\tMOV $dst,0\t\t# do not blow flags\n"
9132 "\tJP,s nan\n"
9133 "\tJEQ,s exit\n"
9134 "\tJA,s inc\n"
9135 "nan:\tDEC $dst\n"
9136 "\tJMP,s exit\n"
9137 "inc:\tINC $dst\n"
9138 "exit:"
9139 %}
9140 opcode(0x66, 0x0F, 0x2F);
9141 ins_encode(OpcP, OpcS, Opcode(tertiary), RegMem(src1, mem),
9142 LdImmI(dst,0x0), CmpX_Result(dst));
9143 ins_pipe( pipe_slow );
9144 %}
9147 instruct subD_reg(regD dst, regD src) %{
9148 predicate (UseSSE <=1);
9149 match(Set dst (SubD dst src));
9151 format %{ "FLD $src\n\t"
9152 "DSUBp $dst,ST" %}
9153 opcode(0xDE, 0x5); /* DE E8+i or DE /5 */
9154 ins_cost(150);
9155 ins_encode( Push_Reg_D(src),
9156 OpcP, RegOpc(dst) );
9157 ins_pipe( fpu_reg_reg );
9158 %}
9160 instruct subD_reg_round(stackSlotD dst, regD src1, regD src2) %{
9161 predicate (UseSSE <=1);
9162 match(Set dst (RoundDouble (SubD src1 src2)));
9163 ins_cost(250);
9165 format %{ "FLD $src2\n\t"
9166 "DSUB ST,$src1\n\t"
9167 "FSTP_D $dst\t# D-round" %}
9168 opcode(0xD8, 0x5);
9169 ins_encode( Push_Reg_D(src2),
9170 OpcP, RegOpc(src1), Pop_Mem_D(dst) );
9171 ins_pipe( fpu_mem_reg_reg );
9172 %}
9175 instruct subD_reg_mem(regD dst, memory src) %{
9176 predicate (UseSSE <=1);
9177 match(Set dst (SubD dst (LoadD src)));
9178 ins_cost(150);
9180 format %{ "FLD $src\n\t"
9181 "DSUBp $dst,ST" %}
9182 opcode(0xDE, 0x5, 0xDD); /* DE C0+i */ /* LoadD DD /0 */
9183 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src),
9184 OpcP, RegOpc(dst) );
9185 ins_pipe( fpu_reg_mem );
9186 %}
9188 instruct absD_reg(regDPR1 dst, regDPR1 src) %{
9189 predicate (UseSSE<=1);
9190 match(Set dst (AbsD src));
9191 ins_cost(100);
9192 format %{ "FABS" %}
9193 opcode(0xE1, 0xD9);
9194 ins_encode( OpcS, OpcP );
9195 ins_pipe( fpu_reg_reg );
9196 %}
9198 instruct absXD_reg( regXD dst ) %{
9199 predicate(UseSSE>=2);
9200 match(Set dst (AbsD dst));
9201 format %{ "ANDPD $dst,[0x7FFFFFFFFFFFFFFF]\t# ABS D by sign masking" %}
9202 ins_encode( AbsXD_encoding(dst));
9203 ins_pipe( pipe_slow );
9204 %}
9206 instruct negD_reg(regDPR1 dst, regDPR1 src) %{
9207 predicate(UseSSE<=1);
9208 match(Set dst (NegD src));
9209 ins_cost(100);
9210 format %{ "FCHS" %}
9211 opcode(0xE0, 0xD9);
9212 ins_encode( OpcS, OpcP );
9213 ins_pipe( fpu_reg_reg );
9214 %}
9216 instruct negXD_reg( regXD dst ) %{
9217 predicate(UseSSE>=2);
9218 match(Set dst (NegD dst));
9219 format %{ "XORPD $dst,[0x8000000000000000]\t# CHS D by sign flipping" %}
9220 ins_encode %{
9221 __ xorpd($dst$$XMMRegister,
9222 ExternalAddress((address)double_signflip_pool));
9223 %}
9224 ins_pipe( pipe_slow );
9225 %}
9227 instruct addD_reg(regD dst, regD src) %{
9228 predicate(UseSSE<=1);
9229 match(Set dst (AddD dst src));
9230 format %{ "FLD $src\n\t"
9231 "DADD $dst,ST" %}
9232 size(4);
9233 ins_cost(150);
9234 opcode(0xDE, 0x0); /* DE C0+i or DE /0*/
9235 ins_encode( Push_Reg_D(src),
9236 OpcP, RegOpc(dst) );
9237 ins_pipe( fpu_reg_reg );
9238 %}
9241 instruct addD_reg_round(stackSlotD dst, regD src1, regD src2) %{
9242 predicate(UseSSE<=1);
9243 match(Set dst (RoundDouble (AddD src1 src2)));
9244 ins_cost(250);
9246 format %{ "FLD $src2\n\t"
9247 "DADD ST,$src1\n\t"
9248 "FSTP_D $dst\t# D-round" %}
9249 opcode(0xD8, 0x0); /* D8 C0+i or D8 /0*/
9250 ins_encode( Push_Reg_D(src2),
9251 OpcP, RegOpc(src1), Pop_Mem_D(dst) );
9252 ins_pipe( fpu_mem_reg_reg );
9253 %}
9256 instruct addD_reg_mem(regD dst, memory src) %{
9257 predicate(UseSSE<=1);
9258 match(Set dst (AddD dst (LoadD src)));
9259 ins_cost(150);
9261 format %{ "FLD $src\n\t"
9262 "DADDp $dst,ST" %}
9263 opcode(0xDE, 0x0, 0xDD); /* DE C0+i */ /* LoadD DD /0 */
9264 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src),
9265 OpcP, RegOpc(dst) );
9266 ins_pipe( fpu_reg_mem );
9267 %}
9269 // add-to-memory
9270 instruct addD_mem_reg(memory dst, regD src) %{
9271 predicate(UseSSE<=1);
9272 match(Set dst (StoreD dst (RoundDouble (AddD (LoadD dst) src))));
9273 ins_cost(150);
9275 format %{ "FLD_D $dst\n\t"
9276 "DADD ST,$src\n\t"
9277 "FST_D $dst" %}
9278 opcode(0xDD, 0x0);
9279 ins_encode( Opcode(0xDD), RMopc_Mem(0x00,dst),
9280 Opcode(0xD8), RegOpc(src),
9281 set_instruction_start,
9282 Opcode(0xDD), RMopc_Mem(0x03,dst) );
9283 ins_pipe( fpu_reg_mem );
9284 %}
9286 instruct addD_reg_imm1(regD dst, immD1 src) %{
9287 predicate(UseSSE<=1);
9288 match(Set dst (AddD dst src));
9289 ins_cost(125);
9290 format %{ "FLD1\n\t"
9291 "DADDp $dst,ST" %}
9292 opcode(0xDE, 0x00);
9293 ins_encode( LdImmD(src),
9294 OpcP, RegOpc(dst) );
9295 ins_pipe( fpu_reg );
9296 %}
9298 instruct addD_reg_imm(regD dst, immD src) %{
9299 predicate(UseSSE<=1 && _kids[1]->_leaf->getd() != 0.0 && _kids[1]->_leaf->getd() != 1.0 );
9300 match(Set dst (AddD dst src));
9301 ins_cost(200);
9302 format %{ "FLD_D [$src]\n\t"
9303 "DADDp $dst,ST" %}
9304 opcode(0xDE, 0x00); /* DE /0 */
9305 ins_encode( LdImmD(src),
9306 OpcP, RegOpc(dst));
9307 ins_pipe( fpu_reg_mem );
9308 %}
9310 instruct addD_reg_imm_round(stackSlotD dst, regD src, immD con) %{
9311 predicate(UseSSE<=1 && _kids[0]->_kids[1]->_leaf->getd() != 0.0 && _kids[0]->_kids[1]->_leaf->getd() != 1.0 );
9312 match(Set dst (RoundDouble (AddD src con)));
9313 ins_cost(200);
9314 format %{ "FLD_D [$con]\n\t"
9315 "DADD ST,$src\n\t"
9316 "FSTP_D $dst\t# D-round" %}
9317 opcode(0xD8, 0x00); /* D8 /0 */
9318 ins_encode( LdImmD(con),
9319 OpcP, RegOpc(src), Pop_Mem_D(dst));
9320 ins_pipe( fpu_mem_reg_con );
9321 %}
9323 // Add two double precision floating point values in xmm
9324 instruct addXD_reg(regXD dst, regXD src) %{
9325 predicate(UseSSE>=2);
9326 match(Set dst (AddD dst src));
9327 format %{ "ADDSD $dst,$src" %}
9328 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x58), RegReg(dst, src));
9329 ins_pipe( pipe_slow );
9330 %}
9332 instruct addXD_imm(regXD dst, immXD con) %{
9333 predicate(UseSSE>=2);
9334 match(Set dst (AddD dst con));
9335 format %{ "ADDSD $dst,[$con]" %}
9336 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x58), LdImmXD(dst, con) );
9337 ins_pipe( pipe_slow );
9338 %}
9340 instruct addXD_mem(regXD dst, memory mem) %{
9341 predicate(UseSSE>=2);
9342 match(Set dst (AddD dst (LoadD mem)));
9343 format %{ "ADDSD $dst,$mem" %}
9344 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x58), RegMem(dst,mem));
9345 ins_pipe( pipe_slow );
9346 %}
9348 // Sub two double precision floating point values in xmm
9349 instruct subXD_reg(regXD dst, regXD src) %{
9350 predicate(UseSSE>=2);
9351 match(Set dst (SubD dst src));
9352 format %{ "SUBSD $dst,$src" %}
9353 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x5C), RegReg(dst, src));
9354 ins_pipe( pipe_slow );
9355 %}
9357 instruct subXD_imm(regXD dst, immXD con) %{
9358 predicate(UseSSE>=2);
9359 match(Set dst (SubD dst con));
9360 format %{ "SUBSD $dst,[$con]" %}
9361 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x5C), LdImmXD(dst, con) );
9362 ins_pipe( pipe_slow );
9363 %}
9365 instruct subXD_mem(regXD dst, memory mem) %{
9366 predicate(UseSSE>=2);
9367 match(Set dst (SubD dst (LoadD mem)));
9368 format %{ "SUBSD $dst,$mem" %}
9369 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x5C), RegMem(dst,mem));
9370 ins_pipe( pipe_slow );
9371 %}
9373 // Mul two double precision floating point values in xmm
9374 instruct mulXD_reg(regXD dst, regXD src) %{
9375 predicate(UseSSE>=2);
9376 match(Set dst (MulD dst src));
9377 format %{ "MULSD $dst,$src" %}
9378 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x59), RegReg(dst, src));
9379 ins_pipe( pipe_slow );
9380 %}
9382 instruct mulXD_imm(regXD dst, immXD con) %{
9383 predicate(UseSSE>=2);
9384 match(Set dst (MulD dst con));
9385 format %{ "MULSD $dst,[$con]" %}
9386 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x59), LdImmXD(dst, con) );
9387 ins_pipe( pipe_slow );
9388 %}
9390 instruct mulXD_mem(regXD dst, memory mem) %{
9391 predicate(UseSSE>=2);
9392 match(Set dst (MulD dst (LoadD mem)));
9393 format %{ "MULSD $dst,$mem" %}
9394 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x59), RegMem(dst,mem));
9395 ins_pipe( pipe_slow );
9396 %}
9398 // Div two double precision floating point values in xmm
9399 instruct divXD_reg(regXD dst, regXD src) %{
9400 predicate(UseSSE>=2);
9401 match(Set dst (DivD dst src));
9402 format %{ "DIVSD $dst,$src" %}
9403 opcode(0xF2, 0x0F, 0x5E);
9404 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x5E), RegReg(dst, src));
9405 ins_pipe( pipe_slow );
9406 %}
9408 instruct divXD_imm(regXD dst, immXD con) %{
9409 predicate(UseSSE>=2);
9410 match(Set dst (DivD dst con));
9411 format %{ "DIVSD $dst,[$con]" %}
9412 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x5E), LdImmXD(dst, con));
9413 ins_pipe( pipe_slow );
9414 %}
9416 instruct divXD_mem(regXD dst, memory mem) %{
9417 predicate(UseSSE>=2);
9418 match(Set dst (DivD dst (LoadD mem)));
9419 format %{ "DIVSD $dst,$mem" %}
9420 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x5E), RegMem(dst,mem));
9421 ins_pipe( pipe_slow );
9422 %}
9425 instruct mulD_reg(regD dst, regD src) %{
9426 predicate(UseSSE<=1);
9427 match(Set dst (MulD dst src));
9428 format %{ "FLD $src\n\t"
9429 "DMULp $dst,ST" %}
9430 opcode(0xDE, 0x1); /* DE C8+i or DE /1*/
9431 ins_cost(150);
9432 ins_encode( Push_Reg_D(src),
9433 OpcP, RegOpc(dst) );
9434 ins_pipe( fpu_reg_reg );
9435 %}
9437 // Strict FP instruction biases argument before multiply then
9438 // biases result to avoid double rounding of subnormals.
9439 //
9440 // scale arg1 by multiplying arg1 by 2^(-15360)
9441 // load arg2
9442 // multiply scaled arg1 by arg2
9443 // rescale product by 2^(15360)
9444 //
9445 instruct strictfp_mulD_reg(regDPR1 dst, regnotDPR1 src) %{
9446 predicate( UseSSE<=1 && Compile::current()->has_method() && Compile::current()->method()->is_strict() );
9447 match(Set dst (MulD dst src));
9448 ins_cost(1); // Select this instruction for all strict FP double multiplies
9450 format %{ "FLD StubRoutines::_fpu_subnormal_bias1\n\t"
9451 "DMULp $dst,ST\n\t"
9452 "FLD $src\n\t"
9453 "DMULp $dst,ST\n\t"
9454 "FLD StubRoutines::_fpu_subnormal_bias2\n\t"
9455 "DMULp $dst,ST\n\t" %}
9456 opcode(0xDE, 0x1); /* DE C8+i or DE /1*/
9457 ins_encode( strictfp_bias1(dst),
9458 Push_Reg_D(src),
9459 OpcP, RegOpc(dst),
9460 strictfp_bias2(dst) );
9461 ins_pipe( fpu_reg_reg );
9462 %}
9464 instruct mulD_reg_imm(regD dst, immD src) %{
9465 predicate( UseSSE<=1 && _kids[1]->_leaf->getd() != 0.0 && _kids[1]->_leaf->getd() != 1.0 );
9466 match(Set dst (MulD dst src));
9467 ins_cost(200);
9468 format %{ "FLD_D [$src]\n\t"
9469 "DMULp $dst,ST" %}
9470 opcode(0xDE, 0x1); /* DE /1 */
9471 ins_encode( LdImmD(src),
9472 OpcP, RegOpc(dst) );
9473 ins_pipe( fpu_reg_mem );
9474 %}
9477 instruct mulD_reg_mem(regD dst, memory src) %{
9478 predicate( UseSSE<=1 );
9479 match(Set dst (MulD dst (LoadD src)));
9480 ins_cost(200);
9481 format %{ "FLD_D $src\n\t"
9482 "DMULp $dst,ST" %}
9483 opcode(0xDE, 0x1, 0xDD); /* DE C8+i or DE /1*/ /* LoadD DD /0 */
9484 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src),
9485 OpcP, RegOpc(dst) );
9486 ins_pipe( fpu_reg_mem );
9487 %}
9489 //
9490 // Cisc-alternate to reg-reg multiply
9491 instruct mulD_reg_mem_cisc(regD dst, regD src, memory mem) %{
9492 predicate( UseSSE<=1 );
9493 match(Set dst (MulD src (LoadD mem)));
9494 ins_cost(250);
9495 format %{ "FLD_D $mem\n\t"
9496 "DMUL ST,$src\n\t"
9497 "FSTP_D $dst" %}
9498 opcode(0xD8, 0x1, 0xD9); /* D8 C8+i */ /* LoadD D9 /0 */
9499 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,mem),
9500 OpcReg_F(src),
9501 Pop_Reg_D(dst) );
9502 ins_pipe( fpu_reg_reg_mem );
9503 %}
9506 // MACRO3 -- addD a mulD
9507 // This instruction is a '2-address' instruction in that the result goes
9508 // back to src2. This eliminates a move from the macro; possibly the
9509 // register allocator will have to add it back (and maybe not).
9510 instruct addD_mulD_reg(regD src2, regD src1, regD src0) %{
9511 predicate( UseSSE<=1 );
9512 match(Set src2 (AddD (MulD src0 src1) src2));
9513 format %{ "FLD $src0\t# ===MACRO3d===\n\t"
9514 "DMUL ST,$src1\n\t"
9515 "DADDp $src2,ST" %}
9516 ins_cost(250);
9517 opcode(0xDD); /* LoadD DD /0 */
9518 ins_encode( Push_Reg_F(src0),
9519 FMul_ST_reg(src1),
9520 FAddP_reg_ST(src2) );
9521 ins_pipe( fpu_reg_reg_reg );
9522 %}
9525 // MACRO3 -- subD a mulD
9526 instruct subD_mulD_reg(regD src2, regD src1, regD src0) %{
9527 predicate( UseSSE<=1 );
9528 match(Set src2 (SubD (MulD src0 src1) src2));
9529 format %{ "FLD $src0\t# ===MACRO3d===\n\t"
9530 "DMUL ST,$src1\n\t"
9531 "DSUBRp $src2,ST" %}
9532 ins_cost(250);
9533 ins_encode( Push_Reg_F(src0),
9534 FMul_ST_reg(src1),
9535 Opcode(0xDE), Opc_plus(0xE0,src2));
9536 ins_pipe( fpu_reg_reg_reg );
9537 %}
9540 instruct divD_reg(regD dst, regD src) %{
9541 predicate( UseSSE<=1 );
9542 match(Set dst (DivD dst src));
9544 format %{ "FLD $src\n\t"
9545 "FDIVp $dst,ST" %}
9546 opcode(0xDE, 0x7); /* DE F8+i or DE /7*/
9547 ins_cost(150);
9548 ins_encode( Push_Reg_D(src),
9549 OpcP, RegOpc(dst) );
9550 ins_pipe( fpu_reg_reg );
9551 %}
9553 // Strict FP instruction biases argument before division then
9554 // biases result, to avoid double rounding of subnormals.
9555 //
9556 // scale dividend by multiplying dividend by 2^(-15360)
9557 // load divisor
9558 // divide scaled dividend by divisor
9559 // rescale quotient by 2^(15360)
9560 //
9561 instruct strictfp_divD_reg(regDPR1 dst, regnotDPR1 src) %{
9562 predicate (UseSSE<=1);
9563 match(Set dst (DivD dst src));
9564 predicate( UseSSE<=1 && Compile::current()->has_method() && Compile::current()->method()->is_strict() );
9565 ins_cost(01);
9567 format %{ "FLD StubRoutines::_fpu_subnormal_bias1\n\t"
9568 "DMULp $dst,ST\n\t"
9569 "FLD $src\n\t"
9570 "FDIVp $dst,ST\n\t"
9571 "FLD StubRoutines::_fpu_subnormal_bias2\n\t"
9572 "DMULp $dst,ST\n\t" %}
9573 opcode(0xDE, 0x7); /* DE F8+i or DE /7*/
9574 ins_encode( strictfp_bias1(dst),
9575 Push_Reg_D(src),
9576 OpcP, RegOpc(dst),
9577 strictfp_bias2(dst) );
9578 ins_pipe( fpu_reg_reg );
9579 %}
9581 instruct divD_reg_round(stackSlotD dst, regD src1, regD src2) %{
9582 predicate( UseSSE<=1 && !(Compile::current()->has_method() && Compile::current()->method()->is_strict()) );
9583 match(Set dst (RoundDouble (DivD src1 src2)));
9585 format %{ "FLD $src1\n\t"
9586 "FDIV ST,$src2\n\t"
9587 "FSTP_D $dst\t# D-round" %}
9588 opcode(0xD8, 0x6); /* D8 F0+i or D8 /6 */
9589 ins_encode( Push_Reg_D(src1),
9590 OpcP, RegOpc(src2), Pop_Mem_D(dst) );
9591 ins_pipe( fpu_mem_reg_reg );
9592 %}
9595 instruct modD_reg(regD dst, regD src, eAXRegI rax, eFlagsReg cr) %{
9596 predicate(UseSSE<=1);
9597 match(Set dst (ModD dst src));
9598 effect(KILL rax, KILL cr); // emitModD() uses EAX and EFLAGS
9600 format %{ "DMOD $dst,$src" %}
9601 ins_cost(250);
9602 ins_encode(Push_Reg_Mod_D(dst, src),
9603 emitModD(),
9604 Push_Result_Mod_D(src),
9605 Pop_Reg_D(dst));
9606 ins_pipe( pipe_slow );
9607 %}
9609 instruct modXD_reg(regXD dst, regXD src0, regXD src1, eAXRegI rax, eFlagsReg cr) %{
9610 predicate(UseSSE>=2);
9611 match(Set dst (ModD src0 src1));
9612 effect(KILL rax, KILL cr);
9614 format %{ "SUB ESP,8\t # DMOD\n"
9615 "\tMOVSD [ESP+0],$src1\n"
9616 "\tFLD_D [ESP+0]\n"
9617 "\tMOVSD [ESP+0],$src0\n"
9618 "\tFLD_D [ESP+0]\n"
9619 "loop:\tFPREM\n"
9620 "\tFWAIT\n"
9621 "\tFNSTSW AX\n"
9622 "\tSAHF\n"
9623 "\tJP loop\n"
9624 "\tFSTP_D [ESP+0]\n"
9625 "\tMOVSD $dst,[ESP+0]\n"
9626 "\tADD ESP,8\n"
9627 "\tFSTP ST0\t # Restore FPU Stack"
9628 %}
9629 ins_cost(250);
9630 ins_encode( Push_ModD_encoding(src0, src1), emitModD(), Push_ResultXD(dst), PopFPU);
9631 ins_pipe( pipe_slow );
9632 %}
9634 instruct sinD_reg(regDPR1 dst, regDPR1 src) %{
9635 predicate (UseSSE<=1);
9636 match(Set dst (SinD src));
9637 ins_cost(1800);
9638 format %{ "DSIN $dst" %}
9639 opcode(0xD9, 0xFE);
9640 ins_encode( OpcP, OpcS );
9641 ins_pipe( pipe_slow );
9642 %}
9644 instruct sinXD_reg(regXD dst, eFlagsReg cr) %{
9645 predicate (UseSSE>=2);
9646 match(Set dst (SinD dst));
9647 effect(KILL cr); // Push_{Src|Result}XD() uses "{SUB|ADD} ESP,8"
9648 ins_cost(1800);
9649 format %{ "DSIN $dst" %}
9650 opcode(0xD9, 0xFE);
9651 ins_encode( Push_SrcXD(dst), OpcP, OpcS, Push_ResultXD(dst) );
9652 ins_pipe( pipe_slow );
9653 %}
9655 instruct cosD_reg(regDPR1 dst, regDPR1 src) %{
9656 predicate (UseSSE<=1);
9657 match(Set dst (CosD src));
9658 ins_cost(1800);
9659 format %{ "DCOS $dst" %}
9660 opcode(0xD9, 0xFF);
9661 ins_encode( OpcP, OpcS );
9662 ins_pipe( pipe_slow );
9663 %}
9665 instruct cosXD_reg(regXD dst, eFlagsReg cr) %{
9666 predicate (UseSSE>=2);
9667 match(Set dst (CosD dst));
9668 effect(KILL cr); // Push_{Src|Result}XD() uses "{SUB|ADD} ESP,8"
9669 ins_cost(1800);
9670 format %{ "DCOS $dst" %}
9671 opcode(0xD9, 0xFF);
9672 ins_encode( Push_SrcXD(dst), OpcP, OpcS, Push_ResultXD(dst) );
9673 ins_pipe( pipe_slow );
9674 %}
9676 instruct tanD_reg(regDPR1 dst, regDPR1 src) %{
9677 predicate (UseSSE<=1);
9678 match(Set dst(TanD src));
9679 format %{ "DTAN $dst" %}
9680 ins_encode( Opcode(0xD9), Opcode(0xF2), // fptan
9681 Opcode(0xDD), Opcode(0xD8)); // fstp st
9682 ins_pipe( pipe_slow );
9683 %}
9685 instruct tanXD_reg(regXD dst, eFlagsReg cr) %{
9686 predicate (UseSSE>=2);
9687 match(Set dst(TanD dst));
9688 effect(KILL cr); // Push_{Src|Result}XD() uses "{SUB|ADD} ESP,8"
9689 format %{ "DTAN $dst" %}
9690 ins_encode( Push_SrcXD(dst),
9691 Opcode(0xD9), Opcode(0xF2), // fptan
9692 Opcode(0xDD), Opcode(0xD8), // fstp st
9693 Push_ResultXD(dst) );
9694 ins_pipe( pipe_slow );
9695 %}
9697 instruct atanD_reg(regD dst, regD src) %{
9698 predicate (UseSSE<=1);
9699 match(Set dst(AtanD dst src));
9700 format %{ "DATA $dst,$src" %}
9701 opcode(0xD9, 0xF3);
9702 ins_encode( Push_Reg_D(src),
9703 OpcP, OpcS, RegOpc(dst) );
9704 ins_pipe( pipe_slow );
9705 %}
9707 instruct atanXD_reg(regXD dst, regXD src, eFlagsReg cr) %{
9708 predicate (UseSSE>=2);
9709 match(Set dst(AtanD dst src));
9710 effect(KILL cr); // Push_{Src|Result}XD() uses "{SUB|ADD} ESP,8"
9711 format %{ "DATA $dst,$src" %}
9712 opcode(0xD9, 0xF3);
9713 ins_encode( Push_SrcXD(src),
9714 OpcP, OpcS, Push_ResultXD(dst) );
9715 ins_pipe( pipe_slow );
9716 %}
9718 instruct sqrtD_reg(regD dst, regD src) %{
9719 predicate (UseSSE<=1);
9720 match(Set dst (SqrtD src));
9721 format %{ "DSQRT $dst,$src" %}
9722 opcode(0xFA, 0xD9);
9723 ins_encode( Push_Reg_D(src),
9724 OpcS, OpcP, Pop_Reg_D(dst) );
9725 ins_pipe( pipe_slow );
9726 %}
9728 instruct powD_reg(regD X, regDPR1 Y, eAXRegI rax, eBXRegI rbx, eCXRegI rcx) %{
9729 predicate (UseSSE<=1);
9730 match(Set Y (PowD X Y)); // Raise X to the Yth power
9731 effect(KILL rax, KILL rbx, KILL rcx);
9732 format %{ "SUB ESP,8\t\t# Fast-path POW encoding\n\t"
9733 "FLD_D $X\n\t"
9734 "FYL2X \t\t\t# Q=Y*ln2(X)\n\t"
9736 "FDUP \t\t\t# Q Q\n\t"
9737 "FRNDINT\t\t\t# int(Q) Q\n\t"
9738 "FSUB ST(1),ST(0)\t# int(Q) frac(Q)\n\t"
9739 "FISTP dword [ESP]\n\t"
9740 "F2XM1 \t\t\t# 2^frac(Q)-1 int(Q)\n\t"
9741 "FLD1 \t\t\t# 1 2^frac(Q)-1 int(Q)\n\t"
9742 "FADDP \t\t\t# 2^frac(Q) int(Q)\n\t" // could use FADD [1.000] instead
9743 "MOV EAX,[ESP]\t# Pick up int(Q)\n\t"
9744 "MOV ECX,0xFFFFF800\t# Overflow mask\n\t"
9745 "ADD EAX,1023\t\t# Double exponent bias\n\t"
9746 "MOV EBX,EAX\t\t# Preshifted biased expo\n\t"
9747 "SHL EAX,20\t\t# Shift exponent into place\n\t"
9748 "TEST EBX,ECX\t\t# Check for overflow\n\t"
9749 "CMOVne EAX,ECX\t\t# If overflow, stuff NaN into EAX\n\t"
9750 "MOV [ESP+4],EAX\t# Marshal 64-bit scaling double\n\t"
9751 "MOV [ESP+0],0\n\t"
9752 "FMUL ST(0),[ESP+0]\t# Scale\n\t"
9754 "ADD ESP,8"
9755 %}
9756 ins_encode( push_stack_temp_qword,
9757 Push_Reg_D(X),
9758 Opcode(0xD9), Opcode(0xF1), // fyl2x
9759 pow_exp_core_encoding,
9760 pop_stack_temp_qword);
9761 ins_pipe( pipe_slow );
9762 %}
9764 instruct powXD_reg(regXD dst, regXD src0, regXD src1, regDPR1 tmp1, eAXRegI rax, eBXRegI rbx, eCXRegI rcx ) %{
9765 predicate (UseSSE>=2);
9766 match(Set dst (PowD src0 src1)); // Raise src0 to the src1'th power
9767 effect(KILL tmp1, KILL rax, KILL rbx, KILL rcx );
9768 format %{ "SUB ESP,8\t\t# Fast-path POW encoding\n\t"
9769 "MOVSD [ESP],$src1\n\t"
9770 "FLD FPR1,$src1\n\t"
9771 "MOVSD [ESP],$src0\n\t"
9772 "FLD FPR1,$src0\n\t"
9773 "FYL2X \t\t\t# Q=Y*ln2(X)\n\t"
9775 "FDUP \t\t\t# Q Q\n\t"
9776 "FRNDINT\t\t\t# int(Q) Q\n\t"
9777 "FSUB ST(1),ST(0)\t# int(Q) frac(Q)\n\t"
9778 "FISTP dword [ESP]\n\t"
9779 "F2XM1 \t\t\t# 2^frac(Q)-1 int(Q)\n\t"
9780 "FLD1 \t\t\t# 1 2^frac(Q)-1 int(Q)\n\t"
9781 "FADDP \t\t\t# 2^frac(Q) int(Q)\n\t" // could use FADD [1.000] instead
9782 "MOV EAX,[ESP]\t# Pick up int(Q)\n\t"
9783 "MOV ECX,0xFFFFF800\t# Overflow mask\n\t"
9784 "ADD EAX,1023\t\t# Double exponent bias\n\t"
9785 "MOV EBX,EAX\t\t# Preshifted biased expo\n\t"
9786 "SHL EAX,20\t\t# Shift exponent into place\n\t"
9787 "TEST EBX,ECX\t\t# Check for overflow\n\t"
9788 "CMOVne EAX,ECX\t\t# If overflow, stuff NaN into EAX\n\t"
9789 "MOV [ESP+4],EAX\t# Marshal 64-bit scaling double\n\t"
9790 "MOV [ESP+0],0\n\t"
9791 "FMUL ST(0),[ESP+0]\t# Scale\n\t"
9793 "FST_D [ESP]\n\t"
9794 "MOVSD $dst,[ESP]\n\t"
9795 "ADD ESP,8"
9796 %}
9797 ins_encode( push_stack_temp_qword,
9798 push_xmm_to_fpr1(src1),
9799 push_xmm_to_fpr1(src0),
9800 Opcode(0xD9), Opcode(0xF1), // fyl2x
9801 pow_exp_core_encoding,
9802 Push_ResultXD(dst) );
9803 ins_pipe( pipe_slow );
9804 %}
9807 instruct expD_reg(regDPR1 dpr1, eAXRegI rax, eBXRegI rbx, eCXRegI rcx) %{
9808 predicate (UseSSE<=1);
9809 match(Set dpr1 (ExpD dpr1));
9810 effect(KILL rax, KILL rbx, KILL rcx);
9811 format %{ "SUB ESP,8\t\t# Fast-path EXP encoding"
9812 "FLDL2E \t\t\t# Ld log2(e) X\n\t"
9813 "FMULP \t\t\t# Q=X*log2(e)\n\t"
9815 "FDUP \t\t\t# Q Q\n\t"
9816 "FRNDINT\t\t\t# int(Q) Q\n\t"
9817 "FSUB ST(1),ST(0)\t# int(Q) frac(Q)\n\t"
9818 "FISTP dword [ESP]\n\t"
9819 "F2XM1 \t\t\t# 2^frac(Q)-1 int(Q)\n\t"
9820 "FLD1 \t\t\t# 1 2^frac(Q)-1 int(Q)\n\t"
9821 "FADDP \t\t\t# 2^frac(Q) int(Q)\n\t" // could use FADD [1.000] instead
9822 "MOV EAX,[ESP]\t# Pick up int(Q)\n\t"
9823 "MOV ECX,0xFFFFF800\t# Overflow mask\n\t"
9824 "ADD EAX,1023\t\t# Double exponent bias\n\t"
9825 "MOV EBX,EAX\t\t# Preshifted biased expo\n\t"
9826 "SHL EAX,20\t\t# Shift exponent into place\n\t"
9827 "TEST EBX,ECX\t\t# Check for overflow\n\t"
9828 "CMOVne EAX,ECX\t\t# If overflow, stuff NaN into EAX\n\t"
9829 "MOV [ESP+4],EAX\t# Marshal 64-bit scaling double\n\t"
9830 "MOV [ESP+0],0\n\t"
9831 "FMUL ST(0),[ESP+0]\t# Scale\n\t"
9833 "ADD ESP,8"
9834 %}
9835 ins_encode( push_stack_temp_qword,
9836 Opcode(0xD9), Opcode(0xEA), // fldl2e
9837 Opcode(0xDE), Opcode(0xC9), // fmulp
9838 pow_exp_core_encoding,
9839 pop_stack_temp_qword);
9840 ins_pipe( pipe_slow );
9841 %}
9843 instruct expXD_reg(regXD dst, regXD src, regDPR1 tmp1, eAXRegI rax, eBXRegI rbx, eCXRegI rcx) %{
9844 predicate (UseSSE>=2);
9845 match(Set dst (ExpD src));
9846 effect(KILL tmp1, KILL rax, KILL rbx, KILL rcx);
9847 format %{ "SUB ESP,8\t\t# Fast-path EXP encoding\n\t"
9848 "MOVSD [ESP],$src\n\t"
9849 "FLDL2E \t\t\t# Ld log2(e) X\n\t"
9850 "FMULP \t\t\t# Q=X*log2(e) X\n\t"
9852 "FDUP \t\t\t# Q Q\n\t"
9853 "FRNDINT\t\t\t# int(Q) Q\n\t"
9854 "FSUB ST(1),ST(0)\t# int(Q) frac(Q)\n\t"
9855 "FISTP dword [ESP]\n\t"
9856 "F2XM1 \t\t\t# 2^frac(Q)-1 int(Q)\n\t"
9857 "FLD1 \t\t\t# 1 2^frac(Q)-1 int(Q)\n\t"
9858 "FADDP \t\t\t# 2^frac(Q) int(Q)\n\t" // could use FADD [1.000] instead
9859 "MOV EAX,[ESP]\t# Pick up int(Q)\n\t"
9860 "MOV ECX,0xFFFFF800\t# Overflow mask\n\t"
9861 "ADD EAX,1023\t\t# Double exponent bias\n\t"
9862 "MOV EBX,EAX\t\t# Preshifted biased expo\n\t"
9863 "SHL EAX,20\t\t# Shift exponent into place\n\t"
9864 "TEST EBX,ECX\t\t# Check for overflow\n\t"
9865 "CMOVne EAX,ECX\t\t# If overflow, stuff NaN into EAX\n\t"
9866 "MOV [ESP+4],EAX\t# Marshal 64-bit scaling double\n\t"
9867 "MOV [ESP+0],0\n\t"
9868 "FMUL ST(0),[ESP+0]\t# Scale\n\t"
9870 "FST_D [ESP]\n\t"
9871 "MOVSD $dst,[ESP]\n\t"
9872 "ADD ESP,8"
9873 %}
9874 ins_encode( Push_SrcXD(src),
9875 Opcode(0xD9), Opcode(0xEA), // fldl2e
9876 Opcode(0xDE), Opcode(0xC9), // fmulp
9877 pow_exp_core_encoding,
9878 Push_ResultXD(dst) );
9879 ins_pipe( pipe_slow );
9880 %}
9884 instruct log10D_reg(regDPR1 dst, regDPR1 src) %{
9885 predicate (UseSSE<=1);
9886 // The source Double operand on FPU stack
9887 match(Set dst (Log10D src));
9888 // fldlg2 ; push log_10(2) on the FPU stack; full 80-bit number
9889 // fxch ; swap ST(0) with ST(1)
9890 // fyl2x ; compute log_10(2) * log_2(x)
9891 format %{ "FLDLG2 \t\t\t#Log10\n\t"
9892 "FXCH \n\t"
9893 "FYL2X \t\t\t# Q=Log10*Log_2(x)"
9894 %}
9895 ins_encode( Opcode(0xD9), Opcode(0xEC), // fldlg2
9896 Opcode(0xD9), Opcode(0xC9), // fxch
9897 Opcode(0xD9), Opcode(0xF1)); // fyl2x
9899 ins_pipe( pipe_slow );
9900 %}
9902 instruct log10XD_reg(regXD dst, regXD src, eFlagsReg cr) %{
9903 predicate (UseSSE>=2);
9904 effect(KILL cr);
9905 match(Set dst (Log10D src));
9906 // fldlg2 ; push log_10(2) on the FPU stack; full 80-bit number
9907 // fyl2x ; compute log_10(2) * log_2(x)
9908 format %{ "FLDLG2 \t\t\t#Log10\n\t"
9909 "FYL2X \t\t\t# Q=Log10*Log_2(x)"
9910 %}
9911 ins_encode( Opcode(0xD9), Opcode(0xEC), // fldlg2
9912 Push_SrcXD(src),
9913 Opcode(0xD9), Opcode(0xF1), // fyl2x
9914 Push_ResultXD(dst));
9916 ins_pipe( pipe_slow );
9917 %}
9919 instruct logD_reg(regDPR1 dst, regDPR1 src) %{
9920 predicate (UseSSE<=1);
9921 // The source Double operand on FPU stack
9922 match(Set dst (LogD src));
9923 // fldln2 ; push log_e(2) on the FPU stack; full 80-bit number
9924 // fxch ; swap ST(0) with ST(1)
9925 // fyl2x ; compute log_e(2) * log_2(x)
9926 format %{ "FLDLN2 \t\t\t#Log_e\n\t"
9927 "FXCH \n\t"
9928 "FYL2X \t\t\t# Q=Log_e*Log_2(x)"
9929 %}
9930 ins_encode( Opcode(0xD9), Opcode(0xED), // fldln2
9931 Opcode(0xD9), Opcode(0xC9), // fxch
9932 Opcode(0xD9), Opcode(0xF1)); // fyl2x
9934 ins_pipe( pipe_slow );
9935 %}
9937 instruct logXD_reg(regXD dst, regXD src, eFlagsReg cr) %{
9938 predicate (UseSSE>=2);
9939 effect(KILL cr);
9940 // The source and result Double operands in XMM registers
9941 match(Set dst (LogD src));
9942 // fldln2 ; push log_e(2) on the FPU stack; full 80-bit number
9943 // fyl2x ; compute log_e(2) * log_2(x)
9944 format %{ "FLDLN2 \t\t\t#Log_e\n\t"
9945 "FYL2X \t\t\t# Q=Log_e*Log_2(x)"
9946 %}
9947 ins_encode( Opcode(0xD9), Opcode(0xED), // fldln2
9948 Push_SrcXD(src),
9949 Opcode(0xD9), Opcode(0xF1), // fyl2x
9950 Push_ResultXD(dst));
9951 ins_pipe( pipe_slow );
9952 %}
9954 //-------------Float Instructions-------------------------------
9955 // Float Math
9957 // Code for float compare:
9958 // fcompp();
9959 // fwait(); fnstsw_ax();
9960 // sahf();
9961 // movl(dst, unordered_result);
9962 // jcc(Assembler::parity, exit);
9963 // movl(dst, less_result);
9964 // jcc(Assembler::below, exit);
9965 // movl(dst, equal_result);
9966 // jcc(Assembler::equal, exit);
9967 // movl(dst, greater_result);
9968 // exit:
9970 // P6 version of float compare, sets condition codes in EFLAGS
9971 instruct cmpF_cc_P6(eFlagsRegU cr, regF src1, regF src2, eAXRegI rax) %{
9972 predicate(VM_Version::supports_cmov() && UseSSE == 0);
9973 match(Set cr (CmpF src1 src2));
9974 effect(KILL rax);
9975 ins_cost(150);
9976 format %{ "FLD $src1\n\t"
9977 "FUCOMIP ST,$src2 // P6 instruction\n\t"
9978 "JNP exit\n\t"
9979 "MOV ah,1 // saw a NaN, set CF (treat as LT)\n\t"
9980 "SAHF\n"
9981 "exit:\tNOP // avoid branch to branch" %}
9982 opcode(0xDF, 0x05); /* DF E8+i or DF /5 */
9983 ins_encode( Push_Reg_D(src1),
9984 OpcP, RegOpc(src2),
9985 cmpF_P6_fixup );
9986 ins_pipe( pipe_slow );
9987 %}
9990 // Compare & branch
9991 instruct cmpF_cc(eFlagsRegU cr, regF src1, regF src2, eAXRegI rax) %{
9992 predicate(UseSSE == 0);
9993 match(Set cr (CmpF src1 src2));
9994 effect(KILL rax);
9995 ins_cost(200);
9996 format %{ "FLD $src1\n\t"
9997 "FCOMp $src2\n\t"
9998 "FNSTSW AX\n\t"
9999 "TEST AX,0x400\n\t"
10000 "JZ,s flags\n\t"
10001 "MOV AH,1\t# unordered treat as LT\n"
10002 "flags:\tSAHF" %}
10003 opcode(0xD8, 0x3); /* D8 D8+i or D8 /3 */
10004 ins_encode( Push_Reg_D(src1),
10005 OpcP, RegOpc(src2),
10006 fpu_flags);
10007 ins_pipe( pipe_slow );
10008 %}
10010 // Compare vs zero into -1,0,1
10011 instruct cmpF_0(eRegI dst, regF src1, immF0 zero, eAXRegI rax, eFlagsReg cr) %{
10012 predicate(UseSSE == 0);
10013 match(Set dst (CmpF3 src1 zero));
10014 effect(KILL cr, KILL rax);
10015 ins_cost(280);
10016 format %{ "FTSTF $dst,$src1" %}
10017 opcode(0xE4, 0xD9);
10018 ins_encode( Push_Reg_D(src1),
10019 OpcS, OpcP, PopFPU,
10020 CmpF_Result(dst));
10021 ins_pipe( pipe_slow );
10022 %}
10024 // Compare into -1,0,1
10025 instruct cmpF_reg(eRegI dst, regF src1, regF src2, eAXRegI rax, eFlagsReg cr) %{
10026 predicate(UseSSE == 0);
10027 match(Set dst (CmpF3 src1 src2));
10028 effect(KILL cr, KILL rax);
10029 ins_cost(300);
10030 format %{ "FCMPF $dst,$src1,$src2" %}
10031 opcode(0xD8, 0x3); /* D8 D8+i or D8 /3 */
10032 ins_encode( Push_Reg_D(src1),
10033 OpcP, RegOpc(src2),
10034 CmpF_Result(dst));
10035 ins_pipe( pipe_slow );
10036 %}
10038 // float compare and set condition codes in EFLAGS by XMM regs
10039 instruct cmpX_cc(eFlagsRegU cr, regX dst, regX src, eAXRegI rax) %{
10040 predicate(UseSSE>=1);
10041 match(Set cr (CmpF dst src));
10042 effect(KILL rax);
10043 ins_cost(145);
10044 format %{ "COMISS $dst,$src\n"
10045 "\tJNP exit\n"
10046 "\tMOV ah,1 // saw a NaN, set CF\n"
10047 "\tSAHF\n"
10048 "exit:\tNOP // avoid branch to branch" %}
10049 opcode(0x0F, 0x2F);
10050 ins_encode(OpcP, OpcS, RegReg(dst, src), cmpF_P6_fixup);
10051 ins_pipe( pipe_slow );
10052 %}
10054 // float compare and set condition codes in EFLAGS by XMM regs
10055 instruct cmpX_ccmem(eFlagsRegU cr, regX dst, memory src, eAXRegI rax) %{
10056 predicate(UseSSE>=1);
10057 match(Set cr (CmpF dst (LoadF src)));
10058 effect(KILL rax);
10059 ins_cost(165);
10060 format %{ "COMISS $dst,$src\n"
10061 "\tJNP exit\n"
10062 "\tMOV ah,1 // saw a NaN, set CF\n"
10063 "\tSAHF\n"
10064 "exit:\tNOP // avoid branch to branch" %}
10065 opcode(0x0F, 0x2F);
10066 ins_encode(OpcP, OpcS, RegMem(dst, src), cmpF_P6_fixup);
10067 ins_pipe( pipe_slow );
10068 %}
10070 // Compare into -1,0,1 in XMM
10071 instruct cmpX_reg(eRegI dst, regX src1, regX src2, eFlagsReg cr) %{
10072 predicate(UseSSE>=1);
10073 match(Set dst (CmpF3 src1 src2));
10074 effect(KILL cr);
10075 ins_cost(255);
10076 format %{ "XOR $dst,$dst\n"
10077 "\tCOMISS $src1,$src2\n"
10078 "\tJP,s nan\n"
10079 "\tJEQ,s exit\n"
10080 "\tJA,s inc\n"
10081 "nan:\tDEC $dst\n"
10082 "\tJMP,s exit\n"
10083 "inc:\tINC $dst\n"
10084 "exit:"
10085 %}
10086 opcode(0x0F, 0x2F);
10087 ins_encode(Xor_Reg(dst), OpcP, OpcS, RegReg(src1, src2), CmpX_Result(dst));
10088 ins_pipe( pipe_slow );
10089 %}
10091 // Compare into -1,0,1 in XMM and memory
10092 instruct cmpX_regmem(eRegI dst, regX src1, memory mem, eFlagsReg cr) %{
10093 predicate(UseSSE>=1);
10094 match(Set dst (CmpF3 src1 (LoadF mem)));
10095 effect(KILL cr);
10096 ins_cost(275);
10097 format %{ "COMISS $src1,$mem\n"
10098 "\tMOV $dst,0\t\t# do not blow flags\n"
10099 "\tJP,s nan\n"
10100 "\tJEQ,s exit\n"
10101 "\tJA,s inc\n"
10102 "nan:\tDEC $dst\n"
10103 "\tJMP,s exit\n"
10104 "inc:\tINC $dst\n"
10105 "exit:"
10106 %}
10107 opcode(0x0F, 0x2F);
10108 ins_encode(OpcP, OpcS, RegMem(src1, mem), LdImmI(dst,0x0), CmpX_Result(dst));
10109 ins_pipe( pipe_slow );
10110 %}
10112 // Spill to obtain 24-bit precision
10113 instruct subF24_reg(stackSlotF dst, regF src1, regF src2) %{
10114 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10115 match(Set dst (SubF src1 src2));
10117 format %{ "FSUB $dst,$src1 - $src2" %}
10118 opcode(0xD8, 0x4); /* D8 E0+i or D8 /4 mod==0x3 ;; result in TOS */
10119 ins_encode( Push_Reg_F(src1),
10120 OpcReg_F(src2),
10121 Pop_Mem_F(dst) );
10122 ins_pipe( fpu_mem_reg_reg );
10123 %}
10124 //
10125 // This instruction does not round to 24-bits
10126 instruct subF_reg(regF dst, regF src) %{
10127 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10128 match(Set dst (SubF dst src));
10130 format %{ "FSUB $dst,$src" %}
10131 opcode(0xDE, 0x5); /* DE E8+i or DE /5 */
10132 ins_encode( Push_Reg_F(src),
10133 OpcP, RegOpc(dst) );
10134 ins_pipe( fpu_reg_reg );
10135 %}
10137 // Spill to obtain 24-bit precision
10138 instruct addF24_reg(stackSlotF dst, regF src1, regF src2) %{
10139 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10140 match(Set dst (AddF src1 src2));
10142 format %{ "FADD $dst,$src1,$src2" %}
10143 opcode(0xD8, 0x0); /* D8 C0+i */
10144 ins_encode( Push_Reg_F(src2),
10145 OpcReg_F(src1),
10146 Pop_Mem_F(dst) );
10147 ins_pipe( fpu_mem_reg_reg );
10148 %}
10149 //
10150 // This instruction does not round to 24-bits
10151 instruct addF_reg(regF dst, regF src) %{
10152 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10153 match(Set dst (AddF dst src));
10155 format %{ "FLD $src\n\t"
10156 "FADDp $dst,ST" %}
10157 opcode(0xDE, 0x0); /* DE C0+i or DE /0*/
10158 ins_encode( Push_Reg_F(src),
10159 OpcP, RegOpc(dst) );
10160 ins_pipe( fpu_reg_reg );
10161 %}
10163 // Add two single precision floating point values in xmm
10164 instruct addX_reg(regX dst, regX src) %{
10165 predicate(UseSSE>=1);
10166 match(Set dst (AddF dst src));
10167 format %{ "ADDSS $dst,$src" %}
10168 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x58), RegReg(dst, src));
10169 ins_pipe( pipe_slow );
10170 %}
10172 instruct addX_imm(regX dst, immXF con) %{
10173 predicate(UseSSE>=1);
10174 match(Set dst (AddF dst con));
10175 format %{ "ADDSS $dst,[$con]" %}
10176 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x58), LdImmX(dst, con) );
10177 ins_pipe( pipe_slow );
10178 %}
10180 instruct addX_mem(regX dst, memory mem) %{
10181 predicate(UseSSE>=1);
10182 match(Set dst (AddF dst (LoadF mem)));
10183 format %{ "ADDSS $dst,$mem" %}
10184 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x58), RegMem(dst, mem));
10185 ins_pipe( pipe_slow );
10186 %}
10188 // Subtract two single precision floating point values in xmm
10189 instruct subX_reg(regX dst, regX src) %{
10190 predicate(UseSSE>=1);
10191 match(Set dst (SubF dst src));
10192 format %{ "SUBSS $dst,$src" %}
10193 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x5C), RegReg(dst, src));
10194 ins_pipe( pipe_slow );
10195 %}
10197 instruct subX_imm(regX dst, immXF con) %{
10198 predicate(UseSSE>=1);
10199 match(Set dst (SubF dst con));
10200 format %{ "SUBSS $dst,[$con]" %}
10201 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x5C), LdImmX(dst, con) );
10202 ins_pipe( pipe_slow );
10203 %}
10205 instruct subX_mem(regX dst, memory mem) %{
10206 predicate(UseSSE>=1);
10207 match(Set dst (SubF dst (LoadF mem)));
10208 format %{ "SUBSS $dst,$mem" %}
10209 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x5C), RegMem(dst,mem));
10210 ins_pipe( pipe_slow );
10211 %}
10213 // Multiply two single precision floating point values in xmm
10214 instruct mulX_reg(regX dst, regX src) %{
10215 predicate(UseSSE>=1);
10216 match(Set dst (MulF dst src));
10217 format %{ "MULSS $dst,$src" %}
10218 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x59), RegReg(dst, src));
10219 ins_pipe( pipe_slow );
10220 %}
10222 instruct mulX_imm(regX dst, immXF con) %{
10223 predicate(UseSSE>=1);
10224 match(Set dst (MulF dst con));
10225 format %{ "MULSS $dst,[$con]" %}
10226 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x59), LdImmX(dst, con) );
10227 ins_pipe( pipe_slow );
10228 %}
10230 instruct mulX_mem(regX dst, memory mem) %{
10231 predicate(UseSSE>=1);
10232 match(Set dst (MulF dst (LoadF mem)));
10233 format %{ "MULSS $dst,$mem" %}
10234 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x59), RegMem(dst,mem));
10235 ins_pipe( pipe_slow );
10236 %}
10238 // Divide two single precision floating point values in xmm
10239 instruct divX_reg(regX dst, regX src) %{
10240 predicate(UseSSE>=1);
10241 match(Set dst (DivF dst src));
10242 format %{ "DIVSS $dst,$src" %}
10243 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x5E), RegReg(dst, src));
10244 ins_pipe( pipe_slow );
10245 %}
10247 instruct divX_imm(regX dst, immXF con) %{
10248 predicate(UseSSE>=1);
10249 match(Set dst (DivF dst con));
10250 format %{ "DIVSS $dst,[$con]" %}
10251 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x5E), LdImmX(dst, con) );
10252 ins_pipe( pipe_slow );
10253 %}
10255 instruct divX_mem(regX dst, memory mem) %{
10256 predicate(UseSSE>=1);
10257 match(Set dst (DivF dst (LoadF mem)));
10258 format %{ "DIVSS $dst,$mem" %}
10259 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x5E), RegMem(dst,mem));
10260 ins_pipe( pipe_slow );
10261 %}
10263 // Get the square root of a single precision floating point values in xmm
10264 instruct sqrtX_reg(regX dst, regX src) %{
10265 predicate(UseSSE>=1);
10266 match(Set dst (ConvD2F (SqrtD (ConvF2D src))));
10267 format %{ "SQRTSS $dst,$src" %}
10268 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x51), RegReg(dst, src));
10269 ins_pipe( pipe_slow );
10270 %}
10272 instruct sqrtX_mem(regX dst, memory mem) %{
10273 predicate(UseSSE>=1);
10274 match(Set dst (ConvD2F (SqrtD (ConvF2D (LoadF mem)))));
10275 format %{ "SQRTSS $dst,$mem" %}
10276 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x51), RegMem(dst, mem));
10277 ins_pipe( pipe_slow );
10278 %}
10280 // Get the square root of a double precision floating point values in xmm
10281 instruct sqrtXD_reg(regXD dst, regXD src) %{
10282 predicate(UseSSE>=2);
10283 match(Set dst (SqrtD src));
10284 format %{ "SQRTSD $dst,$src" %}
10285 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x51), RegReg(dst, src));
10286 ins_pipe( pipe_slow );
10287 %}
10289 instruct sqrtXD_mem(regXD dst, memory mem) %{
10290 predicate(UseSSE>=2);
10291 match(Set dst (SqrtD (LoadD mem)));
10292 format %{ "SQRTSD $dst,$mem" %}
10293 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x51), RegMem(dst, mem));
10294 ins_pipe( pipe_slow );
10295 %}
10297 instruct absF_reg(regFPR1 dst, regFPR1 src) %{
10298 predicate(UseSSE==0);
10299 match(Set dst (AbsF src));
10300 ins_cost(100);
10301 format %{ "FABS" %}
10302 opcode(0xE1, 0xD9);
10303 ins_encode( OpcS, OpcP );
10304 ins_pipe( fpu_reg_reg );
10305 %}
10307 instruct absX_reg(regX dst ) %{
10308 predicate(UseSSE>=1);
10309 match(Set dst (AbsF dst));
10310 format %{ "ANDPS $dst,[0x7FFFFFFF]\t# ABS F by sign masking" %}
10311 ins_encode( AbsXF_encoding(dst));
10312 ins_pipe( pipe_slow );
10313 %}
10315 instruct negF_reg(regFPR1 dst, regFPR1 src) %{
10316 predicate(UseSSE==0);
10317 match(Set dst (NegF src));
10318 ins_cost(100);
10319 format %{ "FCHS" %}
10320 opcode(0xE0, 0xD9);
10321 ins_encode( OpcS, OpcP );
10322 ins_pipe( fpu_reg_reg );
10323 %}
10325 instruct negX_reg( regX dst ) %{
10326 predicate(UseSSE>=1);
10327 match(Set dst (NegF dst));
10328 format %{ "XORPS $dst,[0x80000000]\t# CHS F by sign flipping" %}
10329 ins_encode( NegXF_encoding(dst));
10330 ins_pipe( pipe_slow );
10331 %}
10333 // Cisc-alternate to addF_reg
10334 // Spill to obtain 24-bit precision
10335 instruct addF24_reg_mem(stackSlotF dst, regF src1, memory src2) %{
10336 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10337 match(Set dst (AddF src1 (LoadF src2)));
10339 format %{ "FLD $src2\n\t"
10340 "FADD ST,$src1\n\t"
10341 "FSTP_S $dst" %}
10342 opcode(0xD8, 0x0, 0xD9); /* D8 C0+i */ /* LoadF D9 /0 */
10343 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10344 OpcReg_F(src1),
10345 Pop_Mem_F(dst) );
10346 ins_pipe( fpu_mem_reg_mem );
10347 %}
10348 //
10349 // Cisc-alternate to addF_reg
10350 // This instruction does not round to 24-bits
10351 instruct addF_reg_mem(regF dst, memory src) %{
10352 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10353 match(Set dst (AddF dst (LoadF src)));
10355 format %{ "FADD $dst,$src" %}
10356 opcode(0xDE, 0x0, 0xD9); /* DE C0+i or DE /0*/ /* LoadF D9 /0 */
10357 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src),
10358 OpcP, RegOpc(dst) );
10359 ins_pipe( fpu_reg_mem );
10360 %}
10362 // // Following two instructions for _222_mpegaudio
10363 // Spill to obtain 24-bit precision
10364 instruct addF24_mem_reg(stackSlotF dst, regF src2, memory src1 ) %{
10365 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10366 match(Set dst (AddF src1 src2));
10368 format %{ "FADD $dst,$src1,$src2" %}
10369 opcode(0xD8, 0x0, 0xD9); /* D8 C0+i */ /* LoadF D9 /0 */
10370 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src1),
10371 OpcReg_F(src2),
10372 Pop_Mem_F(dst) );
10373 ins_pipe( fpu_mem_reg_mem );
10374 %}
10376 // Cisc-spill variant
10377 // Spill to obtain 24-bit precision
10378 instruct addF24_mem_cisc(stackSlotF dst, memory src1, memory src2) %{
10379 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10380 match(Set dst (AddF src1 (LoadF src2)));
10382 format %{ "FADD $dst,$src1,$src2 cisc" %}
10383 opcode(0xD8, 0x0, 0xD9); /* D8 C0+i */ /* LoadF D9 /0 */
10384 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10385 set_instruction_start,
10386 OpcP, RMopc_Mem(secondary,src1),
10387 Pop_Mem_F(dst) );
10388 ins_pipe( fpu_mem_mem_mem );
10389 %}
10391 // Spill to obtain 24-bit precision
10392 instruct addF24_mem_mem(stackSlotF dst, memory src1, memory src2) %{
10393 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10394 match(Set dst (AddF src1 src2));
10396 format %{ "FADD $dst,$src1,$src2" %}
10397 opcode(0xD8, 0x0, 0xD9); /* D8 /0 */ /* LoadF D9 /0 */
10398 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10399 set_instruction_start,
10400 OpcP, RMopc_Mem(secondary,src1),
10401 Pop_Mem_F(dst) );
10402 ins_pipe( fpu_mem_mem_mem );
10403 %}
10406 // Spill to obtain 24-bit precision
10407 instruct addF24_reg_imm(stackSlotF dst, regF src1, immF src2) %{
10408 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10409 match(Set dst (AddF src1 src2));
10410 format %{ "FLD $src1\n\t"
10411 "FADD $src2\n\t"
10412 "FSTP_S $dst" %}
10413 opcode(0xD8, 0x00); /* D8 /0 */
10414 ins_encode( Push_Reg_F(src1),
10415 Opc_MemImm_F(src2),
10416 Pop_Mem_F(dst));
10417 ins_pipe( fpu_mem_reg_con );
10418 %}
10419 //
10420 // This instruction does not round to 24-bits
10421 instruct addF_reg_imm(regF dst, regF src1, immF src2) %{
10422 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10423 match(Set dst (AddF src1 src2));
10424 format %{ "FLD $src1\n\t"
10425 "FADD $src2\n\t"
10426 "FSTP_S $dst" %}
10427 opcode(0xD8, 0x00); /* D8 /0 */
10428 ins_encode( Push_Reg_F(src1),
10429 Opc_MemImm_F(src2),
10430 Pop_Reg_F(dst));
10431 ins_pipe( fpu_reg_reg_con );
10432 %}
10434 // Spill to obtain 24-bit precision
10435 instruct mulF24_reg(stackSlotF dst, regF src1, regF src2) %{
10436 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10437 match(Set dst (MulF src1 src2));
10439 format %{ "FLD $src1\n\t"
10440 "FMUL $src2\n\t"
10441 "FSTP_S $dst" %}
10442 opcode(0xD8, 0x1); /* D8 C8+i or D8 /1 ;; result in TOS */
10443 ins_encode( Push_Reg_F(src1),
10444 OpcReg_F(src2),
10445 Pop_Mem_F(dst) );
10446 ins_pipe( fpu_mem_reg_reg );
10447 %}
10448 //
10449 // This instruction does not round to 24-bits
10450 instruct mulF_reg(regF dst, regF src1, regF src2) %{
10451 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10452 match(Set dst (MulF src1 src2));
10454 format %{ "FLD $src1\n\t"
10455 "FMUL $src2\n\t"
10456 "FSTP_S $dst" %}
10457 opcode(0xD8, 0x1); /* D8 C8+i */
10458 ins_encode( Push_Reg_F(src2),
10459 OpcReg_F(src1),
10460 Pop_Reg_F(dst) );
10461 ins_pipe( fpu_reg_reg_reg );
10462 %}
10465 // Spill to obtain 24-bit precision
10466 // Cisc-alternate to reg-reg multiply
10467 instruct mulF24_reg_mem(stackSlotF dst, regF src1, memory src2) %{
10468 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10469 match(Set dst (MulF src1 (LoadF src2)));
10471 format %{ "FLD_S $src2\n\t"
10472 "FMUL $src1\n\t"
10473 "FSTP_S $dst" %}
10474 opcode(0xD8, 0x1, 0xD9); /* D8 C8+i or DE /1*/ /* LoadF D9 /0 */
10475 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10476 OpcReg_F(src1),
10477 Pop_Mem_F(dst) );
10478 ins_pipe( fpu_mem_reg_mem );
10479 %}
10480 //
10481 // This instruction does not round to 24-bits
10482 // Cisc-alternate to reg-reg multiply
10483 instruct mulF_reg_mem(regF dst, regF src1, memory src2) %{
10484 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10485 match(Set dst (MulF src1 (LoadF src2)));
10487 format %{ "FMUL $dst,$src1,$src2" %}
10488 opcode(0xD8, 0x1, 0xD9); /* D8 C8+i */ /* LoadF D9 /0 */
10489 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10490 OpcReg_F(src1),
10491 Pop_Reg_F(dst) );
10492 ins_pipe( fpu_reg_reg_mem );
10493 %}
10495 // Spill to obtain 24-bit precision
10496 instruct mulF24_mem_mem(stackSlotF dst, memory src1, memory src2) %{
10497 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10498 match(Set dst (MulF src1 src2));
10500 format %{ "FMUL $dst,$src1,$src2" %}
10501 opcode(0xD8, 0x1, 0xD9); /* D8 /1 */ /* LoadF D9 /0 */
10502 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10503 set_instruction_start,
10504 OpcP, RMopc_Mem(secondary,src1),
10505 Pop_Mem_F(dst) );
10506 ins_pipe( fpu_mem_mem_mem );
10507 %}
10509 // Spill to obtain 24-bit precision
10510 instruct mulF24_reg_imm(stackSlotF dst, regF src1, immF src2) %{
10511 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10512 match(Set dst (MulF src1 src2));
10514 format %{ "FMULc $dst,$src1,$src2" %}
10515 opcode(0xD8, 0x1); /* D8 /1*/
10516 ins_encode( Push_Reg_F(src1),
10517 Opc_MemImm_F(src2),
10518 Pop_Mem_F(dst));
10519 ins_pipe( fpu_mem_reg_con );
10520 %}
10521 //
10522 // This instruction does not round to 24-bits
10523 instruct mulF_reg_imm(regF dst, regF src1, immF src2) %{
10524 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10525 match(Set dst (MulF src1 src2));
10527 format %{ "FMULc $dst. $src1, $src2" %}
10528 opcode(0xD8, 0x1); /* D8 /1*/
10529 ins_encode( Push_Reg_F(src1),
10530 Opc_MemImm_F(src2),
10531 Pop_Reg_F(dst));
10532 ins_pipe( fpu_reg_reg_con );
10533 %}
10536 //
10537 // MACRO1 -- subsume unshared load into mulF
10538 // This instruction does not round to 24-bits
10539 instruct mulF_reg_load1(regF dst, regF src, memory mem1 ) %{
10540 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10541 match(Set dst (MulF (LoadF mem1) src));
10543 format %{ "FLD $mem1 ===MACRO1===\n\t"
10544 "FMUL ST,$src\n\t"
10545 "FSTP $dst" %}
10546 opcode(0xD8, 0x1, 0xD9); /* D8 C8+i or D8 /1 */ /* LoadF D9 /0 */
10547 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,mem1),
10548 OpcReg_F(src),
10549 Pop_Reg_F(dst) );
10550 ins_pipe( fpu_reg_reg_mem );
10551 %}
10552 //
10553 // MACRO2 -- addF a mulF which subsumed an unshared load
10554 // This instruction does not round to 24-bits
10555 instruct addF_mulF_reg_load1(regF dst, memory mem1, regF src1, regF src2) %{
10556 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10557 match(Set dst (AddF (MulF (LoadF mem1) src1) src2));
10558 ins_cost(95);
10560 format %{ "FLD $mem1 ===MACRO2===\n\t"
10561 "FMUL ST,$src1 subsume mulF left load\n\t"
10562 "FADD ST,$src2\n\t"
10563 "FSTP $dst" %}
10564 opcode(0xD9); /* LoadF D9 /0 */
10565 ins_encode( OpcP, RMopc_Mem(0x00,mem1),
10566 FMul_ST_reg(src1),
10567 FAdd_ST_reg(src2),
10568 Pop_Reg_F(dst) );
10569 ins_pipe( fpu_reg_mem_reg_reg );
10570 %}
10572 // MACRO3 -- addF a mulF
10573 // This instruction does not round to 24-bits. It is a '2-address'
10574 // instruction in that the result goes back to src2. This eliminates
10575 // a move from the macro; possibly the register allocator will have
10576 // to add it back (and maybe not).
10577 instruct addF_mulF_reg(regF src2, regF src1, regF src0) %{
10578 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10579 match(Set src2 (AddF (MulF src0 src1) src2));
10581 format %{ "FLD $src0 ===MACRO3===\n\t"
10582 "FMUL ST,$src1\n\t"
10583 "FADDP $src2,ST" %}
10584 opcode(0xD9); /* LoadF D9 /0 */
10585 ins_encode( Push_Reg_F(src0),
10586 FMul_ST_reg(src1),
10587 FAddP_reg_ST(src2) );
10588 ins_pipe( fpu_reg_reg_reg );
10589 %}
10591 // MACRO4 -- divF subF
10592 // This instruction does not round to 24-bits
10593 instruct subF_divF_reg(regF dst, regF src1, regF src2, regF src3) %{
10594 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10595 match(Set dst (DivF (SubF src2 src1) src3));
10597 format %{ "FLD $src2 ===MACRO4===\n\t"
10598 "FSUB ST,$src1\n\t"
10599 "FDIV ST,$src3\n\t"
10600 "FSTP $dst" %}
10601 opcode(0xDE, 0x7); /* DE F8+i or DE /7*/
10602 ins_encode( Push_Reg_F(src2),
10603 subF_divF_encode(src1,src3),
10604 Pop_Reg_F(dst) );
10605 ins_pipe( fpu_reg_reg_reg_reg );
10606 %}
10608 // Spill to obtain 24-bit precision
10609 instruct divF24_reg(stackSlotF dst, regF src1, regF src2) %{
10610 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10611 match(Set dst (DivF src1 src2));
10613 format %{ "FDIV $dst,$src1,$src2" %}
10614 opcode(0xD8, 0x6); /* D8 F0+i or DE /6*/
10615 ins_encode( Push_Reg_F(src1),
10616 OpcReg_F(src2),
10617 Pop_Mem_F(dst) );
10618 ins_pipe( fpu_mem_reg_reg );
10619 %}
10620 //
10621 // This instruction does not round to 24-bits
10622 instruct divF_reg(regF dst, regF src) %{
10623 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10624 match(Set dst (DivF dst src));
10626 format %{ "FDIV $dst,$src" %}
10627 opcode(0xDE, 0x7); /* DE F8+i or DE /7*/
10628 ins_encode( Push_Reg_F(src),
10629 OpcP, RegOpc(dst) );
10630 ins_pipe( fpu_reg_reg );
10631 %}
10634 // Spill to obtain 24-bit precision
10635 instruct modF24_reg(stackSlotF dst, regF src1, regF src2, eAXRegI rax, eFlagsReg cr) %{
10636 predicate( UseSSE==0 && Compile::current()->select_24_bit_instr());
10637 match(Set dst (ModF src1 src2));
10638 effect(KILL rax, KILL cr); // emitModD() uses EAX and EFLAGS
10640 format %{ "FMOD $dst,$src1,$src2" %}
10641 ins_encode( Push_Reg_Mod_D(src1, src2),
10642 emitModD(),
10643 Push_Result_Mod_D(src2),
10644 Pop_Mem_F(dst));
10645 ins_pipe( pipe_slow );
10646 %}
10647 //
10648 // This instruction does not round to 24-bits
10649 instruct modF_reg(regF dst, regF src, eAXRegI rax, eFlagsReg cr) %{
10650 predicate( UseSSE==0 && !Compile::current()->select_24_bit_instr());
10651 match(Set dst (ModF dst src));
10652 effect(KILL rax, KILL cr); // emitModD() uses EAX and EFLAGS
10654 format %{ "FMOD $dst,$src" %}
10655 ins_encode(Push_Reg_Mod_D(dst, src),
10656 emitModD(),
10657 Push_Result_Mod_D(src),
10658 Pop_Reg_F(dst));
10659 ins_pipe( pipe_slow );
10660 %}
10662 instruct modX_reg(regX dst, regX src0, regX src1, eAXRegI rax, eFlagsReg cr) %{
10663 predicate(UseSSE>=1);
10664 match(Set dst (ModF src0 src1));
10665 effect(KILL rax, KILL cr);
10666 format %{ "SUB ESP,4\t # FMOD\n"
10667 "\tMOVSS [ESP+0],$src1\n"
10668 "\tFLD_S [ESP+0]\n"
10669 "\tMOVSS [ESP+0],$src0\n"
10670 "\tFLD_S [ESP+0]\n"
10671 "loop:\tFPREM\n"
10672 "\tFWAIT\n"
10673 "\tFNSTSW AX\n"
10674 "\tSAHF\n"
10675 "\tJP loop\n"
10676 "\tFSTP_S [ESP+0]\n"
10677 "\tMOVSS $dst,[ESP+0]\n"
10678 "\tADD ESP,4\n"
10679 "\tFSTP ST0\t # Restore FPU Stack"
10680 %}
10681 ins_cost(250);
10682 ins_encode( Push_ModX_encoding(src0, src1), emitModD(), Push_ResultX(dst,0x4), PopFPU);
10683 ins_pipe( pipe_slow );
10684 %}
10687 //----------Arithmetic Conversion Instructions---------------------------------
10688 // The conversions operations are all Alpha sorted. Please keep it that way!
10690 instruct roundFloat_mem_reg(stackSlotF dst, regF src) %{
10691 predicate(UseSSE==0);
10692 match(Set dst (RoundFloat src));
10693 ins_cost(125);
10694 format %{ "FST_S $dst,$src\t# F-round" %}
10695 ins_encode( Pop_Mem_Reg_F(dst, src) );
10696 ins_pipe( fpu_mem_reg );
10697 %}
10699 instruct roundDouble_mem_reg(stackSlotD dst, regD src) %{
10700 predicate(UseSSE<=1);
10701 match(Set dst (RoundDouble src));
10702 ins_cost(125);
10703 format %{ "FST_D $dst,$src\t# D-round" %}
10704 ins_encode( Pop_Mem_Reg_D(dst, src) );
10705 ins_pipe( fpu_mem_reg );
10706 %}
10708 // Force rounding to 24-bit precision and 6-bit exponent
10709 instruct convD2F_reg(stackSlotF dst, regD src) %{
10710 predicate(UseSSE==0);
10711 match(Set dst (ConvD2F src));
10712 format %{ "FST_S $dst,$src\t# F-round" %}
10713 expand %{
10714 roundFloat_mem_reg(dst,src);
10715 %}
10716 %}
10718 // Force rounding to 24-bit precision and 6-bit exponent
10719 instruct convD2X_reg(regX dst, regD src, eFlagsReg cr) %{
10720 predicate(UseSSE==1);
10721 match(Set dst (ConvD2F src));
10722 effect( KILL cr );
10723 format %{ "SUB ESP,4\n\t"
10724 "FST_S [ESP],$src\t# F-round\n\t"
10725 "MOVSS $dst,[ESP]\n\t"
10726 "ADD ESP,4" %}
10727 ins_encode( D2X_encoding(dst, src) );
10728 ins_pipe( pipe_slow );
10729 %}
10731 // Force rounding double precision to single precision
10732 instruct convXD2X_reg(regX dst, regXD src) %{
10733 predicate(UseSSE>=2);
10734 match(Set dst (ConvD2F src));
10735 format %{ "CVTSD2SS $dst,$src\t# F-round" %}
10736 opcode(0xF2, 0x0F, 0x5A);
10737 ins_encode( OpcP, OpcS, Opcode(tertiary), RegReg(dst, src));
10738 ins_pipe( pipe_slow );
10739 %}
10741 instruct convF2D_reg_reg(regD dst, regF src) %{
10742 predicate(UseSSE==0);
10743 match(Set dst (ConvF2D src));
10744 format %{ "FST_S $dst,$src\t# D-round" %}
10745 ins_encode( Pop_Reg_Reg_D(dst, src));
10746 ins_pipe( fpu_reg_reg );
10747 %}
10749 instruct convF2D_reg(stackSlotD dst, regF src) %{
10750 predicate(UseSSE==1);
10751 match(Set dst (ConvF2D src));
10752 format %{ "FST_D $dst,$src\t# D-round" %}
10753 expand %{
10754 roundDouble_mem_reg(dst,src);
10755 %}
10756 %}
10758 instruct convX2D_reg(regD dst, regX src, eFlagsReg cr) %{
10759 predicate(UseSSE==1);
10760 match(Set dst (ConvF2D src));
10761 effect( KILL cr );
10762 format %{ "SUB ESP,4\n\t"
10763 "MOVSS [ESP] $src\n\t"
10764 "FLD_S [ESP]\n\t"
10765 "ADD ESP,4\n\t"
10766 "FSTP $dst\t# D-round" %}
10767 ins_encode( X2D_encoding(dst, src), Pop_Reg_D(dst));
10768 ins_pipe( pipe_slow );
10769 %}
10771 instruct convX2XD_reg(regXD dst, regX src) %{
10772 predicate(UseSSE>=2);
10773 match(Set dst (ConvF2D src));
10774 format %{ "CVTSS2SD $dst,$src\t# D-round" %}
10775 opcode(0xF3, 0x0F, 0x5A);
10776 ins_encode( OpcP, OpcS, Opcode(tertiary), RegReg(dst, src));
10777 ins_pipe( pipe_slow );
10778 %}
10780 // Convert a double to an int. If the double is a NAN, stuff a zero in instead.
10781 instruct convD2I_reg_reg( eAXRegI dst, eDXRegI tmp, regD src, eFlagsReg cr ) %{
10782 predicate(UseSSE<=1);
10783 match(Set dst (ConvD2I src));
10784 effect( KILL tmp, KILL cr );
10785 format %{ "FLD $src\t# Convert double to int \n\t"
10786 "FLDCW trunc mode\n\t"
10787 "SUB ESP,4\n\t"
10788 "FISTp [ESP + #0]\n\t"
10789 "FLDCW std/24-bit mode\n\t"
10790 "POP EAX\n\t"
10791 "CMP EAX,0x80000000\n\t"
10792 "JNE,s fast\n\t"
10793 "FLD_D $src\n\t"
10794 "CALL d2i_wrapper\n"
10795 "fast:" %}
10796 ins_encode( Push_Reg_D(src), D2I_encoding(src) );
10797 ins_pipe( pipe_slow );
10798 %}
10800 // Convert a double to an int. If the double is a NAN, stuff a zero in instead.
10801 instruct convXD2I_reg_reg( eAXRegI dst, eDXRegI tmp, regXD src, eFlagsReg cr ) %{
10802 predicate(UseSSE>=2);
10803 match(Set dst (ConvD2I src));
10804 effect( KILL tmp, KILL cr );
10805 format %{ "CVTTSD2SI $dst, $src\n\t"
10806 "CMP $dst,0x80000000\n\t"
10807 "JNE,s fast\n\t"
10808 "SUB ESP, 8\n\t"
10809 "MOVSD [ESP], $src\n\t"
10810 "FLD_D [ESP]\n\t"
10811 "ADD ESP, 8\n\t"
10812 "CALL d2i_wrapper\n"
10813 "fast:" %}
10814 opcode(0x1); // double-precision conversion
10815 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x2C), FX2I_encoding(src,dst));
10816 ins_pipe( pipe_slow );
10817 %}
10819 instruct convD2L_reg_reg( eADXRegL dst, regD src, eFlagsReg cr ) %{
10820 predicate(UseSSE<=1);
10821 match(Set dst (ConvD2L src));
10822 effect( KILL cr );
10823 format %{ "FLD $src\t# Convert double to long\n\t"
10824 "FLDCW trunc mode\n\t"
10825 "SUB ESP,8\n\t"
10826 "FISTp [ESP + #0]\n\t"
10827 "FLDCW std/24-bit mode\n\t"
10828 "POP EAX\n\t"
10829 "POP EDX\n\t"
10830 "CMP EDX,0x80000000\n\t"
10831 "JNE,s fast\n\t"
10832 "TEST EAX,EAX\n\t"
10833 "JNE,s fast\n\t"
10834 "FLD $src\n\t"
10835 "CALL d2l_wrapper\n"
10836 "fast:" %}
10837 ins_encode( Push_Reg_D(src), D2L_encoding(src) );
10838 ins_pipe( pipe_slow );
10839 %}
10841 // XMM lacks a float/double->long conversion, so use the old FPU stack.
10842 instruct convXD2L_reg_reg( eADXRegL dst, regXD src, eFlagsReg cr ) %{
10843 predicate (UseSSE>=2);
10844 match(Set dst (ConvD2L src));
10845 effect( KILL cr );
10846 format %{ "SUB ESP,8\t# Convert double to long\n\t"
10847 "MOVSD [ESP],$src\n\t"
10848 "FLD_D [ESP]\n\t"
10849 "FLDCW trunc mode\n\t"
10850 "FISTp [ESP + #0]\n\t"
10851 "FLDCW std/24-bit mode\n\t"
10852 "POP EAX\n\t"
10853 "POP EDX\n\t"
10854 "CMP EDX,0x80000000\n\t"
10855 "JNE,s fast\n\t"
10856 "TEST EAX,EAX\n\t"
10857 "JNE,s fast\n\t"
10858 "SUB ESP,8\n\t"
10859 "MOVSD [ESP],$src\n\t"
10860 "FLD_D [ESP]\n\t"
10861 "CALL d2l_wrapper\n"
10862 "fast:" %}
10863 ins_encode( XD2L_encoding(src) );
10864 ins_pipe( pipe_slow );
10865 %}
10867 // Convert a double to an int. Java semantics require we do complex
10868 // manglations in the corner cases. So we set the rounding mode to
10869 // 'zero', store the darned double down as an int, and reset the
10870 // rounding mode to 'nearest'. The hardware stores a flag value down
10871 // if we would overflow or converted a NAN; we check for this and
10872 // and go the slow path if needed.
10873 instruct convF2I_reg_reg(eAXRegI dst, eDXRegI tmp, regF src, eFlagsReg cr ) %{
10874 predicate(UseSSE==0);
10875 match(Set dst (ConvF2I src));
10876 effect( KILL tmp, KILL cr );
10877 format %{ "FLD $src\t# Convert float to int \n\t"
10878 "FLDCW trunc mode\n\t"
10879 "SUB ESP,4\n\t"
10880 "FISTp [ESP + #0]\n\t"
10881 "FLDCW std/24-bit mode\n\t"
10882 "POP EAX\n\t"
10883 "CMP EAX,0x80000000\n\t"
10884 "JNE,s fast\n\t"
10885 "FLD $src\n\t"
10886 "CALL d2i_wrapper\n"
10887 "fast:" %}
10888 // D2I_encoding works for F2I
10889 ins_encode( Push_Reg_F(src), D2I_encoding(src) );
10890 ins_pipe( pipe_slow );
10891 %}
10893 // Convert a float in xmm to an int reg.
10894 instruct convX2I_reg(eAXRegI dst, eDXRegI tmp, regX src, eFlagsReg cr ) %{
10895 predicate(UseSSE>=1);
10896 match(Set dst (ConvF2I src));
10897 effect( KILL tmp, KILL cr );
10898 format %{ "CVTTSS2SI $dst, $src\n\t"
10899 "CMP $dst,0x80000000\n\t"
10900 "JNE,s fast\n\t"
10901 "SUB ESP, 4\n\t"
10902 "MOVSS [ESP], $src\n\t"
10903 "FLD [ESP]\n\t"
10904 "ADD ESP, 4\n\t"
10905 "CALL d2i_wrapper\n"
10906 "fast:" %}
10907 opcode(0x0); // single-precision conversion
10908 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x2C), FX2I_encoding(src,dst));
10909 ins_pipe( pipe_slow );
10910 %}
10912 instruct convF2L_reg_reg( eADXRegL dst, regF src, eFlagsReg cr ) %{
10913 predicate(UseSSE==0);
10914 match(Set dst (ConvF2L src));
10915 effect( KILL cr );
10916 format %{ "FLD $src\t# Convert float to long\n\t"
10917 "FLDCW trunc mode\n\t"
10918 "SUB ESP,8\n\t"
10919 "FISTp [ESP + #0]\n\t"
10920 "FLDCW std/24-bit mode\n\t"
10921 "POP EAX\n\t"
10922 "POP EDX\n\t"
10923 "CMP EDX,0x80000000\n\t"
10924 "JNE,s fast\n\t"
10925 "TEST EAX,EAX\n\t"
10926 "JNE,s fast\n\t"
10927 "FLD $src\n\t"
10928 "CALL d2l_wrapper\n"
10929 "fast:" %}
10930 // D2L_encoding works for F2L
10931 ins_encode( Push_Reg_F(src), D2L_encoding(src) );
10932 ins_pipe( pipe_slow );
10933 %}
10935 // XMM lacks a float/double->long conversion, so use the old FPU stack.
10936 instruct convX2L_reg_reg( eADXRegL dst, regX src, eFlagsReg cr ) %{
10937 predicate (UseSSE>=1);
10938 match(Set dst (ConvF2L src));
10939 effect( KILL cr );
10940 format %{ "SUB ESP,8\t# Convert float to long\n\t"
10941 "MOVSS [ESP],$src\n\t"
10942 "FLD_S [ESP]\n\t"
10943 "FLDCW trunc mode\n\t"
10944 "FISTp [ESP + #0]\n\t"
10945 "FLDCW std/24-bit mode\n\t"
10946 "POP EAX\n\t"
10947 "POP EDX\n\t"
10948 "CMP EDX,0x80000000\n\t"
10949 "JNE,s fast\n\t"
10950 "TEST EAX,EAX\n\t"
10951 "JNE,s fast\n\t"
10952 "SUB ESP,4\t# Convert float to long\n\t"
10953 "MOVSS [ESP],$src\n\t"
10954 "FLD_S [ESP]\n\t"
10955 "ADD ESP,4\n\t"
10956 "CALL d2l_wrapper\n"
10957 "fast:" %}
10958 ins_encode( X2L_encoding(src) );
10959 ins_pipe( pipe_slow );
10960 %}
10962 instruct convI2D_reg(regD dst, stackSlotI src) %{
10963 predicate( UseSSE<=1 );
10964 match(Set dst (ConvI2D src));
10965 format %{ "FILD $src\n\t"
10966 "FSTP $dst" %}
10967 opcode(0xDB, 0x0); /* DB /0 */
10968 ins_encode(Push_Mem_I(src), Pop_Reg_D(dst));
10969 ins_pipe( fpu_reg_mem );
10970 %}
10972 instruct convI2XD_reg(regXD dst, eRegI src) %{
10973 predicate( UseSSE>=2 && !UseXmmI2D );
10974 match(Set dst (ConvI2D src));
10975 format %{ "CVTSI2SD $dst,$src" %}
10976 opcode(0xF2, 0x0F, 0x2A);
10977 ins_encode( OpcP, OpcS, Opcode(tertiary), RegReg(dst, src));
10978 ins_pipe( pipe_slow );
10979 %}
10981 instruct convI2XD_mem(regXD dst, memory mem) %{
10982 predicate( UseSSE>=2 );
10983 match(Set dst (ConvI2D (LoadI mem)));
10984 format %{ "CVTSI2SD $dst,$mem" %}
10985 opcode(0xF2, 0x0F, 0x2A);
10986 ins_encode( OpcP, OpcS, Opcode(tertiary), RegMem(dst, mem));
10987 ins_pipe( pipe_slow );
10988 %}
10990 instruct convXI2XD_reg(regXD dst, eRegI src)
10991 %{
10992 predicate( UseSSE>=2 && UseXmmI2D );
10993 match(Set dst (ConvI2D src));
10995 format %{ "MOVD $dst,$src\n\t"
10996 "CVTDQ2PD $dst,$dst\t# i2d" %}
10997 ins_encode %{
10998 __ movd($dst$$XMMRegister, $src$$Register);
10999 __ cvtdq2pd($dst$$XMMRegister, $dst$$XMMRegister);
11000 %}
11001 ins_pipe(pipe_slow); // XXX
11002 %}
11004 instruct convI2D_mem(regD dst, memory mem) %{
11005 predicate( UseSSE<=1 && !Compile::current()->select_24_bit_instr());
11006 match(Set dst (ConvI2D (LoadI mem)));
11007 format %{ "FILD $mem\n\t"
11008 "FSTP $dst" %}
11009 opcode(0xDB); /* DB /0 */
11010 ins_encode( OpcP, RMopc_Mem(0x00,mem),
11011 Pop_Reg_D(dst));
11012 ins_pipe( fpu_reg_mem );
11013 %}
11015 // Convert a byte to a float; no rounding step needed.
11016 instruct conv24I2F_reg(regF dst, stackSlotI src) %{
11017 predicate( UseSSE==0 && n->in(1)->Opcode() == Op_AndI && n->in(1)->in(2)->is_Con() && n->in(1)->in(2)->get_int() == 255 );
11018 match(Set dst (ConvI2F src));
11019 format %{ "FILD $src\n\t"
11020 "FSTP $dst" %}
11022 opcode(0xDB, 0x0); /* DB /0 */
11023 ins_encode(Push_Mem_I(src), Pop_Reg_F(dst));
11024 ins_pipe( fpu_reg_mem );
11025 %}
11027 // In 24-bit mode, force exponent rounding by storing back out
11028 instruct convI2F_SSF(stackSlotF dst, stackSlotI src) %{
11029 predicate( UseSSE==0 && Compile::current()->select_24_bit_instr());
11030 match(Set dst (ConvI2F src));
11031 ins_cost(200);
11032 format %{ "FILD $src\n\t"
11033 "FSTP_S $dst" %}
11034 opcode(0xDB, 0x0); /* DB /0 */
11035 ins_encode( Push_Mem_I(src),
11036 Pop_Mem_F(dst));
11037 ins_pipe( fpu_mem_mem );
11038 %}
11040 // In 24-bit mode, force exponent rounding by storing back out
11041 instruct convI2F_SSF_mem(stackSlotF dst, memory mem) %{
11042 predicate( UseSSE==0 && Compile::current()->select_24_bit_instr());
11043 match(Set dst (ConvI2F (LoadI mem)));
11044 ins_cost(200);
11045 format %{ "FILD $mem\n\t"
11046 "FSTP_S $dst" %}
11047 opcode(0xDB); /* DB /0 */
11048 ins_encode( OpcP, RMopc_Mem(0x00,mem),
11049 Pop_Mem_F(dst));
11050 ins_pipe( fpu_mem_mem );
11051 %}
11053 // This instruction does not round to 24-bits
11054 instruct convI2F_reg(regF dst, stackSlotI src) %{
11055 predicate( UseSSE==0 && !Compile::current()->select_24_bit_instr());
11056 match(Set dst (ConvI2F src));
11057 format %{ "FILD $src\n\t"
11058 "FSTP $dst" %}
11059 opcode(0xDB, 0x0); /* DB /0 */
11060 ins_encode( Push_Mem_I(src),
11061 Pop_Reg_F(dst));
11062 ins_pipe( fpu_reg_mem );
11063 %}
11065 // This instruction does not round to 24-bits
11066 instruct convI2F_mem(regF dst, memory mem) %{
11067 predicate( UseSSE==0 && !Compile::current()->select_24_bit_instr());
11068 match(Set dst (ConvI2F (LoadI mem)));
11069 format %{ "FILD $mem\n\t"
11070 "FSTP $dst" %}
11071 opcode(0xDB); /* DB /0 */
11072 ins_encode( OpcP, RMopc_Mem(0x00,mem),
11073 Pop_Reg_F(dst));
11074 ins_pipe( fpu_reg_mem );
11075 %}
11077 // Convert an int to a float in xmm; no rounding step needed.
11078 instruct convI2X_reg(regX dst, eRegI src) %{
11079 predicate( UseSSE==1 || UseSSE>=2 && !UseXmmI2F );
11080 match(Set dst (ConvI2F src));
11081 format %{ "CVTSI2SS $dst, $src" %}
11083 opcode(0xF3, 0x0F, 0x2A); /* F3 0F 2A /r */
11084 ins_encode( OpcP, OpcS, Opcode(tertiary), RegReg(dst, src));
11085 ins_pipe( pipe_slow );
11086 %}
11088 instruct convXI2X_reg(regX dst, eRegI src)
11089 %{
11090 predicate( UseSSE>=2 && UseXmmI2F );
11091 match(Set dst (ConvI2F src));
11093 format %{ "MOVD $dst,$src\n\t"
11094 "CVTDQ2PS $dst,$dst\t# i2f" %}
11095 ins_encode %{
11096 __ movd($dst$$XMMRegister, $src$$Register);
11097 __ cvtdq2ps($dst$$XMMRegister, $dst$$XMMRegister);
11098 %}
11099 ins_pipe(pipe_slow); // XXX
11100 %}
11102 instruct convI2L_reg( eRegL dst, eRegI src, eFlagsReg cr) %{
11103 match(Set dst (ConvI2L src));
11104 effect(KILL cr);
11105 format %{ "MOV $dst.lo,$src\n\t"
11106 "MOV $dst.hi,$src\n\t"
11107 "SAR $dst.hi,31" %}
11108 ins_encode(convert_int_long(dst,src));
11109 ins_pipe( ialu_reg_reg_long );
11110 %}
11112 // Zero-extend convert int to long
11113 instruct convI2L_reg_zex(eRegL dst, eRegI src, immL_32bits mask, eFlagsReg flags ) %{
11114 match(Set dst (AndL (ConvI2L src) mask) );
11115 effect( KILL flags );
11116 format %{ "MOV $dst.lo,$src\n\t"
11117 "XOR $dst.hi,$dst.hi" %}
11118 opcode(0x33); // XOR
11119 ins_encode(enc_Copy(dst,src), OpcP, RegReg_Hi2(dst,dst) );
11120 ins_pipe( ialu_reg_reg_long );
11121 %}
11123 // Zero-extend long
11124 instruct zerox_long(eRegL dst, eRegL src, immL_32bits mask, eFlagsReg flags ) %{
11125 match(Set dst (AndL src mask) );
11126 effect( KILL flags );
11127 format %{ "MOV $dst.lo,$src.lo\n\t"
11128 "XOR $dst.hi,$dst.hi\n\t" %}
11129 opcode(0x33); // XOR
11130 ins_encode(enc_Copy(dst,src), OpcP, RegReg_Hi2(dst,dst) );
11131 ins_pipe( ialu_reg_reg_long );
11132 %}
11134 instruct convL2D_reg( stackSlotD dst, eRegL src, eFlagsReg cr) %{
11135 predicate (UseSSE<=1);
11136 match(Set dst (ConvL2D src));
11137 effect( KILL cr );
11138 format %{ "PUSH $src.hi\t# Convert long to double\n\t"
11139 "PUSH $src.lo\n\t"
11140 "FILD ST,[ESP + #0]\n\t"
11141 "ADD ESP,8\n\t"
11142 "FSTP_D $dst\t# D-round" %}
11143 opcode(0xDF, 0x5); /* DF /5 */
11144 ins_encode(convert_long_double(src), Pop_Mem_D(dst));
11145 ins_pipe( pipe_slow );
11146 %}
11148 instruct convL2XD_reg( regXD dst, eRegL src, eFlagsReg cr) %{
11149 predicate (UseSSE>=2);
11150 match(Set dst (ConvL2D src));
11151 effect( KILL cr );
11152 format %{ "PUSH $src.hi\t# Convert long to double\n\t"
11153 "PUSH $src.lo\n\t"
11154 "FILD_D [ESP]\n\t"
11155 "FSTP_D [ESP]\n\t"
11156 "MOVSD $dst,[ESP]\n\t"
11157 "ADD ESP,8" %}
11158 opcode(0xDF, 0x5); /* DF /5 */
11159 ins_encode(convert_long_double2(src), Push_ResultXD(dst));
11160 ins_pipe( pipe_slow );
11161 %}
11163 instruct convL2X_reg( regX dst, eRegL src, eFlagsReg cr) %{
11164 predicate (UseSSE>=1);
11165 match(Set dst (ConvL2F src));
11166 effect( KILL cr );
11167 format %{ "PUSH $src.hi\t# Convert long to single float\n\t"
11168 "PUSH $src.lo\n\t"
11169 "FILD_D [ESP]\n\t"
11170 "FSTP_S [ESP]\n\t"
11171 "MOVSS $dst,[ESP]\n\t"
11172 "ADD ESP,8" %}
11173 opcode(0xDF, 0x5); /* DF /5 */
11174 ins_encode(convert_long_double2(src), Push_ResultX(dst,0x8));
11175 ins_pipe( pipe_slow );
11176 %}
11178 instruct convL2F_reg( stackSlotF dst, eRegL src, eFlagsReg cr) %{
11179 match(Set dst (ConvL2F src));
11180 effect( KILL cr );
11181 format %{ "PUSH $src.hi\t# Convert long to single float\n\t"
11182 "PUSH $src.lo\n\t"
11183 "FILD ST,[ESP + #0]\n\t"
11184 "ADD ESP,8\n\t"
11185 "FSTP_S $dst\t# F-round" %}
11186 opcode(0xDF, 0x5); /* DF /5 */
11187 ins_encode(convert_long_double(src), Pop_Mem_F(dst));
11188 ins_pipe( pipe_slow );
11189 %}
11191 instruct convL2I_reg( eRegI dst, eRegL src ) %{
11192 match(Set dst (ConvL2I src));
11193 effect( DEF dst, USE src );
11194 format %{ "MOV $dst,$src.lo" %}
11195 ins_encode(enc_CopyL_Lo(dst,src));
11196 ins_pipe( ialu_reg_reg );
11197 %}
11200 instruct MoveF2I_stack_reg(eRegI dst, stackSlotF src) %{
11201 match(Set dst (MoveF2I src));
11202 effect( DEF dst, USE src );
11203 ins_cost(100);
11204 format %{ "MOV $dst,$src\t# MoveF2I_stack_reg" %}
11205 opcode(0x8B);
11206 ins_encode( OpcP, RegMem(dst,src));
11207 ins_pipe( ialu_reg_mem );
11208 %}
11210 instruct MoveF2I_reg_stack(stackSlotI dst, regF src) %{
11211 predicate(UseSSE==0);
11212 match(Set dst (MoveF2I src));
11213 effect( DEF dst, USE src );
11215 ins_cost(125);
11216 format %{ "FST_S $dst,$src\t# MoveF2I_reg_stack" %}
11217 ins_encode( Pop_Mem_Reg_F(dst, src) );
11218 ins_pipe( fpu_mem_reg );
11219 %}
11221 instruct MoveF2I_reg_stack_sse(stackSlotI dst, regX src) %{
11222 predicate(UseSSE>=1);
11223 match(Set dst (MoveF2I src));
11224 effect( DEF dst, USE src );
11226 ins_cost(95);
11227 format %{ "MOVSS $dst,$src\t# MoveF2I_reg_stack_sse" %}
11228 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x11), RegMem(src, dst));
11229 ins_pipe( pipe_slow );
11230 %}
11232 instruct MoveF2I_reg_reg_sse(eRegI dst, regX src) %{
11233 predicate(UseSSE>=2);
11234 match(Set dst (MoveF2I src));
11235 effect( DEF dst, USE src );
11236 ins_cost(85);
11237 format %{ "MOVD $dst,$src\t# MoveF2I_reg_reg_sse" %}
11238 ins_encode( MovX2I_reg(dst, src));
11239 ins_pipe( pipe_slow );
11240 %}
11242 instruct MoveI2F_reg_stack(stackSlotF dst, eRegI src) %{
11243 match(Set dst (MoveI2F src));
11244 effect( DEF dst, USE src );
11246 ins_cost(100);
11247 format %{ "MOV $dst,$src\t# MoveI2F_reg_stack" %}
11248 opcode(0x89);
11249 ins_encode( OpcPRegSS( dst, src ) );
11250 ins_pipe( ialu_mem_reg );
11251 %}
11254 instruct MoveI2F_stack_reg(regF dst, stackSlotI src) %{
11255 predicate(UseSSE==0);
11256 match(Set dst (MoveI2F src));
11257 effect(DEF dst, USE src);
11259 ins_cost(125);
11260 format %{ "FLD_S $src\n\t"
11261 "FSTP $dst\t# MoveI2F_stack_reg" %}
11262 opcode(0xD9); /* D9 /0, FLD m32real */
11263 ins_encode( OpcP, RMopc_Mem_no_oop(0x00,src),
11264 Pop_Reg_F(dst) );
11265 ins_pipe( fpu_reg_mem );
11266 %}
11268 instruct MoveI2F_stack_reg_sse(regX dst, stackSlotI src) %{
11269 predicate(UseSSE>=1);
11270 match(Set dst (MoveI2F src));
11271 effect( DEF dst, USE src );
11273 ins_cost(95);
11274 format %{ "MOVSS $dst,$src\t# MoveI2F_stack_reg_sse" %}
11275 ins_encode( Opcode(0xF3), Opcode(0x0F), Opcode(0x10), RegMem(dst,src));
11276 ins_pipe( pipe_slow );
11277 %}
11279 instruct MoveI2F_reg_reg_sse(regX dst, eRegI src) %{
11280 predicate(UseSSE>=2);
11281 match(Set dst (MoveI2F src));
11282 effect( DEF dst, USE src );
11284 ins_cost(85);
11285 format %{ "MOVD $dst,$src\t# MoveI2F_reg_reg_sse" %}
11286 ins_encode( MovI2X_reg(dst, src) );
11287 ins_pipe( pipe_slow );
11288 %}
11290 instruct MoveD2L_stack_reg(eRegL dst, stackSlotD src) %{
11291 match(Set dst (MoveD2L src));
11292 effect(DEF dst, USE src);
11294 ins_cost(250);
11295 format %{ "MOV $dst.lo,$src\n\t"
11296 "MOV $dst.hi,$src+4\t# MoveD2L_stack_reg" %}
11297 opcode(0x8B, 0x8B);
11298 ins_encode( OpcP, RegMem(dst,src), OpcS, RegMem_Hi(dst,src));
11299 ins_pipe( ialu_mem_long_reg );
11300 %}
11302 instruct MoveD2L_reg_stack(stackSlotL dst, regD src) %{
11303 predicate(UseSSE<=1);
11304 match(Set dst (MoveD2L src));
11305 effect(DEF dst, USE src);
11307 ins_cost(125);
11308 format %{ "FST_D $dst,$src\t# MoveD2L_reg_stack" %}
11309 ins_encode( Pop_Mem_Reg_D(dst, src) );
11310 ins_pipe( fpu_mem_reg );
11311 %}
11313 instruct MoveD2L_reg_stack_sse(stackSlotL dst, regXD src) %{
11314 predicate(UseSSE>=2);
11315 match(Set dst (MoveD2L src));
11316 effect(DEF dst, USE src);
11317 ins_cost(95);
11319 format %{ "MOVSD $dst,$src\t# MoveD2L_reg_stack_sse" %}
11320 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x11), RegMem(src,dst));
11321 ins_pipe( pipe_slow );
11322 %}
11324 instruct MoveD2L_reg_reg_sse(eRegL dst, regXD src, regXD tmp) %{
11325 predicate(UseSSE>=2);
11326 match(Set dst (MoveD2L src));
11327 effect(DEF dst, USE src, TEMP tmp);
11328 ins_cost(85);
11329 format %{ "MOVD $dst.lo,$src\n\t"
11330 "PSHUFLW $tmp,$src,0x4E\n\t"
11331 "MOVD $dst.hi,$tmp\t# MoveD2L_reg_reg_sse" %}
11332 ins_encode( MovXD2L_reg(dst, src, tmp) );
11333 ins_pipe( pipe_slow );
11334 %}
11336 instruct MoveL2D_reg_stack(stackSlotD dst, eRegL src) %{
11337 match(Set dst (MoveL2D src));
11338 effect(DEF dst, USE src);
11340 ins_cost(200);
11341 format %{ "MOV $dst,$src.lo\n\t"
11342 "MOV $dst+4,$src.hi\t# MoveL2D_reg_stack" %}
11343 opcode(0x89, 0x89);
11344 ins_encode( OpcP, RegMem( src, dst ), OpcS, RegMem_Hi( src, dst ) );
11345 ins_pipe( ialu_mem_long_reg );
11346 %}
11349 instruct MoveL2D_stack_reg(regD dst, stackSlotL src) %{
11350 predicate(UseSSE<=1);
11351 match(Set dst (MoveL2D src));
11352 effect(DEF dst, USE src);
11353 ins_cost(125);
11355 format %{ "FLD_D $src\n\t"
11356 "FSTP $dst\t# MoveL2D_stack_reg" %}
11357 opcode(0xDD); /* DD /0, FLD m64real */
11358 ins_encode( OpcP, RMopc_Mem_no_oop(0x00,src),
11359 Pop_Reg_D(dst) );
11360 ins_pipe( fpu_reg_mem );
11361 %}
11364 instruct MoveL2D_stack_reg_sse(regXD dst, stackSlotL src) %{
11365 predicate(UseSSE>=2 && UseXmmLoadAndClearUpper);
11366 match(Set dst (MoveL2D src));
11367 effect(DEF dst, USE src);
11369 ins_cost(95);
11370 format %{ "MOVSD $dst,$src\t# MoveL2D_stack_reg_sse" %}
11371 ins_encode( Opcode(0xF2), Opcode(0x0F), Opcode(0x10), RegMem(dst,src));
11372 ins_pipe( pipe_slow );
11373 %}
11375 instruct MoveL2D_stack_reg_sse_partial(regXD dst, stackSlotL src) %{
11376 predicate(UseSSE>=2 && !UseXmmLoadAndClearUpper);
11377 match(Set dst (MoveL2D src));
11378 effect(DEF dst, USE src);
11380 ins_cost(95);
11381 format %{ "MOVLPD $dst,$src\t# MoveL2D_stack_reg_sse" %}
11382 ins_encode( Opcode(0x66), Opcode(0x0F), Opcode(0x12), RegMem(dst,src));
11383 ins_pipe( pipe_slow );
11384 %}
11386 instruct MoveL2D_reg_reg_sse(regXD dst, eRegL src, regXD tmp) %{
11387 predicate(UseSSE>=2);
11388 match(Set dst (MoveL2D src));
11389 effect(TEMP dst, USE src, TEMP tmp);
11390 ins_cost(85);
11391 format %{ "MOVD $dst,$src.lo\n\t"
11392 "MOVD $tmp,$src.hi\n\t"
11393 "PUNPCKLDQ $dst,$tmp\t# MoveL2D_reg_reg_sse" %}
11394 ins_encode( MovL2XD_reg(dst, src, tmp) );
11395 ins_pipe( pipe_slow );
11396 %}
11398 // Replicate scalar to packed byte (1 byte) values in xmm
11399 instruct Repl8B_reg(regXD dst, regXD src) %{
11400 predicate(UseSSE>=2);
11401 match(Set dst (Replicate8B src));
11402 format %{ "MOVDQA $dst,$src\n\t"
11403 "PUNPCKLBW $dst,$dst\n\t"
11404 "PSHUFLW $dst,$dst,0x00\t! replicate8B" %}
11405 ins_encode( pshufd_8x8(dst, src));
11406 ins_pipe( pipe_slow );
11407 %}
11409 // Replicate scalar to packed byte (1 byte) values in xmm
11410 instruct Repl8B_eRegI(regXD dst, eRegI src) %{
11411 predicate(UseSSE>=2);
11412 match(Set dst (Replicate8B src));
11413 format %{ "MOVD $dst,$src\n\t"
11414 "PUNPCKLBW $dst,$dst\n\t"
11415 "PSHUFLW $dst,$dst,0x00\t! replicate8B" %}
11416 ins_encode( mov_i2x(dst, src), pshufd_8x8(dst, dst));
11417 ins_pipe( pipe_slow );
11418 %}
11420 // Replicate scalar zero to packed byte (1 byte) values in xmm
11421 instruct Repl8B_immI0(regXD dst, immI0 zero) %{
11422 predicate(UseSSE>=2);
11423 match(Set dst (Replicate8B zero));
11424 format %{ "PXOR $dst,$dst\t! replicate8B" %}
11425 ins_encode( pxor(dst, dst));
11426 ins_pipe( fpu_reg_reg );
11427 %}
11429 // Replicate scalar to packed shore (2 byte) values in xmm
11430 instruct Repl4S_reg(regXD dst, regXD src) %{
11431 predicate(UseSSE>=2);
11432 match(Set dst (Replicate4S src));
11433 format %{ "PSHUFLW $dst,$src,0x00\t! replicate4S" %}
11434 ins_encode( pshufd_4x16(dst, src));
11435 ins_pipe( fpu_reg_reg );
11436 %}
11438 // Replicate scalar to packed shore (2 byte) values in xmm
11439 instruct Repl4S_eRegI(regXD dst, eRegI src) %{
11440 predicate(UseSSE>=2);
11441 match(Set dst (Replicate4S src));
11442 format %{ "MOVD $dst,$src\n\t"
11443 "PSHUFLW $dst,$dst,0x00\t! replicate4S" %}
11444 ins_encode( mov_i2x(dst, src), pshufd_4x16(dst, dst));
11445 ins_pipe( fpu_reg_reg );
11446 %}
11448 // Replicate scalar zero to packed short (2 byte) values in xmm
11449 instruct Repl4S_immI0(regXD dst, immI0 zero) %{
11450 predicate(UseSSE>=2);
11451 match(Set dst (Replicate4S zero));
11452 format %{ "PXOR $dst,$dst\t! replicate4S" %}
11453 ins_encode( pxor(dst, dst));
11454 ins_pipe( fpu_reg_reg );
11455 %}
11457 // Replicate scalar to packed char (2 byte) values in xmm
11458 instruct Repl4C_reg(regXD dst, regXD src) %{
11459 predicate(UseSSE>=2);
11460 match(Set dst (Replicate4C src));
11461 format %{ "PSHUFLW $dst,$src,0x00\t! replicate4C" %}
11462 ins_encode( pshufd_4x16(dst, src));
11463 ins_pipe( fpu_reg_reg );
11464 %}
11466 // Replicate scalar to packed char (2 byte) values in xmm
11467 instruct Repl4C_eRegI(regXD dst, eRegI src) %{
11468 predicate(UseSSE>=2);
11469 match(Set dst (Replicate4C src));
11470 format %{ "MOVD $dst,$src\n\t"
11471 "PSHUFLW $dst,$dst,0x00\t! replicate4C" %}
11472 ins_encode( mov_i2x(dst, src), pshufd_4x16(dst, dst));
11473 ins_pipe( fpu_reg_reg );
11474 %}
11476 // Replicate scalar zero to packed char (2 byte) values in xmm
11477 instruct Repl4C_immI0(regXD dst, immI0 zero) %{
11478 predicate(UseSSE>=2);
11479 match(Set dst (Replicate4C zero));
11480 format %{ "PXOR $dst,$dst\t! replicate4C" %}
11481 ins_encode( pxor(dst, dst));
11482 ins_pipe( fpu_reg_reg );
11483 %}
11485 // Replicate scalar to packed integer (4 byte) values in xmm
11486 instruct Repl2I_reg(regXD dst, regXD src) %{
11487 predicate(UseSSE>=2);
11488 match(Set dst (Replicate2I src));
11489 format %{ "PSHUFD $dst,$src,0x00\t! replicate2I" %}
11490 ins_encode( pshufd(dst, src, 0x00));
11491 ins_pipe( fpu_reg_reg );
11492 %}
11494 // Replicate scalar to packed integer (4 byte) values in xmm
11495 instruct Repl2I_eRegI(regXD dst, eRegI src) %{
11496 predicate(UseSSE>=2);
11497 match(Set dst (Replicate2I src));
11498 format %{ "MOVD $dst,$src\n\t"
11499 "PSHUFD $dst,$dst,0x00\t! replicate2I" %}
11500 ins_encode( mov_i2x(dst, src), pshufd(dst, dst, 0x00));
11501 ins_pipe( fpu_reg_reg );
11502 %}
11504 // Replicate scalar zero to packed integer (2 byte) values in xmm
11505 instruct Repl2I_immI0(regXD dst, immI0 zero) %{
11506 predicate(UseSSE>=2);
11507 match(Set dst (Replicate2I zero));
11508 format %{ "PXOR $dst,$dst\t! replicate2I" %}
11509 ins_encode( pxor(dst, dst));
11510 ins_pipe( fpu_reg_reg );
11511 %}
11513 // Replicate scalar to packed single precision floating point values in xmm
11514 instruct Repl2F_reg(regXD dst, regXD src) %{
11515 predicate(UseSSE>=2);
11516 match(Set dst (Replicate2F src));
11517 format %{ "PSHUFD $dst,$src,0xe0\t! replicate2F" %}
11518 ins_encode( pshufd(dst, src, 0xe0));
11519 ins_pipe( fpu_reg_reg );
11520 %}
11522 // Replicate scalar to packed single precision floating point values in xmm
11523 instruct Repl2F_regX(regXD dst, regX src) %{
11524 predicate(UseSSE>=2);
11525 match(Set dst (Replicate2F src));
11526 format %{ "PSHUFD $dst,$src,0xe0\t! replicate2F" %}
11527 ins_encode( pshufd(dst, src, 0xe0));
11528 ins_pipe( fpu_reg_reg );
11529 %}
11531 // Replicate scalar to packed single precision floating point values in xmm
11532 instruct Repl2F_immXF0(regXD dst, immXF0 zero) %{
11533 predicate(UseSSE>=2);
11534 match(Set dst (Replicate2F zero));
11535 format %{ "PXOR $dst,$dst\t! replicate2F" %}
11536 ins_encode( pxor(dst, dst));
11537 ins_pipe( fpu_reg_reg );
11538 %}
11542 // =======================================================================
11543 // fast clearing of an array
11545 instruct rep_stos(eCXRegI cnt, eDIRegP base, eAXRegI zero, Universe dummy, eFlagsReg cr) %{
11546 match(Set dummy (ClearArray cnt base));
11547 effect(USE_KILL cnt, USE_KILL base, KILL zero, KILL cr);
11548 format %{ "SHL ECX,1\t# Convert doublewords to words\n\t"
11549 "XOR EAX,EAX\n\t"
11550 "REP STOS\t# store EAX into [EDI++] while ECX--" %}
11551 opcode(0,0x4);
11552 ins_encode( Opcode(0xD1), RegOpc(ECX),
11553 OpcRegReg(0x33,EAX,EAX),
11554 Opcode(0xF3), Opcode(0xAB) );
11555 ins_pipe( pipe_slow );
11556 %}
11558 instruct string_compare(eDIRegP str1, eSIRegP str2, eAXRegI tmp1, eBXRegI tmp2, eCXRegI result, eFlagsReg cr) %{
11559 match(Set result (StrComp str1 str2));
11560 effect(USE_KILL str1, USE_KILL str2, KILL tmp1, KILL tmp2, KILL cr);
11561 //ins_cost(300);
11563 format %{ "String Compare $str1,$str2 -> $result // KILL EAX, EBX" %}
11564 ins_encode( enc_String_Compare() );
11565 ins_pipe( pipe_slow );
11566 %}
11568 //----------Control Flow Instructions------------------------------------------
11569 // Signed compare Instructions
11570 instruct compI_eReg(eFlagsReg cr, eRegI op1, eRegI op2) %{
11571 match(Set cr (CmpI op1 op2));
11572 effect( DEF cr, USE op1, USE op2 );
11573 format %{ "CMP $op1,$op2" %}
11574 opcode(0x3B); /* Opcode 3B /r */
11575 ins_encode( OpcP, RegReg( op1, op2) );
11576 ins_pipe( ialu_cr_reg_reg );
11577 %}
11579 instruct compI_eReg_imm(eFlagsReg cr, eRegI op1, immI op2) %{
11580 match(Set cr (CmpI op1 op2));
11581 effect( DEF cr, USE op1 );
11582 format %{ "CMP $op1,$op2" %}
11583 opcode(0x81,0x07); /* Opcode 81 /7 */
11584 // ins_encode( RegImm( op1, op2) ); /* Was CmpImm */
11585 ins_encode( OpcSErm( op1, op2 ), Con8or32( op2 ) );
11586 ins_pipe( ialu_cr_reg_imm );
11587 %}
11589 // Cisc-spilled version of cmpI_eReg
11590 instruct compI_eReg_mem(eFlagsReg cr, eRegI op1, memory op2) %{
11591 match(Set cr (CmpI op1 (LoadI op2)));
11593 format %{ "CMP $op1,$op2" %}
11594 ins_cost(500);
11595 opcode(0x3B); /* Opcode 3B /r */
11596 ins_encode( OpcP, RegMem( op1, op2) );
11597 ins_pipe( ialu_cr_reg_mem );
11598 %}
11600 instruct testI_reg( eFlagsReg cr, eRegI src, immI0 zero ) %{
11601 match(Set cr (CmpI src zero));
11602 effect( DEF cr, USE src );
11604 format %{ "TEST $src,$src" %}
11605 opcode(0x85);
11606 ins_encode( OpcP, RegReg( src, src ) );
11607 ins_pipe( ialu_cr_reg_imm );
11608 %}
11610 instruct testI_reg_imm( eFlagsReg cr, eRegI src, immI con, immI0 zero ) %{
11611 match(Set cr (CmpI (AndI src con) zero));
11613 format %{ "TEST $src,$con" %}
11614 opcode(0xF7,0x00);
11615 ins_encode( OpcP, RegOpc(src), Con32(con) );
11616 ins_pipe( ialu_cr_reg_imm );
11617 %}
11619 instruct testI_reg_mem( eFlagsReg cr, eRegI src, memory mem, immI0 zero ) %{
11620 match(Set cr (CmpI (AndI src mem) zero));
11622 format %{ "TEST $src,$mem" %}
11623 opcode(0x85);
11624 ins_encode( OpcP, RegMem( src, mem ) );
11625 ins_pipe( ialu_cr_reg_mem );
11626 %}
11628 // Unsigned compare Instructions; really, same as signed except they
11629 // produce an eFlagsRegU instead of eFlagsReg.
11630 instruct compU_eReg(eFlagsRegU cr, eRegI op1, eRegI op2) %{
11631 match(Set cr (CmpU op1 op2));
11633 format %{ "CMPu $op1,$op2" %}
11634 opcode(0x3B); /* Opcode 3B /r */
11635 ins_encode( OpcP, RegReg( op1, op2) );
11636 ins_pipe( ialu_cr_reg_reg );
11637 %}
11639 instruct compU_eReg_imm(eFlagsRegU cr, eRegI op1, immI op2) %{
11640 match(Set cr (CmpU op1 op2));
11642 format %{ "CMPu $op1,$op2" %}
11643 opcode(0x81,0x07); /* Opcode 81 /7 */
11644 ins_encode( OpcSErm( op1, op2 ), Con8or32( op2 ) );
11645 ins_pipe( ialu_cr_reg_imm );
11646 %}
11648 // // Cisc-spilled version of cmpU_eReg
11649 instruct compU_eReg_mem(eFlagsRegU cr, eRegI op1, memory op2) %{
11650 match(Set cr (CmpU op1 (LoadI op2)));
11652 format %{ "CMPu $op1,$op2" %}
11653 ins_cost(500);
11654 opcode(0x3B); /* Opcode 3B /r */
11655 ins_encode( OpcP, RegMem( op1, op2) );
11656 ins_pipe( ialu_cr_reg_mem );
11657 %}
11659 // // Cisc-spilled version of cmpU_eReg
11660 //instruct compU_mem_eReg(eFlagsRegU cr, memory op1, eRegI op2) %{
11661 // match(Set cr (CmpU (LoadI op1) op2));
11662 //
11663 // format %{ "CMPu $op1,$op2" %}
11664 // ins_cost(500);
11665 // opcode(0x39); /* Opcode 39 /r */
11666 // ins_encode( OpcP, RegMem( op1, op2) );
11667 //%}
11669 instruct testU_reg( eFlagsRegU cr, eRegI src, immI0 zero ) %{
11670 match(Set cr (CmpU src zero));
11672 format %{ "TESTu $src,$src" %}
11673 opcode(0x85);
11674 ins_encode( OpcP, RegReg( src, src ) );
11675 ins_pipe( ialu_cr_reg_imm );
11676 %}
11678 // Unsigned pointer compare Instructions
11679 instruct compP_eReg(eFlagsRegU cr, eRegP op1, eRegP op2) %{
11680 match(Set cr (CmpP op1 op2));
11682 format %{ "CMPu $op1,$op2" %}
11683 opcode(0x3B); /* Opcode 3B /r */
11684 ins_encode( OpcP, RegReg( op1, op2) );
11685 ins_pipe( ialu_cr_reg_reg );
11686 %}
11688 instruct compP_eReg_imm(eFlagsRegU cr, eRegP op1, immP op2) %{
11689 match(Set cr (CmpP op1 op2));
11691 format %{ "CMPu $op1,$op2" %}
11692 opcode(0x81,0x07); /* Opcode 81 /7 */
11693 ins_encode( OpcSErm( op1, op2 ), Con8or32( op2 ) );
11694 ins_pipe( ialu_cr_reg_imm );
11695 %}
11697 // // Cisc-spilled version of cmpP_eReg
11698 instruct compP_eReg_mem(eFlagsRegU cr, eRegP op1, memory op2) %{
11699 match(Set cr (CmpP op1 (LoadP op2)));
11701 format %{ "CMPu $op1,$op2" %}
11702 ins_cost(500);
11703 opcode(0x3B); /* Opcode 3B /r */
11704 ins_encode( OpcP, RegMem( op1, op2) );
11705 ins_pipe( ialu_cr_reg_mem );
11706 %}
11708 // // Cisc-spilled version of cmpP_eReg
11709 //instruct compP_mem_eReg(eFlagsRegU cr, memory op1, eRegP op2) %{
11710 // match(Set cr (CmpP (LoadP op1) op2));
11711 //
11712 // format %{ "CMPu $op1,$op2" %}
11713 // ins_cost(500);
11714 // opcode(0x39); /* Opcode 39 /r */
11715 // ins_encode( OpcP, RegMem( op1, op2) );
11716 //%}
11718 // Compare raw pointer (used in out-of-heap check).
11719 // Only works because non-oop pointers must be raw pointers
11720 // and raw pointers have no anti-dependencies.
11721 instruct compP_mem_eReg( eFlagsRegU cr, eRegP op1, memory op2 ) %{
11722 predicate( !n->in(2)->in(2)->bottom_type()->isa_oop_ptr() );
11723 match(Set cr (CmpP op1 (LoadP op2)));
11725 format %{ "CMPu $op1,$op2" %}
11726 opcode(0x3B); /* Opcode 3B /r */
11727 ins_encode( OpcP, RegMem( op1, op2) );
11728 ins_pipe( ialu_cr_reg_mem );
11729 %}
11731 //
11732 // This will generate a signed flags result. This should be ok
11733 // since any compare to a zero should be eq/neq.
11734 instruct testP_reg( eFlagsReg cr, eRegP src, immP0 zero ) %{
11735 match(Set cr (CmpP src zero));
11737 format %{ "TEST $src,$src" %}
11738 opcode(0x85);
11739 ins_encode( OpcP, RegReg( src, src ) );
11740 ins_pipe( ialu_cr_reg_imm );
11741 %}
11743 // Cisc-spilled version of testP_reg
11744 // This will generate a signed flags result. This should be ok
11745 // since any compare to a zero should be eq/neq.
11746 instruct testP_Reg_mem( eFlagsReg cr, memory op, immI0 zero ) %{
11747 match(Set cr (CmpP (LoadP op) zero));
11749 format %{ "TEST $op,0xFFFFFFFF" %}
11750 ins_cost(500);
11751 opcode(0xF7); /* Opcode F7 /0 */
11752 ins_encode( OpcP, RMopc_Mem(0x00,op), Con_d32(0xFFFFFFFF) );
11753 ins_pipe( ialu_cr_reg_imm );
11754 %}
11756 // Yanked all unsigned pointer compare operations.
11757 // Pointer compares are done with CmpP which is already unsigned.
11759 //----------Max and Min--------------------------------------------------------
11760 // Min Instructions
11761 ////
11762 // *** Min and Max using the conditional move are slower than the
11763 // *** branch version on a Pentium III.
11764 // // Conditional move for min
11765 //instruct cmovI_reg_lt( eRegI op2, eRegI op1, eFlagsReg cr ) %{
11766 // effect( USE_DEF op2, USE op1, USE cr );
11767 // format %{ "CMOVlt $op2,$op1\t! min" %}
11768 // opcode(0x4C,0x0F);
11769 // ins_encode( OpcS, OpcP, RegReg( op2, op1 ) );
11770 // ins_pipe( pipe_cmov_reg );
11771 //%}
11772 //
11773 //// Min Register with Register (P6 version)
11774 //instruct minI_eReg_p6( eRegI op1, eRegI op2 ) %{
11775 // predicate(VM_Version::supports_cmov() );
11776 // match(Set op2 (MinI op1 op2));
11777 // ins_cost(200);
11778 // expand %{
11779 // eFlagsReg cr;
11780 // compI_eReg(cr,op1,op2);
11781 // cmovI_reg_lt(op2,op1,cr);
11782 // %}
11783 //%}
11785 // Min Register with Register (generic version)
11786 instruct minI_eReg(eRegI dst, eRegI src, eFlagsReg flags) %{
11787 match(Set dst (MinI dst src));
11788 effect(KILL flags);
11789 ins_cost(300);
11791 format %{ "MIN $dst,$src" %}
11792 opcode(0xCC);
11793 ins_encode( min_enc(dst,src) );
11794 ins_pipe( pipe_slow );
11795 %}
11797 // Max Register with Register
11798 // *** Min and Max using the conditional move are slower than the
11799 // *** branch version on a Pentium III.
11800 // // Conditional move for max
11801 //instruct cmovI_reg_gt( eRegI op2, eRegI op1, eFlagsReg cr ) %{
11802 // effect( USE_DEF op2, USE op1, USE cr );
11803 // format %{ "CMOVgt $op2,$op1\t! max" %}
11804 // opcode(0x4F,0x0F);
11805 // ins_encode( OpcS, OpcP, RegReg( op2, op1 ) );
11806 // ins_pipe( pipe_cmov_reg );
11807 //%}
11808 //
11809 // // Max Register with Register (P6 version)
11810 //instruct maxI_eReg_p6( eRegI op1, eRegI op2 ) %{
11811 // predicate(VM_Version::supports_cmov() );
11812 // match(Set op2 (MaxI op1 op2));
11813 // ins_cost(200);
11814 // expand %{
11815 // eFlagsReg cr;
11816 // compI_eReg(cr,op1,op2);
11817 // cmovI_reg_gt(op2,op1,cr);
11818 // %}
11819 //%}
11821 // Max Register with Register (generic version)
11822 instruct maxI_eReg(eRegI dst, eRegI src, eFlagsReg flags) %{
11823 match(Set dst (MaxI dst src));
11824 effect(KILL flags);
11825 ins_cost(300);
11827 format %{ "MAX $dst,$src" %}
11828 opcode(0xCC);
11829 ins_encode( max_enc(dst,src) );
11830 ins_pipe( pipe_slow );
11831 %}
11833 // ============================================================================
11834 // Branch Instructions
11835 // Jump Table
11836 instruct jumpXtnd(eRegI switch_val) %{
11837 match(Jump switch_val);
11838 ins_cost(350);
11840 format %{ "JMP [table_base](,$switch_val,1)\n\t" %}
11842 ins_encode %{
11843 address table_base = __ address_table_constant(_index2label);
11845 // Jump to Address(table_base + switch_reg)
11846 InternalAddress table(table_base);
11847 Address index(noreg, $switch_val$$Register, Address::times_1);
11848 __ jump(ArrayAddress(table, index));
11849 %}
11850 ins_pc_relative(1);
11851 ins_pipe(pipe_jmp);
11852 %}
11854 // Jump Direct - Label defines a relative address from JMP+1
11855 instruct jmpDir(label labl) %{
11856 match(Goto);
11857 effect(USE labl);
11859 ins_cost(300);
11860 format %{ "JMP $labl" %}
11861 size(5);
11862 opcode(0xE9);
11863 ins_encode( OpcP, Lbl( labl ) );
11864 ins_pipe( pipe_jmp );
11865 ins_pc_relative(1);
11866 %}
11868 // Jump Direct Conditional - Label defines a relative address from Jcc+1
11869 instruct jmpCon(cmpOp cop, eFlagsReg cr, label labl) %{
11870 match(If cop cr);
11871 effect(USE labl);
11873 ins_cost(300);
11874 format %{ "J$cop $labl" %}
11875 size(6);
11876 opcode(0x0F, 0x80);
11877 ins_encode( Jcc( cop, labl) );
11878 ins_pipe( pipe_jcc );
11879 ins_pc_relative(1);
11880 %}
11882 // Jump Direct Conditional - Label defines a relative address from Jcc+1
11883 instruct jmpLoopEnd(cmpOp cop, eFlagsReg cr, label labl) %{
11884 match(CountedLoopEnd cop cr);
11885 effect(USE labl);
11887 ins_cost(300);
11888 format %{ "J$cop $labl\t# Loop end" %}
11889 size(6);
11890 opcode(0x0F, 0x80);
11891 ins_encode( Jcc( cop, labl) );
11892 ins_pipe( pipe_jcc );
11893 ins_pc_relative(1);
11894 %}
11896 // Jump Direct Conditional - Label defines a relative address from Jcc+1
11897 instruct jmpLoopEndU(cmpOpU cop, eFlagsRegU cmp, label labl) %{
11898 match(CountedLoopEnd cop cmp);
11899 effect(USE labl);
11901 ins_cost(300);
11902 format %{ "J$cop,u $labl\t# Loop end" %}
11903 size(6);
11904 opcode(0x0F, 0x80);
11905 ins_encode( Jcc( cop, labl) );
11906 ins_pipe( pipe_jcc );
11907 ins_pc_relative(1);
11908 %}
11910 // Jump Direct Conditional - using unsigned comparison
11911 instruct jmpConU(cmpOpU cop, eFlagsRegU cmp, label labl) %{
11912 match(If cop cmp);
11913 effect(USE labl);
11915 ins_cost(300);
11916 format %{ "J$cop,u $labl" %}
11917 size(6);
11918 opcode(0x0F, 0x80);
11919 ins_encode( Jcc( cop, labl) );
11920 ins_pipe( pipe_jcc );
11921 ins_pc_relative(1);
11922 %}
11924 // ============================================================================
11925 // The 2nd slow-half of a subtype check. Scan the subklass's 2ndary superklass
11926 // array for an instance of the superklass. Set a hidden internal cache on a
11927 // hit (cache is checked with exposed code in gen_subtype_check()). Return
11928 // NZ for a miss or zero for a hit. The encoding ALSO sets flags.
11929 instruct partialSubtypeCheck( eDIRegP result, eSIRegP sub, eAXRegP super, eCXRegI rcx, eFlagsReg cr ) %{
11930 match(Set result (PartialSubtypeCheck sub super));
11931 effect( KILL rcx, KILL cr );
11933 ins_cost(1100); // slightly larger than the next version
11934 format %{ "CMPL EAX,ESI\n\t"
11935 "JEQ,s hit\n\t"
11936 "MOV EDI,[$sub+Klass::secondary_supers]\n\t"
11937 "MOV ECX,[EDI+arrayKlass::length]\t# length to scan\n\t"
11938 "ADD EDI,arrayKlass::base_offset\t# Skip to start of data; set NZ in case count is zero\n\t"
11939 "REPNE SCASD\t# Scan *EDI++ for a match with EAX while CX-- != 0\n\t"
11940 "JNE,s miss\t\t# Missed: EDI not-zero\n\t"
11941 "MOV [$sub+Klass::secondary_super_cache],$super\t# Hit: update cache\n\t"
11942 "hit:\n\t"
11943 "XOR $result,$result\t\t Hit: EDI zero\n\t"
11944 "miss:\t" %}
11946 opcode(0x1); // Force a XOR of EDI
11947 ins_encode( enc_PartialSubtypeCheck() );
11948 ins_pipe( pipe_slow );
11949 %}
11951 instruct partialSubtypeCheck_vs_Zero( eFlagsReg cr, eSIRegP sub, eAXRegP super, eCXRegI rcx, eDIRegP result, immP0 zero ) %{
11952 match(Set cr (CmpP (PartialSubtypeCheck sub super) zero));
11953 effect( KILL rcx, KILL result );
11955 ins_cost(1000);
11956 format %{ "CMPL EAX,ESI\n\t"
11957 "JEQ,s miss\t# Actually a hit; we are done.\n\t"
11958 "MOV EDI,[$sub+Klass::secondary_supers]\n\t"
11959 "MOV ECX,[EDI+arrayKlass::length]\t# length to scan\n\t"
11960 "ADD EDI,arrayKlass::base_offset\t# Skip to start of data; set NZ in case count is zero\n\t"
11961 "REPNE SCASD\t# Scan *EDI++ for a match with EAX while CX-- != 0\n\t"
11962 "JNE,s miss\t\t# Missed: flags NZ\n\t"
11963 "MOV [$sub+Klass::secondary_super_cache],$super\t# Hit: update cache, flags Z\n\t"
11964 "miss:\t" %}
11966 opcode(0x0); // No need to XOR EDI
11967 ins_encode( enc_PartialSubtypeCheck() );
11968 ins_pipe( pipe_slow );
11969 %}
11971 // ============================================================================
11972 // Branch Instructions -- short offset versions
11973 //
11974 // These instructions are used to replace jumps of a long offset (the default
11975 // match) with jumps of a shorter offset. These instructions are all tagged
11976 // with the ins_short_branch attribute, which causes the ADLC to suppress the
11977 // match rules in general matching. Instead, the ADLC generates a conversion
11978 // method in the MachNode which can be used to do in-place replacement of the
11979 // long variant with the shorter variant. The compiler will determine if a
11980 // branch can be taken by the is_short_branch_offset() predicate in the machine
11981 // specific code section of the file.
11983 // Jump Direct - Label defines a relative address from JMP+1
11984 instruct jmpDir_short(label labl) %{
11985 match(Goto);
11986 effect(USE labl);
11988 ins_cost(300);
11989 format %{ "JMP,s $labl" %}
11990 size(2);
11991 opcode(0xEB);
11992 ins_encode( OpcP, LblShort( labl ) );
11993 ins_pipe( pipe_jmp );
11994 ins_pc_relative(1);
11995 ins_short_branch(1);
11996 %}
11998 // Jump Direct Conditional - Label defines a relative address from Jcc+1
11999 instruct jmpCon_short(cmpOp cop, eFlagsReg cr, label labl) %{
12000 match(If cop cr);
12001 effect(USE labl);
12003 ins_cost(300);
12004 format %{ "J$cop,s $labl" %}
12005 size(2);
12006 opcode(0x70);
12007 ins_encode( JccShort( cop, labl) );
12008 ins_pipe( pipe_jcc );
12009 ins_pc_relative(1);
12010 ins_short_branch(1);
12011 %}
12013 // Jump Direct Conditional - Label defines a relative address from Jcc+1
12014 instruct jmpLoopEnd_short(cmpOp cop, eFlagsReg cr, label labl) %{
12015 match(CountedLoopEnd cop cr);
12016 effect(USE labl);
12018 ins_cost(300);
12019 format %{ "J$cop,s $labl" %}
12020 size(2);
12021 opcode(0x70);
12022 ins_encode( JccShort( cop, labl) );
12023 ins_pipe( pipe_jcc );
12024 ins_pc_relative(1);
12025 ins_short_branch(1);
12026 %}
12028 // Jump Direct Conditional - Label defines a relative address from Jcc+1
12029 instruct jmpLoopEndU_short(cmpOpU cop, eFlagsRegU cmp, label labl) %{
12030 match(CountedLoopEnd cop cmp);
12031 effect(USE labl);
12033 ins_cost(300);
12034 format %{ "J$cop,us $labl" %}
12035 size(2);
12036 opcode(0x70);
12037 ins_encode( JccShort( cop, labl) );
12038 ins_pipe( pipe_jcc );
12039 ins_pc_relative(1);
12040 ins_short_branch(1);
12041 %}
12043 // Jump Direct Conditional - using unsigned comparison
12044 instruct jmpConU_short(cmpOpU cop, eFlagsRegU cmp, label labl) %{
12045 match(If cop cmp);
12046 effect(USE labl);
12048 ins_cost(300);
12049 format %{ "J$cop,us $labl" %}
12050 size(2);
12051 opcode(0x70);
12052 ins_encode( JccShort( cop, labl) );
12053 ins_pipe( pipe_jcc );
12054 ins_pc_relative(1);
12055 ins_short_branch(1);
12056 %}
12058 // ============================================================================
12059 // Long Compare
12060 //
12061 // Currently we hold longs in 2 registers. Comparing such values efficiently
12062 // is tricky. The flavor of compare used depends on whether we are testing
12063 // for LT, LE, or EQ. For a simple LT test we can check just the sign bit.
12064 // The GE test is the negated LT test. The LE test can be had by commuting
12065 // the operands (yielding a GE test) and then negating; negate again for the
12066 // GT test. The EQ test is done by ORcc'ing the high and low halves, and the
12067 // NE test is negated from that.
12069 // Due to a shortcoming in the ADLC, it mixes up expressions like:
12070 // (foo (CmpI (CmpL X Y) 0)) and (bar (CmpI (CmpL X 0L) 0)). Note the
12071 // difference between 'Y' and '0L'. The tree-matches for the CmpI sections
12072 // are collapsed internally in the ADLC's dfa-gen code. The match for
12073 // (CmpI (CmpL X Y) 0) is silently replaced with (CmpI (CmpL X 0L) 0) and the
12074 // foo match ends up with the wrong leaf. One fix is to not match both
12075 // reg-reg and reg-zero forms of long-compare. This is unfortunate because
12076 // both forms beat the trinary form of long-compare and both are very useful
12077 // on Intel which has so few registers.
12079 // Manifest a CmpL result in an integer register. Very painful.
12080 // This is the test to avoid.
12081 instruct cmpL3_reg_reg(eSIRegI dst, eRegL src1, eRegL src2, eFlagsReg flags ) %{
12082 match(Set dst (CmpL3 src1 src2));
12083 effect( KILL flags );
12084 ins_cost(1000);
12085 format %{ "XOR $dst,$dst\n\t"
12086 "CMP $src1.hi,$src2.hi\n\t"
12087 "JLT,s m_one\n\t"
12088 "JGT,s p_one\n\t"
12089 "CMP $src1.lo,$src2.lo\n\t"
12090 "JB,s m_one\n\t"
12091 "JEQ,s done\n"
12092 "p_one:\tINC $dst\n\t"
12093 "JMP,s done\n"
12094 "m_one:\tDEC $dst\n"
12095 "done:" %}
12096 ins_encode %{
12097 Label p_one, m_one, done;
12098 __ xorl($dst$$Register, $dst$$Register);
12099 __ cmpl(HIGH_FROM_LOW($src1$$Register), HIGH_FROM_LOW($src2$$Register));
12100 __ jccb(Assembler::less, m_one);
12101 __ jccb(Assembler::greater, p_one);
12102 __ cmpl($src1$$Register, $src2$$Register);
12103 __ jccb(Assembler::below, m_one);
12104 __ jccb(Assembler::equal, done);
12105 __ bind(p_one);
12106 __ increment($dst$$Register);
12107 __ jmpb(done);
12108 __ bind(m_one);
12109 __ decrement($dst$$Register);
12110 __ bind(done);
12111 %}
12112 ins_pipe( pipe_slow );
12113 %}
12115 //======
12116 // Manifest a CmpL result in the normal flags. Only good for LT or GE
12117 // compares. Can be used for LE or GT compares by reversing arguments.
12118 // NOT GOOD FOR EQ/NE tests.
12119 instruct cmpL_zero_flags_LTGE( flagsReg_long_LTGE flags, eRegL src, immL0 zero ) %{
12120 match( Set flags (CmpL src zero ));
12121 ins_cost(100);
12122 format %{ "TEST $src.hi,$src.hi" %}
12123 opcode(0x85);
12124 ins_encode( OpcP, RegReg_Hi2( src, src ) );
12125 ins_pipe( ialu_cr_reg_reg );
12126 %}
12128 // Manifest a CmpL result in the normal flags. Only good for LT or GE
12129 // compares. Can be used for LE or GT compares by reversing arguments.
12130 // NOT GOOD FOR EQ/NE tests.
12131 instruct cmpL_reg_flags_LTGE( flagsReg_long_LTGE flags, eRegL src1, eRegL src2, eRegI tmp ) %{
12132 match( Set flags (CmpL src1 src2 ));
12133 effect( TEMP tmp );
12134 ins_cost(300);
12135 format %{ "CMP $src1.lo,$src2.lo\t! Long compare; set flags for low bits\n\t"
12136 "MOV $tmp,$src1.hi\n\t"
12137 "SBB $tmp,$src2.hi\t! Compute flags for long compare" %}
12138 ins_encode( long_cmp_flags2( src1, src2, tmp ) );
12139 ins_pipe( ialu_cr_reg_reg );
12140 %}
12142 // Long compares reg < zero/req OR reg >= zero/req.
12143 // Just a wrapper for a normal branch, plus the predicate test.
12144 instruct cmpL_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, label labl) %{
12145 match(If cmp flags);
12146 effect(USE labl);
12147 predicate( _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::lt || _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ge );
12148 expand %{
12149 jmpCon(cmp,flags,labl); // JLT or JGE...
12150 %}
12151 %}
12153 // Compare 2 longs and CMOVE longs.
12154 instruct cmovLL_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, eRegL dst, eRegL src) %{
12155 match(Set dst (CMoveL (Binary cmp flags) (Binary dst src)));
12156 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 ));
12157 ins_cost(400);
12158 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
12159 "CMOV$cmp $dst.hi,$src.hi" %}
12160 opcode(0x0F,0x40);
12161 ins_encode( enc_cmov(cmp), RegReg_Lo2( dst, src ), enc_cmov(cmp), RegReg_Hi2( dst, src ) );
12162 ins_pipe( pipe_cmov_reg_long );
12163 %}
12165 instruct cmovLL_mem_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, eRegL dst, load_long_memory src) %{
12166 match(Set dst (CMoveL (Binary cmp flags) (Binary dst (LoadL src))));
12167 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 ));
12168 ins_cost(500);
12169 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
12170 "CMOV$cmp $dst.hi,$src.hi" %}
12171 opcode(0x0F,0x40);
12172 ins_encode( enc_cmov(cmp), RegMem(dst, src), enc_cmov(cmp), RegMem_Hi(dst, src) );
12173 ins_pipe( pipe_cmov_reg_long );
12174 %}
12176 // Compare 2 longs and CMOVE ints.
12177 instruct cmovII_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, eRegI dst, eRegI src) %{
12178 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 ));
12179 match(Set dst (CMoveI (Binary cmp flags) (Binary dst src)));
12180 ins_cost(200);
12181 format %{ "CMOV$cmp $dst,$src" %}
12182 opcode(0x0F,0x40);
12183 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
12184 ins_pipe( pipe_cmov_reg );
12185 %}
12187 instruct cmovII_mem_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, eRegI dst, memory src) %{
12188 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 ));
12189 match(Set dst (CMoveI (Binary cmp flags) (Binary dst (LoadI src))));
12190 ins_cost(250);
12191 format %{ "CMOV$cmp $dst,$src" %}
12192 opcode(0x0F,0x40);
12193 ins_encode( enc_cmov(cmp), RegMem( dst, src ) );
12194 ins_pipe( pipe_cmov_mem );
12195 %}
12197 // Compare 2 longs and CMOVE ints.
12198 instruct cmovPP_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, eRegP dst, eRegP src) %{
12199 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 ));
12200 match(Set dst (CMoveP (Binary cmp flags) (Binary dst src)));
12201 ins_cost(200);
12202 format %{ "CMOV$cmp $dst,$src" %}
12203 opcode(0x0F,0x40);
12204 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
12205 ins_pipe( pipe_cmov_reg );
12206 %}
12208 // Compare 2 longs and CMOVE doubles
12209 instruct cmovDD_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, regD dst, regD src) %{
12210 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 );
12211 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
12212 ins_cost(200);
12213 expand %{
12214 fcmovD_regS(cmp,flags,dst,src);
12215 %}
12216 %}
12218 // Compare 2 longs and CMOVE doubles
12219 instruct cmovXDD_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, regXD dst, regXD src) %{
12220 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 );
12221 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
12222 ins_cost(200);
12223 expand %{
12224 fcmovXD_regS(cmp,flags,dst,src);
12225 %}
12226 %}
12228 instruct cmovFF_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, regF dst, regF src) %{
12229 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 );
12230 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
12231 ins_cost(200);
12232 expand %{
12233 fcmovF_regS(cmp,flags,dst,src);
12234 %}
12235 %}
12237 instruct cmovXX_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, regX dst, regX src) %{
12238 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 );
12239 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
12240 ins_cost(200);
12241 expand %{
12242 fcmovX_regS(cmp,flags,dst,src);
12243 %}
12244 %}
12246 //======
12247 // Manifest a CmpL result in the normal flags. Only good for EQ/NE compares.
12248 instruct cmpL_zero_flags_EQNE( flagsReg_long_EQNE flags, eRegL src, immL0 zero, eRegI tmp ) %{
12249 match( Set flags (CmpL src zero ));
12250 effect(TEMP tmp);
12251 ins_cost(200);
12252 format %{ "MOV $tmp,$src.lo\n\t"
12253 "OR $tmp,$src.hi\t! Long is EQ/NE 0?" %}
12254 ins_encode( long_cmp_flags0( src, tmp ) );
12255 ins_pipe( ialu_reg_reg_long );
12256 %}
12258 // Manifest a CmpL result in the normal flags. Only good for EQ/NE compares.
12259 instruct cmpL_reg_flags_EQNE( flagsReg_long_EQNE flags, eRegL src1, eRegL src2 ) %{
12260 match( Set flags (CmpL src1 src2 ));
12261 ins_cost(200+300);
12262 format %{ "CMP $src1.lo,$src2.lo\t! Long compare; set flags for low bits\n\t"
12263 "JNE,s skip\n\t"
12264 "CMP $src1.hi,$src2.hi\n\t"
12265 "skip:\t" %}
12266 ins_encode( long_cmp_flags1( src1, src2 ) );
12267 ins_pipe( ialu_cr_reg_reg );
12268 %}
12270 // Long compare reg == zero/reg OR reg != zero/reg
12271 // Just a wrapper for a normal branch, plus the predicate test.
12272 instruct cmpL_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, label labl) %{
12273 match(If cmp flags);
12274 effect(USE labl);
12275 predicate( _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::eq || _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ne );
12276 expand %{
12277 jmpCon(cmp,flags,labl); // JEQ or JNE...
12278 %}
12279 %}
12281 // Compare 2 longs and CMOVE longs.
12282 instruct cmovLL_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, eRegL dst, eRegL src) %{
12283 match(Set dst (CMoveL (Binary cmp flags) (Binary dst src)));
12284 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 ));
12285 ins_cost(400);
12286 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
12287 "CMOV$cmp $dst.hi,$src.hi" %}
12288 opcode(0x0F,0x40);
12289 ins_encode( enc_cmov(cmp), RegReg_Lo2( dst, src ), enc_cmov(cmp), RegReg_Hi2( dst, src ) );
12290 ins_pipe( pipe_cmov_reg_long );
12291 %}
12293 instruct cmovLL_mem_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, eRegL dst, load_long_memory src) %{
12294 match(Set dst (CMoveL (Binary cmp flags) (Binary dst (LoadL src))));
12295 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 ));
12296 ins_cost(500);
12297 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
12298 "CMOV$cmp $dst.hi,$src.hi" %}
12299 opcode(0x0F,0x40);
12300 ins_encode( enc_cmov(cmp), RegMem(dst, src), enc_cmov(cmp), RegMem_Hi(dst, src) );
12301 ins_pipe( pipe_cmov_reg_long );
12302 %}
12304 // Compare 2 longs and CMOVE ints.
12305 instruct cmovII_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, eRegI dst, eRegI src) %{
12306 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 ));
12307 match(Set dst (CMoveI (Binary cmp flags) (Binary dst src)));
12308 ins_cost(200);
12309 format %{ "CMOV$cmp $dst,$src" %}
12310 opcode(0x0F,0x40);
12311 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
12312 ins_pipe( pipe_cmov_reg );
12313 %}
12315 instruct cmovII_mem_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, eRegI dst, memory src) %{
12316 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 ));
12317 match(Set dst (CMoveI (Binary cmp flags) (Binary dst (LoadI src))));
12318 ins_cost(250);
12319 format %{ "CMOV$cmp $dst,$src" %}
12320 opcode(0x0F,0x40);
12321 ins_encode( enc_cmov(cmp), RegMem( dst, src ) );
12322 ins_pipe( pipe_cmov_mem );
12323 %}
12325 // Compare 2 longs and CMOVE ints.
12326 instruct cmovPP_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, eRegP dst, eRegP src) %{
12327 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 ));
12328 match(Set dst (CMoveP (Binary cmp flags) (Binary dst src)));
12329 ins_cost(200);
12330 format %{ "CMOV$cmp $dst,$src" %}
12331 opcode(0x0F,0x40);
12332 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
12333 ins_pipe( pipe_cmov_reg );
12334 %}
12336 // Compare 2 longs and CMOVE doubles
12337 instruct cmovDD_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, regD dst, regD src) %{
12338 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 );
12339 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
12340 ins_cost(200);
12341 expand %{
12342 fcmovD_regS(cmp,flags,dst,src);
12343 %}
12344 %}
12346 // Compare 2 longs and CMOVE doubles
12347 instruct cmovXDD_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, regXD dst, regXD src) %{
12348 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 );
12349 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
12350 ins_cost(200);
12351 expand %{
12352 fcmovXD_regS(cmp,flags,dst,src);
12353 %}
12354 %}
12356 instruct cmovFF_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, regF dst, regF src) %{
12357 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 );
12358 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
12359 ins_cost(200);
12360 expand %{
12361 fcmovF_regS(cmp,flags,dst,src);
12362 %}
12363 %}
12365 instruct cmovXX_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, regX dst, regX src) %{
12366 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 );
12367 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
12368 ins_cost(200);
12369 expand %{
12370 fcmovX_regS(cmp,flags,dst,src);
12371 %}
12372 %}
12374 //======
12375 // Manifest a CmpL result in the normal flags. Only good for LE or GT compares.
12376 // Same as cmpL_reg_flags_LEGT except must negate src
12377 instruct cmpL_zero_flags_LEGT( flagsReg_long_LEGT flags, eRegL src, immL0 zero, eRegI tmp ) %{
12378 match( Set flags (CmpL src zero ));
12379 effect( TEMP tmp );
12380 ins_cost(300);
12381 format %{ "XOR $tmp,$tmp\t# Long compare for -$src < 0, use commuted test\n\t"
12382 "CMP $tmp,$src.lo\n\t"
12383 "SBB $tmp,$src.hi\n\t" %}
12384 ins_encode( long_cmp_flags3(src, tmp) );
12385 ins_pipe( ialu_reg_reg_long );
12386 %}
12388 // Manifest a CmpL result in the normal flags. Only good for LE or GT compares.
12389 // Same as cmpL_reg_flags_LTGE except operands swapped. Swapping operands
12390 // requires a commuted test to get the same result.
12391 instruct cmpL_reg_flags_LEGT( flagsReg_long_LEGT flags, eRegL src1, eRegL src2, eRegI tmp ) %{
12392 match( Set flags (CmpL src1 src2 ));
12393 effect( TEMP tmp );
12394 ins_cost(300);
12395 format %{ "CMP $src2.lo,$src1.lo\t! Long compare, swapped operands, use with commuted test\n\t"
12396 "MOV $tmp,$src2.hi\n\t"
12397 "SBB $tmp,$src1.hi\t! Compute flags for long compare" %}
12398 ins_encode( long_cmp_flags2( src2, src1, tmp ) );
12399 ins_pipe( ialu_cr_reg_reg );
12400 %}
12402 // Long compares reg < zero/req OR reg >= zero/req.
12403 // Just a wrapper for a normal branch, plus the predicate test
12404 instruct cmpL_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, label labl) %{
12405 match(If cmp flags);
12406 effect(USE labl);
12407 predicate( _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::gt || _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::le );
12408 ins_cost(300);
12409 expand %{
12410 jmpCon(cmp,flags,labl); // JGT or JLE...
12411 %}
12412 %}
12414 // Compare 2 longs and CMOVE longs.
12415 instruct cmovLL_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, eRegL dst, eRegL src) %{
12416 match(Set dst (CMoveL (Binary cmp flags) (Binary dst src)));
12417 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 ));
12418 ins_cost(400);
12419 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
12420 "CMOV$cmp $dst.hi,$src.hi" %}
12421 opcode(0x0F,0x40);
12422 ins_encode( enc_cmov(cmp), RegReg_Lo2( dst, src ), enc_cmov(cmp), RegReg_Hi2( dst, src ) );
12423 ins_pipe( pipe_cmov_reg_long );
12424 %}
12426 instruct cmovLL_mem_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, eRegL dst, load_long_memory src) %{
12427 match(Set dst (CMoveL (Binary cmp flags) (Binary dst (LoadL src))));
12428 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 ));
12429 ins_cost(500);
12430 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
12431 "CMOV$cmp $dst.hi,$src.hi+4" %}
12432 opcode(0x0F,0x40);
12433 ins_encode( enc_cmov(cmp), RegMem(dst, src), enc_cmov(cmp), RegMem_Hi(dst, src) );
12434 ins_pipe( pipe_cmov_reg_long );
12435 %}
12437 // Compare 2 longs and CMOVE ints.
12438 instruct cmovII_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, eRegI dst, eRegI src) %{
12439 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 ));
12440 match(Set dst (CMoveI (Binary cmp flags) (Binary dst src)));
12441 ins_cost(200);
12442 format %{ "CMOV$cmp $dst,$src" %}
12443 opcode(0x0F,0x40);
12444 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
12445 ins_pipe( pipe_cmov_reg );
12446 %}
12448 instruct cmovII_mem_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, eRegI dst, memory src) %{
12449 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 ));
12450 match(Set dst (CMoveI (Binary cmp flags) (Binary dst (LoadI src))));
12451 ins_cost(250);
12452 format %{ "CMOV$cmp $dst,$src" %}
12453 opcode(0x0F,0x40);
12454 ins_encode( enc_cmov(cmp), RegMem( dst, src ) );
12455 ins_pipe( pipe_cmov_mem );
12456 %}
12458 // Compare 2 longs and CMOVE ptrs.
12459 instruct cmovPP_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, eRegP dst, eRegP src) %{
12460 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 ));
12461 match(Set dst (CMoveP (Binary cmp flags) (Binary dst src)));
12462 ins_cost(200);
12463 format %{ "CMOV$cmp $dst,$src" %}
12464 opcode(0x0F,0x40);
12465 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
12466 ins_pipe( pipe_cmov_reg );
12467 %}
12469 // Compare 2 longs and CMOVE doubles
12470 instruct cmovDD_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, regD dst, regD src) %{
12471 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 );
12472 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
12473 ins_cost(200);
12474 expand %{
12475 fcmovD_regS(cmp,flags,dst,src);
12476 %}
12477 %}
12479 // Compare 2 longs and CMOVE doubles
12480 instruct cmovXDD_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, regXD dst, regXD src) %{
12481 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 );
12482 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
12483 ins_cost(200);
12484 expand %{
12485 fcmovXD_regS(cmp,flags,dst,src);
12486 %}
12487 %}
12489 instruct cmovFF_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, regF dst, regF src) %{
12490 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 );
12491 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
12492 ins_cost(200);
12493 expand %{
12494 fcmovF_regS(cmp,flags,dst,src);
12495 %}
12496 %}
12499 instruct cmovXX_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, regX dst, regX src) %{
12500 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 );
12501 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
12502 ins_cost(200);
12503 expand %{
12504 fcmovX_regS(cmp,flags,dst,src);
12505 %}
12506 %}
12509 // ============================================================================
12510 // Procedure Call/Return Instructions
12511 // Call Java Static Instruction
12512 // Note: If this code changes, the corresponding ret_addr_offset() and
12513 // compute_padding() functions will have to be adjusted.
12514 instruct CallStaticJavaDirect(method meth) %{
12515 match(CallStaticJava);
12516 effect(USE meth);
12518 ins_cost(300);
12519 format %{ "CALL,static " %}
12520 opcode(0xE8); /* E8 cd */
12521 ins_encode( pre_call_FPU,
12522 Java_Static_Call( meth ),
12523 call_epilog,
12524 post_call_FPU );
12525 ins_pipe( pipe_slow );
12526 ins_pc_relative(1);
12527 ins_alignment(4);
12528 %}
12530 // Call Java Dynamic Instruction
12531 // Note: If this code changes, the corresponding ret_addr_offset() and
12532 // compute_padding() functions will have to be adjusted.
12533 instruct CallDynamicJavaDirect(method meth) %{
12534 match(CallDynamicJava);
12535 effect(USE meth);
12537 ins_cost(300);
12538 format %{ "MOV EAX,(oop)-1\n\t"
12539 "CALL,dynamic" %}
12540 opcode(0xE8); /* E8 cd */
12541 ins_encode( pre_call_FPU,
12542 Java_Dynamic_Call( meth ),
12543 call_epilog,
12544 post_call_FPU );
12545 ins_pipe( pipe_slow );
12546 ins_pc_relative(1);
12547 ins_alignment(4);
12548 %}
12550 // Call Runtime Instruction
12551 instruct CallRuntimeDirect(method meth) %{
12552 match(CallRuntime );
12553 effect(USE meth);
12555 ins_cost(300);
12556 format %{ "CALL,runtime " %}
12557 opcode(0xE8); /* E8 cd */
12558 // Use FFREEs to clear entries in float stack
12559 ins_encode( pre_call_FPU,
12560 FFree_Float_Stack_All,
12561 Java_To_Runtime( meth ),
12562 post_call_FPU );
12563 ins_pipe( pipe_slow );
12564 ins_pc_relative(1);
12565 %}
12567 // Call runtime without safepoint
12568 instruct CallLeafDirect(method meth) %{
12569 match(CallLeaf);
12570 effect(USE meth);
12572 ins_cost(300);
12573 format %{ "CALL_LEAF,runtime " %}
12574 opcode(0xE8); /* E8 cd */
12575 ins_encode( pre_call_FPU,
12576 FFree_Float_Stack_All,
12577 Java_To_Runtime( meth ),
12578 Verify_FPU_For_Leaf, post_call_FPU );
12579 ins_pipe( pipe_slow );
12580 ins_pc_relative(1);
12581 %}
12583 instruct CallLeafNoFPDirect(method meth) %{
12584 match(CallLeafNoFP);
12585 effect(USE meth);
12587 ins_cost(300);
12588 format %{ "CALL_LEAF_NOFP,runtime " %}
12589 opcode(0xE8); /* E8 cd */
12590 ins_encode(Java_To_Runtime(meth));
12591 ins_pipe( pipe_slow );
12592 ins_pc_relative(1);
12593 %}
12596 // Return Instruction
12597 // Remove the return address & jump to it.
12598 instruct Ret() %{
12599 match(Return);
12600 format %{ "RET" %}
12601 opcode(0xC3);
12602 ins_encode(OpcP);
12603 ins_pipe( pipe_jmp );
12604 %}
12606 // Tail Call; Jump from runtime stub to Java code.
12607 // Also known as an 'interprocedural jump'.
12608 // Target of jump will eventually return to caller.
12609 // TailJump below removes the return address.
12610 instruct TailCalljmpInd(eRegP_no_EBP jump_target, eBXRegP method_oop) %{
12611 match(TailCall jump_target method_oop );
12612 ins_cost(300);
12613 format %{ "JMP $jump_target \t# EBX holds method oop" %}
12614 opcode(0xFF, 0x4); /* Opcode FF /4 */
12615 ins_encode( OpcP, RegOpc(jump_target) );
12616 ins_pipe( pipe_jmp );
12617 %}
12620 // Tail Jump; remove the return address; jump to target.
12621 // TailCall above leaves the return address around.
12622 instruct tailjmpInd(eRegP_no_EBP jump_target, eAXRegP ex_oop) %{
12623 match( TailJump jump_target ex_oop );
12624 ins_cost(300);
12625 format %{ "POP EDX\t# pop return address into dummy\n\t"
12626 "JMP $jump_target " %}
12627 opcode(0xFF, 0x4); /* Opcode FF /4 */
12628 ins_encode( enc_pop_rdx,
12629 OpcP, RegOpc(jump_target) );
12630 ins_pipe( pipe_jmp );
12631 %}
12633 // Create exception oop: created by stack-crawling runtime code.
12634 // Created exception is now available to this handler, and is setup
12635 // just prior to jumping to this handler. No code emitted.
12636 instruct CreateException( eAXRegP ex_oop )
12637 %{
12638 match(Set ex_oop (CreateEx));
12640 size(0);
12641 // use the following format syntax
12642 format %{ "# exception oop is in EAX; no code emitted" %}
12643 ins_encode();
12644 ins_pipe( empty );
12645 %}
12648 // Rethrow exception:
12649 // The exception oop will come in the first argument position.
12650 // Then JUMP (not call) to the rethrow stub code.
12651 instruct RethrowException()
12652 %{
12653 match(Rethrow);
12655 // use the following format syntax
12656 format %{ "JMP rethrow_stub" %}
12657 ins_encode(enc_rethrow);
12658 ins_pipe( pipe_jmp );
12659 %}
12661 // inlined locking and unlocking
12664 instruct cmpFastLock( eFlagsReg cr, eRegP object, eRegP box, eAXRegI tmp, eRegP scr) %{
12665 match( Set cr (FastLock object box) );
12666 effect( TEMP tmp, TEMP scr );
12667 ins_cost(300);
12668 format %{ "FASTLOCK $object, $box KILLS $tmp,$scr" %}
12669 ins_encode( Fast_Lock(object,box,tmp,scr) );
12670 ins_pipe( pipe_slow );
12671 ins_pc_relative(1);
12672 %}
12674 instruct cmpFastUnlock( eFlagsReg cr, eRegP object, eAXRegP box, eRegP tmp ) %{
12675 match( Set cr (FastUnlock object box) );
12676 effect( TEMP tmp );
12677 ins_cost(300);
12678 format %{ "FASTUNLOCK $object, $box, $tmp" %}
12679 ins_encode( Fast_Unlock(object,box,tmp) );
12680 ins_pipe( pipe_slow );
12681 ins_pc_relative(1);
12682 %}
12686 // ============================================================================
12687 // Safepoint Instruction
12688 instruct safePoint_poll(eFlagsReg cr) %{
12689 match(SafePoint);
12690 effect(KILL cr);
12692 // TODO-FIXME: we currently poll at offset 0 of the safepoint polling page.
12693 // On SPARC that might be acceptable as we can generate the address with
12694 // just a sethi, saving an or. By polling at offset 0 we can end up
12695 // putting additional pressure on the index-0 in the D$. Because of
12696 // alignment (just like the situation at hand) the lower indices tend
12697 // to see more traffic. It'd be better to change the polling address
12698 // to offset 0 of the last $line in the polling page.
12700 format %{ "TSTL #polladdr,EAX\t! Safepoint: poll for GC" %}
12701 ins_cost(125);
12702 size(6) ;
12703 ins_encode( Safepoint_Poll() );
12704 ins_pipe( ialu_reg_mem );
12705 %}
12707 //----------PEEPHOLE RULES-----------------------------------------------------
12708 // These must follow all instruction definitions as they use the names
12709 // defined in the instructions definitions.
12710 //
12711 // peepmatch ( root_instr_name [preceeding_instruction]* );
12712 //
12713 // peepconstraint %{
12714 // (instruction_number.operand_name relational_op instruction_number.operand_name
12715 // [, ...] );
12716 // // instruction numbers are zero-based using left to right order in peepmatch
12717 //
12718 // peepreplace ( instr_name ( [instruction_number.operand_name]* ) );
12719 // // provide an instruction_number.operand_name for each operand that appears
12720 // // in the replacement instruction's match rule
12721 //
12722 // ---------VM FLAGS---------------------------------------------------------
12723 //
12724 // All peephole optimizations can be turned off using -XX:-OptoPeephole
12725 //
12726 // Each peephole rule is given an identifying number starting with zero and
12727 // increasing by one in the order seen by the parser. An individual peephole
12728 // can be enabled, and all others disabled, by using -XX:OptoPeepholeAt=#
12729 // on the command-line.
12730 //
12731 // ---------CURRENT LIMITATIONS----------------------------------------------
12732 //
12733 // Only match adjacent instructions in same basic block
12734 // Only equality constraints
12735 // Only constraints between operands, not (0.dest_reg == EAX_enc)
12736 // Only one replacement instruction
12737 //
12738 // ---------EXAMPLE----------------------------------------------------------
12739 //
12740 // // pertinent parts of existing instructions in architecture description
12741 // instruct movI(eRegI dst, eRegI src) %{
12742 // match(Set dst (CopyI src));
12743 // %}
12744 //
12745 // instruct incI_eReg(eRegI dst, immI1 src, eFlagsReg cr) %{
12746 // match(Set dst (AddI dst src));
12747 // effect(KILL cr);
12748 // %}
12749 //
12750 // // Change (inc mov) to lea
12751 // peephole %{
12752 // // increment preceeded by register-register move
12753 // peepmatch ( incI_eReg movI );
12754 // // require that the destination register of the increment
12755 // // match the destination register of the move
12756 // peepconstraint ( 0.dst == 1.dst );
12757 // // construct a replacement instruction that sets
12758 // // the destination to ( move's source register + one )
12759 // peepreplace ( leaI_eReg_immI( 0.dst 1.src 0.src ) );
12760 // %}
12761 //
12762 // Implementation no longer uses movX instructions since
12763 // machine-independent system no longer uses CopyX nodes.
12764 //
12765 // peephole %{
12766 // peepmatch ( incI_eReg movI );
12767 // peepconstraint ( 0.dst == 1.dst );
12768 // peepreplace ( leaI_eReg_immI( 0.dst 1.src 0.src ) );
12769 // %}
12770 //
12771 // peephole %{
12772 // peepmatch ( decI_eReg movI );
12773 // peepconstraint ( 0.dst == 1.dst );
12774 // peepreplace ( leaI_eReg_immI( 0.dst 1.src 0.src ) );
12775 // %}
12776 //
12777 // peephole %{
12778 // peepmatch ( addI_eReg_imm movI );
12779 // peepconstraint ( 0.dst == 1.dst );
12780 // peepreplace ( leaI_eReg_immI( 0.dst 1.src 0.src ) );
12781 // %}
12782 //
12783 // peephole %{
12784 // peepmatch ( addP_eReg_imm movP );
12785 // peepconstraint ( 0.dst == 1.dst );
12786 // peepreplace ( leaP_eReg_immI( 0.dst 1.src 0.src ) );
12787 // %}
12789 // // Change load of spilled value to only a spill
12790 // instruct storeI(memory mem, eRegI src) %{
12791 // match(Set mem (StoreI mem src));
12792 // %}
12793 //
12794 // instruct loadI(eRegI dst, memory mem) %{
12795 // match(Set dst (LoadI mem));
12796 // %}
12797 //
12798 peephole %{
12799 peepmatch ( loadI storeI );
12800 peepconstraint ( 1.src == 0.dst, 1.mem == 0.mem );
12801 peepreplace ( storeI( 1.mem 1.mem 1.src ) );
12802 %}
12804 //----------SMARTSPILL RULES---------------------------------------------------
12805 // These must follow all instruction definitions as they use the names
12806 // defined in the instructions definitions.