Thu, 03 Jan 2013 16:30:47 -0800
8005544: Use 256bit YMM registers in arraycopy stubs on x86
Summary: Use YMM registers in arraycopy and array_fill stubs.
Reviewed-by: roland, twisti
1 //
2 // Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved.
3 // DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 //
5 // This code is free software; you can redistribute it and/or modify it
6 // under the terms of the GNU General Public License version 2 only, as
7 // published by the Free Software Foundation.
8 //
9 // This code is distributed in the hope that it will be useful, but WITHOUT
10 // ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 // FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 // version 2 for more details (a copy is included in the LICENSE file that
13 // accompanied this code).
14 //
15 // You should have received a copy of the GNU General Public License version
16 // 2 along with this work; if not, write to the Free Software Foundation,
17 // Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 //
19 // Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 // or visit www.oracle.com if you need additional information or have any
21 // questions.
22 //
23 //
25 // X86 Architecture Description File
27 //----------REGISTER DEFINITION BLOCK------------------------------------------
28 // This information is used by the matcher and the register allocator to
29 // describe individual registers and classes of registers within the target
30 // archtecture.
32 register %{
33 //----------Architecture Description Register Definitions----------------------
34 // General Registers
35 // "reg_def" name ( register save type, C convention save type,
36 // ideal register type, encoding );
37 // Register Save Types:
38 //
39 // NS = No-Save: The register allocator assumes that these registers
40 // can be used without saving upon entry to the method, &
41 // that they do not need to be saved at call sites.
42 //
43 // SOC = Save-On-Call: The register allocator assumes that these registers
44 // can be used without saving upon entry to the method,
45 // but that they must be saved at call sites.
46 //
47 // SOE = Save-On-Entry: The register allocator assumes that these registers
48 // must be saved before using them upon entry to the
49 // method, but they do not need to be saved at call
50 // sites.
51 //
52 // AS = Always-Save: The register allocator assumes that these registers
53 // must be saved before using them upon entry to the
54 // method, & that they must be saved at call sites.
55 //
56 // Ideal Register Type is used to determine how to save & restore a
57 // register. Op_RegI will get spilled with LoadI/StoreI, Op_RegP will get
58 // spilled with LoadP/StoreP. If the register supports both, use Op_RegI.
59 //
60 // The encoding number is the actual bit-pattern placed into the opcodes.
62 // General Registers
63 // Previously set EBX, ESI, and EDI as save-on-entry for java code
64 // Turn off SOE in java-code due to frequent use of uncommon-traps.
65 // Now that allocator is better, turn on ESI and EDI as SOE registers.
67 reg_def EBX(SOC, SOE, Op_RegI, 3, rbx->as_VMReg());
68 reg_def ECX(SOC, SOC, Op_RegI, 1, rcx->as_VMReg());
69 reg_def ESI(SOC, SOE, Op_RegI, 6, rsi->as_VMReg());
70 reg_def EDI(SOC, SOE, Op_RegI, 7, rdi->as_VMReg());
71 // now that adapter frames are gone EBP is always saved and restored by the prolog/epilog code
72 reg_def EBP(NS, SOE, Op_RegI, 5, rbp->as_VMReg());
73 reg_def EDX(SOC, SOC, Op_RegI, 2, rdx->as_VMReg());
74 reg_def EAX(SOC, SOC, Op_RegI, 0, rax->as_VMReg());
75 reg_def ESP( NS, NS, Op_RegI, 4, rsp->as_VMReg());
77 // Float registers. We treat TOS/FPR0 special. It is invisible to the
78 // allocator, and only shows up in the encodings.
79 reg_def FPR0L( SOC, SOC, Op_RegF, 0, VMRegImpl::Bad());
80 reg_def FPR0H( SOC, SOC, Op_RegF, 0, VMRegImpl::Bad());
81 // Ok so here's the trick FPR1 is really st(0) except in the midst
82 // of emission of assembly for a machnode. During the emission the fpu stack
83 // is pushed making FPR1 == st(1) temporarily. However at any safepoint
84 // the stack will not have this element so FPR1 == st(0) from the
85 // oopMap viewpoint. This same weirdness with numbering causes
86 // instruction encoding to have to play games with the register
87 // encode to correct for this 0/1 issue. See MachSpillCopyNode::implementation
88 // where it does flt->flt moves to see an example
89 //
90 reg_def FPR1L( SOC, SOC, Op_RegF, 1, as_FloatRegister(0)->as_VMReg());
91 reg_def FPR1H( SOC, SOC, Op_RegF, 1, as_FloatRegister(0)->as_VMReg()->next());
92 reg_def FPR2L( SOC, SOC, Op_RegF, 2, as_FloatRegister(1)->as_VMReg());
93 reg_def FPR2H( SOC, SOC, Op_RegF, 2, as_FloatRegister(1)->as_VMReg()->next());
94 reg_def FPR3L( SOC, SOC, Op_RegF, 3, as_FloatRegister(2)->as_VMReg());
95 reg_def FPR3H( SOC, SOC, Op_RegF, 3, as_FloatRegister(2)->as_VMReg()->next());
96 reg_def FPR4L( SOC, SOC, Op_RegF, 4, as_FloatRegister(3)->as_VMReg());
97 reg_def FPR4H( SOC, SOC, Op_RegF, 4, as_FloatRegister(3)->as_VMReg()->next());
98 reg_def FPR5L( SOC, SOC, Op_RegF, 5, as_FloatRegister(4)->as_VMReg());
99 reg_def FPR5H( SOC, SOC, Op_RegF, 5, as_FloatRegister(4)->as_VMReg()->next());
100 reg_def FPR6L( SOC, SOC, Op_RegF, 6, as_FloatRegister(5)->as_VMReg());
101 reg_def FPR6H( SOC, SOC, Op_RegF, 6, as_FloatRegister(5)->as_VMReg()->next());
102 reg_def FPR7L( SOC, SOC, Op_RegF, 7, as_FloatRegister(6)->as_VMReg());
103 reg_def FPR7H( SOC, SOC, Op_RegF, 7, as_FloatRegister(6)->as_VMReg()->next());
105 // Specify priority of register selection within phases of register
106 // allocation. Highest priority is first. A useful heuristic is to
107 // give registers a low priority when they are required by machine
108 // instructions, like EAX and EDX. Registers which are used as
109 // pairs must fall on an even boundary (witness the FPR#L's in this list).
110 // For the Intel integer registers, the equivalent Long pairs are
111 // EDX:EAX, EBX:ECX, and EDI:EBP.
112 alloc_class chunk0( ECX, EBX, EBP, EDI, EAX, EDX, ESI, ESP,
113 FPR0L, FPR0H, FPR1L, FPR1H, FPR2L, FPR2H,
114 FPR3L, FPR3H, FPR4L, FPR4H, FPR5L, FPR5H,
115 FPR6L, FPR6H, FPR7L, FPR7H );
118 //----------Architecture Description Register Classes--------------------------
119 // Several register classes are automatically defined based upon information in
120 // this architecture description.
121 // 1) reg_class inline_cache_reg ( /* as def'd in frame section */ )
122 // 2) reg_class compiler_method_oop_reg ( /* as def'd in frame section */ )
123 // 2) reg_class interpreter_method_oop_reg ( /* as def'd in frame section */ )
124 // 3) reg_class stack_slots( /* one chunk of stack-based "registers" */ )
125 //
126 // Class for all registers
127 reg_class any_reg(EAX, EDX, EBP, EDI, ESI, ECX, EBX, ESP);
128 // Class for general registers
129 reg_class int_reg(EAX, EDX, EBP, EDI, ESI, ECX, EBX);
130 // Class for general registers which may be used for implicit null checks on win95
131 // Also safe for use by tailjump. We don't want to allocate in rbp,
132 reg_class int_reg_no_rbp(EAX, EDX, EDI, ESI, ECX, EBX);
133 // Class of "X" registers
134 reg_class int_x_reg(EBX, ECX, EDX, EAX);
135 // Class of registers that can appear in an address with no offset.
136 // EBP and ESP require an extra instruction byte for zero offset.
137 // Used in fast-unlock
138 reg_class p_reg(EDX, EDI, ESI, EBX);
139 // Class for general registers not including ECX
140 reg_class ncx_reg(EAX, EDX, EBP, EDI, ESI, EBX);
141 // Class for general registers not including EAX
142 reg_class nax_reg(EDX, EDI, ESI, ECX, EBX);
143 // Class for general registers not including EAX or EBX.
144 reg_class nabx_reg(EDX, EDI, ESI, ECX, EBP);
145 // Class of EAX (for multiply and divide operations)
146 reg_class eax_reg(EAX);
147 // Class of EBX (for atomic add)
148 reg_class ebx_reg(EBX);
149 // Class of ECX (for shift and JCXZ operations and cmpLTMask)
150 reg_class ecx_reg(ECX);
151 // Class of EDX (for multiply and divide operations)
152 reg_class edx_reg(EDX);
153 // Class of EDI (for synchronization)
154 reg_class edi_reg(EDI);
155 // Class of ESI (for synchronization)
156 reg_class esi_reg(ESI);
157 // Singleton class for interpreter's stack pointer
158 reg_class ebp_reg(EBP);
159 // Singleton class for stack pointer
160 reg_class sp_reg(ESP);
161 // Singleton class for instruction pointer
162 // reg_class ip_reg(EIP);
163 // Class of integer register pairs
164 reg_class long_reg( EAX,EDX, ECX,EBX, EBP,EDI );
165 // Class of integer register pairs that aligns with calling convention
166 reg_class eadx_reg( EAX,EDX );
167 reg_class ebcx_reg( ECX,EBX );
168 // Not AX or DX, used in divides
169 reg_class nadx_reg( EBX,ECX,ESI,EDI,EBP );
171 // Floating point registers. Notice FPR0 is not a choice.
172 // FPR0 is not ever allocated; we use clever encodings to fake
173 // a 2-address instructions out of Intels FP stack.
174 reg_class fp_flt_reg( FPR1L,FPR2L,FPR3L,FPR4L,FPR5L,FPR6L,FPR7L );
176 reg_class fp_dbl_reg( FPR1L,FPR1H, FPR2L,FPR2H, FPR3L,FPR3H,
177 FPR4L,FPR4H, FPR5L,FPR5H, FPR6L,FPR6H,
178 FPR7L,FPR7H );
180 reg_class fp_flt_reg0( FPR1L );
181 reg_class fp_dbl_reg0( FPR1L,FPR1H );
182 reg_class fp_dbl_reg1( FPR2L,FPR2H );
183 reg_class fp_dbl_notreg0( FPR2L,FPR2H, FPR3L,FPR3H, FPR4L,FPR4H,
184 FPR5L,FPR5H, FPR6L,FPR6H, FPR7L,FPR7H );
186 %}
189 //----------SOURCE BLOCK-------------------------------------------------------
190 // This is a block of C++ code which provides values, functions, and
191 // definitions necessary in the rest of the architecture description
192 source_hpp %{
193 // Must be visible to the DFA in dfa_x86_32.cpp
194 extern bool is_operand_hi32_zero(Node* n);
195 %}
197 source %{
198 #define RELOC_IMM32 Assembler::imm_operand
199 #define RELOC_DISP32 Assembler::disp32_operand
201 #define __ _masm.
203 // How to find the high register of a Long pair, given the low register
204 #define HIGH_FROM_LOW(x) ((x)+2)
206 // These masks are used to provide 128-bit aligned bitmasks to the XMM
207 // instructions, to allow sign-masking or sign-bit flipping. They allow
208 // fast versions of NegF/NegD and AbsF/AbsD.
210 // Note: 'double' and 'long long' have 32-bits alignment on x86.
211 static jlong* double_quadword(jlong *adr, jlong lo, jlong hi) {
212 // Use the expression (adr)&(~0xF) to provide 128-bits aligned address
213 // of 128-bits operands for SSE instructions.
214 jlong *operand = (jlong*)(((uintptr_t)adr)&((uintptr_t)(~0xF)));
215 // Store the value to a 128-bits operand.
216 operand[0] = lo;
217 operand[1] = hi;
218 return operand;
219 }
221 // Buffer for 128-bits masks used by SSE instructions.
222 static jlong fp_signmask_pool[(4+1)*2]; // 4*128bits(data) + 128bits(alignment)
224 // Static initialization during VM startup.
225 static jlong *float_signmask_pool = double_quadword(&fp_signmask_pool[1*2], CONST64(0x7FFFFFFF7FFFFFFF), CONST64(0x7FFFFFFF7FFFFFFF));
226 static jlong *double_signmask_pool = double_quadword(&fp_signmask_pool[2*2], CONST64(0x7FFFFFFFFFFFFFFF), CONST64(0x7FFFFFFFFFFFFFFF));
227 static jlong *float_signflip_pool = double_quadword(&fp_signmask_pool[3*2], CONST64(0x8000000080000000), CONST64(0x8000000080000000));
228 static jlong *double_signflip_pool = double_quadword(&fp_signmask_pool[4*2], CONST64(0x8000000000000000), CONST64(0x8000000000000000));
230 // Offset hacking within calls.
231 static int pre_call_FPU_size() {
232 if (Compile::current()->in_24_bit_fp_mode())
233 return 6; // fldcw
234 return 0;
235 }
237 static int preserve_SP_size() {
238 return 2; // op, rm(reg/reg)
239 }
241 // !!!!! Special hack to get all type of calls to specify the byte offset
242 // from the start of the call to the point where the return address
243 // will point.
244 int MachCallStaticJavaNode::ret_addr_offset() {
245 int offset = 5 + pre_call_FPU_size(); // 5 bytes from start of call to where return address points
246 if (_method_handle_invoke)
247 offset += preserve_SP_size();
248 return offset;
249 }
251 int MachCallDynamicJavaNode::ret_addr_offset() {
252 return 10 + pre_call_FPU_size(); // 10 bytes from start of call to where return address points
253 }
255 static int sizeof_FFree_Float_Stack_All = -1;
257 int MachCallRuntimeNode::ret_addr_offset() {
258 assert(sizeof_FFree_Float_Stack_All != -1, "must have been emitted already");
259 return sizeof_FFree_Float_Stack_All + 5 + pre_call_FPU_size();
260 }
262 // Indicate if the safepoint node needs the polling page as an input.
263 // Since x86 does have absolute addressing, it doesn't.
264 bool SafePointNode::needs_polling_address_input() {
265 return false;
266 }
268 //
269 // Compute padding required for nodes which need alignment
270 //
272 // The address of the call instruction needs to be 4-byte aligned to
273 // ensure that it does not span a cache line so that it can be patched.
274 int CallStaticJavaDirectNode::compute_padding(int current_offset) const {
275 current_offset += pre_call_FPU_size(); // skip fldcw, if any
276 current_offset += 1; // skip call opcode byte
277 return round_to(current_offset, alignment_required()) - current_offset;
278 }
280 // The address of the call instruction needs to be 4-byte aligned to
281 // ensure that it does not span a cache line so that it can be patched.
282 int CallStaticJavaHandleNode::compute_padding(int current_offset) const {
283 current_offset += pre_call_FPU_size(); // skip fldcw, if any
284 current_offset += preserve_SP_size(); // skip mov rbp, rsp
285 current_offset += 1; // skip call opcode byte
286 return round_to(current_offset, alignment_required()) - current_offset;
287 }
289 // The address of the call instruction needs to be 4-byte aligned to
290 // ensure that it does not span a cache line so that it can be patched.
291 int CallDynamicJavaDirectNode::compute_padding(int current_offset) const {
292 current_offset += pre_call_FPU_size(); // skip fldcw, if any
293 current_offset += 5; // skip MOV instruction
294 current_offset += 1; // skip call opcode byte
295 return round_to(current_offset, alignment_required()) - current_offset;
296 }
298 // EMIT_RM()
299 void emit_rm(CodeBuffer &cbuf, int f1, int f2, int f3) {
300 unsigned char c = (unsigned char)((f1 << 6) | (f2 << 3) | f3);
301 cbuf.insts()->emit_int8(c);
302 }
304 // EMIT_CC()
305 void emit_cc(CodeBuffer &cbuf, int f1, int f2) {
306 unsigned char c = (unsigned char)( f1 | f2 );
307 cbuf.insts()->emit_int8(c);
308 }
310 // EMIT_OPCODE()
311 void emit_opcode(CodeBuffer &cbuf, int code) {
312 cbuf.insts()->emit_int8((unsigned char) code);
313 }
315 // EMIT_OPCODE() w/ relocation information
316 void emit_opcode(CodeBuffer &cbuf, int code, relocInfo::relocType reloc, int offset = 0) {
317 cbuf.relocate(cbuf.insts_mark() + offset, reloc);
318 emit_opcode(cbuf, code);
319 }
321 // EMIT_D8()
322 void emit_d8(CodeBuffer &cbuf, int d8) {
323 cbuf.insts()->emit_int8((unsigned char) d8);
324 }
326 // EMIT_D16()
327 void emit_d16(CodeBuffer &cbuf, int d16) {
328 cbuf.insts()->emit_int16(d16);
329 }
331 // EMIT_D32()
332 void emit_d32(CodeBuffer &cbuf, int d32) {
333 cbuf.insts()->emit_int32(d32);
334 }
336 // emit 32 bit value and construct relocation entry from relocInfo::relocType
337 void emit_d32_reloc(CodeBuffer &cbuf, int d32, relocInfo::relocType reloc,
338 int format) {
339 cbuf.relocate(cbuf.insts_mark(), reloc, format);
340 cbuf.insts()->emit_int32(d32);
341 }
343 // emit 32 bit value and construct relocation entry from RelocationHolder
344 void emit_d32_reloc(CodeBuffer &cbuf, int d32, RelocationHolder const& rspec,
345 int format) {
346 #ifdef ASSERT
347 if (rspec.reloc()->type() == relocInfo::oop_type && d32 != 0 && d32 != (int)Universe::non_oop_word()) {
348 assert(oop(d32)->is_oop() && (ScavengeRootsInCode || !oop(d32)->is_scavengable()), "cannot embed scavengable oops in code");
349 }
350 #endif
351 cbuf.relocate(cbuf.insts_mark(), rspec, format);
352 cbuf.insts()->emit_int32(d32);
353 }
355 // Access stack slot for load or store
356 void store_to_stackslot(CodeBuffer &cbuf, int opcode, int rm_field, int disp) {
357 emit_opcode( cbuf, opcode ); // (e.g., FILD [ESP+src])
358 if( -128 <= disp && disp <= 127 ) {
359 emit_rm( cbuf, 0x01, rm_field, ESP_enc ); // R/M byte
360 emit_rm( cbuf, 0x00, ESP_enc, ESP_enc); // SIB byte
361 emit_d8 (cbuf, disp); // Displacement // R/M byte
362 } else {
363 emit_rm( cbuf, 0x02, rm_field, ESP_enc ); // R/M byte
364 emit_rm( cbuf, 0x00, ESP_enc, ESP_enc); // SIB byte
365 emit_d32(cbuf, disp); // Displacement // R/M byte
366 }
367 }
369 // rRegI ereg, memory mem) %{ // emit_reg_mem
370 void encode_RegMem( CodeBuffer &cbuf, int reg_encoding, int base, int index, int scale, int displace, relocInfo::relocType disp_reloc ) {
371 // There is no index & no scale, use form without SIB byte
372 if ((index == 0x4) &&
373 (scale == 0) && (base != ESP_enc)) {
374 // If no displacement, mode is 0x0; unless base is [EBP]
375 if ( (displace == 0) && (base != EBP_enc) ) {
376 emit_rm(cbuf, 0x0, reg_encoding, base);
377 }
378 else { // If 8-bit displacement, mode 0x1
379 if ((displace >= -128) && (displace <= 127)
380 && (disp_reloc == relocInfo::none) ) {
381 emit_rm(cbuf, 0x1, reg_encoding, base);
382 emit_d8(cbuf, displace);
383 }
384 else { // If 32-bit displacement
385 if (base == -1) { // Special flag for absolute address
386 emit_rm(cbuf, 0x0, reg_encoding, 0x5);
387 // (manual lies; no SIB needed here)
388 if ( disp_reloc != relocInfo::none ) {
389 emit_d32_reloc(cbuf, displace, disp_reloc, 1);
390 } else {
391 emit_d32 (cbuf, displace);
392 }
393 }
394 else { // Normal base + offset
395 emit_rm(cbuf, 0x2, reg_encoding, base);
396 if ( disp_reloc != relocInfo::none ) {
397 emit_d32_reloc(cbuf, displace, disp_reloc, 1);
398 } else {
399 emit_d32 (cbuf, displace);
400 }
401 }
402 }
403 }
404 }
405 else { // Else, encode with the SIB byte
406 // If no displacement, mode is 0x0; unless base is [EBP]
407 if (displace == 0 && (base != EBP_enc)) { // If no displacement
408 emit_rm(cbuf, 0x0, reg_encoding, 0x4);
409 emit_rm(cbuf, scale, index, base);
410 }
411 else { // If 8-bit displacement, mode 0x1
412 if ((displace >= -128) && (displace <= 127)
413 && (disp_reloc == relocInfo::none) ) {
414 emit_rm(cbuf, 0x1, reg_encoding, 0x4);
415 emit_rm(cbuf, scale, index, base);
416 emit_d8(cbuf, displace);
417 }
418 else { // If 32-bit displacement
419 if (base == 0x04 ) {
420 emit_rm(cbuf, 0x2, reg_encoding, 0x4);
421 emit_rm(cbuf, scale, index, 0x04);
422 } else {
423 emit_rm(cbuf, 0x2, reg_encoding, 0x4);
424 emit_rm(cbuf, scale, index, base);
425 }
426 if ( disp_reloc != relocInfo::none ) {
427 emit_d32_reloc(cbuf, displace, disp_reloc, 1);
428 } else {
429 emit_d32 (cbuf, displace);
430 }
431 }
432 }
433 }
434 }
437 void encode_Copy( CodeBuffer &cbuf, int dst_encoding, int src_encoding ) {
438 if( dst_encoding == src_encoding ) {
439 // reg-reg copy, use an empty encoding
440 } else {
441 emit_opcode( cbuf, 0x8B );
442 emit_rm(cbuf, 0x3, dst_encoding, src_encoding );
443 }
444 }
446 void emit_cmpfp_fixup(MacroAssembler& _masm) {
447 Label exit;
448 __ jccb(Assembler::noParity, exit);
449 __ pushf();
450 //
451 // comiss/ucomiss instructions set ZF,PF,CF flags and
452 // zero OF,AF,SF for NaN values.
453 // Fixup flags by zeroing ZF,PF so that compare of NaN
454 // values returns 'less than' result (CF is set).
455 // Leave the rest of flags unchanged.
456 //
457 // 7 6 5 4 3 2 1 0
458 // |S|Z|r|A|r|P|r|C| (r - reserved bit)
459 // 0 0 1 0 1 0 1 1 (0x2B)
460 //
461 __ andl(Address(rsp, 0), 0xffffff2b);
462 __ popf();
463 __ bind(exit);
464 }
466 void emit_cmpfp3(MacroAssembler& _masm, Register dst) {
467 Label done;
468 __ movl(dst, -1);
469 __ jcc(Assembler::parity, done);
470 __ jcc(Assembler::below, done);
471 __ setb(Assembler::notEqual, dst);
472 __ movzbl(dst, dst);
473 __ bind(done);
474 }
477 //=============================================================================
478 const RegMask& MachConstantBaseNode::_out_RegMask = RegMask::Empty;
480 int Compile::ConstantTable::calculate_table_base_offset() const {
481 return 0; // absolute addressing, no offset
482 }
484 void MachConstantBaseNode::emit(CodeBuffer& cbuf, PhaseRegAlloc* ra_) const {
485 // Empty encoding
486 }
488 uint MachConstantBaseNode::size(PhaseRegAlloc* ra_) const {
489 return 0;
490 }
492 #ifndef PRODUCT
493 void MachConstantBaseNode::format(PhaseRegAlloc* ra_, outputStream* st) const {
494 st->print("# MachConstantBaseNode (empty encoding)");
495 }
496 #endif
499 //=============================================================================
500 #ifndef PRODUCT
501 void MachPrologNode::format(PhaseRegAlloc* ra_, outputStream* st) const {
502 Compile* C = ra_->C;
504 int framesize = C->frame_slots() << LogBytesPerInt;
505 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
506 // Remove wordSize for return addr which is already pushed.
507 framesize -= wordSize;
509 if (C->need_stack_bang(framesize)) {
510 framesize -= wordSize;
511 st->print("# stack bang");
512 st->print("\n\t");
513 st->print("PUSH EBP\t# Save EBP");
514 if (framesize) {
515 st->print("\n\t");
516 st->print("SUB ESP, #%d\t# Create frame",framesize);
517 }
518 } else {
519 st->print("SUB ESP, #%d\t# Create frame",framesize);
520 st->print("\n\t");
521 framesize -= wordSize;
522 st->print("MOV [ESP + #%d], EBP\t# Save EBP",framesize);
523 }
525 if (VerifyStackAtCalls) {
526 st->print("\n\t");
527 framesize -= wordSize;
528 st->print("MOV [ESP + #%d], 0xBADB100D\t# Majik cookie for stack depth check",framesize);
529 }
531 if( C->in_24_bit_fp_mode() ) {
532 st->print("\n\t");
533 st->print("FLDCW \t# load 24 bit fpu control word");
534 }
535 if (UseSSE >= 2 && VerifyFPU) {
536 st->print("\n\t");
537 st->print("# verify FPU stack (must be clean on entry)");
538 }
540 #ifdef ASSERT
541 if (VerifyStackAtCalls) {
542 st->print("\n\t");
543 st->print("# stack alignment check");
544 }
545 #endif
546 st->cr();
547 }
548 #endif
551 void MachPrologNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
552 Compile* C = ra_->C;
553 MacroAssembler _masm(&cbuf);
555 int framesize = C->frame_slots() << LogBytesPerInt;
557 __ verified_entry(framesize, C->need_stack_bang(framesize), C->in_24_bit_fp_mode());
559 C->set_frame_complete(cbuf.insts_size());
561 if (C->has_mach_constant_base_node()) {
562 // NOTE: We set the table base offset here because users might be
563 // emitted before MachConstantBaseNode.
564 Compile::ConstantTable& constant_table = C->constant_table();
565 constant_table.set_table_base_offset(constant_table.calculate_table_base_offset());
566 }
567 }
569 uint MachPrologNode::size(PhaseRegAlloc *ra_) const {
570 return MachNode::size(ra_); // too many variables; just compute it the hard way
571 }
573 int MachPrologNode::reloc() const {
574 return 0; // a large enough number
575 }
577 //=============================================================================
578 #ifndef PRODUCT
579 void MachEpilogNode::format( PhaseRegAlloc *ra_, outputStream* st ) const {
580 Compile *C = ra_->C;
581 int framesize = C->frame_slots() << LogBytesPerInt;
582 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
583 // Remove two words for return addr and rbp,
584 framesize -= 2*wordSize;
586 if( C->in_24_bit_fp_mode() ) {
587 st->print("FLDCW standard control word");
588 st->cr(); st->print("\t");
589 }
590 if( framesize ) {
591 st->print("ADD ESP,%d\t# Destroy frame",framesize);
592 st->cr(); st->print("\t");
593 }
594 st->print_cr("POPL EBP"); st->print("\t");
595 if( do_polling() && C->is_method_compilation() ) {
596 st->print("TEST PollPage,EAX\t! Poll Safepoint");
597 st->cr(); st->print("\t");
598 }
599 }
600 #endif
602 void MachEpilogNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
603 Compile *C = ra_->C;
605 // If method set FPU control word, restore to standard control word
606 if( C->in_24_bit_fp_mode() ) {
607 MacroAssembler masm(&cbuf);
608 masm.fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
609 }
611 int framesize = C->frame_slots() << LogBytesPerInt;
612 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
613 // Remove two words for return addr and rbp,
614 framesize -= 2*wordSize;
616 // Note that VerifyStackAtCalls' Majik cookie does not change the frame size popped here
618 if( framesize >= 128 ) {
619 emit_opcode(cbuf, 0x81); // add SP, #framesize
620 emit_rm(cbuf, 0x3, 0x00, ESP_enc);
621 emit_d32(cbuf, framesize);
622 }
623 else if( framesize ) {
624 emit_opcode(cbuf, 0x83); // add SP, #framesize
625 emit_rm(cbuf, 0x3, 0x00, ESP_enc);
626 emit_d8(cbuf, framesize);
627 }
629 emit_opcode(cbuf, 0x58 | EBP_enc);
631 if( do_polling() && C->is_method_compilation() ) {
632 cbuf.relocate(cbuf.insts_end(), relocInfo::poll_return_type, 0);
633 emit_opcode(cbuf,0x85);
634 emit_rm(cbuf, 0x0, EAX_enc, 0x5); // EAX
635 emit_d32(cbuf, (intptr_t)os::get_polling_page());
636 }
637 }
639 uint MachEpilogNode::size(PhaseRegAlloc *ra_) const {
640 Compile *C = ra_->C;
641 // If method set FPU control word, restore to standard control word
642 int size = C->in_24_bit_fp_mode() ? 6 : 0;
643 if( do_polling() && C->is_method_compilation() ) size += 6;
645 int framesize = C->frame_slots() << LogBytesPerInt;
646 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
647 // Remove two words for return addr and rbp,
648 framesize -= 2*wordSize;
650 size++; // popl rbp,
652 if( framesize >= 128 ) {
653 size += 6;
654 } else {
655 size += framesize ? 3 : 0;
656 }
657 return size;
658 }
660 int MachEpilogNode::reloc() const {
661 return 0; // a large enough number
662 }
664 const Pipeline * MachEpilogNode::pipeline() const {
665 return MachNode::pipeline_class();
666 }
668 int MachEpilogNode::safepoint_offset() const { return 0; }
670 //=============================================================================
672 enum RC { rc_bad, rc_int, rc_float, rc_xmm, rc_stack };
673 static enum RC rc_class( OptoReg::Name reg ) {
675 if( !OptoReg::is_valid(reg) ) return rc_bad;
676 if (OptoReg::is_stack(reg)) return rc_stack;
678 VMReg r = OptoReg::as_VMReg(reg);
679 if (r->is_Register()) return rc_int;
680 if (r->is_FloatRegister()) {
681 assert(UseSSE < 2, "shouldn't be used in SSE2+ mode");
682 return rc_float;
683 }
684 assert(r->is_XMMRegister(), "must be");
685 return rc_xmm;
686 }
688 static int impl_helper( CodeBuffer *cbuf, bool do_size, bool is_load, int offset, int reg,
689 int opcode, const char *op_str, int size, outputStream* st ) {
690 if( cbuf ) {
691 emit_opcode (*cbuf, opcode );
692 encode_RegMem(*cbuf, Matcher::_regEncode[reg], ESP_enc, 0x4, 0, offset, relocInfo::none);
693 #ifndef PRODUCT
694 } else if( !do_size ) {
695 if( size != 0 ) st->print("\n\t");
696 if( opcode == 0x8B || opcode == 0x89 ) { // MOV
697 if( is_load ) st->print("%s %s,[ESP + #%d]",op_str,Matcher::regName[reg],offset);
698 else st->print("%s [ESP + #%d],%s",op_str,offset,Matcher::regName[reg]);
699 } else { // FLD, FST, PUSH, POP
700 st->print("%s [ESP + #%d]",op_str,offset);
701 }
702 #endif
703 }
704 int offset_size = (offset == 0) ? 0 : ((offset <= 127) ? 1 : 4);
705 return size+3+offset_size;
706 }
708 // Helper for XMM registers. Extra opcode bits, limited syntax.
709 static int impl_x_helper( CodeBuffer *cbuf, bool do_size, bool is_load,
710 int offset, int reg_lo, int reg_hi, int size, outputStream* st ) {
711 if (cbuf) {
712 MacroAssembler _masm(cbuf);
713 if (reg_lo+1 == reg_hi) { // double move?
714 if (is_load) {
715 __ movdbl(as_XMMRegister(Matcher::_regEncode[reg_lo]), Address(rsp, offset));
716 } else {
717 __ movdbl(Address(rsp, offset), as_XMMRegister(Matcher::_regEncode[reg_lo]));
718 }
719 } else {
720 if (is_load) {
721 __ movflt(as_XMMRegister(Matcher::_regEncode[reg_lo]), Address(rsp, offset));
722 } else {
723 __ movflt(Address(rsp, offset), as_XMMRegister(Matcher::_regEncode[reg_lo]));
724 }
725 }
726 #ifndef PRODUCT
727 } else if (!do_size) {
728 if (size != 0) st->print("\n\t");
729 if (reg_lo+1 == reg_hi) { // double move?
730 if (is_load) st->print("%s %s,[ESP + #%d]",
731 UseXmmLoadAndClearUpper ? "MOVSD " : "MOVLPD",
732 Matcher::regName[reg_lo], offset);
733 else st->print("MOVSD [ESP + #%d],%s",
734 offset, Matcher::regName[reg_lo]);
735 } else {
736 if (is_load) st->print("MOVSS %s,[ESP + #%d]",
737 Matcher::regName[reg_lo], offset);
738 else st->print("MOVSS [ESP + #%d],%s",
739 offset, Matcher::regName[reg_lo]);
740 }
741 #endif
742 }
743 int offset_size = (offset == 0) ? 0 : ((offset <= 127) ? 1 : 4);
744 // VEX_2bytes prefix is used if UseAVX > 0, so it takes the same 2 bytes as SIMD prefix.
745 return size+5+offset_size;
746 }
749 static int impl_movx_helper( CodeBuffer *cbuf, bool do_size, int src_lo, int dst_lo,
750 int src_hi, int dst_hi, int size, outputStream* st ) {
751 if (cbuf) {
752 MacroAssembler _masm(cbuf);
753 if (src_lo+1 == src_hi && dst_lo+1 == dst_hi) { // double move?
754 __ movdbl(as_XMMRegister(Matcher::_regEncode[dst_lo]),
755 as_XMMRegister(Matcher::_regEncode[src_lo]));
756 } else {
757 __ movflt(as_XMMRegister(Matcher::_regEncode[dst_lo]),
758 as_XMMRegister(Matcher::_regEncode[src_lo]));
759 }
760 #ifndef PRODUCT
761 } else if (!do_size) {
762 if (size != 0) st->print("\n\t");
763 if (UseXmmRegToRegMoveAll) {//Use movaps,movapd to move between xmm registers
764 if (src_lo+1 == src_hi && dst_lo+1 == dst_hi) { // double move?
765 st->print("MOVAPD %s,%s",Matcher::regName[dst_lo],Matcher::regName[src_lo]);
766 } else {
767 st->print("MOVAPS %s,%s",Matcher::regName[dst_lo],Matcher::regName[src_lo]);
768 }
769 } else {
770 if( src_lo+1 == src_hi && dst_lo+1 == dst_hi ) { // double move?
771 st->print("MOVSD %s,%s",Matcher::regName[dst_lo],Matcher::regName[src_lo]);
772 } else {
773 st->print("MOVSS %s,%s",Matcher::regName[dst_lo],Matcher::regName[src_lo]);
774 }
775 }
776 #endif
777 }
778 // VEX_2bytes prefix is used if UseAVX > 0, and it takes the same 2 bytes as SIMD prefix.
779 // Only MOVAPS SSE prefix uses 1 byte.
780 int sz = 4;
781 if (!(src_lo+1 == src_hi && dst_lo+1 == dst_hi) &&
782 UseXmmRegToRegMoveAll && (UseAVX == 0)) sz = 3;
783 return size + sz;
784 }
786 static int impl_movgpr2x_helper( CodeBuffer *cbuf, bool do_size, int src_lo, int dst_lo,
787 int src_hi, int dst_hi, int size, outputStream* st ) {
788 // 32-bit
789 if (cbuf) {
790 MacroAssembler _masm(cbuf);
791 __ movdl(as_XMMRegister(Matcher::_regEncode[dst_lo]),
792 as_Register(Matcher::_regEncode[src_lo]));
793 #ifndef PRODUCT
794 } else if (!do_size) {
795 st->print("movdl %s, %s\t# spill", Matcher::regName[dst_lo], Matcher::regName[src_lo]);
796 #endif
797 }
798 return 4;
799 }
802 static int impl_movx2gpr_helper( CodeBuffer *cbuf, bool do_size, int src_lo, int dst_lo,
803 int src_hi, int dst_hi, int size, outputStream* st ) {
804 // 32-bit
805 if (cbuf) {
806 MacroAssembler _masm(cbuf);
807 __ movdl(as_Register(Matcher::_regEncode[dst_lo]),
808 as_XMMRegister(Matcher::_regEncode[src_lo]));
809 #ifndef PRODUCT
810 } else if (!do_size) {
811 st->print("movdl %s, %s\t# spill", Matcher::regName[dst_lo], Matcher::regName[src_lo]);
812 #endif
813 }
814 return 4;
815 }
817 static int impl_mov_helper( CodeBuffer *cbuf, bool do_size, int src, int dst, int size, outputStream* st ) {
818 if( cbuf ) {
819 emit_opcode(*cbuf, 0x8B );
820 emit_rm (*cbuf, 0x3, Matcher::_regEncode[dst], Matcher::_regEncode[src] );
821 #ifndef PRODUCT
822 } else if( !do_size ) {
823 if( size != 0 ) st->print("\n\t");
824 st->print("MOV %s,%s",Matcher::regName[dst],Matcher::regName[src]);
825 #endif
826 }
827 return size+2;
828 }
830 static int impl_fp_store_helper( CodeBuffer *cbuf, bool do_size, int src_lo, int src_hi, int dst_lo, int dst_hi,
831 int offset, int size, outputStream* st ) {
832 if( src_lo != FPR1L_num ) { // Move value to top of FP stack, if not already there
833 if( cbuf ) {
834 emit_opcode( *cbuf, 0xD9 ); // FLD (i.e., push it)
835 emit_d8( *cbuf, 0xC0-1+Matcher::_regEncode[src_lo] );
836 #ifndef PRODUCT
837 } else if( !do_size ) {
838 if( size != 0 ) st->print("\n\t");
839 st->print("FLD %s",Matcher::regName[src_lo]);
840 #endif
841 }
842 size += 2;
843 }
845 int st_op = (src_lo != FPR1L_num) ? EBX_num /*store & pop*/ : EDX_num /*store no pop*/;
846 const char *op_str;
847 int op;
848 if( src_lo+1 == src_hi && dst_lo+1 == dst_hi ) { // double store?
849 op_str = (src_lo != FPR1L_num) ? "FSTP_D" : "FST_D ";
850 op = 0xDD;
851 } else { // 32-bit store
852 op_str = (src_lo != FPR1L_num) ? "FSTP_S" : "FST_S ";
853 op = 0xD9;
854 assert( !OptoReg::is_valid(src_hi) && !OptoReg::is_valid(dst_hi), "no non-adjacent float-stores" );
855 }
857 return impl_helper(cbuf,do_size,false,offset,st_op,op,op_str,size, st);
858 }
860 // Next two methods are shared by 32- and 64-bit VM. They are defined in x86.ad.
861 static int vec_mov_helper(CodeBuffer *cbuf, bool do_size, int src_lo, int dst_lo,
862 int src_hi, int dst_hi, uint ireg, outputStream* st);
864 static int vec_spill_helper(CodeBuffer *cbuf, bool do_size, bool is_load,
865 int stack_offset, int reg, uint ireg, outputStream* st);
867 static int vec_stack_to_stack_helper(CodeBuffer *cbuf, bool do_size, int src_offset,
868 int dst_offset, uint ireg, outputStream* st) {
869 int calc_size = 0;
870 int src_offset_size = (src_offset == 0) ? 0 : ((src_offset < 0x80) ? 1 : 4);
871 int dst_offset_size = (dst_offset == 0) ? 0 : ((dst_offset < 0x80) ? 1 : 4);
872 switch (ireg) {
873 case Op_VecS:
874 calc_size = 3+src_offset_size + 3+dst_offset_size;
875 break;
876 case Op_VecD:
877 calc_size = 3+src_offset_size + 3+dst_offset_size;
878 src_offset += 4;
879 dst_offset += 4;
880 src_offset_size = (src_offset == 0) ? 0 : ((src_offset < 0x80) ? 1 : 4);
881 dst_offset_size = (dst_offset == 0) ? 0 : ((dst_offset < 0x80) ? 1 : 4);
882 calc_size += 3+src_offset_size + 3+dst_offset_size;
883 break;
884 case Op_VecX:
885 calc_size = 6 + 6 + 5+src_offset_size + 5+dst_offset_size;
886 break;
887 case Op_VecY:
888 calc_size = 6 + 6 + 5+src_offset_size + 5+dst_offset_size;
889 break;
890 default:
891 ShouldNotReachHere();
892 }
893 if (cbuf) {
894 MacroAssembler _masm(cbuf);
895 int offset = __ offset();
896 switch (ireg) {
897 case Op_VecS:
898 __ pushl(Address(rsp, src_offset));
899 __ popl (Address(rsp, dst_offset));
900 break;
901 case Op_VecD:
902 __ pushl(Address(rsp, src_offset));
903 __ popl (Address(rsp, dst_offset));
904 __ pushl(Address(rsp, src_offset+4));
905 __ popl (Address(rsp, dst_offset+4));
906 break;
907 case Op_VecX:
908 __ movdqu(Address(rsp, -16), xmm0);
909 __ movdqu(xmm0, Address(rsp, src_offset));
910 __ movdqu(Address(rsp, dst_offset), xmm0);
911 __ movdqu(xmm0, Address(rsp, -16));
912 break;
913 case Op_VecY:
914 __ vmovdqu(Address(rsp, -32), xmm0);
915 __ vmovdqu(xmm0, Address(rsp, src_offset));
916 __ vmovdqu(Address(rsp, dst_offset), xmm0);
917 __ vmovdqu(xmm0, Address(rsp, -32));
918 break;
919 default:
920 ShouldNotReachHere();
921 }
922 int size = __ offset() - offset;
923 assert(size == calc_size, "incorrect size calculattion");
924 return size;
925 #ifndef PRODUCT
926 } else if (!do_size) {
927 switch (ireg) {
928 case Op_VecS:
929 st->print("pushl [rsp + #%d]\t# 32-bit mem-mem spill\n\t"
930 "popl [rsp + #%d]",
931 src_offset, dst_offset);
932 break;
933 case Op_VecD:
934 st->print("pushl [rsp + #%d]\t# 64-bit mem-mem spill\n\t"
935 "popq [rsp + #%d]\n\t"
936 "pushl [rsp + #%d]\n\t"
937 "popq [rsp + #%d]",
938 src_offset, dst_offset, src_offset+4, dst_offset+4);
939 break;
940 case Op_VecX:
941 st->print("movdqu [rsp - #16], xmm0\t# 128-bit mem-mem spill\n\t"
942 "movdqu xmm0, [rsp + #%d]\n\t"
943 "movdqu [rsp + #%d], xmm0\n\t"
944 "movdqu xmm0, [rsp - #16]",
945 src_offset, dst_offset);
946 break;
947 case Op_VecY:
948 st->print("vmovdqu [rsp - #32], xmm0\t# 256-bit mem-mem spill\n\t"
949 "vmovdqu xmm0, [rsp + #%d]\n\t"
950 "vmovdqu [rsp + #%d], xmm0\n\t"
951 "vmovdqu xmm0, [rsp - #32]",
952 src_offset, dst_offset);
953 break;
954 default:
955 ShouldNotReachHere();
956 }
957 #endif
958 }
959 return calc_size;
960 }
962 uint MachSpillCopyNode::implementation( CodeBuffer *cbuf, PhaseRegAlloc *ra_, bool do_size, outputStream* st ) const {
963 // Get registers to move
964 OptoReg::Name src_second = ra_->get_reg_second(in(1));
965 OptoReg::Name src_first = ra_->get_reg_first(in(1));
966 OptoReg::Name dst_second = ra_->get_reg_second(this );
967 OptoReg::Name dst_first = ra_->get_reg_first(this );
969 enum RC src_second_rc = rc_class(src_second);
970 enum RC src_first_rc = rc_class(src_first);
971 enum RC dst_second_rc = rc_class(dst_second);
972 enum RC dst_first_rc = rc_class(dst_first);
974 assert( OptoReg::is_valid(src_first) && OptoReg::is_valid(dst_first), "must move at least 1 register" );
976 // Generate spill code!
977 int size = 0;
979 if( src_first == dst_first && src_second == dst_second )
980 return size; // Self copy, no move
982 if (bottom_type()->isa_vect() != NULL) {
983 uint ireg = ideal_reg();
984 assert((src_first_rc != rc_int && dst_first_rc != rc_int), "sanity");
985 assert((src_first_rc != rc_float && dst_first_rc != rc_float), "sanity");
986 assert((ireg == Op_VecS || ireg == Op_VecD || ireg == Op_VecX || ireg == Op_VecY), "sanity");
987 if( src_first_rc == rc_stack && dst_first_rc == rc_stack ) {
988 // mem -> mem
989 int src_offset = ra_->reg2offset(src_first);
990 int dst_offset = ra_->reg2offset(dst_first);
991 return vec_stack_to_stack_helper(cbuf, do_size, src_offset, dst_offset, ireg, st);
992 } else if (src_first_rc == rc_xmm && dst_first_rc == rc_xmm ) {
993 return vec_mov_helper(cbuf, do_size, src_first, dst_first, src_second, dst_second, ireg, st);
994 } else if (src_first_rc == rc_xmm && dst_first_rc == rc_stack ) {
995 int stack_offset = ra_->reg2offset(dst_first);
996 return vec_spill_helper(cbuf, do_size, false, stack_offset, src_first, ireg, st);
997 } else if (src_first_rc == rc_stack && dst_first_rc == rc_xmm ) {
998 int stack_offset = ra_->reg2offset(src_first);
999 return vec_spill_helper(cbuf, do_size, true, stack_offset, dst_first, ireg, st);
1000 } else {
1001 ShouldNotReachHere();
1002 }
1003 }
1005 // --------------------------------------
1006 // Check for mem-mem move. push/pop to move.
1007 if( src_first_rc == rc_stack && dst_first_rc == rc_stack ) {
1008 if( src_second == dst_first ) { // overlapping stack copy ranges
1009 assert( src_second_rc == rc_stack && dst_second_rc == rc_stack, "we only expect a stk-stk copy here" );
1010 size = impl_helper(cbuf,do_size,true ,ra_->reg2offset(src_second),ESI_num,0xFF,"PUSH ",size, st);
1011 size = impl_helper(cbuf,do_size,false,ra_->reg2offset(dst_second),EAX_num,0x8F,"POP ",size, st);
1012 src_second_rc = dst_second_rc = rc_bad; // flag as already moved the second bits
1013 }
1014 // move low bits
1015 size = impl_helper(cbuf,do_size,true ,ra_->reg2offset(src_first),ESI_num,0xFF,"PUSH ",size, st);
1016 size = impl_helper(cbuf,do_size,false,ra_->reg2offset(dst_first),EAX_num,0x8F,"POP ",size, st);
1017 if( src_second_rc == rc_stack && dst_second_rc == rc_stack ) { // mov second bits
1018 size = impl_helper(cbuf,do_size,true ,ra_->reg2offset(src_second),ESI_num,0xFF,"PUSH ",size, st);
1019 size = impl_helper(cbuf,do_size,false,ra_->reg2offset(dst_second),EAX_num,0x8F,"POP ",size, st);
1020 }
1021 return size;
1022 }
1024 // --------------------------------------
1025 // Check for integer reg-reg copy
1026 if( src_first_rc == rc_int && dst_first_rc == rc_int )
1027 size = impl_mov_helper(cbuf,do_size,src_first,dst_first,size, st);
1029 // Check for integer store
1030 if( src_first_rc == rc_int && dst_first_rc == rc_stack )
1031 size = impl_helper(cbuf,do_size,false,ra_->reg2offset(dst_first),src_first,0x89,"MOV ",size, st);
1033 // Check for integer load
1034 if( dst_first_rc == rc_int && src_first_rc == rc_stack )
1035 size = impl_helper(cbuf,do_size,true ,ra_->reg2offset(src_first),dst_first,0x8B,"MOV ",size, st);
1037 // Check for integer reg-xmm reg copy
1038 if( src_first_rc == rc_int && dst_first_rc == rc_xmm ) {
1039 assert( (src_second_rc == rc_bad && dst_second_rc == rc_bad),
1040 "no 64 bit integer-float reg moves" );
1041 return impl_movgpr2x_helper(cbuf,do_size,src_first,dst_first,src_second, dst_second, size, st);
1042 }
1043 // --------------------------------------
1044 // Check for float reg-reg copy
1045 if( src_first_rc == rc_float && dst_first_rc == rc_float ) {
1046 assert( (src_second_rc == rc_bad && dst_second_rc == rc_bad) ||
1047 (src_first+1 == src_second && dst_first+1 == dst_second), "no non-adjacent float-moves" );
1048 if( cbuf ) {
1050 // Note the mucking with the register encode to compensate for the 0/1
1051 // indexing issue mentioned in a comment in the reg_def sections
1052 // for FPR registers many lines above here.
1054 if( src_first != FPR1L_num ) {
1055 emit_opcode (*cbuf, 0xD9 ); // FLD ST(i)
1056 emit_d8 (*cbuf, 0xC0+Matcher::_regEncode[src_first]-1 );
1057 emit_opcode (*cbuf, 0xDD ); // FSTP ST(i)
1058 emit_d8 (*cbuf, 0xD8+Matcher::_regEncode[dst_first] );
1059 } else {
1060 emit_opcode (*cbuf, 0xDD ); // FST ST(i)
1061 emit_d8 (*cbuf, 0xD0+Matcher::_regEncode[dst_first]-1 );
1062 }
1063 #ifndef PRODUCT
1064 } else if( !do_size ) {
1065 if( size != 0 ) st->print("\n\t");
1066 if( src_first != FPR1L_num ) st->print("FLD %s\n\tFSTP %s",Matcher::regName[src_first],Matcher::regName[dst_first]);
1067 else st->print( "FST %s", Matcher::regName[dst_first]);
1068 #endif
1069 }
1070 return size + ((src_first != FPR1L_num) ? 2+2 : 2);
1071 }
1073 // Check for float store
1074 if( src_first_rc == rc_float && dst_first_rc == rc_stack ) {
1075 return impl_fp_store_helper(cbuf,do_size,src_first,src_second,dst_first,dst_second,ra_->reg2offset(dst_first),size, st);
1076 }
1078 // Check for float load
1079 if( dst_first_rc == rc_float && src_first_rc == rc_stack ) {
1080 int offset = ra_->reg2offset(src_first);
1081 const char *op_str;
1082 int op;
1083 if( src_first+1 == src_second && dst_first+1 == dst_second ) { // double load?
1084 op_str = "FLD_D";
1085 op = 0xDD;
1086 } else { // 32-bit load
1087 op_str = "FLD_S";
1088 op = 0xD9;
1089 assert( src_second_rc == rc_bad && dst_second_rc == rc_bad, "no non-adjacent float-loads" );
1090 }
1091 if( cbuf ) {
1092 emit_opcode (*cbuf, op );
1093 encode_RegMem(*cbuf, 0x0, ESP_enc, 0x4, 0, offset, relocInfo::none);
1094 emit_opcode (*cbuf, 0xDD ); // FSTP ST(i)
1095 emit_d8 (*cbuf, 0xD8+Matcher::_regEncode[dst_first] );
1096 #ifndef PRODUCT
1097 } else if( !do_size ) {
1098 if( size != 0 ) st->print("\n\t");
1099 st->print("%s ST,[ESP + #%d]\n\tFSTP %s",op_str, offset,Matcher::regName[dst_first]);
1100 #endif
1101 }
1102 int offset_size = (offset == 0) ? 0 : ((offset <= 127) ? 1 : 4);
1103 return size + 3+offset_size+2;
1104 }
1106 // Check for xmm reg-reg copy
1107 if( src_first_rc == rc_xmm && dst_first_rc == rc_xmm ) {
1108 assert( (src_second_rc == rc_bad && dst_second_rc == rc_bad) ||
1109 (src_first+1 == src_second && dst_first+1 == dst_second),
1110 "no non-adjacent float-moves" );
1111 return impl_movx_helper(cbuf,do_size,src_first,dst_first,src_second, dst_second, size, st);
1112 }
1114 // Check for xmm reg-integer reg copy
1115 if( src_first_rc == rc_xmm && dst_first_rc == rc_int ) {
1116 assert( (src_second_rc == rc_bad && dst_second_rc == rc_bad),
1117 "no 64 bit float-integer reg moves" );
1118 return impl_movx2gpr_helper(cbuf,do_size,src_first,dst_first,src_second, dst_second, size, st);
1119 }
1121 // Check for xmm store
1122 if( src_first_rc == rc_xmm && dst_first_rc == rc_stack ) {
1123 return impl_x_helper(cbuf,do_size,false,ra_->reg2offset(dst_first),src_first, src_second, size, st);
1124 }
1126 // Check for float xmm load
1127 if( dst_first_rc == rc_xmm && src_first_rc == rc_stack ) {
1128 return impl_x_helper(cbuf,do_size,true ,ra_->reg2offset(src_first),dst_first, dst_second, size, st);
1129 }
1131 // Copy from float reg to xmm reg
1132 if( dst_first_rc == rc_xmm && src_first_rc == rc_float ) {
1133 // copy to the top of stack from floating point reg
1134 // and use LEA to preserve flags
1135 if( cbuf ) {
1136 emit_opcode(*cbuf,0x8D); // LEA ESP,[ESP-8]
1137 emit_rm(*cbuf, 0x1, ESP_enc, 0x04);
1138 emit_rm(*cbuf, 0x0, 0x04, ESP_enc);
1139 emit_d8(*cbuf,0xF8);
1140 #ifndef PRODUCT
1141 } else if( !do_size ) {
1142 if( size != 0 ) st->print("\n\t");
1143 st->print("LEA ESP,[ESP-8]");
1144 #endif
1145 }
1146 size += 4;
1148 size = impl_fp_store_helper(cbuf,do_size,src_first,src_second,dst_first,dst_second,0,size, st);
1150 // Copy from the temp memory to the xmm reg.
1151 size = impl_x_helper(cbuf,do_size,true ,0,dst_first, dst_second, size, st);
1153 if( cbuf ) {
1154 emit_opcode(*cbuf,0x8D); // LEA ESP,[ESP+8]
1155 emit_rm(*cbuf, 0x1, ESP_enc, 0x04);
1156 emit_rm(*cbuf, 0x0, 0x04, ESP_enc);
1157 emit_d8(*cbuf,0x08);
1158 #ifndef PRODUCT
1159 } else if( !do_size ) {
1160 if( size != 0 ) st->print("\n\t");
1161 st->print("LEA ESP,[ESP+8]");
1162 #endif
1163 }
1164 size += 4;
1165 return size;
1166 }
1168 assert( size > 0, "missed a case" );
1170 // --------------------------------------------------------------------
1171 // Check for second bits still needing moving.
1172 if( src_second == dst_second )
1173 return size; // Self copy; no move
1174 assert( src_second_rc != rc_bad && dst_second_rc != rc_bad, "src_second & dst_second cannot be Bad" );
1176 // Check for second word int-int move
1177 if( src_second_rc == rc_int && dst_second_rc == rc_int )
1178 return impl_mov_helper(cbuf,do_size,src_second,dst_second,size, st);
1180 // Check for second word integer store
1181 if( src_second_rc == rc_int && dst_second_rc == rc_stack )
1182 return impl_helper(cbuf,do_size,false,ra_->reg2offset(dst_second),src_second,0x89,"MOV ",size, st);
1184 // Check for second word integer load
1185 if( dst_second_rc == rc_int && src_second_rc == rc_stack )
1186 return impl_helper(cbuf,do_size,true ,ra_->reg2offset(src_second),dst_second,0x8B,"MOV ",size, st);
1189 Unimplemented();
1190 }
1192 #ifndef PRODUCT
1193 void MachSpillCopyNode::format(PhaseRegAlloc *ra_, outputStream* st) const {
1194 implementation( NULL, ra_, false, st );
1195 }
1196 #endif
1198 void MachSpillCopyNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
1199 implementation( &cbuf, ra_, false, NULL );
1200 }
1202 uint MachSpillCopyNode::size(PhaseRegAlloc *ra_) const {
1203 return implementation( NULL, ra_, true, NULL );
1204 }
1207 //=============================================================================
1208 #ifndef PRODUCT
1209 void BoxLockNode::format( PhaseRegAlloc *ra_, outputStream* st ) const {
1210 int offset = ra_->reg2offset(in_RegMask(0).find_first_elem());
1211 int reg = ra_->get_reg_first(this);
1212 st->print("LEA %s,[ESP + #%d]",Matcher::regName[reg],offset);
1213 }
1214 #endif
1216 void BoxLockNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
1217 int offset = ra_->reg2offset(in_RegMask(0).find_first_elem());
1218 int reg = ra_->get_encode(this);
1219 if( offset >= 128 ) {
1220 emit_opcode(cbuf, 0x8D); // LEA reg,[SP+offset]
1221 emit_rm(cbuf, 0x2, reg, 0x04);
1222 emit_rm(cbuf, 0x0, 0x04, ESP_enc);
1223 emit_d32(cbuf, offset);
1224 }
1225 else {
1226 emit_opcode(cbuf, 0x8D); // LEA reg,[SP+offset]
1227 emit_rm(cbuf, 0x1, reg, 0x04);
1228 emit_rm(cbuf, 0x0, 0x04, ESP_enc);
1229 emit_d8(cbuf, offset);
1230 }
1231 }
1233 uint BoxLockNode::size(PhaseRegAlloc *ra_) const {
1234 int offset = ra_->reg2offset(in_RegMask(0).find_first_elem());
1235 if( offset >= 128 ) {
1236 return 7;
1237 }
1238 else {
1239 return 4;
1240 }
1241 }
1243 //=============================================================================
1245 // emit call stub, compiled java to interpreter
1246 void emit_java_to_interp(CodeBuffer &cbuf ) {
1247 // Stub is fixed up when the corresponding call is converted from calling
1248 // compiled code to calling interpreted code.
1249 // mov rbx,0
1250 // jmp -1
1252 address mark = cbuf.insts_mark(); // get mark within main instrs section
1254 // Note that the code buffer's insts_mark is always relative to insts.
1255 // That's why we must use the macroassembler to generate a stub.
1256 MacroAssembler _masm(&cbuf);
1258 address base =
1259 __ start_a_stub(Compile::MAX_stubs_size);
1260 if (base == NULL) return; // CodeBuffer::expand failed
1261 // static stub relocation stores the instruction address of the call
1262 __ relocate(static_stub_Relocation::spec(mark), RELOC_IMM32);
1263 // static stub relocation also tags the Method* in the code-stream.
1264 __ mov_metadata(rbx, (Metadata*)NULL); // method is zapped till fixup time
1265 // This is recognized as unresolved by relocs/nativeInst/ic code
1266 __ jump(RuntimeAddress(__ pc()));
1268 __ end_a_stub();
1269 // Update current stubs pointer and restore insts_end.
1270 }
1271 // size of call stub, compiled java to interpretor
1272 uint size_java_to_interp() {
1273 return 10; // movl; jmp
1274 }
1275 // relocation entries for call stub, compiled java to interpretor
1276 uint reloc_java_to_interp() {
1277 return 4; // 3 in emit_java_to_interp + 1 in Java_Static_Call
1278 }
1280 //=============================================================================
1281 #ifndef PRODUCT
1282 void MachUEPNode::format( PhaseRegAlloc *ra_, outputStream* st ) const {
1283 st->print_cr( "CMP EAX,[ECX+4]\t# Inline cache check");
1284 st->print_cr("\tJNE SharedRuntime::handle_ic_miss_stub");
1285 st->print_cr("\tNOP");
1286 st->print_cr("\tNOP");
1287 if( !OptoBreakpoint )
1288 st->print_cr("\tNOP");
1289 }
1290 #endif
1292 void MachUEPNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
1293 MacroAssembler masm(&cbuf);
1294 #ifdef ASSERT
1295 uint insts_size = cbuf.insts_size();
1296 #endif
1297 masm.cmpptr(rax, Address(rcx, oopDesc::klass_offset_in_bytes()));
1298 masm.jump_cc(Assembler::notEqual,
1299 RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1300 /* WARNING these NOPs are critical so that verified entry point is properly
1301 aligned for patching by NativeJump::patch_verified_entry() */
1302 int nops_cnt = 2;
1303 if( !OptoBreakpoint ) // Leave space for int3
1304 nops_cnt += 1;
1305 masm.nop(nops_cnt);
1307 assert(cbuf.insts_size() - insts_size == size(ra_), "checking code size of inline cache node");
1308 }
1310 uint MachUEPNode::size(PhaseRegAlloc *ra_) const {
1311 return OptoBreakpoint ? 11 : 12;
1312 }
1315 //=============================================================================
1316 uint size_exception_handler() {
1317 // NativeCall instruction size is the same as NativeJump.
1318 // exception handler starts out as jump and can be patched to
1319 // a call be deoptimization. (4932387)
1320 // Note that this value is also credited (in output.cpp) to
1321 // the size of the code section.
1322 return NativeJump::instruction_size;
1323 }
1325 // Emit exception handler code. Stuff framesize into a register
1326 // and call a VM stub routine.
1327 int emit_exception_handler(CodeBuffer& cbuf) {
1329 // Note that the code buffer's insts_mark is always relative to insts.
1330 // That's why we must use the macroassembler to generate a handler.
1331 MacroAssembler _masm(&cbuf);
1332 address base =
1333 __ start_a_stub(size_exception_handler());
1334 if (base == NULL) return 0; // CodeBuffer::expand failed
1335 int offset = __ offset();
1336 __ jump(RuntimeAddress(OptoRuntime::exception_blob()->entry_point()));
1337 assert(__ offset() - offset <= (int) size_exception_handler(), "overflow");
1338 __ end_a_stub();
1339 return offset;
1340 }
1342 uint size_deopt_handler() {
1343 // NativeCall instruction size is the same as NativeJump.
1344 // exception handler starts out as jump and can be patched to
1345 // a call be deoptimization. (4932387)
1346 // Note that this value is also credited (in output.cpp) to
1347 // the size of the code section.
1348 return 5 + NativeJump::instruction_size; // pushl(); jmp;
1349 }
1351 // Emit deopt handler code.
1352 int emit_deopt_handler(CodeBuffer& cbuf) {
1354 // Note that the code buffer's insts_mark is always relative to insts.
1355 // That's why we must use the macroassembler to generate a handler.
1356 MacroAssembler _masm(&cbuf);
1357 address base =
1358 __ start_a_stub(size_exception_handler());
1359 if (base == NULL) return 0; // CodeBuffer::expand failed
1360 int offset = __ offset();
1361 InternalAddress here(__ pc());
1362 __ pushptr(here.addr());
1364 __ jump(RuntimeAddress(SharedRuntime::deopt_blob()->unpack()));
1365 assert(__ offset() - offset <= (int) size_deopt_handler(), "overflow");
1366 __ end_a_stub();
1367 return offset;
1368 }
1370 int Matcher::regnum_to_fpu_offset(int regnum) {
1371 return regnum - 32; // The FP registers are in the second chunk
1372 }
1374 // This is UltraSparc specific, true just means we have fast l2f conversion
1375 const bool Matcher::convL2FSupported(void) {
1376 return true;
1377 }
1379 // Is this branch offset short enough that a short branch can be used?
1380 //
1381 // NOTE: If the platform does not provide any short branch variants, then
1382 // this method should return false for offset 0.
1383 bool Matcher::is_short_branch_offset(int rule, int br_size, int offset) {
1384 // The passed offset is relative to address of the branch.
1385 // On 86 a branch displacement is calculated relative to address
1386 // of a next instruction.
1387 offset -= br_size;
1389 // the short version of jmpConUCF2 contains multiple branches,
1390 // making the reach slightly less
1391 if (rule == jmpConUCF2_rule)
1392 return (-126 <= offset && offset <= 125);
1393 return (-128 <= offset && offset <= 127);
1394 }
1396 const bool Matcher::isSimpleConstant64(jlong value) {
1397 // Will one (StoreL ConL) be cheaper than two (StoreI ConI)?.
1398 return false;
1399 }
1401 // The ecx parameter to rep stos for the ClearArray node is in dwords.
1402 const bool Matcher::init_array_count_is_in_bytes = false;
1404 // Threshold size for cleararray.
1405 const int Matcher::init_array_short_size = 8 * BytesPerLong;
1407 // Needs 2 CMOV's for longs.
1408 const int Matcher::long_cmove_cost() { return 1; }
1410 // No CMOVF/CMOVD with SSE/SSE2
1411 const int Matcher::float_cmove_cost() { return (UseSSE>=1) ? ConditionalMoveLimit : 0; }
1413 // Should the Matcher clone shifts on addressing modes, expecting them to
1414 // be subsumed into complex addressing expressions or compute them into
1415 // registers? True for Intel but false for most RISCs
1416 const bool Matcher::clone_shift_expressions = true;
1418 // Do we need to mask the count passed to shift instructions or does
1419 // the cpu only look at the lower 5/6 bits anyway?
1420 const bool Matcher::need_masked_shift_count = false;
1422 bool Matcher::narrow_oop_use_complex_address() {
1423 ShouldNotCallThis();
1424 return true;
1425 }
1427 bool Matcher::narrow_klass_use_complex_address() {
1428 ShouldNotCallThis();
1429 return true;
1430 }
1433 // Is it better to copy float constants, or load them directly from memory?
1434 // Intel can load a float constant from a direct address, requiring no
1435 // extra registers. Most RISCs will have to materialize an address into a
1436 // register first, so they would do better to copy the constant from stack.
1437 const bool Matcher::rematerialize_float_constants = true;
1439 // If CPU can load and store mis-aligned doubles directly then no fixup is
1440 // needed. Else we split the double into 2 integer pieces and move it
1441 // piece-by-piece. Only happens when passing doubles into C code as the
1442 // Java calling convention forces doubles to be aligned.
1443 const bool Matcher::misaligned_doubles_ok = true;
1446 void Matcher::pd_implicit_null_fixup(MachNode *node, uint idx) {
1447 // Get the memory operand from the node
1448 uint numopnds = node->num_opnds(); // Virtual call for number of operands
1449 uint skipped = node->oper_input_base(); // Sum of leaves skipped so far
1450 assert( idx >= skipped, "idx too low in pd_implicit_null_fixup" );
1451 uint opcnt = 1; // First operand
1452 uint num_edges = node->_opnds[1]->num_edges(); // leaves for first operand
1453 while( idx >= skipped+num_edges ) {
1454 skipped += num_edges;
1455 opcnt++; // Bump operand count
1456 assert( opcnt < numopnds, "Accessing non-existent operand" );
1457 num_edges = node->_opnds[opcnt]->num_edges(); // leaves for next operand
1458 }
1460 MachOper *memory = node->_opnds[opcnt];
1461 MachOper *new_memory = NULL;
1462 switch (memory->opcode()) {
1463 case DIRECT:
1464 case INDOFFSET32X:
1465 // No transformation necessary.
1466 return;
1467 case INDIRECT:
1468 new_memory = new (C) indirect_win95_safeOper( );
1469 break;
1470 case INDOFFSET8:
1471 new_memory = new (C) indOffset8_win95_safeOper(memory->disp(NULL, NULL, 0));
1472 break;
1473 case INDOFFSET32:
1474 new_memory = new (C) indOffset32_win95_safeOper(memory->disp(NULL, NULL, 0));
1475 break;
1476 case INDINDEXOFFSET:
1477 new_memory = new (C) indIndexOffset_win95_safeOper(memory->disp(NULL, NULL, 0));
1478 break;
1479 case INDINDEXSCALE:
1480 new_memory = new (C) indIndexScale_win95_safeOper(memory->scale());
1481 break;
1482 case INDINDEXSCALEOFFSET:
1483 new_memory = new (C) indIndexScaleOffset_win95_safeOper(memory->scale(), memory->disp(NULL, NULL, 0));
1484 break;
1485 case LOAD_LONG_INDIRECT:
1486 case LOAD_LONG_INDOFFSET32:
1487 // Does not use EBP as address register, use { EDX, EBX, EDI, ESI}
1488 return;
1489 default:
1490 assert(false, "unexpected memory operand in pd_implicit_null_fixup()");
1491 return;
1492 }
1493 node->_opnds[opcnt] = new_memory;
1494 }
1496 // Advertise here if the CPU requires explicit rounding operations
1497 // to implement the UseStrictFP mode.
1498 const bool Matcher::strict_fp_requires_explicit_rounding = true;
1500 // Are floats conerted to double when stored to stack during deoptimization?
1501 // On x32 it is stored with convertion only when FPU is used for floats.
1502 bool Matcher::float_in_double() { return (UseSSE == 0); }
1504 // Do ints take an entire long register or just half?
1505 const bool Matcher::int_in_long = false;
1507 // Return whether or not this register is ever used as an argument. This
1508 // function is used on startup to build the trampoline stubs in generateOptoStub.
1509 // Registers not mentioned will be killed by the VM call in the trampoline, and
1510 // arguments in those registers not be available to the callee.
1511 bool Matcher::can_be_java_arg( int reg ) {
1512 if( reg == ECX_num || reg == EDX_num ) return true;
1513 if( (reg == XMM0_num || reg == XMM1_num ) && UseSSE>=1 ) return true;
1514 if( (reg == XMM0b_num || reg == XMM1b_num) && UseSSE>=2 ) return true;
1515 return false;
1516 }
1518 bool Matcher::is_spillable_arg( int reg ) {
1519 return can_be_java_arg(reg);
1520 }
1522 bool Matcher::use_asm_for_ldiv_by_con( jlong divisor ) {
1523 // Use hardware integer DIV instruction when
1524 // it is faster than a code which use multiply.
1525 // Only when constant divisor fits into 32 bit
1526 // (min_jint is excluded to get only correct
1527 // positive 32 bit values from negative).
1528 return VM_Version::has_fast_idiv() &&
1529 (divisor == (int)divisor && divisor != min_jint);
1530 }
1532 // Register for DIVI projection of divmodI
1533 RegMask Matcher::divI_proj_mask() {
1534 return EAX_REG_mask();
1535 }
1537 // Register for MODI projection of divmodI
1538 RegMask Matcher::modI_proj_mask() {
1539 return EDX_REG_mask();
1540 }
1542 // Register for DIVL projection of divmodL
1543 RegMask Matcher::divL_proj_mask() {
1544 ShouldNotReachHere();
1545 return RegMask();
1546 }
1548 // Register for MODL projection of divmodL
1549 RegMask Matcher::modL_proj_mask() {
1550 ShouldNotReachHere();
1551 return RegMask();
1552 }
1554 const RegMask Matcher::method_handle_invoke_SP_save_mask() {
1555 return EBP_REG_mask();
1556 }
1558 // Returns true if the high 32 bits of the value is known to be zero.
1559 bool is_operand_hi32_zero(Node* n) {
1560 int opc = n->Opcode();
1561 if (opc == Op_AndL) {
1562 Node* o2 = n->in(2);
1563 if (o2->is_Con() && (o2->get_long() & 0xFFFFFFFF00000000LL) == 0LL) {
1564 return true;
1565 }
1566 }
1567 if (opc == Op_ConL && (n->get_long() & 0xFFFFFFFF00000000LL) == 0LL) {
1568 return true;
1569 }
1570 return false;
1571 }
1573 %}
1575 //----------ENCODING BLOCK-----------------------------------------------------
1576 // This block specifies the encoding classes used by the compiler to output
1577 // byte streams. Encoding classes generate functions which are called by
1578 // Machine Instruction Nodes in order to generate the bit encoding of the
1579 // instruction. Operands specify their base encoding interface with the
1580 // interface keyword. There are currently supported four interfaces,
1581 // REG_INTER, CONST_INTER, MEMORY_INTER, & COND_INTER. REG_INTER causes an
1582 // operand to generate a function which returns its register number when
1583 // queried. CONST_INTER causes an operand to generate a function which
1584 // returns the value of the constant when queried. MEMORY_INTER causes an
1585 // operand to generate four functions which return the Base Register, the
1586 // Index Register, the Scale Value, and the Offset Value of the operand when
1587 // queried. COND_INTER causes an operand to generate six functions which
1588 // return the encoding code (ie - encoding bits for the instruction)
1589 // associated with each basic boolean condition for a conditional instruction.
1590 // Instructions specify two basic values for encoding. They use the
1591 // ins_encode keyword to specify their encoding class (which must be one of
1592 // the class names specified in the encoding block), and they use the
1593 // opcode keyword to specify, in order, their primary, secondary, and
1594 // tertiary opcode. Only the opcode sections which a particular instruction
1595 // needs for encoding need to be specified.
1596 encode %{
1597 // Build emit functions for each basic byte or larger field in the intel
1598 // encoding scheme (opcode, rm, sib, immediate), and call them from C++
1599 // code in the enc_class source block. Emit functions will live in the
1600 // main source block for now. In future, we can generalize this by
1601 // adding a syntax that specifies the sizes of fields in an order,
1602 // so that the adlc can build the emit functions automagically
1604 // Emit primary opcode
1605 enc_class OpcP %{
1606 emit_opcode(cbuf, $primary);
1607 %}
1609 // Emit secondary opcode
1610 enc_class OpcS %{
1611 emit_opcode(cbuf, $secondary);
1612 %}
1614 // Emit opcode directly
1615 enc_class Opcode(immI d8) %{
1616 emit_opcode(cbuf, $d8$$constant);
1617 %}
1619 enc_class SizePrefix %{
1620 emit_opcode(cbuf,0x66);
1621 %}
1623 enc_class RegReg (rRegI dst, rRegI src) %{ // RegReg(Many)
1624 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
1625 %}
1627 enc_class OpcRegReg (immI opcode, rRegI dst, rRegI src) %{ // OpcRegReg(Many)
1628 emit_opcode(cbuf,$opcode$$constant);
1629 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
1630 %}
1632 enc_class mov_r32_imm0( rRegI dst ) %{
1633 emit_opcode( cbuf, 0xB8 + $dst$$reg ); // 0xB8+ rd -- MOV r32 ,imm32
1634 emit_d32 ( cbuf, 0x0 ); // imm32==0x0
1635 %}
1637 enc_class cdq_enc %{
1638 // Full implementation of Java idiv and irem; checks for
1639 // special case as described in JVM spec., p.243 & p.271.
1640 //
1641 // normal case special case
1642 //
1643 // input : rax,: dividend min_int
1644 // reg: divisor -1
1645 //
1646 // output: rax,: quotient (= rax, idiv reg) min_int
1647 // rdx: remainder (= rax, irem reg) 0
1648 //
1649 // Code sequnce:
1650 //
1651 // 81 F8 00 00 00 80 cmp rax,80000000h
1652 // 0F 85 0B 00 00 00 jne normal_case
1653 // 33 D2 xor rdx,edx
1654 // 83 F9 FF cmp rcx,0FFh
1655 // 0F 84 03 00 00 00 je done
1656 // normal_case:
1657 // 99 cdq
1658 // F7 F9 idiv rax,ecx
1659 // done:
1660 //
1661 emit_opcode(cbuf,0x81); emit_d8(cbuf,0xF8);
1662 emit_opcode(cbuf,0x00); emit_d8(cbuf,0x00);
1663 emit_opcode(cbuf,0x00); emit_d8(cbuf,0x80); // cmp rax,80000000h
1664 emit_opcode(cbuf,0x0F); emit_d8(cbuf,0x85);
1665 emit_opcode(cbuf,0x0B); emit_d8(cbuf,0x00);
1666 emit_opcode(cbuf,0x00); emit_d8(cbuf,0x00); // jne normal_case
1667 emit_opcode(cbuf,0x33); emit_d8(cbuf,0xD2); // xor rdx,edx
1668 emit_opcode(cbuf,0x83); emit_d8(cbuf,0xF9); emit_d8(cbuf,0xFF); // cmp rcx,0FFh
1669 emit_opcode(cbuf,0x0F); emit_d8(cbuf,0x84);
1670 emit_opcode(cbuf,0x03); emit_d8(cbuf,0x00);
1671 emit_opcode(cbuf,0x00); emit_d8(cbuf,0x00); // je done
1672 // normal_case:
1673 emit_opcode(cbuf,0x99); // cdq
1674 // idiv (note: must be emitted by the user of this rule)
1675 // normal:
1676 %}
1678 // Dense encoding for older common ops
1679 enc_class Opc_plus(immI opcode, rRegI reg) %{
1680 emit_opcode(cbuf, $opcode$$constant + $reg$$reg);
1681 %}
1684 // Opcde enc_class for 8/32 bit immediate instructions with sign-extension
1685 enc_class OpcSE (immI imm) %{ // Emit primary opcode and set sign-extend bit
1686 // Check for 8-bit immediate, and set sign extend bit in opcode
1687 if (($imm$$constant >= -128) && ($imm$$constant <= 127)) {
1688 emit_opcode(cbuf, $primary | 0x02);
1689 }
1690 else { // If 32-bit immediate
1691 emit_opcode(cbuf, $primary);
1692 }
1693 %}
1695 enc_class OpcSErm (rRegI dst, immI imm) %{ // OpcSEr/m
1696 // Emit primary opcode and set sign-extend bit
1697 // Check for 8-bit immediate, and set sign extend bit in opcode
1698 if (($imm$$constant >= -128) && ($imm$$constant <= 127)) {
1699 emit_opcode(cbuf, $primary | 0x02); }
1700 else { // If 32-bit immediate
1701 emit_opcode(cbuf, $primary);
1702 }
1703 // Emit r/m byte with secondary opcode, after primary opcode.
1704 emit_rm(cbuf, 0x3, $secondary, $dst$$reg);
1705 %}
1707 enc_class Con8or32 (immI imm) %{ // Con8or32(storeImmI), 8 or 32 bits
1708 // Check for 8-bit immediate, and set sign extend bit in opcode
1709 if (($imm$$constant >= -128) && ($imm$$constant <= 127)) {
1710 $$$emit8$imm$$constant;
1711 }
1712 else { // If 32-bit immediate
1713 // Output immediate
1714 $$$emit32$imm$$constant;
1715 }
1716 %}
1718 enc_class Long_OpcSErm_Lo(eRegL dst, immL imm) %{
1719 // Emit primary opcode and set sign-extend bit
1720 // Check for 8-bit immediate, and set sign extend bit in opcode
1721 int con = (int)$imm$$constant; // Throw away top bits
1722 emit_opcode(cbuf, ((con >= -128) && (con <= 127)) ? ($primary | 0x02) : $primary);
1723 // Emit r/m byte with secondary opcode, after primary opcode.
1724 emit_rm(cbuf, 0x3, $secondary, $dst$$reg);
1725 if ((con >= -128) && (con <= 127)) emit_d8 (cbuf,con);
1726 else emit_d32(cbuf,con);
1727 %}
1729 enc_class Long_OpcSErm_Hi(eRegL dst, immL imm) %{
1730 // Emit primary opcode and set sign-extend bit
1731 // Check for 8-bit immediate, and set sign extend bit in opcode
1732 int con = (int)($imm$$constant >> 32); // Throw away bottom bits
1733 emit_opcode(cbuf, ((con >= -128) && (con <= 127)) ? ($primary | 0x02) : $primary);
1734 // Emit r/m byte with tertiary opcode, after primary opcode.
1735 emit_rm(cbuf, 0x3, $tertiary, HIGH_FROM_LOW($dst$$reg));
1736 if ((con >= -128) && (con <= 127)) emit_d8 (cbuf,con);
1737 else emit_d32(cbuf,con);
1738 %}
1740 enc_class OpcSReg (rRegI dst) %{ // BSWAP
1741 emit_cc(cbuf, $secondary, $dst$$reg );
1742 %}
1744 enc_class bswap_long_bytes(eRegL dst) %{ // BSWAP
1745 int destlo = $dst$$reg;
1746 int desthi = HIGH_FROM_LOW(destlo);
1747 // bswap lo
1748 emit_opcode(cbuf, 0x0F);
1749 emit_cc(cbuf, 0xC8, destlo);
1750 // bswap hi
1751 emit_opcode(cbuf, 0x0F);
1752 emit_cc(cbuf, 0xC8, desthi);
1753 // xchg lo and hi
1754 emit_opcode(cbuf, 0x87);
1755 emit_rm(cbuf, 0x3, destlo, desthi);
1756 %}
1758 enc_class RegOpc (rRegI div) %{ // IDIV, IMOD, JMP indirect, ...
1759 emit_rm(cbuf, 0x3, $secondary, $div$$reg );
1760 %}
1762 enc_class enc_cmov(cmpOp cop ) %{ // CMOV
1763 $$$emit8$primary;
1764 emit_cc(cbuf, $secondary, $cop$$cmpcode);
1765 %}
1767 enc_class enc_cmov_dpr(cmpOp cop, regDPR src ) %{ // CMOV
1768 int op = 0xDA00 + $cop$$cmpcode + ($src$$reg-1);
1769 emit_d8(cbuf, op >> 8 );
1770 emit_d8(cbuf, op & 255);
1771 %}
1773 // emulate a CMOV with a conditional branch around a MOV
1774 enc_class enc_cmov_branch( cmpOp cop, immI brOffs ) %{ // CMOV
1775 // Invert sense of branch from sense of CMOV
1776 emit_cc( cbuf, 0x70, ($cop$$cmpcode^1) );
1777 emit_d8( cbuf, $brOffs$$constant );
1778 %}
1780 enc_class enc_PartialSubtypeCheck( ) %{
1781 Register Redi = as_Register(EDI_enc); // result register
1782 Register Reax = as_Register(EAX_enc); // super class
1783 Register Recx = as_Register(ECX_enc); // killed
1784 Register Resi = as_Register(ESI_enc); // sub class
1785 Label miss;
1787 MacroAssembler _masm(&cbuf);
1788 __ check_klass_subtype_slow_path(Resi, Reax, Recx, Redi,
1789 NULL, &miss,
1790 /*set_cond_codes:*/ true);
1791 if ($primary) {
1792 __ xorptr(Redi, Redi);
1793 }
1794 __ bind(miss);
1795 %}
1797 enc_class FFree_Float_Stack_All %{ // Free_Float_Stack_All
1798 MacroAssembler masm(&cbuf);
1799 int start = masm.offset();
1800 if (UseSSE >= 2) {
1801 if (VerifyFPU) {
1802 masm.verify_FPU(0, "must be empty in SSE2+ mode");
1803 }
1804 } else {
1805 // External c_calling_convention expects the FPU stack to be 'clean'.
1806 // Compiled code leaves it dirty. Do cleanup now.
1807 masm.empty_FPU_stack();
1808 }
1809 if (sizeof_FFree_Float_Stack_All == -1) {
1810 sizeof_FFree_Float_Stack_All = masm.offset() - start;
1811 } else {
1812 assert(masm.offset() - start == sizeof_FFree_Float_Stack_All, "wrong size");
1813 }
1814 %}
1816 enc_class Verify_FPU_For_Leaf %{
1817 if( VerifyFPU ) {
1818 MacroAssembler masm(&cbuf);
1819 masm.verify_FPU( -3, "Returning from Runtime Leaf call");
1820 }
1821 %}
1823 enc_class Java_To_Runtime (method meth) %{ // CALL Java_To_Runtime, Java_To_Runtime_Leaf
1824 // This is the instruction starting address for relocation info.
1825 cbuf.set_insts_mark();
1826 $$$emit8$primary;
1827 // CALL directly to the runtime
1828 emit_d32_reloc(cbuf, ($meth$$method - (int)(cbuf.insts_end()) - 4),
1829 runtime_call_Relocation::spec(), RELOC_IMM32 );
1831 if (UseSSE >= 2) {
1832 MacroAssembler _masm(&cbuf);
1833 BasicType rt = tf()->return_type();
1835 if ((rt == T_FLOAT || rt == T_DOUBLE) && !return_value_is_used()) {
1836 // A C runtime call where the return value is unused. In SSE2+
1837 // mode the result needs to be removed from the FPU stack. It's
1838 // likely that this function call could be removed by the
1839 // optimizer if the C function is a pure function.
1840 __ ffree(0);
1841 } else if (rt == T_FLOAT) {
1842 __ lea(rsp, Address(rsp, -4));
1843 __ fstp_s(Address(rsp, 0));
1844 __ movflt(xmm0, Address(rsp, 0));
1845 __ lea(rsp, Address(rsp, 4));
1846 } else if (rt == T_DOUBLE) {
1847 __ lea(rsp, Address(rsp, -8));
1848 __ fstp_d(Address(rsp, 0));
1849 __ movdbl(xmm0, Address(rsp, 0));
1850 __ lea(rsp, Address(rsp, 8));
1851 }
1852 }
1853 %}
1856 enc_class pre_call_FPU %{
1857 // If method sets FPU control word restore it here
1858 debug_only(int off0 = cbuf.insts_size());
1859 if( Compile::current()->in_24_bit_fp_mode() ) {
1860 MacroAssembler masm(&cbuf);
1861 masm.fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
1862 }
1863 debug_only(int off1 = cbuf.insts_size());
1864 assert(off1 - off0 == pre_call_FPU_size(), "correct size prediction");
1865 %}
1867 enc_class post_call_FPU %{
1868 // If method sets FPU control word do it here also
1869 if( Compile::current()->in_24_bit_fp_mode() ) {
1870 MacroAssembler masm(&cbuf);
1871 masm.fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
1872 }
1873 %}
1875 enc_class Java_Static_Call (method meth) %{ // JAVA STATIC CALL
1876 // CALL to fixup routine. Fixup routine uses ScopeDesc info to determine
1877 // who we intended to call.
1878 cbuf.set_insts_mark();
1879 $$$emit8$primary;
1880 if ( !_method ) {
1881 emit_d32_reloc(cbuf, ($meth$$method - (int)(cbuf.insts_end()) - 4),
1882 runtime_call_Relocation::spec(), RELOC_IMM32 );
1883 } else if(_optimized_virtual) {
1884 emit_d32_reloc(cbuf, ($meth$$method - (int)(cbuf.insts_end()) - 4),
1885 opt_virtual_call_Relocation::spec(), RELOC_IMM32 );
1886 } else {
1887 emit_d32_reloc(cbuf, ($meth$$method - (int)(cbuf.insts_end()) - 4),
1888 static_call_Relocation::spec(), RELOC_IMM32 );
1889 }
1890 if( _method ) { // Emit stub for static call
1891 emit_java_to_interp(cbuf);
1892 }
1893 %}
1895 enc_class Java_Dynamic_Call (method meth) %{ // JAVA DYNAMIC CALL
1896 MacroAssembler _masm(&cbuf);
1897 __ ic_call((address)$meth$$method);
1898 %}
1900 enc_class Java_Compiled_Call (method meth) %{ // JAVA COMPILED CALL
1901 int disp = in_bytes(Method::from_compiled_offset());
1902 assert( -128 <= disp && disp <= 127, "compiled_code_offset isn't small");
1904 // CALL *[EAX+in_bytes(Method::from_compiled_code_entry_point_offset())]
1905 cbuf.set_insts_mark();
1906 $$$emit8$primary;
1907 emit_rm(cbuf, 0x01, $secondary, EAX_enc ); // R/M byte
1908 emit_d8(cbuf, disp); // Displacement
1910 %}
1912 // Following encoding is no longer used, but may be restored if calling
1913 // convention changes significantly.
1914 // Became: Xor_Reg(EBP), Java_To_Runtime( labl )
1915 //
1916 // enc_class Java_Interpreter_Call (label labl) %{ // JAVA INTERPRETER CALL
1917 // // int ic_reg = Matcher::inline_cache_reg();
1918 // // int ic_encode = Matcher::_regEncode[ic_reg];
1919 // // int imo_reg = Matcher::interpreter_method_oop_reg();
1920 // // int imo_encode = Matcher::_regEncode[imo_reg];
1921 //
1922 // // // Interpreter expects method_oop in EBX, currently a callee-saved register,
1923 // // // so we load it immediately before the call
1924 // // emit_opcode(cbuf, 0x8B); // MOV imo_reg,ic_reg # method_oop
1925 // // emit_rm(cbuf, 0x03, imo_encode, ic_encode ); // R/M byte
1926 //
1927 // // xor rbp,ebp
1928 // emit_opcode(cbuf, 0x33);
1929 // emit_rm(cbuf, 0x3, EBP_enc, EBP_enc);
1930 //
1931 // // CALL to interpreter.
1932 // cbuf.set_insts_mark();
1933 // $$$emit8$primary;
1934 // emit_d32_reloc(cbuf, ($labl$$label - (int)(cbuf.insts_end()) - 4),
1935 // runtime_call_Relocation::spec(), RELOC_IMM32 );
1936 // %}
1938 enc_class RegOpcImm (rRegI dst, immI8 shift) %{ // SHL, SAR, SHR
1939 $$$emit8$primary;
1940 emit_rm(cbuf, 0x3, $secondary, $dst$$reg);
1941 $$$emit8$shift$$constant;
1942 %}
1944 enc_class LdImmI (rRegI dst, immI src) %{ // Load Immediate
1945 // Load immediate does not have a zero or sign extended version
1946 // for 8-bit immediates
1947 emit_opcode(cbuf, 0xB8 + $dst$$reg);
1948 $$$emit32$src$$constant;
1949 %}
1951 enc_class LdImmP (rRegI dst, immI src) %{ // Load Immediate
1952 // Load immediate does not have a zero or sign extended version
1953 // for 8-bit immediates
1954 emit_opcode(cbuf, $primary + $dst$$reg);
1955 $$$emit32$src$$constant;
1956 %}
1958 enc_class LdImmL_Lo( eRegL dst, immL src) %{ // Load Immediate
1959 // Load immediate does not have a zero or sign extended version
1960 // for 8-bit immediates
1961 int dst_enc = $dst$$reg;
1962 int src_con = $src$$constant & 0x0FFFFFFFFL;
1963 if (src_con == 0) {
1964 // xor dst, dst
1965 emit_opcode(cbuf, 0x33);
1966 emit_rm(cbuf, 0x3, dst_enc, dst_enc);
1967 } else {
1968 emit_opcode(cbuf, $primary + dst_enc);
1969 emit_d32(cbuf, src_con);
1970 }
1971 %}
1973 enc_class LdImmL_Hi( eRegL dst, immL src) %{ // Load Immediate
1974 // Load immediate does not have a zero or sign extended version
1975 // for 8-bit immediates
1976 int dst_enc = $dst$$reg + 2;
1977 int src_con = ((julong)($src$$constant)) >> 32;
1978 if (src_con == 0) {
1979 // xor dst, dst
1980 emit_opcode(cbuf, 0x33);
1981 emit_rm(cbuf, 0x3, dst_enc, dst_enc);
1982 } else {
1983 emit_opcode(cbuf, $primary + dst_enc);
1984 emit_d32(cbuf, src_con);
1985 }
1986 %}
1989 // Encode a reg-reg copy. If it is useless, then empty encoding.
1990 enc_class enc_Copy( rRegI dst, rRegI src ) %{
1991 encode_Copy( cbuf, $dst$$reg, $src$$reg );
1992 %}
1994 enc_class enc_CopyL_Lo( rRegI dst, eRegL src ) %{
1995 encode_Copy( cbuf, $dst$$reg, $src$$reg );
1996 %}
1998 enc_class RegReg (rRegI dst, rRegI src) %{ // RegReg(Many)
1999 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2000 %}
2002 enc_class RegReg_Lo(eRegL dst, eRegL src) %{ // RegReg(Many)
2003 $$$emit8$primary;
2004 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2005 %}
2007 enc_class RegReg_Hi(eRegL dst, eRegL src) %{ // RegReg(Many)
2008 $$$emit8$secondary;
2009 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), HIGH_FROM_LOW($src$$reg));
2010 %}
2012 enc_class RegReg_Lo2(eRegL dst, eRegL src) %{ // RegReg(Many)
2013 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2014 %}
2016 enc_class RegReg_Hi2(eRegL dst, eRegL src) %{ // RegReg(Many)
2017 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), HIGH_FROM_LOW($src$$reg));
2018 %}
2020 enc_class RegReg_HiLo( eRegL src, rRegI dst ) %{
2021 emit_rm(cbuf, 0x3, $dst$$reg, HIGH_FROM_LOW($src$$reg));
2022 %}
2024 enc_class Con32 (immI src) %{ // Con32(storeImmI)
2025 // Output immediate
2026 $$$emit32$src$$constant;
2027 %}
2029 enc_class Con32FPR_as_bits(immFPR src) %{ // storeF_imm
2030 // Output Float immediate bits
2031 jfloat jf = $src$$constant;
2032 int jf_as_bits = jint_cast( jf );
2033 emit_d32(cbuf, jf_as_bits);
2034 %}
2036 enc_class Con32F_as_bits(immF src) %{ // storeX_imm
2037 // Output Float immediate bits
2038 jfloat jf = $src$$constant;
2039 int jf_as_bits = jint_cast( jf );
2040 emit_d32(cbuf, jf_as_bits);
2041 %}
2043 enc_class Con16 (immI src) %{ // Con16(storeImmI)
2044 // Output immediate
2045 $$$emit16$src$$constant;
2046 %}
2048 enc_class Con_d32(immI src) %{
2049 emit_d32(cbuf,$src$$constant);
2050 %}
2052 enc_class conmemref (eRegP t1) %{ // Con32(storeImmI)
2053 // Output immediate memory reference
2054 emit_rm(cbuf, 0x00, $t1$$reg, 0x05 );
2055 emit_d32(cbuf, 0x00);
2056 %}
2058 enc_class lock_prefix( ) %{
2059 if( os::is_MP() )
2060 emit_opcode(cbuf,0xF0); // [Lock]
2061 %}
2063 // Cmp-xchg long value.
2064 // Note: we need to swap rbx, and rcx before and after the
2065 // cmpxchg8 instruction because the instruction uses
2066 // rcx as the high order word of the new value to store but
2067 // our register encoding uses rbx,.
2068 enc_class enc_cmpxchg8(eSIRegP mem_ptr) %{
2070 // XCHG rbx,ecx
2071 emit_opcode(cbuf,0x87);
2072 emit_opcode(cbuf,0xD9);
2073 // [Lock]
2074 if( os::is_MP() )
2075 emit_opcode(cbuf,0xF0);
2076 // CMPXCHG8 [Eptr]
2077 emit_opcode(cbuf,0x0F);
2078 emit_opcode(cbuf,0xC7);
2079 emit_rm( cbuf, 0x0, 1, $mem_ptr$$reg );
2080 // XCHG rbx,ecx
2081 emit_opcode(cbuf,0x87);
2082 emit_opcode(cbuf,0xD9);
2083 %}
2085 enc_class enc_cmpxchg(eSIRegP mem_ptr) %{
2086 // [Lock]
2087 if( os::is_MP() )
2088 emit_opcode(cbuf,0xF0);
2090 // CMPXCHG [Eptr]
2091 emit_opcode(cbuf,0x0F);
2092 emit_opcode(cbuf,0xB1);
2093 emit_rm( cbuf, 0x0, 1, $mem_ptr$$reg );
2094 %}
2096 enc_class enc_flags_ne_to_boolean( iRegI res ) %{
2097 int res_encoding = $res$$reg;
2099 // MOV res,0
2100 emit_opcode( cbuf, 0xB8 + res_encoding);
2101 emit_d32( cbuf, 0 );
2102 // JNE,s fail
2103 emit_opcode(cbuf,0x75);
2104 emit_d8(cbuf, 5 );
2105 // MOV res,1
2106 emit_opcode( cbuf, 0xB8 + res_encoding);
2107 emit_d32( cbuf, 1 );
2108 // fail:
2109 %}
2111 enc_class set_instruction_start( ) %{
2112 cbuf.set_insts_mark(); // Mark start of opcode for reloc info in mem operand
2113 %}
2115 enc_class RegMem (rRegI ereg, memory mem) %{ // emit_reg_mem
2116 int reg_encoding = $ereg$$reg;
2117 int base = $mem$$base;
2118 int index = $mem$$index;
2119 int scale = $mem$$scale;
2120 int displace = $mem$$disp;
2121 relocInfo::relocType disp_reloc = $mem->disp_reloc();
2122 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, disp_reloc);
2123 %}
2125 enc_class RegMem_Hi(eRegL ereg, memory mem) %{ // emit_reg_mem
2126 int reg_encoding = HIGH_FROM_LOW($ereg$$reg); // Hi register of pair, computed from lo
2127 int base = $mem$$base;
2128 int index = $mem$$index;
2129 int scale = $mem$$scale;
2130 int displace = $mem$$disp + 4; // Offset is 4 further in memory
2131 assert( $mem->disp_reloc() == relocInfo::none, "Cannot add 4 to oop" );
2132 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, relocInfo::none);
2133 %}
2135 enc_class move_long_small_shift( eRegL dst, immI_1_31 cnt ) %{
2136 int r1, r2;
2137 if( $tertiary == 0xA4 ) { r1 = $dst$$reg; r2 = HIGH_FROM_LOW($dst$$reg); }
2138 else { r2 = $dst$$reg; r1 = HIGH_FROM_LOW($dst$$reg); }
2139 emit_opcode(cbuf,0x0F);
2140 emit_opcode(cbuf,$tertiary);
2141 emit_rm(cbuf, 0x3, r1, r2);
2142 emit_d8(cbuf,$cnt$$constant);
2143 emit_d8(cbuf,$primary);
2144 emit_rm(cbuf, 0x3, $secondary, r1);
2145 emit_d8(cbuf,$cnt$$constant);
2146 %}
2148 enc_class move_long_big_shift_sign( eRegL dst, immI_32_63 cnt ) %{
2149 emit_opcode( cbuf, 0x8B ); // Move
2150 emit_rm(cbuf, 0x3, $dst$$reg, HIGH_FROM_LOW($dst$$reg));
2151 if( $cnt$$constant > 32 ) { // Shift, if not by zero
2152 emit_d8(cbuf,$primary);
2153 emit_rm(cbuf, 0x3, $secondary, $dst$$reg);
2154 emit_d8(cbuf,$cnt$$constant-32);
2155 }
2156 emit_d8(cbuf,$primary);
2157 emit_rm(cbuf, 0x3, $secondary, HIGH_FROM_LOW($dst$$reg));
2158 emit_d8(cbuf,31);
2159 %}
2161 enc_class move_long_big_shift_clr( eRegL dst, immI_32_63 cnt ) %{
2162 int r1, r2;
2163 if( $secondary == 0x5 ) { r1 = $dst$$reg; r2 = HIGH_FROM_LOW($dst$$reg); }
2164 else { r2 = $dst$$reg; r1 = HIGH_FROM_LOW($dst$$reg); }
2166 emit_opcode( cbuf, 0x8B ); // Move r1,r2
2167 emit_rm(cbuf, 0x3, r1, r2);
2168 if( $cnt$$constant > 32 ) { // Shift, if not by zero
2169 emit_opcode(cbuf,$primary);
2170 emit_rm(cbuf, 0x3, $secondary, r1);
2171 emit_d8(cbuf,$cnt$$constant-32);
2172 }
2173 emit_opcode(cbuf,0x33); // XOR r2,r2
2174 emit_rm(cbuf, 0x3, r2, r2);
2175 %}
2177 // Clone of RegMem but accepts an extra parameter to access each
2178 // half of a double in memory; it never needs relocation info.
2179 enc_class Mov_MemD_half_to_Reg (immI opcode, memory mem, immI disp_for_half, rRegI rm_reg) %{
2180 emit_opcode(cbuf,$opcode$$constant);
2181 int reg_encoding = $rm_reg$$reg;
2182 int base = $mem$$base;
2183 int index = $mem$$index;
2184 int scale = $mem$$scale;
2185 int displace = $mem$$disp + $disp_for_half$$constant;
2186 relocInfo::relocType disp_reloc = relocInfo::none;
2187 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, disp_reloc);
2188 %}
2190 // !!!!! Special Custom Code used by MemMove, and stack access instructions !!!!!
2191 //
2192 // Clone of RegMem except the RM-byte's reg/opcode field is an ADLC-time constant
2193 // and it never needs relocation information.
2194 // Frequently used to move data between FPU's Stack Top and memory.
2195 enc_class RMopc_Mem_no_oop (immI rm_opcode, memory mem) %{
2196 int rm_byte_opcode = $rm_opcode$$constant;
2197 int base = $mem$$base;
2198 int index = $mem$$index;
2199 int scale = $mem$$scale;
2200 int displace = $mem$$disp;
2201 assert( $mem->disp_reloc() == relocInfo::none, "No oops here because no reloc info allowed" );
2202 encode_RegMem(cbuf, rm_byte_opcode, base, index, scale, displace, relocInfo::none);
2203 %}
2205 enc_class RMopc_Mem (immI rm_opcode, memory mem) %{
2206 int rm_byte_opcode = $rm_opcode$$constant;
2207 int base = $mem$$base;
2208 int index = $mem$$index;
2209 int scale = $mem$$scale;
2210 int displace = $mem$$disp;
2211 relocInfo::relocType disp_reloc = $mem->disp_reloc(); // disp-as-oop when working with static globals
2212 encode_RegMem(cbuf, rm_byte_opcode, base, index, scale, displace, disp_reloc);
2213 %}
2215 enc_class RegLea (rRegI dst, rRegI src0, immI src1 ) %{ // emit_reg_lea
2216 int reg_encoding = $dst$$reg;
2217 int base = $src0$$reg; // 0xFFFFFFFF indicates no base
2218 int index = 0x04; // 0x04 indicates no index
2219 int scale = 0x00; // 0x00 indicates no scale
2220 int displace = $src1$$constant; // 0x00 indicates no displacement
2221 relocInfo::relocType disp_reloc = relocInfo::none;
2222 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, disp_reloc);
2223 %}
2225 enc_class min_enc (rRegI dst, rRegI src) %{ // MIN
2226 // Compare dst,src
2227 emit_opcode(cbuf,0x3B);
2228 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2229 // jmp dst < src around move
2230 emit_opcode(cbuf,0x7C);
2231 emit_d8(cbuf,2);
2232 // move dst,src
2233 emit_opcode(cbuf,0x8B);
2234 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2235 %}
2237 enc_class max_enc (rRegI dst, rRegI src) %{ // MAX
2238 // Compare dst,src
2239 emit_opcode(cbuf,0x3B);
2240 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2241 // jmp dst > src around move
2242 emit_opcode(cbuf,0x7F);
2243 emit_d8(cbuf,2);
2244 // move dst,src
2245 emit_opcode(cbuf,0x8B);
2246 emit_rm(cbuf, 0x3, $dst$$reg, $src$$reg);
2247 %}
2249 enc_class enc_FPR_store(memory mem, regDPR src) %{
2250 // If src is FPR1, we can just FST to store it.
2251 // Else we need to FLD it to FPR1, then FSTP to store/pop it.
2252 int reg_encoding = 0x2; // Just store
2253 int base = $mem$$base;
2254 int index = $mem$$index;
2255 int scale = $mem$$scale;
2256 int displace = $mem$$disp;
2257 relocInfo::relocType disp_reloc = $mem->disp_reloc(); // disp-as-oop when working with static globals
2258 if( $src$$reg != FPR1L_enc ) {
2259 reg_encoding = 0x3; // Store & pop
2260 emit_opcode( cbuf, 0xD9 ); // FLD (i.e., push it)
2261 emit_d8( cbuf, 0xC0-1+$src$$reg );
2262 }
2263 cbuf.set_insts_mark(); // Mark start of opcode for reloc info in mem operand
2264 emit_opcode(cbuf,$primary);
2265 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, disp_reloc);
2266 %}
2268 enc_class neg_reg(rRegI dst) %{
2269 // NEG $dst
2270 emit_opcode(cbuf,0xF7);
2271 emit_rm(cbuf, 0x3, 0x03, $dst$$reg );
2272 %}
2274 enc_class setLT_reg(eCXRegI dst) %{
2275 // SETLT $dst
2276 emit_opcode(cbuf,0x0F);
2277 emit_opcode(cbuf,0x9C);
2278 emit_rm( cbuf, 0x3, 0x4, $dst$$reg );
2279 %}
2281 enc_class enc_cmpLTP(ncxRegI p, ncxRegI q, ncxRegI y, eCXRegI tmp) %{ // cadd_cmpLT
2282 int tmpReg = $tmp$$reg;
2284 // SUB $p,$q
2285 emit_opcode(cbuf,0x2B);
2286 emit_rm(cbuf, 0x3, $p$$reg, $q$$reg);
2287 // SBB $tmp,$tmp
2288 emit_opcode(cbuf,0x1B);
2289 emit_rm(cbuf, 0x3, tmpReg, tmpReg);
2290 // AND $tmp,$y
2291 emit_opcode(cbuf,0x23);
2292 emit_rm(cbuf, 0x3, tmpReg, $y$$reg);
2293 // ADD $p,$tmp
2294 emit_opcode(cbuf,0x03);
2295 emit_rm(cbuf, 0x3, $p$$reg, tmpReg);
2296 %}
2298 enc_class enc_cmpLTP_mem(rRegI p, rRegI q, memory mem, eCXRegI tmp) %{ // cadd_cmpLT
2299 int tmpReg = $tmp$$reg;
2301 // SUB $p,$q
2302 emit_opcode(cbuf,0x2B);
2303 emit_rm(cbuf, 0x3, $p$$reg, $q$$reg);
2304 // SBB $tmp,$tmp
2305 emit_opcode(cbuf,0x1B);
2306 emit_rm(cbuf, 0x3, tmpReg, tmpReg);
2307 // AND $tmp,$y
2308 cbuf.set_insts_mark(); // Mark start of opcode for reloc info in mem operand
2309 emit_opcode(cbuf,0x23);
2310 int reg_encoding = tmpReg;
2311 int base = $mem$$base;
2312 int index = $mem$$index;
2313 int scale = $mem$$scale;
2314 int displace = $mem$$disp;
2315 relocInfo::relocType disp_reloc = $mem->disp_reloc();
2316 encode_RegMem(cbuf, reg_encoding, base, index, scale, displace, disp_reloc);
2317 // ADD $p,$tmp
2318 emit_opcode(cbuf,0x03);
2319 emit_rm(cbuf, 0x3, $p$$reg, tmpReg);
2320 %}
2322 enc_class shift_left_long( eRegL dst, eCXRegI shift ) %{
2323 // TEST shift,32
2324 emit_opcode(cbuf,0xF7);
2325 emit_rm(cbuf, 0x3, 0, ECX_enc);
2326 emit_d32(cbuf,0x20);
2327 // JEQ,s small
2328 emit_opcode(cbuf, 0x74);
2329 emit_d8(cbuf, 0x04);
2330 // MOV $dst.hi,$dst.lo
2331 emit_opcode( cbuf, 0x8B );
2332 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), $dst$$reg );
2333 // CLR $dst.lo
2334 emit_opcode(cbuf, 0x33);
2335 emit_rm(cbuf, 0x3, $dst$$reg, $dst$$reg);
2336 // small:
2337 // SHLD $dst.hi,$dst.lo,$shift
2338 emit_opcode(cbuf,0x0F);
2339 emit_opcode(cbuf,0xA5);
2340 emit_rm(cbuf, 0x3, $dst$$reg, HIGH_FROM_LOW($dst$$reg));
2341 // SHL $dst.lo,$shift"
2342 emit_opcode(cbuf,0xD3);
2343 emit_rm(cbuf, 0x3, 0x4, $dst$$reg );
2344 %}
2346 enc_class shift_right_long( eRegL dst, eCXRegI shift ) %{
2347 // TEST shift,32
2348 emit_opcode(cbuf,0xF7);
2349 emit_rm(cbuf, 0x3, 0, ECX_enc);
2350 emit_d32(cbuf,0x20);
2351 // JEQ,s small
2352 emit_opcode(cbuf, 0x74);
2353 emit_d8(cbuf, 0x04);
2354 // MOV $dst.lo,$dst.hi
2355 emit_opcode( cbuf, 0x8B );
2356 emit_rm(cbuf, 0x3, $dst$$reg, HIGH_FROM_LOW($dst$$reg) );
2357 // CLR $dst.hi
2358 emit_opcode(cbuf, 0x33);
2359 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), HIGH_FROM_LOW($dst$$reg));
2360 // small:
2361 // SHRD $dst.lo,$dst.hi,$shift
2362 emit_opcode(cbuf,0x0F);
2363 emit_opcode(cbuf,0xAD);
2364 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), $dst$$reg);
2365 // SHR $dst.hi,$shift"
2366 emit_opcode(cbuf,0xD3);
2367 emit_rm(cbuf, 0x3, 0x5, HIGH_FROM_LOW($dst$$reg) );
2368 %}
2370 enc_class shift_right_arith_long( eRegL dst, eCXRegI shift ) %{
2371 // TEST shift,32
2372 emit_opcode(cbuf,0xF7);
2373 emit_rm(cbuf, 0x3, 0, ECX_enc);
2374 emit_d32(cbuf,0x20);
2375 // JEQ,s small
2376 emit_opcode(cbuf, 0x74);
2377 emit_d8(cbuf, 0x05);
2378 // MOV $dst.lo,$dst.hi
2379 emit_opcode( cbuf, 0x8B );
2380 emit_rm(cbuf, 0x3, $dst$$reg, HIGH_FROM_LOW($dst$$reg) );
2381 // SAR $dst.hi,31
2382 emit_opcode(cbuf, 0xC1);
2383 emit_rm(cbuf, 0x3, 7, HIGH_FROM_LOW($dst$$reg) );
2384 emit_d8(cbuf, 0x1F );
2385 // small:
2386 // SHRD $dst.lo,$dst.hi,$shift
2387 emit_opcode(cbuf,0x0F);
2388 emit_opcode(cbuf,0xAD);
2389 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), $dst$$reg);
2390 // SAR $dst.hi,$shift"
2391 emit_opcode(cbuf,0xD3);
2392 emit_rm(cbuf, 0x3, 0x7, HIGH_FROM_LOW($dst$$reg) );
2393 %}
2396 // ----------------- Encodings for floating point unit -----------------
2397 // May leave result in FPU-TOS or FPU reg depending on opcodes
2398 enc_class OpcReg_FPR(regFPR src) %{ // FMUL, FDIV
2399 $$$emit8$primary;
2400 emit_rm(cbuf, 0x3, $secondary, $src$$reg );
2401 %}
2403 // Pop argument in FPR0 with FSTP ST(0)
2404 enc_class PopFPU() %{
2405 emit_opcode( cbuf, 0xDD );
2406 emit_d8( cbuf, 0xD8 );
2407 %}
2409 // !!!!! equivalent to Pop_Reg_F
2410 enc_class Pop_Reg_DPR( regDPR dst ) %{
2411 emit_opcode( cbuf, 0xDD ); // FSTP ST(i)
2412 emit_d8( cbuf, 0xD8+$dst$$reg );
2413 %}
2415 enc_class Push_Reg_DPR( regDPR dst ) %{
2416 emit_opcode( cbuf, 0xD9 );
2417 emit_d8( cbuf, 0xC0-1+$dst$$reg ); // FLD ST(i-1)
2418 %}
2420 enc_class strictfp_bias1( regDPR dst ) %{
2421 emit_opcode( cbuf, 0xDB ); // FLD m80real
2422 emit_opcode( cbuf, 0x2D );
2423 emit_d32( cbuf, (int)StubRoutines::addr_fpu_subnormal_bias1() );
2424 emit_opcode( cbuf, 0xDE ); // FMULP ST(dst), ST0
2425 emit_opcode( cbuf, 0xC8+$dst$$reg );
2426 %}
2428 enc_class strictfp_bias2( regDPR dst ) %{
2429 emit_opcode( cbuf, 0xDB ); // FLD m80real
2430 emit_opcode( cbuf, 0x2D );
2431 emit_d32( cbuf, (int)StubRoutines::addr_fpu_subnormal_bias2() );
2432 emit_opcode( cbuf, 0xDE ); // FMULP ST(dst), ST0
2433 emit_opcode( cbuf, 0xC8+$dst$$reg );
2434 %}
2436 // Special case for moving an integer register to a stack slot.
2437 enc_class OpcPRegSS( stackSlotI dst, rRegI src ) %{ // RegSS
2438 store_to_stackslot( cbuf, $primary, $src$$reg, $dst$$disp );
2439 %}
2441 // Special case for moving a register to a stack slot.
2442 enc_class RegSS( stackSlotI dst, rRegI src ) %{ // RegSS
2443 // Opcode already emitted
2444 emit_rm( cbuf, 0x02, $src$$reg, ESP_enc ); // R/M byte
2445 emit_rm( cbuf, 0x00, ESP_enc, ESP_enc); // SIB byte
2446 emit_d32(cbuf, $dst$$disp); // Displacement
2447 %}
2449 // Push the integer in stackSlot 'src' onto FP-stack
2450 enc_class Push_Mem_I( memory src ) %{ // FILD [ESP+src]
2451 store_to_stackslot( cbuf, $primary, $secondary, $src$$disp );
2452 %}
2454 // Push FPU's TOS float to a stack-slot, and pop FPU-stack
2455 enc_class Pop_Mem_FPR( stackSlotF dst ) %{ // FSTP_S [ESP+dst]
2456 store_to_stackslot( cbuf, 0xD9, 0x03, $dst$$disp );
2457 %}
2459 // Same as Pop_Mem_F except for opcode
2460 // Push FPU's TOS double to a stack-slot, and pop FPU-stack
2461 enc_class Pop_Mem_DPR( stackSlotD dst ) %{ // FSTP_D [ESP+dst]
2462 store_to_stackslot( cbuf, 0xDD, 0x03, $dst$$disp );
2463 %}
2465 enc_class Pop_Reg_FPR( regFPR dst ) %{
2466 emit_opcode( cbuf, 0xDD ); // FSTP ST(i)
2467 emit_d8( cbuf, 0xD8+$dst$$reg );
2468 %}
2470 enc_class Push_Reg_FPR( regFPR dst ) %{
2471 emit_opcode( cbuf, 0xD9 ); // FLD ST(i-1)
2472 emit_d8( cbuf, 0xC0-1+$dst$$reg );
2473 %}
2475 // Push FPU's float to a stack-slot, and pop FPU-stack
2476 enc_class Pop_Mem_Reg_FPR( stackSlotF dst, regFPR src ) %{
2477 int pop = 0x02;
2478 if ($src$$reg != FPR1L_enc) {
2479 emit_opcode( cbuf, 0xD9 ); // FLD ST(i-1)
2480 emit_d8( cbuf, 0xC0-1+$src$$reg );
2481 pop = 0x03;
2482 }
2483 store_to_stackslot( cbuf, 0xD9, pop, $dst$$disp ); // FST<P>_S [ESP+dst]
2484 %}
2486 // Push FPU's double to a stack-slot, and pop FPU-stack
2487 enc_class Pop_Mem_Reg_DPR( stackSlotD dst, regDPR src ) %{
2488 int pop = 0x02;
2489 if ($src$$reg != FPR1L_enc) {
2490 emit_opcode( cbuf, 0xD9 ); // FLD ST(i-1)
2491 emit_d8( cbuf, 0xC0-1+$src$$reg );
2492 pop = 0x03;
2493 }
2494 store_to_stackslot( cbuf, 0xDD, pop, $dst$$disp ); // FST<P>_D [ESP+dst]
2495 %}
2497 // Push FPU's double to a FPU-stack-slot, and pop FPU-stack
2498 enc_class Pop_Reg_Reg_DPR( regDPR dst, regFPR src ) %{
2499 int pop = 0xD0 - 1; // -1 since we skip FLD
2500 if ($src$$reg != FPR1L_enc) {
2501 emit_opcode( cbuf, 0xD9 ); // FLD ST(src-1)
2502 emit_d8( cbuf, 0xC0-1+$src$$reg );
2503 pop = 0xD8;
2504 }
2505 emit_opcode( cbuf, 0xDD );
2506 emit_d8( cbuf, pop+$dst$$reg ); // FST<P> ST(i)
2507 %}
2510 enc_class Push_Reg_Mod_DPR( regDPR dst, regDPR src) %{
2511 // load dst in FPR0
2512 emit_opcode( cbuf, 0xD9 );
2513 emit_d8( cbuf, 0xC0-1+$dst$$reg );
2514 if ($src$$reg != FPR1L_enc) {
2515 // fincstp
2516 emit_opcode (cbuf, 0xD9);
2517 emit_opcode (cbuf, 0xF7);
2518 // swap src with FPR1:
2519 // FXCH FPR1 with src
2520 emit_opcode(cbuf, 0xD9);
2521 emit_d8(cbuf, 0xC8-1+$src$$reg );
2522 // fdecstp
2523 emit_opcode (cbuf, 0xD9);
2524 emit_opcode (cbuf, 0xF6);
2525 }
2526 %}
2528 enc_class Push_ModD_encoding(regD src0, regD src1) %{
2529 MacroAssembler _masm(&cbuf);
2530 __ subptr(rsp, 8);
2531 __ movdbl(Address(rsp, 0), $src1$$XMMRegister);
2532 __ fld_d(Address(rsp, 0));
2533 __ movdbl(Address(rsp, 0), $src0$$XMMRegister);
2534 __ fld_d(Address(rsp, 0));
2535 %}
2537 enc_class Push_ModF_encoding(regF src0, regF src1) %{
2538 MacroAssembler _masm(&cbuf);
2539 __ subptr(rsp, 4);
2540 __ movflt(Address(rsp, 0), $src1$$XMMRegister);
2541 __ fld_s(Address(rsp, 0));
2542 __ movflt(Address(rsp, 0), $src0$$XMMRegister);
2543 __ fld_s(Address(rsp, 0));
2544 %}
2546 enc_class Push_ResultD(regD dst) %{
2547 MacroAssembler _masm(&cbuf);
2548 __ fstp_d(Address(rsp, 0));
2549 __ movdbl($dst$$XMMRegister, Address(rsp, 0));
2550 __ addptr(rsp, 8);
2551 %}
2553 enc_class Push_ResultF(regF dst, immI d8) %{
2554 MacroAssembler _masm(&cbuf);
2555 __ fstp_s(Address(rsp, 0));
2556 __ movflt($dst$$XMMRegister, Address(rsp, 0));
2557 __ addptr(rsp, $d8$$constant);
2558 %}
2560 enc_class Push_SrcD(regD src) %{
2561 MacroAssembler _masm(&cbuf);
2562 __ subptr(rsp, 8);
2563 __ movdbl(Address(rsp, 0), $src$$XMMRegister);
2564 __ fld_d(Address(rsp, 0));
2565 %}
2567 enc_class push_stack_temp_qword() %{
2568 MacroAssembler _masm(&cbuf);
2569 __ subptr(rsp, 8);
2570 %}
2572 enc_class pop_stack_temp_qword() %{
2573 MacroAssembler _masm(&cbuf);
2574 __ addptr(rsp, 8);
2575 %}
2577 enc_class push_xmm_to_fpr1(regD src) %{
2578 MacroAssembler _masm(&cbuf);
2579 __ movdbl(Address(rsp, 0), $src$$XMMRegister);
2580 __ fld_d(Address(rsp, 0));
2581 %}
2583 enc_class Push_Result_Mod_DPR( regDPR src) %{
2584 if ($src$$reg != FPR1L_enc) {
2585 // fincstp
2586 emit_opcode (cbuf, 0xD9);
2587 emit_opcode (cbuf, 0xF7);
2588 // FXCH FPR1 with src
2589 emit_opcode(cbuf, 0xD9);
2590 emit_d8(cbuf, 0xC8-1+$src$$reg );
2591 // fdecstp
2592 emit_opcode (cbuf, 0xD9);
2593 emit_opcode (cbuf, 0xF6);
2594 }
2595 // // following asm replaced with Pop_Reg_F or Pop_Mem_F
2596 // // FSTP FPR$dst$$reg
2597 // emit_opcode( cbuf, 0xDD );
2598 // emit_d8( cbuf, 0xD8+$dst$$reg );
2599 %}
2601 enc_class fnstsw_sahf_skip_parity() %{
2602 // fnstsw ax
2603 emit_opcode( cbuf, 0xDF );
2604 emit_opcode( cbuf, 0xE0 );
2605 // sahf
2606 emit_opcode( cbuf, 0x9E );
2607 // jnp ::skip
2608 emit_opcode( cbuf, 0x7B );
2609 emit_opcode( cbuf, 0x05 );
2610 %}
2612 enc_class emitModDPR() %{
2613 // fprem must be iterative
2614 // :: loop
2615 // fprem
2616 emit_opcode( cbuf, 0xD9 );
2617 emit_opcode( cbuf, 0xF8 );
2618 // wait
2619 emit_opcode( cbuf, 0x9b );
2620 // fnstsw ax
2621 emit_opcode( cbuf, 0xDF );
2622 emit_opcode( cbuf, 0xE0 );
2623 // sahf
2624 emit_opcode( cbuf, 0x9E );
2625 // jp ::loop
2626 emit_opcode( cbuf, 0x0F );
2627 emit_opcode( cbuf, 0x8A );
2628 emit_opcode( cbuf, 0xF4 );
2629 emit_opcode( cbuf, 0xFF );
2630 emit_opcode( cbuf, 0xFF );
2631 emit_opcode( cbuf, 0xFF );
2632 %}
2634 enc_class fpu_flags() %{
2635 // fnstsw_ax
2636 emit_opcode( cbuf, 0xDF);
2637 emit_opcode( cbuf, 0xE0);
2638 // test ax,0x0400
2639 emit_opcode( cbuf, 0x66 ); // operand-size prefix for 16-bit immediate
2640 emit_opcode( cbuf, 0xA9 );
2641 emit_d16 ( cbuf, 0x0400 );
2642 // // // This sequence works, but stalls for 12-16 cycles on PPro
2643 // // test rax,0x0400
2644 // emit_opcode( cbuf, 0xA9 );
2645 // emit_d32 ( cbuf, 0x00000400 );
2646 //
2647 // jz exit (no unordered comparison)
2648 emit_opcode( cbuf, 0x74 );
2649 emit_d8 ( cbuf, 0x02 );
2650 // mov ah,1 - treat as LT case (set carry flag)
2651 emit_opcode( cbuf, 0xB4 );
2652 emit_d8 ( cbuf, 0x01 );
2653 // sahf
2654 emit_opcode( cbuf, 0x9E);
2655 %}
2657 enc_class cmpF_P6_fixup() %{
2658 // Fixup the integer flags in case comparison involved a NaN
2659 //
2660 // JNP exit (no unordered comparison, P-flag is set by NaN)
2661 emit_opcode( cbuf, 0x7B );
2662 emit_d8 ( cbuf, 0x03 );
2663 // MOV AH,1 - treat as LT case (set carry flag)
2664 emit_opcode( cbuf, 0xB4 );
2665 emit_d8 ( cbuf, 0x01 );
2666 // SAHF
2667 emit_opcode( cbuf, 0x9E);
2668 // NOP // target for branch to avoid branch to branch
2669 emit_opcode( cbuf, 0x90);
2670 %}
2672 // fnstsw_ax();
2673 // sahf();
2674 // movl(dst, nan_result);
2675 // jcc(Assembler::parity, exit);
2676 // movl(dst, less_result);
2677 // jcc(Assembler::below, exit);
2678 // movl(dst, equal_result);
2679 // jcc(Assembler::equal, exit);
2680 // movl(dst, greater_result);
2682 // less_result = 1;
2683 // greater_result = -1;
2684 // equal_result = 0;
2685 // nan_result = -1;
2687 enc_class CmpF_Result(rRegI dst) %{
2688 // fnstsw_ax();
2689 emit_opcode( cbuf, 0xDF);
2690 emit_opcode( cbuf, 0xE0);
2691 // sahf
2692 emit_opcode( cbuf, 0x9E);
2693 // movl(dst, nan_result);
2694 emit_opcode( cbuf, 0xB8 + $dst$$reg);
2695 emit_d32( cbuf, -1 );
2696 // jcc(Assembler::parity, exit);
2697 emit_opcode( cbuf, 0x7A );
2698 emit_d8 ( cbuf, 0x13 );
2699 // movl(dst, less_result);
2700 emit_opcode( cbuf, 0xB8 + $dst$$reg);
2701 emit_d32( cbuf, -1 );
2702 // jcc(Assembler::below, exit);
2703 emit_opcode( cbuf, 0x72 );
2704 emit_d8 ( cbuf, 0x0C );
2705 // movl(dst, equal_result);
2706 emit_opcode( cbuf, 0xB8 + $dst$$reg);
2707 emit_d32( cbuf, 0 );
2708 // jcc(Assembler::equal, exit);
2709 emit_opcode( cbuf, 0x74 );
2710 emit_d8 ( cbuf, 0x05 );
2711 // movl(dst, greater_result);
2712 emit_opcode( cbuf, 0xB8 + $dst$$reg);
2713 emit_d32( cbuf, 1 );
2714 %}
2717 // Compare the longs and set flags
2718 // BROKEN! Do Not use as-is
2719 enc_class cmpl_test( eRegL src1, eRegL src2 ) %{
2720 // CMP $src1.hi,$src2.hi
2721 emit_opcode( cbuf, 0x3B );
2722 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($src1$$reg), HIGH_FROM_LOW($src2$$reg) );
2723 // JNE,s done
2724 emit_opcode(cbuf,0x75);
2725 emit_d8(cbuf, 2 );
2726 // CMP $src1.lo,$src2.lo
2727 emit_opcode( cbuf, 0x3B );
2728 emit_rm(cbuf, 0x3, $src1$$reg, $src2$$reg );
2729 // done:
2730 %}
2732 enc_class convert_int_long( regL dst, rRegI src ) %{
2733 // mov $dst.lo,$src
2734 int dst_encoding = $dst$$reg;
2735 int src_encoding = $src$$reg;
2736 encode_Copy( cbuf, dst_encoding , src_encoding );
2737 // mov $dst.hi,$src
2738 encode_Copy( cbuf, HIGH_FROM_LOW(dst_encoding), src_encoding );
2739 // sar $dst.hi,31
2740 emit_opcode( cbuf, 0xC1 );
2741 emit_rm(cbuf, 0x3, 7, HIGH_FROM_LOW(dst_encoding) );
2742 emit_d8(cbuf, 0x1F );
2743 %}
2745 enc_class convert_long_double( eRegL src ) %{
2746 // push $src.hi
2747 emit_opcode(cbuf, 0x50+HIGH_FROM_LOW($src$$reg));
2748 // push $src.lo
2749 emit_opcode(cbuf, 0x50+$src$$reg );
2750 // fild 64-bits at [SP]
2751 emit_opcode(cbuf,0xdf);
2752 emit_d8(cbuf, 0x6C);
2753 emit_d8(cbuf, 0x24);
2754 emit_d8(cbuf, 0x00);
2755 // pop stack
2756 emit_opcode(cbuf, 0x83); // add SP, #8
2757 emit_rm(cbuf, 0x3, 0x00, ESP_enc);
2758 emit_d8(cbuf, 0x8);
2759 %}
2761 enc_class multiply_con_and_shift_high( eDXRegI dst, nadxRegI src1, eADXRegL_low_only src2, immI_32_63 cnt, eFlagsReg cr ) %{
2762 // IMUL EDX:EAX,$src1
2763 emit_opcode( cbuf, 0xF7 );
2764 emit_rm( cbuf, 0x3, 0x5, $src1$$reg );
2765 // SAR EDX,$cnt-32
2766 int shift_count = ((int)$cnt$$constant) - 32;
2767 if (shift_count > 0) {
2768 emit_opcode(cbuf, 0xC1);
2769 emit_rm(cbuf, 0x3, 7, $dst$$reg );
2770 emit_d8(cbuf, shift_count);
2771 }
2772 %}
2774 // this version doesn't have add sp, 8
2775 enc_class convert_long_double2( eRegL src ) %{
2776 // push $src.hi
2777 emit_opcode(cbuf, 0x50+HIGH_FROM_LOW($src$$reg));
2778 // push $src.lo
2779 emit_opcode(cbuf, 0x50+$src$$reg );
2780 // fild 64-bits at [SP]
2781 emit_opcode(cbuf,0xdf);
2782 emit_d8(cbuf, 0x6C);
2783 emit_d8(cbuf, 0x24);
2784 emit_d8(cbuf, 0x00);
2785 %}
2787 enc_class long_int_multiply( eADXRegL dst, nadxRegI src) %{
2788 // Basic idea: long = (long)int * (long)int
2789 // IMUL EDX:EAX, src
2790 emit_opcode( cbuf, 0xF7 );
2791 emit_rm( cbuf, 0x3, 0x5, $src$$reg);
2792 %}
2794 enc_class long_uint_multiply( eADXRegL dst, nadxRegI src) %{
2795 // Basic Idea: long = (int & 0xffffffffL) * (int & 0xffffffffL)
2796 // MUL EDX:EAX, src
2797 emit_opcode( cbuf, 0xF7 );
2798 emit_rm( cbuf, 0x3, 0x4, $src$$reg);
2799 %}
2801 enc_class long_multiply( eADXRegL dst, eRegL src, rRegI tmp ) %{
2802 // Basic idea: lo(result) = lo(x_lo * y_lo)
2803 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi)
2804 // MOV $tmp,$src.lo
2805 encode_Copy( cbuf, $tmp$$reg, $src$$reg );
2806 // IMUL $tmp,EDX
2807 emit_opcode( cbuf, 0x0F );
2808 emit_opcode( cbuf, 0xAF );
2809 emit_rm( cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($dst$$reg) );
2810 // MOV EDX,$src.hi
2811 encode_Copy( cbuf, HIGH_FROM_LOW($dst$$reg), HIGH_FROM_LOW($src$$reg) );
2812 // IMUL EDX,EAX
2813 emit_opcode( cbuf, 0x0F );
2814 emit_opcode( cbuf, 0xAF );
2815 emit_rm( cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), $dst$$reg );
2816 // ADD $tmp,EDX
2817 emit_opcode( cbuf, 0x03 );
2818 emit_rm( cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($dst$$reg) );
2819 // MUL EDX:EAX,$src.lo
2820 emit_opcode( cbuf, 0xF7 );
2821 emit_rm( cbuf, 0x3, 0x4, $src$$reg );
2822 // ADD EDX,ESI
2823 emit_opcode( cbuf, 0x03 );
2824 emit_rm( cbuf, 0x3, HIGH_FROM_LOW($dst$$reg), $tmp$$reg );
2825 %}
2827 enc_class long_multiply_con( eADXRegL dst, immL_127 src, rRegI tmp ) %{
2828 // Basic idea: lo(result) = lo(src * y_lo)
2829 // hi(result) = hi(src * y_lo) + lo(src * y_hi)
2830 // IMUL $tmp,EDX,$src
2831 emit_opcode( cbuf, 0x6B );
2832 emit_rm( cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($dst$$reg) );
2833 emit_d8( cbuf, (int)$src$$constant );
2834 // MOV EDX,$src
2835 emit_opcode(cbuf, 0xB8 + EDX_enc);
2836 emit_d32( cbuf, (int)$src$$constant );
2837 // MUL EDX:EAX,EDX
2838 emit_opcode( cbuf, 0xF7 );
2839 emit_rm( cbuf, 0x3, 0x4, EDX_enc );
2840 // ADD EDX,ESI
2841 emit_opcode( cbuf, 0x03 );
2842 emit_rm( cbuf, 0x3, EDX_enc, $tmp$$reg );
2843 %}
2845 enc_class long_div( eRegL src1, eRegL src2 ) %{
2846 // PUSH src1.hi
2847 emit_opcode(cbuf, HIGH_FROM_LOW(0x50+$src1$$reg) );
2848 // PUSH src1.lo
2849 emit_opcode(cbuf, 0x50+$src1$$reg );
2850 // PUSH src2.hi
2851 emit_opcode(cbuf, HIGH_FROM_LOW(0x50+$src2$$reg) );
2852 // PUSH src2.lo
2853 emit_opcode(cbuf, 0x50+$src2$$reg );
2854 // CALL directly to the runtime
2855 cbuf.set_insts_mark();
2856 emit_opcode(cbuf,0xE8); // Call into runtime
2857 emit_d32_reloc(cbuf, (CAST_FROM_FN_PTR(address, SharedRuntime::ldiv) - cbuf.insts_end()) - 4, runtime_call_Relocation::spec(), RELOC_IMM32 );
2858 // Restore stack
2859 emit_opcode(cbuf, 0x83); // add SP, #framesize
2860 emit_rm(cbuf, 0x3, 0x00, ESP_enc);
2861 emit_d8(cbuf, 4*4);
2862 %}
2864 enc_class long_mod( eRegL src1, eRegL src2 ) %{
2865 // PUSH src1.hi
2866 emit_opcode(cbuf, HIGH_FROM_LOW(0x50+$src1$$reg) );
2867 // PUSH src1.lo
2868 emit_opcode(cbuf, 0x50+$src1$$reg );
2869 // PUSH src2.hi
2870 emit_opcode(cbuf, HIGH_FROM_LOW(0x50+$src2$$reg) );
2871 // PUSH src2.lo
2872 emit_opcode(cbuf, 0x50+$src2$$reg );
2873 // CALL directly to the runtime
2874 cbuf.set_insts_mark();
2875 emit_opcode(cbuf,0xE8); // Call into runtime
2876 emit_d32_reloc(cbuf, (CAST_FROM_FN_PTR(address, SharedRuntime::lrem ) - cbuf.insts_end()) - 4, runtime_call_Relocation::spec(), RELOC_IMM32 );
2877 // Restore stack
2878 emit_opcode(cbuf, 0x83); // add SP, #framesize
2879 emit_rm(cbuf, 0x3, 0x00, ESP_enc);
2880 emit_d8(cbuf, 4*4);
2881 %}
2883 enc_class long_cmp_flags0( eRegL src, rRegI tmp ) %{
2884 // MOV $tmp,$src.lo
2885 emit_opcode(cbuf, 0x8B);
2886 emit_rm(cbuf, 0x3, $tmp$$reg, $src$$reg);
2887 // OR $tmp,$src.hi
2888 emit_opcode(cbuf, 0x0B);
2889 emit_rm(cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($src$$reg));
2890 %}
2892 enc_class long_cmp_flags1( eRegL src1, eRegL src2 ) %{
2893 // CMP $src1.lo,$src2.lo
2894 emit_opcode( cbuf, 0x3B );
2895 emit_rm(cbuf, 0x3, $src1$$reg, $src2$$reg );
2896 // JNE,s skip
2897 emit_cc(cbuf, 0x70, 0x5);
2898 emit_d8(cbuf,2);
2899 // CMP $src1.hi,$src2.hi
2900 emit_opcode( cbuf, 0x3B );
2901 emit_rm(cbuf, 0x3, HIGH_FROM_LOW($src1$$reg), HIGH_FROM_LOW($src2$$reg) );
2902 %}
2904 enc_class long_cmp_flags2( eRegL src1, eRegL src2, rRegI tmp ) %{
2905 // CMP $src1.lo,$src2.lo\t! Long compare; set flags for low bits
2906 emit_opcode( cbuf, 0x3B );
2907 emit_rm(cbuf, 0x3, $src1$$reg, $src2$$reg );
2908 // MOV $tmp,$src1.hi
2909 emit_opcode( cbuf, 0x8B );
2910 emit_rm(cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($src1$$reg) );
2911 // SBB $tmp,$src2.hi\t! Compute flags for long compare
2912 emit_opcode( cbuf, 0x1B );
2913 emit_rm(cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($src2$$reg) );
2914 %}
2916 enc_class long_cmp_flags3( eRegL src, rRegI tmp ) %{
2917 // XOR $tmp,$tmp
2918 emit_opcode(cbuf,0x33); // XOR
2919 emit_rm(cbuf,0x3, $tmp$$reg, $tmp$$reg);
2920 // CMP $tmp,$src.lo
2921 emit_opcode( cbuf, 0x3B );
2922 emit_rm(cbuf, 0x3, $tmp$$reg, $src$$reg );
2923 // SBB $tmp,$src.hi
2924 emit_opcode( cbuf, 0x1B );
2925 emit_rm(cbuf, 0x3, $tmp$$reg, HIGH_FROM_LOW($src$$reg) );
2926 %}
2928 // Sniff, sniff... smells like Gnu Superoptimizer
2929 enc_class neg_long( eRegL dst ) %{
2930 emit_opcode(cbuf,0xF7); // NEG hi
2931 emit_rm (cbuf,0x3, 0x3, HIGH_FROM_LOW($dst$$reg));
2932 emit_opcode(cbuf,0xF7); // NEG lo
2933 emit_rm (cbuf,0x3, 0x3, $dst$$reg );
2934 emit_opcode(cbuf,0x83); // SBB hi,0
2935 emit_rm (cbuf,0x3, 0x3, HIGH_FROM_LOW($dst$$reg));
2936 emit_d8 (cbuf,0 );
2937 %}
2940 // Because the transitions from emitted code to the runtime
2941 // monitorenter/exit helper stubs are so slow it's critical that
2942 // we inline both the stack-locking fast-path and the inflated fast path.
2943 //
2944 // See also: cmpFastLock and cmpFastUnlock.
2945 //
2946 // What follows is a specialized inline transliteration of the code
2947 // in slow_enter() and slow_exit(). If we're concerned about I$ bloat
2948 // another option would be to emit TrySlowEnter and TrySlowExit methods
2949 // at startup-time. These methods would accept arguments as
2950 // (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure
2951 // indications in the icc.ZFlag. Fast_Lock and Fast_Unlock would simply
2952 // marshal the arguments and emit calls to TrySlowEnter and TrySlowExit.
2953 // In practice, however, the # of lock sites is bounded and is usually small.
2954 // Besides the call overhead, TrySlowEnter and TrySlowExit might suffer
2955 // if the processor uses simple bimodal branch predictors keyed by EIP
2956 // Since the helper routines would be called from multiple synchronization
2957 // sites.
2958 //
2959 // An even better approach would be write "MonitorEnter()" and "MonitorExit()"
2960 // in java - using j.u.c and unsafe - and just bind the lock and unlock sites
2961 // to those specialized methods. That'd give us a mostly platform-independent
2962 // implementation that the JITs could optimize and inline at their pleasure.
2963 // Done correctly, the only time we'd need to cross to native could would be
2964 // to park() or unpark() threads. We'd also need a few more unsafe operators
2965 // to (a) prevent compiler-JIT reordering of non-volatile accesses, and
2966 // (b) explicit barriers or fence operations.
2967 //
2968 // TODO:
2969 //
2970 // * Arrange for C2 to pass "Self" into Fast_Lock and Fast_Unlock in one of the registers (scr).
2971 // This avoids manifesting the Self pointer in the Fast_Lock and Fast_Unlock terminals.
2972 // Given TLAB allocation, Self is usually manifested in a register, so passing it into
2973 // the lock operators would typically be faster than reifying Self.
2974 //
2975 // * Ideally I'd define the primitives as:
2976 // fast_lock (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED.
2977 // fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED
2978 // Unfortunately ADLC bugs prevent us from expressing the ideal form.
2979 // Instead, we're stuck with a rather awkward and brittle register assignments below.
2980 // Furthermore the register assignments are overconstrained, possibly resulting in
2981 // sub-optimal code near the synchronization site.
2982 //
2983 // * Eliminate the sp-proximity tests and just use "== Self" tests instead.
2984 // Alternately, use a better sp-proximity test.
2985 //
2986 // * Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value.
2987 // Either one is sufficient to uniquely identify a thread.
2988 // TODO: eliminate use of sp in _owner and use get_thread(tr) instead.
2989 //
2990 // * Intrinsify notify() and notifyAll() for the common cases where the
2991 // object is locked by the calling thread but the waitlist is empty.
2992 // avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll().
2993 //
2994 // * use jccb and jmpb instead of jcc and jmp to improve code density.
2995 // But beware of excessive branch density on AMD Opterons.
2996 //
2997 // * Both Fast_Lock and Fast_Unlock set the ICC.ZF to indicate success
2998 // or failure of the fast-path. If the fast-path fails then we pass
2999 // control to the slow-path, typically in C. In Fast_Lock and
3000 // Fast_Unlock we often branch to DONE_LABEL, just to find that C2
3001 // will emit a conditional branch immediately after the node.
3002 // So we have branches to branches and lots of ICC.ZF games.
3003 // Instead, it might be better to have C2 pass a "FailureLabel"
3004 // into Fast_Lock and Fast_Unlock. In the case of success, control
3005 // will drop through the node. ICC.ZF is undefined at exit.
3006 // In the case of failure, the node will branch directly to the
3007 // FailureLabel
3010 // obj: object to lock
3011 // box: on-stack box address (displaced header location) - KILLED
3012 // rax,: tmp -- KILLED
3013 // scr: tmp -- KILLED
3014 enc_class Fast_Lock( eRegP obj, eRegP box, eAXRegI tmp, eRegP scr ) %{
3016 Register objReg = as_Register($obj$$reg);
3017 Register boxReg = as_Register($box$$reg);
3018 Register tmpReg = as_Register($tmp$$reg);
3019 Register scrReg = as_Register($scr$$reg);
3021 // Ensure the register assignents are disjoint
3022 guarantee (objReg != boxReg, "") ;
3023 guarantee (objReg != tmpReg, "") ;
3024 guarantee (objReg != scrReg, "") ;
3025 guarantee (boxReg != tmpReg, "") ;
3026 guarantee (boxReg != scrReg, "") ;
3027 guarantee (tmpReg == as_Register(EAX_enc), "") ;
3029 MacroAssembler masm(&cbuf);
3031 if (_counters != NULL) {
3032 masm.atomic_incl(ExternalAddress((address) _counters->total_entry_count_addr()));
3033 }
3034 if (EmitSync & 1) {
3035 // set box->dhw = unused_mark (3)
3036 // Force all sync thru slow-path: slow_enter() and slow_exit()
3037 masm.movptr (Address(boxReg, 0), int32_t(markOopDesc::unused_mark())) ;
3038 masm.cmpptr (rsp, (int32_t)0) ;
3039 } else
3040 if (EmitSync & 2) {
3041 Label DONE_LABEL ;
3042 if (UseBiasedLocking) {
3043 // Note: tmpReg maps to the swap_reg argument and scrReg to the tmp_reg argument.
3044 masm.biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, _counters);
3045 }
3047 masm.movptr(tmpReg, Address(objReg, 0)) ; // fetch markword
3048 masm.orptr (tmpReg, 0x1);
3049 masm.movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS
3050 if (os::is_MP()) { masm.lock(); }
3051 masm.cmpxchgptr(boxReg, Address(objReg, 0)); // Updates tmpReg
3052 masm.jcc(Assembler::equal, DONE_LABEL);
3053 // Recursive locking
3054 masm.subptr(tmpReg, rsp);
3055 masm.andptr(tmpReg, (int32_t) 0xFFFFF003 );
3056 masm.movptr(Address(boxReg, 0), tmpReg);
3057 masm.bind(DONE_LABEL) ;
3058 } else {
3059 // Possible cases that we'll encounter in fast_lock
3060 // ------------------------------------------------
3061 // * Inflated
3062 // -- unlocked
3063 // -- Locked
3064 // = by self
3065 // = by other
3066 // * biased
3067 // -- by Self
3068 // -- by other
3069 // * neutral
3070 // * stack-locked
3071 // -- by self
3072 // = sp-proximity test hits
3073 // = sp-proximity test generates false-negative
3074 // -- by other
3075 //
3077 Label IsInflated, DONE_LABEL, PopDone ;
3079 // TODO: optimize away redundant LDs of obj->mark and improve the markword triage
3080 // order to reduce the number of conditional branches in the most common cases.
3081 // Beware -- there's a subtle invariant that fetch of the markword
3082 // at [FETCH], below, will never observe a biased encoding (*101b).
3083 // If this invariant is not held we risk exclusion (safety) failure.
3084 if (UseBiasedLocking && !UseOptoBiasInlining) {
3085 masm.biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, _counters);
3086 }
3088 masm.movptr(tmpReg, Address(objReg, 0)) ; // [FETCH]
3089 masm.testptr(tmpReg, 0x02) ; // Inflated v (Stack-locked or neutral)
3090 masm.jccb (Assembler::notZero, IsInflated) ;
3092 // Attempt stack-locking ...
3093 masm.orptr (tmpReg, 0x1);
3094 masm.movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS
3095 if (os::is_MP()) { masm.lock(); }
3096 masm.cmpxchgptr(boxReg, Address(objReg, 0)); // Updates tmpReg
3097 if (_counters != NULL) {
3098 masm.cond_inc32(Assembler::equal,
3099 ExternalAddress((address)_counters->fast_path_entry_count_addr()));
3100 }
3101 masm.jccb (Assembler::equal, DONE_LABEL);
3103 // Recursive locking
3104 masm.subptr(tmpReg, rsp);
3105 masm.andptr(tmpReg, 0xFFFFF003 );
3106 masm.movptr(Address(boxReg, 0), tmpReg);
3107 if (_counters != NULL) {
3108 masm.cond_inc32(Assembler::equal,
3109 ExternalAddress((address)_counters->fast_path_entry_count_addr()));
3110 }
3111 masm.jmp (DONE_LABEL) ;
3113 masm.bind (IsInflated) ;
3115 // The object is inflated.
3116 //
3117 // TODO-FIXME: eliminate the ugly use of manifest constants:
3118 // Use markOopDesc::monitor_value instead of "2".
3119 // use markOop::unused_mark() instead of "3".
3120 // The tmpReg value is an objectMonitor reference ORed with
3121 // markOopDesc::monitor_value (2). We can either convert tmpReg to an
3122 // objectmonitor pointer by masking off the "2" bit or we can just
3123 // use tmpReg as an objectmonitor pointer but bias the objectmonitor
3124 // field offsets with "-2" to compensate for and annul the low-order tag bit.
3125 //
3126 // I use the latter as it avoids AGI stalls.
3127 // As such, we write "mov r, [tmpReg+OFFSETOF(Owner)-2]"
3128 // instead of "mov r, [tmpReg+OFFSETOF(Owner)]".
3129 //
3130 #define OFFSET_SKEWED(f) ((ObjectMonitor::f ## _offset_in_bytes())-2)
3132 // boxReg refers to the on-stack BasicLock in the current frame.
3133 // We'd like to write:
3134 // set box->_displaced_header = markOop::unused_mark(). Any non-0 value suffices.
3135 // This is convenient but results a ST-before-CAS penalty. The following CAS suffers
3136 // additional latency as we have another ST in the store buffer that must drain.
3138 if (EmitSync & 8192) {
3139 masm.movptr(Address(boxReg, 0), 3) ; // results in ST-before-CAS penalty
3140 masm.get_thread (scrReg) ;
3141 masm.movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2]
3142 masm.movptr(tmpReg, NULL_WORD); // consider: xor vs mov
3143 if (os::is_MP()) { masm.lock(); }
3144 masm.cmpxchgptr(scrReg, Address(boxReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
3145 } else
3146 if ((EmitSync & 128) == 0) { // avoid ST-before-CAS
3147 masm.movptr(scrReg, boxReg) ;
3148 masm.movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2]
3150 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
3151 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
3152 // prefetchw [eax + Offset(_owner)-2]
3153 masm.prefetchw(Address(rax, ObjectMonitor::owner_offset_in_bytes()-2));
3154 }
3156 if ((EmitSync & 64) == 0) {
3157 // Optimistic form: consider XORL tmpReg,tmpReg
3158 masm.movptr(tmpReg, NULL_WORD) ;
3159 } else {
3160 // Can suffer RTS->RTO upgrades on shared or cold $ lines
3161 // Test-And-CAS instead of CAS
3162 masm.movptr(tmpReg, Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ; // rax, = m->_owner
3163 masm.testptr(tmpReg, tmpReg) ; // Locked ?
3164 masm.jccb (Assembler::notZero, DONE_LABEL) ;
3165 }
3167 // Appears unlocked - try to swing _owner from null to non-null.
3168 // Ideally, I'd manifest "Self" with get_thread and then attempt
3169 // to CAS the register containing Self into m->Owner.
3170 // But we don't have enough registers, so instead we can either try to CAS
3171 // rsp or the address of the box (in scr) into &m->owner. If the CAS succeeds
3172 // we later store "Self" into m->Owner. Transiently storing a stack address
3173 // (rsp or the address of the box) into m->owner is harmless.
3174 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
3175 if (os::is_MP()) { masm.lock(); }
3176 masm.cmpxchgptr(scrReg, Address(boxReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
3177 masm.movptr(Address(scrReg, 0), 3) ; // box->_displaced_header = 3
3178 masm.jccb (Assembler::notZero, DONE_LABEL) ;
3179 masm.get_thread (scrReg) ; // beware: clobbers ICCs
3180 masm.movptr(Address(boxReg, ObjectMonitor::owner_offset_in_bytes()-2), scrReg) ;
3181 masm.xorptr(boxReg, boxReg) ; // set icc.ZFlag = 1 to indicate success
3183 // If the CAS fails we can either retry or pass control to the slow-path.
3184 // We use the latter tactic.
3185 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
3186 // If the CAS was successful ...
3187 // Self has acquired the lock
3188 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
3189 // Intentional fall-through into DONE_LABEL ...
3190 } else {
3191 masm.movptr(Address(boxReg, 0), 3) ; // results in ST-before-CAS penalty
3192 masm.movptr(boxReg, tmpReg) ;
3194 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
3195 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
3196 // prefetchw [eax + Offset(_owner)-2]
3197 masm.prefetchw(Address(rax, ObjectMonitor::owner_offset_in_bytes()-2));
3198 }
3200 if ((EmitSync & 64) == 0) {
3201 // Optimistic form
3202 masm.xorptr (tmpReg, tmpReg) ;
3203 } else {
3204 // Can suffer RTS->RTO upgrades on shared or cold $ lines
3205 masm.movptr(tmpReg, Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ; // rax, = m->_owner
3206 masm.testptr(tmpReg, tmpReg) ; // Locked ?
3207 masm.jccb (Assembler::notZero, DONE_LABEL) ;
3208 }
3210 // Appears unlocked - try to swing _owner from null to non-null.
3211 // Use either "Self" (in scr) or rsp as thread identity in _owner.
3212 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
3213 masm.get_thread (scrReg) ;
3214 if (os::is_MP()) { masm.lock(); }
3215 masm.cmpxchgptr(scrReg, Address(boxReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
3217 // If the CAS fails we can either retry or pass control to the slow-path.
3218 // We use the latter tactic.
3219 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
3220 // If the CAS was successful ...
3221 // Self has acquired the lock
3222 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
3223 // Intentional fall-through into DONE_LABEL ...
3224 }
3226 // DONE_LABEL is a hot target - we'd really like to place it at the
3227 // start of cache line by padding with NOPs.
3228 // See the AMD and Intel software optimization manuals for the
3229 // most efficient "long" NOP encodings.
3230 // Unfortunately none of our alignment mechanisms suffice.
3231 masm.bind(DONE_LABEL);
3233 // Avoid branch-to-branch on AMD processors
3234 // This appears to be superstition.
3235 if (EmitSync & 32) masm.nop() ;
3238 // At DONE_LABEL the icc ZFlag is set as follows ...
3239 // Fast_Unlock uses the same protocol.
3240 // ZFlag == 1 -> Success
3241 // ZFlag == 0 -> Failure - force control through the slow-path
3242 }
3243 %}
3245 // obj: object to unlock
3246 // box: box address (displaced header location), killed. Must be EAX.
3247 // rbx,: killed tmp; cannot be obj nor box.
3248 //
3249 // Some commentary on balanced locking:
3250 //
3251 // Fast_Lock and Fast_Unlock are emitted only for provably balanced lock sites.
3252 // Methods that don't have provably balanced locking are forced to run in the
3253 // interpreter - such methods won't be compiled to use fast_lock and fast_unlock.
3254 // The interpreter provides two properties:
3255 // I1: At return-time the interpreter automatically and quietly unlocks any
3256 // objects acquired the current activation (frame). Recall that the
3257 // interpreter maintains an on-stack list of locks currently held by
3258 // a frame.
3259 // I2: If a method attempts to unlock an object that is not held by the
3260 // the frame the interpreter throws IMSX.
3261 //
3262 // Lets say A(), which has provably balanced locking, acquires O and then calls B().
3263 // B() doesn't have provably balanced locking so it runs in the interpreter.
3264 // Control returns to A() and A() unlocks O. By I1 and I2, above, we know that O
3265 // is still locked by A().
3266 //
3267 // The only other source of unbalanced locking would be JNI. The "Java Native Interface:
3268 // Programmer's Guide and Specification" claims that an object locked by jni_monitorenter
3269 // should not be unlocked by "normal" java-level locking and vice-versa. The specification
3270 // doesn't specify what will occur if a program engages in such mixed-mode locking, however.
3272 enc_class Fast_Unlock( nabxRegP obj, eAXRegP box, eRegP tmp) %{
3274 Register objReg = as_Register($obj$$reg);
3275 Register boxReg = as_Register($box$$reg);
3276 Register tmpReg = as_Register($tmp$$reg);
3278 guarantee (objReg != boxReg, "") ;
3279 guarantee (objReg != tmpReg, "") ;
3280 guarantee (boxReg != tmpReg, "") ;
3281 guarantee (boxReg == as_Register(EAX_enc), "") ;
3282 MacroAssembler masm(&cbuf);
3284 if (EmitSync & 4) {
3285 // Disable - inhibit all inlining. Force control through the slow-path
3286 masm.cmpptr (rsp, 0) ;
3287 } else
3288 if (EmitSync & 8) {
3289 Label DONE_LABEL ;
3290 if (UseBiasedLocking) {
3291 masm.biased_locking_exit(objReg, tmpReg, DONE_LABEL);
3292 }
3293 // classic stack-locking code ...
3294 masm.movptr(tmpReg, Address(boxReg, 0)) ;
3295 masm.testptr(tmpReg, tmpReg) ;
3296 masm.jcc (Assembler::zero, DONE_LABEL) ;
3297 if (os::is_MP()) { masm.lock(); }
3298 masm.cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses EAX which is box
3299 masm.bind(DONE_LABEL);
3300 } else {
3301 Label DONE_LABEL, Stacked, CheckSucc, Inflated ;
3303 // Critically, the biased locking test must have precedence over
3304 // and appear before the (box->dhw == 0) recursive stack-lock test.
3305 if (UseBiasedLocking && !UseOptoBiasInlining) {
3306 masm.biased_locking_exit(objReg, tmpReg, DONE_LABEL);
3307 }
3309 masm.cmpptr(Address(boxReg, 0), 0) ; // Examine the displaced header
3310 masm.movptr(tmpReg, Address(objReg, 0)) ; // Examine the object's markword
3311 masm.jccb (Assembler::zero, DONE_LABEL) ; // 0 indicates recursive stack-lock
3313 masm.testptr(tmpReg, 0x02) ; // Inflated?
3314 masm.jccb (Assembler::zero, Stacked) ;
3316 masm.bind (Inflated) ;
3317 // It's inflated.
3318 // Despite our balanced locking property we still check that m->_owner == Self
3319 // as java routines or native JNI code called by this thread might
3320 // have released the lock.
3321 // Refer to the comments in synchronizer.cpp for how we might encode extra
3322 // state in _succ so we can avoid fetching EntryList|cxq.
3323 //
3324 // I'd like to add more cases in fast_lock() and fast_unlock() --
3325 // such as recursive enter and exit -- but we have to be wary of
3326 // I$ bloat, T$ effects and BP$ effects.
3327 //
3328 // If there's no contention try a 1-0 exit. That is, exit without
3329 // a costly MEMBAR or CAS. See synchronizer.cpp for details on how
3330 // we detect and recover from the race that the 1-0 exit admits.
3331 //
3332 // Conceptually Fast_Unlock() must execute a STST|LDST "release" barrier
3333 // before it STs null into _owner, releasing the lock. Updates
3334 // to data protected by the critical section must be visible before
3335 // we drop the lock (and thus before any other thread could acquire
3336 // the lock and observe the fields protected by the lock).
3337 // IA32's memory-model is SPO, so STs are ordered with respect to
3338 // each other and there's no need for an explicit barrier (fence).
3339 // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
3341 masm.get_thread (boxReg) ;
3342 if ((EmitSync & 4096) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
3343 // prefetchw [ebx + Offset(_owner)-2]
3344 masm.prefetchw(Address(rbx, ObjectMonitor::owner_offset_in_bytes()-2));
3345 }
3347 // Note that we could employ various encoding schemes to reduce
3348 // the number of loads below (currently 4) to just 2 or 3.
3349 // Refer to the comments in synchronizer.cpp.
3350 // In practice the chain of fetches doesn't seem to impact performance, however.
3351 if ((EmitSync & 65536) == 0 && (EmitSync & 256)) {
3352 // Attempt to reduce branch density - AMD's branch predictor.
3353 masm.xorptr(boxReg, Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
3354 masm.orptr(boxReg, Address (tmpReg, ObjectMonitor::recursions_offset_in_bytes()-2)) ;
3355 masm.orptr(boxReg, Address (tmpReg, ObjectMonitor::EntryList_offset_in_bytes()-2)) ;
3356 masm.orptr(boxReg, Address (tmpReg, ObjectMonitor::cxq_offset_in_bytes()-2)) ;
3357 masm.jccb (Assembler::notZero, DONE_LABEL) ;
3358 masm.movptr(Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2), NULL_WORD) ;
3359 masm.jmpb (DONE_LABEL) ;
3360 } else {
3361 masm.xorptr(boxReg, Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
3362 masm.orptr(boxReg, Address (tmpReg, ObjectMonitor::recursions_offset_in_bytes()-2)) ;
3363 masm.jccb (Assembler::notZero, DONE_LABEL) ;
3364 masm.movptr(boxReg, Address (tmpReg, ObjectMonitor::EntryList_offset_in_bytes()-2)) ;
3365 masm.orptr(boxReg, Address (tmpReg, ObjectMonitor::cxq_offset_in_bytes()-2)) ;
3366 masm.jccb (Assembler::notZero, CheckSucc) ;
3367 masm.movptr(Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2), NULL_WORD) ;
3368 masm.jmpb (DONE_LABEL) ;
3369 }
3371 // The Following code fragment (EmitSync & 65536) improves the performance of
3372 // contended applications and contended synchronization microbenchmarks.
3373 // Unfortunately the emission of the code - even though not executed - causes regressions
3374 // in scimark and jetstream, evidently because of $ effects. Replacing the code
3375 // with an equal number of never-executed NOPs results in the same regression.
3376 // We leave it off by default.
3378 if ((EmitSync & 65536) != 0) {
3379 Label LSuccess, LGoSlowPath ;
3381 masm.bind (CheckSucc) ;
3383 // Optional pre-test ... it's safe to elide this
3384 if ((EmitSync & 16) == 0) {
3385 masm.cmpptr(Address (tmpReg, ObjectMonitor::succ_offset_in_bytes()-2), 0) ;
3386 masm.jccb (Assembler::zero, LGoSlowPath) ;
3387 }
3389 // We have a classic Dekker-style idiom:
3390 // ST m->_owner = 0 ; MEMBAR; LD m->_succ
3391 // There are a number of ways to implement the barrier:
3392 // (1) lock:andl &m->_owner, 0
3393 // is fast, but mask doesn't currently support the "ANDL M,IMM32" form.
3394 // LOCK: ANDL [ebx+Offset(_Owner)-2], 0
3395 // Encodes as 81 31 OFF32 IMM32 or 83 63 OFF8 IMM8
3396 // (2) If supported, an explicit MFENCE is appealing.
3397 // In older IA32 processors MFENCE is slower than lock:add or xchg
3398 // particularly if the write-buffer is full as might be the case if
3399 // if stores closely precede the fence or fence-equivalent instruction.
3400 // In more modern implementations MFENCE appears faster, however.
3401 // (3) In lieu of an explicit fence, use lock:addl to the top-of-stack
3402 // The $lines underlying the top-of-stack should be in M-state.
3403 // The locked add instruction is serializing, of course.
3404 // (4) Use xchg, which is serializing
3405 // mov boxReg, 0; xchgl boxReg, [tmpReg + Offset(_owner)-2] also works
3406 // (5) ST m->_owner = 0 and then execute lock:orl &m->_succ, 0.
3407 // The integer condition codes will tell us if succ was 0.
3408 // Since _succ and _owner should reside in the same $line and
3409 // we just stored into _owner, it's likely that the $line
3410 // remains in M-state for the lock:orl.
3411 //
3412 // We currently use (3), although it's likely that switching to (2)
3413 // is correct for the future.
3415 masm.movptr(Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2), NULL_WORD) ;
3416 if (os::is_MP()) {
3417 if (VM_Version::supports_sse2() && 1 == FenceInstruction) {
3418 masm.mfence();
3419 } else {
3420 masm.lock () ; masm.addptr(Address(rsp, 0), 0) ;
3421 }
3422 }
3423 // Ratify _succ remains non-null
3424 masm.cmpptr(Address (tmpReg, ObjectMonitor::succ_offset_in_bytes()-2), 0) ;
3425 masm.jccb (Assembler::notZero, LSuccess) ;
3427 masm.xorptr(boxReg, boxReg) ; // box is really EAX
3428 if (os::is_MP()) { masm.lock(); }
3429 masm.cmpxchgptr(rsp, Address(tmpReg, ObjectMonitor::owner_offset_in_bytes()-2));
3430 masm.jccb (Assembler::notEqual, LSuccess) ;
3431 // Since we're low on registers we installed rsp as a placeholding in _owner.
3432 // Now install Self over rsp. This is safe as we're transitioning from
3433 // non-null to non=null
3434 masm.get_thread (boxReg) ;
3435 masm.movptr(Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2), boxReg) ;
3436 // Intentional fall-through into LGoSlowPath ...
3438 masm.bind (LGoSlowPath) ;
3439 masm.orptr(boxReg, 1) ; // set ICC.ZF=0 to indicate failure
3440 masm.jmpb (DONE_LABEL) ;
3442 masm.bind (LSuccess) ;
3443 masm.xorptr(boxReg, boxReg) ; // set ICC.ZF=1 to indicate success
3444 masm.jmpb (DONE_LABEL) ;
3445 }
3447 masm.bind (Stacked) ;
3448 // It's not inflated and it's not recursively stack-locked and it's not biased.
3449 // It must be stack-locked.
3450 // Try to reset the header to displaced header.
3451 // The "box" value on the stack is stable, so we can reload
3452 // and be assured we observe the same value as above.
3453 masm.movptr(tmpReg, Address(boxReg, 0)) ;
3454 if (os::is_MP()) { masm.lock(); }
3455 masm.cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses EAX which is box
3456 // Intention fall-thru into DONE_LABEL
3459 // DONE_LABEL is a hot target - we'd really like to place it at the
3460 // start of cache line by padding with NOPs.
3461 // See the AMD and Intel software optimization manuals for the
3462 // most efficient "long" NOP encodings.
3463 // Unfortunately none of our alignment mechanisms suffice.
3464 if ((EmitSync & 65536) == 0) {
3465 masm.bind (CheckSucc) ;
3466 }
3467 masm.bind(DONE_LABEL);
3469 // Avoid branch to branch on AMD processors
3470 if (EmitSync & 32768) { masm.nop() ; }
3471 }
3472 %}
3475 enc_class enc_pop_rdx() %{
3476 emit_opcode(cbuf,0x5A);
3477 %}
3479 enc_class enc_rethrow() %{
3480 cbuf.set_insts_mark();
3481 emit_opcode(cbuf, 0xE9); // jmp entry
3482 emit_d32_reloc(cbuf, (int)OptoRuntime::rethrow_stub() - ((int)cbuf.insts_end())-4,
3483 runtime_call_Relocation::spec(), RELOC_IMM32 );
3484 %}
3487 // Convert a double to an int. Java semantics require we do complex
3488 // manglelations in the corner cases. So we set the rounding mode to
3489 // 'zero', store the darned double down as an int, and reset the
3490 // rounding mode to 'nearest'. The hardware throws an exception which
3491 // patches up the correct value directly to the stack.
3492 enc_class DPR2I_encoding( regDPR src ) %{
3493 // Flip to round-to-zero mode. We attempted to allow invalid-op
3494 // exceptions here, so that a NAN or other corner-case value will
3495 // thrown an exception (but normal values get converted at full speed).
3496 // However, I2C adapters and other float-stack manglers leave pending
3497 // invalid-op exceptions hanging. We would have to clear them before
3498 // enabling them and that is more expensive than just testing for the
3499 // invalid value Intel stores down in the corner cases.
3500 emit_opcode(cbuf,0xD9); // FLDCW trunc
3501 emit_opcode(cbuf,0x2D);
3502 emit_d32(cbuf,(int)StubRoutines::addr_fpu_cntrl_wrd_trunc());
3503 // Allocate a word
3504 emit_opcode(cbuf,0x83); // SUB ESP,4
3505 emit_opcode(cbuf,0xEC);
3506 emit_d8(cbuf,0x04);
3507 // Encoding assumes a double has been pushed into FPR0.
3508 // Store down the double as an int, popping the FPU stack
3509 emit_opcode(cbuf,0xDB); // FISTP [ESP]
3510 emit_opcode(cbuf,0x1C);
3511 emit_d8(cbuf,0x24);
3512 // Restore the rounding mode; mask the exception
3513 emit_opcode(cbuf,0xD9); // FLDCW std/24-bit mode
3514 emit_opcode(cbuf,0x2D);
3515 emit_d32( cbuf, Compile::current()->in_24_bit_fp_mode()
3516 ? (int)StubRoutines::addr_fpu_cntrl_wrd_24()
3517 : (int)StubRoutines::addr_fpu_cntrl_wrd_std());
3519 // Load the converted int; adjust CPU stack
3520 emit_opcode(cbuf,0x58); // POP EAX
3521 emit_opcode(cbuf,0x3D); // CMP EAX,imm
3522 emit_d32 (cbuf,0x80000000); // 0x80000000
3523 emit_opcode(cbuf,0x75); // JNE around_slow_call
3524 emit_d8 (cbuf,0x07); // Size of slow_call
3525 // Push src onto stack slow-path
3526 emit_opcode(cbuf,0xD9 ); // FLD ST(i)
3527 emit_d8 (cbuf,0xC0-1+$src$$reg );
3528 // CALL directly to the runtime
3529 cbuf.set_insts_mark();
3530 emit_opcode(cbuf,0xE8); // Call into runtime
3531 emit_d32_reloc(cbuf, (StubRoutines::d2i_wrapper() - cbuf.insts_end()) - 4, runtime_call_Relocation::spec(), RELOC_IMM32 );
3532 // Carry on here...
3533 %}
3535 enc_class DPR2L_encoding( regDPR src ) %{
3536 emit_opcode(cbuf,0xD9); // FLDCW trunc
3537 emit_opcode(cbuf,0x2D);
3538 emit_d32(cbuf,(int)StubRoutines::addr_fpu_cntrl_wrd_trunc());
3539 // Allocate a word
3540 emit_opcode(cbuf,0x83); // SUB ESP,8
3541 emit_opcode(cbuf,0xEC);
3542 emit_d8(cbuf,0x08);
3543 // Encoding assumes a double has been pushed into FPR0.
3544 // Store down the double as a long, popping the FPU stack
3545 emit_opcode(cbuf,0xDF); // FISTP [ESP]
3546 emit_opcode(cbuf,0x3C);
3547 emit_d8(cbuf,0x24);
3548 // Restore the rounding mode; mask the exception
3549 emit_opcode(cbuf,0xD9); // FLDCW std/24-bit mode
3550 emit_opcode(cbuf,0x2D);
3551 emit_d32( cbuf, Compile::current()->in_24_bit_fp_mode()
3552 ? (int)StubRoutines::addr_fpu_cntrl_wrd_24()
3553 : (int)StubRoutines::addr_fpu_cntrl_wrd_std());
3555 // Load the converted int; adjust CPU stack
3556 emit_opcode(cbuf,0x58); // POP EAX
3557 emit_opcode(cbuf,0x5A); // POP EDX
3558 emit_opcode(cbuf,0x81); // CMP EDX,imm
3559 emit_d8 (cbuf,0xFA); // rdx
3560 emit_d32 (cbuf,0x80000000); // 0x80000000
3561 emit_opcode(cbuf,0x75); // JNE around_slow_call
3562 emit_d8 (cbuf,0x07+4); // Size of slow_call
3563 emit_opcode(cbuf,0x85); // TEST EAX,EAX
3564 emit_opcode(cbuf,0xC0); // 2/rax,/rax,
3565 emit_opcode(cbuf,0x75); // JNE around_slow_call
3566 emit_d8 (cbuf,0x07); // Size of slow_call
3567 // Push src onto stack slow-path
3568 emit_opcode(cbuf,0xD9 ); // FLD ST(i)
3569 emit_d8 (cbuf,0xC0-1+$src$$reg );
3570 // CALL directly to the runtime
3571 cbuf.set_insts_mark();
3572 emit_opcode(cbuf,0xE8); // Call into runtime
3573 emit_d32_reloc(cbuf, (StubRoutines::d2l_wrapper() - cbuf.insts_end()) - 4, runtime_call_Relocation::spec(), RELOC_IMM32 );
3574 // Carry on here...
3575 %}
3577 enc_class FMul_ST_reg( eRegFPR src1 ) %{
3578 // Operand was loaded from memory into fp ST (stack top)
3579 // FMUL ST,$src /* D8 C8+i */
3580 emit_opcode(cbuf, 0xD8);
3581 emit_opcode(cbuf, 0xC8 + $src1$$reg);
3582 %}
3584 enc_class FAdd_ST_reg( eRegFPR src2 ) %{
3585 // FADDP ST,src2 /* D8 C0+i */
3586 emit_opcode(cbuf, 0xD8);
3587 emit_opcode(cbuf, 0xC0 + $src2$$reg);
3588 //could use FADDP src2,fpST /* DE C0+i */
3589 %}
3591 enc_class FAddP_reg_ST( eRegFPR src2 ) %{
3592 // FADDP src2,ST /* DE C0+i */
3593 emit_opcode(cbuf, 0xDE);
3594 emit_opcode(cbuf, 0xC0 + $src2$$reg);
3595 %}
3597 enc_class subFPR_divFPR_encode( eRegFPR src1, eRegFPR src2) %{
3598 // Operand has been loaded into fp ST (stack top)
3599 // FSUB ST,$src1
3600 emit_opcode(cbuf, 0xD8);
3601 emit_opcode(cbuf, 0xE0 + $src1$$reg);
3603 // FDIV
3604 emit_opcode(cbuf, 0xD8);
3605 emit_opcode(cbuf, 0xF0 + $src2$$reg);
3606 %}
3608 enc_class MulFAddF (eRegFPR src1, eRegFPR src2) %{
3609 // Operand was loaded from memory into fp ST (stack top)
3610 // FADD ST,$src /* D8 C0+i */
3611 emit_opcode(cbuf, 0xD8);
3612 emit_opcode(cbuf, 0xC0 + $src1$$reg);
3614 // FMUL ST,src2 /* D8 C*+i */
3615 emit_opcode(cbuf, 0xD8);
3616 emit_opcode(cbuf, 0xC8 + $src2$$reg);
3617 %}
3620 enc_class MulFAddFreverse (eRegFPR src1, eRegFPR src2) %{
3621 // Operand was loaded from memory into fp ST (stack top)
3622 // FADD ST,$src /* D8 C0+i */
3623 emit_opcode(cbuf, 0xD8);
3624 emit_opcode(cbuf, 0xC0 + $src1$$reg);
3626 // FMULP src2,ST /* DE C8+i */
3627 emit_opcode(cbuf, 0xDE);
3628 emit_opcode(cbuf, 0xC8 + $src2$$reg);
3629 %}
3631 // Atomically load the volatile long
3632 enc_class enc_loadL_volatile( memory mem, stackSlotL dst ) %{
3633 emit_opcode(cbuf,0xDF);
3634 int rm_byte_opcode = 0x05;
3635 int base = $mem$$base;
3636 int index = $mem$$index;
3637 int scale = $mem$$scale;
3638 int displace = $mem$$disp;
3639 relocInfo::relocType disp_reloc = $mem->disp_reloc(); // disp-as-oop when working with static globals
3640 encode_RegMem(cbuf, rm_byte_opcode, base, index, scale, displace, disp_reloc);
3641 store_to_stackslot( cbuf, 0x0DF, 0x07, $dst$$disp );
3642 %}
3644 // Volatile Store Long. Must be atomic, so move it into
3645 // the FP TOS and then do a 64-bit FIST. Has to probe the
3646 // target address before the store (for null-ptr checks)
3647 // so the memory operand is used twice in the encoding.
3648 enc_class enc_storeL_volatile( memory mem, stackSlotL src ) %{
3649 store_to_stackslot( cbuf, 0x0DF, 0x05, $src$$disp );
3650 cbuf.set_insts_mark(); // Mark start of FIST in case $mem has an oop
3651 emit_opcode(cbuf,0xDF);
3652 int rm_byte_opcode = 0x07;
3653 int base = $mem$$base;
3654 int index = $mem$$index;
3655 int scale = $mem$$scale;
3656 int displace = $mem$$disp;
3657 relocInfo::relocType disp_reloc = $mem->disp_reloc(); // disp-as-oop when working with static globals
3658 encode_RegMem(cbuf, rm_byte_opcode, base, index, scale, displace, disp_reloc);
3659 %}
3661 // Safepoint Poll. This polls the safepoint page, and causes an
3662 // exception if it is not readable. Unfortunately, it kills the condition code
3663 // in the process
3664 // We current use TESTL [spp],EDI
3665 // A better choice might be TESTB [spp + pagesize() - CacheLineSize()],0
3667 enc_class Safepoint_Poll() %{
3668 cbuf.relocate(cbuf.insts_mark(), relocInfo::poll_type, 0);
3669 emit_opcode(cbuf,0x85);
3670 emit_rm (cbuf, 0x0, 0x7, 0x5);
3671 emit_d32(cbuf, (intptr_t)os::get_polling_page());
3672 %}
3673 %}
3676 //----------FRAME--------------------------------------------------------------
3677 // Definition of frame structure and management information.
3678 //
3679 // S T A C K L A Y O U T Allocators stack-slot number
3680 // | (to get allocators register number
3681 // G Owned by | | v add OptoReg::stack0())
3682 // r CALLER | |
3683 // o | +--------+ pad to even-align allocators stack-slot
3684 // w V | pad0 | numbers; owned by CALLER
3685 // t -----------+--------+----> Matcher::_in_arg_limit, unaligned
3686 // h ^ | in | 5
3687 // | | args | 4 Holes in incoming args owned by SELF
3688 // | | | | 3
3689 // | | +--------+
3690 // V | | old out| Empty on Intel, window on Sparc
3691 // | old |preserve| Must be even aligned.
3692 // | SP-+--------+----> Matcher::_old_SP, even aligned
3693 // | | in | 3 area for Intel ret address
3694 // Owned by |preserve| Empty on Sparc.
3695 // SELF +--------+
3696 // | | pad2 | 2 pad to align old SP
3697 // | +--------+ 1
3698 // | | locks | 0
3699 // | +--------+----> OptoReg::stack0(), even aligned
3700 // | | pad1 | 11 pad to align new SP
3701 // | +--------+
3702 // | | | 10
3703 // | | spills | 9 spills
3704 // V | | 8 (pad0 slot for callee)
3705 // -----------+--------+----> Matcher::_out_arg_limit, unaligned
3706 // ^ | out | 7
3707 // | | args | 6 Holes in outgoing args owned by CALLEE
3708 // Owned by +--------+
3709 // CALLEE | new out| 6 Empty on Intel, window on Sparc
3710 // | new |preserve| Must be even-aligned.
3711 // | SP-+--------+----> Matcher::_new_SP, even aligned
3712 // | | |
3713 //
3714 // Note 1: Only region 8-11 is determined by the allocator. Region 0-5 is
3715 // known from SELF's arguments and the Java calling convention.
3716 // Region 6-7 is determined per call site.
3717 // Note 2: If the calling convention leaves holes in the incoming argument
3718 // area, those holes are owned by SELF. Holes in the outgoing area
3719 // are owned by the CALLEE. Holes should not be nessecary in the
3720 // incoming area, as the Java calling convention is completely under
3721 // the control of the AD file. Doubles can be sorted and packed to
3722 // avoid holes. Holes in the outgoing arguments may be nessecary for
3723 // varargs C calling conventions.
3724 // Note 3: Region 0-3 is even aligned, with pad2 as needed. Region 3-5 is
3725 // even aligned with pad0 as needed.
3726 // Region 6 is even aligned. Region 6-7 is NOT even aligned;
3727 // region 6-11 is even aligned; it may be padded out more so that
3728 // the region from SP to FP meets the minimum stack alignment.
3730 frame %{
3731 // What direction does stack grow in (assumed to be same for C & Java)
3732 stack_direction(TOWARDS_LOW);
3734 // These three registers define part of the calling convention
3735 // between compiled code and the interpreter.
3736 inline_cache_reg(EAX); // Inline Cache Register
3737 interpreter_method_oop_reg(EBX); // Method Oop Register when calling interpreter
3739 // Optional: name the operand used by cisc-spilling to access [stack_pointer + offset]
3740 cisc_spilling_operand_name(indOffset32);
3742 // Number of stack slots consumed by locking an object
3743 sync_stack_slots(1);
3745 // Compiled code's Frame Pointer
3746 frame_pointer(ESP);
3747 // Interpreter stores its frame pointer in a register which is
3748 // stored to the stack by I2CAdaptors.
3749 // I2CAdaptors convert from interpreted java to compiled java.
3750 interpreter_frame_pointer(EBP);
3752 // Stack alignment requirement
3753 // Alignment size in bytes (128-bit -> 16 bytes)
3754 stack_alignment(StackAlignmentInBytes);
3756 // Number of stack slots between incoming argument block and the start of
3757 // a new frame. The PROLOG must add this many slots to the stack. The
3758 // EPILOG must remove this many slots. Intel needs one slot for
3759 // return address and one for rbp, (must save rbp)
3760 in_preserve_stack_slots(2+VerifyStackAtCalls);
3762 // Number of outgoing stack slots killed above the out_preserve_stack_slots
3763 // for calls to C. Supports the var-args backing area for register parms.
3764 varargs_C_out_slots_killed(0);
3766 // The after-PROLOG location of the return address. Location of
3767 // return address specifies a type (REG or STACK) and a number
3768 // representing the register number (i.e. - use a register name) or
3769 // stack slot.
3770 // Ret Addr is on stack in slot 0 if no locks or verification or alignment.
3771 // Otherwise, it is above the locks and verification slot and alignment word
3772 return_addr(STACK - 1 +
3773 round_to((Compile::current()->in_preserve_stack_slots() +
3774 Compile::current()->fixed_slots()),
3775 stack_alignment_in_slots()));
3777 // Body of function which returns an integer array locating
3778 // arguments either in registers or in stack slots. Passed an array
3779 // of ideal registers called "sig" and a "length" count. Stack-slot
3780 // offsets are based on outgoing arguments, i.e. a CALLER setting up
3781 // arguments for a CALLEE. Incoming stack arguments are
3782 // automatically biased by the preserve_stack_slots field above.
3783 calling_convention %{
3784 // No difference between ingoing/outgoing just pass false
3785 SharedRuntime::java_calling_convention(sig_bt, regs, length, false);
3786 %}
3789 // Body of function which returns an integer array locating
3790 // arguments either in registers or in stack slots. Passed an array
3791 // of ideal registers called "sig" and a "length" count. Stack-slot
3792 // offsets are based on outgoing arguments, i.e. a CALLER setting up
3793 // arguments for a CALLEE. Incoming stack arguments are
3794 // automatically biased by the preserve_stack_slots field above.
3795 c_calling_convention %{
3796 // This is obviously always outgoing
3797 (void) SharedRuntime::c_calling_convention(sig_bt, regs, length);
3798 %}
3800 // Location of C & interpreter return values
3801 c_return_value %{
3802 assert( ideal_reg >= Op_RegI && ideal_reg <= Op_RegL, "only return normal values" );
3803 static int lo[Op_RegL+1] = { 0, 0, OptoReg::Bad, EAX_num, EAX_num, FPR1L_num, FPR1L_num, EAX_num };
3804 static int hi[Op_RegL+1] = { 0, 0, OptoReg::Bad, OptoReg::Bad, OptoReg::Bad, OptoReg::Bad, FPR1H_num, EDX_num };
3806 // in SSE2+ mode we want to keep the FPU stack clean so pretend
3807 // that C functions return float and double results in XMM0.
3808 if( ideal_reg == Op_RegD && UseSSE>=2 )
3809 return OptoRegPair(XMM0b_num,XMM0_num);
3810 if( ideal_reg == Op_RegF && UseSSE>=2 )
3811 return OptoRegPair(OptoReg::Bad,XMM0_num);
3813 return OptoRegPair(hi[ideal_reg],lo[ideal_reg]);
3814 %}
3816 // Location of return values
3817 return_value %{
3818 assert( ideal_reg >= Op_RegI && ideal_reg <= Op_RegL, "only return normal values" );
3819 static int lo[Op_RegL+1] = { 0, 0, OptoReg::Bad, EAX_num, EAX_num, FPR1L_num, FPR1L_num, EAX_num };
3820 static int hi[Op_RegL+1] = { 0, 0, OptoReg::Bad, OptoReg::Bad, OptoReg::Bad, OptoReg::Bad, FPR1H_num, EDX_num };
3821 if( ideal_reg == Op_RegD && UseSSE>=2 )
3822 return OptoRegPair(XMM0b_num,XMM0_num);
3823 if( ideal_reg == Op_RegF && UseSSE>=1 )
3824 return OptoRegPair(OptoReg::Bad,XMM0_num);
3825 return OptoRegPair(hi[ideal_reg],lo[ideal_reg]);
3826 %}
3828 %}
3830 //----------ATTRIBUTES---------------------------------------------------------
3831 //----------Operand Attributes-------------------------------------------------
3832 op_attrib op_cost(0); // Required cost attribute
3834 //----------Instruction Attributes---------------------------------------------
3835 ins_attrib ins_cost(100); // Required cost attribute
3836 ins_attrib ins_size(8); // Required size attribute (in bits)
3837 ins_attrib ins_short_branch(0); // Required flag: is this instruction a
3838 // non-matching short branch variant of some
3839 // long branch?
3840 ins_attrib ins_alignment(1); // Required alignment attribute (must be a power of 2)
3841 // specifies the alignment that some part of the instruction (not
3842 // necessarily the start) requires. If > 1, a compute_padding()
3843 // function must be provided for the instruction
3845 //----------OPERANDS-----------------------------------------------------------
3846 // Operand definitions must precede instruction definitions for correct parsing
3847 // in the ADLC because operands constitute user defined types which are used in
3848 // instruction definitions.
3850 //----------Simple Operands----------------------------------------------------
3851 // Immediate Operands
3852 // Integer Immediate
3853 operand immI() %{
3854 match(ConI);
3856 op_cost(10);
3857 format %{ %}
3858 interface(CONST_INTER);
3859 %}
3861 // Constant for test vs zero
3862 operand immI0() %{
3863 predicate(n->get_int() == 0);
3864 match(ConI);
3866 op_cost(0);
3867 format %{ %}
3868 interface(CONST_INTER);
3869 %}
3871 // Constant for increment
3872 operand immI1() %{
3873 predicate(n->get_int() == 1);
3874 match(ConI);
3876 op_cost(0);
3877 format %{ %}
3878 interface(CONST_INTER);
3879 %}
3881 // Constant for decrement
3882 operand immI_M1() %{
3883 predicate(n->get_int() == -1);
3884 match(ConI);
3886 op_cost(0);
3887 format %{ %}
3888 interface(CONST_INTER);
3889 %}
3891 // Valid scale values for addressing modes
3892 operand immI2() %{
3893 predicate(0 <= n->get_int() && (n->get_int() <= 3));
3894 match(ConI);
3896 format %{ %}
3897 interface(CONST_INTER);
3898 %}
3900 operand immI8() %{
3901 predicate((-128 <= n->get_int()) && (n->get_int() <= 127));
3902 match(ConI);
3904 op_cost(5);
3905 format %{ %}
3906 interface(CONST_INTER);
3907 %}
3909 operand immI16() %{
3910 predicate((-32768 <= n->get_int()) && (n->get_int() <= 32767));
3911 match(ConI);
3913 op_cost(10);
3914 format %{ %}
3915 interface(CONST_INTER);
3916 %}
3918 // Constant for long shifts
3919 operand immI_32() %{
3920 predicate( n->get_int() == 32 );
3921 match(ConI);
3923 op_cost(0);
3924 format %{ %}
3925 interface(CONST_INTER);
3926 %}
3928 operand immI_1_31() %{
3929 predicate( n->get_int() >= 1 && n->get_int() <= 31 );
3930 match(ConI);
3932 op_cost(0);
3933 format %{ %}
3934 interface(CONST_INTER);
3935 %}
3937 operand immI_32_63() %{
3938 predicate( n->get_int() >= 32 && n->get_int() <= 63 );
3939 match(ConI);
3940 op_cost(0);
3942 format %{ %}
3943 interface(CONST_INTER);
3944 %}
3946 operand immI_1() %{
3947 predicate( n->get_int() == 1 );
3948 match(ConI);
3950 op_cost(0);
3951 format %{ %}
3952 interface(CONST_INTER);
3953 %}
3955 operand immI_2() %{
3956 predicate( n->get_int() == 2 );
3957 match(ConI);
3959 op_cost(0);
3960 format %{ %}
3961 interface(CONST_INTER);
3962 %}
3964 operand immI_3() %{
3965 predicate( n->get_int() == 3 );
3966 match(ConI);
3968 op_cost(0);
3969 format %{ %}
3970 interface(CONST_INTER);
3971 %}
3973 // Pointer Immediate
3974 operand immP() %{
3975 match(ConP);
3977 op_cost(10);
3978 format %{ %}
3979 interface(CONST_INTER);
3980 %}
3982 // NULL Pointer Immediate
3983 operand immP0() %{
3984 predicate( n->get_ptr() == 0 );
3985 match(ConP);
3986 op_cost(0);
3988 format %{ %}
3989 interface(CONST_INTER);
3990 %}
3992 // Long Immediate
3993 operand immL() %{
3994 match(ConL);
3996 op_cost(20);
3997 format %{ %}
3998 interface(CONST_INTER);
3999 %}
4001 // Long Immediate zero
4002 operand immL0() %{
4003 predicate( n->get_long() == 0L );
4004 match(ConL);
4005 op_cost(0);
4007 format %{ %}
4008 interface(CONST_INTER);
4009 %}
4011 // Long Immediate zero
4012 operand immL_M1() %{
4013 predicate( n->get_long() == -1L );
4014 match(ConL);
4015 op_cost(0);
4017 format %{ %}
4018 interface(CONST_INTER);
4019 %}
4021 // Long immediate from 0 to 127.
4022 // Used for a shorter form of long mul by 10.
4023 operand immL_127() %{
4024 predicate((0 <= n->get_long()) && (n->get_long() <= 127));
4025 match(ConL);
4026 op_cost(0);
4028 format %{ %}
4029 interface(CONST_INTER);
4030 %}
4032 // Long Immediate: low 32-bit mask
4033 operand immL_32bits() %{
4034 predicate(n->get_long() == 0xFFFFFFFFL);
4035 match(ConL);
4036 op_cost(0);
4038 format %{ %}
4039 interface(CONST_INTER);
4040 %}
4042 // Long Immediate: low 32-bit mask
4043 operand immL32() %{
4044 predicate(n->get_long() == (int)(n->get_long()));
4045 match(ConL);
4046 op_cost(20);
4048 format %{ %}
4049 interface(CONST_INTER);
4050 %}
4052 //Double Immediate zero
4053 operand immDPR0() %{
4054 // Do additional (and counter-intuitive) test against NaN to work around VC++
4055 // bug that generates code such that NaNs compare equal to 0.0
4056 predicate( UseSSE<=1 && n->getd() == 0.0 && !g_isnan(n->getd()) );
4057 match(ConD);
4059 op_cost(5);
4060 format %{ %}
4061 interface(CONST_INTER);
4062 %}
4064 // Double Immediate one
4065 operand immDPR1() %{
4066 predicate( UseSSE<=1 && n->getd() == 1.0 );
4067 match(ConD);
4069 op_cost(5);
4070 format %{ %}
4071 interface(CONST_INTER);
4072 %}
4074 // Double Immediate
4075 operand immDPR() %{
4076 predicate(UseSSE<=1);
4077 match(ConD);
4079 op_cost(5);
4080 format %{ %}
4081 interface(CONST_INTER);
4082 %}
4084 operand immD() %{
4085 predicate(UseSSE>=2);
4086 match(ConD);
4088 op_cost(5);
4089 format %{ %}
4090 interface(CONST_INTER);
4091 %}
4093 // Double Immediate zero
4094 operand immD0() %{
4095 // Do additional (and counter-intuitive) test against NaN to work around VC++
4096 // bug that generates code such that NaNs compare equal to 0.0 AND do not
4097 // compare equal to -0.0.
4098 predicate( UseSSE>=2 && jlong_cast(n->getd()) == 0 );
4099 match(ConD);
4101 format %{ %}
4102 interface(CONST_INTER);
4103 %}
4105 // Float Immediate zero
4106 operand immFPR0() %{
4107 predicate(UseSSE == 0 && n->getf() == 0.0F);
4108 match(ConF);
4110 op_cost(5);
4111 format %{ %}
4112 interface(CONST_INTER);
4113 %}
4115 // Float Immediate one
4116 operand immFPR1() %{
4117 predicate(UseSSE == 0 && n->getf() == 1.0F);
4118 match(ConF);
4120 op_cost(5);
4121 format %{ %}
4122 interface(CONST_INTER);
4123 %}
4125 // Float Immediate
4126 operand immFPR() %{
4127 predicate( UseSSE == 0 );
4128 match(ConF);
4130 op_cost(5);
4131 format %{ %}
4132 interface(CONST_INTER);
4133 %}
4135 // Float Immediate
4136 operand immF() %{
4137 predicate(UseSSE >= 1);
4138 match(ConF);
4140 op_cost(5);
4141 format %{ %}
4142 interface(CONST_INTER);
4143 %}
4145 // Float Immediate zero. Zero and not -0.0
4146 operand immF0() %{
4147 predicate( UseSSE >= 1 && jint_cast(n->getf()) == 0 );
4148 match(ConF);
4150 op_cost(5);
4151 format %{ %}
4152 interface(CONST_INTER);
4153 %}
4155 // Immediates for special shifts (sign extend)
4157 // Constants for increment
4158 operand immI_16() %{
4159 predicate( n->get_int() == 16 );
4160 match(ConI);
4162 format %{ %}
4163 interface(CONST_INTER);
4164 %}
4166 operand immI_24() %{
4167 predicate( n->get_int() == 24 );
4168 match(ConI);
4170 format %{ %}
4171 interface(CONST_INTER);
4172 %}
4174 // Constant for byte-wide masking
4175 operand immI_255() %{
4176 predicate( n->get_int() == 255 );
4177 match(ConI);
4179 format %{ %}
4180 interface(CONST_INTER);
4181 %}
4183 // Constant for short-wide masking
4184 operand immI_65535() %{
4185 predicate(n->get_int() == 65535);
4186 match(ConI);
4188 format %{ %}
4189 interface(CONST_INTER);
4190 %}
4192 // Register Operands
4193 // Integer Register
4194 operand rRegI() %{
4195 constraint(ALLOC_IN_RC(int_reg));
4196 match(RegI);
4197 match(xRegI);
4198 match(eAXRegI);
4199 match(eBXRegI);
4200 match(eCXRegI);
4201 match(eDXRegI);
4202 match(eDIRegI);
4203 match(eSIRegI);
4205 format %{ %}
4206 interface(REG_INTER);
4207 %}
4209 // Subset of Integer Register
4210 operand xRegI(rRegI reg) %{
4211 constraint(ALLOC_IN_RC(int_x_reg));
4212 match(reg);
4213 match(eAXRegI);
4214 match(eBXRegI);
4215 match(eCXRegI);
4216 match(eDXRegI);
4218 format %{ %}
4219 interface(REG_INTER);
4220 %}
4222 // Special Registers
4223 operand eAXRegI(xRegI reg) %{
4224 constraint(ALLOC_IN_RC(eax_reg));
4225 match(reg);
4226 match(rRegI);
4228 format %{ "EAX" %}
4229 interface(REG_INTER);
4230 %}
4232 // Special Registers
4233 operand eBXRegI(xRegI reg) %{
4234 constraint(ALLOC_IN_RC(ebx_reg));
4235 match(reg);
4236 match(rRegI);
4238 format %{ "EBX" %}
4239 interface(REG_INTER);
4240 %}
4242 operand eCXRegI(xRegI reg) %{
4243 constraint(ALLOC_IN_RC(ecx_reg));
4244 match(reg);
4245 match(rRegI);
4247 format %{ "ECX" %}
4248 interface(REG_INTER);
4249 %}
4251 operand eDXRegI(xRegI reg) %{
4252 constraint(ALLOC_IN_RC(edx_reg));
4253 match(reg);
4254 match(rRegI);
4256 format %{ "EDX" %}
4257 interface(REG_INTER);
4258 %}
4260 operand eDIRegI(xRegI reg) %{
4261 constraint(ALLOC_IN_RC(edi_reg));
4262 match(reg);
4263 match(rRegI);
4265 format %{ "EDI" %}
4266 interface(REG_INTER);
4267 %}
4269 operand naxRegI() %{
4270 constraint(ALLOC_IN_RC(nax_reg));
4271 match(RegI);
4272 match(eCXRegI);
4273 match(eDXRegI);
4274 match(eSIRegI);
4275 match(eDIRegI);
4277 format %{ %}
4278 interface(REG_INTER);
4279 %}
4281 operand nadxRegI() %{
4282 constraint(ALLOC_IN_RC(nadx_reg));
4283 match(RegI);
4284 match(eBXRegI);
4285 match(eCXRegI);
4286 match(eSIRegI);
4287 match(eDIRegI);
4289 format %{ %}
4290 interface(REG_INTER);
4291 %}
4293 operand ncxRegI() %{
4294 constraint(ALLOC_IN_RC(ncx_reg));
4295 match(RegI);
4296 match(eAXRegI);
4297 match(eDXRegI);
4298 match(eSIRegI);
4299 match(eDIRegI);
4301 format %{ %}
4302 interface(REG_INTER);
4303 %}
4305 // // This operand was used by cmpFastUnlock, but conflicted with 'object' reg
4306 // //
4307 operand eSIRegI(xRegI reg) %{
4308 constraint(ALLOC_IN_RC(esi_reg));
4309 match(reg);
4310 match(rRegI);
4312 format %{ "ESI" %}
4313 interface(REG_INTER);
4314 %}
4316 // Pointer Register
4317 operand anyRegP() %{
4318 constraint(ALLOC_IN_RC(any_reg));
4319 match(RegP);
4320 match(eAXRegP);
4321 match(eBXRegP);
4322 match(eCXRegP);
4323 match(eDIRegP);
4324 match(eRegP);
4326 format %{ %}
4327 interface(REG_INTER);
4328 %}
4330 operand eRegP() %{
4331 constraint(ALLOC_IN_RC(int_reg));
4332 match(RegP);
4333 match(eAXRegP);
4334 match(eBXRegP);
4335 match(eCXRegP);
4336 match(eDIRegP);
4338 format %{ %}
4339 interface(REG_INTER);
4340 %}
4342 // On windows95, EBP is not safe to use for implicit null tests.
4343 operand eRegP_no_EBP() %{
4344 constraint(ALLOC_IN_RC(int_reg_no_rbp));
4345 match(RegP);
4346 match(eAXRegP);
4347 match(eBXRegP);
4348 match(eCXRegP);
4349 match(eDIRegP);
4351 op_cost(100);
4352 format %{ %}
4353 interface(REG_INTER);
4354 %}
4356 operand naxRegP() %{
4357 constraint(ALLOC_IN_RC(nax_reg));
4358 match(RegP);
4359 match(eBXRegP);
4360 match(eDXRegP);
4361 match(eCXRegP);
4362 match(eSIRegP);
4363 match(eDIRegP);
4365 format %{ %}
4366 interface(REG_INTER);
4367 %}
4369 operand nabxRegP() %{
4370 constraint(ALLOC_IN_RC(nabx_reg));
4371 match(RegP);
4372 match(eCXRegP);
4373 match(eDXRegP);
4374 match(eSIRegP);
4375 match(eDIRegP);
4377 format %{ %}
4378 interface(REG_INTER);
4379 %}
4381 operand pRegP() %{
4382 constraint(ALLOC_IN_RC(p_reg));
4383 match(RegP);
4384 match(eBXRegP);
4385 match(eDXRegP);
4386 match(eSIRegP);
4387 match(eDIRegP);
4389 format %{ %}
4390 interface(REG_INTER);
4391 %}
4393 // Special Registers
4394 // Return a pointer value
4395 operand eAXRegP(eRegP reg) %{
4396 constraint(ALLOC_IN_RC(eax_reg));
4397 match(reg);
4398 format %{ "EAX" %}
4399 interface(REG_INTER);
4400 %}
4402 // Used in AtomicAdd
4403 operand eBXRegP(eRegP reg) %{
4404 constraint(ALLOC_IN_RC(ebx_reg));
4405 match(reg);
4406 format %{ "EBX" %}
4407 interface(REG_INTER);
4408 %}
4410 // Tail-call (interprocedural jump) to interpreter
4411 operand eCXRegP(eRegP reg) %{
4412 constraint(ALLOC_IN_RC(ecx_reg));
4413 match(reg);
4414 format %{ "ECX" %}
4415 interface(REG_INTER);
4416 %}
4418 operand eSIRegP(eRegP reg) %{
4419 constraint(ALLOC_IN_RC(esi_reg));
4420 match(reg);
4421 format %{ "ESI" %}
4422 interface(REG_INTER);
4423 %}
4425 // Used in rep stosw
4426 operand eDIRegP(eRegP reg) %{
4427 constraint(ALLOC_IN_RC(edi_reg));
4428 match(reg);
4429 format %{ "EDI" %}
4430 interface(REG_INTER);
4431 %}
4433 operand eBPRegP() %{
4434 constraint(ALLOC_IN_RC(ebp_reg));
4435 match(RegP);
4436 format %{ "EBP" %}
4437 interface(REG_INTER);
4438 %}
4440 operand eRegL() %{
4441 constraint(ALLOC_IN_RC(long_reg));
4442 match(RegL);
4443 match(eADXRegL);
4445 format %{ %}
4446 interface(REG_INTER);
4447 %}
4449 operand eADXRegL( eRegL reg ) %{
4450 constraint(ALLOC_IN_RC(eadx_reg));
4451 match(reg);
4453 format %{ "EDX:EAX" %}
4454 interface(REG_INTER);
4455 %}
4457 operand eBCXRegL( eRegL reg ) %{
4458 constraint(ALLOC_IN_RC(ebcx_reg));
4459 match(reg);
4461 format %{ "EBX:ECX" %}
4462 interface(REG_INTER);
4463 %}
4465 // Special case for integer high multiply
4466 operand eADXRegL_low_only() %{
4467 constraint(ALLOC_IN_RC(eadx_reg));
4468 match(RegL);
4470 format %{ "EAX" %}
4471 interface(REG_INTER);
4472 %}
4474 // Flags register, used as output of compare instructions
4475 operand eFlagsReg() %{
4476 constraint(ALLOC_IN_RC(int_flags));
4477 match(RegFlags);
4479 format %{ "EFLAGS" %}
4480 interface(REG_INTER);
4481 %}
4483 // Flags register, used as output of FLOATING POINT compare instructions
4484 operand eFlagsRegU() %{
4485 constraint(ALLOC_IN_RC(int_flags));
4486 match(RegFlags);
4488 format %{ "EFLAGS_U" %}
4489 interface(REG_INTER);
4490 %}
4492 operand eFlagsRegUCF() %{
4493 constraint(ALLOC_IN_RC(int_flags));
4494 match(RegFlags);
4495 predicate(false);
4497 format %{ "EFLAGS_U_CF" %}
4498 interface(REG_INTER);
4499 %}
4501 // Condition Code Register used by long compare
4502 operand flagsReg_long_LTGE() %{
4503 constraint(ALLOC_IN_RC(int_flags));
4504 match(RegFlags);
4505 format %{ "FLAGS_LTGE" %}
4506 interface(REG_INTER);
4507 %}
4508 operand flagsReg_long_EQNE() %{
4509 constraint(ALLOC_IN_RC(int_flags));
4510 match(RegFlags);
4511 format %{ "FLAGS_EQNE" %}
4512 interface(REG_INTER);
4513 %}
4514 operand flagsReg_long_LEGT() %{
4515 constraint(ALLOC_IN_RC(int_flags));
4516 match(RegFlags);
4517 format %{ "FLAGS_LEGT" %}
4518 interface(REG_INTER);
4519 %}
4521 // Float register operands
4522 operand regDPR() %{
4523 predicate( UseSSE < 2 );
4524 constraint(ALLOC_IN_RC(fp_dbl_reg));
4525 match(RegD);
4526 match(regDPR1);
4527 match(regDPR2);
4528 format %{ %}
4529 interface(REG_INTER);
4530 %}
4532 operand regDPR1(regDPR reg) %{
4533 predicate( UseSSE < 2 );
4534 constraint(ALLOC_IN_RC(fp_dbl_reg0));
4535 match(reg);
4536 format %{ "FPR1" %}
4537 interface(REG_INTER);
4538 %}
4540 operand regDPR2(regDPR reg) %{
4541 predicate( UseSSE < 2 );
4542 constraint(ALLOC_IN_RC(fp_dbl_reg1));
4543 match(reg);
4544 format %{ "FPR2" %}
4545 interface(REG_INTER);
4546 %}
4548 operand regnotDPR1(regDPR reg) %{
4549 predicate( UseSSE < 2 );
4550 constraint(ALLOC_IN_RC(fp_dbl_notreg0));
4551 match(reg);
4552 format %{ %}
4553 interface(REG_INTER);
4554 %}
4556 // Float register operands
4557 operand regFPR() %{
4558 predicate( UseSSE < 2 );
4559 constraint(ALLOC_IN_RC(fp_flt_reg));
4560 match(RegF);
4561 match(regFPR1);
4562 format %{ %}
4563 interface(REG_INTER);
4564 %}
4566 // Float register operands
4567 operand regFPR1(regFPR reg) %{
4568 predicate( UseSSE < 2 );
4569 constraint(ALLOC_IN_RC(fp_flt_reg0));
4570 match(reg);
4571 format %{ "FPR1" %}
4572 interface(REG_INTER);
4573 %}
4575 // XMM Float register operands
4576 operand regF() %{
4577 predicate( UseSSE>=1 );
4578 constraint(ALLOC_IN_RC(float_reg));
4579 match(RegF);
4580 format %{ %}
4581 interface(REG_INTER);
4582 %}
4584 // XMM Double register operands
4585 operand regD() %{
4586 predicate( UseSSE>=2 );
4587 constraint(ALLOC_IN_RC(double_reg));
4588 match(RegD);
4589 format %{ %}
4590 interface(REG_INTER);
4591 %}
4594 //----------Memory Operands----------------------------------------------------
4595 // Direct Memory Operand
4596 operand direct(immP addr) %{
4597 match(addr);
4599 format %{ "[$addr]" %}
4600 interface(MEMORY_INTER) %{
4601 base(0xFFFFFFFF);
4602 index(0x4);
4603 scale(0x0);
4604 disp($addr);
4605 %}
4606 %}
4608 // Indirect Memory Operand
4609 operand indirect(eRegP reg) %{
4610 constraint(ALLOC_IN_RC(int_reg));
4611 match(reg);
4613 format %{ "[$reg]" %}
4614 interface(MEMORY_INTER) %{
4615 base($reg);
4616 index(0x4);
4617 scale(0x0);
4618 disp(0x0);
4619 %}
4620 %}
4622 // Indirect Memory Plus Short Offset Operand
4623 operand indOffset8(eRegP reg, immI8 off) %{
4624 match(AddP reg off);
4626 format %{ "[$reg + $off]" %}
4627 interface(MEMORY_INTER) %{
4628 base($reg);
4629 index(0x4);
4630 scale(0x0);
4631 disp($off);
4632 %}
4633 %}
4635 // Indirect Memory Plus Long Offset Operand
4636 operand indOffset32(eRegP reg, immI off) %{
4637 match(AddP reg off);
4639 format %{ "[$reg + $off]" %}
4640 interface(MEMORY_INTER) %{
4641 base($reg);
4642 index(0x4);
4643 scale(0x0);
4644 disp($off);
4645 %}
4646 %}
4648 // Indirect Memory Plus Long Offset Operand
4649 operand indOffset32X(rRegI reg, immP off) %{
4650 match(AddP off reg);
4652 format %{ "[$reg + $off]" %}
4653 interface(MEMORY_INTER) %{
4654 base($reg);
4655 index(0x4);
4656 scale(0x0);
4657 disp($off);
4658 %}
4659 %}
4661 // Indirect Memory Plus Index Register Plus Offset Operand
4662 operand indIndexOffset(eRegP reg, rRegI ireg, immI off) %{
4663 match(AddP (AddP reg ireg) off);
4665 op_cost(10);
4666 format %{"[$reg + $off + $ireg]" %}
4667 interface(MEMORY_INTER) %{
4668 base($reg);
4669 index($ireg);
4670 scale(0x0);
4671 disp($off);
4672 %}
4673 %}
4675 // Indirect Memory Plus Index Register Plus Offset Operand
4676 operand indIndex(eRegP reg, rRegI ireg) %{
4677 match(AddP reg ireg);
4679 op_cost(10);
4680 format %{"[$reg + $ireg]" %}
4681 interface(MEMORY_INTER) %{
4682 base($reg);
4683 index($ireg);
4684 scale(0x0);
4685 disp(0x0);
4686 %}
4687 %}
4689 // // -------------------------------------------------------------------------
4690 // // 486 architecture doesn't support "scale * index + offset" with out a base
4691 // // -------------------------------------------------------------------------
4692 // // Scaled Memory Operands
4693 // // Indirect Memory Times Scale Plus Offset Operand
4694 // operand indScaleOffset(immP off, rRegI ireg, immI2 scale) %{
4695 // match(AddP off (LShiftI ireg scale));
4696 //
4697 // op_cost(10);
4698 // format %{"[$off + $ireg << $scale]" %}
4699 // interface(MEMORY_INTER) %{
4700 // base(0x4);
4701 // index($ireg);
4702 // scale($scale);
4703 // disp($off);
4704 // %}
4705 // %}
4707 // Indirect Memory Times Scale Plus Index Register
4708 operand indIndexScale(eRegP reg, rRegI ireg, immI2 scale) %{
4709 match(AddP reg (LShiftI ireg scale));
4711 op_cost(10);
4712 format %{"[$reg + $ireg << $scale]" %}
4713 interface(MEMORY_INTER) %{
4714 base($reg);
4715 index($ireg);
4716 scale($scale);
4717 disp(0x0);
4718 %}
4719 %}
4721 // Indirect Memory Times Scale Plus Index Register Plus Offset Operand
4722 operand indIndexScaleOffset(eRegP reg, immI off, rRegI ireg, immI2 scale) %{
4723 match(AddP (AddP reg (LShiftI ireg scale)) off);
4725 op_cost(10);
4726 format %{"[$reg + $off + $ireg << $scale]" %}
4727 interface(MEMORY_INTER) %{
4728 base($reg);
4729 index($ireg);
4730 scale($scale);
4731 disp($off);
4732 %}
4733 %}
4735 //----------Load Long Memory Operands------------------------------------------
4736 // The load-long idiom will use it's address expression again after loading
4737 // the first word of the long. If the load-long destination overlaps with
4738 // registers used in the addressing expression, the 2nd half will be loaded
4739 // from a clobbered address. Fix this by requiring that load-long use
4740 // address registers that do not overlap with the load-long target.
4742 // load-long support
4743 operand load_long_RegP() %{
4744 constraint(ALLOC_IN_RC(esi_reg));
4745 match(RegP);
4746 match(eSIRegP);
4747 op_cost(100);
4748 format %{ %}
4749 interface(REG_INTER);
4750 %}
4752 // Indirect Memory Operand Long
4753 operand load_long_indirect(load_long_RegP reg) %{
4754 constraint(ALLOC_IN_RC(esi_reg));
4755 match(reg);
4757 format %{ "[$reg]" %}
4758 interface(MEMORY_INTER) %{
4759 base($reg);
4760 index(0x4);
4761 scale(0x0);
4762 disp(0x0);
4763 %}
4764 %}
4766 // Indirect Memory Plus Long Offset Operand
4767 operand load_long_indOffset32(load_long_RegP reg, immI off) %{
4768 match(AddP reg off);
4770 format %{ "[$reg + $off]" %}
4771 interface(MEMORY_INTER) %{
4772 base($reg);
4773 index(0x4);
4774 scale(0x0);
4775 disp($off);
4776 %}
4777 %}
4779 opclass load_long_memory(load_long_indirect, load_long_indOffset32);
4782 //----------Special Memory Operands--------------------------------------------
4783 // Stack Slot Operand - This operand is used for loading and storing temporary
4784 // values on the stack where a match requires a value to
4785 // flow through memory.
4786 operand stackSlotP(sRegP reg) %{
4787 constraint(ALLOC_IN_RC(stack_slots));
4788 // No match rule because this operand is only generated in matching
4789 format %{ "[$reg]" %}
4790 interface(MEMORY_INTER) %{
4791 base(0x4); // ESP
4792 index(0x4); // No Index
4793 scale(0x0); // No Scale
4794 disp($reg); // Stack Offset
4795 %}
4796 %}
4798 operand stackSlotI(sRegI reg) %{
4799 constraint(ALLOC_IN_RC(stack_slots));
4800 // No match rule because this operand is only generated in matching
4801 format %{ "[$reg]" %}
4802 interface(MEMORY_INTER) %{
4803 base(0x4); // ESP
4804 index(0x4); // No Index
4805 scale(0x0); // No Scale
4806 disp($reg); // Stack Offset
4807 %}
4808 %}
4810 operand stackSlotF(sRegF reg) %{
4811 constraint(ALLOC_IN_RC(stack_slots));
4812 // No match rule because this operand is only generated in matching
4813 format %{ "[$reg]" %}
4814 interface(MEMORY_INTER) %{
4815 base(0x4); // ESP
4816 index(0x4); // No Index
4817 scale(0x0); // No Scale
4818 disp($reg); // Stack Offset
4819 %}
4820 %}
4822 operand stackSlotD(sRegD reg) %{
4823 constraint(ALLOC_IN_RC(stack_slots));
4824 // No match rule because this operand is only generated in matching
4825 format %{ "[$reg]" %}
4826 interface(MEMORY_INTER) %{
4827 base(0x4); // ESP
4828 index(0x4); // No Index
4829 scale(0x0); // No Scale
4830 disp($reg); // Stack Offset
4831 %}
4832 %}
4834 operand stackSlotL(sRegL reg) %{
4835 constraint(ALLOC_IN_RC(stack_slots));
4836 // No match rule because this operand is only generated in matching
4837 format %{ "[$reg]" %}
4838 interface(MEMORY_INTER) %{
4839 base(0x4); // ESP
4840 index(0x4); // No Index
4841 scale(0x0); // No Scale
4842 disp($reg); // Stack Offset
4843 %}
4844 %}
4846 //----------Memory Operands - Win95 Implicit Null Variants----------------
4847 // Indirect Memory Operand
4848 operand indirect_win95_safe(eRegP_no_EBP reg)
4849 %{
4850 constraint(ALLOC_IN_RC(int_reg));
4851 match(reg);
4853 op_cost(100);
4854 format %{ "[$reg]" %}
4855 interface(MEMORY_INTER) %{
4856 base($reg);
4857 index(0x4);
4858 scale(0x0);
4859 disp(0x0);
4860 %}
4861 %}
4863 // Indirect Memory Plus Short Offset Operand
4864 operand indOffset8_win95_safe(eRegP_no_EBP reg, immI8 off)
4865 %{
4866 match(AddP reg off);
4868 op_cost(100);
4869 format %{ "[$reg + $off]" %}
4870 interface(MEMORY_INTER) %{
4871 base($reg);
4872 index(0x4);
4873 scale(0x0);
4874 disp($off);
4875 %}
4876 %}
4878 // Indirect Memory Plus Long Offset Operand
4879 operand indOffset32_win95_safe(eRegP_no_EBP reg, immI off)
4880 %{
4881 match(AddP reg off);
4883 op_cost(100);
4884 format %{ "[$reg + $off]" %}
4885 interface(MEMORY_INTER) %{
4886 base($reg);
4887 index(0x4);
4888 scale(0x0);
4889 disp($off);
4890 %}
4891 %}
4893 // Indirect Memory Plus Index Register Plus Offset Operand
4894 operand indIndexOffset_win95_safe(eRegP_no_EBP reg, rRegI ireg, immI off)
4895 %{
4896 match(AddP (AddP reg ireg) off);
4898 op_cost(100);
4899 format %{"[$reg + $off + $ireg]" %}
4900 interface(MEMORY_INTER) %{
4901 base($reg);
4902 index($ireg);
4903 scale(0x0);
4904 disp($off);
4905 %}
4906 %}
4908 // Indirect Memory Times Scale Plus Index Register
4909 operand indIndexScale_win95_safe(eRegP_no_EBP reg, rRegI ireg, immI2 scale)
4910 %{
4911 match(AddP reg (LShiftI ireg scale));
4913 op_cost(100);
4914 format %{"[$reg + $ireg << $scale]" %}
4915 interface(MEMORY_INTER) %{
4916 base($reg);
4917 index($ireg);
4918 scale($scale);
4919 disp(0x0);
4920 %}
4921 %}
4923 // Indirect Memory Times Scale Plus Index Register Plus Offset Operand
4924 operand indIndexScaleOffset_win95_safe(eRegP_no_EBP reg, immI off, rRegI ireg, immI2 scale)
4925 %{
4926 match(AddP (AddP reg (LShiftI ireg scale)) off);
4928 op_cost(100);
4929 format %{"[$reg + $off + $ireg << $scale]" %}
4930 interface(MEMORY_INTER) %{
4931 base($reg);
4932 index($ireg);
4933 scale($scale);
4934 disp($off);
4935 %}
4936 %}
4938 //----------Conditional Branch Operands----------------------------------------
4939 // Comparison Op - This is the operation of the comparison, and is limited to
4940 // the following set of codes:
4941 // L (<), LE (<=), G (>), GE (>=), E (==), NE (!=)
4942 //
4943 // Other attributes of the comparison, such as unsignedness, are specified
4944 // by the comparison instruction that sets a condition code flags register.
4945 // That result is represented by a flags operand whose subtype is appropriate
4946 // to the unsignedness (etc.) of the comparison.
4947 //
4948 // Later, the instruction which matches both the Comparison Op (a Bool) and
4949 // the flags (produced by the Cmp) specifies the coding of the comparison op
4950 // by matching a specific subtype of Bool operand below, such as cmpOpU.
4952 // Comparision Code
4953 operand cmpOp() %{
4954 match(Bool);
4956 format %{ "" %}
4957 interface(COND_INTER) %{
4958 equal(0x4, "e");
4959 not_equal(0x5, "ne");
4960 less(0xC, "l");
4961 greater_equal(0xD, "ge");
4962 less_equal(0xE, "le");
4963 greater(0xF, "g");
4964 %}
4965 %}
4967 // Comparison Code, unsigned compare. Used by FP also, with
4968 // C2 (unordered) turned into GT or LT already. The other bits
4969 // C0 and C3 are turned into Carry & Zero flags.
4970 operand cmpOpU() %{
4971 match(Bool);
4973 format %{ "" %}
4974 interface(COND_INTER) %{
4975 equal(0x4, "e");
4976 not_equal(0x5, "ne");
4977 less(0x2, "b");
4978 greater_equal(0x3, "nb");
4979 less_equal(0x6, "be");
4980 greater(0x7, "nbe");
4981 %}
4982 %}
4984 // Floating comparisons that don't require any fixup for the unordered case
4985 operand cmpOpUCF() %{
4986 match(Bool);
4987 predicate(n->as_Bool()->_test._test == BoolTest::lt ||
4988 n->as_Bool()->_test._test == BoolTest::ge ||
4989 n->as_Bool()->_test._test == BoolTest::le ||
4990 n->as_Bool()->_test._test == BoolTest::gt);
4991 format %{ "" %}
4992 interface(COND_INTER) %{
4993 equal(0x4, "e");
4994 not_equal(0x5, "ne");
4995 less(0x2, "b");
4996 greater_equal(0x3, "nb");
4997 less_equal(0x6, "be");
4998 greater(0x7, "nbe");
4999 %}
5000 %}
5003 // Floating comparisons that can be fixed up with extra conditional jumps
5004 operand cmpOpUCF2() %{
5005 match(Bool);
5006 predicate(n->as_Bool()->_test._test == BoolTest::ne ||
5007 n->as_Bool()->_test._test == BoolTest::eq);
5008 format %{ "" %}
5009 interface(COND_INTER) %{
5010 equal(0x4, "e");
5011 not_equal(0x5, "ne");
5012 less(0x2, "b");
5013 greater_equal(0x3, "nb");
5014 less_equal(0x6, "be");
5015 greater(0x7, "nbe");
5016 %}
5017 %}
5019 // Comparison Code for FP conditional move
5020 operand cmpOp_fcmov() %{
5021 match(Bool);
5023 format %{ "" %}
5024 interface(COND_INTER) %{
5025 equal (0x0C8);
5026 not_equal (0x1C8);
5027 less (0x0C0);
5028 greater_equal(0x1C0);
5029 less_equal (0x0D0);
5030 greater (0x1D0);
5031 %}
5032 %}
5034 // Comparision Code used in long compares
5035 operand cmpOp_commute() %{
5036 match(Bool);
5038 format %{ "" %}
5039 interface(COND_INTER) %{
5040 equal(0x4, "e");
5041 not_equal(0x5, "ne");
5042 less(0xF, "g");
5043 greater_equal(0xE, "le");
5044 less_equal(0xD, "ge");
5045 greater(0xC, "l");
5046 %}
5047 %}
5049 //----------OPERAND CLASSES----------------------------------------------------
5050 // Operand Classes are groups of operands that are used as to simplify
5051 // instruction definitions by not requiring the AD writer to specify separate
5052 // instructions for every form of operand when the instruction accepts
5053 // multiple operand types with the same basic encoding and format. The classic
5054 // case of this is memory operands.
5056 opclass memory(direct, indirect, indOffset8, indOffset32, indOffset32X, indIndexOffset,
5057 indIndex, indIndexScale, indIndexScaleOffset);
5059 // Long memory operations are encoded in 2 instructions and a +4 offset.
5060 // This means some kind of offset is always required and you cannot use
5061 // an oop as the offset (done when working on static globals).
5062 opclass long_memory(direct, indirect, indOffset8, indOffset32, indIndexOffset,
5063 indIndex, indIndexScale, indIndexScaleOffset);
5066 //----------PIPELINE-----------------------------------------------------------
5067 // Rules which define the behavior of the target architectures pipeline.
5068 pipeline %{
5070 //----------ATTRIBUTES---------------------------------------------------------
5071 attributes %{
5072 variable_size_instructions; // Fixed size instructions
5073 max_instructions_per_bundle = 3; // Up to 3 instructions per bundle
5074 instruction_unit_size = 1; // An instruction is 1 bytes long
5075 instruction_fetch_unit_size = 16; // The processor fetches one line
5076 instruction_fetch_units = 1; // of 16 bytes
5078 // List of nop instructions
5079 nops( MachNop );
5080 %}
5082 //----------RESOURCES----------------------------------------------------------
5083 // Resources are the functional units available to the machine
5085 // Generic P2/P3 pipeline
5086 // 3 decoders, only D0 handles big operands; a "bundle" is the limit of
5087 // 3 instructions decoded per cycle.
5088 // 2 load/store ops per cycle, 1 branch, 1 FPU,
5089 // 2 ALU op, only ALU0 handles mul/div instructions.
5090 resources( D0, D1, D2, DECODE = D0 | D1 | D2,
5091 MS0, MS1, MEM = MS0 | MS1,
5092 BR, FPU,
5093 ALU0, ALU1, ALU = ALU0 | ALU1 );
5095 //----------PIPELINE DESCRIPTION-----------------------------------------------
5096 // Pipeline Description specifies the stages in the machine's pipeline
5098 // Generic P2/P3 pipeline
5099 pipe_desc(S0, S1, S2, S3, S4, S5);
5101 //----------PIPELINE CLASSES---------------------------------------------------
5102 // Pipeline Classes describe the stages in which input and output are
5103 // referenced by the hardware pipeline.
5105 // Naming convention: ialu or fpu
5106 // Then: _reg
5107 // Then: _reg if there is a 2nd register
5108 // Then: _long if it's a pair of instructions implementing a long
5109 // Then: _fat if it requires the big decoder
5110 // Or: _mem if it requires the big decoder and a memory unit.
5112 // Integer ALU reg operation
5113 pipe_class ialu_reg(rRegI dst) %{
5114 single_instruction;
5115 dst : S4(write);
5116 dst : S3(read);
5117 DECODE : S0; // any decoder
5118 ALU : S3; // any alu
5119 %}
5121 // Long ALU reg operation
5122 pipe_class ialu_reg_long(eRegL dst) %{
5123 instruction_count(2);
5124 dst : S4(write);
5125 dst : S3(read);
5126 DECODE : S0(2); // any 2 decoders
5127 ALU : S3(2); // both alus
5128 %}
5130 // Integer ALU reg operation using big decoder
5131 pipe_class ialu_reg_fat(rRegI dst) %{
5132 single_instruction;
5133 dst : S4(write);
5134 dst : S3(read);
5135 D0 : S0; // big decoder only
5136 ALU : S3; // any alu
5137 %}
5139 // Long ALU reg operation using big decoder
5140 pipe_class ialu_reg_long_fat(eRegL dst) %{
5141 instruction_count(2);
5142 dst : S4(write);
5143 dst : S3(read);
5144 D0 : S0(2); // big decoder only; twice
5145 ALU : S3(2); // any 2 alus
5146 %}
5148 // Integer ALU reg-reg operation
5149 pipe_class ialu_reg_reg(rRegI dst, rRegI src) %{
5150 single_instruction;
5151 dst : S4(write);
5152 src : S3(read);
5153 DECODE : S0; // any decoder
5154 ALU : S3; // any alu
5155 %}
5157 // Long ALU reg-reg operation
5158 pipe_class ialu_reg_reg_long(eRegL dst, eRegL src) %{
5159 instruction_count(2);
5160 dst : S4(write);
5161 src : S3(read);
5162 DECODE : S0(2); // any 2 decoders
5163 ALU : S3(2); // both alus
5164 %}
5166 // Integer ALU reg-reg operation
5167 pipe_class ialu_reg_reg_fat(rRegI dst, memory src) %{
5168 single_instruction;
5169 dst : S4(write);
5170 src : S3(read);
5171 D0 : S0; // big decoder only
5172 ALU : S3; // any alu
5173 %}
5175 // Long ALU reg-reg operation
5176 pipe_class ialu_reg_reg_long_fat(eRegL dst, eRegL src) %{
5177 instruction_count(2);
5178 dst : S4(write);
5179 src : S3(read);
5180 D0 : S0(2); // big decoder only; twice
5181 ALU : S3(2); // both alus
5182 %}
5184 // Integer ALU reg-mem operation
5185 pipe_class ialu_reg_mem(rRegI dst, memory mem) %{
5186 single_instruction;
5187 dst : S5(write);
5188 mem : S3(read);
5189 D0 : S0; // big decoder only
5190 ALU : S4; // any alu
5191 MEM : S3; // any mem
5192 %}
5194 // Long ALU reg-mem operation
5195 pipe_class ialu_reg_long_mem(eRegL dst, load_long_memory mem) %{
5196 instruction_count(2);
5197 dst : S5(write);
5198 mem : S3(read);
5199 D0 : S0(2); // big decoder only; twice
5200 ALU : S4(2); // any 2 alus
5201 MEM : S3(2); // both mems
5202 %}
5204 // Integer mem operation (prefetch)
5205 pipe_class ialu_mem(memory mem)
5206 %{
5207 single_instruction;
5208 mem : S3(read);
5209 D0 : S0; // big decoder only
5210 MEM : S3; // any mem
5211 %}
5213 // Integer Store to Memory
5214 pipe_class ialu_mem_reg(memory mem, rRegI src) %{
5215 single_instruction;
5216 mem : S3(read);
5217 src : S5(read);
5218 D0 : S0; // big decoder only
5219 ALU : S4; // any alu
5220 MEM : S3;
5221 %}
5223 // Long Store to Memory
5224 pipe_class ialu_mem_long_reg(memory mem, eRegL src) %{
5225 instruction_count(2);
5226 mem : S3(read);
5227 src : S5(read);
5228 D0 : S0(2); // big decoder only; twice
5229 ALU : S4(2); // any 2 alus
5230 MEM : S3(2); // Both mems
5231 %}
5233 // Integer Store to Memory
5234 pipe_class ialu_mem_imm(memory mem) %{
5235 single_instruction;
5236 mem : S3(read);
5237 D0 : S0; // big decoder only
5238 ALU : S4; // any alu
5239 MEM : S3;
5240 %}
5242 // Integer ALU0 reg-reg operation
5243 pipe_class ialu_reg_reg_alu0(rRegI dst, rRegI src) %{
5244 single_instruction;
5245 dst : S4(write);
5246 src : S3(read);
5247 D0 : S0; // Big decoder only
5248 ALU0 : S3; // only alu0
5249 %}
5251 // Integer ALU0 reg-mem operation
5252 pipe_class ialu_reg_mem_alu0(rRegI dst, memory mem) %{
5253 single_instruction;
5254 dst : S5(write);
5255 mem : S3(read);
5256 D0 : S0; // big decoder only
5257 ALU0 : S4; // ALU0 only
5258 MEM : S3; // any mem
5259 %}
5261 // Integer ALU reg-reg operation
5262 pipe_class ialu_cr_reg_reg(eFlagsReg cr, rRegI src1, rRegI src2) %{
5263 single_instruction;
5264 cr : S4(write);
5265 src1 : S3(read);
5266 src2 : S3(read);
5267 DECODE : S0; // any decoder
5268 ALU : S3; // any alu
5269 %}
5271 // Integer ALU reg-imm operation
5272 pipe_class ialu_cr_reg_imm(eFlagsReg cr, rRegI src1) %{
5273 single_instruction;
5274 cr : S4(write);
5275 src1 : S3(read);
5276 DECODE : S0; // any decoder
5277 ALU : S3; // any alu
5278 %}
5280 // Integer ALU reg-mem operation
5281 pipe_class ialu_cr_reg_mem(eFlagsReg cr, rRegI src1, memory src2) %{
5282 single_instruction;
5283 cr : S4(write);
5284 src1 : S3(read);
5285 src2 : S3(read);
5286 D0 : S0; // big decoder only
5287 ALU : S4; // any alu
5288 MEM : S3;
5289 %}
5291 // Conditional move reg-reg
5292 pipe_class pipe_cmplt( rRegI p, rRegI q, rRegI y ) %{
5293 instruction_count(4);
5294 y : S4(read);
5295 q : S3(read);
5296 p : S3(read);
5297 DECODE : S0(4); // any decoder
5298 %}
5300 // Conditional move reg-reg
5301 pipe_class pipe_cmov_reg( rRegI dst, rRegI src, eFlagsReg cr ) %{
5302 single_instruction;
5303 dst : S4(write);
5304 src : S3(read);
5305 cr : S3(read);
5306 DECODE : S0; // any decoder
5307 %}
5309 // Conditional move reg-mem
5310 pipe_class pipe_cmov_mem( eFlagsReg cr, rRegI dst, memory src) %{
5311 single_instruction;
5312 dst : S4(write);
5313 src : S3(read);
5314 cr : S3(read);
5315 DECODE : S0; // any decoder
5316 MEM : S3;
5317 %}
5319 // Conditional move reg-reg long
5320 pipe_class pipe_cmov_reg_long( eFlagsReg cr, eRegL dst, eRegL src) %{
5321 single_instruction;
5322 dst : S4(write);
5323 src : S3(read);
5324 cr : S3(read);
5325 DECODE : S0(2); // any 2 decoders
5326 %}
5328 // Conditional move double reg-reg
5329 pipe_class pipe_cmovDPR_reg( eFlagsReg cr, regDPR1 dst, regDPR src) %{
5330 single_instruction;
5331 dst : S4(write);
5332 src : S3(read);
5333 cr : S3(read);
5334 DECODE : S0; // any decoder
5335 %}
5337 // Float reg-reg operation
5338 pipe_class fpu_reg(regDPR dst) %{
5339 instruction_count(2);
5340 dst : S3(read);
5341 DECODE : S0(2); // any 2 decoders
5342 FPU : S3;
5343 %}
5345 // Float reg-reg operation
5346 pipe_class fpu_reg_reg(regDPR dst, regDPR src) %{
5347 instruction_count(2);
5348 dst : S4(write);
5349 src : S3(read);
5350 DECODE : S0(2); // any 2 decoders
5351 FPU : S3;
5352 %}
5354 // Float reg-reg operation
5355 pipe_class fpu_reg_reg_reg(regDPR dst, regDPR src1, regDPR src2) %{
5356 instruction_count(3);
5357 dst : S4(write);
5358 src1 : S3(read);
5359 src2 : S3(read);
5360 DECODE : S0(3); // any 3 decoders
5361 FPU : S3(2);
5362 %}
5364 // Float reg-reg operation
5365 pipe_class fpu_reg_reg_reg_reg(regDPR dst, regDPR src1, regDPR src2, regDPR src3) %{
5366 instruction_count(4);
5367 dst : S4(write);
5368 src1 : S3(read);
5369 src2 : S3(read);
5370 src3 : S3(read);
5371 DECODE : S0(4); // any 3 decoders
5372 FPU : S3(2);
5373 %}
5375 // Float reg-reg operation
5376 pipe_class fpu_reg_mem_reg_reg(regDPR dst, memory src1, regDPR src2, regDPR src3) %{
5377 instruction_count(4);
5378 dst : S4(write);
5379 src1 : S3(read);
5380 src2 : S3(read);
5381 src3 : S3(read);
5382 DECODE : S1(3); // any 3 decoders
5383 D0 : S0; // Big decoder only
5384 FPU : S3(2);
5385 MEM : S3;
5386 %}
5388 // Float reg-mem operation
5389 pipe_class fpu_reg_mem(regDPR dst, memory mem) %{
5390 instruction_count(2);
5391 dst : S5(write);
5392 mem : S3(read);
5393 D0 : S0; // big decoder only
5394 DECODE : S1; // any decoder for FPU POP
5395 FPU : S4;
5396 MEM : S3; // any mem
5397 %}
5399 // Float reg-mem operation
5400 pipe_class fpu_reg_reg_mem(regDPR dst, regDPR src1, memory mem) %{
5401 instruction_count(3);
5402 dst : S5(write);
5403 src1 : S3(read);
5404 mem : S3(read);
5405 D0 : S0; // big decoder only
5406 DECODE : S1(2); // any decoder for FPU POP
5407 FPU : S4;
5408 MEM : S3; // any mem
5409 %}
5411 // Float mem-reg operation
5412 pipe_class fpu_mem_reg(memory mem, regDPR src) %{
5413 instruction_count(2);
5414 src : S5(read);
5415 mem : S3(read);
5416 DECODE : S0; // any decoder for FPU PUSH
5417 D0 : S1; // big decoder only
5418 FPU : S4;
5419 MEM : S3; // any mem
5420 %}
5422 pipe_class fpu_mem_reg_reg(memory mem, regDPR src1, regDPR src2) %{
5423 instruction_count(3);
5424 src1 : S3(read);
5425 src2 : S3(read);
5426 mem : S3(read);
5427 DECODE : S0(2); // any decoder for FPU PUSH
5428 D0 : S1; // big decoder only
5429 FPU : S4;
5430 MEM : S3; // any mem
5431 %}
5433 pipe_class fpu_mem_reg_mem(memory mem, regDPR src1, memory src2) %{
5434 instruction_count(3);
5435 src1 : S3(read);
5436 src2 : S3(read);
5437 mem : S4(read);
5438 DECODE : S0; // any decoder for FPU PUSH
5439 D0 : S0(2); // big decoder only
5440 FPU : S4;
5441 MEM : S3(2); // any mem
5442 %}
5444 pipe_class fpu_mem_mem(memory dst, memory src1) %{
5445 instruction_count(2);
5446 src1 : S3(read);
5447 dst : S4(read);
5448 D0 : S0(2); // big decoder only
5449 MEM : S3(2); // any mem
5450 %}
5452 pipe_class fpu_mem_mem_mem(memory dst, memory src1, memory src2) %{
5453 instruction_count(3);
5454 src1 : S3(read);
5455 src2 : S3(read);
5456 dst : S4(read);
5457 D0 : S0(3); // big decoder only
5458 FPU : S4;
5459 MEM : S3(3); // any mem
5460 %}
5462 pipe_class fpu_mem_reg_con(memory mem, regDPR src1) %{
5463 instruction_count(3);
5464 src1 : S4(read);
5465 mem : S4(read);
5466 DECODE : S0; // any decoder for FPU PUSH
5467 D0 : S0(2); // big decoder only
5468 FPU : S4;
5469 MEM : S3(2); // any mem
5470 %}
5472 // Float load constant
5473 pipe_class fpu_reg_con(regDPR dst) %{
5474 instruction_count(2);
5475 dst : S5(write);
5476 D0 : S0; // big decoder only for the load
5477 DECODE : S1; // any decoder for FPU POP
5478 FPU : S4;
5479 MEM : S3; // any mem
5480 %}
5482 // Float load constant
5483 pipe_class fpu_reg_reg_con(regDPR dst, regDPR src) %{
5484 instruction_count(3);
5485 dst : S5(write);
5486 src : S3(read);
5487 D0 : S0; // big decoder only for the load
5488 DECODE : S1(2); // any decoder for FPU POP
5489 FPU : S4;
5490 MEM : S3; // any mem
5491 %}
5493 // UnConditional branch
5494 pipe_class pipe_jmp( label labl ) %{
5495 single_instruction;
5496 BR : S3;
5497 %}
5499 // Conditional branch
5500 pipe_class pipe_jcc( cmpOp cmp, eFlagsReg cr, label labl ) %{
5501 single_instruction;
5502 cr : S1(read);
5503 BR : S3;
5504 %}
5506 // Allocation idiom
5507 pipe_class pipe_cmpxchg( eRegP dst, eRegP heap_ptr ) %{
5508 instruction_count(1); force_serialization;
5509 fixed_latency(6);
5510 heap_ptr : S3(read);
5511 DECODE : S0(3);
5512 D0 : S2;
5513 MEM : S3;
5514 ALU : S3(2);
5515 dst : S5(write);
5516 BR : S5;
5517 %}
5519 // Generic big/slow expanded idiom
5520 pipe_class pipe_slow( ) %{
5521 instruction_count(10); multiple_bundles; force_serialization;
5522 fixed_latency(100);
5523 D0 : S0(2);
5524 MEM : S3(2);
5525 %}
5527 // The real do-nothing guy
5528 pipe_class empty( ) %{
5529 instruction_count(0);
5530 %}
5532 // Define the class for the Nop node
5533 define %{
5534 MachNop = empty;
5535 %}
5537 %}
5539 //----------INSTRUCTIONS-------------------------------------------------------
5540 //
5541 // match -- States which machine-independent subtree may be replaced
5542 // by this instruction.
5543 // ins_cost -- The estimated cost of this instruction is used by instruction
5544 // selection to identify a minimum cost tree of machine
5545 // instructions that matches a tree of machine-independent
5546 // instructions.
5547 // format -- A string providing the disassembly for this instruction.
5548 // The value of an instruction's operand may be inserted
5549 // by referring to it with a '$' prefix.
5550 // opcode -- Three instruction opcodes may be provided. These are referred
5551 // to within an encode class as $primary, $secondary, and $tertiary
5552 // respectively. The primary opcode is commonly used to
5553 // indicate the type of machine instruction, while secondary
5554 // and tertiary are often used for prefix options or addressing
5555 // modes.
5556 // ins_encode -- A list of encode classes with parameters. The encode class
5557 // name must have been defined in an 'enc_class' specification
5558 // in the encode section of the architecture description.
5560 //----------BSWAP-Instruction--------------------------------------------------
5561 instruct bytes_reverse_int(rRegI dst) %{
5562 match(Set dst (ReverseBytesI dst));
5564 format %{ "BSWAP $dst" %}
5565 opcode(0x0F, 0xC8);
5566 ins_encode( OpcP, OpcSReg(dst) );
5567 ins_pipe( ialu_reg );
5568 %}
5570 instruct bytes_reverse_long(eRegL dst) %{
5571 match(Set dst (ReverseBytesL dst));
5573 format %{ "BSWAP $dst.lo\n\t"
5574 "BSWAP $dst.hi\n\t"
5575 "XCHG $dst.lo $dst.hi" %}
5577 ins_cost(125);
5578 ins_encode( bswap_long_bytes(dst) );
5579 ins_pipe( ialu_reg_reg);
5580 %}
5582 instruct bytes_reverse_unsigned_short(rRegI dst, eFlagsReg cr) %{
5583 match(Set dst (ReverseBytesUS dst));
5584 effect(KILL cr);
5586 format %{ "BSWAP $dst\n\t"
5587 "SHR $dst,16\n\t" %}
5588 ins_encode %{
5589 __ bswapl($dst$$Register);
5590 __ shrl($dst$$Register, 16);
5591 %}
5592 ins_pipe( ialu_reg );
5593 %}
5595 instruct bytes_reverse_short(rRegI dst, eFlagsReg cr) %{
5596 match(Set dst (ReverseBytesS dst));
5597 effect(KILL cr);
5599 format %{ "BSWAP $dst\n\t"
5600 "SAR $dst,16\n\t" %}
5601 ins_encode %{
5602 __ bswapl($dst$$Register);
5603 __ sarl($dst$$Register, 16);
5604 %}
5605 ins_pipe( ialu_reg );
5606 %}
5609 //---------- Zeros Count Instructions ------------------------------------------
5611 instruct countLeadingZerosI(rRegI dst, rRegI src, eFlagsReg cr) %{
5612 predicate(UseCountLeadingZerosInstruction);
5613 match(Set dst (CountLeadingZerosI src));
5614 effect(KILL cr);
5616 format %{ "LZCNT $dst, $src\t# count leading zeros (int)" %}
5617 ins_encode %{
5618 __ lzcntl($dst$$Register, $src$$Register);
5619 %}
5620 ins_pipe(ialu_reg);
5621 %}
5623 instruct countLeadingZerosI_bsr(rRegI dst, rRegI src, eFlagsReg cr) %{
5624 predicate(!UseCountLeadingZerosInstruction);
5625 match(Set dst (CountLeadingZerosI src));
5626 effect(KILL cr);
5628 format %{ "BSR $dst, $src\t# count leading zeros (int)\n\t"
5629 "JNZ skip\n\t"
5630 "MOV $dst, -1\n"
5631 "skip:\n\t"
5632 "NEG $dst\n\t"
5633 "ADD $dst, 31" %}
5634 ins_encode %{
5635 Register Rdst = $dst$$Register;
5636 Register Rsrc = $src$$Register;
5637 Label skip;
5638 __ bsrl(Rdst, Rsrc);
5639 __ jccb(Assembler::notZero, skip);
5640 __ movl(Rdst, -1);
5641 __ bind(skip);
5642 __ negl(Rdst);
5643 __ addl(Rdst, BitsPerInt - 1);
5644 %}
5645 ins_pipe(ialu_reg);
5646 %}
5648 instruct countLeadingZerosL(rRegI dst, eRegL src, eFlagsReg cr) %{
5649 predicate(UseCountLeadingZerosInstruction);
5650 match(Set dst (CountLeadingZerosL src));
5651 effect(TEMP dst, KILL cr);
5653 format %{ "LZCNT $dst, $src.hi\t# count leading zeros (long)\n\t"
5654 "JNC done\n\t"
5655 "LZCNT $dst, $src.lo\n\t"
5656 "ADD $dst, 32\n"
5657 "done:" %}
5658 ins_encode %{
5659 Register Rdst = $dst$$Register;
5660 Register Rsrc = $src$$Register;
5661 Label done;
5662 __ lzcntl(Rdst, HIGH_FROM_LOW(Rsrc));
5663 __ jccb(Assembler::carryClear, done);
5664 __ lzcntl(Rdst, Rsrc);
5665 __ addl(Rdst, BitsPerInt);
5666 __ bind(done);
5667 %}
5668 ins_pipe(ialu_reg);
5669 %}
5671 instruct countLeadingZerosL_bsr(rRegI dst, eRegL src, eFlagsReg cr) %{
5672 predicate(!UseCountLeadingZerosInstruction);
5673 match(Set dst (CountLeadingZerosL src));
5674 effect(TEMP dst, KILL cr);
5676 format %{ "BSR $dst, $src.hi\t# count leading zeros (long)\n\t"
5677 "JZ msw_is_zero\n\t"
5678 "ADD $dst, 32\n\t"
5679 "JMP not_zero\n"
5680 "msw_is_zero:\n\t"
5681 "BSR $dst, $src.lo\n\t"
5682 "JNZ not_zero\n\t"
5683 "MOV $dst, -1\n"
5684 "not_zero:\n\t"
5685 "NEG $dst\n\t"
5686 "ADD $dst, 63\n" %}
5687 ins_encode %{
5688 Register Rdst = $dst$$Register;
5689 Register Rsrc = $src$$Register;
5690 Label msw_is_zero;
5691 Label not_zero;
5692 __ bsrl(Rdst, HIGH_FROM_LOW(Rsrc));
5693 __ jccb(Assembler::zero, msw_is_zero);
5694 __ addl(Rdst, BitsPerInt);
5695 __ jmpb(not_zero);
5696 __ bind(msw_is_zero);
5697 __ bsrl(Rdst, Rsrc);
5698 __ jccb(Assembler::notZero, not_zero);
5699 __ movl(Rdst, -1);
5700 __ bind(not_zero);
5701 __ negl(Rdst);
5702 __ addl(Rdst, BitsPerLong - 1);
5703 %}
5704 ins_pipe(ialu_reg);
5705 %}
5707 instruct countTrailingZerosI(rRegI dst, rRegI src, eFlagsReg cr) %{
5708 match(Set dst (CountTrailingZerosI src));
5709 effect(KILL cr);
5711 format %{ "BSF $dst, $src\t# count trailing zeros (int)\n\t"
5712 "JNZ done\n\t"
5713 "MOV $dst, 32\n"
5714 "done:" %}
5715 ins_encode %{
5716 Register Rdst = $dst$$Register;
5717 Label done;
5718 __ bsfl(Rdst, $src$$Register);
5719 __ jccb(Assembler::notZero, done);
5720 __ movl(Rdst, BitsPerInt);
5721 __ bind(done);
5722 %}
5723 ins_pipe(ialu_reg);
5724 %}
5726 instruct countTrailingZerosL(rRegI dst, eRegL src, eFlagsReg cr) %{
5727 match(Set dst (CountTrailingZerosL src));
5728 effect(TEMP dst, KILL cr);
5730 format %{ "BSF $dst, $src.lo\t# count trailing zeros (long)\n\t"
5731 "JNZ done\n\t"
5732 "BSF $dst, $src.hi\n\t"
5733 "JNZ msw_not_zero\n\t"
5734 "MOV $dst, 32\n"
5735 "msw_not_zero:\n\t"
5736 "ADD $dst, 32\n"
5737 "done:" %}
5738 ins_encode %{
5739 Register Rdst = $dst$$Register;
5740 Register Rsrc = $src$$Register;
5741 Label msw_not_zero;
5742 Label done;
5743 __ bsfl(Rdst, Rsrc);
5744 __ jccb(Assembler::notZero, done);
5745 __ bsfl(Rdst, HIGH_FROM_LOW(Rsrc));
5746 __ jccb(Assembler::notZero, msw_not_zero);
5747 __ movl(Rdst, BitsPerInt);
5748 __ bind(msw_not_zero);
5749 __ addl(Rdst, BitsPerInt);
5750 __ bind(done);
5751 %}
5752 ins_pipe(ialu_reg);
5753 %}
5756 //---------- Population Count Instructions -------------------------------------
5758 instruct popCountI(rRegI dst, rRegI src, eFlagsReg cr) %{
5759 predicate(UsePopCountInstruction);
5760 match(Set dst (PopCountI src));
5761 effect(KILL cr);
5763 format %{ "POPCNT $dst, $src" %}
5764 ins_encode %{
5765 __ popcntl($dst$$Register, $src$$Register);
5766 %}
5767 ins_pipe(ialu_reg);
5768 %}
5770 instruct popCountI_mem(rRegI dst, memory mem, eFlagsReg cr) %{
5771 predicate(UsePopCountInstruction);
5772 match(Set dst (PopCountI (LoadI mem)));
5773 effect(KILL cr);
5775 format %{ "POPCNT $dst, $mem" %}
5776 ins_encode %{
5777 __ popcntl($dst$$Register, $mem$$Address);
5778 %}
5779 ins_pipe(ialu_reg);
5780 %}
5782 // Note: Long.bitCount(long) returns an int.
5783 instruct popCountL(rRegI dst, eRegL src, rRegI tmp, eFlagsReg cr) %{
5784 predicate(UsePopCountInstruction);
5785 match(Set dst (PopCountL src));
5786 effect(KILL cr, TEMP tmp, TEMP dst);
5788 format %{ "POPCNT $dst, $src.lo\n\t"
5789 "POPCNT $tmp, $src.hi\n\t"
5790 "ADD $dst, $tmp" %}
5791 ins_encode %{
5792 __ popcntl($dst$$Register, $src$$Register);
5793 __ popcntl($tmp$$Register, HIGH_FROM_LOW($src$$Register));
5794 __ addl($dst$$Register, $tmp$$Register);
5795 %}
5796 ins_pipe(ialu_reg);
5797 %}
5799 // Note: Long.bitCount(long) returns an int.
5800 instruct popCountL_mem(rRegI dst, memory mem, rRegI tmp, eFlagsReg cr) %{
5801 predicate(UsePopCountInstruction);
5802 match(Set dst (PopCountL (LoadL mem)));
5803 effect(KILL cr, TEMP tmp, TEMP dst);
5805 format %{ "POPCNT $dst, $mem\n\t"
5806 "POPCNT $tmp, $mem+4\n\t"
5807 "ADD $dst, $tmp" %}
5808 ins_encode %{
5809 //__ popcntl($dst$$Register, $mem$$Address$$first);
5810 //__ popcntl($tmp$$Register, $mem$$Address$$second);
5811 __ popcntl($dst$$Register, Address::make_raw($mem$$base, $mem$$index, $mem$$scale, $mem$$disp, relocInfo::none));
5812 __ popcntl($tmp$$Register, Address::make_raw($mem$$base, $mem$$index, $mem$$scale, $mem$$disp + 4, relocInfo::none));
5813 __ addl($dst$$Register, $tmp$$Register);
5814 %}
5815 ins_pipe(ialu_reg);
5816 %}
5819 //----------Load/Store/Move Instructions---------------------------------------
5820 //----------Load Instructions--------------------------------------------------
5821 // Load Byte (8bit signed)
5822 instruct loadB(xRegI dst, memory mem) %{
5823 match(Set dst (LoadB mem));
5825 ins_cost(125);
5826 format %{ "MOVSX8 $dst,$mem\t# byte" %}
5828 ins_encode %{
5829 __ movsbl($dst$$Register, $mem$$Address);
5830 %}
5832 ins_pipe(ialu_reg_mem);
5833 %}
5835 // Load Byte (8bit signed) into Long Register
5836 instruct loadB2L(eRegL dst, memory mem, eFlagsReg cr) %{
5837 match(Set dst (ConvI2L (LoadB mem)));
5838 effect(KILL cr);
5840 ins_cost(375);
5841 format %{ "MOVSX8 $dst.lo,$mem\t# byte -> long\n\t"
5842 "MOV $dst.hi,$dst.lo\n\t"
5843 "SAR $dst.hi,7" %}
5845 ins_encode %{
5846 __ movsbl($dst$$Register, $mem$$Address);
5847 __ movl(HIGH_FROM_LOW($dst$$Register), $dst$$Register); // This is always a different register.
5848 __ sarl(HIGH_FROM_LOW($dst$$Register), 7); // 24+1 MSB are already signed extended.
5849 %}
5851 ins_pipe(ialu_reg_mem);
5852 %}
5854 // Load Unsigned Byte (8bit UNsigned)
5855 instruct loadUB(xRegI dst, memory mem) %{
5856 match(Set dst (LoadUB mem));
5858 ins_cost(125);
5859 format %{ "MOVZX8 $dst,$mem\t# ubyte -> int" %}
5861 ins_encode %{
5862 __ movzbl($dst$$Register, $mem$$Address);
5863 %}
5865 ins_pipe(ialu_reg_mem);
5866 %}
5868 // Load Unsigned Byte (8 bit UNsigned) into Long Register
5869 instruct loadUB2L(eRegL dst, memory mem, eFlagsReg cr) %{
5870 match(Set dst (ConvI2L (LoadUB mem)));
5871 effect(KILL cr);
5873 ins_cost(250);
5874 format %{ "MOVZX8 $dst.lo,$mem\t# ubyte -> long\n\t"
5875 "XOR $dst.hi,$dst.hi" %}
5877 ins_encode %{
5878 Register Rdst = $dst$$Register;
5879 __ movzbl(Rdst, $mem$$Address);
5880 __ xorl(HIGH_FROM_LOW(Rdst), HIGH_FROM_LOW(Rdst));
5881 %}
5883 ins_pipe(ialu_reg_mem);
5884 %}
5886 // Load Unsigned Byte (8 bit UNsigned) with mask into Long Register
5887 instruct loadUB2L_immI8(eRegL dst, memory mem, immI8 mask, eFlagsReg cr) %{
5888 match(Set dst (ConvI2L (AndI (LoadUB mem) mask)));
5889 effect(KILL cr);
5891 format %{ "MOVZX8 $dst.lo,$mem\t# ubyte & 8-bit mask -> long\n\t"
5892 "XOR $dst.hi,$dst.hi\n\t"
5893 "AND $dst.lo,$mask" %}
5894 ins_encode %{
5895 Register Rdst = $dst$$Register;
5896 __ movzbl(Rdst, $mem$$Address);
5897 __ xorl(HIGH_FROM_LOW(Rdst), HIGH_FROM_LOW(Rdst));
5898 __ andl(Rdst, $mask$$constant);
5899 %}
5900 ins_pipe(ialu_reg_mem);
5901 %}
5903 // Load Short (16bit signed)
5904 instruct loadS(rRegI dst, memory mem) %{
5905 match(Set dst (LoadS mem));
5907 ins_cost(125);
5908 format %{ "MOVSX $dst,$mem\t# short" %}
5910 ins_encode %{
5911 __ movswl($dst$$Register, $mem$$Address);
5912 %}
5914 ins_pipe(ialu_reg_mem);
5915 %}
5917 // Load Short (16 bit signed) to Byte (8 bit signed)
5918 instruct loadS2B(rRegI dst, memory mem, immI_24 twentyfour) %{
5919 match(Set dst (RShiftI (LShiftI (LoadS mem) twentyfour) twentyfour));
5921 ins_cost(125);
5922 format %{ "MOVSX $dst, $mem\t# short -> byte" %}
5923 ins_encode %{
5924 __ movsbl($dst$$Register, $mem$$Address);
5925 %}
5926 ins_pipe(ialu_reg_mem);
5927 %}
5929 // Load Short (16bit signed) into Long Register
5930 instruct loadS2L(eRegL dst, memory mem, eFlagsReg cr) %{
5931 match(Set dst (ConvI2L (LoadS mem)));
5932 effect(KILL cr);
5934 ins_cost(375);
5935 format %{ "MOVSX $dst.lo,$mem\t# short -> long\n\t"
5936 "MOV $dst.hi,$dst.lo\n\t"
5937 "SAR $dst.hi,15" %}
5939 ins_encode %{
5940 __ movswl($dst$$Register, $mem$$Address);
5941 __ movl(HIGH_FROM_LOW($dst$$Register), $dst$$Register); // This is always a different register.
5942 __ sarl(HIGH_FROM_LOW($dst$$Register), 15); // 16+1 MSB are already signed extended.
5943 %}
5945 ins_pipe(ialu_reg_mem);
5946 %}
5948 // Load Unsigned Short/Char (16bit unsigned)
5949 instruct loadUS(rRegI dst, memory mem) %{
5950 match(Set dst (LoadUS mem));
5952 ins_cost(125);
5953 format %{ "MOVZX $dst,$mem\t# ushort/char -> int" %}
5955 ins_encode %{
5956 __ movzwl($dst$$Register, $mem$$Address);
5957 %}
5959 ins_pipe(ialu_reg_mem);
5960 %}
5962 // Load Unsigned Short/Char (16 bit UNsigned) to Byte (8 bit signed)
5963 instruct loadUS2B(rRegI dst, memory mem, immI_24 twentyfour) %{
5964 match(Set dst (RShiftI (LShiftI (LoadUS mem) twentyfour) twentyfour));
5966 ins_cost(125);
5967 format %{ "MOVSX $dst, $mem\t# ushort -> byte" %}
5968 ins_encode %{
5969 __ movsbl($dst$$Register, $mem$$Address);
5970 %}
5971 ins_pipe(ialu_reg_mem);
5972 %}
5974 // Load Unsigned Short/Char (16 bit UNsigned) into Long Register
5975 instruct loadUS2L(eRegL dst, memory mem, eFlagsReg cr) %{
5976 match(Set dst (ConvI2L (LoadUS mem)));
5977 effect(KILL cr);
5979 ins_cost(250);
5980 format %{ "MOVZX $dst.lo,$mem\t# ushort/char -> long\n\t"
5981 "XOR $dst.hi,$dst.hi" %}
5983 ins_encode %{
5984 __ movzwl($dst$$Register, $mem$$Address);
5985 __ xorl(HIGH_FROM_LOW($dst$$Register), HIGH_FROM_LOW($dst$$Register));
5986 %}
5988 ins_pipe(ialu_reg_mem);
5989 %}
5991 // Load Unsigned Short/Char (16 bit UNsigned) with mask 0xFF into Long Register
5992 instruct loadUS2L_immI_255(eRegL dst, memory mem, immI_255 mask, eFlagsReg cr) %{
5993 match(Set dst (ConvI2L (AndI (LoadUS mem) mask)));
5994 effect(KILL cr);
5996 format %{ "MOVZX8 $dst.lo,$mem\t# ushort/char & 0xFF -> long\n\t"
5997 "XOR $dst.hi,$dst.hi" %}
5998 ins_encode %{
5999 Register Rdst = $dst$$Register;
6000 __ movzbl(Rdst, $mem$$Address);
6001 __ xorl(HIGH_FROM_LOW(Rdst), HIGH_FROM_LOW(Rdst));
6002 %}
6003 ins_pipe(ialu_reg_mem);
6004 %}
6006 // Load Unsigned Short/Char (16 bit UNsigned) with a 16-bit mask into Long Register
6007 instruct loadUS2L_immI16(eRegL dst, memory mem, immI16 mask, eFlagsReg cr) %{
6008 match(Set dst (ConvI2L (AndI (LoadUS mem) mask)));
6009 effect(KILL cr);
6011 format %{ "MOVZX $dst.lo, $mem\t# ushort/char & 16-bit mask -> long\n\t"
6012 "XOR $dst.hi,$dst.hi\n\t"
6013 "AND $dst.lo,$mask" %}
6014 ins_encode %{
6015 Register Rdst = $dst$$Register;
6016 __ movzwl(Rdst, $mem$$Address);
6017 __ xorl(HIGH_FROM_LOW(Rdst), HIGH_FROM_LOW(Rdst));
6018 __ andl(Rdst, $mask$$constant);
6019 %}
6020 ins_pipe(ialu_reg_mem);
6021 %}
6023 // Load Integer
6024 instruct loadI(rRegI dst, memory mem) %{
6025 match(Set dst (LoadI mem));
6027 ins_cost(125);
6028 format %{ "MOV $dst,$mem\t# int" %}
6030 ins_encode %{
6031 __ movl($dst$$Register, $mem$$Address);
6032 %}
6034 ins_pipe(ialu_reg_mem);
6035 %}
6037 // Load Integer (32 bit signed) to Byte (8 bit signed)
6038 instruct loadI2B(rRegI dst, memory mem, immI_24 twentyfour) %{
6039 match(Set dst (RShiftI (LShiftI (LoadI mem) twentyfour) twentyfour));
6041 ins_cost(125);
6042 format %{ "MOVSX $dst, $mem\t# int -> byte" %}
6043 ins_encode %{
6044 __ movsbl($dst$$Register, $mem$$Address);
6045 %}
6046 ins_pipe(ialu_reg_mem);
6047 %}
6049 // Load Integer (32 bit signed) to Unsigned Byte (8 bit UNsigned)
6050 instruct loadI2UB(rRegI dst, memory mem, immI_255 mask) %{
6051 match(Set dst (AndI (LoadI mem) mask));
6053 ins_cost(125);
6054 format %{ "MOVZX $dst, $mem\t# int -> ubyte" %}
6055 ins_encode %{
6056 __ movzbl($dst$$Register, $mem$$Address);
6057 %}
6058 ins_pipe(ialu_reg_mem);
6059 %}
6061 // Load Integer (32 bit signed) to Short (16 bit signed)
6062 instruct loadI2S(rRegI dst, memory mem, immI_16 sixteen) %{
6063 match(Set dst (RShiftI (LShiftI (LoadI mem) sixteen) sixteen));
6065 ins_cost(125);
6066 format %{ "MOVSX $dst, $mem\t# int -> short" %}
6067 ins_encode %{
6068 __ movswl($dst$$Register, $mem$$Address);
6069 %}
6070 ins_pipe(ialu_reg_mem);
6071 %}
6073 // Load Integer (32 bit signed) to Unsigned Short/Char (16 bit UNsigned)
6074 instruct loadI2US(rRegI dst, memory mem, immI_65535 mask) %{
6075 match(Set dst (AndI (LoadI mem) mask));
6077 ins_cost(125);
6078 format %{ "MOVZX $dst, $mem\t# int -> ushort/char" %}
6079 ins_encode %{
6080 __ movzwl($dst$$Register, $mem$$Address);
6081 %}
6082 ins_pipe(ialu_reg_mem);
6083 %}
6085 // Load Integer into Long Register
6086 instruct loadI2L(eRegL dst, memory mem, eFlagsReg cr) %{
6087 match(Set dst (ConvI2L (LoadI mem)));
6088 effect(KILL cr);
6090 ins_cost(375);
6091 format %{ "MOV $dst.lo,$mem\t# int -> long\n\t"
6092 "MOV $dst.hi,$dst.lo\n\t"
6093 "SAR $dst.hi,31" %}
6095 ins_encode %{
6096 __ movl($dst$$Register, $mem$$Address);
6097 __ movl(HIGH_FROM_LOW($dst$$Register), $dst$$Register); // This is always a different register.
6098 __ sarl(HIGH_FROM_LOW($dst$$Register), 31);
6099 %}
6101 ins_pipe(ialu_reg_mem);
6102 %}
6104 // Load Integer with mask 0xFF into Long Register
6105 instruct loadI2L_immI_255(eRegL dst, memory mem, immI_255 mask, eFlagsReg cr) %{
6106 match(Set dst (ConvI2L (AndI (LoadI mem) mask)));
6107 effect(KILL cr);
6109 format %{ "MOVZX8 $dst.lo,$mem\t# int & 0xFF -> long\n\t"
6110 "XOR $dst.hi,$dst.hi" %}
6111 ins_encode %{
6112 Register Rdst = $dst$$Register;
6113 __ movzbl(Rdst, $mem$$Address);
6114 __ xorl(HIGH_FROM_LOW(Rdst), HIGH_FROM_LOW(Rdst));
6115 %}
6116 ins_pipe(ialu_reg_mem);
6117 %}
6119 // Load Integer with mask 0xFFFF into Long Register
6120 instruct loadI2L_immI_65535(eRegL dst, memory mem, immI_65535 mask, eFlagsReg cr) %{
6121 match(Set dst (ConvI2L (AndI (LoadI mem) mask)));
6122 effect(KILL cr);
6124 format %{ "MOVZX $dst.lo,$mem\t# int & 0xFFFF -> long\n\t"
6125 "XOR $dst.hi,$dst.hi" %}
6126 ins_encode %{
6127 Register Rdst = $dst$$Register;
6128 __ movzwl(Rdst, $mem$$Address);
6129 __ xorl(HIGH_FROM_LOW(Rdst), HIGH_FROM_LOW(Rdst));
6130 %}
6131 ins_pipe(ialu_reg_mem);
6132 %}
6134 // Load Integer with 32-bit mask into Long Register
6135 instruct loadI2L_immI(eRegL dst, memory mem, immI mask, eFlagsReg cr) %{
6136 match(Set dst (ConvI2L (AndI (LoadI mem) mask)));
6137 effect(KILL cr);
6139 format %{ "MOV $dst.lo,$mem\t# int & 32-bit mask -> long\n\t"
6140 "XOR $dst.hi,$dst.hi\n\t"
6141 "AND $dst.lo,$mask" %}
6142 ins_encode %{
6143 Register Rdst = $dst$$Register;
6144 __ movl(Rdst, $mem$$Address);
6145 __ xorl(HIGH_FROM_LOW(Rdst), HIGH_FROM_LOW(Rdst));
6146 __ andl(Rdst, $mask$$constant);
6147 %}
6148 ins_pipe(ialu_reg_mem);
6149 %}
6151 // Load Unsigned Integer into Long Register
6152 instruct loadUI2L(eRegL dst, memory mem, immL_32bits mask, eFlagsReg cr) %{
6153 match(Set dst (AndL (ConvI2L (LoadI mem)) mask));
6154 effect(KILL cr);
6156 ins_cost(250);
6157 format %{ "MOV $dst.lo,$mem\t# uint -> long\n\t"
6158 "XOR $dst.hi,$dst.hi" %}
6160 ins_encode %{
6161 __ movl($dst$$Register, $mem$$Address);
6162 __ xorl(HIGH_FROM_LOW($dst$$Register), HIGH_FROM_LOW($dst$$Register));
6163 %}
6165 ins_pipe(ialu_reg_mem);
6166 %}
6168 // Load Long. Cannot clobber address while loading, so restrict address
6169 // register to ESI
6170 instruct loadL(eRegL dst, load_long_memory mem) %{
6171 predicate(!((LoadLNode*)n)->require_atomic_access());
6172 match(Set dst (LoadL mem));
6174 ins_cost(250);
6175 format %{ "MOV $dst.lo,$mem\t# long\n\t"
6176 "MOV $dst.hi,$mem+4" %}
6178 ins_encode %{
6179 Address Amemlo = Address::make_raw($mem$$base, $mem$$index, $mem$$scale, $mem$$disp, relocInfo::none);
6180 Address Amemhi = Address::make_raw($mem$$base, $mem$$index, $mem$$scale, $mem$$disp + 4, relocInfo::none);
6181 __ movl($dst$$Register, Amemlo);
6182 __ movl(HIGH_FROM_LOW($dst$$Register), Amemhi);
6183 %}
6185 ins_pipe(ialu_reg_long_mem);
6186 %}
6188 // Volatile Load Long. Must be atomic, so do 64-bit FILD
6189 // then store it down to the stack and reload on the int
6190 // side.
6191 instruct loadL_volatile(stackSlotL dst, memory mem) %{
6192 predicate(UseSSE<=1 && ((LoadLNode*)n)->require_atomic_access());
6193 match(Set dst (LoadL mem));
6195 ins_cost(200);
6196 format %{ "FILD $mem\t# Atomic volatile long load\n\t"
6197 "FISTp $dst" %}
6198 ins_encode(enc_loadL_volatile(mem,dst));
6199 ins_pipe( fpu_reg_mem );
6200 %}
6202 instruct loadLX_volatile(stackSlotL dst, memory mem, regD tmp) %{
6203 predicate(UseSSE>=2 && ((LoadLNode*)n)->require_atomic_access());
6204 match(Set dst (LoadL mem));
6205 effect(TEMP tmp);
6206 ins_cost(180);
6207 format %{ "MOVSD $tmp,$mem\t# Atomic volatile long load\n\t"
6208 "MOVSD $dst,$tmp" %}
6209 ins_encode %{
6210 __ movdbl($tmp$$XMMRegister, $mem$$Address);
6211 __ movdbl(Address(rsp, $dst$$disp), $tmp$$XMMRegister);
6212 %}
6213 ins_pipe( pipe_slow );
6214 %}
6216 instruct loadLX_reg_volatile(eRegL dst, memory mem, regD tmp) %{
6217 predicate(UseSSE>=2 && ((LoadLNode*)n)->require_atomic_access());
6218 match(Set dst (LoadL mem));
6219 effect(TEMP tmp);
6220 ins_cost(160);
6221 format %{ "MOVSD $tmp,$mem\t# Atomic volatile long load\n\t"
6222 "MOVD $dst.lo,$tmp\n\t"
6223 "PSRLQ $tmp,32\n\t"
6224 "MOVD $dst.hi,$tmp" %}
6225 ins_encode %{
6226 __ movdbl($tmp$$XMMRegister, $mem$$Address);
6227 __ movdl($dst$$Register, $tmp$$XMMRegister);
6228 __ psrlq($tmp$$XMMRegister, 32);
6229 __ movdl(HIGH_FROM_LOW($dst$$Register), $tmp$$XMMRegister);
6230 %}
6231 ins_pipe( pipe_slow );
6232 %}
6234 // Load Range
6235 instruct loadRange(rRegI dst, memory mem) %{
6236 match(Set dst (LoadRange mem));
6238 ins_cost(125);
6239 format %{ "MOV $dst,$mem" %}
6240 opcode(0x8B);
6241 ins_encode( OpcP, RegMem(dst,mem));
6242 ins_pipe( ialu_reg_mem );
6243 %}
6246 // Load Pointer
6247 instruct loadP(eRegP dst, memory mem) %{
6248 match(Set dst (LoadP mem));
6250 ins_cost(125);
6251 format %{ "MOV $dst,$mem" %}
6252 opcode(0x8B);
6253 ins_encode( OpcP, RegMem(dst,mem));
6254 ins_pipe( ialu_reg_mem );
6255 %}
6257 // Load Klass Pointer
6258 instruct loadKlass(eRegP dst, memory mem) %{
6259 match(Set dst (LoadKlass mem));
6261 ins_cost(125);
6262 format %{ "MOV $dst,$mem" %}
6263 opcode(0x8B);
6264 ins_encode( OpcP, RegMem(dst,mem));
6265 ins_pipe( ialu_reg_mem );
6266 %}
6268 // Load Double
6269 instruct loadDPR(regDPR dst, memory mem) %{
6270 predicate(UseSSE<=1);
6271 match(Set dst (LoadD mem));
6273 ins_cost(150);
6274 format %{ "FLD_D ST,$mem\n\t"
6275 "FSTP $dst" %}
6276 opcode(0xDD); /* DD /0 */
6277 ins_encode( OpcP, RMopc_Mem(0x00,mem),
6278 Pop_Reg_DPR(dst) );
6279 ins_pipe( fpu_reg_mem );
6280 %}
6282 // Load Double to XMM
6283 instruct loadD(regD dst, memory mem) %{
6284 predicate(UseSSE>=2 && UseXmmLoadAndClearUpper);
6285 match(Set dst (LoadD mem));
6286 ins_cost(145);
6287 format %{ "MOVSD $dst,$mem" %}
6288 ins_encode %{
6289 __ movdbl ($dst$$XMMRegister, $mem$$Address);
6290 %}
6291 ins_pipe( pipe_slow );
6292 %}
6294 instruct loadD_partial(regD dst, memory mem) %{
6295 predicate(UseSSE>=2 && !UseXmmLoadAndClearUpper);
6296 match(Set dst (LoadD mem));
6297 ins_cost(145);
6298 format %{ "MOVLPD $dst,$mem" %}
6299 ins_encode %{
6300 __ movdbl ($dst$$XMMRegister, $mem$$Address);
6301 %}
6302 ins_pipe( pipe_slow );
6303 %}
6305 // Load to XMM register (single-precision floating point)
6306 // MOVSS instruction
6307 instruct loadF(regF dst, memory mem) %{
6308 predicate(UseSSE>=1);
6309 match(Set dst (LoadF mem));
6310 ins_cost(145);
6311 format %{ "MOVSS $dst,$mem" %}
6312 ins_encode %{
6313 __ movflt ($dst$$XMMRegister, $mem$$Address);
6314 %}
6315 ins_pipe( pipe_slow );
6316 %}
6318 // Load Float
6319 instruct loadFPR(regFPR dst, memory mem) %{
6320 predicate(UseSSE==0);
6321 match(Set dst (LoadF mem));
6323 ins_cost(150);
6324 format %{ "FLD_S ST,$mem\n\t"
6325 "FSTP $dst" %}
6326 opcode(0xD9); /* D9 /0 */
6327 ins_encode( OpcP, RMopc_Mem(0x00,mem),
6328 Pop_Reg_FPR(dst) );
6329 ins_pipe( fpu_reg_mem );
6330 %}
6332 // Load Effective Address
6333 instruct leaP8(eRegP dst, indOffset8 mem) %{
6334 match(Set dst mem);
6336 ins_cost(110);
6337 format %{ "LEA $dst,$mem" %}
6338 opcode(0x8D);
6339 ins_encode( OpcP, RegMem(dst,mem));
6340 ins_pipe( ialu_reg_reg_fat );
6341 %}
6343 instruct leaP32(eRegP dst, indOffset32 mem) %{
6344 match(Set dst mem);
6346 ins_cost(110);
6347 format %{ "LEA $dst,$mem" %}
6348 opcode(0x8D);
6349 ins_encode( OpcP, RegMem(dst,mem));
6350 ins_pipe( ialu_reg_reg_fat );
6351 %}
6353 instruct leaPIdxOff(eRegP dst, indIndexOffset mem) %{
6354 match(Set dst mem);
6356 ins_cost(110);
6357 format %{ "LEA $dst,$mem" %}
6358 opcode(0x8D);
6359 ins_encode( OpcP, RegMem(dst,mem));
6360 ins_pipe( ialu_reg_reg_fat );
6361 %}
6363 instruct leaPIdxScale(eRegP dst, indIndexScale mem) %{
6364 match(Set dst mem);
6366 ins_cost(110);
6367 format %{ "LEA $dst,$mem" %}
6368 opcode(0x8D);
6369 ins_encode( OpcP, RegMem(dst,mem));
6370 ins_pipe( ialu_reg_reg_fat );
6371 %}
6373 instruct leaPIdxScaleOff(eRegP dst, indIndexScaleOffset mem) %{
6374 match(Set dst mem);
6376 ins_cost(110);
6377 format %{ "LEA $dst,$mem" %}
6378 opcode(0x8D);
6379 ins_encode( OpcP, RegMem(dst,mem));
6380 ins_pipe( ialu_reg_reg_fat );
6381 %}
6383 // Load Constant
6384 instruct loadConI(rRegI dst, immI src) %{
6385 match(Set dst src);
6387 format %{ "MOV $dst,$src" %}
6388 ins_encode( LdImmI(dst, src) );
6389 ins_pipe( ialu_reg_fat );
6390 %}
6392 // Load Constant zero
6393 instruct loadConI0(rRegI dst, immI0 src, eFlagsReg cr) %{
6394 match(Set dst src);
6395 effect(KILL cr);
6397 ins_cost(50);
6398 format %{ "XOR $dst,$dst" %}
6399 opcode(0x33); /* + rd */
6400 ins_encode( OpcP, RegReg( dst, dst ) );
6401 ins_pipe( ialu_reg );
6402 %}
6404 instruct loadConP(eRegP dst, immP src) %{
6405 match(Set dst src);
6407 format %{ "MOV $dst,$src" %}
6408 opcode(0xB8); /* + rd */
6409 ins_encode( LdImmP(dst, src) );
6410 ins_pipe( ialu_reg_fat );
6411 %}
6413 instruct loadConL(eRegL dst, immL src, eFlagsReg cr) %{
6414 match(Set dst src);
6415 effect(KILL cr);
6416 ins_cost(200);
6417 format %{ "MOV $dst.lo,$src.lo\n\t"
6418 "MOV $dst.hi,$src.hi" %}
6419 opcode(0xB8);
6420 ins_encode( LdImmL_Lo(dst, src), LdImmL_Hi(dst, src) );
6421 ins_pipe( ialu_reg_long_fat );
6422 %}
6424 instruct loadConL0(eRegL dst, immL0 src, eFlagsReg cr) %{
6425 match(Set dst src);
6426 effect(KILL cr);
6427 ins_cost(150);
6428 format %{ "XOR $dst.lo,$dst.lo\n\t"
6429 "XOR $dst.hi,$dst.hi" %}
6430 opcode(0x33,0x33);
6431 ins_encode( RegReg_Lo(dst,dst), RegReg_Hi(dst, dst) );
6432 ins_pipe( ialu_reg_long );
6433 %}
6435 // The instruction usage is guarded by predicate in operand immFPR().
6436 instruct loadConFPR(regFPR dst, immFPR con) %{
6437 match(Set dst con);
6438 ins_cost(125);
6439 format %{ "FLD_S ST,[$constantaddress]\t# load from constant table: float=$con\n\t"
6440 "FSTP $dst" %}
6441 ins_encode %{
6442 __ fld_s($constantaddress($con));
6443 __ fstp_d($dst$$reg);
6444 %}
6445 ins_pipe(fpu_reg_con);
6446 %}
6448 // The instruction usage is guarded by predicate in operand immFPR0().
6449 instruct loadConFPR0(regFPR dst, immFPR0 con) %{
6450 match(Set dst con);
6451 ins_cost(125);
6452 format %{ "FLDZ ST\n\t"
6453 "FSTP $dst" %}
6454 ins_encode %{
6455 __ fldz();
6456 __ fstp_d($dst$$reg);
6457 %}
6458 ins_pipe(fpu_reg_con);
6459 %}
6461 // The instruction usage is guarded by predicate in operand immFPR1().
6462 instruct loadConFPR1(regFPR dst, immFPR1 con) %{
6463 match(Set dst con);
6464 ins_cost(125);
6465 format %{ "FLD1 ST\n\t"
6466 "FSTP $dst" %}
6467 ins_encode %{
6468 __ fld1();
6469 __ fstp_d($dst$$reg);
6470 %}
6471 ins_pipe(fpu_reg_con);
6472 %}
6474 // The instruction usage is guarded by predicate in operand immF().
6475 instruct loadConF(regF dst, immF con) %{
6476 match(Set dst con);
6477 ins_cost(125);
6478 format %{ "MOVSS $dst,[$constantaddress]\t# load from constant table: float=$con" %}
6479 ins_encode %{
6480 __ movflt($dst$$XMMRegister, $constantaddress($con));
6481 %}
6482 ins_pipe(pipe_slow);
6483 %}
6485 // The instruction usage is guarded by predicate in operand immF0().
6486 instruct loadConF0(regF dst, immF0 src) %{
6487 match(Set dst src);
6488 ins_cost(100);
6489 format %{ "XORPS $dst,$dst\t# float 0.0" %}
6490 ins_encode %{
6491 __ xorps($dst$$XMMRegister, $dst$$XMMRegister);
6492 %}
6493 ins_pipe(pipe_slow);
6494 %}
6496 // The instruction usage is guarded by predicate in operand immDPR().
6497 instruct loadConDPR(regDPR dst, immDPR con) %{
6498 match(Set dst con);
6499 ins_cost(125);
6501 format %{ "FLD_D ST,[$constantaddress]\t# load from constant table: double=$con\n\t"
6502 "FSTP $dst" %}
6503 ins_encode %{
6504 __ fld_d($constantaddress($con));
6505 __ fstp_d($dst$$reg);
6506 %}
6507 ins_pipe(fpu_reg_con);
6508 %}
6510 // The instruction usage is guarded by predicate in operand immDPR0().
6511 instruct loadConDPR0(regDPR dst, immDPR0 con) %{
6512 match(Set dst con);
6513 ins_cost(125);
6515 format %{ "FLDZ ST\n\t"
6516 "FSTP $dst" %}
6517 ins_encode %{
6518 __ fldz();
6519 __ fstp_d($dst$$reg);
6520 %}
6521 ins_pipe(fpu_reg_con);
6522 %}
6524 // The instruction usage is guarded by predicate in operand immDPR1().
6525 instruct loadConDPR1(regDPR dst, immDPR1 con) %{
6526 match(Set dst con);
6527 ins_cost(125);
6529 format %{ "FLD1 ST\n\t"
6530 "FSTP $dst" %}
6531 ins_encode %{
6532 __ fld1();
6533 __ fstp_d($dst$$reg);
6534 %}
6535 ins_pipe(fpu_reg_con);
6536 %}
6538 // The instruction usage is guarded by predicate in operand immD().
6539 instruct loadConD(regD dst, immD con) %{
6540 match(Set dst con);
6541 ins_cost(125);
6542 format %{ "MOVSD $dst,[$constantaddress]\t# load from constant table: double=$con" %}
6543 ins_encode %{
6544 __ movdbl($dst$$XMMRegister, $constantaddress($con));
6545 %}
6546 ins_pipe(pipe_slow);
6547 %}
6549 // The instruction usage is guarded by predicate in operand immD0().
6550 instruct loadConD0(regD dst, immD0 src) %{
6551 match(Set dst src);
6552 ins_cost(100);
6553 format %{ "XORPD $dst,$dst\t# double 0.0" %}
6554 ins_encode %{
6555 __ xorpd ($dst$$XMMRegister, $dst$$XMMRegister);
6556 %}
6557 ins_pipe( pipe_slow );
6558 %}
6560 // Load Stack Slot
6561 instruct loadSSI(rRegI dst, stackSlotI src) %{
6562 match(Set dst src);
6563 ins_cost(125);
6565 format %{ "MOV $dst,$src" %}
6566 opcode(0x8B);
6567 ins_encode( OpcP, RegMem(dst,src));
6568 ins_pipe( ialu_reg_mem );
6569 %}
6571 instruct loadSSL(eRegL dst, stackSlotL src) %{
6572 match(Set dst src);
6574 ins_cost(200);
6575 format %{ "MOV $dst,$src.lo\n\t"
6576 "MOV $dst+4,$src.hi" %}
6577 opcode(0x8B, 0x8B);
6578 ins_encode( OpcP, RegMem( dst, src ), OpcS, RegMem_Hi( dst, src ) );
6579 ins_pipe( ialu_mem_long_reg );
6580 %}
6582 // Load Stack Slot
6583 instruct loadSSP(eRegP dst, stackSlotP src) %{
6584 match(Set dst src);
6585 ins_cost(125);
6587 format %{ "MOV $dst,$src" %}
6588 opcode(0x8B);
6589 ins_encode( OpcP, RegMem(dst,src));
6590 ins_pipe( ialu_reg_mem );
6591 %}
6593 // Load Stack Slot
6594 instruct loadSSF(regFPR dst, stackSlotF src) %{
6595 match(Set dst src);
6596 ins_cost(125);
6598 format %{ "FLD_S $src\n\t"
6599 "FSTP $dst" %}
6600 opcode(0xD9); /* D9 /0, FLD m32real */
6601 ins_encode( OpcP, RMopc_Mem_no_oop(0x00,src),
6602 Pop_Reg_FPR(dst) );
6603 ins_pipe( fpu_reg_mem );
6604 %}
6606 // Load Stack Slot
6607 instruct loadSSD(regDPR dst, stackSlotD src) %{
6608 match(Set dst src);
6609 ins_cost(125);
6611 format %{ "FLD_D $src\n\t"
6612 "FSTP $dst" %}
6613 opcode(0xDD); /* DD /0, FLD m64real */
6614 ins_encode( OpcP, RMopc_Mem_no_oop(0x00,src),
6615 Pop_Reg_DPR(dst) );
6616 ins_pipe( fpu_reg_mem );
6617 %}
6619 // Prefetch instructions.
6620 // Must be safe to execute with invalid address (cannot fault).
6622 instruct prefetchr0( memory mem ) %{
6623 predicate(UseSSE==0 && !VM_Version::supports_3dnow_prefetch());
6624 match(PrefetchRead mem);
6625 ins_cost(0);
6626 size(0);
6627 format %{ "PREFETCHR (non-SSE is empty encoding)" %}
6628 ins_encode();
6629 ins_pipe(empty);
6630 %}
6632 instruct prefetchr( memory mem ) %{
6633 predicate(UseSSE==0 && VM_Version::supports_3dnow_prefetch() || ReadPrefetchInstr==3);
6634 match(PrefetchRead mem);
6635 ins_cost(100);
6637 format %{ "PREFETCHR $mem\t! Prefetch into level 1 cache for read" %}
6638 ins_encode %{
6639 __ prefetchr($mem$$Address);
6640 %}
6641 ins_pipe(ialu_mem);
6642 %}
6644 instruct prefetchrNTA( memory mem ) %{
6645 predicate(UseSSE>=1 && ReadPrefetchInstr==0);
6646 match(PrefetchRead mem);
6647 ins_cost(100);
6649 format %{ "PREFETCHNTA $mem\t! Prefetch into non-temporal cache for read" %}
6650 ins_encode %{
6651 __ prefetchnta($mem$$Address);
6652 %}
6653 ins_pipe(ialu_mem);
6654 %}
6656 instruct prefetchrT0( memory mem ) %{
6657 predicate(UseSSE>=1 && ReadPrefetchInstr==1);
6658 match(PrefetchRead mem);
6659 ins_cost(100);
6661 format %{ "PREFETCHT0 $mem\t! Prefetch into L1 and L2 caches for read" %}
6662 ins_encode %{
6663 __ prefetcht0($mem$$Address);
6664 %}
6665 ins_pipe(ialu_mem);
6666 %}
6668 instruct prefetchrT2( memory mem ) %{
6669 predicate(UseSSE>=1 && ReadPrefetchInstr==2);
6670 match(PrefetchRead mem);
6671 ins_cost(100);
6673 format %{ "PREFETCHT2 $mem\t! Prefetch into L2 cache for read" %}
6674 ins_encode %{
6675 __ prefetcht2($mem$$Address);
6676 %}
6677 ins_pipe(ialu_mem);
6678 %}
6680 instruct prefetchw0( memory mem ) %{
6681 predicate(UseSSE==0 && !VM_Version::supports_3dnow_prefetch());
6682 match(PrefetchWrite mem);
6683 ins_cost(0);
6684 size(0);
6685 format %{ "Prefetch (non-SSE is empty encoding)" %}
6686 ins_encode();
6687 ins_pipe(empty);
6688 %}
6690 instruct prefetchw( memory mem ) %{
6691 predicate(UseSSE==0 && VM_Version::supports_3dnow_prefetch());
6692 match( PrefetchWrite mem );
6693 ins_cost(100);
6695 format %{ "PREFETCHW $mem\t! Prefetch into L1 cache and mark modified" %}
6696 ins_encode %{
6697 __ prefetchw($mem$$Address);
6698 %}
6699 ins_pipe(ialu_mem);
6700 %}
6702 instruct prefetchwNTA( memory mem ) %{
6703 predicate(UseSSE>=1);
6704 match(PrefetchWrite mem);
6705 ins_cost(100);
6707 format %{ "PREFETCHNTA $mem\t! Prefetch into non-temporal cache for write" %}
6708 ins_encode %{
6709 __ prefetchnta($mem$$Address);
6710 %}
6711 ins_pipe(ialu_mem);
6712 %}
6714 // Prefetch instructions for allocation.
6716 instruct prefetchAlloc0( memory mem ) %{
6717 predicate(UseSSE==0 && AllocatePrefetchInstr!=3);
6718 match(PrefetchAllocation mem);
6719 ins_cost(0);
6720 size(0);
6721 format %{ "Prefetch allocation (non-SSE is empty encoding)" %}
6722 ins_encode();
6723 ins_pipe(empty);
6724 %}
6726 instruct prefetchAlloc( memory mem ) %{
6727 predicate(AllocatePrefetchInstr==3);
6728 match( PrefetchAllocation mem );
6729 ins_cost(100);
6731 format %{ "PREFETCHW $mem\t! Prefetch allocation into L1 cache and mark modified" %}
6732 ins_encode %{
6733 __ prefetchw($mem$$Address);
6734 %}
6735 ins_pipe(ialu_mem);
6736 %}
6738 instruct prefetchAllocNTA( memory mem ) %{
6739 predicate(UseSSE>=1 && AllocatePrefetchInstr==0);
6740 match(PrefetchAllocation mem);
6741 ins_cost(100);
6743 format %{ "PREFETCHNTA $mem\t! Prefetch allocation into non-temporal cache for write" %}
6744 ins_encode %{
6745 __ prefetchnta($mem$$Address);
6746 %}
6747 ins_pipe(ialu_mem);
6748 %}
6750 instruct prefetchAllocT0( memory mem ) %{
6751 predicate(UseSSE>=1 && AllocatePrefetchInstr==1);
6752 match(PrefetchAllocation mem);
6753 ins_cost(100);
6755 format %{ "PREFETCHT0 $mem\t! Prefetch allocation into L1 and L2 caches for write" %}
6756 ins_encode %{
6757 __ prefetcht0($mem$$Address);
6758 %}
6759 ins_pipe(ialu_mem);
6760 %}
6762 instruct prefetchAllocT2( memory mem ) %{
6763 predicate(UseSSE>=1 && AllocatePrefetchInstr==2);
6764 match(PrefetchAllocation mem);
6765 ins_cost(100);
6767 format %{ "PREFETCHT2 $mem\t! Prefetch allocation into L2 cache for write" %}
6768 ins_encode %{
6769 __ prefetcht2($mem$$Address);
6770 %}
6771 ins_pipe(ialu_mem);
6772 %}
6774 //----------Store Instructions-------------------------------------------------
6776 // Store Byte
6777 instruct storeB(memory mem, xRegI src) %{
6778 match(Set mem (StoreB mem src));
6780 ins_cost(125);
6781 format %{ "MOV8 $mem,$src" %}
6782 opcode(0x88);
6783 ins_encode( OpcP, RegMem( src, mem ) );
6784 ins_pipe( ialu_mem_reg );
6785 %}
6787 // Store Char/Short
6788 instruct storeC(memory mem, rRegI src) %{
6789 match(Set mem (StoreC mem src));
6791 ins_cost(125);
6792 format %{ "MOV16 $mem,$src" %}
6793 opcode(0x89, 0x66);
6794 ins_encode( OpcS, OpcP, RegMem( src, mem ) );
6795 ins_pipe( ialu_mem_reg );
6796 %}
6798 // Store Integer
6799 instruct storeI(memory mem, rRegI src) %{
6800 match(Set mem (StoreI mem src));
6802 ins_cost(125);
6803 format %{ "MOV $mem,$src" %}
6804 opcode(0x89);
6805 ins_encode( OpcP, RegMem( src, mem ) );
6806 ins_pipe( ialu_mem_reg );
6807 %}
6809 // Store Long
6810 instruct storeL(long_memory mem, eRegL src) %{
6811 predicate(!((StoreLNode*)n)->require_atomic_access());
6812 match(Set mem (StoreL mem src));
6814 ins_cost(200);
6815 format %{ "MOV $mem,$src.lo\n\t"
6816 "MOV $mem+4,$src.hi" %}
6817 opcode(0x89, 0x89);
6818 ins_encode( OpcP, RegMem( src, mem ), OpcS, RegMem_Hi( src, mem ) );
6819 ins_pipe( ialu_mem_long_reg );
6820 %}
6822 // Store Long to Integer
6823 instruct storeL2I(memory mem, eRegL src) %{
6824 match(Set mem (StoreI mem (ConvL2I src)));
6826 format %{ "MOV $mem,$src.lo\t# long -> int" %}
6827 ins_encode %{
6828 __ movl($mem$$Address, $src$$Register);
6829 %}
6830 ins_pipe(ialu_mem_reg);
6831 %}
6833 // Volatile Store Long. Must be atomic, so move it into
6834 // the FP TOS and then do a 64-bit FIST. Has to probe the
6835 // target address before the store (for null-ptr checks)
6836 // so the memory operand is used twice in the encoding.
6837 instruct storeL_volatile(memory mem, stackSlotL src, eFlagsReg cr ) %{
6838 predicate(UseSSE<=1 && ((StoreLNode*)n)->require_atomic_access());
6839 match(Set mem (StoreL mem src));
6840 effect( KILL cr );
6841 ins_cost(400);
6842 format %{ "CMP $mem,EAX\t# Probe address for implicit null check\n\t"
6843 "FILD $src\n\t"
6844 "FISTp $mem\t # 64-bit atomic volatile long store" %}
6845 opcode(0x3B);
6846 ins_encode( OpcP, RegMem( EAX, mem ), enc_storeL_volatile(mem,src));
6847 ins_pipe( fpu_reg_mem );
6848 %}
6850 instruct storeLX_volatile(memory mem, stackSlotL src, regD tmp, eFlagsReg cr) %{
6851 predicate(UseSSE>=2 && ((StoreLNode*)n)->require_atomic_access());
6852 match(Set mem (StoreL mem src));
6853 effect( TEMP tmp, KILL cr );
6854 ins_cost(380);
6855 format %{ "CMP $mem,EAX\t# Probe address for implicit null check\n\t"
6856 "MOVSD $tmp,$src\n\t"
6857 "MOVSD $mem,$tmp\t # 64-bit atomic volatile long store" %}
6858 ins_encode %{
6859 __ cmpl(rax, $mem$$Address);
6860 __ movdbl($tmp$$XMMRegister, Address(rsp, $src$$disp));
6861 __ movdbl($mem$$Address, $tmp$$XMMRegister);
6862 %}
6863 ins_pipe( pipe_slow );
6864 %}
6866 instruct storeLX_reg_volatile(memory mem, eRegL src, regD tmp2, regD tmp, eFlagsReg cr) %{
6867 predicate(UseSSE>=2 && ((StoreLNode*)n)->require_atomic_access());
6868 match(Set mem (StoreL mem src));
6869 effect( TEMP tmp2 , TEMP tmp, KILL cr );
6870 ins_cost(360);
6871 format %{ "CMP $mem,EAX\t# Probe address for implicit null check\n\t"
6872 "MOVD $tmp,$src.lo\n\t"
6873 "MOVD $tmp2,$src.hi\n\t"
6874 "PUNPCKLDQ $tmp,$tmp2\n\t"
6875 "MOVSD $mem,$tmp\t # 64-bit atomic volatile long store" %}
6876 ins_encode %{
6877 __ cmpl(rax, $mem$$Address);
6878 __ movdl($tmp$$XMMRegister, $src$$Register);
6879 __ movdl($tmp2$$XMMRegister, HIGH_FROM_LOW($src$$Register));
6880 __ punpckldq($tmp$$XMMRegister, $tmp2$$XMMRegister);
6881 __ movdbl($mem$$Address, $tmp$$XMMRegister);
6882 %}
6883 ins_pipe( pipe_slow );
6884 %}
6886 // Store Pointer; for storing unknown oops and raw pointers
6887 instruct storeP(memory mem, anyRegP src) %{
6888 match(Set mem (StoreP mem src));
6890 ins_cost(125);
6891 format %{ "MOV $mem,$src" %}
6892 opcode(0x89);
6893 ins_encode( OpcP, RegMem( src, mem ) );
6894 ins_pipe( ialu_mem_reg );
6895 %}
6897 // Store Integer Immediate
6898 instruct storeImmI(memory mem, immI src) %{
6899 match(Set mem (StoreI mem src));
6901 ins_cost(150);
6902 format %{ "MOV $mem,$src" %}
6903 opcode(0xC7); /* C7 /0 */
6904 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con32( src ));
6905 ins_pipe( ialu_mem_imm );
6906 %}
6908 // Store Short/Char Immediate
6909 instruct storeImmI16(memory mem, immI16 src) %{
6910 predicate(UseStoreImmI16);
6911 match(Set mem (StoreC mem src));
6913 ins_cost(150);
6914 format %{ "MOV16 $mem,$src" %}
6915 opcode(0xC7); /* C7 /0 Same as 32 store immediate with prefix */
6916 ins_encode( SizePrefix, OpcP, RMopc_Mem(0x00,mem), Con16( src ));
6917 ins_pipe( ialu_mem_imm );
6918 %}
6920 // Store Pointer Immediate; null pointers or constant oops that do not
6921 // need card-mark barriers.
6922 instruct storeImmP(memory mem, immP src) %{
6923 match(Set mem (StoreP mem src));
6925 ins_cost(150);
6926 format %{ "MOV $mem,$src" %}
6927 opcode(0xC7); /* C7 /0 */
6928 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con32( src ));
6929 ins_pipe( ialu_mem_imm );
6930 %}
6932 // Store Byte Immediate
6933 instruct storeImmB(memory mem, immI8 src) %{
6934 match(Set mem (StoreB mem src));
6936 ins_cost(150);
6937 format %{ "MOV8 $mem,$src" %}
6938 opcode(0xC6); /* C6 /0 */
6939 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con8or32( src ));
6940 ins_pipe( ialu_mem_imm );
6941 %}
6943 // Store CMS card-mark Immediate
6944 instruct storeImmCM(memory mem, immI8 src) %{
6945 match(Set mem (StoreCM mem src));
6947 ins_cost(150);
6948 format %{ "MOV8 $mem,$src\t! CMS card-mark imm0" %}
6949 opcode(0xC6); /* C6 /0 */
6950 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con8or32( src ));
6951 ins_pipe( ialu_mem_imm );
6952 %}
6954 // Store Double
6955 instruct storeDPR( memory mem, regDPR1 src) %{
6956 predicate(UseSSE<=1);
6957 match(Set mem (StoreD mem src));
6959 ins_cost(100);
6960 format %{ "FST_D $mem,$src" %}
6961 opcode(0xDD); /* DD /2 */
6962 ins_encode( enc_FPR_store(mem,src) );
6963 ins_pipe( fpu_mem_reg );
6964 %}
6966 // Store double does rounding on x86
6967 instruct storeDPR_rounded( memory mem, regDPR1 src) %{
6968 predicate(UseSSE<=1);
6969 match(Set mem (StoreD mem (RoundDouble src)));
6971 ins_cost(100);
6972 format %{ "FST_D $mem,$src\t# round" %}
6973 opcode(0xDD); /* DD /2 */
6974 ins_encode( enc_FPR_store(mem,src) );
6975 ins_pipe( fpu_mem_reg );
6976 %}
6978 // Store XMM register to memory (double-precision floating points)
6979 // MOVSD instruction
6980 instruct storeD(memory mem, regD src) %{
6981 predicate(UseSSE>=2);
6982 match(Set mem (StoreD mem src));
6983 ins_cost(95);
6984 format %{ "MOVSD $mem,$src" %}
6985 ins_encode %{
6986 __ movdbl($mem$$Address, $src$$XMMRegister);
6987 %}
6988 ins_pipe( pipe_slow );
6989 %}
6991 // Store XMM register to memory (single-precision floating point)
6992 // MOVSS instruction
6993 instruct storeF(memory mem, regF src) %{
6994 predicate(UseSSE>=1);
6995 match(Set mem (StoreF mem src));
6996 ins_cost(95);
6997 format %{ "MOVSS $mem,$src" %}
6998 ins_encode %{
6999 __ movflt($mem$$Address, $src$$XMMRegister);
7000 %}
7001 ins_pipe( pipe_slow );
7002 %}
7004 // Store Float
7005 instruct storeFPR( memory mem, regFPR1 src) %{
7006 predicate(UseSSE==0);
7007 match(Set mem (StoreF mem src));
7009 ins_cost(100);
7010 format %{ "FST_S $mem,$src" %}
7011 opcode(0xD9); /* D9 /2 */
7012 ins_encode( enc_FPR_store(mem,src) );
7013 ins_pipe( fpu_mem_reg );
7014 %}
7016 // Store Float does rounding on x86
7017 instruct storeFPR_rounded( memory mem, regFPR1 src) %{
7018 predicate(UseSSE==0);
7019 match(Set mem (StoreF mem (RoundFloat src)));
7021 ins_cost(100);
7022 format %{ "FST_S $mem,$src\t# round" %}
7023 opcode(0xD9); /* D9 /2 */
7024 ins_encode( enc_FPR_store(mem,src) );
7025 ins_pipe( fpu_mem_reg );
7026 %}
7028 // Store Float does rounding on x86
7029 instruct storeFPR_Drounded( memory mem, regDPR1 src) %{
7030 predicate(UseSSE<=1);
7031 match(Set mem (StoreF mem (ConvD2F src)));
7033 ins_cost(100);
7034 format %{ "FST_S $mem,$src\t# D-round" %}
7035 opcode(0xD9); /* D9 /2 */
7036 ins_encode( enc_FPR_store(mem,src) );
7037 ins_pipe( fpu_mem_reg );
7038 %}
7040 // Store immediate Float value (it is faster than store from FPU register)
7041 // The instruction usage is guarded by predicate in operand immFPR().
7042 instruct storeFPR_imm( memory mem, immFPR src) %{
7043 match(Set mem (StoreF mem src));
7045 ins_cost(50);
7046 format %{ "MOV $mem,$src\t# store float" %}
7047 opcode(0xC7); /* C7 /0 */
7048 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con32FPR_as_bits( src ));
7049 ins_pipe( ialu_mem_imm );
7050 %}
7052 // Store immediate Float value (it is faster than store from XMM register)
7053 // The instruction usage is guarded by predicate in operand immF().
7054 instruct storeF_imm( memory mem, immF src) %{
7055 match(Set mem (StoreF mem src));
7057 ins_cost(50);
7058 format %{ "MOV $mem,$src\t# store float" %}
7059 opcode(0xC7); /* C7 /0 */
7060 ins_encode( OpcP, RMopc_Mem(0x00,mem), Con32F_as_bits( src ));
7061 ins_pipe( ialu_mem_imm );
7062 %}
7064 // Store Integer to stack slot
7065 instruct storeSSI(stackSlotI dst, rRegI src) %{
7066 match(Set dst src);
7068 ins_cost(100);
7069 format %{ "MOV $dst,$src" %}
7070 opcode(0x89);
7071 ins_encode( OpcPRegSS( dst, src ) );
7072 ins_pipe( ialu_mem_reg );
7073 %}
7075 // Store Integer to stack slot
7076 instruct storeSSP(stackSlotP dst, eRegP src) %{
7077 match(Set dst src);
7079 ins_cost(100);
7080 format %{ "MOV $dst,$src" %}
7081 opcode(0x89);
7082 ins_encode( OpcPRegSS( dst, src ) );
7083 ins_pipe( ialu_mem_reg );
7084 %}
7086 // Store Long to stack slot
7087 instruct storeSSL(stackSlotL dst, eRegL src) %{
7088 match(Set dst src);
7090 ins_cost(200);
7091 format %{ "MOV $dst,$src.lo\n\t"
7092 "MOV $dst+4,$src.hi" %}
7093 opcode(0x89, 0x89);
7094 ins_encode( OpcP, RegMem( src, dst ), OpcS, RegMem_Hi( src, dst ) );
7095 ins_pipe( ialu_mem_long_reg );
7096 %}
7098 //----------MemBar Instructions-----------------------------------------------
7099 // Memory barrier flavors
7101 instruct membar_acquire() %{
7102 match(MemBarAcquire);
7103 ins_cost(400);
7105 size(0);
7106 format %{ "MEMBAR-acquire ! (empty encoding)" %}
7107 ins_encode();
7108 ins_pipe(empty);
7109 %}
7111 instruct membar_acquire_lock() %{
7112 match(MemBarAcquireLock);
7113 ins_cost(0);
7115 size(0);
7116 format %{ "MEMBAR-acquire (prior CMPXCHG in FastLock so empty encoding)" %}
7117 ins_encode( );
7118 ins_pipe(empty);
7119 %}
7121 instruct membar_release() %{
7122 match(MemBarRelease);
7123 ins_cost(400);
7125 size(0);
7126 format %{ "MEMBAR-release ! (empty encoding)" %}
7127 ins_encode( );
7128 ins_pipe(empty);
7129 %}
7131 instruct membar_release_lock() %{
7132 match(MemBarReleaseLock);
7133 ins_cost(0);
7135 size(0);
7136 format %{ "MEMBAR-release (a FastUnlock follows so empty encoding)" %}
7137 ins_encode( );
7138 ins_pipe(empty);
7139 %}
7141 instruct membar_volatile(eFlagsReg cr) %{
7142 match(MemBarVolatile);
7143 effect(KILL cr);
7144 ins_cost(400);
7146 format %{
7147 $$template
7148 if (os::is_MP()) {
7149 $$emit$$"LOCK ADDL [ESP + #0], 0\t! membar_volatile"
7150 } else {
7151 $$emit$$"MEMBAR-volatile ! (empty encoding)"
7152 }
7153 %}
7154 ins_encode %{
7155 __ membar(Assembler::StoreLoad);
7156 %}
7157 ins_pipe(pipe_slow);
7158 %}
7160 instruct unnecessary_membar_volatile() %{
7161 match(MemBarVolatile);
7162 predicate(Matcher::post_store_load_barrier(n));
7163 ins_cost(0);
7165 size(0);
7166 format %{ "MEMBAR-volatile (unnecessary so empty encoding)" %}
7167 ins_encode( );
7168 ins_pipe(empty);
7169 %}
7171 instruct membar_storestore() %{
7172 match(MemBarStoreStore);
7173 ins_cost(0);
7175 size(0);
7176 format %{ "MEMBAR-storestore (empty encoding)" %}
7177 ins_encode( );
7178 ins_pipe(empty);
7179 %}
7181 //----------Move Instructions--------------------------------------------------
7182 instruct castX2P(eAXRegP dst, eAXRegI src) %{
7183 match(Set dst (CastX2P src));
7184 format %{ "# X2P $dst, $src" %}
7185 ins_encode( /*empty encoding*/ );
7186 ins_cost(0);
7187 ins_pipe(empty);
7188 %}
7190 instruct castP2X(rRegI dst, eRegP src ) %{
7191 match(Set dst (CastP2X src));
7192 ins_cost(50);
7193 format %{ "MOV $dst, $src\t# CastP2X" %}
7194 ins_encode( enc_Copy( dst, src) );
7195 ins_pipe( ialu_reg_reg );
7196 %}
7198 //----------Conditional Move---------------------------------------------------
7199 // Conditional move
7200 instruct jmovI_reg(cmpOp cop, eFlagsReg cr, rRegI dst, rRegI src) %{
7201 predicate(!VM_Version::supports_cmov() );
7202 match(Set dst (CMoveI (Binary cop cr) (Binary dst src)));
7203 ins_cost(200);
7204 format %{ "J$cop,us skip\t# signed cmove\n\t"
7205 "MOV $dst,$src\n"
7206 "skip:" %}
7207 ins_encode %{
7208 Label Lskip;
7209 // Invert sense of branch from sense of CMOV
7210 __ jccb((Assembler::Condition)($cop$$cmpcode^1), Lskip);
7211 __ movl($dst$$Register, $src$$Register);
7212 __ bind(Lskip);
7213 %}
7214 ins_pipe( pipe_cmov_reg );
7215 %}
7217 instruct jmovI_regU(cmpOpU cop, eFlagsRegU cr, rRegI dst, rRegI src) %{
7218 predicate(!VM_Version::supports_cmov() );
7219 match(Set dst (CMoveI (Binary cop cr) (Binary dst src)));
7220 ins_cost(200);
7221 format %{ "J$cop,us skip\t# unsigned cmove\n\t"
7222 "MOV $dst,$src\n"
7223 "skip:" %}
7224 ins_encode %{
7225 Label Lskip;
7226 // Invert sense of branch from sense of CMOV
7227 __ jccb((Assembler::Condition)($cop$$cmpcode^1), Lskip);
7228 __ movl($dst$$Register, $src$$Register);
7229 __ bind(Lskip);
7230 %}
7231 ins_pipe( pipe_cmov_reg );
7232 %}
7234 instruct cmovI_reg(rRegI dst, rRegI src, eFlagsReg cr, cmpOp cop ) %{
7235 predicate(VM_Version::supports_cmov() );
7236 match(Set dst (CMoveI (Binary cop cr) (Binary dst src)));
7237 ins_cost(200);
7238 format %{ "CMOV$cop $dst,$src" %}
7239 opcode(0x0F,0x40);
7240 ins_encode( enc_cmov(cop), RegReg( dst, src ) );
7241 ins_pipe( pipe_cmov_reg );
7242 %}
7244 instruct cmovI_regU( cmpOpU cop, eFlagsRegU cr, rRegI dst, rRegI src ) %{
7245 predicate(VM_Version::supports_cmov() );
7246 match(Set dst (CMoveI (Binary cop cr) (Binary dst src)));
7247 ins_cost(200);
7248 format %{ "CMOV$cop $dst,$src" %}
7249 opcode(0x0F,0x40);
7250 ins_encode( enc_cmov(cop), RegReg( dst, src ) );
7251 ins_pipe( pipe_cmov_reg );
7252 %}
7254 instruct cmovI_regUCF( cmpOpUCF cop, eFlagsRegUCF cr, rRegI dst, rRegI src ) %{
7255 predicate(VM_Version::supports_cmov() );
7256 match(Set dst (CMoveI (Binary cop cr) (Binary dst src)));
7257 ins_cost(200);
7258 expand %{
7259 cmovI_regU(cop, cr, dst, src);
7260 %}
7261 %}
7263 // Conditional move
7264 instruct cmovI_mem(cmpOp cop, eFlagsReg cr, rRegI dst, memory src) %{
7265 predicate(VM_Version::supports_cmov() );
7266 match(Set dst (CMoveI (Binary cop cr) (Binary dst (LoadI src))));
7267 ins_cost(250);
7268 format %{ "CMOV$cop $dst,$src" %}
7269 opcode(0x0F,0x40);
7270 ins_encode( enc_cmov(cop), RegMem( dst, src ) );
7271 ins_pipe( pipe_cmov_mem );
7272 %}
7274 // Conditional move
7275 instruct cmovI_memU(cmpOpU cop, eFlagsRegU cr, rRegI dst, memory src) %{
7276 predicate(VM_Version::supports_cmov() );
7277 match(Set dst (CMoveI (Binary cop cr) (Binary dst (LoadI src))));
7278 ins_cost(250);
7279 format %{ "CMOV$cop $dst,$src" %}
7280 opcode(0x0F,0x40);
7281 ins_encode( enc_cmov(cop), RegMem( dst, src ) );
7282 ins_pipe( pipe_cmov_mem );
7283 %}
7285 instruct cmovI_memUCF(cmpOpUCF cop, eFlagsRegUCF cr, rRegI dst, memory src) %{
7286 predicate(VM_Version::supports_cmov() );
7287 match(Set dst (CMoveI (Binary cop cr) (Binary dst (LoadI src))));
7288 ins_cost(250);
7289 expand %{
7290 cmovI_memU(cop, cr, dst, src);
7291 %}
7292 %}
7294 // Conditional move
7295 instruct cmovP_reg(eRegP dst, eRegP src, eFlagsReg cr, cmpOp cop ) %{
7296 predicate(VM_Version::supports_cmov() );
7297 match(Set dst (CMoveP (Binary cop cr) (Binary dst src)));
7298 ins_cost(200);
7299 format %{ "CMOV$cop $dst,$src\t# ptr" %}
7300 opcode(0x0F,0x40);
7301 ins_encode( enc_cmov(cop), RegReg( dst, src ) );
7302 ins_pipe( pipe_cmov_reg );
7303 %}
7305 // Conditional move (non-P6 version)
7306 // Note: a CMoveP is generated for stubs and native wrappers
7307 // regardless of whether we are on a P6, so we
7308 // emulate a cmov here
7309 instruct cmovP_reg_nonP6(eRegP dst, eRegP src, eFlagsReg cr, cmpOp cop ) %{
7310 match(Set dst (CMoveP (Binary cop cr) (Binary dst src)));
7311 ins_cost(300);
7312 format %{ "Jn$cop skip\n\t"
7313 "MOV $dst,$src\t# pointer\n"
7314 "skip:" %}
7315 opcode(0x8b);
7316 ins_encode( enc_cmov_branch(cop, 0x2), OpcP, RegReg(dst, src));
7317 ins_pipe( pipe_cmov_reg );
7318 %}
7320 // Conditional move
7321 instruct cmovP_regU(cmpOpU cop, eFlagsRegU cr, eRegP dst, eRegP src ) %{
7322 predicate(VM_Version::supports_cmov() );
7323 match(Set dst (CMoveP (Binary cop cr) (Binary dst src)));
7324 ins_cost(200);
7325 format %{ "CMOV$cop $dst,$src\t# ptr" %}
7326 opcode(0x0F,0x40);
7327 ins_encode( enc_cmov(cop), RegReg( dst, src ) );
7328 ins_pipe( pipe_cmov_reg );
7329 %}
7331 instruct cmovP_regUCF(cmpOpUCF cop, eFlagsRegUCF cr, eRegP dst, eRegP src ) %{
7332 predicate(VM_Version::supports_cmov() );
7333 match(Set dst (CMoveP (Binary cop cr) (Binary dst src)));
7334 ins_cost(200);
7335 expand %{
7336 cmovP_regU(cop, cr, dst, src);
7337 %}
7338 %}
7340 // DISABLED: Requires the ADLC to emit a bottom_type call that
7341 // correctly meets the two pointer arguments; one is an incoming
7342 // register but the other is a memory operand. ALSO appears to
7343 // be buggy with implicit null checks.
7344 //
7345 //// Conditional move
7346 //instruct cmovP_mem(cmpOp cop, eFlagsReg 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 //%}
7355 //
7356 //// Conditional move
7357 //instruct cmovP_memU(cmpOpU cop, eFlagsRegU cr, eRegP dst, memory src) %{
7358 // predicate(VM_Version::supports_cmov() );
7359 // match(Set dst (CMoveP (Binary cop cr) (Binary dst (LoadP src))));
7360 // ins_cost(250);
7361 // format %{ "CMOV$cop $dst,$src\t# ptr" %}
7362 // opcode(0x0F,0x40);
7363 // ins_encode( enc_cmov(cop), RegMem( dst, src ) );
7364 // ins_pipe( pipe_cmov_mem );
7365 //%}
7367 // Conditional move
7368 instruct fcmovDPR_regU(cmpOp_fcmov cop, eFlagsRegU cr, regDPR1 dst, regDPR src) %{
7369 predicate(UseSSE<=1);
7370 match(Set dst (CMoveD (Binary cop cr) (Binary dst src)));
7371 ins_cost(200);
7372 format %{ "FCMOV$cop $dst,$src\t# double" %}
7373 opcode(0xDA);
7374 ins_encode( enc_cmov_dpr(cop,src) );
7375 ins_pipe( pipe_cmovDPR_reg );
7376 %}
7378 // Conditional move
7379 instruct fcmovFPR_regU(cmpOp_fcmov cop, eFlagsRegU cr, regFPR1 dst, regFPR src) %{
7380 predicate(UseSSE==0);
7381 match(Set dst (CMoveF (Binary cop cr) (Binary dst src)));
7382 ins_cost(200);
7383 format %{ "FCMOV$cop $dst,$src\t# float" %}
7384 opcode(0xDA);
7385 ins_encode( enc_cmov_dpr(cop,src) );
7386 ins_pipe( pipe_cmovDPR_reg );
7387 %}
7389 // Float CMOV on Intel doesn't handle *signed* compares, only unsigned.
7390 instruct fcmovDPR_regS(cmpOp cop, eFlagsReg cr, regDPR dst, regDPR src) %{
7391 predicate(UseSSE<=1);
7392 match(Set dst (CMoveD (Binary cop cr) (Binary dst src)));
7393 ins_cost(200);
7394 format %{ "Jn$cop skip\n\t"
7395 "MOV $dst,$src\t# double\n"
7396 "skip:" %}
7397 opcode (0xdd, 0x3); /* DD D8+i or DD /3 */
7398 ins_encode( enc_cmov_branch( cop, 0x4 ), Push_Reg_DPR(src), OpcP, RegOpc(dst) );
7399 ins_pipe( pipe_cmovDPR_reg );
7400 %}
7402 // Float CMOV on Intel doesn't handle *signed* compares, only unsigned.
7403 instruct fcmovFPR_regS(cmpOp cop, eFlagsReg cr, regFPR dst, regFPR src) %{
7404 predicate(UseSSE==0);
7405 match(Set dst (CMoveF (Binary cop cr) (Binary dst src)));
7406 ins_cost(200);
7407 format %{ "Jn$cop skip\n\t"
7408 "MOV $dst,$src\t# float\n"
7409 "skip:" %}
7410 opcode (0xdd, 0x3); /* DD D8+i or DD /3 */
7411 ins_encode( enc_cmov_branch( cop, 0x4 ), Push_Reg_FPR(src), OpcP, RegOpc(dst) );
7412 ins_pipe( pipe_cmovDPR_reg );
7413 %}
7415 // No CMOVE with SSE/SSE2
7416 instruct fcmovF_regS(cmpOp cop, eFlagsReg cr, regF dst, regF src) %{
7417 predicate (UseSSE>=1);
7418 match(Set dst (CMoveF (Binary cop cr) (Binary dst src)));
7419 ins_cost(200);
7420 format %{ "Jn$cop skip\n\t"
7421 "MOVSS $dst,$src\t# float\n"
7422 "skip:" %}
7423 ins_encode %{
7424 Label skip;
7425 // Invert sense of branch from sense of CMOV
7426 __ jccb((Assembler::Condition)($cop$$cmpcode^1), skip);
7427 __ movflt($dst$$XMMRegister, $src$$XMMRegister);
7428 __ bind(skip);
7429 %}
7430 ins_pipe( pipe_slow );
7431 %}
7433 // No CMOVE with SSE/SSE2
7434 instruct fcmovD_regS(cmpOp cop, eFlagsReg cr, regD dst, regD src) %{
7435 predicate (UseSSE>=2);
7436 match(Set dst (CMoveD (Binary cop cr) (Binary dst src)));
7437 ins_cost(200);
7438 format %{ "Jn$cop skip\n\t"
7439 "MOVSD $dst,$src\t# float\n"
7440 "skip:" %}
7441 ins_encode %{
7442 Label skip;
7443 // Invert sense of branch from sense of CMOV
7444 __ jccb((Assembler::Condition)($cop$$cmpcode^1), skip);
7445 __ movdbl($dst$$XMMRegister, $src$$XMMRegister);
7446 __ bind(skip);
7447 %}
7448 ins_pipe( pipe_slow );
7449 %}
7451 // unsigned version
7452 instruct fcmovF_regU(cmpOpU cop, eFlagsRegU cr, regF dst, regF src) %{
7453 predicate (UseSSE>=1);
7454 match(Set dst (CMoveF (Binary cop cr) (Binary dst src)));
7455 ins_cost(200);
7456 format %{ "Jn$cop skip\n\t"
7457 "MOVSS $dst,$src\t# float\n"
7458 "skip:" %}
7459 ins_encode %{
7460 Label skip;
7461 // Invert sense of branch from sense of CMOV
7462 __ jccb((Assembler::Condition)($cop$$cmpcode^1), skip);
7463 __ movflt($dst$$XMMRegister, $src$$XMMRegister);
7464 __ bind(skip);
7465 %}
7466 ins_pipe( pipe_slow );
7467 %}
7469 instruct fcmovF_regUCF(cmpOpUCF cop, eFlagsRegUCF cr, regF dst, regF src) %{
7470 predicate (UseSSE>=1);
7471 match(Set dst (CMoveF (Binary cop cr) (Binary dst src)));
7472 ins_cost(200);
7473 expand %{
7474 fcmovF_regU(cop, cr, dst, src);
7475 %}
7476 %}
7478 // unsigned version
7479 instruct fcmovD_regU(cmpOpU cop, eFlagsRegU cr, regD dst, regD src) %{
7480 predicate (UseSSE>=2);
7481 match(Set dst (CMoveD (Binary cop cr) (Binary dst src)));
7482 ins_cost(200);
7483 format %{ "Jn$cop skip\n\t"
7484 "MOVSD $dst,$src\t# float\n"
7485 "skip:" %}
7486 ins_encode %{
7487 Label skip;
7488 // Invert sense of branch from sense of CMOV
7489 __ jccb((Assembler::Condition)($cop$$cmpcode^1), skip);
7490 __ movdbl($dst$$XMMRegister, $src$$XMMRegister);
7491 __ bind(skip);
7492 %}
7493 ins_pipe( pipe_slow );
7494 %}
7496 instruct fcmovD_regUCF(cmpOpUCF cop, eFlagsRegUCF cr, regD dst, regD src) %{
7497 predicate (UseSSE>=2);
7498 match(Set dst (CMoveD (Binary cop cr) (Binary dst src)));
7499 ins_cost(200);
7500 expand %{
7501 fcmovD_regU(cop, cr, dst, src);
7502 %}
7503 %}
7505 instruct cmovL_reg(cmpOp cop, eFlagsReg cr, eRegL dst, eRegL src) %{
7506 predicate(VM_Version::supports_cmov() );
7507 match(Set dst (CMoveL (Binary cop cr) (Binary dst src)));
7508 ins_cost(200);
7509 format %{ "CMOV$cop $dst.lo,$src.lo\n\t"
7510 "CMOV$cop $dst.hi,$src.hi" %}
7511 opcode(0x0F,0x40);
7512 ins_encode( enc_cmov(cop), RegReg_Lo2( dst, src ), enc_cmov(cop), RegReg_Hi2( dst, src ) );
7513 ins_pipe( pipe_cmov_reg_long );
7514 %}
7516 instruct cmovL_regU(cmpOpU cop, eFlagsRegU cr, eRegL dst, eRegL src) %{
7517 predicate(VM_Version::supports_cmov() );
7518 match(Set dst (CMoveL (Binary cop cr) (Binary dst src)));
7519 ins_cost(200);
7520 format %{ "CMOV$cop $dst.lo,$src.lo\n\t"
7521 "CMOV$cop $dst.hi,$src.hi" %}
7522 opcode(0x0F,0x40);
7523 ins_encode( enc_cmov(cop), RegReg_Lo2( dst, src ), enc_cmov(cop), RegReg_Hi2( dst, src ) );
7524 ins_pipe( pipe_cmov_reg_long );
7525 %}
7527 instruct cmovL_regUCF(cmpOpUCF cop, eFlagsRegUCF cr, eRegL dst, eRegL src) %{
7528 predicate(VM_Version::supports_cmov() );
7529 match(Set dst (CMoveL (Binary cop cr) (Binary dst src)));
7530 ins_cost(200);
7531 expand %{
7532 cmovL_regU(cop, cr, dst, src);
7533 %}
7534 %}
7536 //----------Arithmetic Instructions--------------------------------------------
7537 //----------Addition Instructions----------------------------------------------
7538 // Integer Addition Instructions
7539 instruct addI_eReg(rRegI dst, rRegI src, eFlagsReg cr) %{
7540 match(Set dst (AddI dst src));
7541 effect(KILL cr);
7543 size(2);
7544 format %{ "ADD $dst,$src" %}
7545 opcode(0x03);
7546 ins_encode( OpcP, RegReg( dst, src) );
7547 ins_pipe( ialu_reg_reg );
7548 %}
7550 instruct addI_eReg_imm(rRegI dst, immI src, eFlagsReg cr) %{
7551 match(Set dst (AddI dst src));
7552 effect(KILL cr);
7554 format %{ "ADD $dst,$src" %}
7555 opcode(0x81, 0x00); /* /0 id */
7556 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
7557 ins_pipe( ialu_reg );
7558 %}
7560 instruct incI_eReg(rRegI dst, immI1 src, eFlagsReg cr) %{
7561 predicate(UseIncDec);
7562 match(Set dst (AddI dst src));
7563 effect(KILL cr);
7565 size(1);
7566 format %{ "INC $dst" %}
7567 opcode(0x40); /* */
7568 ins_encode( Opc_plus( primary, dst ) );
7569 ins_pipe( ialu_reg );
7570 %}
7572 instruct leaI_eReg_immI(rRegI dst, rRegI src0, immI src1) %{
7573 match(Set dst (AddI src0 src1));
7574 ins_cost(110);
7576 format %{ "LEA $dst,[$src0 + $src1]" %}
7577 opcode(0x8D); /* 0x8D /r */
7578 ins_encode( OpcP, RegLea( dst, src0, src1 ) );
7579 ins_pipe( ialu_reg_reg );
7580 %}
7582 instruct leaP_eReg_immI(eRegP dst, eRegP src0, immI src1) %{
7583 match(Set dst (AddP src0 src1));
7584 ins_cost(110);
7586 format %{ "LEA $dst,[$src0 + $src1]\t# ptr" %}
7587 opcode(0x8D); /* 0x8D /r */
7588 ins_encode( OpcP, RegLea( dst, src0, src1 ) );
7589 ins_pipe( ialu_reg_reg );
7590 %}
7592 instruct decI_eReg(rRegI dst, immI_M1 src, eFlagsReg cr) %{
7593 predicate(UseIncDec);
7594 match(Set dst (AddI dst src));
7595 effect(KILL cr);
7597 size(1);
7598 format %{ "DEC $dst" %}
7599 opcode(0x48); /* */
7600 ins_encode( Opc_plus( primary, dst ) );
7601 ins_pipe( ialu_reg );
7602 %}
7604 instruct addP_eReg(eRegP dst, rRegI src, eFlagsReg cr) %{
7605 match(Set dst (AddP dst src));
7606 effect(KILL cr);
7608 size(2);
7609 format %{ "ADD $dst,$src" %}
7610 opcode(0x03);
7611 ins_encode( OpcP, RegReg( dst, src) );
7612 ins_pipe( ialu_reg_reg );
7613 %}
7615 instruct addP_eReg_imm(eRegP dst, immI src, eFlagsReg cr) %{
7616 match(Set dst (AddP dst src));
7617 effect(KILL cr);
7619 format %{ "ADD $dst,$src" %}
7620 opcode(0x81,0x00); /* Opcode 81 /0 id */
7621 // ins_encode( RegImm( dst, src) );
7622 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
7623 ins_pipe( ialu_reg );
7624 %}
7626 instruct addI_eReg_mem(rRegI dst, memory src, eFlagsReg cr) %{
7627 match(Set dst (AddI dst (LoadI src)));
7628 effect(KILL cr);
7630 ins_cost(125);
7631 format %{ "ADD $dst,$src" %}
7632 opcode(0x03);
7633 ins_encode( OpcP, RegMem( dst, src) );
7634 ins_pipe( ialu_reg_mem );
7635 %}
7637 instruct addI_mem_eReg(memory dst, rRegI src, eFlagsReg cr) %{
7638 match(Set dst (StoreI dst (AddI (LoadI dst) src)));
7639 effect(KILL cr);
7641 ins_cost(150);
7642 format %{ "ADD $dst,$src" %}
7643 opcode(0x01); /* Opcode 01 /r */
7644 ins_encode( OpcP, RegMem( src, dst ) );
7645 ins_pipe( ialu_mem_reg );
7646 %}
7648 // Add Memory with Immediate
7649 instruct addI_mem_imm(memory dst, immI src, eFlagsReg cr) %{
7650 match(Set dst (StoreI dst (AddI (LoadI dst) src)));
7651 effect(KILL cr);
7653 ins_cost(125);
7654 format %{ "ADD $dst,$src" %}
7655 opcode(0x81); /* Opcode 81 /0 id */
7656 ins_encode( OpcSE( src ), RMopc_Mem(0x00,dst), Con8or32( src ) );
7657 ins_pipe( ialu_mem_imm );
7658 %}
7660 instruct incI_mem(memory dst, immI1 src, eFlagsReg cr) %{
7661 match(Set dst (StoreI dst (AddI (LoadI dst) src)));
7662 effect(KILL cr);
7664 ins_cost(125);
7665 format %{ "INC $dst" %}
7666 opcode(0xFF); /* Opcode FF /0 */
7667 ins_encode( OpcP, RMopc_Mem(0x00,dst));
7668 ins_pipe( ialu_mem_imm );
7669 %}
7671 instruct decI_mem(memory dst, immI_M1 src, eFlagsReg cr) %{
7672 match(Set dst (StoreI dst (AddI (LoadI dst) src)));
7673 effect(KILL cr);
7675 ins_cost(125);
7676 format %{ "DEC $dst" %}
7677 opcode(0xFF); /* Opcode FF /1 */
7678 ins_encode( OpcP, RMopc_Mem(0x01,dst));
7679 ins_pipe( ialu_mem_imm );
7680 %}
7683 instruct checkCastPP( eRegP dst ) %{
7684 match(Set dst (CheckCastPP dst));
7686 size(0);
7687 format %{ "#checkcastPP of $dst" %}
7688 ins_encode( /*empty encoding*/ );
7689 ins_pipe( empty );
7690 %}
7692 instruct castPP( eRegP dst ) %{
7693 match(Set dst (CastPP dst));
7694 format %{ "#castPP of $dst" %}
7695 ins_encode( /*empty encoding*/ );
7696 ins_pipe( empty );
7697 %}
7699 instruct castII( rRegI dst ) %{
7700 match(Set dst (CastII dst));
7701 format %{ "#castII of $dst" %}
7702 ins_encode( /*empty encoding*/ );
7703 ins_cost(0);
7704 ins_pipe( empty );
7705 %}
7708 // Load-locked - same as a regular pointer load when used with compare-swap
7709 instruct loadPLocked(eRegP dst, memory mem) %{
7710 match(Set dst (LoadPLocked mem));
7712 ins_cost(125);
7713 format %{ "MOV $dst,$mem\t# Load ptr. locked" %}
7714 opcode(0x8B);
7715 ins_encode( OpcP, RegMem(dst,mem));
7716 ins_pipe( ialu_reg_mem );
7717 %}
7719 // Conditional-store of the updated heap-top.
7720 // Used during allocation of the shared heap.
7721 // Sets flags (EQ) on success. Implemented with a CMPXCHG on Intel.
7722 instruct storePConditional( memory heap_top_ptr, eAXRegP oldval, eRegP newval, eFlagsReg cr ) %{
7723 match(Set cr (StorePConditional heap_top_ptr (Binary oldval newval)));
7724 // EAX is killed if there is contention, but then it's also unused.
7725 // In the common case of no contention, EAX holds the new oop address.
7726 format %{ "CMPXCHG $heap_top_ptr,$newval\t# If EAX==$heap_top_ptr Then store $newval into $heap_top_ptr" %}
7727 ins_encode( lock_prefix, Opcode(0x0F), Opcode(0xB1), RegMem(newval,heap_top_ptr) );
7728 ins_pipe( pipe_cmpxchg );
7729 %}
7731 // Conditional-store of an int value.
7732 // ZF flag is set on success, reset otherwise. Implemented with a CMPXCHG on Intel.
7733 instruct storeIConditional( memory mem, eAXRegI oldval, rRegI newval, eFlagsReg cr ) %{
7734 match(Set cr (StoreIConditional mem (Binary oldval newval)));
7735 effect(KILL oldval);
7736 format %{ "CMPXCHG $mem,$newval\t# If EAX==$mem Then store $newval into $mem" %}
7737 ins_encode( lock_prefix, Opcode(0x0F), Opcode(0xB1), RegMem(newval, mem) );
7738 ins_pipe( pipe_cmpxchg );
7739 %}
7741 // Conditional-store of a long value.
7742 // ZF flag is set on success, reset otherwise. Implemented with a CMPXCHG8 on Intel.
7743 instruct storeLConditional( memory mem, eADXRegL oldval, eBCXRegL newval, eFlagsReg cr ) %{
7744 match(Set cr (StoreLConditional mem (Binary oldval newval)));
7745 effect(KILL oldval);
7746 format %{ "XCHG EBX,ECX\t# correct order for CMPXCHG8 instruction\n\t"
7747 "CMPXCHG8 $mem,ECX:EBX\t# If EDX:EAX==$mem Then store ECX:EBX into $mem\n\t"
7748 "XCHG EBX,ECX"
7749 %}
7750 ins_encode %{
7751 // Note: we need to swap rbx, and rcx before and after the
7752 // cmpxchg8 instruction because the instruction uses
7753 // rcx as the high order word of the new value to store but
7754 // our register encoding uses rbx.
7755 __ xchgl(as_Register(EBX_enc), as_Register(ECX_enc));
7756 if( os::is_MP() )
7757 __ lock();
7758 __ cmpxchg8($mem$$Address);
7759 __ xchgl(as_Register(EBX_enc), as_Register(ECX_enc));
7760 %}
7761 ins_pipe( pipe_cmpxchg );
7762 %}
7764 // No flag versions for CompareAndSwap{P,I,L} because matcher can't match them
7766 instruct compareAndSwapL( rRegI res, eSIRegP mem_ptr, eADXRegL oldval, eBCXRegL newval, eFlagsReg cr ) %{
7767 predicate(VM_Version::supports_cx8());
7768 match(Set res (CompareAndSwapL mem_ptr (Binary oldval newval)));
7769 effect(KILL cr, KILL oldval);
7770 format %{ "CMPXCHG8 [$mem_ptr],$newval\t# If EDX:EAX==[$mem_ptr] Then store $newval into [$mem_ptr]\n\t"
7771 "MOV $res,0\n\t"
7772 "JNE,s fail\n\t"
7773 "MOV $res,1\n"
7774 "fail:" %}
7775 ins_encode( enc_cmpxchg8(mem_ptr),
7776 enc_flags_ne_to_boolean(res) );
7777 ins_pipe( pipe_cmpxchg );
7778 %}
7780 instruct compareAndSwapP( rRegI res, pRegP mem_ptr, eAXRegP oldval, eCXRegP newval, eFlagsReg cr) %{
7781 match(Set res (CompareAndSwapP mem_ptr (Binary oldval newval)));
7782 effect(KILL cr, KILL oldval);
7783 format %{ "CMPXCHG [$mem_ptr],$newval\t# If EAX==[$mem_ptr] Then store $newval into [$mem_ptr]\n\t"
7784 "MOV $res,0\n\t"
7785 "JNE,s fail\n\t"
7786 "MOV $res,1\n"
7787 "fail:" %}
7788 ins_encode( enc_cmpxchg(mem_ptr), enc_flags_ne_to_boolean(res) );
7789 ins_pipe( pipe_cmpxchg );
7790 %}
7792 instruct compareAndSwapI( rRegI res, pRegP mem_ptr, eAXRegI oldval, eCXRegI newval, eFlagsReg cr) %{
7793 match(Set res (CompareAndSwapI mem_ptr (Binary oldval newval)));
7794 effect(KILL cr, KILL oldval);
7795 format %{ "CMPXCHG [$mem_ptr],$newval\t# If EAX==[$mem_ptr] Then store $newval into [$mem_ptr]\n\t"
7796 "MOV $res,0\n\t"
7797 "JNE,s fail\n\t"
7798 "MOV $res,1\n"
7799 "fail:" %}
7800 ins_encode( enc_cmpxchg(mem_ptr), enc_flags_ne_to_boolean(res) );
7801 ins_pipe( pipe_cmpxchg );
7802 %}
7804 instruct xaddI_no_res( memory mem, Universe dummy, immI add, eFlagsReg cr) %{
7805 predicate(n->as_LoadStore()->result_not_used());
7806 match(Set dummy (GetAndAddI mem add));
7807 effect(KILL cr);
7808 format %{ "ADDL [$mem],$add" %}
7809 ins_encode %{
7810 if (os::is_MP()) { __ lock(); }
7811 __ addl($mem$$Address, $add$$constant);
7812 %}
7813 ins_pipe( pipe_cmpxchg );
7814 %}
7816 instruct xaddI( memory mem, rRegI newval, eFlagsReg cr) %{
7817 match(Set newval (GetAndAddI mem newval));
7818 effect(KILL cr);
7819 format %{ "XADDL [$mem],$newval" %}
7820 ins_encode %{
7821 if (os::is_MP()) { __ lock(); }
7822 __ xaddl($mem$$Address, $newval$$Register);
7823 %}
7824 ins_pipe( pipe_cmpxchg );
7825 %}
7827 instruct xchgI( memory mem, rRegI newval) %{
7828 match(Set newval (GetAndSetI mem newval));
7829 format %{ "XCHGL $newval,[$mem]" %}
7830 ins_encode %{
7831 __ xchgl($newval$$Register, $mem$$Address);
7832 %}
7833 ins_pipe( pipe_cmpxchg );
7834 %}
7836 instruct xchgP( memory mem, pRegP newval) %{
7837 match(Set newval (GetAndSetP mem newval));
7838 format %{ "XCHGL $newval,[$mem]" %}
7839 ins_encode %{
7840 __ xchgl($newval$$Register, $mem$$Address);
7841 %}
7842 ins_pipe( pipe_cmpxchg );
7843 %}
7845 //----------Subtraction Instructions-------------------------------------------
7846 // Integer Subtraction Instructions
7847 instruct subI_eReg(rRegI dst, rRegI src, eFlagsReg cr) %{
7848 match(Set dst (SubI dst src));
7849 effect(KILL cr);
7851 size(2);
7852 format %{ "SUB $dst,$src" %}
7853 opcode(0x2B);
7854 ins_encode( OpcP, RegReg( dst, src) );
7855 ins_pipe( ialu_reg_reg );
7856 %}
7858 instruct subI_eReg_imm(rRegI dst, immI src, eFlagsReg cr) %{
7859 match(Set dst (SubI dst src));
7860 effect(KILL cr);
7862 format %{ "SUB $dst,$src" %}
7863 opcode(0x81,0x05); /* Opcode 81 /5 */
7864 // ins_encode( RegImm( dst, src) );
7865 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
7866 ins_pipe( ialu_reg );
7867 %}
7869 instruct subI_eReg_mem(rRegI dst, memory src, eFlagsReg cr) %{
7870 match(Set dst (SubI dst (LoadI src)));
7871 effect(KILL cr);
7873 ins_cost(125);
7874 format %{ "SUB $dst,$src" %}
7875 opcode(0x2B);
7876 ins_encode( OpcP, RegMem( dst, src) );
7877 ins_pipe( ialu_reg_mem );
7878 %}
7880 instruct subI_mem_eReg(memory dst, rRegI src, eFlagsReg cr) %{
7881 match(Set dst (StoreI dst (SubI (LoadI dst) src)));
7882 effect(KILL cr);
7884 ins_cost(150);
7885 format %{ "SUB $dst,$src" %}
7886 opcode(0x29); /* Opcode 29 /r */
7887 ins_encode( OpcP, RegMem( src, dst ) );
7888 ins_pipe( ialu_mem_reg );
7889 %}
7891 // Subtract from a pointer
7892 instruct subP_eReg(eRegP dst, rRegI src, immI0 zero, eFlagsReg cr) %{
7893 match(Set dst (AddP dst (SubI zero src)));
7894 effect(KILL cr);
7896 size(2);
7897 format %{ "SUB $dst,$src" %}
7898 opcode(0x2B);
7899 ins_encode( OpcP, RegReg( dst, src) );
7900 ins_pipe( ialu_reg_reg );
7901 %}
7903 instruct negI_eReg(rRegI dst, immI0 zero, eFlagsReg cr) %{
7904 match(Set dst (SubI zero dst));
7905 effect(KILL cr);
7907 size(2);
7908 format %{ "NEG $dst" %}
7909 opcode(0xF7,0x03); // Opcode F7 /3
7910 ins_encode( OpcP, RegOpc( dst ) );
7911 ins_pipe( ialu_reg );
7912 %}
7915 //----------Multiplication/Division Instructions-------------------------------
7916 // Integer Multiplication Instructions
7917 // Multiply Register
7918 instruct mulI_eReg(rRegI dst, rRegI src, eFlagsReg cr) %{
7919 match(Set dst (MulI dst src));
7920 effect(KILL cr);
7922 size(3);
7923 ins_cost(300);
7924 format %{ "IMUL $dst,$src" %}
7925 opcode(0xAF, 0x0F);
7926 ins_encode( OpcS, OpcP, RegReg( dst, src) );
7927 ins_pipe( ialu_reg_reg_alu0 );
7928 %}
7930 // Multiply 32-bit Immediate
7931 instruct mulI_eReg_imm(rRegI dst, rRegI src, immI imm, eFlagsReg cr) %{
7932 match(Set dst (MulI src imm));
7933 effect(KILL cr);
7935 ins_cost(300);
7936 format %{ "IMUL $dst,$src,$imm" %}
7937 opcode(0x69); /* 69 /r id */
7938 ins_encode( OpcSE(imm), RegReg( dst, src ), Con8or32( imm ) );
7939 ins_pipe( ialu_reg_reg_alu0 );
7940 %}
7942 instruct loadConL_low_only(eADXRegL_low_only dst, immL32 src, eFlagsReg cr) %{
7943 match(Set dst src);
7944 effect(KILL cr);
7946 // Note that this is artificially increased to make it more expensive than loadConL
7947 ins_cost(250);
7948 format %{ "MOV EAX,$src\t// low word only" %}
7949 opcode(0xB8);
7950 ins_encode( LdImmL_Lo(dst, src) );
7951 ins_pipe( ialu_reg_fat );
7952 %}
7954 // Multiply by 32-bit Immediate, taking the shifted high order results
7955 // (special case for shift by 32)
7956 instruct mulI_imm_high(eDXRegI dst, nadxRegI src1, eADXRegL_low_only src2, immI_32 cnt, eFlagsReg cr) %{
7957 match(Set dst (ConvL2I (RShiftL (MulL (ConvI2L src1) src2) cnt)));
7958 predicate( _kids[0]->_kids[0]->_kids[1]->_leaf->Opcode() == Op_ConL &&
7959 _kids[0]->_kids[0]->_kids[1]->_leaf->as_Type()->type()->is_long()->get_con() >= min_jint &&
7960 _kids[0]->_kids[0]->_kids[1]->_leaf->as_Type()->type()->is_long()->get_con() <= max_jint );
7961 effect(USE src1, KILL cr);
7963 // Note that this is adjusted by 150 to compensate for the overcosting of loadConL_low_only
7964 ins_cost(0*100 + 1*400 - 150);
7965 format %{ "IMUL EDX:EAX,$src1" %}
7966 ins_encode( multiply_con_and_shift_high( dst, src1, src2, cnt, cr ) );
7967 ins_pipe( pipe_slow );
7968 %}
7970 // Multiply by 32-bit Immediate, taking the shifted high order results
7971 instruct mulI_imm_RShift_high(eDXRegI dst, nadxRegI src1, eADXRegL_low_only src2, immI_32_63 cnt, eFlagsReg cr) %{
7972 match(Set dst (ConvL2I (RShiftL (MulL (ConvI2L src1) src2) cnt)));
7973 predicate( _kids[0]->_kids[0]->_kids[1]->_leaf->Opcode() == Op_ConL &&
7974 _kids[0]->_kids[0]->_kids[1]->_leaf->as_Type()->type()->is_long()->get_con() >= min_jint &&
7975 _kids[0]->_kids[0]->_kids[1]->_leaf->as_Type()->type()->is_long()->get_con() <= max_jint );
7976 effect(USE src1, KILL cr);
7978 // Note that this is adjusted by 150 to compensate for the overcosting of loadConL_low_only
7979 ins_cost(1*100 + 1*400 - 150);
7980 format %{ "IMUL EDX:EAX,$src1\n\t"
7981 "SAR EDX,$cnt-32" %}
7982 ins_encode( multiply_con_and_shift_high( dst, src1, src2, cnt, cr ) );
7983 ins_pipe( pipe_slow );
7984 %}
7986 // Multiply Memory 32-bit Immediate
7987 instruct mulI_mem_imm(rRegI dst, memory src, immI imm, eFlagsReg cr) %{
7988 match(Set dst (MulI (LoadI src) imm));
7989 effect(KILL cr);
7991 ins_cost(300);
7992 format %{ "IMUL $dst,$src,$imm" %}
7993 opcode(0x69); /* 69 /r id */
7994 ins_encode( OpcSE(imm), RegMem( dst, src ), Con8or32( imm ) );
7995 ins_pipe( ialu_reg_mem_alu0 );
7996 %}
7998 // Multiply Memory
7999 instruct mulI(rRegI dst, memory src, eFlagsReg cr) %{
8000 match(Set dst (MulI dst (LoadI src)));
8001 effect(KILL cr);
8003 ins_cost(350);
8004 format %{ "IMUL $dst,$src" %}
8005 opcode(0xAF, 0x0F);
8006 ins_encode( OpcS, OpcP, RegMem( dst, src) );
8007 ins_pipe( ialu_reg_mem_alu0 );
8008 %}
8010 // Multiply Register Int to Long
8011 instruct mulI2L(eADXRegL dst, eAXRegI src, nadxRegI src1, eFlagsReg flags) %{
8012 // Basic Idea: long = (long)int * (long)int
8013 match(Set dst (MulL (ConvI2L src) (ConvI2L src1)));
8014 effect(DEF dst, USE src, USE src1, KILL flags);
8016 ins_cost(300);
8017 format %{ "IMUL $dst,$src1" %}
8019 ins_encode( long_int_multiply( dst, src1 ) );
8020 ins_pipe( ialu_reg_reg_alu0 );
8021 %}
8023 instruct mulIS_eReg(eADXRegL dst, immL_32bits mask, eFlagsReg flags, eAXRegI src, nadxRegI src1) %{
8024 // Basic Idea: long = (int & 0xffffffffL) * (int & 0xffffffffL)
8025 match(Set dst (MulL (AndL (ConvI2L src) mask) (AndL (ConvI2L src1) mask)));
8026 effect(KILL flags);
8028 ins_cost(300);
8029 format %{ "MUL $dst,$src1" %}
8031 ins_encode( long_uint_multiply(dst, src1) );
8032 ins_pipe( ialu_reg_reg_alu0 );
8033 %}
8035 // Multiply Register Long
8036 instruct mulL_eReg(eADXRegL dst, eRegL src, rRegI tmp, eFlagsReg cr) %{
8037 match(Set dst (MulL dst src));
8038 effect(KILL cr, TEMP tmp);
8039 ins_cost(4*100+3*400);
8040 // Basic idea: lo(result) = lo(x_lo * y_lo)
8041 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi)
8042 format %{ "MOV $tmp,$src.lo\n\t"
8043 "IMUL $tmp,EDX\n\t"
8044 "MOV EDX,$src.hi\n\t"
8045 "IMUL EDX,EAX\n\t"
8046 "ADD $tmp,EDX\n\t"
8047 "MUL EDX:EAX,$src.lo\n\t"
8048 "ADD EDX,$tmp" %}
8049 ins_encode( long_multiply( dst, src, tmp ) );
8050 ins_pipe( pipe_slow );
8051 %}
8053 // Multiply Register Long where the left operand's high 32 bits are zero
8054 instruct mulL_eReg_lhi0(eADXRegL dst, eRegL src, rRegI tmp, eFlagsReg cr) %{
8055 predicate(is_operand_hi32_zero(n->in(1)));
8056 match(Set dst (MulL dst src));
8057 effect(KILL cr, TEMP tmp);
8058 ins_cost(2*100+2*400);
8059 // Basic idea: lo(result) = lo(x_lo * y_lo)
8060 // hi(result) = hi(x_lo * y_lo) + lo(x_lo * y_hi) where lo(x_hi * y_lo) = 0 because x_hi = 0
8061 format %{ "MOV $tmp,$src.hi\n\t"
8062 "IMUL $tmp,EAX\n\t"
8063 "MUL EDX:EAX,$src.lo\n\t"
8064 "ADD EDX,$tmp" %}
8065 ins_encode %{
8066 __ movl($tmp$$Register, HIGH_FROM_LOW($src$$Register));
8067 __ imull($tmp$$Register, rax);
8068 __ mull($src$$Register);
8069 __ addl(rdx, $tmp$$Register);
8070 %}
8071 ins_pipe( pipe_slow );
8072 %}
8074 // Multiply Register Long where the right operand's high 32 bits are zero
8075 instruct mulL_eReg_rhi0(eADXRegL dst, eRegL src, rRegI tmp, eFlagsReg cr) %{
8076 predicate(is_operand_hi32_zero(n->in(2)));
8077 match(Set dst (MulL dst src));
8078 effect(KILL cr, TEMP tmp);
8079 ins_cost(2*100+2*400);
8080 // Basic idea: lo(result) = lo(x_lo * y_lo)
8081 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) where lo(x_lo * y_hi) = 0 because y_hi = 0
8082 format %{ "MOV $tmp,$src.lo\n\t"
8083 "IMUL $tmp,EDX\n\t"
8084 "MUL EDX:EAX,$src.lo\n\t"
8085 "ADD EDX,$tmp" %}
8086 ins_encode %{
8087 __ movl($tmp$$Register, $src$$Register);
8088 __ imull($tmp$$Register, rdx);
8089 __ mull($src$$Register);
8090 __ addl(rdx, $tmp$$Register);
8091 %}
8092 ins_pipe( pipe_slow );
8093 %}
8095 // Multiply Register Long where the left and the right operands' high 32 bits are zero
8096 instruct mulL_eReg_hi0(eADXRegL dst, eRegL src, eFlagsReg cr) %{
8097 predicate(is_operand_hi32_zero(n->in(1)) && is_operand_hi32_zero(n->in(2)));
8098 match(Set dst (MulL dst src));
8099 effect(KILL cr);
8100 ins_cost(1*400);
8101 // Basic idea: lo(result) = lo(x_lo * y_lo)
8102 // hi(result) = hi(x_lo * y_lo) where lo(x_hi * y_lo) = 0 and lo(x_lo * y_hi) = 0 because x_hi = 0 and y_hi = 0
8103 format %{ "MUL EDX:EAX,$src.lo\n\t" %}
8104 ins_encode %{
8105 __ mull($src$$Register);
8106 %}
8107 ins_pipe( pipe_slow );
8108 %}
8110 // Multiply Register Long by small constant
8111 instruct mulL_eReg_con(eADXRegL dst, immL_127 src, rRegI tmp, eFlagsReg cr) %{
8112 match(Set dst (MulL dst src));
8113 effect(KILL cr, TEMP tmp);
8114 ins_cost(2*100+2*400);
8115 size(12);
8116 // Basic idea: lo(result) = lo(src * EAX)
8117 // hi(result) = hi(src * EAX) + lo(src * EDX)
8118 format %{ "IMUL $tmp,EDX,$src\n\t"
8119 "MOV EDX,$src\n\t"
8120 "MUL EDX\t# EDX*EAX -> EDX:EAX\n\t"
8121 "ADD EDX,$tmp" %}
8122 ins_encode( long_multiply_con( dst, src, tmp ) );
8123 ins_pipe( pipe_slow );
8124 %}
8126 // Integer DIV with Register
8127 instruct divI_eReg(eAXRegI rax, eDXRegI rdx, eCXRegI div, eFlagsReg cr) %{
8128 match(Set rax (DivI rax div));
8129 effect(KILL rdx, KILL cr);
8130 size(26);
8131 ins_cost(30*100+10*100);
8132 format %{ "CMP EAX,0x80000000\n\t"
8133 "JNE,s normal\n\t"
8134 "XOR EDX,EDX\n\t"
8135 "CMP ECX,-1\n\t"
8136 "JE,s done\n"
8137 "normal: CDQ\n\t"
8138 "IDIV $div\n\t"
8139 "done:" %}
8140 opcode(0xF7, 0x7); /* Opcode F7 /7 */
8141 ins_encode( cdq_enc, OpcP, RegOpc(div) );
8142 ins_pipe( ialu_reg_reg_alu0 );
8143 %}
8145 // Divide Register Long
8146 instruct divL_eReg( eADXRegL dst, eRegL src1, eRegL src2, eFlagsReg cr, eCXRegI cx, eBXRegI bx ) %{
8147 match(Set dst (DivL src1 src2));
8148 effect( KILL cr, KILL cx, KILL bx );
8149 ins_cost(10000);
8150 format %{ "PUSH $src1.hi\n\t"
8151 "PUSH $src1.lo\n\t"
8152 "PUSH $src2.hi\n\t"
8153 "PUSH $src2.lo\n\t"
8154 "CALL SharedRuntime::ldiv\n\t"
8155 "ADD ESP,16" %}
8156 ins_encode( long_div(src1,src2) );
8157 ins_pipe( pipe_slow );
8158 %}
8160 // Integer DIVMOD with Register, both quotient and mod results
8161 instruct divModI_eReg_divmod(eAXRegI rax, eDXRegI rdx, eCXRegI div, eFlagsReg cr) %{
8162 match(DivModI rax div);
8163 effect(KILL cr);
8164 size(26);
8165 ins_cost(30*100+10*100);
8166 format %{ "CMP EAX,0x80000000\n\t"
8167 "JNE,s normal\n\t"
8168 "XOR EDX,EDX\n\t"
8169 "CMP ECX,-1\n\t"
8170 "JE,s done\n"
8171 "normal: CDQ\n\t"
8172 "IDIV $div\n\t"
8173 "done:" %}
8174 opcode(0xF7, 0x7); /* Opcode F7 /7 */
8175 ins_encode( cdq_enc, OpcP, RegOpc(div) );
8176 ins_pipe( pipe_slow );
8177 %}
8179 // Integer MOD with Register
8180 instruct modI_eReg(eDXRegI rdx, eAXRegI rax, eCXRegI div, eFlagsReg cr) %{
8181 match(Set rdx (ModI rax div));
8182 effect(KILL rax, KILL cr);
8184 size(26);
8185 ins_cost(300);
8186 format %{ "CDQ\n\t"
8187 "IDIV $div" %}
8188 opcode(0xF7, 0x7); /* Opcode F7 /7 */
8189 ins_encode( cdq_enc, OpcP, RegOpc(div) );
8190 ins_pipe( ialu_reg_reg_alu0 );
8191 %}
8193 // Remainder Register Long
8194 instruct modL_eReg( eADXRegL dst, eRegL src1, eRegL src2, eFlagsReg cr, eCXRegI cx, eBXRegI bx ) %{
8195 match(Set dst (ModL src1 src2));
8196 effect( KILL cr, KILL cx, KILL bx );
8197 ins_cost(10000);
8198 format %{ "PUSH $src1.hi\n\t"
8199 "PUSH $src1.lo\n\t"
8200 "PUSH $src2.hi\n\t"
8201 "PUSH $src2.lo\n\t"
8202 "CALL SharedRuntime::lrem\n\t"
8203 "ADD ESP,16" %}
8204 ins_encode( long_mod(src1,src2) );
8205 ins_pipe( pipe_slow );
8206 %}
8208 // Divide Register Long (no special case since divisor != -1)
8209 instruct divL_eReg_imm32( eADXRegL dst, immL32 imm, rRegI tmp, rRegI tmp2, eFlagsReg cr ) %{
8210 match(Set dst (DivL dst imm));
8211 effect( TEMP tmp, TEMP tmp2, KILL cr );
8212 ins_cost(1000);
8213 format %{ "MOV $tmp,abs($imm) # ldiv EDX:EAX,$imm\n\t"
8214 "XOR $tmp2,$tmp2\n\t"
8215 "CMP $tmp,EDX\n\t"
8216 "JA,s fast\n\t"
8217 "MOV $tmp2,EAX\n\t"
8218 "MOV EAX,EDX\n\t"
8219 "MOV EDX,0\n\t"
8220 "JLE,s pos\n\t"
8221 "LNEG EAX : $tmp2\n\t"
8222 "DIV $tmp # unsigned division\n\t"
8223 "XCHG EAX,$tmp2\n\t"
8224 "DIV $tmp\n\t"
8225 "LNEG $tmp2 : EAX\n\t"
8226 "JMP,s done\n"
8227 "pos:\n\t"
8228 "DIV $tmp\n\t"
8229 "XCHG EAX,$tmp2\n"
8230 "fast:\n\t"
8231 "DIV $tmp\n"
8232 "done:\n\t"
8233 "MOV EDX,$tmp2\n\t"
8234 "NEG EDX:EAX # if $imm < 0" %}
8235 ins_encode %{
8236 int con = (int)$imm$$constant;
8237 assert(con != 0 && con != -1 && con != min_jint, "wrong divisor");
8238 int pcon = (con > 0) ? con : -con;
8239 Label Lfast, Lpos, Ldone;
8241 __ movl($tmp$$Register, pcon);
8242 __ xorl($tmp2$$Register,$tmp2$$Register);
8243 __ cmpl($tmp$$Register, HIGH_FROM_LOW($dst$$Register));
8244 __ jccb(Assembler::above, Lfast); // result fits into 32 bit
8246 __ movl($tmp2$$Register, $dst$$Register); // save
8247 __ movl($dst$$Register, HIGH_FROM_LOW($dst$$Register));
8248 __ movl(HIGH_FROM_LOW($dst$$Register),0); // preserve flags
8249 __ jccb(Assembler::lessEqual, Lpos); // result is positive
8251 // Negative dividend.
8252 // convert value to positive to use unsigned division
8253 __ lneg($dst$$Register, $tmp2$$Register);
8254 __ divl($tmp$$Register);
8255 __ xchgl($dst$$Register, $tmp2$$Register);
8256 __ divl($tmp$$Register);
8257 // revert result back to negative
8258 __ lneg($tmp2$$Register, $dst$$Register);
8259 __ jmpb(Ldone);
8261 __ bind(Lpos);
8262 __ divl($tmp$$Register); // Use unsigned division
8263 __ xchgl($dst$$Register, $tmp2$$Register);
8264 // Fallthrow for final divide, tmp2 has 32 bit hi result
8266 __ bind(Lfast);
8267 // fast path: src is positive
8268 __ divl($tmp$$Register); // Use unsigned division
8270 __ bind(Ldone);
8271 __ movl(HIGH_FROM_LOW($dst$$Register),$tmp2$$Register);
8272 if (con < 0) {
8273 __ lneg(HIGH_FROM_LOW($dst$$Register), $dst$$Register);
8274 }
8275 %}
8276 ins_pipe( pipe_slow );
8277 %}
8279 // Remainder Register Long (remainder fit into 32 bits)
8280 instruct modL_eReg_imm32( eADXRegL dst, immL32 imm, rRegI tmp, rRegI tmp2, eFlagsReg cr ) %{
8281 match(Set dst (ModL dst imm));
8282 effect( TEMP tmp, TEMP tmp2, KILL cr );
8283 ins_cost(1000);
8284 format %{ "MOV $tmp,abs($imm) # lrem EDX:EAX,$imm\n\t"
8285 "CMP $tmp,EDX\n\t"
8286 "JA,s fast\n\t"
8287 "MOV $tmp2,EAX\n\t"
8288 "MOV EAX,EDX\n\t"
8289 "MOV EDX,0\n\t"
8290 "JLE,s pos\n\t"
8291 "LNEG EAX : $tmp2\n\t"
8292 "DIV $tmp # unsigned division\n\t"
8293 "MOV EAX,$tmp2\n\t"
8294 "DIV $tmp\n\t"
8295 "NEG EDX\n\t"
8296 "JMP,s done\n"
8297 "pos:\n\t"
8298 "DIV $tmp\n\t"
8299 "MOV EAX,$tmp2\n"
8300 "fast:\n\t"
8301 "DIV $tmp\n"
8302 "done:\n\t"
8303 "MOV EAX,EDX\n\t"
8304 "SAR EDX,31\n\t" %}
8305 ins_encode %{
8306 int con = (int)$imm$$constant;
8307 assert(con != 0 && con != -1 && con != min_jint, "wrong divisor");
8308 int pcon = (con > 0) ? con : -con;
8309 Label Lfast, Lpos, Ldone;
8311 __ movl($tmp$$Register, pcon);
8312 __ cmpl($tmp$$Register, HIGH_FROM_LOW($dst$$Register));
8313 __ jccb(Assembler::above, Lfast); // src is positive and result fits into 32 bit
8315 __ movl($tmp2$$Register, $dst$$Register); // save
8316 __ movl($dst$$Register, HIGH_FROM_LOW($dst$$Register));
8317 __ movl(HIGH_FROM_LOW($dst$$Register),0); // preserve flags
8318 __ jccb(Assembler::lessEqual, Lpos); // result is positive
8320 // Negative dividend.
8321 // convert value to positive to use unsigned division
8322 __ lneg($dst$$Register, $tmp2$$Register);
8323 __ divl($tmp$$Register);
8324 __ movl($dst$$Register, $tmp2$$Register);
8325 __ divl($tmp$$Register);
8326 // revert remainder back to negative
8327 __ negl(HIGH_FROM_LOW($dst$$Register));
8328 __ jmpb(Ldone);
8330 __ bind(Lpos);
8331 __ divl($tmp$$Register);
8332 __ movl($dst$$Register, $tmp2$$Register);
8334 __ bind(Lfast);
8335 // fast path: src is positive
8336 __ divl($tmp$$Register);
8338 __ bind(Ldone);
8339 __ movl($dst$$Register, HIGH_FROM_LOW($dst$$Register));
8340 __ sarl(HIGH_FROM_LOW($dst$$Register), 31); // result sign
8342 %}
8343 ins_pipe( pipe_slow );
8344 %}
8346 // Integer Shift Instructions
8347 // Shift Left by one
8348 instruct shlI_eReg_1(rRegI dst, immI1 shift, eFlagsReg cr) %{
8349 match(Set dst (LShiftI dst shift));
8350 effect(KILL cr);
8352 size(2);
8353 format %{ "SHL $dst,$shift" %}
8354 opcode(0xD1, 0x4); /* D1 /4 */
8355 ins_encode( OpcP, RegOpc( dst ) );
8356 ins_pipe( ialu_reg );
8357 %}
8359 // Shift Left by 8-bit immediate
8360 instruct salI_eReg_imm(rRegI dst, immI8 shift, eFlagsReg cr) %{
8361 match(Set dst (LShiftI dst shift));
8362 effect(KILL cr);
8364 size(3);
8365 format %{ "SHL $dst,$shift" %}
8366 opcode(0xC1, 0x4); /* C1 /4 ib */
8367 ins_encode( RegOpcImm( dst, shift) );
8368 ins_pipe( ialu_reg );
8369 %}
8371 // Shift Left by variable
8372 instruct salI_eReg_CL(rRegI dst, eCXRegI shift, eFlagsReg cr) %{
8373 match(Set dst (LShiftI dst shift));
8374 effect(KILL cr);
8376 size(2);
8377 format %{ "SHL $dst,$shift" %}
8378 opcode(0xD3, 0x4); /* D3 /4 */
8379 ins_encode( OpcP, RegOpc( dst ) );
8380 ins_pipe( ialu_reg_reg );
8381 %}
8383 // Arithmetic shift right by one
8384 instruct sarI_eReg_1(rRegI dst, immI1 shift, eFlagsReg cr) %{
8385 match(Set dst (RShiftI dst shift));
8386 effect(KILL cr);
8388 size(2);
8389 format %{ "SAR $dst,$shift" %}
8390 opcode(0xD1, 0x7); /* D1 /7 */
8391 ins_encode( OpcP, RegOpc( dst ) );
8392 ins_pipe( ialu_reg );
8393 %}
8395 // Arithmetic shift right by one
8396 instruct sarI_mem_1(memory dst, immI1 shift, eFlagsReg cr) %{
8397 match(Set dst (StoreI dst (RShiftI (LoadI dst) shift)));
8398 effect(KILL cr);
8399 format %{ "SAR $dst,$shift" %}
8400 opcode(0xD1, 0x7); /* D1 /7 */
8401 ins_encode( OpcP, RMopc_Mem(secondary,dst) );
8402 ins_pipe( ialu_mem_imm );
8403 %}
8405 // Arithmetic Shift Right by 8-bit immediate
8406 instruct sarI_eReg_imm(rRegI dst, immI8 shift, eFlagsReg cr) %{
8407 match(Set dst (RShiftI dst shift));
8408 effect(KILL cr);
8410 size(3);
8411 format %{ "SAR $dst,$shift" %}
8412 opcode(0xC1, 0x7); /* C1 /7 ib */
8413 ins_encode( RegOpcImm( dst, shift ) );
8414 ins_pipe( ialu_mem_imm );
8415 %}
8417 // Arithmetic Shift Right by 8-bit immediate
8418 instruct sarI_mem_imm(memory dst, immI8 shift, eFlagsReg cr) %{
8419 match(Set dst (StoreI dst (RShiftI (LoadI dst) shift)));
8420 effect(KILL cr);
8422 format %{ "SAR $dst,$shift" %}
8423 opcode(0xC1, 0x7); /* C1 /7 ib */
8424 ins_encode( OpcP, RMopc_Mem(secondary, dst ), Con8or32( shift ) );
8425 ins_pipe( ialu_mem_imm );
8426 %}
8428 // Arithmetic Shift Right by variable
8429 instruct sarI_eReg_CL(rRegI dst, eCXRegI shift, eFlagsReg cr) %{
8430 match(Set dst (RShiftI dst shift));
8431 effect(KILL cr);
8433 size(2);
8434 format %{ "SAR $dst,$shift" %}
8435 opcode(0xD3, 0x7); /* D3 /7 */
8436 ins_encode( OpcP, RegOpc( dst ) );
8437 ins_pipe( ialu_reg_reg );
8438 %}
8440 // Logical shift right by one
8441 instruct shrI_eReg_1(rRegI dst, immI1 shift, eFlagsReg cr) %{
8442 match(Set dst (URShiftI dst shift));
8443 effect(KILL cr);
8445 size(2);
8446 format %{ "SHR $dst,$shift" %}
8447 opcode(0xD1, 0x5); /* D1 /5 */
8448 ins_encode( OpcP, RegOpc( dst ) );
8449 ins_pipe( ialu_reg );
8450 %}
8452 // Logical Shift Right by 8-bit immediate
8453 instruct shrI_eReg_imm(rRegI dst, immI8 shift, eFlagsReg cr) %{
8454 match(Set dst (URShiftI dst shift));
8455 effect(KILL cr);
8457 size(3);
8458 format %{ "SHR $dst,$shift" %}
8459 opcode(0xC1, 0x5); /* C1 /5 ib */
8460 ins_encode( RegOpcImm( dst, shift) );
8461 ins_pipe( ialu_reg );
8462 %}
8465 // Logical Shift Right by 24, followed by Arithmetic Shift Left by 24.
8466 // This idiom is used by the compiler for the i2b bytecode.
8467 instruct i2b(rRegI dst, xRegI src, immI_24 twentyfour) %{
8468 match(Set dst (RShiftI (LShiftI src twentyfour) twentyfour));
8470 size(3);
8471 format %{ "MOVSX $dst,$src :8" %}
8472 ins_encode %{
8473 __ movsbl($dst$$Register, $src$$Register);
8474 %}
8475 ins_pipe(ialu_reg_reg);
8476 %}
8478 // Logical Shift Right by 16, followed by Arithmetic Shift Left by 16.
8479 // This idiom is used by the compiler the i2s bytecode.
8480 instruct i2s(rRegI dst, xRegI src, immI_16 sixteen) %{
8481 match(Set dst (RShiftI (LShiftI src sixteen) sixteen));
8483 size(3);
8484 format %{ "MOVSX $dst,$src :16" %}
8485 ins_encode %{
8486 __ movswl($dst$$Register, $src$$Register);
8487 %}
8488 ins_pipe(ialu_reg_reg);
8489 %}
8492 // Logical Shift Right by variable
8493 instruct shrI_eReg_CL(rRegI dst, eCXRegI shift, eFlagsReg cr) %{
8494 match(Set dst (URShiftI dst shift));
8495 effect(KILL cr);
8497 size(2);
8498 format %{ "SHR $dst,$shift" %}
8499 opcode(0xD3, 0x5); /* D3 /5 */
8500 ins_encode( OpcP, RegOpc( dst ) );
8501 ins_pipe( ialu_reg_reg );
8502 %}
8505 //----------Logical Instructions-----------------------------------------------
8506 //----------Integer Logical Instructions---------------------------------------
8507 // And Instructions
8508 // And Register with Register
8509 instruct andI_eReg(rRegI dst, rRegI src, eFlagsReg cr) %{
8510 match(Set dst (AndI dst src));
8511 effect(KILL cr);
8513 size(2);
8514 format %{ "AND $dst,$src" %}
8515 opcode(0x23);
8516 ins_encode( OpcP, RegReg( dst, src) );
8517 ins_pipe( ialu_reg_reg );
8518 %}
8520 // And Register with Immediate
8521 instruct andI_eReg_imm(rRegI dst, immI src, eFlagsReg cr) %{
8522 match(Set dst (AndI dst src));
8523 effect(KILL cr);
8525 format %{ "AND $dst,$src" %}
8526 opcode(0x81,0x04); /* Opcode 81 /4 */
8527 // ins_encode( RegImm( dst, src) );
8528 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
8529 ins_pipe( ialu_reg );
8530 %}
8532 // And Register with Memory
8533 instruct andI_eReg_mem(rRegI dst, memory src, eFlagsReg cr) %{
8534 match(Set dst (AndI dst (LoadI src)));
8535 effect(KILL cr);
8537 ins_cost(125);
8538 format %{ "AND $dst,$src" %}
8539 opcode(0x23);
8540 ins_encode( OpcP, RegMem( dst, src) );
8541 ins_pipe( ialu_reg_mem );
8542 %}
8544 // And Memory with Register
8545 instruct andI_mem_eReg(memory dst, rRegI src, eFlagsReg cr) %{
8546 match(Set dst (StoreI dst (AndI (LoadI dst) src)));
8547 effect(KILL cr);
8549 ins_cost(150);
8550 format %{ "AND $dst,$src" %}
8551 opcode(0x21); /* Opcode 21 /r */
8552 ins_encode( OpcP, RegMem( src, dst ) );
8553 ins_pipe( ialu_mem_reg );
8554 %}
8556 // And Memory with Immediate
8557 instruct andI_mem_imm(memory dst, immI src, eFlagsReg cr) %{
8558 match(Set dst (StoreI dst (AndI (LoadI dst) src)));
8559 effect(KILL cr);
8561 ins_cost(125);
8562 format %{ "AND $dst,$src" %}
8563 opcode(0x81, 0x4); /* Opcode 81 /4 id */
8564 // ins_encode( MemImm( dst, src) );
8565 ins_encode( OpcSE( src ), RMopc_Mem(secondary, dst ), Con8or32( src ) );
8566 ins_pipe( ialu_mem_imm );
8567 %}
8569 // Or Instructions
8570 // Or Register with Register
8571 instruct orI_eReg(rRegI dst, rRegI src, eFlagsReg cr) %{
8572 match(Set dst (OrI dst src));
8573 effect(KILL cr);
8575 size(2);
8576 format %{ "OR $dst,$src" %}
8577 opcode(0x0B);
8578 ins_encode( OpcP, RegReg( dst, src) );
8579 ins_pipe( ialu_reg_reg );
8580 %}
8582 instruct orI_eReg_castP2X(rRegI dst, eRegP src, eFlagsReg cr) %{
8583 match(Set dst (OrI dst (CastP2X src)));
8584 effect(KILL cr);
8586 size(2);
8587 format %{ "OR $dst,$src" %}
8588 opcode(0x0B);
8589 ins_encode( OpcP, RegReg( dst, src) );
8590 ins_pipe( ialu_reg_reg );
8591 %}
8594 // Or Register with Immediate
8595 instruct orI_eReg_imm(rRegI dst, immI src, eFlagsReg cr) %{
8596 match(Set dst (OrI dst src));
8597 effect(KILL cr);
8599 format %{ "OR $dst,$src" %}
8600 opcode(0x81,0x01); /* Opcode 81 /1 id */
8601 // ins_encode( RegImm( dst, src) );
8602 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
8603 ins_pipe( ialu_reg );
8604 %}
8606 // Or Register with Memory
8607 instruct orI_eReg_mem(rRegI dst, memory src, eFlagsReg cr) %{
8608 match(Set dst (OrI dst (LoadI src)));
8609 effect(KILL cr);
8611 ins_cost(125);
8612 format %{ "OR $dst,$src" %}
8613 opcode(0x0B);
8614 ins_encode( OpcP, RegMem( dst, src) );
8615 ins_pipe( ialu_reg_mem );
8616 %}
8618 // Or Memory with Register
8619 instruct orI_mem_eReg(memory dst, rRegI src, eFlagsReg cr) %{
8620 match(Set dst (StoreI dst (OrI (LoadI dst) src)));
8621 effect(KILL cr);
8623 ins_cost(150);
8624 format %{ "OR $dst,$src" %}
8625 opcode(0x09); /* Opcode 09 /r */
8626 ins_encode( OpcP, RegMem( src, dst ) );
8627 ins_pipe( ialu_mem_reg );
8628 %}
8630 // Or Memory with Immediate
8631 instruct orI_mem_imm(memory dst, immI src, eFlagsReg cr) %{
8632 match(Set dst (StoreI dst (OrI (LoadI dst) src)));
8633 effect(KILL cr);
8635 ins_cost(125);
8636 format %{ "OR $dst,$src" %}
8637 opcode(0x81,0x1); /* Opcode 81 /1 id */
8638 // ins_encode( MemImm( dst, src) );
8639 ins_encode( OpcSE( src ), RMopc_Mem(secondary, dst ), Con8or32( src ) );
8640 ins_pipe( ialu_mem_imm );
8641 %}
8643 // ROL/ROR
8644 // ROL expand
8645 instruct rolI_eReg_imm1(rRegI dst, immI1 shift, eFlagsReg cr) %{
8646 effect(USE_DEF dst, USE shift, KILL cr);
8648 format %{ "ROL $dst, $shift" %}
8649 opcode(0xD1, 0x0); /* Opcode D1 /0 */
8650 ins_encode( OpcP, RegOpc( dst ));
8651 ins_pipe( ialu_reg );
8652 %}
8654 instruct rolI_eReg_imm8(rRegI dst, immI8 shift, eFlagsReg cr) %{
8655 effect(USE_DEF dst, USE shift, KILL cr);
8657 format %{ "ROL $dst, $shift" %}
8658 opcode(0xC1, 0x0); /*Opcode /C1 /0 */
8659 ins_encode( RegOpcImm(dst, shift) );
8660 ins_pipe(ialu_reg);
8661 %}
8663 instruct rolI_eReg_CL(ncxRegI dst, eCXRegI shift, eFlagsReg cr) %{
8664 effect(USE_DEF dst, USE shift, KILL cr);
8666 format %{ "ROL $dst, $shift" %}
8667 opcode(0xD3, 0x0); /* Opcode D3 /0 */
8668 ins_encode(OpcP, RegOpc(dst));
8669 ins_pipe( ialu_reg_reg );
8670 %}
8671 // end of ROL expand
8673 // ROL 32bit by one once
8674 instruct rolI_eReg_i1(rRegI dst, immI1 lshift, immI_M1 rshift, eFlagsReg cr) %{
8675 match(Set dst ( OrI (LShiftI dst lshift) (URShiftI dst rshift)));
8677 expand %{
8678 rolI_eReg_imm1(dst, lshift, cr);
8679 %}
8680 %}
8682 // ROL 32bit var by imm8 once
8683 instruct rolI_eReg_i8(rRegI dst, immI8 lshift, immI8 rshift, eFlagsReg cr) %{
8684 predicate( 0 == ((n->in(1)->in(2)->get_int() + n->in(2)->in(2)->get_int()) & 0x1f));
8685 match(Set dst ( OrI (LShiftI dst lshift) (URShiftI dst rshift)));
8687 expand %{
8688 rolI_eReg_imm8(dst, lshift, cr);
8689 %}
8690 %}
8692 // ROL 32bit var by var once
8693 instruct rolI_eReg_Var_C0(ncxRegI dst, eCXRegI shift, immI0 zero, eFlagsReg cr) %{
8694 match(Set dst ( OrI (LShiftI dst shift) (URShiftI dst (SubI zero shift))));
8696 expand %{
8697 rolI_eReg_CL(dst, shift, cr);
8698 %}
8699 %}
8701 // ROL 32bit var by var once
8702 instruct rolI_eReg_Var_C32(ncxRegI dst, eCXRegI shift, immI_32 c32, eFlagsReg cr) %{
8703 match(Set dst ( OrI (LShiftI dst shift) (URShiftI dst (SubI c32 shift))));
8705 expand %{
8706 rolI_eReg_CL(dst, shift, cr);
8707 %}
8708 %}
8710 // ROR expand
8711 instruct rorI_eReg_imm1(rRegI dst, immI1 shift, eFlagsReg cr) %{
8712 effect(USE_DEF dst, USE shift, KILL cr);
8714 format %{ "ROR $dst, $shift" %}
8715 opcode(0xD1,0x1); /* Opcode D1 /1 */
8716 ins_encode( OpcP, RegOpc( dst ) );
8717 ins_pipe( ialu_reg );
8718 %}
8720 instruct rorI_eReg_imm8(rRegI dst, immI8 shift, eFlagsReg cr) %{
8721 effect (USE_DEF dst, USE shift, KILL cr);
8723 format %{ "ROR $dst, $shift" %}
8724 opcode(0xC1, 0x1); /* Opcode /C1 /1 ib */
8725 ins_encode( RegOpcImm(dst, shift) );
8726 ins_pipe( ialu_reg );
8727 %}
8729 instruct rorI_eReg_CL(ncxRegI dst, eCXRegI shift, eFlagsReg cr)%{
8730 effect(USE_DEF dst, USE shift, KILL cr);
8732 format %{ "ROR $dst, $shift" %}
8733 opcode(0xD3, 0x1); /* Opcode D3 /1 */
8734 ins_encode(OpcP, RegOpc(dst));
8735 ins_pipe( ialu_reg_reg );
8736 %}
8737 // end of ROR expand
8739 // ROR right once
8740 instruct rorI_eReg_i1(rRegI dst, immI1 rshift, immI_M1 lshift, eFlagsReg cr) %{
8741 match(Set dst ( OrI (URShiftI dst rshift) (LShiftI dst lshift)));
8743 expand %{
8744 rorI_eReg_imm1(dst, rshift, cr);
8745 %}
8746 %}
8748 // ROR 32bit by immI8 once
8749 instruct rorI_eReg_i8(rRegI dst, immI8 rshift, immI8 lshift, eFlagsReg cr) %{
8750 predicate( 0 == ((n->in(1)->in(2)->get_int() + n->in(2)->in(2)->get_int()) & 0x1f));
8751 match(Set dst ( OrI (URShiftI dst rshift) (LShiftI dst lshift)));
8753 expand %{
8754 rorI_eReg_imm8(dst, rshift, cr);
8755 %}
8756 %}
8758 // ROR 32bit var by var once
8759 instruct rorI_eReg_Var_C0(ncxRegI dst, eCXRegI shift, immI0 zero, eFlagsReg cr) %{
8760 match(Set dst ( OrI (URShiftI dst shift) (LShiftI dst (SubI zero shift))));
8762 expand %{
8763 rorI_eReg_CL(dst, shift, cr);
8764 %}
8765 %}
8767 // ROR 32bit var by var once
8768 instruct rorI_eReg_Var_C32(ncxRegI dst, eCXRegI shift, immI_32 c32, eFlagsReg cr) %{
8769 match(Set dst ( OrI (URShiftI dst shift) (LShiftI dst (SubI c32 shift))));
8771 expand %{
8772 rorI_eReg_CL(dst, shift, cr);
8773 %}
8774 %}
8776 // Xor Instructions
8777 // Xor Register with Register
8778 instruct xorI_eReg(rRegI dst, rRegI src, eFlagsReg cr) %{
8779 match(Set dst (XorI dst src));
8780 effect(KILL cr);
8782 size(2);
8783 format %{ "XOR $dst,$src" %}
8784 opcode(0x33);
8785 ins_encode( OpcP, RegReg( dst, src) );
8786 ins_pipe( ialu_reg_reg );
8787 %}
8789 // Xor Register with Immediate -1
8790 instruct xorI_eReg_im1(rRegI dst, immI_M1 imm) %{
8791 match(Set dst (XorI dst imm));
8793 size(2);
8794 format %{ "NOT $dst" %}
8795 ins_encode %{
8796 __ notl($dst$$Register);
8797 %}
8798 ins_pipe( ialu_reg );
8799 %}
8801 // Xor Register with Immediate
8802 instruct xorI_eReg_imm(rRegI dst, immI src, eFlagsReg cr) %{
8803 match(Set dst (XorI dst src));
8804 effect(KILL cr);
8806 format %{ "XOR $dst,$src" %}
8807 opcode(0x81,0x06); /* Opcode 81 /6 id */
8808 // ins_encode( RegImm( dst, src) );
8809 ins_encode( OpcSErm( dst, src ), Con8or32( src ) );
8810 ins_pipe( ialu_reg );
8811 %}
8813 // Xor Register with Memory
8814 instruct xorI_eReg_mem(rRegI dst, memory src, eFlagsReg cr) %{
8815 match(Set dst (XorI dst (LoadI src)));
8816 effect(KILL cr);
8818 ins_cost(125);
8819 format %{ "XOR $dst,$src" %}
8820 opcode(0x33);
8821 ins_encode( OpcP, RegMem(dst, src) );
8822 ins_pipe( ialu_reg_mem );
8823 %}
8825 // Xor Memory with Register
8826 instruct xorI_mem_eReg(memory dst, rRegI src, eFlagsReg cr) %{
8827 match(Set dst (StoreI dst (XorI (LoadI dst) src)));
8828 effect(KILL cr);
8830 ins_cost(150);
8831 format %{ "XOR $dst,$src" %}
8832 opcode(0x31); /* Opcode 31 /r */
8833 ins_encode( OpcP, RegMem( src, dst ) );
8834 ins_pipe( ialu_mem_reg );
8835 %}
8837 // Xor Memory with Immediate
8838 instruct xorI_mem_imm(memory dst, immI src, eFlagsReg cr) %{
8839 match(Set dst (StoreI dst (XorI (LoadI dst) src)));
8840 effect(KILL cr);
8842 ins_cost(125);
8843 format %{ "XOR $dst,$src" %}
8844 opcode(0x81,0x6); /* Opcode 81 /6 id */
8845 ins_encode( OpcSE( src ), RMopc_Mem(secondary, dst ), Con8or32( src ) );
8846 ins_pipe( ialu_mem_imm );
8847 %}
8849 //----------Convert Int to Boolean---------------------------------------------
8851 instruct movI_nocopy(rRegI dst, rRegI src) %{
8852 effect( DEF dst, USE src );
8853 format %{ "MOV $dst,$src" %}
8854 ins_encode( enc_Copy( dst, src) );
8855 ins_pipe( ialu_reg_reg );
8856 %}
8858 instruct ci2b( rRegI dst, rRegI src, eFlagsReg cr ) %{
8859 effect( USE_DEF dst, USE src, KILL cr );
8861 size(4);
8862 format %{ "NEG $dst\n\t"
8863 "ADC $dst,$src" %}
8864 ins_encode( neg_reg(dst),
8865 OpcRegReg(0x13,dst,src) );
8866 ins_pipe( ialu_reg_reg_long );
8867 %}
8869 instruct convI2B( rRegI dst, rRegI src, eFlagsReg cr ) %{
8870 match(Set dst (Conv2B src));
8872 expand %{
8873 movI_nocopy(dst,src);
8874 ci2b(dst,src,cr);
8875 %}
8876 %}
8878 instruct movP_nocopy(rRegI dst, eRegP src) %{
8879 effect( DEF dst, USE src );
8880 format %{ "MOV $dst,$src" %}
8881 ins_encode( enc_Copy( dst, src) );
8882 ins_pipe( ialu_reg_reg );
8883 %}
8885 instruct cp2b( rRegI dst, eRegP src, eFlagsReg cr ) %{
8886 effect( USE_DEF dst, USE src, KILL cr );
8887 format %{ "NEG $dst\n\t"
8888 "ADC $dst,$src" %}
8889 ins_encode( neg_reg(dst),
8890 OpcRegReg(0x13,dst,src) );
8891 ins_pipe( ialu_reg_reg_long );
8892 %}
8894 instruct convP2B( rRegI dst, eRegP src, eFlagsReg cr ) %{
8895 match(Set dst (Conv2B src));
8897 expand %{
8898 movP_nocopy(dst,src);
8899 cp2b(dst,src,cr);
8900 %}
8901 %}
8903 instruct cmpLTMask( eCXRegI dst, ncxRegI p, ncxRegI q, eFlagsReg cr ) %{
8904 match(Set dst (CmpLTMask p q));
8905 effect( KILL cr );
8906 ins_cost(400);
8908 // SETlt can only use low byte of EAX,EBX, ECX, or EDX as destination
8909 format %{ "XOR $dst,$dst\n\t"
8910 "CMP $p,$q\n\t"
8911 "SETlt $dst\n\t"
8912 "NEG $dst" %}
8913 ins_encode( OpcRegReg(0x33,dst,dst),
8914 OpcRegReg(0x3B,p,q),
8915 setLT_reg(dst), neg_reg(dst) );
8916 ins_pipe( pipe_slow );
8917 %}
8919 instruct cmpLTMask0( rRegI dst, immI0 zero, eFlagsReg cr ) %{
8920 match(Set dst (CmpLTMask dst zero));
8921 effect( DEF dst, KILL cr );
8922 ins_cost(100);
8924 format %{ "SAR $dst,31" %}
8925 opcode(0xC1, 0x7); /* C1 /7 ib */
8926 ins_encode( RegOpcImm( dst, 0x1F ) );
8927 ins_pipe( ialu_reg );
8928 %}
8931 instruct cadd_cmpLTMask( ncxRegI p, ncxRegI q, ncxRegI y, eCXRegI tmp, eFlagsReg cr ) %{
8932 match(Set p (AddI (AndI (CmpLTMask p q) y) (SubI p q)));
8933 effect( KILL tmp, KILL cr );
8934 ins_cost(400);
8935 // annoyingly, $tmp has no edges so you cant ask for it in
8936 // any format or encoding
8937 format %{ "SUB $p,$q\n\t"
8938 "SBB ECX,ECX\n\t"
8939 "AND ECX,$y\n\t"
8940 "ADD $p,ECX" %}
8941 ins_encode( enc_cmpLTP(p,q,y,tmp) );
8942 ins_pipe( pipe_cmplt );
8943 %}
8945 /* If I enable this, I encourage spilling in the inner loop of compress.
8946 instruct cadd_cmpLTMask_mem( ncxRegI p, ncxRegI q, memory y, eCXRegI tmp, eFlagsReg cr ) %{
8947 match(Set p (AddI (AndI (CmpLTMask p q) (LoadI y)) (SubI p q)));
8948 effect( USE_KILL tmp, KILL cr );
8949 ins_cost(400);
8951 format %{ "SUB $p,$q\n\t"
8952 "SBB ECX,ECX\n\t"
8953 "AND ECX,$y\n\t"
8954 "ADD $p,ECX" %}
8955 ins_encode( enc_cmpLTP_mem(p,q,y,tmp) );
8956 %}
8957 */
8959 //----------Long Instructions------------------------------------------------
8960 // Add Long Register with Register
8961 instruct addL_eReg(eRegL dst, eRegL src, eFlagsReg cr) %{
8962 match(Set dst (AddL dst src));
8963 effect(KILL cr);
8964 ins_cost(200);
8965 format %{ "ADD $dst.lo,$src.lo\n\t"
8966 "ADC $dst.hi,$src.hi" %}
8967 opcode(0x03, 0x13);
8968 ins_encode( RegReg_Lo(dst, src), RegReg_Hi(dst,src) );
8969 ins_pipe( ialu_reg_reg_long );
8970 %}
8972 // Add Long Register with Immediate
8973 instruct addL_eReg_imm(eRegL dst, immL src, eFlagsReg cr) %{
8974 match(Set dst (AddL dst src));
8975 effect(KILL cr);
8976 format %{ "ADD $dst.lo,$src.lo\n\t"
8977 "ADC $dst.hi,$src.hi" %}
8978 opcode(0x81,0x00,0x02); /* Opcode 81 /0, 81 /2 */
8979 ins_encode( Long_OpcSErm_Lo( dst, src ), Long_OpcSErm_Hi( dst, src ) );
8980 ins_pipe( ialu_reg_long );
8981 %}
8983 // Add Long Register with Memory
8984 instruct addL_eReg_mem(eRegL dst, load_long_memory mem, eFlagsReg cr) %{
8985 match(Set dst (AddL dst (LoadL mem)));
8986 effect(KILL cr);
8987 ins_cost(125);
8988 format %{ "ADD $dst.lo,$mem\n\t"
8989 "ADC $dst.hi,$mem+4" %}
8990 opcode(0x03, 0x13);
8991 ins_encode( OpcP, RegMem( dst, mem), OpcS, RegMem_Hi(dst,mem) );
8992 ins_pipe( ialu_reg_long_mem );
8993 %}
8995 // Subtract Long Register with Register.
8996 instruct subL_eReg(eRegL dst, eRegL src, eFlagsReg cr) %{
8997 match(Set dst (SubL dst src));
8998 effect(KILL cr);
8999 ins_cost(200);
9000 format %{ "SUB $dst.lo,$src.lo\n\t"
9001 "SBB $dst.hi,$src.hi" %}
9002 opcode(0x2B, 0x1B);
9003 ins_encode( RegReg_Lo(dst, src), RegReg_Hi(dst,src) );
9004 ins_pipe( ialu_reg_reg_long );
9005 %}
9007 // Subtract Long Register with Immediate
9008 instruct subL_eReg_imm(eRegL dst, immL src, eFlagsReg cr) %{
9009 match(Set dst (SubL dst src));
9010 effect(KILL cr);
9011 format %{ "SUB $dst.lo,$src.lo\n\t"
9012 "SBB $dst.hi,$src.hi" %}
9013 opcode(0x81,0x05,0x03); /* Opcode 81 /5, 81 /3 */
9014 ins_encode( Long_OpcSErm_Lo( dst, src ), Long_OpcSErm_Hi( dst, src ) );
9015 ins_pipe( ialu_reg_long );
9016 %}
9018 // Subtract Long Register with Memory
9019 instruct subL_eReg_mem(eRegL dst, load_long_memory mem, eFlagsReg cr) %{
9020 match(Set dst (SubL dst (LoadL mem)));
9021 effect(KILL cr);
9022 ins_cost(125);
9023 format %{ "SUB $dst.lo,$mem\n\t"
9024 "SBB $dst.hi,$mem+4" %}
9025 opcode(0x2B, 0x1B);
9026 ins_encode( OpcP, RegMem( dst, mem), OpcS, RegMem_Hi(dst,mem) );
9027 ins_pipe( ialu_reg_long_mem );
9028 %}
9030 instruct negL_eReg(eRegL dst, immL0 zero, eFlagsReg cr) %{
9031 match(Set dst (SubL zero dst));
9032 effect(KILL cr);
9033 ins_cost(300);
9034 format %{ "NEG $dst.hi\n\tNEG $dst.lo\n\tSBB $dst.hi,0" %}
9035 ins_encode( neg_long(dst) );
9036 ins_pipe( ialu_reg_reg_long );
9037 %}
9039 // And Long Register with Register
9040 instruct andL_eReg(eRegL dst, eRegL src, eFlagsReg cr) %{
9041 match(Set dst (AndL dst src));
9042 effect(KILL cr);
9043 format %{ "AND $dst.lo,$src.lo\n\t"
9044 "AND $dst.hi,$src.hi" %}
9045 opcode(0x23,0x23);
9046 ins_encode( RegReg_Lo( dst, src), RegReg_Hi( dst, src) );
9047 ins_pipe( ialu_reg_reg_long );
9048 %}
9050 // And Long Register with Immediate
9051 instruct andL_eReg_imm(eRegL dst, immL src, eFlagsReg cr) %{
9052 match(Set dst (AndL dst src));
9053 effect(KILL cr);
9054 format %{ "AND $dst.lo,$src.lo\n\t"
9055 "AND $dst.hi,$src.hi" %}
9056 opcode(0x81,0x04,0x04); /* Opcode 81 /4, 81 /4 */
9057 ins_encode( Long_OpcSErm_Lo( dst, src ), Long_OpcSErm_Hi( dst, src ) );
9058 ins_pipe( ialu_reg_long );
9059 %}
9061 // And Long Register with Memory
9062 instruct andL_eReg_mem(eRegL dst, load_long_memory mem, eFlagsReg cr) %{
9063 match(Set dst (AndL dst (LoadL mem)));
9064 effect(KILL cr);
9065 ins_cost(125);
9066 format %{ "AND $dst.lo,$mem\n\t"
9067 "AND $dst.hi,$mem+4" %}
9068 opcode(0x23, 0x23);
9069 ins_encode( OpcP, RegMem( dst, mem), OpcS, RegMem_Hi(dst,mem) );
9070 ins_pipe( ialu_reg_long_mem );
9071 %}
9073 // Or Long Register with Register
9074 instruct orl_eReg(eRegL dst, eRegL src, eFlagsReg cr) %{
9075 match(Set dst (OrL dst src));
9076 effect(KILL cr);
9077 format %{ "OR $dst.lo,$src.lo\n\t"
9078 "OR $dst.hi,$src.hi" %}
9079 opcode(0x0B,0x0B);
9080 ins_encode( RegReg_Lo( dst, src), RegReg_Hi( dst, src) );
9081 ins_pipe( ialu_reg_reg_long );
9082 %}
9084 // Or Long Register with Immediate
9085 instruct orl_eReg_imm(eRegL dst, immL src, eFlagsReg cr) %{
9086 match(Set dst (OrL dst src));
9087 effect(KILL cr);
9088 format %{ "OR $dst.lo,$src.lo\n\t"
9089 "OR $dst.hi,$src.hi" %}
9090 opcode(0x81,0x01,0x01); /* Opcode 81 /1, 81 /1 */
9091 ins_encode( Long_OpcSErm_Lo( dst, src ), Long_OpcSErm_Hi( dst, src ) );
9092 ins_pipe( ialu_reg_long );
9093 %}
9095 // Or Long Register with Memory
9096 instruct orl_eReg_mem(eRegL dst, load_long_memory mem, eFlagsReg cr) %{
9097 match(Set dst (OrL dst (LoadL mem)));
9098 effect(KILL cr);
9099 ins_cost(125);
9100 format %{ "OR $dst.lo,$mem\n\t"
9101 "OR $dst.hi,$mem+4" %}
9102 opcode(0x0B,0x0B);
9103 ins_encode( OpcP, RegMem( dst, mem), OpcS, RegMem_Hi(dst,mem) );
9104 ins_pipe( ialu_reg_long_mem );
9105 %}
9107 // Xor Long Register with Register
9108 instruct xorl_eReg(eRegL dst, eRegL src, eFlagsReg cr) %{
9109 match(Set dst (XorL dst src));
9110 effect(KILL cr);
9111 format %{ "XOR $dst.lo,$src.lo\n\t"
9112 "XOR $dst.hi,$src.hi" %}
9113 opcode(0x33,0x33);
9114 ins_encode( RegReg_Lo( dst, src), RegReg_Hi( dst, src) );
9115 ins_pipe( ialu_reg_reg_long );
9116 %}
9118 // Xor Long Register with Immediate -1
9119 instruct xorl_eReg_im1(eRegL dst, immL_M1 imm) %{
9120 match(Set dst (XorL dst imm));
9121 format %{ "NOT $dst.lo\n\t"
9122 "NOT $dst.hi" %}
9123 ins_encode %{
9124 __ notl($dst$$Register);
9125 __ notl(HIGH_FROM_LOW($dst$$Register));
9126 %}
9127 ins_pipe( ialu_reg_long );
9128 %}
9130 // Xor Long Register with Immediate
9131 instruct xorl_eReg_imm(eRegL dst, immL src, eFlagsReg cr) %{
9132 match(Set dst (XorL dst src));
9133 effect(KILL cr);
9134 format %{ "XOR $dst.lo,$src.lo\n\t"
9135 "XOR $dst.hi,$src.hi" %}
9136 opcode(0x81,0x06,0x06); /* Opcode 81 /6, 81 /6 */
9137 ins_encode( Long_OpcSErm_Lo( dst, src ), Long_OpcSErm_Hi( dst, src ) );
9138 ins_pipe( ialu_reg_long );
9139 %}
9141 // Xor Long Register with Memory
9142 instruct xorl_eReg_mem(eRegL dst, load_long_memory mem, eFlagsReg cr) %{
9143 match(Set dst (XorL dst (LoadL mem)));
9144 effect(KILL cr);
9145 ins_cost(125);
9146 format %{ "XOR $dst.lo,$mem\n\t"
9147 "XOR $dst.hi,$mem+4" %}
9148 opcode(0x33,0x33);
9149 ins_encode( OpcP, RegMem( dst, mem), OpcS, RegMem_Hi(dst,mem) );
9150 ins_pipe( ialu_reg_long_mem );
9151 %}
9153 // Shift Left Long by 1
9154 instruct shlL_eReg_1(eRegL dst, immI_1 cnt, eFlagsReg cr) %{
9155 predicate(UseNewLongLShift);
9156 match(Set dst (LShiftL dst cnt));
9157 effect(KILL cr);
9158 ins_cost(100);
9159 format %{ "ADD $dst.lo,$dst.lo\n\t"
9160 "ADC $dst.hi,$dst.hi" %}
9161 ins_encode %{
9162 __ addl($dst$$Register,$dst$$Register);
9163 __ adcl(HIGH_FROM_LOW($dst$$Register),HIGH_FROM_LOW($dst$$Register));
9164 %}
9165 ins_pipe( ialu_reg_long );
9166 %}
9168 // Shift Left Long by 2
9169 instruct shlL_eReg_2(eRegL dst, immI_2 cnt, eFlagsReg cr) %{
9170 predicate(UseNewLongLShift);
9171 match(Set dst (LShiftL dst cnt));
9172 effect(KILL cr);
9173 ins_cost(100);
9174 format %{ "ADD $dst.lo,$dst.lo\n\t"
9175 "ADC $dst.hi,$dst.hi\n\t"
9176 "ADD $dst.lo,$dst.lo\n\t"
9177 "ADC $dst.hi,$dst.hi" %}
9178 ins_encode %{
9179 __ addl($dst$$Register,$dst$$Register);
9180 __ adcl(HIGH_FROM_LOW($dst$$Register),HIGH_FROM_LOW($dst$$Register));
9181 __ addl($dst$$Register,$dst$$Register);
9182 __ adcl(HIGH_FROM_LOW($dst$$Register),HIGH_FROM_LOW($dst$$Register));
9183 %}
9184 ins_pipe( ialu_reg_long );
9185 %}
9187 // Shift Left Long by 3
9188 instruct shlL_eReg_3(eRegL dst, immI_3 cnt, eFlagsReg cr) %{
9189 predicate(UseNewLongLShift);
9190 match(Set dst (LShiftL dst cnt));
9191 effect(KILL cr);
9192 ins_cost(100);
9193 format %{ "ADD $dst.lo,$dst.lo\n\t"
9194 "ADC $dst.hi,$dst.hi\n\t"
9195 "ADD $dst.lo,$dst.lo\n\t"
9196 "ADC $dst.hi,$dst.hi\n\t"
9197 "ADD $dst.lo,$dst.lo\n\t"
9198 "ADC $dst.hi,$dst.hi" %}
9199 ins_encode %{
9200 __ addl($dst$$Register,$dst$$Register);
9201 __ adcl(HIGH_FROM_LOW($dst$$Register),HIGH_FROM_LOW($dst$$Register));
9202 __ addl($dst$$Register,$dst$$Register);
9203 __ adcl(HIGH_FROM_LOW($dst$$Register),HIGH_FROM_LOW($dst$$Register));
9204 __ addl($dst$$Register,$dst$$Register);
9205 __ adcl(HIGH_FROM_LOW($dst$$Register),HIGH_FROM_LOW($dst$$Register));
9206 %}
9207 ins_pipe( ialu_reg_long );
9208 %}
9210 // Shift Left Long by 1-31
9211 instruct shlL_eReg_1_31(eRegL dst, immI_1_31 cnt, eFlagsReg cr) %{
9212 match(Set dst (LShiftL dst cnt));
9213 effect(KILL cr);
9214 ins_cost(200);
9215 format %{ "SHLD $dst.hi,$dst.lo,$cnt\n\t"
9216 "SHL $dst.lo,$cnt" %}
9217 opcode(0xC1, 0x4, 0xA4); /* 0F/A4, then C1 /4 ib */
9218 ins_encode( move_long_small_shift(dst,cnt) );
9219 ins_pipe( ialu_reg_long );
9220 %}
9222 // Shift Left Long by 32-63
9223 instruct shlL_eReg_32_63(eRegL dst, immI_32_63 cnt, eFlagsReg cr) %{
9224 match(Set dst (LShiftL dst cnt));
9225 effect(KILL cr);
9226 ins_cost(300);
9227 format %{ "MOV $dst.hi,$dst.lo\n"
9228 "\tSHL $dst.hi,$cnt-32\n"
9229 "\tXOR $dst.lo,$dst.lo" %}
9230 opcode(0xC1, 0x4); /* C1 /4 ib */
9231 ins_encode( move_long_big_shift_clr(dst,cnt) );
9232 ins_pipe( ialu_reg_long );
9233 %}
9235 // Shift Left Long by variable
9236 instruct salL_eReg_CL(eRegL dst, eCXRegI shift, eFlagsReg cr) %{
9237 match(Set dst (LShiftL dst shift));
9238 effect(KILL cr);
9239 ins_cost(500+200);
9240 size(17);
9241 format %{ "TEST $shift,32\n\t"
9242 "JEQ,s small\n\t"
9243 "MOV $dst.hi,$dst.lo\n\t"
9244 "XOR $dst.lo,$dst.lo\n"
9245 "small:\tSHLD $dst.hi,$dst.lo,$shift\n\t"
9246 "SHL $dst.lo,$shift" %}
9247 ins_encode( shift_left_long( dst, shift ) );
9248 ins_pipe( pipe_slow );
9249 %}
9251 // Shift Right Long by 1-31
9252 instruct shrL_eReg_1_31(eRegL dst, immI_1_31 cnt, eFlagsReg cr) %{
9253 match(Set dst (URShiftL dst cnt));
9254 effect(KILL cr);
9255 ins_cost(200);
9256 format %{ "SHRD $dst.lo,$dst.hi,$cnt\n\t"
9257 "SHR $dst.hi,$cnt" %}
9258 opcode(0xC1, 0x5, 0xAC); /* 0F/AC, then C1 /5 ib */
9259 ins_encode( move_long_small_shift(dst,cnt) );
9260 ins_pipe( ialu_reg_long );
9261 %}
9263 // Shift Right Long by 32-63
9264 instruct shrL_eReg_32_63(eRegL dst, immI_32_63 cnt, eFlagsReg cr) %{
9265 match(Set dst (URShiftL dst cnt));
9266 effect(KILL cr);
9267 ins_cost(300);
9268 format %{ "MOV $dst.lo,$dst.hi\n"
9269 "\tSHR $dst.lo,$cnt-32\n"
9270 "\tXOR $dst.hi,$dst.hi" %}
9271 opcode(0xC1, 0x5); /* C1 /5 ib */
9272 ins_encode( move_long_big_shift_clr(dst,cnt) );
9273 ins_pipe( ialu_reg_long );
9274 %}
9276 // Shift Right Long by variable
9277 instruct shrL_eReg_CL(eRegL dst, eCXRegI shift, eFlagsReg cr) %{
9278 match(Set dst (URShiftL dst shift));
9279 effect(KILL cr);
9280 ins_cost(600);
9281 size(17);
9282 format %{ "TEST $shift,32\n\t"
9283 "JEQ,s small\n\t"
9284 "MOV $dst.lo,$dst.hi\n\t"
9285 "XOR $dst.hi,$dst.hi\n"
9286 "small:\tSHRD $dst.lo,$dst.hi,$shift\n\t"
9287 "SHR $dst.hi,$shift" %}
9288 ins_encode( shift_right_long( dst, shift ) );
9289 ins_pipe( pipe_slow );
9290 %}
9292 // Shift Right Long by 1-31
9293 instruct sarL_eReg_1_31(eRegL dst, immI_1_31 cnt, eFlagsReg cr) %{
9294 match(Set dst (RShiftL dst cnt));
9295 effect(KILL cr);
9296 ins_cost(200);
9297 format %{ "SHRD $dst.lo,$dst.hi,$cnt\n\t"
9298 "SAR $dst.hi,$cnt" %}
9299 opcode(0xC1, 0x7, 0xAC); /* 0F/AC, then C1 /7 ib */
9300 ins_encode( move_long_small_shift(dst,cnt) );
9301 ins_pipe( ialu_reg_long );
9302 %}
9304 // Shift Right Long by 32-63
9305 instruct sarL_eReg_32_63( eRegL dst, immI_32_63 cnt, eFlagsReg cr) %{
9306 match(Set dst (RShiftL dst cnt));
9307 effect(KILL cr);
9308 ins_cost(300);
9309 format %{ "MOV $dst.lo,$dst.hi\n"
9310 "\tSAR $dst.lo,$cnt-32\n"
9311 "\tSAR $dst.hi,31" %}
9312 opcode(0xC1, 0x7); /* C1 /7 ib */
9313 ins_encode( move_long_big_shift_sign(dst,cnt) );
9314 ins_pipe( ialu_reg_long );
9315 %}
9317 // Shift Right arithmetic Long by variable
9318 instruct sarL_eReg_CL(eRegL dst, eCXRegI shift, eFlagsReg cr) %{
9319 match(Set dst (RShiftL dst shift));
9320 effect(KILL cr);
9321 ins_cost(600);
9322 size(18);
9323 format %{ "TEST $shift,32\n\t"
9324 "JEQ,s small\n\t"
9325 "MOV $dst.lo,$dst.hi\n\t"
9326 "SAR $dst.hi,31\n"
9327 "small:\tSHRD $dst.lo,$dst.hi,$shift\n\t"
9328 "SAR $dst.hi,$shift" %}
9329 ins_encode( shift_right_arith_long( dst, shift ) );
9330 ins_pipe( pipe_slow );
9331 %}
9334 //----------Double Instructions------------------------------------------------
9335 // Double Math
9337 // Compare & branch
9339 // P6 version of float compare, sets condition codes in EFLAGS
9340 instruct cmpDPR_cc_P6(eFlagsRegU cr, regDPR src1, regDPR src2, eAXRegI rax) %{
9341 predicate(VM_Version::supports_cmov() && UseSSE <=1);
9342 match(Set cr (CmpD src1 src2));
9343 effect(KILL rax);
9344 ins_cost(150);
9345 format %{ "FLD $src1\n\t"
9346 "FUCOMIP ST,$src2 // P6 instruction\n\t"
9347 "JNP exit\n\t"
9348 "MOV ah,1 // saw a NaN, set CF\n\t"
9349 "SAHF\n"
9350 "exit:\tNOP // avoid branch to branch" %}
9351 opcode(0xDF, 0x05); /* DF E8+i or DF /5 */
9352 ins_encode( Push_Reg_DPR(src1),
9353 OpcP, RegOpc(src2),
9354 cmpF_P6_fixup );
9355 ins_pipe( pipe_slow );
9356 %}
9358 instruct cmpDPR_cc_P6CF(eFlagsRegUCF cr, regDPR src1, regDPR src2) %{
9359 predicate(VM_Version::supports_cmov() && UseSSE <=1);
9360 match(Set cr (CmpD src1 src2));
9361 ins_cost(150);
9362 format %{ "FLD $src1\n\t"
9363 "FUCOMIP ST,$src2 // P6 instruction" %}
9364 opcode(0xDF, 0x05); /* DF E8+i or DF /5 */
9365 ins_encode( Push_Reg_DPR(src1),
9366 OpcP, RegOpc(src2));
9367 ins_pipe( pipe_slow );
9368 %}
9370 // Compare & branch
9371 instruct cmpDPR_cc(eFlagsRegU cr, regDPR src1, regDPR src2, eAXRegI rax) %{
9372 predicate(UseSSE<=1);
9373 match(Set cr (CmpD src1 src2));
9374 effect(KILL rax);
9375 ins_cost(200);
9376 format %{ "FLD $src1\n\t"
9377 "FCOMp $src2\n\t"
9378 "FNSTSW AX\n\t"
9379 "TEST AX,0x400\n\t"
9380 "JZ,s flags\n\t"
9381 "MOV AH,1\t# unordered treat as LT\n"
9382 "flags:\tSAHF" %}
9383 opcode(0xD8, 0x3); /* D8 D8+i or D8 /3 */
9384 ins_encode( Push_Reg_DPR(src1),
9385 OpcP, RegOpc(src2),
9386 fpu_flags);
9387 ins_pipe( pipe_slow );
9388 %}
9390 // Compare vs zero into -1,0,1
9391 instruct cmpDPR_0(rRegI dst, regDPR src1, immDPR0 zero, eAXRegI rax, eFlagsReg cr) %{
9392 predicate(UseSSE<=1);
9393 match(Set dst (CmpD3 src1 zero));
9394 effect(KILL cr, KILL rax);
9395 ins_cost(280);
9396 format %{ "FTSTD $dst,$src1" %}
9397 opcode(0xE4, 0xD9);
9398 ins_encode( Push_Reg_DPR(src1),
9399 OpcS, OpcP, PopFPU,
9400 CmpF_Result(dst));
9401 ins_pipe( pipe_slow );
9402 %}
9404 // Compare into -1,0,1
9405 instruct cmpDPR_reg(rRegI dst, regDPR src1, regDPR src2, eAXRegI rax, eFlagsReg cr) %{
9406 predicate(UseSSE<=1);
9407 match(Set dst (CmpD3 src1 src2));
9408 effect(KILL cr, KILL rax);
9409 ins_cost(300);
9410 format %{ "FCMPD $dst,$src1,$src2" %}
9411 opcode(0xD8, 0x3); /* D8 D8+i or D8 /3 */
9412 ins_encode( Push_Reg_DPR(src1),
9413 OpcP, RegOpc(src2),
9414 CmpF_Result(dst));
9415 ins_pipe( pipe_slow );
9416 %}
9418 // float compare and set condition codes in EFLAGS by XMM regs
9419 instruct cmpD_cc(eFlagsRegU cr, regD src1, regD src2) %{
9420 predicate(UseSSE>=2);
9421 match(Set cr (CmpD src1 src2));
9422 ins_cost(145);
9423 format %{ "UCOMISD $src1,$src2\n\t"
9424 "JNP,s exit\n\t"
9425 "PUSHF\t# saw NaN, set CF\n\t"
9426 "AND [rsp], #0xffffff2b\n\t"
9427 "POPF\n"
9428 "exit:" %}
9429 ins_encode %{
9430 __ ucomisd($src1$$XMMRegister, $src2$$XMMRegister);
9431 emit_cmpfp_fixup(_masm);
9432 %}
9433 ins_pipe( pipe_slow );
9434 %}
9436 instruct cmpD_ccCF(eFlagsRegUCF cr, regD src1, regD src2) %{
9437 predicate(UseSSE>=2);
9438 match(Set cr (CmpD src1 src2));
9439 ins_cost(100);
9440 format %{ "UCOMISD $src1,$src2" %}
9441 ins_encode %{
9442 __ ucomisd($src1$$XMMRegister, $src2$$XMMRegister);
9443 %}
9444 ins_pipe( pipe_slow );
9445 %}
9447 // float compare and set condition codes in EFLAGS by XMM regs
9448 instruct cmpD_ccmem(eFlagsRegU cr, regD src1, memory src2) %{
9449 predicate(UseSSE>=2);
9450 match(Set cr (CmpD src1 (LoadD src2)));
9451 ins_cost(145);
9452 format %{ "UCOMISD $src1,$src2\n\t"
9453 "JNP,s exit\n\t"
9454 "PUSHF\t# saw NaN, set CF\n\t"
9455 "AND [rsp], #0xffffff2b\n\t"
9456 "POPF\n"
9457 "exit:" %}
9458 ins_encode %{
9459 __ ucomisd($src1$$XMMRegister, $src2$$Address);
9460 emit_cmpfp_fixup(_masm);
9461 %}
9462 ins_pipe( pipe_slow );
9463 %}
9465 instruct cmpD_ccmemCF(eFlagsRegUCF cr, regD src1, memory src2) %{
9466 predicate(UseSSE>=2);
9467 match(Set cr (CmpD src1 (LoadD src2)));
9468 ins_cost(100);
9469 format %{ "UCOMISD $src1,$src2" %}
9470 ins_encode %{
9471 __ ucomisd($src1$$XMMRegister, $src2$$Address);
9472 %}
9473 ins_pipe( pipe_slow );
9474 %}
9476 // Compare into -1,0,1 in XMM
9477 instruct cmpD_reg(xRegI dst, regD src1, regD src2, eFlagsReg cr) %{
9478 predicate(UseSSE>=2);
9479 match(Set dst (CmpD3 src1 src2));
9480 effect(KILL cr);
9481 ins_cost(255);
9482 format %{ "UCOMISD $src1, $src2\n\t"
9483 "MOV $dst, #-1\n\t"
9484 "JP,s done\n\t"
9485 "JB,s done\n\t"
9486 "SETNE $dst\n\t"
9487 "MOVZB $dst, $dst\n"
9488 "done:" %}
9489 ins_encode %{
9490 __ ucomisd($src1$$XMMRegister, $src2$$XMMRegister);
9491 emit_cmpfp3(_masm, $dst$$Register);
9492 %}
9493 ins_pipe( pipe_slow );
9494 %}
9496 // Compare into -1,0,1 in XMM and memory
9497 instruct cmpD_regmem(xRegI dst, regD src1, memory src2, eFlagsReg cr) %{
9498 predicate(UseSSE>=2);
9499 match(Set dst (CmpD3 src1 (LoadD src2)));
9500 effect(KILL cr);
9501 ins_cost(275);
9502 format %{ "UCOMISD $src1, $src2\n\t"
9503 "MOV $dst, #-1\n\t"
9504 "JP,s done\n\t"
9505 "JB,s done\n\t"
9506 "SETNE $dst\n\t"
9507 "MOVZB $dst, $dst\n"
9508 "done:" %}
9509 ins_encode %{
9510 __ ucomisd($src1$$XMMRegister, $src2$$Address);
9511 emit_cmpfp3(_masm, $dst$$Register);
9512 %}
9513 ins_pipe( pipe_slow );
9514 %}
9517 instruct subDPR_reg(regDPR dst, regDPR src) %{
9518 predicate (UseSSE <=1);
9519 match(Set dst (SubD dst src));
9521 format %{ "FLD $src\n\t"
9522 "DSUBp $dst,ST" %}
9523 opcode(0xDE, 0x5); /* DE E8+i or DE /5 */
9524 ins_cost(150);
9525 ins_encode( Push_Reg_DPR(src),
9526 OpcP, RegOpc(dst) );
9527 ins_pipe( fpu_reg_reg );
9528 %}
9530 instruct subDPR_reg_round(stackSlotD dst, regDPR src1, regDPR src2) %{
9531 predicate (UseSSE <=1);
9532 match(Set dst (RoundDouble (SubD src1 src2)));
9533 ins_cost(250);
9535 format %{ "FLD $src2\n\t"
9536 "DSUB ST,$src1\n\t"
9537 "FSTP_D $dst\t# D-round" %}
9538 opcode(0xD8, 0x5);
9539 ins_encode( Push_Reg_DPR(src2),
9540 OpcP, RegOpc(src1), Pop_Mem_DPR(dst) );
9541 ins_pipe( fpu_mem_reg_reg );
9542 %}
9545 instruct subDPR_reg_mem(regDPR dst, memory src) %{
9546 predicate (UseSSE <=1);
9547 match(Set dst (SubD dst (LoadD src)));
9548 ins_cost(150);
9550 format %{ "FLD $src\n\t"
9551 "DSUBp $dst,ST" %}
9552 opcode(0xDE, 0x5, 0xDD); /* DE C0+i */ /* LoadD DD /0 */
9553 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src),
9554 OpcP, RegOpc(dst) );
9555 ins_pipe( fpu_reg_mem );
9556 %}
9558 instruct absDPR_reg(regDPR1 dst, regDPR1 src) %{
9559 predicate (UseSSE<=1);
9560 match(Set dst (AbsD src));
9561 ins_cost(100);
9562 format %{ "FABS" %}
9563 opcode(0xE1, 0xD9);
9564 ins_encode( OpcS, OpcP );
9565 ins_pipe( fpu_reg_reg );
9566 %}
9568 instruct negDPR_reg(regDPR1 dst, regDPR1 src) %{
9569 predicate(UseSSE<=1);
9570 match(Set dst (NegD src));
9571 ins_cost(100);
9572 format %{ "FCHS" %}
9573 opcode(0xE0, 0xD9);
9574 ins_encode( OpcS, OpcP );
9575 ins_pipe( fpu_reg_reg );
9576 %}
9578 instruct addDPR_reg(regDPR dst, regDPR src) %{
9579 predicate(UseSSE<=1);
9580 match(Set dst (AddD dst src));
9581 format %{ "FLD $src\n\t"
9582 "DADD $dst,ST" %}
9583 size(4);
9584 ins_cost(150);
9585 opcode(0xDE, 0x0); /* DE C0+i or DE /0*/
9586 ins_encode( Push_Reg_DPR(src),
9587 OpcP, RegOpc(dst) );
9588 ins_pipe( fpu_reg_reg );
9589 %}
9592 instruct addDPR_reg_round(stackSlotD dst, regDPR src1, regDPR src2) %{
9593 predicate(UseSSE<=1);
9594 match(Set dst (RoundDouble (AddD src1 src2)));
9595 ins_cost(250);
9597 format %{ "FLD $src2\n\t"
9598 "DADD ST,$src1\n\t"
9599 "FSTP_D $dst\t# D-round" %}
9600 opcode(0xD8, 0x0); /* D8 C0+i or D8 /0*/
9601 ins_encode( Push_Reg_DPR(src2),
9602 OpcP, RegOpc(src1), Pop_Mem_DPR(dst) );
9603 ins_pipe( fpu_mem_reg_reg );
9604 %}
9607 instruct addDPR_reg_mem(regDPR dst, memory src) %{
9608 predicate(UseSSE<=1);
9609 match(Set dst (AddD dst (LoadD src)));
9610 ins_cost(150);
9612 format %{ "FLD $src\n\t"
9613 "DADDp $dst,ST" %}
9614 opcode(0xDE, 0x0, 0xDD); /* DE C0+i */ /* LoadD DD /0 */
9615 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src),
9616 OpcP, RegOpc(dst) );
9617 ins_pipe( fpu_reg_mem );
9618 %}
9620 // add-to-memory
9621 instruct addDPR_mem_reg(memory dst, regDPR src) %{
9622 predicate(UseSSE<=1);
9623 match(Set dst (StoreD dst (RoundDouble (AddD (LoadD dst) src))));
9624 ins_cost(150);
9626 format %{ "FLD_D $dst\n\t"
9627 "DADD ST,$src\n\t"
9628 "FST_D $dst" %}
9629 opcode(0xDD, 0x0);
9630 ins_encode( Opcode(0xDD), RMopc_Mem(0x00,dst),
9631 Opcode(0xD8), RegOpc(src),
9632 set_instruction_start,
9633 Opcode(0xDD), RMopc_Mem(0x03,dst) );
9634 ins_pipe( fpu_reg_mem );
9635 %}
9637 instruct addDPR_reg_imm1(regDPR dst, immDPR1 con) %{
9638 predicate(UseSSE<=1);
9639 match(Set dst (AddD dst con));
9640 ins_cost(125);
9641 format %{ "FLD1\n\t"
9642 "DADDp $dst,ST" %}
9643 ins_encode %{
9644 __ fld1();
9645 __ faddp($dst$$reg);
9646 %}
9647 ins_pipe(fpu_reg);
9648 %}
9650 instruct addDPR_reg_imm(regDPR dst, immDPR con) %{
9651 predicate(UseSSE<=1 && _kids[1]->_leaf->getd() != 0.0 && _kids[1]->_leaf->getd() != 1.0 );
9652 match(Set dst (AddD dst con));
9653 ins_cost(200);
9654 format %{ "FLD_D [$constantaddress]\t# load from constant table: double=$con\n\t"
9655 "DADDp $dst,ST" %}
9656 ins_encode %{
9657 __ fld_d($constantaddress($con));
9658 __ faddp($dst$$reg);
9659 %}
9660 ins_pipe(fpu_reg_mem);
9661 %}
9663 instruct addDPR_reg_imm_round(stackSlotD dst, regDPR src, immDPR con) %{
9664 predicate(UseSSE<=1 && _kids[0]->_kids[1]->_leaf->getd() != 0.0 && _kids[0]->_kids[1]->_leaf->getd() != 1.0 );
9665 match(Set dst (RoundDouble (AddD src con)));
9666 ins_cost(200);
9667 format %{ "FLD_D [$constantaddress]\t# load from constant table: double=$con\n\t"
9668 "DADD ST,$src\n\t"
9669 "FSTP_D $dst\t# D-round" %}
9670 ins_encode %{
9671 __ fld_d($constantaddress($con));
9672 __ fadd($src$$reg);
9673 __ fstp_d(Address(rsp, $dst$$disp));
9674 %}
9675 ins_pipe(fpu_mem_reg_con);
9676 %}
9678 instruct mulDPR_reg(regDPR dst, regDPR src) %{
9679 predicate(UseSSE<=1);
9680 match(Set dst (MulD dst src));
9681 format %{ "FLD $src\n\t"
9682 "DMULp $dst,ST" %}
9683 opcode(0xDE, 0x1); /* DE C8+i or DE /1*/
9684 ins_cost(150);
9685 ins_encode( Push_Reg_DPR(src),
9686 OpcP, RegOpc(dst) );
9687 ins_pipe( fpu_reg_reg );
9688 %}
9690 // Strict FP instruction biases argument before multiply then
9691 // biases result to avoid double rounding of subnormals.
9692 //
9693 // scale arg1 by multiplying arg1 by 2^(-15360)
9694 // load arg2
9695 // multiply scaled arg1 by arg2
9696 // rescale product by 2^(15360)
9697 //
9698 instruct strictfp_mulDPR_reg(regDPR1 dst, regnotDPR1 src) %{
9699 predicate( UseSSE<=1 && Compile::current()->has_method() && Compile::current()->method()->is_strict() );
9700 match(Set dst (MulD dst src));
9701 ins_cost(1); // Select this instruction for all strict FP double multiplies
9703 format %{ "FLD StubRoutines::_fpu_subnormal_bias1\n\t"
9704 "DMULp $dst,ST\n\t"
9705 "FLD $src\n\t"
9706 "DMULp $dst,ST\n\t"
9707 "FLD StubRoutines::_fpu_subnormal_bias2\n\t"
9708 "DMULp $dst,ST\n\t" %}
9709 opcode(0xDE, 0x1); /* DE C8+i or DE /1*/
9710 ins_encode( strictfp_bias1(dst),
9711 Push_Reg_DPR(src),
9712 OpcP, RegOpc(dst),
9713 strictfp_bias2(dst) );
9714 ins_pipe( fpu_reg_reg );
9715 %}
9717 instruct mulDPR_reg_imm(regDPR dst, immDPR con) %{
9718 predicate( UseSSE<=1 && _kids[1]->_leaf->getd() != 0.0 && _kids[1]->_leaf->getd() != 1.0 );
9719 match(Set dst (MulD dst con));
9720 ins_cost(200);
9721 format %{ "FLD_D [$constantaddress]\t# load from constant table: double=$con\n\t"
9722 "DMULp $dst,ST" %}
9723 ins_encode %{
9724 __ fld_d($constantaddress($con));
9725 __ fmulp($dst$$reg);
9726 %}
9727 ins_pipe(fpu_reg_mem);
9728 %}
9731 instruct mulDPR_reg_mem(regDPR dst, memory src) %{
9732 predicate( UseSSE<=1 );
9733 match(Set dst (MulD dst (LoadD src)));
9734 ins_cost(200);
9735 format %{ "FLD_D $src\n\t"
9736 "DMULp $dst,ST" %}
9737 opcode(0xDE, 0x1, 0xDD); /* DE C8+i or DE /1*/ /* LoadD DD /0 */
9738 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src),
9739 OpcP, RegOpc(dst) );
9740 ins_pipe( fpu_reg_mem );
9741 %}
9743 //
9744 // Cisc-alternate to reg-reg multiply
9745 instruct mulDPR_reg_mem_cisc(regDPR dst, regDPR src, memory mem) %{
9746 predicate( UseSSE<=1 );
9747 match(Set dst (MulD src (LoadD mem)));
9748 ins_cost(250);
9749 format %{ "FLD_D $mem\n\t"
9750 "DMUL ST,$src\n\t"
9751 "FSTP_D $dst" %}
9752 opcode(0xD8, 0x1, 0xD9); /* D8 C8+i */ /* LoadD D9 /0 */
9753 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,mem),
9754 OpcReg_FPR(src),
9755 Pop_Reg_DPR(dst) );
9756 ins_pipe( fpu_reg_reg_mem );
9757 %}
9760 // MACRO3 -- addDPR a mulDPR
9761 // This instruction is a '2-address' instruction in that the result goes
9762 // back to src2. This eliminates a move from the macro; possibly the
9763 // register allocator will have to add it back (and maybe not).
9764 instruct addDPR_mulDPR_reg(regDPR src2, regDPR src1, regDPR src0) %{
9765 predicate( UseSSE<=1 );
9766 match(Set src2 (AddD (MulD src0 src1) src2));
9767 format %{ "FLD $src0\t# ===MACRO3d===\n\t"
9768 "DMUL ST,$src1\n\t"
9769 "DADDp $src2,ST" %}
9770 ins_cost(250);
9771 opcode(0xDD); /* LoadD DD /0 */
9772 ins_encode( Push_Reg_FPR(src0),
9773 FMul_ST_reg(src1),
9774 FAddP_reg_ST(src2) );
9775 ins_pipe( fpu_reg_reg_reg );
9776 %}
9779 // MACRO3 -- subDPR a mulDPR
9780 instruct subDPR_mulDPR_reg(regDPR src2, regDPR src1, regDPR src0) %{
9781 predicate( UseSSE<=1 );
9782 match(Set src2 (SubD (MulD src0 src1) src2));
9783 format %{ "FLD $src0\t# ===MACRO3d===\n\t"
9784 "DMUL ST,$src1\n\t"
9785 "DSUBRp $src2,ST" %}
9786 ins_cost(250);
9787 ins_encode( Push_Reg_FPR(src0),
9788 FMul_ST_reg(src1),
9789 Opcode(0xDE), Opc_plus(0xE0,src2));
9790 ins_pipe( fpu_reg_reg_reg );
9791 %}
9794 instruct divDPR_reg(regDPR dst, regDPR src) %{
9795 predicate( UseSSE<=1 );
9796 match(Set dst (DivD dst src));
9798 format %{ "FLD $src\n\t"
9799 "FDIVp $dst,ST" %}
9800 opcode(0xDE, 0x7); /* DE F8+i or DE /7*/
9801 ins_cost(150);
9802 ins_encode( Push_Reg_DPR(src),
9803 OpcP, RegOpc(dst) );
9804 ins_pipe( fpu_reg_reg );
9805 %}
9807 // Strict FP instruction biases argument before division then
9808 // biases result, to avoid double rounding of subnormals.
9809 //
9810 // scale dividend by multiplying dividend by 2^(-15360)
9811 // load divisor
9812 // divide scaled dividend by divisor
9813 // rescale quotient by 2^(15360)
9814 //
9815 instruct strictfp_divDPR_reg(regDPR1 dst, regnotDPR1 src) %{
9816 predicate (UseSSE<=1);
9817 match(Set dst (DivD dst src));
9818 predicate( UseSSE<=1 && Compile::current()->has_method() && Compile::current()->method()->is_strict() );
9819 ins_cost(01);
9821 format %{ "FLD StubRoutines::_fpu_subnormal_bias1\n\t"
9822 "DMULp $dst,ST\n\t"
9823 "FLD $src\n\t"
9824 "FDIVp $dst,ST\n\t"
9825 "FLD StubRoutines::_fpu_subnormal_bias2\n\t"
9826 "DMULp $dst,ST\n\t" %}
9827 opcode(0xDE, 0x7); /* DE F8+i or DE /7*/
9828 ins_encode( strictfp_bias1(dst),
9829 Push_Reg_DPR(src),
9830 OpcP, RegOpc(dst),
9831 strictfp_bias2(dst) );
9832 ins_pipe( fpu_reg_reg );
9833 %}
9835 instruct divDPR_reg_round(stackSlotD dst, regDPR src1, regDPR src2) %{
9836 predicate( UseSSE<=1 && !(Compile::current()->has_method() && Compile::current()->method()->is_strict()) );
9837 match(Set dst (RoundDouble (DivD src1 src2)));
9839 format %{ "FLD $src1\n\t"
9840 "FDIV ST,$src2\n\t"
9841 "FSTP_D $dst\t# D-round" %}
9842 opcode(0xD8, 0x6); /* D8 F0+i or D8 /6 */
9843 ins_encode( Push_Reg_DPR(src1),
9844 OpcP, RegOpc(src2), Pop_Mem_DPR(dst) );
9845 ins_pipe( fpu_mem_reg_reg );
9846 %}
9849 instruct modDPR_reg(regDPR dst, regDPR src, eAXRegI rax, eFlagsReg cr) %{
9850 predicate(UseSSE<=1);
9851 match(Set dst (ModD dst src));
9852 effect(KILL rax, KILL cr); // emitModDPR() uses EAX and EFLAGS
9854 format %{ "DMOD $dst,$src" %}
9855 ins_cost(250);
9856 ins_encode(Push_Reg_Mod_DPR(dst, src),
9857 emitModDPR(),
9858 Push_Result_Mod_DPR(src),
9859 Pop_Reg_DPR(dst));
9860 ins_pipe( pipe_slow );
9861 %}
9863 instruct modD_reg(regD dst, regD src0, regD src1, eAXRegI rax, eFlagsReg cr) %{
9864 predicate(UseSSE>=2);
9865 match(Set dst (ModD src0 src1));
9866 effect(KILL rax, KILL cr);
9868 format %{ "SUB ESP,8\t # DMOD\n"
9869 "\tMOVSD [ESP+0],$src1\n"
9870 "\tFLD_D [ESP+0]\n"
9871 "\tMOVSD [ESP+0],$src0\n"
9872 "\tFLD_D [ESP+0]\n"
9873 "loop:\tFPREM\n"
9874 "\tFWAIT\n"
9875 "\tFNSTSW AX\n"
9876 "\tSAHF\n"
9877 "\tJP loop\n"
9878 "\tFSTP_D [ESP+0]\n"
9879 "\tMOVSD $dst,[ESP+0]\n"
9880 "\tADD ESP,8\n"
9881 "\tFSTP ST0\t # Restore FPU Stack"
9882 %}
9883 ins_cost(250);
9884 ins_encode( Push_ModD_encoding(src0, src1), emitModDPR(), Push_ResultD(dst), PopFPU);
9885 ins_pipe( pipe_slow );
9886 %}
9888 instruct sinDPR_reg(regDPR1 dst, regDPR1 src) %{
9889 predicate (UseSSE<=1);
9890 match(Set dst (SinD src));
9891 ins_cost(1800);
9892 format %{ "DSIN $dst" %}
9893 opcode(0xD9, 0xFE);
9894 ins_encode( OpcP, OpcS );
9895 ins_pipe( pipe_slow );
9896 %}
9898 instruct sinD_reg(regD dst, eFlagsReg cr) %{
9899 predicate (UseSSE>=2);
9900 match(Set dst (SinD dst));
9901 effect(KILL cr); // Push_{Src|Result}D() uses "{SUB|ADD} ESP,8"
9902 ins_cost(1800);
9903 format %{ "DSIN $dst" %}
9904 opcode(0xD9, 0xFE);
9905 ins_encode( Push_SrcD(dst), OpcP, OpcS, Push_ResultD(dst) );
9906 ins_pipe( pipe_slow );
9907 %}
9909 instruct cosDPR_reg(regDPR1 dst, regDPR1 src) %{
9910 predicate (UseSSE<=1);
9911 match(Set dst (CosD src));
9912 ins_cost(1800);
9913 format %{ "DCOS $dst" %}
9914 opcode(0xD9, 0xFF);
9915 ins_encode( OpcP, OpcS );
9916 ins_pipe( pipe_slow );
9917 %}
9919 instruct cosD_reg(regD dst, eFlagsReg cr) %{
9920 predicate (UseSSE>=2);
9921 match(Set dst (CosD dst));
9922 effect(KILL cr); // Push_{Src|Result}D() uses "{SUB|ADD} ESP,8"
9923 ins_cost(1800);
9924 format %{ "DCOS $dst" %}
9925 opcode(0xD9, 0xFF);
9926 ins_encode( Push_SrcD(dst), OpcP, OpcS, Push_ResultD(dst) );
9927 ins_pipe( pipe_slow );
9928 %}
9930 instruct tanDPR_reg(regDPR1 dst, regDPR1 src) %{
9931 predicate (UseSSE<=1);
9932 match(Set dst(TanD src));
9933 format %{ "DTAN $dst" %}
9934 ins_encode( Opcode(0xD9), Opcode(0xF2), // fptan
9935 Opcode(0xDD), Opcode(0xD8)); // fstp st
9936 ins_pipe( pipe_slow );
9937 %}
9939 instruct tanD_reg(regD dst, eFlagsReg cr) %{
9940 predicate (UseSSE>=2);
9941 match(Set dst(TanD dst));
9942 effect(KILL cr); // Push_{Src|Result}D() uses "{SUB|ADD} ESP,8"
9943 format %{ "DTAN $dst" %}
9944 ins_encode( Push_SrcD(dst),
9945 Opcode(0xD9), Opcode(0xF2), // fptan
9946 Opcode(0xDD), Opcode(0xD8), // fstp st
9947 Push_ResultD(dst) );
9948 ins_pipe( pipe_slow );
9949 %}
9951 instruct atanDPR_reg(regDPR dst, regDPR src) %{
9952 predicate (UseSSE<=1);
9953 match(Set dst(AtanD dst src));
9954 format %{ "DATA $dst,$src" %}
9955 opcode(0xD9, 0xF3);
9956 ins_encode( Push_Reg_DPR(src),
9957 OpcP, OpcS, RegOpc(dst) );
9958 ins_pipe( pipe_slow );
9959 %}
9961 instruct atanD_reg(regD dst, regD src, eFlagsReg cr) %{
9962 predicate (UseSSE>=2);
9963 match(Set dst(AtanD dst src));
9964 effect(KILL cr); // Push_{Src|Result}D() uses "{SUB|ADD} ESP,8"
9965 format %{ "DATA $dst,$src" %}
9966 opcode(0xD9, 0xF3);
9967 ins_encode( Push_SrcD(src),
9968 OpcP, OpcS, Push_ResultD(dst) );
9969 ins_pipe( pipe_slow );
9970 %}
9972 instruct sqrtDPR_reg(regDPR dst, regDPR src) %{
9973 predicate (UseSSE<=1);
9974 match(Set dst (SqrtD src));
9975 format %{ "DSQRT $dst,$src" %}
9976 opcode(0xFA, 0xD9);
9977 ins_encode( Push_Reg_DPR(src),
9978 OpcS, OpcP, Pop_Reg_DPR(dst) );
9979 ins_pipe( pipe_slow );
9980 %}
9982 instruct powDPR_reg(regDPR X, regDPR1 Y, eAXRegI rax, eDXRegI rdx, eCXRegI rcx, eFlagsReg cr) %{
9983 predicate (UseSSE<=1);
9984 match(Set Y (PowD X Y)); // Raise X to the Yth power
9985 effect(KILL rax, KILL rdx, KILL rcx, KILL cr);
9986 format %{ "fast_pow $X $Y -> $Y // KILL $rax, $rcx, $rdx" %}
9987 ins_encode %{
9988 __ subptr(rsp, 8);
9989 __ fld_s($X$$reg - 1);
9990 __ fast_pow();
9991 __ addptr(rsp, 8);
9992 %}
9993 ins_pipe( pipe_slow );
9994 %}
9996 instruct powD_reg(regD dst, regD src0, regD src1, eAXRegI rax, eDXRegI rdx, eCXRegI rcx, eFlagsReg cr) %{
9997 predicate (UseSSE>=2);
9998 match(Set dst (PowD src0 src1)); // Raise src0 to the src1'th power
9999 effect(KILL rax, KILL rdx, KILL rcx, KILL cr);
10000 format %{ "fast_pow $src0 $src1 -> $dst // KILL $rax, $rcx, $rdx" %}
10001 ins_encode %{
10002 __ subptr(rsp, 8);
10003 __ movdbl(Address(rsp, 0), $src1$$XMMRegister);
10004 __ fld_d(Address(rsp, 0));
10005 __ movdbl(Address(rsp, 0), $src0$$XMMRegister);
10006 __ fld_d(Address(rsp, 0));
10007 __ fast_pow();
10008 __ fstp_d(Address(rsp, 0));
10009 __ movdbl($dst$$XMMRegister, Address(rsp, 0));
10010 __ addptr(rsp, 8);
10011 %}
10012 ins_pipe( pipe_slow );
10013 %}
10016 instruct expDPR_reg(regDPR1 dpr1, eAXRegI rax, eDXRegI rdx, eCXRegI rcx, eFlagsReg cr) %{
10017 predicate (UseSSE<=1);
10018 match(Set dpr1 (ExpD dpr1));
10019 effect(KILL rax, KILL rcx, KILL rdx, KILL cr);
10020 format %{ "fast_exp $dpr1 -> $dpr1 // KILL $rax, $rcx, $rdx" %}
10021 ins_encode %{
10022 __ fast_exp();
10023 %}
10024 ins_pipe( pipe_slow );
10025 %}
10027 instruct expD_reg(regD dst, regD src, eAXRegI rax, eDXRegI rdx, eCXRegI rcx, eFlagsReg cr) %{
10028 predicate (UseSSE>=2);
10029 match(Set dst (ExpD src));
10030 effect(KILL rax, KILL rcx, KILL rdx, KILL cr);
10031 format %{ "fast_exp $dst -> $src // KILL $rax, $rcx, $rdx" %}
10032 ins_encode %{
10033 __ subptr(rsp, 8);
10034 __ movdbl(Address(rsp, 0), $src$$XMMRegister);
10035 __ fld_d(Address(rsp, 0));
10036 __ fast_exp();
10037 __ fstp_d(Address(rsp, 0));
10038 __ movdbl($dst$$XMMRegister, Address(rsp, 0));
10039 __ addptr(rsp, 8);
10040 %}
10041 ins_pipe( pipe_slow );
10042 %}
10044 instruct log10DPR_reg(regDPR1 dst, regDPR1 src) %{
10045 predicate (UseSSE<=1);
10046 // The source Double operand on FPU stack
10047 match(Set dst (Log10D src));
10048 // fldlg2 ; push log_10(2) on the FPU stack; full 80-bit number
10049 // fxch ; swap ST(0) with ST(1)
10050 // fyl2x ; compute log_10(2) * log_2(x)
10051 format %{ "FLDLG2 \t\t\t#Log10\n\t"
10052 "FXCH \n\t"
10053 "FYL2X \t\t\t# Q=Log10*Log_2(x)"
10054 %}
10055 ins_encode( Opcode(0xD9), Opcode(0xEC), // fldlg2
10056 Opcode(0xD9), Opcode(0xC9), // fxch
10057 Opcode(0xD9), Opcode(0xF1)); // fyl2x
10059 ins_pipe( pipe_slow );
10060 %}
10062 instruct log10D_reg(regD dst, regD src, eFlagsReg cr) %{
10063 predicate (UseSSE>=2);
10064 effect(KILL cr);
10065 match(Set dst (Log10D src));
10066 // fldlg2 ; push log_10(2) on the FPU stack; full 80-bit number
10067 // fyl2x ; compute log_10(2) * log_2(x)
10068 format %{ "FLDLG2 \t\t\t#Log10\n\t"
10069 "FYL2X \t\t\t# Q=Log10*Log_2(x)"
10070 %}
10071 ins_encode( Opcode(0xD9), Opcode(0xEC), // fldlg2
10072 Push_SrcD(src),
10073 Opcode(0xD9), Opcode(0xF1), // fyl2x
10074 Push_ResultD(dst));
10076 ins_pipe( pipe_slow );
10077 %}
10079 instruct logDPR_reg(regDPR1 dst, regDPR1 src) %{
10080 predicate (UseSSE<=1);
10081 // The source Double operand on FPU stack
10082 match(Set dst (LogD src));
10083 // fldln2 ; push log_e(2) on the FPU stack; full 80-bit number
10084 // fxch ; swap ST(0) with ST(1)
10085 // fyl2x ; compute log_e(2) * log_2(x)
10086 format %{ "FLDLN2 \t\t\t#Log_e\n\t"
10087 "FXCH \n\t"
10088 "FYL2X \t\t\t# Q=Log_e*Log_2(x)"
10089 %}
10090 ins_encode( Opcode(0xD9), Opcode(0xED), // fldln2
10091 Opcode(0xD9), Opcode(0xC9), // fxch
10092 Opcode(0xD9), Opcode(0xF1)); // fyl2x
10094 ins_pipe( pipe_slow );
10095 %}
10097 instruct logD_reg(regD dst, regD src, eFlagsReg cr) %{
10098 predicate (UseSSE>=2);
10099 effect(KILL cr);
10100 // The source and result Double operands in XMM registers
10101 match(Set dst (LogD src));
10102 // fldln2 ; push log_e(2) on the FPU stack; full 80-bit number
10103 // fyl2x ; compute log_e(2) * log_2(x)
10104 format %{ "FLDLN2 \t\t\t#Log_e\n\t"
10105 "FYL2X \t\t\t# Q=Log_e*Log_2(x)"
10106 %}
10107 ins_encode( Opcode(0xD9), Opcode(0xED), // fldln2
10108 Push_SrcD(src),
10109 Opcode(0xD9), Opcode(0xF1), // fyl2x
10110 Push_ResultD(dst));
10111 ins_pipe( pipe_slow );
10112 %}
10114 //-------------Float Instructions-------------------------------
10115 // Float Math
10117 // Code for float compare:
10118 // fcompp();
10119 // fwait(); fnstsw_ax();
10120 // sahf();
10121 // movl(dst, unordered_result);
10122 // jcc(Assembler::parity, exit);
10123 // movl(dst, less_result);
10124 // jcc(Assembler::below, exit);
10125 // movl(dst, equal_result);
10126 // jcc(Assembler::equal, exit);
10127 // movl(dst, greater_result);
10128 // exit:
10130 // P6 version of float compare, sets condition codes in EFLAGS
10131 instruct cmpFPR_cc_P6(eFlagsRegU cr, regFPR src1, regFPR src2, eAXRegI rax) %{
10132 predicate(VM_Version::supports_cmov() && UseSSE == 0);
10133 match(Set cr (CmpF src1 src2));
10134 effect(KILL rax);
10135 ins_cost(150);
10136 format %{ "FLD $src1\n\t"
10137 "FUCOMIP ST,$src2 // P6 instruction\n\t"
10138 "JNP exit\n\t"
10139 "MOV ah,1 // saw a NaN, set CF (treat as LT)\n\t"
10140 "SAHF\n"
10141 "exit:\tNOP // avoid branch to branch" %}
10142 opcode(0xDF, 0x05); /* DF E8+i or DF /5 */
10143 ins_encode( Push_Reg_DPR(src1),
10144 OpcP, RegOpc(src2),
10145 cmpF_P6_fixup );
10146 ins_pipe( pipe_slow );
10147 %}
10149 instruct cmpFPR_cc_P6CF(eFlagsRegUCF cr, regFPR src1, regFPR src2) %{
10150 predicate(VM_Version::supports_cmov() && UseSSE == 0);
10151 match(Set cr (CmpF src1 src2));
10152 ins_cost(100);
10153 format %{ "FLD $src1\n\t"
10154 "FUCOMIP ST,$src2 // P6 instruction" %}
10155 opcode(0xDF, 0x05); /* DF E8+i or DF /5 */
10156 ins_encode( Push_Reg_DPR(src1),
10157 OpcP, RegOpc(src2));
10158 ins_pipe( pipe_slow );
10159 %}
10162 // Compare & branch
10163 instruct cmpFPR_cc(eFlagsRegU cr, regFPR src1, regFPR src2, eAXRegI rax) %{
10164 predicate(UseSSE == 0);
10165 match(Set cr (CmpF src1 src2));
10166 effect(KILL rax);
10167 ins_cost(200);
10168 format %{ "FLD $src1\n\t"
10169 "FCOMp $src2\n\t"
10170 "FNSTSW AX\n\t"
10171 "TEST AX,0x400\n\t"
10172 "JZ,s flags\n\t"
10173 "MOV AH,1\t# unordered treat as LT\n"
10174 "flags:\tSAHF" %}
10175 opcode(0xD8, 0x3); /* D8 D8+i or D8 /3 */
10176 ins_encode( Push_Reg_DPR(src1),
10177 OpcP, RegOpc(src2),
10178 fpu_flags);
10179 ins_pipe( pipe_slow );
10180 %}
10182 // Compare vs zero into -1,0,1
10183 instruct cmpFPR_0(rRegI dst, regFPR src1, immFPR0 zero, eAXRegI rax, eFlagsReg cr) %{
10184 predicate(UseSSE == 0);
10185 match(Set dst (CmpF3 src1 zero));
10186 effect(KILL cr, KILL rax);
10187 ins_cost(280);
10188 format %{ "FTSTF $dst,$src1" %}
10189 opcode(0xE4, 0xD9);
10190 ins_encode( Push_Reg_DPR(src1),
10191 OpcS, OpcP, PopFPU,
10192 CmpF_Result(dst));
10193 ins_pipe( pipe_slow );
10194 %}
10196 // Compare into -1,0,1
10197 instruct cmpFPR_reg(rRegI dst, regFPR src1, regFPR src2, eAXRegI rax, eFlagsReg cr) %{
10198 predicate(UseSSE == 0);
10199 match(Set dst (CmpF3 src1 src2));
10200 effect(KILL cr, KILL rax);
10201 ins_cost(300);
10202 format %{ "FCMPF $dst,$src1,$src2" %}
10203 opcode(0xD8, 0x3); /* D8 D8+i or D8 /3 */
10204 ins_encode( Push_Reg_DPR(src1),
10205 OpcP, RegOpc(src2),
10206 CmpF_Result(dst));
10207 ins_pipe( pipe_slow );
10208 %}
10210 // float compare and set condition codes in EFLAGS by XMM regs
10211 instruct cmpF_cc(eFlagsRegU cr, regF src1, regF src2) %{
10212 predicate(UseSSE>=1);
10213 match(Set cr (CmpF src1 src2));
10214 ins_cost(145);
10215 format %{ "UCOMISS $src1,$src2\n\t"
10216 "JNP,s exit\n\t"
10217 "PUSHF\t# saw NaN, set CF\n\t"
10218 "AND [rsp], #0xffffff2b\n\t"
10219 "POPF\n"
10220 "exit:" %}
10221 ins_encode %{
10222 __ ucomiss($src1$$XMMRegister, $src2$$XMMRegister);
10223 emit_cmpfp_fixup(_masm);
10224 %}
10225 ins_pipe( pipe_slow );
10226 %}
10228 instruct cmpF_ccCF(eFlagsRegUCF cr, regF src1, regF src2) %{
10229 predicate(UseSSE>=1);
10230 match(Set cr (CmpF src1 src2));
10231 ins_cost(100);
10232 format %{ "UCOMISS $src1,$src2" %}
10233 ins_encode %{
10234 __ ucomiss($src1$$XMMRegister, $src2$$XMMRegister);
10235 %}
10236 ins_pipe( pipe_slow );
10237 %}
10239 // float compare and set condition codes in EFLAGS by XMM regs
10240 instruct cmpF_ccmem(eFlagsRegU cr, regF src1, memory src2) %{
10241 predicate(UseSSE>=1);
10242 match(Set cr (CmpF src1 (LoadF src2)));
10243 ins_cost(165);
10244 format %{ "UCOMISS $src1,$src2\n\t"
10245 "JNP,s exit\n\t"
10246 "PUSHF\t# saw NaN, set CF\n\t"
10247 "AND [rsp], #0xffffff2b\n\t"
10248 "POPF\n"
10249 "exit:" %}
10250 ins_encode %{
10251 __ ucomiss($src1$$XMMRegister, $src2$$Address);
10252 emit_cmpfp_fixup(_masm);
10253 %}
10254 ins_pipe( pipe_slow );
10255 %}
10257 instruct cmpF_ccmemCF(eFlagsRegUCF cr, regF src1, memory src2) %{
10258 predicate(UseSSE>=1);
10259 match(Set cr (CmpF src1 (LoadF src2)));
10260 ins_cost(100);
10261 format %{ "UCOMISS $src1,$src2" %}
10262 ins_encode %{
10263 __ ucomiss($src1$$XMMRegister, $src2$$Address);
10264 %}
10265 ins_pipe( pipe_slow );
10266 %}
10268 // Compare into -1,0,1 in XMM
10269 instruct cmpF_reg(xRegI dst, regF src1, regF src2, eFlagsReg cr) %{
10270 predicate(UseSSE>=1);
10271 match(Set dst (CmpF3 src1 src2));
10272 effect(KILL cr);
10273 ins_cost(255);
10274 format %{ "UCOMISS $src1, $src2\n\t"
10275 "MOV $dst, #-1\n\t"
10276 "JP,s done\n\t"
10277 "JB,s done\n\t"
10278 "SETNE $dst\n\t"
10279 "MOVZB $dst, $dst\n"
10280 "done:" %}
10281 ins_encode %{
10282 __ ucomiss($src1$$XMMRegister, $src2$$XMMRegister);
10283 emit_cmpfp3(_masm, $dst$$Register);
10284 %}
10285 ins_pipe( pipe_slow );
10286 %}
10288 // Compare into -1,0,1 in XMM and memory
10289 instruct cmpF_regmem(xRegI dst, regF src1, memory src2, eFlagsReg cr) %{
10290 predicate(UseSSE>=1);
10291 match(Set dst (CmpF3 src1 (LoadF src2)));
10292 effect(KILL cr);
10293 ins_cost(275);
10294 format %{ "UCOMISS $src1, $src2\n\t"
10295 "MOV $dst, #-1\n\t"
10296 "JP,s done\n\t"
10297 "JB,s done\n\t"
10298 "SETNE $dst\n\t"
10299 "MOVZB $dst, $dst\n"
10300 "done:" %}
10301 ins_encode %{
10302 __ ucomiss($src1$$XMMRegister, $src2$$Address);
10303 emit_cmpfp3(_masm, $dst$$Register);
10304 %}
10305 ins_pipe( pipe_slow );
10306 %}
10308 // Spill to obtain 24-bit precision
10309 instruct subFPR24_reg(stackSlotF dst, regFPR src1, regFPR src2) %{
10310 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10311 match(Set dst (SubF src1 src2));
10313 format %{ "FSUB $dst,$src1 - $src2" %}
10314 opcode(0xD8, 0x4); /* D8 E0+i or D8 /4 mod==0x3 ;; result in TOS */
10315 ins_encode( Push_Reg_FPR(src1),
10316 OpcReg_FPR(src2),
10317 Pop_Mem_FPR(dst) );
10318 ins_pipe( fpu_mem_reg_reg );
10319 %}
10320 //
10321 // This instruction does not round to 24-bits
10322 instruct subFPR_reg(regFPR dst, regFPR src) %{
10323 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10324 match(Set dst (SubF dst src));
10326 format %{ "FSUB $dst,$src" %}
10327 opcode(0xDE, 0x5); /* DE E8+i or DE /5 */
10328 ins_encode( Push_Reg_FPR(src),
10329 OpcP, RegOpc(dst) );
10330 ins_pipe( fpu_reg_reg );
10331 %}
10333 // Spill to obtain 24-bit precision
10334 instruct addFPR24_reg(stackSlotF dst, regFPR src1, regFPR src2) %{
10335 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10336 match(Set dst (AddF src1 src2));
10338 format %{ "FADD $dst,$src1,$src2" %}
10339 opcode(0xD8, 0x0); /* D8 C0+i */
10340 ins_encode( Push_Reg_FPR(src2),
10341 OpcReg_FPR(src1),
10342 Pop_Mem_FPR(dst) );
10343 ins_pipe( fpu_mem_reg_reg );
10344 %}
10345 //
10346 // This instruction does not round to 24-bits
10347 instruct addFPR_reg(regFPR dst, regFPR src) %{
10348 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10349 match(Set dst (AddF dst src));
10351 format %{ "FLD $src\n\t"
10352 "FADDp $dst,ST" %}
10353 opcode(0xDE, 0x0); /* DE C0+i or DE /0*/
10354 ins_encode( Push_Reg_FPR(src),
10355 OpcP, RegOpc(dst) );
10356 ins_pipe( fpu_reg_reg );
10357 %}
10359 instruct absFPR_reg(regFPR1 dst, regFPR1 src) %{
10360 predicate(UseSSE==0);
10361 match(Set dst (AbsF src));
10362 ins_cost(100);
10363 format %{ "FABS" %}
10364 opcode(0xE1, 0xD9);
10365 ins_encode( OpcS, OpcP );
10366 ins_pipe( fpu_reg_reg );
10367 %}
10369 instruct negFPR_reg(regFPR1 dst, regFPR1 src) %{
10370 predicate(UseSSE==0);
10371 match(Set dst (NegF src));
10372 ins_cost(100);
10373 format %{ "FCHS" %}
10374 opcode(0xE0, 0xD9);
10375 ins_encode( OpcS, OpcP );
10376 ins_pipe( fpu_reg_reg );
10377 %}
10379 // Cisc-alternate to addFPR_reg
10380 // Spill to obtain 24-bit precision
10381 instruct addFPR24_reg_mem(stackSlotF dst, regFPR src1, memory src2) %{
10382 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10383 match(Set dst (AddF src1 (LoadF src2)));
10385 format %{ "FLD $src2\n\t"
10386 "FADD ST,$src1\n\t"
10387 "FSTP_S $dst" %}
10388 opcode(0xD8, 0x0, 0xD9); /* D8 C0+i */ /* LoadF D9 /0 */
10389 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10390 OpcReg_FPR(src1),
10391 Pop_Mem_FPR(dst) );
10392 ins_pipe( fpu_mem_reg_mem );
10393 %}
10394 //
10395 // Cisc-alternate to addFPR_reg
10396 // This instruction does not round to 24-bits
10397 instruct addFPR_reg_mem(regFPR dst, memory src) %{
10398 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10399 match(Set dst (AddF dst (LoadF src)));
10401 format %{ "FADD $dst,$src" %}
10402 opcode(0xDE, 0x0, 0xD9); /* DE C0+i or DE /0*/ /* LoadF D9 /0 */
10403 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src),
10404 OpcP, RegOpc(dst) );
10405 ins_pipe( fpu_reg_mem );
10406 %}
10408 // // Following two instructions for _222_mpegaudio
10409 // Spill to obtain 24-bit precision
10410 instruct addFPR24_mem_reg(stackSlotF dst, regFPR src2, memory src1 ) %{
10411 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10412 match(Set dst (AddF src1 src2));
10414 format %{ "FADD $dst,$src1,$src2" %}
10415 opcode(0xD8, 0x0, 0xD9); /* D8 C0+i */ /* LoadF D9 /0 */
10416 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src1),
10417 OpcReg_FPR(src2),
10418 Pop_Mem_FPR(dst) );
10419 ins_pipe( fpu_mem_reg_mem );
10420 %}
10422 // Cisc-spill variant
10423 // Spill to obtain 24-bit precision
10424 instruct addFPR24_mem_cisc(stackSlotF dst, memory src1, memory src2) %{
10425 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10426 match(Set dst (AddF src1 (LoadF src2)));
10428 format %{ "FADD $dst,$src1,$src2 cisc" %}
10429 opcode(0xD8, 0x0, 0xD9); /* D8 C0+i */ /* LoadF D9 /0 */
10430 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10431 set_instruction_start,
10432 OpcP, RMopc_Mem(secondary,src1),
10433 Pop_Mem_FPR(dst) );
10434 ins_pipe( fpu_mem_mem_mem );
10435 %}
10437 // Spill to obtain 24-bit precision
10438 instruct addFPR24_mem_mem(stackSlotF dst, memory src1, memory src2) %{
10439 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10440 match(Set dst (AddF src1 src2));
10442 format %{ "FADD $dst,$src1,$src2" %}
10443 opcode(0xD8, 0x0, 0xD9); /* D8 /0 */ /* LoadF D9 /0 */
10444 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10445 set_instruction_start,
10446 OpcP, RMopc_Mem(secondary,src1),
10447 Pop_Mem_FPR(dst) );
10448 ins_pipe( fpu_mem_mem_mem );
10449 %}
10452 // Spill to obtain 24-bit precision
10453 instruct addFPR24_reg_imm(stackSlotF dst, regFPR src, immFPR con) %{
10454 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10455 match(Set dst (AddF src con));
10456 format %{ "FLD $src\n\t"
10457 "FADD_S [$constantaddress]\t# load from constant table: float=$con\n\t"
10458 "FSTP_S $dst" %}
10459 ins_encode %{
10460 __ fld_s($src$$reg - 1); // FLD ST(i-1)
10461 __ fadd_s($constantaddress($con));
10462 __ fstp_s(Address(rsp, $dst$$disp));
10463 %}
10464 ins_pipe(fpu_mem_reg_con);
10465 %}
10466 //
10467 // This instruction does not round to 24-bits
10468 instruct addFPR_reg_imm(regFPR dst, regFPR src, immFPR con) %{
10469 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10470 match(Set dst (AddF src con));
10471 format %{ "FLD $src\n\t"
10472 "FADD_S [$constantaddress]\t# load from constant table: float=$con\n\t"
10473 "FSTP $dst" %}
10474 ins_encode %{
10475 __ fld_s($src$$reg - 1); // FLD ST(i-1)
10476 __ fadd_s($constantaddress($con));
10477 __ fstp_d($dst$$reg);
10478 %}
10479 ins_pipe(fpu_reg_reg_con);
10480 %}
10482 // Spill to obtain 24-bit precision
10483 instruct mulFPR24_reg(stackSlotF dst, regFPR src1, regFPR src2) %{
10484 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10485 match(Set dst (MulF src1 src2));
10487 format %{ "FLD $src1\n\t"
10488 "FMUL $src2\n\t"
10489 "FSTP_S $dst" %}
10490 opcode(0xD8, 0x1); /* D8 C8+i or D8 /1 ;; result in TOS */
10491 ins_encode( Push_Reg_FPR(src1),
10492 OpcReg_FPR(src2),
10493 Pop_Mem_FPR(dst) );
10494 ins_pipe( fpu_mem_reg_reg );
10495 %}
10496 //
10497 // This instruction does not round to 24-bits
10498 instruct mulFPR_reg(regFPR dst, regFPR src1, regFPR src2) %{
10499 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10500 match(Set dst (MulF src1 src2));
10502 format %{ "FLD $src1\n\t"
10503 "FMUL $src2\n\t"
10504 "FSTP_S $dst" %}
10505 opcode(0xD8, 0x1); /* D8 C8+i */
10506 ins_encode( Push_Reg_FPR(src2),
10507 OpcReg_FPR(src1),
10508 Pop_Reg_FPR(dst) );
10509 ins_pipe( fpu_reg_reg_reg );
10510 %}
10513 // Spill to obtain 24-bit precision
10514 // Cisc-alternate to reg-reg multiply
10515 instruct mulFPR24_reg_mem(stackSlotF dst, regFPR src1, memory src2) %{
10516 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10517 match(Set dst (MulF src1 (LoadF src2)));
10519 format %{ "FLD_S $src2\n\t"
10520 "FMUL $src1\n\t"
10521 "FSTP_S $dst" %}
10522 opcode(0xD8, 0x1, 0xD9); /* D8 C8+i or DE /1*/ /* LoadF D9 /0 */
10523 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10524 OpcReg_FPR(src1),
10525 Pop_Mem_FPR(dst) );
10526 ins_pipe( fpu_mem_reg_mem );
10527 %}
10528 //
10529 // This instruction does not round to 24-bits
10530 // Cisc-alternate to reg-reg multiply
10531 instruct mulFPR_reg_mem(regFPR dst, regFPR src1, memory src2) %{
10532 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10533 match(Set dst (MulF src1 (LoadF src2)));
10535 format %{ "FMUL $dst,$src1,$src2" %}
10536 opcode(0xD8, 0x1, 0xD9); /* D8 C8+i */ /* LoadF D9 /0 */
10537 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10538 OpcReg_FPR(src1),
10539 Pop_Reg_FPR(dst) );
10540 ins_pipe( fpu_reg_reg_mem );
10541 %}
10543 // Spill to obtain 24-bit precision
10544 instruct mulFPR24_mem_mem(stackSlotF dst, memory src1, memory src2) %{
10545 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10546 match(Set dst (MulF src1 src2));
10548 format %{ "FMUL $dst,$src1,$src2" %}
10549 opcode(0xD8, 0x1, 0xD9); /* D8 /1 */ /* LoadF D9 /0 */
10550 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,src2),
10551 set_instruction_start,
10552 OpcP, RMopc_Mem(secondary,src1),
10553 Pop_Mem_FPR(dst) );
10554 ins_pipe( fpu_mem_mem_mem );
10555 %}
10557 // Spill to obtain 24-bit precision
10558 instruct mulFPR24_reg_imm(stackSlotF dst, regFPR src, immFPR con) %{
10559 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10560 match(Set dst (MulF src con));
10562 format %{ "FLD $src\n\t"
10563 "FMUL_S [$constantaddress]\t# load from constant table: float=$con\n\t"
10564 "FSTP_S $dst" %}
10565 ins_encode %{
10566 __ fld_s($src$$reg - 1); // FLD ST(i-1)
10567 __ fmul_s($constantaddress($con));
10568 __ fstp_s(Address(rsp, $dst$$disp));
10569 %}
10570 ins_pipe(fpu_mem_reg_con);
10571 %}
10572 //
10573 // This instruction does not round to 24-bits
10574 instruct mulFPR_reg_imm(regFPR dst, regFPR src, immFPR con) %{
10575 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10576 match(Set dst (MulF src con));
10578 format %{ "FLD $src\n\t"
10579 "FMUL_S [$constantaddress]\t# load from constant table: float=$con\n\t"
10580 "FSTP $dst" %}
10581 ins_encode %{
10582 __ fld_s($src$$reg - 1); // FLD ST(i-1)
10583 __ fmul_s($constantaddress($con));
10584 __ fstp_d($dst$$reg);
10585 %}
10586 ins_pipe(fpu_reg_reg_con);
10587 %}
10590 //
10591 // MACRO1 -- subsume unshared load into mulFPR
10592 // This instruction does not round to 24-bits
10593 instruct mulFPR_reg_load1(regFPR dst, regFPR src, memory mem1 ) %{
10594 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10595 match(Set dst (MulF (LoadF mem1) src));
10597 format %{ "FLD $mem1 ===MACRO1===\n\t"
10598 "FMUL ST,$src\n\t"
10599 "FSTP $dst" %}
10600 opcode(0xD8, 0x1, 0xD9); /* D8 C8+i or D8 /1 */ /* LoadF D9 /0 */
10601 ins_encode( Opcode(tertiary), RMopc_Mem(0x00,mem1),
10602 OpcReg_FPR(src),
10603 Pop_Reg_FPR(dst) );
10604 ins_pipe( fpu_reg_reg_mem );
10605 %}
10606 //
10607 // MACRO2 -- addFPR a mulFPR which subsumed an unshared load
10608 // This instruction does not round to 24-bits
10609 instruct addFPR_mulFPR_reg_load1(regFPR dst, memory mem1, regFPR src1, regFPR src2) %{
10610 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10611 match(Set dst (AddF (MulF (LoadF mem1) src1) src2));
10612 ins_cost(95);
10614 format %{ "FLD $mem1 ===MACRO2===\n\t"
10615 "FMUL ST,$src1 subsume mulFPR left load\n\t"
10616 "FADD ST,$src2\n\t"
10617 "FSTP $dst" %}
10618 opcode(0xD9); /* LoadF D9 /0 */
10619 ins_encode( OpcP, RMopc_Mem(0x00,mem1),
10620 FMul_ST_reg(src1),
10621 FAdd_ST_reg(src2),
10622 Pop_Reg_FPR(dst) );
10623 ins_pipe( fpu_reg_mem_reg_reg );
10624 %}
10626 // MACRO3 -- addFPR a mulFPR
10627 // This instruction does not round to 24-bits. It is a '2-address'
10628 // instruction in that the result goes back to src2. This eliminates
10629 // a move from the macro; possibly the register allocator will have
10630 // to add it back (and maybe not).
10631 instruct addFPR_mulFPR_reg(regFPR src2, regFPR src1, regFPR src0) %{
10632 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10633 match(Set src2 (AddF (MulF src0 src1) src2));
10635 format %{ "FLD $src0 ===MACRO3===\n\t"
10636 "FMUL ST,$src1\n\t"
10637 "FADDP $src2,ST" %}
10638 opcode(0xD9); /* LoadF D9 /0 */
10639 ins_encode( Push_Reg_FPR(src0),
10640 FMul_ST_reg(src1),
10641 FAddP_reg_ST(src2) );
10642 ins_pipe( fpu_reg_reg_reg );
10643 %}
10645 // MACRO4 -- divFPR subFPR
10646 // This instruction does not round to 24-bits
10647 instruct subFPR_divFPR_reg(regFPR dst, regFPR src1, regFPR src2, regFPR src3) %{
10648 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10649 match(Set dst (DivF (SubF src2 src1) src3));
10651 format %{ "FLD $src2 ===MACRO4===\n\t"
10652 "FSUB ST,$src1\n\t"
10653 "FDIV ST,$src3\n\t"
10654 "FSTP $dst" %}
10655 opcode(0xDE, 0x7); /* DE F8+i or DE /7*/
10656 ins_encode( Push_Reg_FPR(src2),
10657 subFPR_divFPR_encode(src1,src3),
10658 Pop_Reg_FPR(dst) );
10659 ins_pipe( fpu_reg_reg_reg_reg );
10660 %}
10662 // Spill to obtain 24-bit precision
10663 instruct divFPR24_reg(stackSlotF dst, regFPR src1, regFPR src2) %{
10664 predicate(UseSSE==0 && Compile::current()->select_24_bit_instr());
10665 match(Set dst (DivF src1 src2));
10667 format %{ "FDIV $dst,$src1,$src2" %}
10668 opcode(0xD8, 0x6); /* D8 F0+i or DE /6*/
10669 ins_encode( Push_Reg_FPR(src1),
10670 OpcReg_FPR(src2),
10671 Pop_Mem_FPR(dst) );
10672 ins_pipe( fpu_mem_reg_reg );
10673 %}
10674 //
10675 // This instruction does not round to 24-bits
10676 instruct divFPR_reg(regFPR dst, regFPR src) %{
10677 predicate(UseSSE==0 && !Compile::current()->select_24_bit_instr());
10678 match(Set dst (DivF dst src));
10680 format %{ "FDIV $dst,$src" %}
10681 opcode(0xDE, 0x7); /* DE F8+i or DE /7*/
10682 ins_encode( Push_Reg_FPR(src),
10683 OpcP, RegOpc(dst) );
10684 ins_pipe( fpu_reg_reg );
10685 %}
10688 // Spill to obtain 24-bit precision
10689 instruct modFPR24_reg(stackSlotF dst, regFPR src1, regFPR src2, eAXRegI rax, eFlagsReg cr) %{
10690 predicate( UseSSE==0 && Compile::current()->select_24_bit_instr());
10691 match(Set dst (ModF src1 src2));
10692 effect(KILL rax, KILL cr); // emitModDPR() uses EAX and EFLAGS
10694 format %{ "FMOD $dst,$src1,$src2" %}
10695 ins_encode( Push_Reg_Mod_DPR(src1, src2),
10696 emitModDPR(),
10697 Push_Result_Mod_DPR(src2),
10698 Pop_Mem_FPR(dst));
10699 ins_pipe( pipe_slow );
10700 %}
10701 //
10702 // This instruction does not round to 24-bits
10703 instruct modFPR_reg(regFPR dst, regFPR src, eAXRegI rax, eFlagsReg cr) %{
10704 predicate( UseSSE==0 && !Compile::current()->select_24_bit_instr());
10705 match(Set dst (ModF dst src));
10706 effect(KILL rax, KILL cr); // emitModDPR() uses EAX and EFLAGS
10708 format %{ "FMOD $dst,$src" %}
10709 ins_encode(Push_Reg_Mod_DPR(dst, src),
10710 emitModDPR(),
10711 Push_Result_Mod_DPR(src),
10712 Pop_Reg_FPR(dst));
10713 ins_pipe( pipe_slow );
10714 %}
10716 instruct modF_reg(regF dst, regF src0, regF src1, eAXRegI rax, eFlagsReg cr) %{
10717 predicate(UseSSE>=1);
10718 match(Set dst (ModF src0 src1));
10719 effect(KILL rax, KILL cr);
10720 format %{ "SUB ESP,4\t # FMOD\n"
10721 "\tMOVSS [ESP+0],$src1\n"
10722 "\tFLD_S [ESP+0]\n"
10723 "\tMOVSS [ESP+0],$src0\n"
10724 "\tFLD_S [ESP+0]\n"
10725 "loop:\tFPREM\n"
10726 "\tFWAIT\n"
10727 "\tFNSTSW AX\n"
10728 "\tSAHF\n"
10729 "\tJP loop\n"
10730 "\tFSTP_S [ESP+0]\n"
10731 "\tMOVSS $dst,[ESP+0]\n"
10732 "\tADD ESP,4\n"
10733 "\tFSTP ST0\t # Restore FPU Stack"
10734 %}
10735 ins_cost(250);
10736 ins_encode( Push_ModF_encoding(src0, src1), emitModDPR(), Push_ResultF(dst,0x4), PopFPU);
10737 ins_pipe( pipe_slow );
10738 %}
10741 //----------Arithmetic Conversion Instructions---------------------------------
10742 // The conversions operations are all Alpha sorted. Please keep it that way!
10744 instruct roundFloat_mem_reg(stackSlotF dst, regFPR src) %{
10745 predicate(UseSSE==0);
10746 match(Set dst (RoundFloat src));
10747 ins_cost(125);
10748 format %{ "FST_S $dst,$src\t# F-round" %}
10749 ins_encode( Pop_Mem_Reg_FPR(dst, src) );
10750 ins_pipe( fpu_mem_reg );
10751 %}
10753 instruct roundDouble_mem_reg(stackSlotD dst, regDPR src) %{
10754 predicate(UseSSE<=1);
10755 match(Set dst (RoundDouble src));
10756 ins_cost(125);
10757 format %{ "FST_D $dst,$src\t# D-round" %}
10758 ins_encode( Pop_Mem_Reg_DPR(dst, src) );
10759 ins_pipe( fpu_mem_reg );
10760 %}
10762 // Force rounding to 24-bit precision and 6-bit exponent
10763 instruct convDPR2FPR_reg(stackSlotF dst, regDPR src) %{
10764 predicate(UseSSE==0);
10765 match(Set dst (ConvD2F src));
10766 format %{ "FST_S $dst,$src\t# F-round" %}
10767 expand %{
10768 roundFloat_mem_reg(dst,src);
10769 %}
10770 %}
10772 // Force rounding to 24-bit precision and 6-bit exponent
10773 instruct convDPR2F_reg(regF dst, regDPR src, eFlagsReg cr) %{
10774 predicate(UseSSE==1);
10775 match(Set dst (ConvD2F src));
10776 effect( KILL cr );
10777 format %{ "SUB ESP,4\n\t"
10778 "FST_S [ESP],$src\t# F-round\n\t"
10779 "MOVSS $dst,[ESP]\n\t"
10780 "ADD ESP,4" %}
10781 ins_encode %{
10782 __ subptr(rsp, 4);
10783 if ($src$$reg != FPR1L_enc) {
10784 __ fld_s($src$$reg-1);
10785 __ fstp_s(Address(rsp, 0));
10786 } else {
10787 __ fst_s(Address(rsp, 0));
10788 }
10789 __ movflt($dst$$XMMRegister, Address(rsp, 0));
10790 __ addptr(rsp, 4);
10791 %}
10792 ins_pipe( pipe_slow );
10793 %}
10795 // Force rounding double precision to single precision
10796 instruct convD2F_reg(regF dst, regD src) %{
10797 predicate(UseSSE>=2);
10798 match(Set dst (ConvD2F src));
10799 format %{ "CVTSD2SS $dst,$src\t# F-round" %}
10800 ins_encode %{
10801 __ cvtsd2ss ($dst$$XMMRegister, $src$$XMMRegister);
10802 %}
10803 ins_pipe( pipe_slow );
10804 %}
10806 instruct convFPR2DPR_reg_reg(regDPR dst, regFPR src) %{
10807 predicate(UseSSE==0);
10808 match(Set dst (ConvF2D src));
10809 format %{ "FST_S $dst,$src\t# D-round" %}
10810 ins_encode( Pop_Reg_Reg_DPR(dst, src));
10811 ins_pipe( fpu_reg_reg );
10812 %}
10814 instruct convFPR2D_reg(stackSlotD dst, regFPR src) %{
10815 predicate(UseSSE==1);
10816 match(Set dst (ConvF2D src));
10817 format %{ "FST_D $dst,$src\t# D-round" %}
10818 expand %{
10819 roundDouble_mem_reg(dst,src);
10820 %}
10821 %}
10823 instruct convF2DPR_reg(regDPR dst, regF src, eFlagsReg cr) %{
10824 predicate(UseSSE==1);
10825 match(Set dst (ConvF2D src));
10826 effect( KILL cr );
10827 format %{ "SUB ESP,4\n\t"
10828 "MOVSS [ESP] $src\n\t"
10829 "FLD_S [ESP]\n\t"
10830 "ADD ESP,4\n\t"
10831 "FSTP $dst\t# D-round" %}
10832 ins_encode %{
10833 __ subptr(rsp, 4);
10834 __ movflt(Address(rsp, 0), $src$$XMMRegister);
10835 __ fld_s(Address(rsp, 0));
10836 __ addptr(rsp, 4);
10837 __ fstp_d($dst$$reg);
10838 %}
10839 ins_pipe( pipe_slow );
10840 %}
10842 instruct convF2D_reg(regD dst, regF src) %{
10843 predicate(UseSSE>=2);
10844 match(Set dst (ConvF2D src));
10845 format %{ "CVTSS2SD $dst,$src\t# D-round" %}
10846 ins_encode %{
10847 __ cvtss2sd ($dst$$XMMRegister, $src$$XMMRegister);
10848 %}
10849 ins_pipe( pipe_slow );
10850 %}
10852 // Convert a double to an int. If the double is a NAN, stuff a zero in instead.
10853 instruct convDPR2I_reg_reg( eAXRegI dst, eDXRegI tmp, regDPR src, eFlagsReg cr ) %{
10854 predicate(UseSSE<=1);
10855 match(Set dst (ConvD2I src));
10856 effect( KILL tmp, KILL cr );
10857 format %{ "FLD $src\t# Convert double to int \n\t"
10858 "FLDCW trunc mode\n\t"
10859 "SUB ESP,4\n\t"
10860 "FISTp [ESP + #0]\n\t"
10861 "FLDCW std/24-bit mode\n\t"
10862 "POP EAX\n\t"
10863 "CMP EAX,0x80000000\n\t"
10864 "JNE,s fast\n\t"
10865 "FLD_D $src\n\t"
10866 "CALL d2i_wrapper\n"
10867 "fast:" %}
10868 ins_encode( Push_Reg_DPR(src), DPR2I_encoding(src) );
10869 ins_pipe( pipe_slow );
10870 %}
10872 // Convert a double to an int. If the double is a NAN, stuff a zero in instead.
10873 instruct convD2I_reg_reg( eAXRegI dst, eDXRegI tmp, regD src, eFlagsReg cr ) %{
10874 predicate(UseSSE>=2);
10875 match(Set dst (ConvD2I src));
10876 effect( KILL tmp, KILL cr );
10877 format %{ "CVTTSD2SI $dst, $src\n\t"
10878 "CMP $dst,0x80000000\n\t"
10879 "JNE,s fast\n\t"
10880 "SUB ESP, 8\n\t"
10881 "MOVSD [ESP], $src\n\t"
10882 "FLD_D [ESP]\n\t"
10883 "ADD ESP, 8\n\t"
10884 "CALL d2i_wrapper\n"
10885 "fast:" %}
10886 ins_encode %{
10887 Label fast;
10888 __ cvttsd2sil($dst$$Register, $src$$XMMRegister);
10889 __ cmpl($dst$$Register, 0x80000000);
10890 __ jccb(Assembler::notEqual, fast);
10891 __ subptr(rsp, 8);
10892 __ movdbl(Address(rsp, 0), $src$$XMMRegister);
10893 __ fld_d(Address(rsp, 0));
10894 __ addptr(rsp, 8);
10895 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, StubRoutines::d2i_wrapper())));
10896 __ bind(fast);
10897 %}
10898 ins_pipe( pipe_slow );
10899 %}
10901 instruct convDPR2L_reg_reg( eADXRegL dst, regDPR src, eFlagsReg cr ) %{
10902 predicate(UseSSE<=1);
10903 match(Set dst (ConvD2L src));
10904 effect( KILL cr );
10905 format %{ "FLD $src\t# Convert double to long\n\t"
10906 "FLDCW trunc mode\n\t"
10907 "SUB ESP,8\n\t"
10908 "FISTp [ESP + #0]\n\t"
10909 "FLDCW std/24-bit mode\n\t"
10910 "POP EAX\n\t"
10911 "POP EDX\n\t"
10912 "CMP EDX,0x80000000\n\t"
10913 "JNE,s fast\n\t"
10914 "TEST EAX,EAX\n\t"
10915 "JNE,s fast\n\t"
10916 "FLD $src\n\t"
10917 "CALL d2l_wrapper\n"
10918 "fast:" %}
10919 ins_encode( Push_Reg_DPR(src), DPR2L_encoding(src) );
10920 ins_pipe( pipe_slow );
10921 %}
10923 // XMM lacks a float/double->long conversion, so use the old FPU stack.
10924 instruct convD2L_reg_reg( eADXRegL dst, regD src, eFlagsReg cr ) %{
10925 predicate (UseSSE>=2);
10926 match(Set dst (ConvD2L src));
10927 effect( KILL cr );
10928 format %{ "SUB ESP,8\t# Convert double to long\n\t"
10929 "MOVSD [ESP],$src\n\t"
10930 "FLD_D [ESP]\n\t"
10931 "FLDCW trunc mode\n\t"
10932 "FISTp [ESP + #0]\n\t"
10933 "FLDCW std/24-bit mode\n\t"
10934 "POP EAX\n\t"
10935 "POP EDX\n\t"
10936 "CMP EDX,0x80000000\n\t"
10937 "JNE,s fast\n\t"
10938 "TEST EAX,EAX\n\t"
10939 "JNE,s fast\n\t"
10940 "SUB ESP,8\n\t"
10941 "MOVSD [ESP],$src\n\t"
10942 "FLD_D [ESP]\n\t"
10943 "ADD ESP,8\n\t"
10944 "CALL d2l_wrapper\n"
10945 "fast:" %}
10946 ins_encode %{
10947 Label fast;
10948 __ subptr(rsp, 8);
10949 __ movdbl(Address(rsp, 0), $src$$XMMRegister);
10950 __ fld_d(Address(rsp, 0));
10951 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_trunc()));
10952 __ fistp_d(Address(rsp, 0));
10953 // Restore the rounding mode, mask the exception
10954 if (Compile::current()->in_24_bit_fp_mode()) {
10955 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
10956 } else {
10957 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
10958 }
10959 // Load the converted long, adjust CPU stack
10960 __ pop(rax);
10961 __ pop(rdx);
10962 __ cmpl(rdx, 0x80000000);
10963 __ jccb(Assembler::notEqual, fast);
10964 __ testl(rax, rax);
10965 __ jccb(Assembler::notEqual, fast);
10966 __ subptr(rsp, 8);
10967 __ movdbl(Address(rsp, 0), $src$$XMMRegister);
10968 __ fld_d(Address(rsp, 0));
10969 __ addptr(rsp, 8);
10970 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, StubRoutines::d2l_wrapper())));
10971 __ bind(fast);
10972 %}
10973 ins_pipe( pipe_slow );
10974 %}
10976 // Convert a double to an int. Java semantics require we do complex
10977 // manglations in the corner cases. So we set the rounding mode to
10978 // 'zero', store the darned double down as an int, and reset the
10979 // rounding mode to 'nearest'. The hardware stores a flag value down
10980 // if we would overflow or converted a NAN; we check for this and
10981 // and go the slow path if needed.
10982 instruct convFPR2I_reg_reg(eAXRegI dst, eDXRegI tmp, regFPR src, eFlagsReg cr ) %{
10983 predicate(UseSSE==0);
10984 match(Set dst (ConvF2I src));
10985 effect( KILL tmp, KILL cr );
10986 format %{ "FLD $src\t# Convert float to int \n\t"
10987 "FLDCW trunc mode\n\t"
10988 "SUB ESP,4\n\t"
10989 "FISTp [ESP + #0]\n\t"
10990 "FLDCW std/24-bit mode\n\t"
10991 "POP EAX\n\t"
10992 "CMP EAX,0x80000000\n\t"
10993 "JNE,s fast\n\t"
10994 "FLD $src\n\t"
10995 "CALL d2i_wrapper\n"
10996 "fast:" %}
10997 // DPR2I_encoding works for FPR2I
10998 ins_encode( Push_Reg_FPR(src), DPR2I_encoding(src) );
10999 ins_pipe( pipe_slow );
11000 %}
11002 // Convert a float in xmm to an int reg.
11003 instruct convF2I_reg(eAXRegI dst, eDXRegI tmp, regF src, eFlagsReg cr ) %{
11004 predicate(UseSSE>=1);
11005 match(Set dst (ConvF2I src));
11006 effect( KILL tmp, KILL cr );
11007 format %{ "CVTTSS2SI $dst, $src\n\t"
11008 "CMP $dst,0x80000000\n\t"
11009 "JNE,s fast\n\t"
11010 "SUB ESP, 4\n\t"
11011 "MOVSS [ESP], $src\n\t"
11012 "FLD [ESP]\n\t"
11013 "ADD ESP, 4\n\t"
11014 "CALL d2i_wrapper\n"
11015 "fast:" %}
11016 ins_encode %{
11017 Label fast;
11018 __ cvttss2sil($dst$$Register, $src$$XMMRegister);
11019 __ cmpl($dst$$Register, 0x80000000);
11020 __ jccb(Assembler::notEqual, fast);
11021 __ subptr(rsp, 4);
11022 __ movflt(Address(rsp, 0), $src$$XMMRegister);
11023 __ fld_s(Address(rsp, 0));
11024 __ addptr(rsp, 4);
11025 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, StubRoutines::d2i_wrapper())));
11026 __ bind(fast);
11027 %}
11028 ins_pipe( pipe_slow );
11029 %}
11031 instruct convFPR2L_reg_reg( eADXRegL dst, regFPR src, eFlagsReg cr ) %{
11032 predicate(UseSSE==0);
11033 match(Set dst (ConvF2L src));
11034 effect( KILL cr );
11035 format %{ "FLD $src\t# Convert float to long\n\t"
11036 "FLDCW trunc mode\n\t"
11037 "SUB ESP,8\n\t"
11038 "FISTp [ESP + #0]\n\t"
11039 "FLDCW std/24-bit mode\n\t"
11040 "POP EAX\n\t"
11041 "POP EDX\n\t"
11042 "CMP EDX,0x80000000\n\t"
11043 "JNE,s fast\n\t"
11044 "TEST EAX,EAX\n\t"
11045 "JNE,s fast\n\t"
11046 "FLD $src\n\t"
11047 "CALL d2l_wrapper\n"
11048 "fast:" %}
11049 // DPR2L_encoding works for FPR2L
11050 ins_encode( Push_Reg_FPR(src), DPR2L_encoding(src) );
11051 ins_pipe( pipe_slow );
11052 %}
11054 // XMM lacks a float/double->long conversion, so use the old FPU stack.
11055 instruct convF2L_reg_reg( eADXRegL dst, regF src, eFlagsReg cr ) %{
11056 predicate (UseSSE>=1);
11057 match(Set dst (ConvF2L src));
11058 effect( KILL cr );
11059 format %{ "SUB ESP,8\t# Convert float to long\n\t"
11060 "MOVSS [ESP],$src\n\t"
11061 "FLD_S [ESP]\n\t"
11062 "FLDCW trunc mode\n\t"
11063 "FISTp [ESP + #0]\n\t"
11064 "FLDCW std/24-bit mode\n\t"
11065 "POP EAX\n\t"
11066 "POP EDX\n\t"
11067 "CMP EDX,0x80000000\n\t"
11068 "JNE,s fast\n\t"
11069 "TEST EAX,EAX\n\t"
11070 "JNE,s fast\n\t"
11071 "SUB ESP,4\t# Convert float to long\n\t"
11072 "MOVSS [ESP],$src\n\t"
11073 "FLD_S [ESP]\n\t"
11074 "ADD ESP,4\n\t"
11075 "CALL d2l_wrapper\n"
11076 "fast:" %}
11077 ins_encode %{
11078 Label fast;
11079 __ subptr(rsp, 8);
11080 __ movflt(Address(rsp, 0), $src$$XMMRegister);
11081 __ fld_s(Address(rsp, 0));
11082 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_trunc()));
11083 __ fistp_d(Address(rsp, 0));
11084 // Restore the rounding mode, mask the exception
11085 if (Compile::current()->in_24_bit_fp_mode()) {
11086 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
11087 } else {
11088 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
11089 }
11090 // Load the converted long, adjust CPU stack
11091 __ pop(rax);
11092 __ pop(rdx);
11093 __ cmpl(rdx, 0x80000000);
11094 __ jccb(Assembler::notEqual, fast);
11095 __ testl(rax, rax);
11096 __ jccb(Assembler::notEqual, fast);
11097 __ subptr(rsp, 4);
11098 __ movflt(Address(rsp, 0), $src$$XMMRegister);
11099 __ fld_s(Address(rsp, 0));
11100 __ addptr(rsp, 4);
11101 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, StubRoutines::d2l_wrapper())));
11102 __ bind(fast);
11103 %}
11104 ins_pipe( pipe_slow );
11105 %}
11107 instruct convI2DPR_reg(regDPR dst, stackSlotI src) %{
11108 predicate( UseSSE<=1 );
11109 match(Set dst (ConvI2D src));
11110 format %{ "FILD $src\n\t"
11111 "FSTP $dst" %}
11112 opcode(0xDB, 0x0); /* DB /0 */
11113 ins_encode(Push_Mem_I(src), Pop_Reg_DPR(dst));
11114 ins_pipe( fpu_reg_mem );
11115 %}
11117 instruct convI2D_reg(regD dst, rRegI src) %{
11118 predicate( UseSSE>=2 && !UseXmmI2D );
11119 match(Set dst (ConvI2D src));
11120 format %{ "CVTSI2SD $dst,$src" %}
11121 ins_encode %{
11122 __ cvtsi2sdl ($dst$$XMMRegister, $src$$Register);
11123 %}
11124 ins_pipe( pipe_slow );
11125 %}
11127 instruct convI2D_mem(regD dst, memory mem) %{
11128 predicate( UseSSE>=2 );
11129 match(Set dst (ConvI2D (LoadI mem)));
11130 format %{ "CVTSI2SD $dst,$mem" %}
11131 ins_encode %{
11132 __ cvtsi2sdl ($dst$$XMMRegister, $mem$$Address);
11133 %}
11134 ins_pipe( pipe_slow );
11135 %}
11137 instruct convXI2D_reg(regD dst, rRegI src)
11138 %{
11139 predicate( UseSSE>=2 && UseXmmI2D );
11140 match(Set dst (ConvI2D src));
11142 format %{ "MOVD $dst,$src\n\t"
11143 "CVTDQ2PD $dst,$dst\t# i2d" %}
11144 ins_encode %{
11145 __ movdl($dst$$XMMRegister, $src$$Register);
11146 __ cvtdq2pd($dst$$XMMRegister, $dst$$XMMRegister);
11147 %}
11148 ins_pipe(pipe_slow); // XXX
11149 %}
11151 instruct convI2DPR_mem(regDPR dst, memory mem) %{
11152 predicate( UseSSE<=1 && !Compile::current()->select_24_bit_instr());
11153 match(Set dst (ConvI2D (LoadI mem)));
11154 format %{ "FILD $mem\n\t"
11155 "FSTP $dst" %}
11156 opcode(0xDB); /* DB /0 */
11157 ins_encode( OpcP, RMopc_Mem(0x00,mem),
11158 Pop_Reg_DPR(dst));
11159 ins_pipe( fpu_reg_mem );
11160 %}
11162 // Convert a byte to a float; no rounding step needed.
11163 instruct conv24I2FPR_reg(regFPR dst, stackSlotI src) %{
11164 predicate( UseSSE==0 && n->in(1)->Opcode() == Op_AndI && n->in(1)->in(2)->is_Con() && n->in(1)->in(2)->get_int() == 255 );
11165 match(Set dst (ConvI2F src));
11166 format %{ "FILD $src\n\t"
11167 "FSTP $dst" %}
11169 opcode(0xDB, 0x0); /* DB /0 */
11170 ins_encode(Push_Mem_I(src), Pop_Reg_FPR(dst));
11171 ins_pipe( fpu_reg_mem );
11172 %}
11174 // In 24-bit mode, force exponent rounding by storing back out
11175 instruct convI2FPR_SSF(stackSlotF dst, stackSlotI src) %{
11176 predicate( UseSSE==0 && Compile::current()->select_24_bit_instr());
11177 match(Set dst (ConvI2F src));
11178 ins_cost(200);
11179 format %{ "FILD $src\n\t"
11180 "FSTP_S $dst" %}
11181 opcode(0xDB, 0x0); /* DB /0 */
11182 ins_encode( Push_Mem_I(src),
11183 Pop_Mem_FPR(dst));
11184 ins_pipe( fpu_mem_mem );
11185 %}
11187 // In 24-bit mode, force exponent rounding by storing back out
11188 instruct convI2FPR_SSF_mem(stackSlotF dst, memory mem) %{
11189 predicate( UseSSE==0 && Compile::current()->select_24_bit_instr());
11190 match(Set dst (ConvI2F (LoadI mem)));
11191 ins_cost(200);
11192 format %{ "FILD $mem\n\t"
11193 "FSTP_S $dst" %}
11194 opcode(0xDB); /* DB /0 */
11195 ins_encode( OpcP, RMopc_Mem(0x00,mem),
11196 Pop_Mem_FPR(dst));
11197 ins_pipe( fpu_mem_mem );
11198 %}
11200 // This instruction does not round to 24-bits
11201 instruct convI2FPR_reg(regFPR dst, stackSlotI src) %{
11202 predicate( UseSSE==0 && !Compile::current()->select_24_bit_instr());
11203 match(Set dst (ConvI2F src));
11204 format %{ "FILD $src\n\t"
11205 "FSTP $dst" %}
11206 opcode(0xDB, 0x0); /* DB /0 */
11207 ins_encode( Push_Mem_I(src),
11208 Pop_Reg_FPR(dst));
11209 ins_pipe( fpu_reg_mem );
11210 %}
11212 // This instruction does not round to 24-bits
11213 instruct convI2FPR_mem(regFPR dst, memory mem) %{
11214 predicate( UseSSE==0 && !Compile::current()->select_24_bit_instr());
11215 match(Set dst (ConvI2F (LoadI mem)));
11216 format %{ "FILD $mem\n\t"
11217 "FSTP $dst" %}
11218 opcode(0xDB); /* DB /0 */
11219 ins_encode( OpcP, RMopc_Mem(0x00,mem),
11220 Pop_Reg_FPR(dst));
11221 ins_pipe( fpu_reg_mem );
11222 %}
11224 // Convert an int to a float in xmm; no rounding step needed.
11225 instruct convI2F_reg(regF dst, rRegI src) %{
11226 predicate( UseSSE==1 || UseSSE>=2 && !UseXmmI2F );
11227 match(Set dst (ConvI2F src));
11228 format %{ "CVTSI2SS $dst, $src" %}
11229 ins_encode %{
11230 __ cvtsi2ssl ($dst$$XMMRegister, $src$$Register);
11231 %}
11232 ins_pipe( pipe_slow );
11233 %}
11235 instruct convXI2F_reg(regF dst, rRegI src)
11236 %{
11237 predicate( UseSSE>=2 && UseXmmI2F );
11238 match(Set dst (ConvI2F src));
11240 format %{ "MOVD $dst,$src\n\t"
11241 "CVTDQ2PS $dst,$dst\t# i2f" %}
11242 ins_encode %{
11243 __ movdl($dst$$XMMRegister, $src$$Register);
11244 __ cvtdq2ps($dst$$XMMRegister, $dst$$XMMRegister);
11245 %}
11246 ins_pipe(pipe_slow); // XXX
11247 %}
11249 instruct convI2L_reg( eRegL dst, rRegI src, eFlagsReg cr) %{
11250 match(Set dst (ConvI2L src));
11251 effect(KILL cr);
11252 ins_cost(375);
11253 format %{ "MOV $dst.lo,$src\n\t"
11254 "MOV $dst.hi,$src\n\t"
11255 "SAR $dst.hi,31" %}
11256 ins_encode(convert_int_long(dst,src));
11257 ins_pipe( ialu_reg_reg_long );
11258 %}
11260 // Zero-extend convert int to long
11261 instruct convI2L_reg_zex(eRegL dst, rRegI src, immL_32bits mask, eFlagsReg flags ) %{
11262 match(Set dst (AndL (ConvI2L src) mask) );
11263 effect( KILL flags );
11264 ins_cost(250);
11265 format %{ "MOV $dst.lo,$src\n\t"
11266 "XOR $dst.hi,$dst.hi" %}
11267 opcode(0x33); // XOR
11268 ins_encode(enc_Copy(dst,src), OpcP, RegReg_Hi2(dst,dst) );
11269 ins_pipe( ialu_reg_reg_long );
11270 %}
11272 // Zero-extend long
11273 instruct zerox_long(eRegL dst, eRegL src, immL_32bits mask, eFlagsReg flags ) %{
11274 match(Set dst (AndL src mask) );
11275 effect( KILL flags );
11276 ins_cost(250);
11277 format %{ "MOV $dst.lo,$src.lo\n\t"
11278 "XOR $dst.hi,$dst.hi\n\t" %}
11279 opcode(0x33); // XOR
11280 ins_encode(enc_Copy(dst,src), OpcP, RegReg_Hi2(dst,dst) );
11281 ins_pipe( ialu_reg_reg_long );
11282 %}
11284 instruct convL2DPR_reg( stackSlotD dst, eRegL src, eFlagsReg cr) %{
11285 predicate (UseSSE<=1);
11286 match(Set dst (ConvL2D src));
11287 effect( KILL cr );
11288 format %{ "PUSH $src.hi\t# Convert long to double\n\t"
11289 "PUSH $src.lo\n\t"
11290 "FILD ST,[ESP + #0]\n\t"
11291 "ADD ESP,8\n\t"
11292 "FSTP_D $dst\t# D-round" %}
11293 opcode(0xDF, 0x5); /* DF /5 */
11294 ins_encode(convert_long_double(src), Pop_Mem_DPR(dst));
11295 ins_pipe( pipe_slow );
11296 %}
11298 instruct convL2D_reg( regD dst, eRegL src, eFlagsReg cr) %{
11299 predicate (UseSSE>=2);
11300 match(Set dst (ConvL2D src));
11301 effect( KILL cr );
11302 format %{ "PUSH $src.hi\t# Convert long to double\n\t"
11303 "PUSH $src.lo\n\t"
11304 "FILD_D [ESP]\n\t"
11305 "FSTP_D [ESP]\n\t"
11306 "MOVSD $dst,[ESP]\n\t"
11307 "ADD ESP,8" %}
11308 opcode(0xDF, 0x5); /* DF /5 */
11309 ins_encode(convert_long_double2(src), Push_ResultD(dst));
11310 ins_pipe( pipe_slow );
11311 %}
11313 instruct convL2F_reg( regF dst, eRegL src, eFlagsReg cr) %{
11314 predicate (UseSSE>=1);
11315 match(Set dst (ConvL2F src));
11316 effect( KILL cr );
11317 format %{ "PUSH $src.hi\t# Convert long to single float\n\t"
11318 "PUSH $src.lo\n\t"
11319 "FILD_D [ESP]\n\t"
11320 "FSTP_S [ESP]\n\t"
11321 "MOVSS $dst,[ESP]\n\t"
11322 "ADD ESP,8" %}
11323 opcode(0xDF, 0x5); /* DF /5 */
11324 ins_encode(convert_long_double2(src), Push_ResultF(dst,0x8));
11325 ins_pipe( pipe_slow );
11326 %}
11328 instruct convL2FPR_reg( stackSlotF dst, eRegL src, eFlagsReg cr) %{
11329 match(Set dst (ConvL2F src));
11330 effect( KILL cr );
11331 format %{ "PUSH $src.hi\t# Convert long to single float\n\t"
11332 "PUSH $src.lo\n\t"
11333 "FILD ST,[ESP + #0]\n\t"
11334 "ADD ESP,8\n\t"
11335 "FSTP_S $dst\t# F-round" %}
11336 opcode(0xDF, 0x5); /* DF /5 */
11337 ins_encode(convert_long_double(src), Pop_Mem_FPR(dst));
11338 ins_pipe( pipe_slow );
11339 %}
11341 instruct convL2I_reg( rRegI dst, eRegL src ) %{
11342 match(Set dst (ConvL2I src));
11343 effect( DEF dst, USE src );
11344 format %{ "MOV $dst,$src.lo" %}
11345 ins_encode(enc_CopyL_Lo(dst,src));
11346 ins_pipe( ialu_reg_reg );
11347 %}
11350 instruct MoveF2I_stack_reg(rRegI dst, stackSlotF src) %{
11351 match(Set dst (MoveF2I src));
11352 effect( DEF dst, USE src );
11353 ins_cost(100);
11354 format %{ "MOV $dst,$src\t# MoveF2I_stack_reg" %}
11355 ins_encode %{
11356 __ movl($dst$$Register, Address(rsp, $src$$disp));
11357 %}
11358 ins_pipe( ialu_reg_mem );
11359 %}
11361 instruct MoveFPR2I_reg_stack(stackSlotI dst, regFPR src) %{
11362 predicate(UseSSE==0);
11363 match(Set dst (MoveF2I src));
11364 effect( DEF dst, USE src );
11366 ins_cost(125);
11367 format %{ "FST_S $dst,$src\t# MoveF2I_reg_stack" %}
11368 ins_encode( Pop_Mem_Reg_FPR(dst, src) );
11369 ins_pipe( fpu_mem_reg );
11370 %}
11372 instruct MoveF2I_reg_stack_sse(stackSlotI dst, regF src) %{
11373 predicate(UseSSE>=1);
11374 match(Set dst (MoveF2I src));
11375 effect( DEF dst, USE src );
11377 ins_cost(95);
11378 format %{ "MOVSS $dst,$src\t# MoveF2I_reg_stack_sse" %}
11379 ins_encode %{
11380 __ movflt(Address(rsp, $dst$$disp), $src$$XMMRegister);
11381 %}
11382 ins_pipe( pipe_slow );
11383 %}
11385 instruct MoveF2I_reg_reg_sse(rRegI dst, regF src) %{
11386 predicate(UseSSE>=2);
11387 match(Set dst (MoveF2I src));
11388 effect( DEF dst, USE src );
11389 ins_cost(85);
11390 format %{ "MOVD $dst,$src\t# MoveF2I_reg_reg_sse" %}
11391 ins_encode %{
11392 __ movdl($dst$$Register, $src$$XMMRegister);
11393 %}
11394 ins_pipe( pipe_slow );
11395 %}
11397 instruct MoveI2F_reg_stack(stackSlotF dst, rRegI src) %{
11398 match(Set dst (MoveI2F src));
11399 effect( DEF dst, USE src );
11401 ins_cost(100);
11402 format %{ "MOV $dst,$src\t# MoveI2F_reg_stack" %}
11403 ins_encode %{
11404 __ movl(Address(rsp, $dst$$disp), $src$$Register);
11405 %}
11406 ins_pipe( ialu_mem_reg );
11407 %}
11410 instruct MoveI2FPR_stack_reg(regFPR dst, stackSlotI src) %{
11411 predicate(UseSSE==0);
11412 match(Set dst (MoveI2F src));
11413 effect(DEF dst, USE src);
11415 ins_cost(125);
11416 format %{ "FLD_S $src\n\t"
11417 "FSTP $dst\t# MoveI2F_stack_reg" %}
11418 opcode(0xD9); /* D9 /0, FLD m32real */
11419 ins_encode( OpcP, RMopc_Mem_no_oop(0x00,src),
11420 Pop_Reg_FPR(dst) );
11421 ins_pipe( fpu_reg_mem );
11422 %}
11424 instruct MoveI2F_stack_reg_sse(regF dst, stackSlotI src) %{
11425 predicate(UseSSE>=1);
11426 match(Set dst (MoveI2F src));
11427 effect( DEF dst, USE src );
11429 ins_cost(95);
11430 format %{ "MOVSS $dst,$src\t# MoveI2F_stack_reg_sse" %}
11431 ins_encode %{
11432 __ movflt($dst$$XMMRegister, Address(rsp, $src$$disp));
11433 %}
11434 ins_pipe( pipe_slow );
11435 %}
11437 instruct MoveI2F_reg_reg_sse(regF dst, rRegI src) %{
11438 predicate(UseSSE>=2);
11439 match(Set dst (MoveI2F src));
11440 effect( DEF dst, USE src );
11442 ins_cost(85);
11443 format %{ "MOVD $dst,$src\t# MoveI2F_reg_reg_sse" %}
11444 ins_encode %{
11445 __ movdl($dst$$XMMRegister, $src$$Register);
11446 %}
11447 ins_pipe( pipe_slow );
11448 %}
11450 instruct MoveD2L_stack_reg(eRegL dst, stackSlotD src) %{
11451 match(Set dst (MoveD2L src));
11452 effect(DEF dst, USE src);
11454 ins_cost(250);
11455 format %{ "MOV $dst.lo,$src\n\t"
11456 "MOV $dst.hi,$src+4\t# MoveD2L_stack_reg" %}
11457 opcode(0x8B, 0x8B);
11458 ins_encode( OpcP, RegMem(dst,src), OpcS, RegMem_Hi(dst,src));
11459 ins_pipe( ialu_mem_long_reg );
11460 %}
11462 instruct MoveDPR2L_reg_stack(stackSlotL dst, regDPR src) %{
11463 predicate(UseSSE<=1);
11464 match(Set dst (MoveD2L src));
11465 effect(DEF dst, USE src);
11467 ins_cost(125);
11468 format %{ "FST_D $dst,$src\t# MoveD2L_reg_stack" %}
11469 ins_encode( Pop_Mem_Reg_DPR(dst, src) );
11470 ins_pipe( fpu_mem_reg );
11471 %}
11473 instruct MoveD2L_reg_stack_sse(stackSlotL dst, regD src) %{
11474 predicate(UseSSE>=2);
11475 match(Set dst (MoveD2L src));
11476 effect(DEF dst, USE src);
11477 ins_cost(95);
11478 format %{ "MOVSD $dst,$src\t# MoveD2L_reg_stack_sse" %}
11479 ins_encode %{
11480 __ movdbl(Address(rsp, $dst$$disp), $src$$XMMRegister);
11481 %}
11482 ins_pipe( pipe_slow );
11483 %}
11485 instruct MoveD2L_reg_reg_sse(eRegL dst, regD src, regD tmp) %{
11486 predicate(UseSSE>=2);
11487 match(Set dst (MoveD2L src));
11488 effect(DEF dst, USE src, TEMP tmp);
11489 ins_cost(85);
11490 format %{ "MOVD $dst.lo,$src\n\t"
11491 "PSHUFLW $tmp,$src,0x4E\n\t"
11492 "MOVD $dst.hi,$tmp\t# MoveD2L_reg_reg_sse" %}
11493 ins_encode %{
11494 __ movdl($dst$$Register, $src$$XMMRegister);
11495 __ pshuflw($tmp$$XMMRegister, $src$$XMMRegister, 0x4e);
11496 __ movdl(HIGH_FROM_LOW($dst$$Register), $tmp$$XMMRegister);
11497 %}
11498 ins_pipe( pipe_slow );
11499 %}
11501 instruct MoveL2D_reg_stack(stackSlotD dst, eRegL src) %{
11502 match(Set dst (MoveL2D src));
11503 effect(DEF dst, USE src);
11505 ins_cost(200);
11506 format %{ "MOV $dst,$src.lo\n\t"
11507 "MOV $dst+4,$src.hi\t# MoveL2D_reg_stack" %}
11508 opcode(0x89, 0x89);
11509 ins_encode( OpcP, RegMem( src, dst ), OpcS, RegMem_Hi( src, dst ) );
11510 ins_pipe( ialu_mem_long_reg );
11511 %}
11514 instruct MoveL2DPR_stack_reg(regDPR dst, stackSlotL src) %{
11515 predicate(UseSSE<=1);
11516 match(Set dst (MoveL2D src));
11517 effect(DEF dst, USE src);
11518 ins_cost(125);
11520 format %{ "FLD_D $src\n\t"
11521 "FSTP $dst\t# MoveL2D_stack_reg" %}
11522 opcode(0xDD); /* DD /0, FLD m64real */
11523 ins_encode( OpcP, RMopc_Mem_no_oop(0x00,src),
11524 Pop_Reg_DPR(dst) );
11525 ins_pipe( fpu_reg_mem );
11526 %}
11529 instruct MoveL2D_stack_reg_sse(regD dst, stackSlotL src) %{
11530 predicate(UseSSE>=2 && UseXmmLoadAndClearUpper);
11531 match(Set dst (MoveL2D src));
11532 effect(DEF dst, USE src);
11534 ins_cost(95);
11535 format %{ "MOVSD $dst,$src\t# MoveL2D_stack_reg_sse" %}
11536 ins_encode %{
11537 __ movdbl($dst$$XMMRegister, Address(rsp, $src$$disp));
11538 %}
11539 ins_pipe( pipe_slow );
11540 %}
11542 instruct MoveL2D_stack_reg_sse_partial(regD dst, stackSlotL src) %{
11543 predicate(UseSSE>=2 && !UseXmmLoadAndClearUpper);
11544 match(Set dst (MoveL2D src));
11545 effect(DEF dst, USE src);
11547 ins_cost(95);
11548 format %{ "MOVLPD $dst,$src\t# MoveL2D_stack_reg_sse" %}
11549 ins_encode %{
11550 __ movdbl($dst$$XMMRegister, Address(rsp, $src$$disp));
11551 %}
11552 ins_pipe( pipe_slow );
11553 %}
11555 instruct MoveL2D_reg_reg_sse(regD dst, eRegL src, regD tmp) %{
11556 predicate(UseSSE>=2);
11557 match(Set dst (MoveL2D src));
11558 effect(TEMP dst, USE src, TEMP tmp);
11559 ins_cost(85);
11560 format %{ "MOVD $dst,$src.lo\n\t"
11561 "MOVD $tmp,$src.hi\n\t"
11562 "PUNPCKLDQ $dst,$tmp\t# MoveL2D_reg_reg_sse" %}
11563 ins_encode %{
11564 __ movdl($dst$$XMMRegister, $src$$Register);
11565 __ movdl($tmp$$XMMRegister, HIGH_FROM_LOW($src$$Register));
11566 __ punpckldq($dst$$XMMRegister, $tmp$$XMMRegister);
11567 %}
11568 ins_pipe( pipe_slow );
11569 %}
11572 // =======================================================================
11573 // fast clearing of an array
11574 instruct rep_stos(eCXRegI cnt, eDIRegP base, eAXRegI zero, Universe dummy, eFlagsReg cr) %{
11575 predicate(!UseFastStosb);
11576 match(Set dummy (ClearArray cnt base));
11577 effect(USE_KILL cnt, USE_KILL base, KILL zero, KILL cr);
11578 format %{ "XOR EAX,EAX\t# ClearArray:\n\t"
11579 "SHL ECX,1\t# Convert doublewords to words\n\t"
11580 "REP STOS\t# store EAX into [EDI++] while ECX--" %}
11581 ins_encode %{
11582 __ clear_mem($base$$Register, $cnt$$Register, $zero$$Register);
11583 %}
11584 ins_pipe( pipe_slow );
11585 %}
11587 instruct rep_fast_stosb(eCXRegI cnt, eDIRegP base, eAXRegI zero, Universe dummy, eFlagsReg cr) %{
11588 predicate(UseFastStosb);
11589 match(Set dummy (ClearArray cnt base));
11590 effect(USE_KILL cnt, USE_KILL base, KILL zero, KILL cr);
11591 format %{ "XOR EAX,EAX\t# ClearArray:\n\t"
11592 "SHL ECX,3\t# Convert doublewords to bytes\n\t"
11593 "REP STOSB\t# store EAX into [EDI++] while ECX--" %}
11594 ins_encode %{
11595 __ clear_mem($base$$Register, $cnt$$Register, $zero$$Register);
11596 %}
11597 ins_pipe( pipe_slow );
11598 %}
11600 instruct string_compare(eDIRegP str1, eCXRegI cnt1, eSIRegP str2, eDXRegI cnt2,
11601 eAXRegI result, regD tmp1, eFlagsReg cr) %{
11602 match(Set result (StrComp (Binary str1 cnt1) (Binary str2 cnt2)));
11603 effect(TEMP tmp1, USE_KILL str1, USE_KILL str2, USE_KILL cnt1, USE_KILL cnt2, KILL cr);
11605 format %{ "String Compare $str1,$cnt1,$str2,$cnt2 -> $result // KILL $tmp1" %}
11606 ins_encode %{
11607 __ string_compare($str1$$Register, $str2$$Register,
11608 $cnt1$$Register, $cnt2$$Register, $result$$Register,
11609 $tmp1$$XMMRegister);
11610 %}
11611 ins_pipe( pipe_slow );
11612 %}
11614 // fast string equals
11615 instruct string_equals(eDIRegP str1, eSIRegP str2, eCXRegI cnt, eAXRegI result,
11616 regD tmp1, regD tmp2, eBXRegI tmp3, eFlagsReg cr) %{
11617 match(Set result (StrEquals (Binary str1 str2) cnt));
11618 effect(TEMP tmp1, TEMP tmp2, USE_KILL str1, USE_KILL str2, USE_KILL cnt, KILL tmp3, KILL cr);
11620 format %{ "String Equals $str1,$str2,$cnt -> $result // KILL $tmp1, $tmp2, $tmp3" %}
11621 ins_encode %{
11622 __ char_arrays_equals(false, $str1$$Register, $str2$$Register,
11623 $cnt$$Register, $result$$Register, $tmp3$$Register,
11624 $tmp1$$XMMRegister, $tmp2$$XMMRegister);
11625 %}
11626 ins_pipe( pipe_slow );
11627 %}
11629 // fast search of substring with known size.
11630 instruct string_indexof_con(eDIRegP str1, eDXRegI cnt1, eSIRegP str2, immI int_cnt2,
11631 eBXRegI result, regD vec, eAXRegI cnt2, eCXRegI tmp, eFlagsReg cr) %{
11632 predicate(UseSSE42Intrinsics);
11633 match(Set result (StrIndexOf (Binary str1 cnt1) (Binary str2 int_cnt2)));
11634 effect(TEMP vec, USE_KILL str1, USE_KILL str2, USE_KILL cnt1, KILL cnt2, KILL tmp, KILL cr);
11636 format %{ "String IndexOf $str1,$cnt1,$str2,$int_cnt2 -> $result // KILL $vec, $cnt1, $cnt2, $tmp" %}
11637 ins_encode %{
11638 int icnt2 = (int)$int_cnt2$$constant;
11639 if (icnt2 >= 8) {
11640 // IndexOf for constant substrings with size >= 8 elements
11641 // which don't need to be loaded through stack.
11642 __ string_indexofC8($str1$$Register, $str2$$Register,
11643 $cnt1$$Register, $cnt2$$Register,
11644 icnt2, $result$$Register,
11645 $vec$$XMMRegister, $tmp$$Register);
11646 } else {
11647 // Small strings are loaded through stack if they cross page boundary.
11648 __ string_indexof($str1$$Register, $str2$$Register,
11649 $cnt1$$Register, $cnt2$$Register,
11650 icnt2, $result$$Register,
11651 $vec$$XMMRegister, $tmp$$Register);
11652 }
11653 %}
11654 ins_pipe( pipe_slow );
11655 %}
11657 instruct string_indexof(eDIRegP str1, eDXRegI cnt1, eSIRegP str2, eAXRegI cnt2,
11658 eBXRegI result, regD vec, eCXRegI tmp, eFlagsReg cr) %{
11659 predicate(UseSSE42Intrinsics);
11660 match(Set result (StrIndexOf (Binary str1 cnt1) (Binary str2 cnt2)));
11661 effect(TEMP vec, USE_KILL str1, USE_KILL str2, USE_KILL cnt1, USE_KILL cnt2, KILL tmp, KILL cr);
11663 format %{ "String IndexOf $str1,$cnt1,$str2,$cnt2 -> $result // KILL all" %}
11664 ins_encode %{
11665 __ string_indexof($str1$$Register, $str2$$Register,
11666 $cnt1$$Register, $cnt2$$Register,
11667 (-1), $result$$Register,
11668 $vec$$XMMRegister, $tmp$$Register);
11669 %}
11670 ins_pipe( pipe_slow );
11671 %}
11673 // fast array equals
11674 instruct array_equals(eDIRegP ary1, eSIRegP ary2, eAXRegI result,
11675 regD tmp1, regD tmp2, eCXRegI tmp3, eBXRegI tmp4, eFlagsReg cr)
11676 %{
11677 match(Set result (AryEq ary1 ary2));
11678 effect(TEMP tmp1, TEMP tmp2, USE_KILL ary1, USE_KILL ary2, KILL tmp3, KILL tmp4, KILL cr);
11679 //ins_cost(300);
11681 format %{ "Array Equals $ary1,$ary2 -> $result // KILL $tmp1, $tmp2, $tmp3, $tmp4" %}
11682 ins_encode %{
11683 __ char_arrays_equals(true, $ary1$$Register, $ary2$$Register,
11684 $tmp3$$Register, $result$$Register, $tmp4$$Register,
11685 $tmp1$$XMMRegister, $tmp2$$XMMRegister);
11686 %}
11687 ins_pipe( pipe_slow );
11688 %}
11690 //----------Control Flow Instructions------------------------------------------
11691 // Signed compare Instructions
11692 instruct compI_eReg(eFlagsReg cr, rRegI op1, rRegI op2) %{
11693 match(Set cr (CmpI op1 op2));
11694 effect( DEF cr, USE op1, USE op2 );
11695 format %{ "CMP $op1,$op2" %}
11696 opcode(0x3B); /* Opcode 3B /r */
11697 ins_encode( OpcP, RegReg( op1, op2) );
11698 ins_pipe( ialu_cr_reg_reg );
11699 %}
11701 instruct compI_eReg_imm(eFlagsReg cr, rRegI op1, immI op2) %{
11702 match(Set cr (CmpI op1 op2));
11703 effect( DEF cr, USE op1 );
11704 format %{ "CMP $op1,$op2" %}
11705 opcode(0x81,0x07); /* Opcode 81 /7 */
11706 // ins_encode( RegImm( op1, op2) ); /* Was CmpImm */
11707 ins_encode( OpcSErm( op1, op2 ), Con8or32( op2 ) );
11708 ins_pipe( ialu_cr_reg_imm );
11709 %}
11711 // Cisc-spilled version of cmpI_eReg
11712 instruct compI_eReg_mem(eFlagsReg cr, rRegI op1, memory op2) %{
11713 match(Set cr (CmpI op1 (LoadI op2)));
11715 format %{ "CMP $op1,$op2" %}
11716 ins_cost(500);
11717 opcode(0x3B); /* Opcode 3B /r */
11718 ins_encode( OpcP, RegMem( op1, op2) );
11719 ins_pipe( ialu_cr_reg_mem );
11720 %}
11722 instruct testI_reg( eFlagsReg cr, rRegI src, immI0 zero ) %{
11723 match(Set cr (CmpI src zero));
11724 effect( DEF cr, USE src );
11726 format %{ "TEST $src,$src" %}
11727 opcode(0x85);
11728 ins_encode( OpcP, RegReg( src, src ) );
11729 ins_pipe( ialu_cr_reg_imm );
11730 %}
11732 instruct testI_reg_imm( eFlagsReg cr, rRegI src, immI con, immI0 zero ) %{
11733 match(Set cr (CmpI (AndI src con) zero));
11735 format %{ "TEST $src,$con" %}
11736 opcode(0xF7,0x00);
11737 ins_encode( OpcP, RegOpc(src), Con32(con) );
11738 ins_pipe( ialu_cr_reg_imm );
11739 %}
11741 instruct testI_reg_mem( eFlagsReg cr, rRegI src, memory mem, immI0 zero ) %{
11742 match(Set cr (CmpI (AndI src mem) zero));
11744 format %{ "TEST $src,$mem" %}
11745 opcode(0x85);
11746 ins_encode( OpcP, RegMem( src, mem ) );
11747 ins_pipe( ialu_cr_reg_mem );
11748 %}
11750 // Unsigned compare Instructions; really, same as signed except they
11751 // produce an eFlagsRegU instead of eFlagsReg.
11752 instruct compU_eReg(eFlagsRegU cr, rRegI op1, rRegI op2) %{
11753 match(Set cr (CmpU op1 op2));
11755 format %{ "CMPu $op1,$op2" %}
11756 opcode(0x3B); /* Opcode 3B /r */
11757 ins_encode( OpcP, RegReg( op1, op2) );
11758 ins_pipe( ialu_cr_reg_reg );
11759 %}
11761 instruct compU_eReg_imm(eFlagsRegU cr, rRegI op1, immI op2) %{
11762 match(Set cr (CmpU op1 op2));
11764 format %{ "CMPu $op1,$op2" %}
11765 opcode(0x81,0x07); /* Opcode 81 /7 */
11766 ins_encode( OpcSErm( op1, op2 ), Con8or32( op2 ) );
11767 ins_pipe( ialu_cr_reg_imm );
11768 %}
11770 // // Cisc-spilled version of cmpU_eReg
11771 instruct compU_eReg_mem(eFlagsRegU cr, rRegI op1, memory op2) %{
11772 match(Set cr (CmpU op1 (LoadI op2)));
11774 format %{ "CMPu $op1,$op2" %}
11775 ins_cost(500);
11776 opcode(0x3B); /* Opcode 3B /r */
11777 ins_encode( OpcP, RegMem( op1, op2) );
11778 ins_pipe( ialu_cr_reg_mem );
11779 %}
11781 // // Cisc-spilled version of cmpU_eReg
11782 //instruct compU_mem_eReg(eFlagsRegU cr, memory op1, rRegI op2) %{
11783 // match(Set cr (CmpU (LoadI op1) op2));
11784 //
11785 // format %{ "CMPu $op1,$op2" %}
11786 // ins_cost(500);
11787 // opcode(0x39); /* Opcode 39 /r */
11788 // ins_encode( OpcP, RegMem( op1, op2) );
11789 //%}
11791 instruct testU_reg( eFlagsRegU cr, rRegI src, immI0 zero ) %{
11792 match(Set cr (CmpU src zero));
11794 format %{ "TESTu $src,$src" %}
11795 opcode(0x85);
11796 ins_encode( OpcP, RegReg( src, src ) );
11797 ins_pipe( ialu_cr_reg_imm );
11798 %}
11800 // Unsigned pointer compare Instructions
11801 instruct compP_eReg(eFlagsRegU cr, eRegP op1, eRegP op2) %{
11802 match(Set cr (CmpP op1 op2));
11804 format %{ "CMPu $op1,$op2" %}
11805 opcode(0x3B); /* Opcode 3B /r */
11806 ins_encode( OpcP, RegReg( op1, op2) );
11807 ins_pipe( ialu_cr_reg_reg );
11808 %}
11810 instruct compP_eReg_imm(eFlagsRegU cr, eRegP op1, immP op2) %{
11811 match(Set cr (CmpP op1 op2));
11813 format %{ "CMPu $op1,$op2" %}
11814 opcode(0x81,0x07); /* Opcode 81 /7 */
11815 ins_encode( OpcSErm( op1, op2 ), Con8or32( op2 ) );
11816 ins_pipe( ialu_cr_reg_imm );
11817 %}
11819 // // Cisc-spilled version of cmpP_eReg
11820 instruct compP_eReg_mem(eFlagsRegU cr, eRegP op1, memory op2) %{
11821 match(Set cr (CmpP op1 (LoadP op2)));
11823 format %{ "CMPu $op1,$op2" %}
11824 ins_cost(500);
11825 opcode(0x3B); /* Opcode 3B /r */
11826 ins_encode( OpcP, RegMem( op1, op2) );
11827 ins_pipe( ialu_cr_reg_mem );
11828 %}
11830 // // Cisc-spilled version of cmpP_eReg
11831 //instruct compP_mem_eReg(eFlagsRegU cr, memory op1, eRegP op2) %{
11832 // match(Set cr (CmpP (LoadP op1) op2));
11833 //
11834 // format %{ "CMPu $op1,$op2" %}
11835 // ins_cost(500);
11836 // opcode(0x39); /* Opcode 39 /r */
11837 // ins_encode( OpcP, RegMem( op1, op2) );
11838 //%}
11840 // Compare raw pointer (used in out-of-heap check).
11841 // Only works because non-oop pointers must be raw pointers
11842 // and raw pointers have no anti-dependencies.
11843 instruct compP_mem_eReg( eFlagsRegU cr, eRegP op1, memory op2 ) %{
11844 predicate( n->in(2)->in(2)->bottom_type()->reloc() == relocInfo::none );
11845 match(Set cr (CmpP op1 (LoadP op2)));
11847 format %{ "CMPu $op1,$op2" %}
11848 opcode(0x3B); /* Opcode 3B /r */
11849 ins_encode( OpcP, RegMem( op1, op2) );
11850 ins_pipe( ialu_cr_reg_mem );
11851 %}
11853 //
11854 // This will generate a signed flags result. This should be ok
11855 // since any compare to a zero should be eq/neq.
11856 instruct testP_reg( eFlagsReg cr, eRegP src, immP0 zero ) %{
11857 match(Set cr (CmpP src zero));
11859 format %{ "TEST $src,$src" %}
11860 opcode(0x85);
11861 ins_encode( OpcP, RegReg( src, src ) );
11862 ins_pipe( ialu_cr_reg_imm );
11863 %}
11865 // Cisc-spilled version of testP_reg
11866 // This will generate a signed flags result. This should be ok
11867 // since any compare to a zero should be eq/neq.
11868 instruct testP_Reg_mem( eFlagsReg cr, memory op, immI0 zero ) %{
11869 match(Set cr (CmpP (LoadP op) zero));
11871 format %{ "TEST $op,0xFFFFFFFF" %}
11872 ins_cost(500);
11873 opcode(0xF7); /* Opcode F7 /0 */
11874 ins_encode( OpcP, RMopc_Mem(0x00,op), Con_d32(0xFFFFFFFF) );
11875 ins_pipe( ialu_cr_reg_imm );
11876 %}
11878 // Yanked all unsigned pointer compare operations.
11879 // Pointer compares are done with CmpP which is already unsigned.
11881 //----------Max and Min--------------------------------------------------------
11882 // Min Instructions
11883 ////
11884 // *** Min and Max using the conditional move are slower than the
11885 // *** branch version on a Pentium III.
11886 // // Conditional move for min
11887 //instruct cmovI_reg_lt( rRegI op2, rRegI op1, eFlagsReg cr ) %{
11888 // effect( USE_DEF op2, USE op1, USE cr );
11889 // format %{ "CMOVlt $op2,$op1\t! min" %}
11890 // opcode(0x4C,0x0F);
11891 // ins_encode( OpcS, OpcP, RegReg( op2, op1 ) );
11892 // ins_pipe( pipe_cmov_reg );
11893 //%}
11894 //
11895 //// Min Register with Register (P6 version)
11896 //instruct minI_eReg_p6( rRegI op1, rRegI op2 ) %{
11897 // predicate(VM_Version::supports_cmov() );
11898 // match(Set op2 (MinI op1 op2));
11899 // ins_cost(200);
11900 // expand %{
11901 // eFlagsReg cr;
11902 // compI_eReg(cr,op1,op2);
11903 // cmovI_reg_lt(op2,op1,cr);
11904 // %}
11905 //%}
11907 // Min Register with Register (generic version)
11908 instruct minI_eReg(rRegI dst, rRegI src, eFlagsReg flags) %{
11909 match(Set dst (MinI dst src));
11910 effect(KILL flags);
11911 ins_cost(300);
11913 format %{ "MIN $dst,$src" %}
11914 opcode(0xCC);
11915 ins_encode( min_enc(dst,src) );
11916 ins_pipe( pipe_slow );
11917 %}
11919 // Max Register with Register
11920 // *** Min and Max using the conditional move are slower than the
11921 // *** branch version on a Pentium III.
11922 // // Conditional move for max
11923 //instruct cmovI_reg_gt( rRegI op2, rRegI op1, eFlagsReg cr ) %{
11924 // effect( USE_DEF op2, USE op1, USE cr );
11925 // format %{ "CMOVgt $op2,$op1\t! max" %}
11926 // opcode(0x4F,0x0F);
11927 // ins_encode( OpcS, OpcP, RegReg( op2, op1 ) );
11928 // ins_pipe( pipe_cmov_reg );
11929 //%}
11930 //
11931 // // Max Register with Register (P6 version)
11932 //instruct maxI_eReg_p6( rRegI op1, rRegI op2 ) %{
11933 // predicate(VM_Version::supports_cmov() );
11934 // match(Set op2 (MaxI op1 op2));
11935 // ins_cost(200);
11936 // expand %{
11937 // eFlagsReg cr;
11938 // compI_eReg(cr,op1,op2);
11939 // cmovI_reg_gt(op2,op1,cr);
11940 // %}
11941 //%}
11943 // Max Register with Register (generic version)
11944 instruct maxI_eReg(rRegI dst, rRegI src, eFlagsReg flags) %{
11945 match(Set dst (MaxI dst src));
11946 effect(KILL flags);
11947 ins_cost(300);
11949 format %{ "MAX $dst,$src" %}
11950 opcode(0xCC);
11951 ins_encode( max_enc(dst,src) );
11952 ins_pipe( pipe_slow );
11953 %}
11955 // ============================================================================
11956 // Counted Loop limit node which represents exact final iterator value.
11957 // Note: the resulting value should fit into integer range since
11958 // counted loops have limit check on overflow.
11959 instruct loopLimit_eReg(eAXRegI limit, nadxRegI init, immI stride, eDXRegI limit_hi, nadxRegI tmp, eFlagsReg flags) %{
11960 match(Set limit (LoopLimit (Binary init limit) stride));
11961 effect(TEMP limit_hi, TEMP tmp, KILL flags);
11962 ins_cost(300);
11964 format %{ "loopLimit $init,$limit,$stride # $limit = $init + $stride *( $limit - $init + $stride -1)/ $stride, kills $limit_hi" %}
11965 ins_encode %{
11966 int strd = (int)$stride$$constant;
11967 assert(strd != 1 && strd != -1, "sanity");
11968 int m1 = (strd > 0) ? 1 : -1;
11969 // Convert limit to long (EAX:EDX)
11970 __ cdql();
11971 // Convert init to long (init:tmp)
11972 __ movl($tmp$$Register, $init$$Register);
11973 __ sarl($tmp$$Register, 31);
11974 // $limit - $init
11975 __ subl($limit$$Register, $init$$Register);
11976 __ sbbl($limit_hi$$Register, $tmp$$Register);
11977 // + ($stride - 1)
11978 if (strd > 0) {
11979 __ addl($limit$$Register, (strd - 1));
11980 __ adcl($limit_hi$$Register, 0);
11981 __ movl($tmp$$Register, strd);
11982 } else {
11983 __ addl($limit$$Register, (strd + 1));
11984 __ adcl($limit_hi$$Register, -1);
11985 __ lneg($limit_hi$$Register, $limit$$Register);
11986 __ movl($tmp$$Register, -strd);
11987 }
11988 // signed devision: (EAX:EDX) / pos_stride
11989 __ idivl($tmp$$Register);
11990 if (strd < 0) {
11991 // restore sign
11992 __ negl($tmp$$Register);
11993 }
11994 // (EAX) * stride
11995 __ mull($tmp$$Register);
11996 // + init (ignore upper bits)
11997 __ addl($limit$$Register, $init$$Register);
11998 %}
11999 ins_pipe( pipe_slow );
12000 %}
12002 // ============================================================================
12003 // Branch Instructions
12004 // Jump Table
12005 instruct jumpXtnd(rRegI switch_val) %{
12006 match(Jump switch_val);
12007 ins_cost(350);
12008 format %{ "JMP [$constantaddress](,$switch_val,1)\n\t" %}
12009 ins_encode %{
12010 // Jump to Address(table_base + switch_reg)
12011 Address index(noreg, $switch_val$$Register, Address::times_1);
12012 __ jump(ArrayAddress($constantaddress, index));
12013 %}
12014 ins_pipe(pipe_jmp);
12015 %}
12017 // Jump Direct - Label defines a relative address from JMP+1
12018 instruct jmpDir(label labl) %{
12019 match(Goto);
12020 effect(USE labl);
12022 ins_cost(300);
12023 format %{ "JMP $labl" %}
12024 size(5);
12025 ins_encode %{
12026 Label* L = $labl$$label;
12027 __ jmp(*L, false); // Always long jump
12028 %}
12029 ins_pipe( pipe_jmp );
12030 %}
12032 // Jump Direct Conditional - Label defines a relative address from Jcc+1
12033 instruct jmpCon(cmpOp cop, eFlagsReg cr, label labl) %{
12034 match(If cop cr);
12035 effect(USE labl);
12037 ins_cost(300);
12038 format %{ "J$cop $labl" %}
12039 size(6);
12040 ins_encode %{
12041 Label* L = $labl$$label;
12042 __ jcc((Assembler::Condition)($cop$$cmpcode), *L, false); // Always long jump
12043 %}
12044 ins_pipe( pipe_jcc );
12045 %}
12047 // Jump Direct Conditional - Label defines a relative address from Jcc+1
12048 instruct jmpLoopEnd(cmpOp cop, eFlagsReg cr, label labl) %{
12049 match(CountedLoopEnd cop cr);
12050 effect(USE labl);
12052 ins_cost(300);
12053 format %{ "J$cop $labl\t# Loop end" %}
12054 size(6);
12055 ins_encode %{
12056 Label* L = $labl$$label;
12057 __ jcc((Assembler::Condition)($cop$$cmpcode), *L, false); // Always long jump
12058 %}
12059 ins_pipe( pipe_jcc );
12060 %}
12062 // Jump Direct Conditional - Label defines a relative address from Jcc+1
12063 instruct jmpLoopEndU(cmpOpU cop, eFlagsRegU cmp, label labl) %{
12064 match(CountedLoopEnd cop cmp);
12065 effect(USE labl);
12067 ins_cost(300);
12068 format %{ "J$cop,u $labl\t# Loop end" %}
12069 size(6);
12070 ins_encode %{
12071 Label* L = $labl$$label;
12072 __ jcc((Assembler::Condition)($cop$$cmpcode), *L, false); // Always long jump
12073 %}
12074 ins_pipe( pipe_jcc );
12075 %}
12077 instruct jmpLoopEndUCF(cmpOpUCF cop, eFlagsRegUCF cmp, label labl) %{
12078 match(CountedLoopEnd cop cmp);
12079 effect(USE labl);
12081 ins_cost(200);
12082 format %{ "J$cop,u $labl\t# Loop end" %}
12083 size(6);
12084 ins_encode %{
12085 Label* L = $labl$$label;
12086 __ jcc((Assembler::Condition)($cop$$cmpcode), *L, false); // Always long jump
12087 %}
12088 ins_pipe( pipe_jcc );
12089 %}
12091 // Jump Direct Conditional - using unsigned comparison
12092 instruct jmpConU(cmpOpU cop, eFlagsRegU cmp, label labl) %{
12093 match(If cop cmp);
12094 effect(USE labl);
12096 ins_cost(300);
12097 format %{ "J$cop,u $labl" %}
12098 size(6);
12099 ins_encode %{
12100 Label* L = $labl$$label;
12101 __ jcc((Assembler::Condition)($cop$$cmpcode), *L, false); // Always long jump
12102 %}
12103 ins_pipe(pipe_jcc);
12104 %}
12106 instruct jmpConUCF(cmpOpUCF cop, eFlagsRegUCF cmp, label labl) %{
12107 match(If cop cmp);
12108 effect(USE labl);
12110 ins_cost(200);
12111 format %{ "J$cop,u $labl" %}
12112 size(6);
12113 ins_encode %{
12114 Label* L = $labl$$label;
12115 __ jcc((Assembler::Condition)($cop$$cmpcode), *L, false); // Always long jump
12116 %}
12117 ins_pipe(pipe_jcc);
12118 %}
12120 instruct jmpConUCF2(cmpOpUCF2 cop, eFlagsRegUCF cmp, label labl) %{
12121 match(If cop cmp);
12122 effect(USE labl);
12124 ins_cost(200);
12125 format %{ $$template
12126 if ($cop$$cmpcode == Assembler::notEqual) {
12127 $$emit$$"JP,u $labl\n\t"
12128 $$emit$$"J$cop,u $labl"
12129 } else {
12130 $$emit$$"JP,u done\n\t"
12131 $$emit$$"J$cop,u $labl\n\t"
12132 $$emit$$"done:"
12133 }
12134 %}
12135 ins_encode %{
12136 Label* l = $labl$$label;
12137 if ($cop$$cmpcode == Assembler::notEqual) {
12138 __ jcc(Assembler::parity, *l, false);
12139 __ jcc(Assembler::notEqual, *l, false);
12140 } else if ($cop$$cmpcode == Assembler::equal) {
12141 Label done;
12142 __ jccb(Assembler::parity, done);
12143 __ jcc(Assembler::equal, *l, false);
12144 __ bind(done);
12145 } else {
12146 ShouldNotReachHere();
12147 }
12148 %}
12149 ins_pipe(pipe_jcc);
12150 %}
12152 // ============================================================================
12153 // The 2nd slow-half of a subtype check. Scan the subklass's 2ndary superklass
12154 // array for an instance of the superklass. Set a hidden internal cache on a
12155 // hit (cache is checked with exposed code in gen_subtype_check()). Return
12156 // NZ for a miss or zero for a hit. The encoding ALSO sets flags.
12157 instruct partialSubtypeCheck( eDIRegP result, eSIRegP sub, eAXRegP super, eCXRegI rcx, eFlagsReg cr ) %{
12158 match(Set result (PartialSubtypeCheck sub super));
12159 effect( KILL rcx, KILL cr );
12161 ins_cost(1100); // slightly larger than the next version
12162 format %{ "MOV EDI,[$sub+Klass::secondary_supers]\n\t"
12163 "MOV ECX,[EDI+ArrayKlass::length]\t# length to scan\n\t"
12164 "ADD EDI,ArrayKlass::base_offset\t# Skip to start of data; set NZ in case count is zero\n\t"
12165 "REPNE SCASD\t# Scan *EDI++ for a match with EAX while CX-- != 0\n\t"
12166 "JNE,s miss\t\t# Missed: EDI not-zero\n\t"
12167 "MOV [$sub+Klass::secondary_super_cache],$super\t# Hit: update cache\n\t"
12168 "XOR $result,$result\t\t Hit: EDI zero\n\t"
12169 "miss:\t" %}
12171 opcode(0x1); // Force a XOR of EDI
12172 ins_encode( enc_PartialSubtypeCheck() );
12173 ins_pipe( pipe_slow );
12174 %}
12176 instruct partialSubtypeCheck_vs_Zero( eFlagsReg cr, eSIRegP sub, eAXRegP super, eCXRegI rcx, eDIRegP result, immP0 zero ) %{
12177 match(Set cr (CmpP (PartialSubtypeCheck sub super) zero));
12178 effect( KILL rcx, KILL result );
12180 ins_cost(1000);
12181 format %{ "MOV EDI,[$sub+Klass::secondary_supers]\n\t"
12182 "MOV ECX,[EDI+ArrayKlass::length]\t# length to scan\n\t"
12183 "ADD EDI,ArrayKlass::base_offset\t# Skip to start of data; set NZ in case count is zero\n\t"
12184 "REPNE SCASD\t# Scan *EDI++ for a match with EAX while CX-- != 0\n\t"
12185 "JNE,s miss\t\t# Missed: flags NZ\n\t"
12186 "MOV [$sub+Klass::secondary_super_cache],$super\t# Hit: update cache, flags Z\n\t"
12187 "miss:\t" %}
12189 opcode(0x0); // No need to XOR EDI
12190 ins_encode( enc_PartialSubtypeCheck() );
12191 ins_pipe( pipe_slow );
12192 %}
12194 // ============================================================================
12195 // Branch Instructions -- short offset versions
12196 //
12197 // These instructions are used to replace jumps of a long offset (the default
12198 // match) with jumps of a shorter offset. These instructions are all tagged
12199 // with the ins_short_branch attribute, which causes the ADLC to suppress the
12200 // match rules in general matching. Instead, the ADLC generates a conversion
12201 // method in the MachNode which can be used to do in-place replacement of the
12202 // long variant with the shorter variant. The compiler will determine if a
12203 // branch can be taken by the is_short_branch_offset() predicate in the machine
12204 // specific code section of the file.
12206 // Jump Direct - Label defines a relative address from JMP+1
12207 instruct jmpDir_short(label labl) %{
12208 match(Goto);
12209 effect(USE labl);
12211 ins_cost(300);
12212 format %{ "JMP,s $labl" %}
12213 size(2);
12214 ins_encode %{
12215 Label* L = $labl$$label;
12216 __ jmpb(*L);
12217 %}
12218 ins_pipe( pipe_jmp );
12219 ins_short_branch(1);
12220 %}
12222 // Jump Direct Conditional - Label defines a relative address from Jcc+1
12223 instruct jmpCon_short(cmpOp cop, eFlagsReg cr, label labl) %{
12224 match(If cop cr);
12225 effect(USE labl);
12227 ins_cost(300);
12228 format %{ "J$cop,s $labl" %}
12229 size(2);
12230 ins_encode %{
12231 Label* L = $labl$$label;
12232 __ jccb((Assembler::Condition)($cop$$cmpcode), *L);
12233 %}
12234 ins_pipe( pipe_jcc );
12235 ins_short_branch(1);
12236 %}
12238 // Jump Direct Conditional - Label defines a relative address from Jcc+1
12239 instruct jmpLoopEnd_short(cmpOp cop, eFlagsReg cr, label labl) %{
12240 match(CountedLoopEnd cop cr);
12241 effect(USE labl);
12243 ins_cost(300);
12244 format %{ "J$cop,s $labl\t# Loop end" %}
12245 size(2);
12246 ins_encode %{
12247 Label* L = $labl$$label;
12248 __ jccb((Assembler::Condition)($cop$$cmpcode), *L);
12249 %}
12250 ins_pipe( pipe_jcc );
12251 ins_short_branch(1);
12252 %}
12254 // Jump Direct Conditional - Label defines a relative address from Jcc+1
12255 instruct jmpLoopEndU_short(cmpOpU cop, eFlagsRegU cmp, label labl) %{
12256 match(CountedLoopEnd cop cmp);
12257 effect(USE labl);
12259 ins_cost(300);
12260 format %{ "J$cop,us $labl\t# Loop end" %}
12261 size(2);
12262 ins_encode %{
12263 Label* L = $labl$$label;
12264 __ jccb((Assembler::Condition)($cop$$cmpcode), *L);
12265 %}
12266 ins_pipe( pipe_jcc );
12267 ins_short_branch(1);
12268 %}
12270 instruct jmpLoopEndUCF_short(cmpOpUCF cop, eFlagsRegUCF cmp, label labl) %{
12271 match(CountedLoopEnd cop cmp);
12272 effect(USE labl);
12274 ins_cost(300);
12275 format %{ "J$cop,us $labl\t# Loop end" %}
12276 size(2);
12277 ins_encode %{
12278 Label* L = $labl$$label;
12279 __ jccb((Assembler::Condition)($cop$$cmpcode), *L);
12280 %}
12281 ins_pipe( pipe_jcc );
12282 ins_short_branch(1);
12283 %}
12285 // Jump Direct Conditional - using unsigned comparison
12286 instruct jmpConU_short(cmpOpU cop, eFlagsRegU cmp, label labl) %{
12287 match(If cop cmp);
12288 effect(USE labl);
12290 ins_cost(300);
12291 format %{ "J$cop,us $labl" %}
12292 size(2);
12293 ins_encode %{
12294 Label* L = $labl$$label;
12295 __ jccb((Assembler::Condition)($cop$$cmpcode), *L);
12296 %}
12297 ins_pipe( pipe_jcc );
12298 ins_short_branch(1);
12299 %}
12301 instruct jmpConUCF_short(cmpOpUCF cop, eFlagsRegUCF cmp, label labl) %{
12302 match(If cop cmp);
12303 effect(USE labl);
12305 ins_cost(300);
12306 format %{ "J$cop,us $labl" %}
12307 size(2);
12308 ins_encode %{
12309 Label* L = $labl$$label;
12310 __ jccb((Assembler::Condition)($cop$$cmpcode), *L);
12311 %}
12312 ins_pipe( pipe_jcc );
12313 ins_short_branch(1);
12314 %}
12316 instruct jmpConUCF2_short(cmpOpUCF2 cop, eFlagsRegUCF cmp, label labl) %{
12317 match(If cop cmp);
12318 effect(USE labl);
12320 ins_cost(300);
12321 format %{ $$template
12322 if ($cop$$cmpcode == Assembler::notEqual) {
12323 $$emit$$"JP,u,s $labl\n\t"
12324 $$emit$$"J$cop,u,s $labl"
12325 } else {
12326 $$emit$$"JP,u,s done\n\t"
12327 $$emit$$"J$cop,u,s $labl\n\t"
12328 $$emit$$"done:"
12329 }
12330 %}
12331 size(4);
12332 ins_encode %{
12333 Label* l = $labl$$label;
12334 if ($cop$$cmpcode == Assembler::notEqual) {
12335 __ jccb(Assembler::parity, *l);
12336 __ jccb(Assembler::notEqual, *l);
12337 } else if ($cop$$cmpcode == Assembler::equal) {
12338 Label done;
12339 __ jccb(Assembler::parity, done);
12340 __ jccb(Assembler::equal, *l);
12341 __ bind(done);
12342 } else {
12343 ShouldNotReachHere();
12344 }
12345 %}
12346 ins_pipe(pipe_jcc);
12347 ins_short_branch(1);
12348 %}
12350 // ============================================================================
12351 // Long Compare
12352 //
12353 // Currently we hold longs in 2 registers. Comparing such values efficiently
12354 // is tricky. The flavor of compare used depends on whether we are testing
12355 // for LT, LE, or EQ. For a simple LT test we can check just the sign bit.
12356 // The GE test is the negated LT test. The LE test can be had by commuting
12357 // the operands (yielding a GE test) and then negating; negate again for the
12358 // GT test. The EQ test is done by ORcc'ing the high and low halves, and the
12359 // NE test is negated from that.
12361 // Due to a shortcoming in the ADLC, it mixes up expressions like:
12362 // (foo (CmpI (CmpL X Y) 0)) and (bar (CmpI (CmpL X 0L) 0)). Note the
12363 // difference between 'Y' and '0L'. The tree-matches for the CmpI sections
12364 // are collapsed internally in the ADLC's dfa-gen code. The match for
12365 // (CmpI (CmpL X Y) 0) is silently replaced with (CmpI (CmpL X 0L) 0) and the
12366 // foo match ends up with the wrong leaf. One fix is to not match both
12367 // reg-reg and reg-zero forms of long-compare. This is unfortunate because
12368 // both forms beat the trinary form of long-compare and both are very useful
12369 // on Intel which has so few registers.
12371 // Manifest a CmpL result in an integer register. Very painful.
12372 // This is the test to avoid.
12373 instruct cmpL3_reg_reg(eSIRegI dst, eRegL src1, eRegL src2, eFlagsReg flags ) %{
12374 match(Set dst (CmpL3 src1 src2));
12375 effect( KILL flags );
12376 ins_cost(1000);
12377 format %{ "XOR $dst,$dst\n\t"
12378 "CMP $src1.hi,$src2.hi\n\t"
12379 "JLT,s m_one\n\t"
12380 "JGT,s p_one\n\t"
12381 "CMP $src1.lo,$src2.lo\n\t"
12382 "JB,s m_one\n\t"
12383 "JEQ,s done\n"
12384 "p_one:\tINC $dst\n\t"
12385 "JMP,s done\n"
12386 "m_one:\tDEC $dst\n"
12387 "done:" %}
12388 ins_encode %{
12389 Label p_one, m_one, done;
12390 __ xorptr($dst$$Register, $dst$$Register);
12391 __ cmpl(HIGH_FROM_LOW($src1$$Register), HIGH_FROM_LOW($src2$$Register));
12392 __ jccb(Assembler::less, m_one);
12393 __ jccb(Assembler::greater, p_one);
12394 __ cmpl($src1$$Register, $src2$$Register);
12395 __ jccb(Assembler::below, m_one);
12396 __ jccb(Assembler::equal, done);
12397 __ bind(p_one);
12398 __ incrementl($dst$$Register);
12399 __ jmpb(done);
12400 __ bind(m_one);
12401 __ decrementl($dst$$Register);
12402 __ bind(done);
12403 %}
12404 ins_pipe( pipe_slow );
12405 %}
12407 //======
12408 // Manifest a CmpL result in the normal flags. Only good for LT or GE
12409 // compares. Can be used for LE or GT compares by reversing arguments.
12410 // NOT GOOD FOR EQ/NE tests.
12411 instruct cmpL_zero_flags_LTGE( flagsReg_long_LTGE flags, eRegL src, immL0 zero ) %{
12412 match( Set flags (CmpL src zero ));
12413 ins_cost(100);
12414 format %{ "TEST $src.hi,$src.hi" %}
12415 opcode(0x85);
12416 ins_encode( OpcP, RegReg_Hi2( src, src ) );
12417 ins_pipe( ialu_cr_reg_reg );
12418 %}
12420 // Manifest a CmpL result in the normal flags. Only good for LT or GE
12421 // compares. Can be used for LE or GT compares by reversing arguments.
12422 // NOT GOOD FOR EQ/NE tests.
12423 instruct cmpL_reg_flags_LTGE( flagsReg_long_LTGE flags, eRegL src1, eRegL src2, rRegI tmp ) %{
12424 match( Set flags (CmpL src1 src2 ));
12425 effect( TEMP tmp );
12426 ins_cost(300);
12427 format %{ "CMP $src1.lo,$src2.lo\t! Long compare; set flags for low bits\n\t"
12428 "MOV $tmp,$src1.hi\n\t"
12429 "SBB $tmp,$src2.hi\t! Compute flags for long compare" %}
12430 ins_encode( long_cmp_flags2( src1, src2, tmp ) );
12431 ins_pipe( ialu_cr_reg_reg );
12432 %}
12434 // Long compares reg < zero/req OR reg >= zero/req.
12435 // Just a wrapper for a normal branch, plus the predicate test.
12436 instruct cmpL_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, label labl) %{
12437 match(If cmp flags);
12438 effect(USE labl);
12439 predicate( _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::lt || _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ge );
12440 expand %{
12441 jmpCon(cmp,flags,labl); // JLT or JGE...
12442 %}
12443 %}
12445 // Compare 2 longs and CMOVE longs.
12446 instruct cmovLL_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, eRegL dst, eRegL src) %{
12447 match(Set dst (CMoveL (Binary cmp flags) (Binary dst src)));
12448 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 ));
12449 ins_cost(400);
12450 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
12451 "CMOV$cmp $dst.hi,$src.hi" %}
12452 opcode(0x0F,0x40);
12453 ins_encode( enc_cmov(cmp), RegReg_Lo2( dst, src ), enc_cmov(cmp), RegReg_Hi2( dst, src ) );
12454 ins_pipe( pipe_cmov_reg_long );
12455 %}
12457 instruct cmovLL_mem_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, eRegL dst, load_long_memory src) %{
12458 match(Set dst (CMoveL (Binary cmp flags) (Binary dst (LoadL src))));
12459 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 ));
12460 ins_cost(500);
12461 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
12462 "CMOV$cmp $dst.hi,$src.hi" %}
12463 opcode(0x0F,0x40);
12464 ins_encode( enc_cmov(cmp), RegMem(dst, src), enc_cmov(cmp), RegMem_Hi(dst, src) );
12465 ins_pipe( pipe_cmov_reg_long );
12466 %}
12468 // Compare 2 longs and CMOVE ints.
12469 instruct cmovII_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, rRegI dst, rRegI src) %{
12470 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 ));
12471 match(Set dst (CMoveI (Binary cmp flags) (Binary dst src)));
12472 ins_cost(200);
12473 format %{ "CMOV$cmp $dst,$src" %}
12474 opcode(0x0F,0x40);
12475 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
12476 ins_pipe( pipe_cmov_reg );
12477 %}
12479 instruct cmovII_mem_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, rRegI dst, memory src) %{
12480 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 ));
12481 match(Set dst (CMoveI (Binary cmp flags) (Binary dst (LoadI src))));
12482 ins_cost(250);
12483 format %{ "CMOV$cmp $dst,$src" %}
12484 opcode(0x0F,0x40);
12485 ins_encode( enc_cmov(cmp), RegMem( dst, src ) );
12486 ins_pipe( pipe_cmov_mem );
12487 %}
12489 // Compare 2 longs and CMOVE ints.
12490 instruct cmovPP_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, eRegP dst, eRegP src) %{
12491 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 ));
12492 match(Set dst (CMoveP (Binary cmp flags) (Binary dst src)));
12493 ins_cost(200);
12494 format %{ "CMOV$cmp $dst,$src" %}
12495 opcode(0x0F,0x40);
12496 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
12497 ins_pipe( pipe_cmov_reg );
12498 %}
12500 // Compare 2 longs and CMOVE doubles
12501 instruct cmovDDPR_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, regDPR dst, regDPR src) %{
12502 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 );
12503 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
12504 ins_cost(200);
12505 expand %{
12506 fcmovDPR_regS(cmp,flags,dst,src);
12507 %}
12508 %}
12510 // Compare 2 longs and CMOVE doubles
12511 instruct cmovDD_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, regD dst, regD src) %{
12512 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 );
12513 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
12514 ins_cost(200);
12515 expand %{
12516 fcmovD_regS(cmp,flags,dst,src);
12517 %}
12518 %}
12520 instruct cmovFFPR_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, regFPR dst, regFPR src) %{
12521 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 );
12522 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
12523 ins_cost(200);
12524 expand %{
12525 fcmovFPR_regS(cmp,flags,dst,src);
12526 %}
12527 %}
12529 instruct cmovFF_reg_LTGE(cmpOp cmp, flagsReg_long_LTGE flags, regF dst, regF src) %{
12530 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 );
12531 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
12532 ins_cost(200);
12533 expand %{
12534 fcmovF_regS(cmp,flags,dst,src);
12535 %}
12536 %}
12538 //======
12539 // Manifest a CmpL result in the normal flags. Only good for EQ/NE compares.
12540 instruct cmpL_zero_flags_EQNE( flagsReg_long_EQNE flags, eRegL src, immL0 zero, rRegI tmp ) %{
12541 match( Set flags (CmpL src zero ));
12542 effect(TEMP tmp);
12543 ins_cost(200);
12544 format %{ "MOV $tmp,$src.lo\n\t"
12545 "OR $tmp,$src.hi\t! Long is EQ/NE 0?" %}
12546 ins_encode( long_cmp_flags0( src, tmp ) );
12547 ins_pipe( ialu_reg_reg_long );
12548 %}
12550 // Manifest a CmpL result in the normal flags. Only good for EQ/NE compares.
12551 instruct cmpL_reg_flags_EQNE( flagsReg_long_EQNE flags, eRegL src1, eRegL src2 ) %{
12552 match( Set flags (CmpL src1 src2 ));
12553 ins_cost(200+300);
12554 format %{ "CMP $src1.lo,$src2.lo\t! Long compare; set flags for low bits\n\t"
12555 "JNE,s skip\n\t"
12556 "CMP $src1.hi,$src2.hi\n\t"
12557 "skip:\t" %}
12558 ins_encode( long_cmp_flags1( src1, src2 ) );
12559 ins_pipe( ialu_cr_reg_reg );
12560 %}
12562 // Long compare reg == zero/reg OR reg != zero/reg
12563 // Just a wrapper for a normal branch, plus the predicate test.
12564 instruct cmpL_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, label labl) %{
12565 match(If cmp flags);
12566 effect(USE labl);
12567 predicate( _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::eq || _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::ne );
12568 expand %{
12569 jmpCon(cmp,flags,labl); // JEQ or JNE...
12570 %}
12571 %}
12573 // Compare 2 longs and CMOVE longs.
12574 instruct cmovLL_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, eRegL dst, eRegL src) %{
12575 match(Set dst (CMoveL (Binary cmp flags) (Binary dst src)));
12576 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 ));
12577 ins_cost(400);
12578 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
12579 "CMOV$cmp $dst.hi,$src.hi" %}
12580 opcode(0x0F,0x40);
12581 ins_encode( enc_cmov(cmp), RegReg_Lo2( dst, src ), enc_cmov(cmp), RegReg_Hi2( dst, src ) );
12582 ins_pipe( pipe_cmov_reg_long );
12583 %}
12585 instruct cmovLL_mem_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, eRegL dst, load_long_memory src) %{
12586 match(Set dst (CMoveL (Binary cmp flags) (Binary dst (LoadL src))));
12587 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 ));
12588 ins_cost(500);
12589 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
12590 "CMOV$cmp $dst.hi,$src.hi" %}
12591 opcode(0x0F,0x40);
12592 ins_encode( enc_cmov(cmp), RegMem(dst, src), enc_cmov(cmp), RegMem_Hi(dst, src) );
12593 ins_pipe( pipe_cmov_reg_long );
12594 %}
12596 // Compare 2 longs and CMOVE ints.
12597 instruct cmovII_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, rRegI dst, rRegI src) %{
12598 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 ));
12599 match(Set dst (CMoveI (Binary cmp flags) (Binary dst src)));
12600 ins_cost(200);
12601 format %{ "CMOV$cmp $dst,$src" %}
12602 opcode(0x0F,0x40);
12603 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
12604 ins_pipe( pipe_cmov_reg );
12605 %}
12607 instruct cmovII_mem_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, rRegI dst, memory src) %{
12608 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 ));
12609 match(Set dst (CMoveI (Binary cmp flags) (Binary dst (LoadI src))));
12610 ins_cost(250);
12611 format %{ "CMOV$cmp $dst,$src" %}
12612 opcode(0x0F,0x40);
12613 ins_encode( enc_cmov(cmp), RegMem( dst, src ) );
12614 ins_pipe( pipe_cmov_mem );
12615 %}
12617 // Compare 2 longs and CMOVE ints.
12618 instruct cmovPP_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, eRegP dst, eRegP src) %{
12619 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 ));
12620 match(Set dst (CMoveP (Binary cmp flags) (Binary dst src)));
12621 ins_cost(200);
12622 format %{ "CMOV$cmp $dst,$src" %}
12623 opcode(0x0F,0x40);
12624 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
12625 ins_pipe( pipe_cmov_reg );
12626 %}
12628 // Compare 2 longs and CMOVE doubles
12629 instruct cmovDDPR_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, regDPR dst, regDPR src) %{
12630 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 );
12631 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
12632 ins_cost(200);
12633 expand %{
12634 fcmovDPR_regS(cmp,flags,dst,src);
12635 %}
12636 %}
12638 // Compare 2 longs and CMOVE doubles
12639 instruct cmovDD_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, regD dst, regD src) %{
12640 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 );
12641 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
12642 ins_cost(200);
12643 expand %{
12644 fcmovD_regS(cmp,flags,dst,src);
12645 %}
12646 %}
12648 instruct cmovFFPR_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, regFPR dst, regFPR src) %{
12649 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 );
12650 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
12651 ins_cost(200);
12652 expand %{
12653 fcmovFPR_regS(cmp,flags,dst,src);
12654 %}
12655 %}
12657 instruct cmovFF_reg_EQNE(cmpOp cmp, flagsReg_long_EQNE flags, regF dst, regF src) %{
12658 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 );
12659 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
12660 ins_cost(200);
12661 expand %{
12662 fcmovF_regS(cmp,flags,dst,src);
12663 %}
12664 %}
12666 //======
12667 // Manifest a CmpL result in the normal flags. Only good for LE or GT compares.
12668 // Same as cmpL_reg_flags_LEGT except must negate src
12669 instruct cmpL_zero_flags_LEGT( flagsReg_long_LEGT flags, eRegL src, immL0 zero, rRegI tmp ) %{
12670 match( Set flags (CmpL src zero ));
12671 effect( TEMP tmp );
12672 ins_cost(300);
12673 format %{ "XOR $tmp,$tmp\t# Long compare for -$src < 0, use commuted test\n\t"
12674 "CMP $tmp,$src.lo\n\t"
12675 "SBB $tmp,$src.hi\n\t" %}
12676 ins_encode( long_cmp_flags3(src, tmp) );
12677 ins_pipe( ialu_reg_reg_long );
12678 %}
12680 // Manifest a CmpL result in the normal flags. Only good for LE or GT compares.
12681 // Same as cmpL_reg_flags_LTGE except operands swapped. Swapping operands
12682 // requires a commuted test to get the same result.
12683 instruct cmpL_reg_flags_LEGT( flagsReg_long_LEGT flags, eRegL src1, eRegL src2, rRegI tmp ) %{
12684 match( Set flags (CmpL src1 src2 ));
12685 effect( TEMP tmp );
12686 ins_cost(300);
12687 format %{ "CMP $src2.lo,$src1.lo\t! Long compare, swapped operands, use with commuted test\n\t"
12688 "MOV $tmp,$src2.hi\n\t"
12689 "SBB $tmp,$src1.hi\t! Compute flags for long compare" %}
12690 ins_encode( long_cmp_flags2( src2, src1, tmp ) );
12691 ins_pipe( ialu_cr_reg_reg );
12692 %}
12694 // Long compares reg < zero/req OR reg >= zero/req.
12695 // Just a wrapper for a normal branch, plus the predicate test
12696 instruct cmpL_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, label labl) %{
12697 match(If cmp flags);
12698 effect(USE labl);
12699 predicate( _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::gt || _kids[0]->_leaf->as_Bool()->_test._test == BoolTest::le );
12700 ins_cost(300);
12701 expand %{
12702 jmpCon(cmp,flags,labl); // JGT or JLE...
12703 %}
12704 %}
12706 // Compare 2 longs and CMOVE longs.
12707 instruct cmovLL_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, eRegL dst, eRegL src) %{
12708 match(Set dst (CMoveL (Binary cmp flags) (Binary dst src)));
12709 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 ));
12710 ins_cost(400);
12711 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
12712 "CMOV$cmp $dst.hi,$src.hi" %}
12713 opcode(0x0F,0x40);
12714 ins_encode( enc_cmov(cmp), RegReg_Lo2( dst, src ), enc_cmov(cmp), RegReg_Hi2( dst, src ) );
12715 ins_pipe( pipe_cmov_reg_long );
12716 %}
12718 instruct cmovLL_mem_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, eRegL dst, load_long_memory src) %{
12719 match(Set dst (CMoveL (Binary cmp flags) (Binary dst (LoadL src))));
12720 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 ));
12721 ins_cost(500);
12722 format %{ "CMOV$cmp $dst.lo,$src.lo\n\t"
12723 "CMOV$cmp $dst.hi,$src.hi+4" %}
12724 opcode(0x0F,0x40);
12725 ins_encode( enc_cmov(cmp), RegMem(dst, src), enc_cmov(cmp), RegMem_Hi(dst, src) );
12726 ins_pipe( pipe_cmov_reg_long );
12727 %}
12729 // Compare 2 longs and CMOVE ints.
12730 instruct cmovII_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, rRegI dst, rRegI src) %{
12731 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 ));
12732 match(Set dst (CMoveI (Binary cmp flags) (Binary dst src)));
12733 ins_cost(200);
12734 format %{ "CMOV$cmp $dst,$src" %}
12735 opcode(0x0F,0x40);
12736 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
12737 ins_pipe( pipe_cmov_reg );
12738 %}
12740 instruct cmovII_mem_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, rRegI dst, memory src) %{
12741 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 ));
12742 match(Set dst (CMoveI (Binary cmp flags) (Binary dst (LoadI src))));
12743 ins_cost(250);
12744 format %{ "CMOV$cmp $dst,$src" %}
12745 opcode(0x0F,0x40);
12746 ins_encode( enc_cmov(cmp), RegMem( dst, src ) );
12747 ins_pipe( pipe_cmov_mem );
12748 %}
12750 // Compare 2 longs and CMOVE ptrs.
12751 instruct cmovPP_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, eRegP dst, eRegP src) %{
12752 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 ));
12753 match(Set dst (CMoveP (Binary cmp flags) (Binary dst src)));
12754 ins_cost(200);
12755 format %{ "CMOV$cmp $dst,$src" %}
12756 opcode(0x0F,0x40);
12757 ins_encode( enc_cmov(cmp), RegReg( dst, src ) );
12758 ins_pipe( pipe_cmov_reg );
12759 %}
12761 // Compare 2 longs and CMOVE doubles
12762 instruct cmovDDPR_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, regDPR dst, regDPR src) %{
12763 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 );
12764 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
12765 ins_cost(200);
12766 expand %{
12767 fcmovDPR_regS(cmp,flags,dst,src);
12768 %}
12769 %}
12771 // Compare 2 longs and CMOVE doubles
12772 instruct cmovDD_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, regD dst, regD src) %{
12773 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 );
12774 match(Set dst (CMoveD (Binary cmp flags) (Binary dst src)));
12775 ins_cost(200);
12776 expand %{
12777 fcmovD_regS(cmp,flags,dst,src);
12778 %}
12779 %}
12781 instruct cmovFFPR_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, regFPR dst, regFPR src) %{
12782 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 );
12783 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
12784 ins_cost(200);
12785 expand %{
12786 fcmovFPR_regS(cmp,flags,dst,src);
12787 %}
12788 %}
12791 instruct cmovFF_reg_LEGT(cmpOp_commute cmp, flagsReg_long_LEGT flags, regF dst, regF src) %{
12792 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 );
12793 match(Set dst (CMoveF (Binary cmp flags) (Binary dst src)));
12794 ins_cost(200);
12795 expand %{
12796 fcmovF_regS(cmp,flags,dst,src);
12797 %}
12798 %}
12801 // ============================================================================
12802 // Procedure Call/Return Instructions
12803 // Call Java Static Instruction
12804 // Note: If this code changes, the corresponding ret_addr_offset() and
12805 // compute_padding() functions will have to be adjusted.
12806 instruct CallStaticJavaDirect(method meth) %{
12807 match(CallStaticJava);
12808 predicate(! ((CallStaticJavaNode*)n)->is_method_handle_invoke());
12809 effect(USE meth);
12811 ins_cost(300);
12812 format %{ "CALL,static " %}
12813 opcode(0xE8); /* E8 cd */
12814 ins_encode( pre_call_FPU,
12815 Java_Static_Call( meth ),
12816 call_epilog,
12817 post_call_FPU );
12818 ins_pipe( pipe_slow );
12819 ins_alignment(4);
12820 %}
12822 // Call Java Static Instruction (method handle version)
12823 // Note: If this code changes, the corresponding ret_addr_offset() and
12824 // compute_padding() functions will have to be adjusted.
12825 instruct CallStaticJavaHandle(method meth, eBPRegP ebp_mh_SP_save) %{
12826 match(CallStaticJava);
12827 predicate(((CallStaticJavaNode*)n)->is_method_handle_invoke());
12828 effect(USE meth);
12829 // EBP is saved by all callees (for interpreter stack correction).
12830 // We use it here for a similar purpose, in {preserve,restore}_SP.
12832 ins_cost(300);
12833 format %{ "CALL,static/MethodHandle " %}
12834 opcode(0xE8); /* E8 cd */
12835 ins_encode( pre_call_FPU,
12836 preserve_SP,
12837 Java_Static_Call( meth ),
12838 restore_SP,
12839 call_epilog,
12840 post_call_FPU );
12841 ins_pipe( pipe_slow );
12842 ins_alignment(4);
12843 %}
12845 // Call Java Dynamic Instruction
12846 // Note: If this code changes, the corresponding ret_addr_offset() and
12847 // compute_padding() functions will have to be adjusted.
12848 instruct CallDynamicJavaDirect(method meth) %{
12849 match(CallDynamicJava);
12850 effect(USE meth);
12852 ins_cost(300);
12853 format %{ "MOV EAX,(oop)-1\n\t"
12854 "CALL,dynamic" %}
12855 opcode(0xE8); /* E8 cd */
12856 ins_encode( pre_call_FPU,
12857 Java_Dynamic_Call( meth ),
12858 call_epilog,
12859 post_call_FPU );
12860 ins_pipe( pipe_slow );
12861 ins_alignment(4);
12862 %}
12864 // Call Runtime Instruction
12865 instruct CallRuntimeDirect(method meth) %{
12866 match(CallRuntime );
12867 effect(USE meth);
12869 ins_cost(300);
12870 format %{ "CALL,runtime " %}
12871 opcode(0xE8); /* E8 cd */
12872 // Use FFREEs to clear entries in float stack
12873 ins_encode( pre_call_FPU,
12874 FFree_Float_Stack_All,
12875 Java_To_Runtime( meth ),
12876 post_call_FPU );
12877 ins_pipe( pipe_slow );
12878 %}
12880 // Call runtime without safepoint
12881 instruct CallLeafDirect(method meth) %{
12882 match(CallLeaf);
12883 effect(USE meth);
12885 ins_cost(300);
12886 format %{ "CALL_LEAF,runtime " %}
12887 opcode(0xE8); /* E8 cd */
12888 ins_encode( pre_call_FPU,
12889 FFree_Float_Stack_All,
12890 Java_To_Runtime( meth ),
12891 Verify_FPU_For_Leaf, post_call_FPU );
12892 ins_pipe( pipe_slow );
12893 %}
12895 instruct CallLeafNoFPDirect(method meth) %{
12896 match(CallLeafNoFP);
12897 effect(USE meth);
12899 ins_cost(300);
12900 format %{ "CALL_LEAF_NOFP,runtime " %}
12901 opcode(0xE8); /* E8 cd */
12902 ins_encode(Java_To_Runtime(meth));
12903 ins_pipe( pipe_slow );
12904 %}
12907 // Return Instruction
12908 // Remove the return address & jump to it.
12909 instruct Ret() %{
12910 match(Return);
12911 format %{ "RET" %}
12912 opcode(0xC3);
12913 ins_encode(OpcP);
12914 ins_pipe( pipe_jmp );
12915 %}
12917 // Tail Call; Jump from runtime stub to Java code.
12918 // Also known as an 'interprocedural jump'.
12919 // Target of jump will eventually return to caller.
12920 // TailJump below removes the return address.
12921 instruct TailCalljmpInd(eRegP_no_EBP jump_target, eBXRegP method_oop) %{
12922 match(TailCall jump_target method_oop );
12923 ins_cost(300);
12924 format %{ "JMP $jump_target \t# EBX holds method oop" %}
12925 opcode(0xFF, 0x4); /* Opcode FF /4 */
12926 ins_encode( OpcP, RegOpc(jump_target) );
12927 ins_pipe( pipe_jmp );
12928 %}
12931 // Tail Jump; remove the return address; jump to target.
12932 // TailCall above leaves the return address around.
12933 instruct tailjmpInd(eRegP_no_EBP jump_target, eAXRegP ex_oop) %{
12934 match( TailJump jump_target ex_oop );
12935 ins_cost(300);
12936 format %{ "POP EDX\t# pop return address into dummy\n\t"
12937 "JMP $jump_target " %}
12938 opcode(0xFF, 0x4); /* Opcode FF /4 */
12939 ins_encode( enc_pop_rdx,
12940 OpcP, RegOpc(jump_target) );
12941 ins_pipe( pipe_jmp );
12942 %}
12944 // Create exception oop: created by stack-crawling runtime code.
12945 // Created exception is now available to this handler, and is setup
12946 // just prior to jumping to this handler. No code emitted.
12947 instruct CreateException( eAXRegP ex_oop )
12948 %{
12949 match(Set ex_oop (CreateEx));
12951 size(0);
12952 // use the following format syntax
12953 format %{ "# exception oop is in EAX; no code emitted" %}
12954 ins_encode();
12955 ins_pipe( empty );
12956 %}
12959 // Rethrow exception:
12960 // The exception oop will come in the first argument position.
12961 // Then JUMP (not call) to the rethrow stub code.
12962 instruct RethrowException()
12963 %{
12964 match(Rethrow);
12966 // use the following format syntax
12967 format %{ "JMP rethrow_stub" %}
12968 ins_encode(enc_rethrow);
12969 ins_pipe( pipe_jmp );
12970 %}
12972 // inlined locking and unlocking
12975 instruct cmpFastLock( eFlagsReg cr, eRegP object, eBXRegP box, eAXRegI tmp, eRegP scr) %{
12976 match( Set cr (FastLock object box) );
12977 effect( TEMP tmp, TEMP scr, USE_KILL box );
12978 ins_cost(300);
12979 format %{ "FASTLOCK $object,$box\t! kills $box,$tmp,$scr" %}
12980 ins_encode( Fast_Lock(object,box,tmp,scr) );
12981 ins_pipe( pipe_slow );
12982 %}
12984 instruct cmpFastUnlock( eFlagsReg cr, eRegP object, eAXRegP box, eRegP tmp ) %{
12985 match( Set cr (FastUnlock object box) );
12986 effect( TEMP tmp, USE_KILL box );
12987 ins_cost(300);
12988 format %{ "FASTUNLOCK $object,$box\t! kills $box,$tmp" %}
12989 ins_encode( Fast_Unlock(object,box,tmp) );
12990 ins_pipe( pipe_slow );
12991 %}
12995 // ============================================================================
12996 // Safepoint Instruction
12997 instruct safePoint_poll(eFlagsReg cr) %{
12998 match(SafePoint);
12999 effect(KILL cr);
13001 // TODO-FIXME: we currently poll at offset 0 of the safepoint polling page.
13002 // On SPARC that might be acceptable as we can generate the address with
13003 // just a sethi, saving an or. By polling at offset 0 we can end up
13004 // putting additional pressure on the index-0 in the D$. Because of
13005 // alignment (just like the situation at hand) the lower indices tend
13006 // to see more traffic. It'd be better to change the polling address
13007 // to offset 0 of the last $line in the polling page.
13009 format %{ "TSTL #polladdr,EAX\t! Safepoint: poll for GC" %}
13010 ins_cost(125);
13011 size(6) ;
13012 ins_encode( Safepoint_Poll() );
13013 ins_pipe( ialu_reg_mem );
13014 %}
13017 // ============================================================================
13018 // This name is KNOWN by the ADLC and cannot be changed.
13019 // The ADLC forces a 'TypeRawPtr::BOTTOM' output type
13020 // for this guy.
13021 instruct tlsLoadP(eRegP dst, eFlagsReg cr) %{
13022 match(Set dst (ThreadLocal));
13023 effect(DEF dst, KILL cr);
13025 format %{ "MOV $dst, Thread::current()" %}
13026 ins_encode %{
13027 Register dstReg = as_Register($dst$$reg);
13028 __ get_thread(dstReg);
13029 %}
13030 ins_pipe( ialu_reg_fat );
13031 %}
13035 //----------PEEPHOLE RULES-----------------------------------------------------
13036 // These must follow all instruction definitions as they use the names
13037 // defined in the instructions definitions.
13038 //
13039 // peepmatch ( root_instr_name [preceding_instruction]* );
13040 //
13041 // peepconstraint %{
13042 // (instruction_number.operand_name relational_op instruction_number.operand_name
13043 // [, ...] );
13044 // // instruction numbers are zero-based using left to right order in peepmatch
13045 //
13046 // peepreplace ( instr_name ( [instruction_number.operand_name]* ) );
13047 // // provide an instruction_number.operand_name for each operand that appears
13048 // // in the replacement instruction's match rule
13049 //
13050 // ---------VM FLAGS---------------------------------------------------------
13051 //
13052 // All peephole optimizations can be turned off using -XX:-OptoPeephole
13053 //
13054 // Each peephole rule is given an identifying number starting with zero and
13055 // increasing by one in the order seen by the parser. An individual peephole
13056 // can be enabled, and all others disabled, by using -XX:OptoPeepholeAt=#
13057 // on the command-line.
13058 //
13059 // ---------CURRENT LIMITATIONS----------------------------------------------
13060 //
13061 // Only match adjacent instructions in same basic block
13062 // Only equality constraints
13063 // Only constraints between operands, not (0.dest_reg == EAX_enc)
13064 // Only one replacement instruction
13065 //
13066 // ---------EXAMPLE----------------------------------------------------------
13067 //
13068 // // pertinent parts of existing instructions in architecture description
13069 // instruct movI(rRegI dst, rRegI src) %{
13070 // match(Set dst (CopyI src));
13071 // %}
13072 //
13073 // instruct incI_eReg(rRegI dst, immI1 src, eFlagsReg cr) %{
13074 // match(Set dst (AddI dst src));
13075 // effect(KILL cr);
13076 // %}
13077 //
13078 // // Change (inc mov) to lea
13079 // peephole %{
13080 // // increment preceeded by register-register move
13081 // peepmatch ( incI_eReg movI );
13082 // // require that the destination register of the increment
13083 // // match the destination register of the move
13084 // peepconstraint ( 0.dst == 1.dst );
13085 // // construct a replacement instruction that sets
13086 // // the destination to ( move's source register + one )
13087 // peepreplace ( leaI_eReg_immI( 0.dst 1.src 0.src ) );
13088 // %}
13089 //
13090 // Implementation no longer uses movX instructions since
13091 // machine-independent system no longer uses CopyX nodes.
13092 //
13093 // peephole %{
13094 // peepmatch ( incI_eReg movI );
13095 // peepconstraint ( 0.dst == 1.dst );
13096 // peepreplace ( leaI_eReg_immI( 0.dst 1.src 0.src ) );
13097 // %}
13098 //
13099 // peephole %{
13100 // peepmatch ( decI_eReg movI );
13101 // peepconstraint ( 0.dst == 1.dst );
13102 // peepreplace ( leaI_eReg_immI( 0.dst 1.src 0.src ) );
13103 // %}
13104 //
13105 // peephole %{
13106 // peepmatch ( addI_eReg_imm movI );
13107 // peepconstraint ( 0.dst == 1.dst );
13108 // peepreplace ( leaI_eReg_immI( 0.dst 1.src 0.src ) );
13109 // %}
13110 //
13111 // peephole %{
13112 // peepmatch ( addP_eReg_imm movP );
13113 // peepconstraint ( 0.dst == 1.dst );
13114 // peepreplace ( leaP_eReg_immI( 0.dst 1.src 0.src ) );
13115 // %}
13117 // // Change load of spilled value to only a spill
13118 // instruct storeI(memory mem, rRegI src) %{
13119 // match(Set mem (StoreI mem src));
13120 // %}
13121 //
13122 // instruct loadI(rRegI dst, memory mem) %{
13123 // match(Set dst (LoadI mem));
13124 // %}
13125 //
13126 peephole %{
13127 peepmatch ( loadI storeI );
13128 peepconstraint ( 1.src == 0.dst, 1.mem == 0.mem );
13129 peepreplace ( storeI( 1.mem 1.mem 1.src ) );
13130 %}
13132 //----------SMARTSPILL RULES---------------------------------------------------
13133 // These must follow all instruction definitions as they use the names
13134 // defined in the instructions definitions.