1 // Copyright 2013 the V8 project authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file.
5 #include "src/compiler/code-generator.h"
7 #include "src/compiler/code-generator-impl.h"
8 #include "src/compiler/gap-resolver.h"
9 #include "src/compiler/node-matchers.h"
10 #include "src/scopes.h"
11 #include "src/x64/assembler-x64.h"
12 #include "src/x64/macro-assembler-x64.h"
21 #define kScratchDoubleReg xmm0
24 // Adds X64 specific methods for decoding operands.
25 class X64OperandConverter : public InstructionOperandConverter {
27 X64OperandConverter(CodeGenerator* gen, Instruction* instr)
28 : InstructionOperandConverter(gen, instr) {}
30 Immediate InputImmediate(size_t index) {
31 return ToImmediate(instr_->InputAt(index));
34 Operand InputOperand(size_t index, int extra = 0) {
35 return ToOperand(instr_->InputAt(index), extra);
38 Operand OutputOperand() { return ToOperand(instr_->Output()); }
40 Immediate ToImmediate(InstructionOperand* operand) {
41 return Immediate(ToConstant(operand).ToInt32());
44 Operand ToOperand(InstructionOperand* op, int extra = 0) {
45 DCHECK(op->IsStackSlot() || op->IsDoubleStackSlot());
46 // The linkage computes where all spill slots are located.
47 FrameOffset offset = linkage()->GetFrameOffset(op->index(), frame(), extra);
48 return Operand(offset.from_stack_pointer() ? rsp : rbp, offset.offset());
51 static size_t NextOffset(size_t* offset) {
57 static ScaleFactor ScaleFor(AddressingMode one, AddressingMode mode) {
58 STATIC_ASSERT(0 == static_cast<int>(times_1));
59 STATIC_ASSERT(1 == static_cast<int>(times_2));
60 STATIC_ASSERT(2 == static_cast<int>(times_4));
61 STATIC_ASSERT(3 == static_cast<int>(times_8));
62 int scale = static_cast<int>(mode - one);
63 DCHECK(scale >= 0 && scale < 4);
64 return static_cast<ScaleFactor>(scale);
67 Operand MemoryOperand(size_t* offset) {
68 AddressingMode mode = AddressingModeField::decode(instr_->opcode());
71 Register base = InputRegister(NextOffset(offset));
73 return Operand(base, disp);
76 Register base = InputRegister(NextOffset(offset));
77 int32_t disp = InputInt32(NextOffset(offset));
78 return Operand(base, disp);
84 Register base = InputRegister(NextOffset(offset));
85 Register index = InputRegister(NextOffset(offset));
86 ScaleFactor scale = ScaleFor(kMode_MR1, mode);
88 return Operand(base, index, scale, disp);
94 Register base = InputRegister(NextOffset(offset));
95 Register index = InputRegister(NextOffset(offset));
96 ScaleFactor scale = ScaleFor(kMode_MR1I, mode);
97 int32_t disp = InputInt32(NextOffset(offset));
98 return Operand(base, index, scale, disp);
101 Register base = InputRegister(NextOffset(offset));
103 return Operand(base, disp);
106 UNREACHABLE(); // Should use kModeMR with more compact encoding instead
107 return Operand(no_reg, 0);
110 Register index = InputRegister(NextOffset(offset));
111 ScaleFactor scale = ScaleFor(kMode_M1, mode);
113 return Operand(index, scale, disp);
119 Register index = InputRegister(NextOffset(offset));
120 ScaleFactor scale = ScaleFor(kMode_M1I, mode);
121 int32_t disp = InputInt32(NextOffset(offset));
122 return Operand(index, scale, disp);
126 return Operand(no_reg, 0);
129 return Operand(no_reg, 0);
132 Operand MemoryOperand(size_t first_input = 0) {
133 return MemoryOperand(&first_input);
140 bool HasImmediateInput(Instruction* instr, size_t index) {
141 return instr->InputAt(index)->IsImmediate();
145 class OutOfLineLoadZero FINAL : public OutOfLineCode {
147 OutOfLineLoadZero(CodeGenerator* gen, Register result)
148 : OutOfLineCode(gen), result_(result) {}
150 void Generate() FINAL { __ xorl(result_, result_); }
153 Register const result_;
157 class OutOfLineLoadNaN FINAL : public OutOfLineCode {
159 OutOfLineLoadNaN(CodeGenerator* gen, XMMRegister result)
160 : OutOfLineCode(gen), result_(result) {}
162 void Generate() FINAL { __ pcmpeqd(result_, result_); }
165 XMMRegister const result_;
169 class OutOfLineTruncateDoubleToI FINAL : public OutOfLineCode {
171 OutOfLineTruncateDoubleToI(CodeGenerator* gen, Register result,
173 : OutOfLineCode(gen), result_(result), input_(input) {}
175 void Generate() FINAL {
176 __ subp(rsp, Immediate(kDoubleSize));
177 __ movsd(MemOperand(rsp, 0), input_);
178 __ SlowTruncateToI(result_, rsp, 0);
179 __ addp(rsp, Immediate(kDoubleSize));
183 Register const result_;
184 XMMRegister const input_;
190 #define ASSEMBLE_UNOP(asm_instr) \
192 if (instr->Output()->IsRegister()) { \
193 __ asm_instr(i.OutputRegister()); \
195 __ asm_instr(i.OutputOperand()); \
200 #define ASSEMBLE_BINOP(asm_instr) \
202 if (HasImmediateInput(instr, 1)) { \
203 if (instr->InputAt(0)->IsRegister()) { \
204 __ asm_instr(i.InputRegister(0), i.InputImmediate(1)); \
206 __ asm_instr(i.InputOperand(0), i.InputImmediate(1)); \
209 if (instr->InputAt(1)->IsRegister()) { \
210 __ asm_instr(i.InputRegister(0), i.InputRegister(1)); \
212 __ asm_instr(i.InputRegister(0), i.InputOperand(1)); \
218 #define ASSEMBLE_MULT(asm_instr) \
220 if (HasImmediateInput(instr, 1)) { \
221 if (instr->InputAt(0)->IsRegister()) { \
222 __ asm_instr(i.OutputRegister(), i.InputRegister(0), \
223 i.InputImmediate(1)); \
225 __ asm_instr(i.OutputRegister(), i.InputOperand(0), \
226 i.InputImmediate(1)); \
229 if (instr->InputAt(1)->IsRegister()) { \
230 __ asm_instr(i.OutputRegister(), i.InputRegister(1)); \
232 __ asm_instr(i.OutputRegister(), i.InputOperand(1)); \
238 #define ASSEMBLE_SHIFT(asm_instr, width) \
240 if (HasImmediateInput(instr, 1)) { \
241 if (instr->Output()->IsRegister()) { \
242 __ asm_instr(i.OutputRegister(), Immediate(i.InputInt##width(1))); \
244 __ asm_instr(i.OutputOperand(), Immediate(i.InputInt##width(1))); \
247 if (instr->Output()->IsRegister()) { \
248 __ asm_instr##_cl(i.OutputRegister()); \
250 __ asm_instr##_cl(i.OutputOperand()); \
256 #define ASSEMBLE_MOVX(asm_instr) \
258 if (instr->addressing_mode() != kMode_None) { \
259 __ asm_instr(i.OutputRegister(), i.MemoryOperand()); \
260 } else if (instr->InputAt(0)->IsRegister()) { \
261 __ asm_instr(i.OutputRegister(), i.InputRegister(0)); \
263 __ asm_instr(i.OutputRegister(), i.InputOperand(0)); \
268 #define ASSEMBLE_DOUBLE_BINOP(asm_instr) \
270 if (instr->InputAt(1)->IsDoubleRegister()) { \
271 __ asm_instr(i.InputDoubleRegister(0), i.InputDoubleRegister(1)); \
273 __ asm_instr(i.InputDoubleRegister(0), i.InputOperand(1)); \
278 #define ASSEMBLE_AVX_DOUBLE_BINOP(asm_instr) \
280 CpuFeatureScope avx_scope(masm(), AVX); \
281 if (instr->InputAt(1)->IsDoubleRegister()) { \
282 __ asm_instr(i.OutputDoubleRegister(), i.InputDoubleRegister(0), \
283 i.InputDoubleRegister(1)); \
285 __ asm_instr(i.OutputDoubleRegister(), i.InputDoubleRegister(0), \
286 i.InputOperand(1)); \
291 #define ASSEMBLE_CHECKED_LOAD_FLOAT(asm_instr) \
293 auto result = i.OutputDoubleRegister(); \
294 auto buffer = i.InputRegister(0); \
295 auto index1 = i.InputRegister(1); \
296 auto index2 = i.InputInt32(2); \
297 OutOfLineCode* ool; \
298 if (instr->InputAt(3)->IsRegister()) { \
299 auto length = i.InputRegister(3); \
300 DCHECK_EQ(0, index2); \
301 __ cmpl(index1, length); \
302 ool = new (zone()) OutOfLineLoadNaN(this, result); \
304 auto length = i.InputInt32(3); \
305 DCHECK_LE(index2, length); \
306 __ cmpq(index1, Immediate(length - index2)); \
307 class OutOfLineLoadFloat FINAL : public OutOfLineCode { \
309 OutOfLineLoadFloat(CodeGenerator* gen, XMMRegister result, \
310 Register buffer, Register index1, int32_t index2, \
312 : OutOfLineCode(gen), \
319 void Generate() FINAL { \
320 __ leal(kScratchRegister, Operand(index1_, index2_)); \
321 __ pcmpeqd(result_, result_); \
322 __ cmpl(kScratchRegister, Immediate(length_)); \
323 __ j(above_equal, exit()); \
324 __ asm_instr(result_, \
325 Operand(buffer_, kScratchRegister, times_1, 0)); \
329 XMMRegister const result_; \
330 Register const buffer_; \
331 Register const index1_; \
332 int32_t const index2_; \
333 int32_t const length_; \
336 OutOfLineLoadFloat(this, result, buffer, index1, index2, length); \
338 __ j(above_equal, ool->entry()); \
339 __ asm_instr(result, Operand(buffer, index1, times_1, index2)); \
340 __ bind(ool->exit()); \
344 #define ASSEMBLE_CHECKED_LOAD_INTEGER(asm_instr) \
346 auto result = i.OutputRegister(); \
347 auto buffer = i.InputRegister(0); \
348 auto index1 = i.InputRegister(1); \
349 auto index2 = i.InputInt32(2); \
350 OutOfLineCode* ool; \
351 if (instr->InputAt(3)->IsRegister()) { \
352 auto length = i.InputRegister(3); \
353 DCHECK_EQ(0, index2); \
354 __ cmpl(index1, length); \
355 ool = new (zone()) OutOfLineLoadZero(this, result); \
357 auto length = i.InputInt32(3); \
358 DCHECK_LE(index2, length); \
359 __ cmpq(index1, Immediate(length - index2)); \
360 class OutOfLineLoadInteger FINAL : public OutOfLineCode { \
362 OutOfLineLoadInteger(CodeGenerator* gen, Register result, \
363 Register buffer, Register index1, int32_t index2, \
365 : OutOfLineCode(gen), \
372 void Generate() FINAL { \
374 __ leal(kScratchRegister, Operand(index1_, index2_)); \
375 __ cmpl(kScratchRegister, Immediate(length_)); \
376 __ j(above_equal, &oob, Label::kNear); \
377 __ asm_instr(result_, \
378 Operand(buffer_, kScratchRegister, times_1, 0)); \
381 __ xorl(result_, result_); \
385 Register const result_; \
386 Register const buffer_; \
387 Register const index1_; \
388 int32_t const index2_; \
389 int32_t const length_; \
392 OutOfLineLoadInteger(this, result, buffer, index1, index2, length); \
394 __ j(above_equal, ool->entry()); \
395 __ asm_instr(result, Operand(buffer, index1, times_1, index2)); \
396 __ bind(ool->exit()); \
400 #define ASSEMBLE_CHECKED_STORE_FLOAT(asm_instr) \
402 auto buffer = i.InputRegister(0); \
403 auto index1 = i.InputRegister(1); \
404 auto index2 = i.InputInt32(2); \
405 auto value = i.InputDoubleRegister(4); \
406 if (instr->InputAt(3)->IsRegister()) { \
407 auto length = i.InputRegister(3); \
408 DCHECK_EQ(0, index2); \
410 __ cmpl(index1, length); \
411 __ j(above_equal, &done, Label::kNear); \
412 __ asm_instr(Operand(buffer, index1, times_1, index2), value); \
415 auto length = i.InputInt32(3); \
416 DCHECK_LE(index2, length); \
417 __ cmpq(index1, Immediate(length - index2)); \
418 class OutOfLineStoreFloat FINAL : public OutOfLineCode { \
420 OutOfLineStoreFloat(CodeGenerator* gen, Register buffer, \
421 Register index1, int32_t index2, int32_t length, \
423 : OutOfLineCode(gen), \
430 void Generate() FINAL { \
431 __ leal(kScratchRegister, Operand(index1_, index2_)); \
432 __ cmpl(kScratchRegister, Immediate(length_)); \
433 __ j(above_equal, exit()); \
434 __ asm_instr(Operand(buffer_, kScratchRegister, times_1, 0), \
439 Register const buffer_; \
440 Register const index1_; \
441 int32_t const index2_; \
442 int32_t const length_; \
443 XMMRegister const value_; \
445 auto ool = new (zone()) \
446 OutOfLineStoreFloat(this, buffer, index1, index2, length, value); \
447 __ j(above_equal, ool->entry()); \
448 __ asm_instr(Operand(buffer, index1, times_1, index2), value); \
449 __ bind(ool->exit()); \
454 #define ASSEMBLE_CHECKED_STORE_INTEGER_IMPL(asm_instr, Value) \
456 auto buffer = i.InputRegister(0); \
457 auto index1 = i.InputRegister(1); \
458 auto index2 = i.InputInt32(2); \
459 if (instr->InputAt(3)->IsRegister()) { \
460 auto length = i.InputRegister(3); \
461 DCHECK_EQ(0, index2); \
463 __ cmpl(index1, length); \
464 __ j(above_equal, &done, Label::kNear); \
465 __ asm_instr(Operand(buffer, index1, times_1, index2), value); \
468 auto length = i.InputInt32(3); \
469 DCHECK_LE(index2, length); \
470 __ cmpq(index1, Immediate(length - index2)); \
471 class OutOfLineStoreInteger FINAL : public OutOfLineCode { \
473 OutOfLineStoreInteger(CodeGenerator* gen, Register buffer, \
474 Register index1, int32_t index2, int32_t length, \
476 : OutOfLineCode(gen), \
483 void Generate() FINAL { \
484 __ leal(kScratchRegister, Operand(index1_, index2_)); \
485 __ cmpl(kScratchRegister, Immediate(length_)); \
486 __ j(above_equal, exit()); \
487 __ asm_instr(Operand(buffer_, kScratchRegister, times_1, 0), \
492 Register const buffer_; \
493 Register const index1_; \
494 int32_t const index2_; \
495 int32_t const length_; \
496 Value const value_; \
498 auto ool = new (zone()) \
499 OutOfLineStoreInteger(this, buffer, index1, index2, length, value); \
500 __ j(above_equal, ool->entry()); \
501 __ asm_instr(Operand(buffer, index1, times_1, index2), value); \
502 __ bind(ool->exit()); \
507 #define ASSEMBLE_CHECKED_STORE_INTEGER(asm_instr) \
509 if (instr->InputAt(4)->IsRegister()) { \
510 Register value = i.InputRegister(4); \
511 ASSEMBLE_CHECKED_STORE_INTEGER_IMPL(asm_instr, Register); \
513 Immediate value = i.InputImmediate(4); \
514 ASSEMBLE_CHECKED_STORE_INTEGER_IMPL(asm_instr, Immediate); \
519 // Assembles an instruction after register allocation, producing machine code.
520 void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
521 X64OperandConverter i(this, instr);
523 switch (ArchOpcodeField::decode(instr->opcode())) {
524 case kArchCallCodeObject: {
525 EnsureSpaceForLazyDeopt();
526 if (HasImmediateInput(instr, 0)) {
527 Handle<Code> code = Handle<Code>::cast(i.InputHeapObject(0));
528 __ Call(code, RelocInfo::CODE_TARGET);
530 Register reg = i.InputRegister(0);
531 int entry = Code::kHeaderSize - kHeapObjectTag;
532 __ Call(Operand(reg, entry));
534 RecordCallPosition(instr);
537 case kArchCallJSFunction: {
538 EnsureSpaceForLazyDeopt();
539 Register func = i.InputRegister(0);
540 if (FLAG_debug_code) {
541 // Check the function's context matches the context argument.
542 __ cmpp(rsi, FieldOperand(func, JSFunction::kContextOffset));
543 __ Assert(equal, kWrongFunctionContext);
545 __ Call(FieldOperand(func, JSFunction::kCodeEntryOffset));
546 RecordCallPosition(instr);
550 AssembleArchJump(i.InputRpo(0));
552 case kArchLookupSwitch:
553 AssembleArchLookupSwitch(instr);
555 case kArchTableSwitch:
556 AssembleArchTableSwitch(instr);
559 // don't emit code for nops.
561 case kArchDeoptimize: {
563 BuildTranslation(instr, -1, 0, OutputFrameStateCombine::Ignore());
564 AssembleDeoptimizerCall(deopt_state_id, Deoptimizer::EAGER);
570 case kArchStackPointer:
571 __ movq(i.OutputRegister(), rsp);
573 case kArchTruncateDoubleToI: {
574 auto result = i.OutputRegister();
575 auto input = i.InputDoubleRegister(0);
576 auto ool = new (zone()) OutOfLineTruncateDoubleToI(this, result, input);
577 __ cvttsd2siq(result, input);
578 __ cmpq(result, Immediate(1));
579 __ j(overflow, ool->entry());
580 __ bind(ool->exit());
584 ASSEMBLE_BINOP(addl);
587 ASSEMBLE_BINOP(addq);
590 ASSEMBLE_BINOP(subl);
593 ASSEMBLE_BINOP(subq);
596 ASSEMBLE_BINOP(andl);
599 ASSEMBLE_BINOP(andq);
602 ASSEMBLE_BINOP(cmpl);
605 ASSEMBLE_BINOP(cmpq);
608 ASSEMBLE_BINOP(testl);
611 ASSEMBLE_BINOP(testq);
614 ASSEMBLE_MULT(imull);
617 ASSEMBLE_MULT(imulq);
620 if (instr->InputAt(1)->IsRegister()) {
621 __ imull(i.InputRegister(1));
623 __ imull(i.InputOperand(1));
627 if (instr->InputAt(1)->IsRegister()) {
628 __ mull(i.InputRegister(1));
630 __ mull(i.InputOperand(1));
635 __ idivl(i.InputRegister(1));
639 __ idivq(i.InputRegister(1));
643 __ divl(i.InputRegister(1));
647 __ divq(i.InputRegister(1));
668 ASSEMBLE_BINOP(xorl);
671 ASSEMBLE_BINOP(xorq);
674 ASSEMBLE_SHIFT(shll, 5);
677 ASSEMBLE_SHIFT(shlq, 6);
680 ASSEMBLE_SHIFT(shrl, 5);
683 ASSEMBLE_SHIFT(shrq, 6);
686 ASSEMBLE_SHIFT(sarl, 5);
689 ASSEMBLE_SHIFT(sarq, 6);
692 ASSEMBLE_SHIFT(rorl, 5);
695 ASSEMBLE_SHIFT(rorq, 6);
698 ASSEMBLE_DOUBLE_BINOP(ucomisd);
701 ASSEMBLE_DOUBLE_BINOP(addsd);
704 ASSEMBLE_DOUBLE_BINOP(subsd);
707 ASSEMBLE_DOUBLE_BINOP(mulsd);
710 ASSEMBLE_DOUBLE_BINOP(divsd);
712 case kSSEFloat64Mod: {
713 __ subq(rsp, Immediate(kDoubleSize));
714 // Move values to st(0) and st(1).
715 __ movsd(Operand(rsp, 0), i.InputDoubleRegister(1));
716 __ fld_d(Operand(rsp, 0));
717 __ movsd(Operand(rsp, 0), i.InputDoubleRegister(0));
718 __ fld_d(Operand(rsp, 0));
719 // Loop while fprem isn't done.
722 // This instructions traps on all kinds inputs, but we are assuming the
723 // floating point control word is set to ignore them all.
725 // The following 2 instruction implicitly use rax.
727 if (CpuFeatures::IsSupported(SAHF)) {
728 CpuFeatureScope sahf_scope(masm(), SAHF);
731 __ shrl(rax, Immediate(8));
732 __ andl(rax, Immediate(0xFF));
736 __ j(parity_even, &mod_loop);
737 // Move output to stack and clean up.
739 __ fstp_d(Operand(rsp, 0));
740 __ movsd(i.OutputDoubleRegister(), Operand(rsp, 0));
741 __ addq(rsp, Immediate(kDoubleSize));
745 ASSEMBLE_DOUBLE_BINOP(maxsd);
748 ASSEMBLE_DOUBLE_BINOP(minsd);
750 case kSSEFloat64Sqrt:
751 if (instr->InputAt(0)->IsDoubleRegister()) {
752 __ sqrtsd(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
754 __ sqrtsd(i.OutputDoubleRegister(), i.InputOperand(0));
757 case kSSEFloat64Round: {
758 CpuFeatureScope sse_scope(masm(), SSE4_1);
759 RoundingMode const mode =
760 static_cast<RoundingMode>(MiscField::decode(instr->opcode()));
761 __ roundsd(i.OutputDoubleRegister(), i.InputDoubleRegister(0), mode);
765 if (instr->InputAt(0)->IsDoubleRegister()) {
766 __ cvtss2sd(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
768 __ cvtss2sd(i.OutputDoubleRegister(), i.InputOperand(0));
772 if (instr->InputAt(0)->IsDoubleRegister()) {
773 __ cvtsd2ss(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
775 __ cvtsd2ss(i.OutputDoubleRegister(), i.InputOperand(0));
778 case kSSEFloat64ToInt32:
779 if (instr->InputAt(0)->IsDoubleRegister()) {
780 __ cvttsd2si(i.OutputRegister(), i.InputDoubleRegister(0));
782 __ cvttsd2si(i.OutputRegister(), i.InputOperand(0));
785 case kSSEFloat64ToUint32: {
786 if (instr->InputAt(0)->IsDoubleRegister()) {
787 __ cvttsd2siq(i.OutputRegister(), i.InputDoubleRegister(0));
789 __ cvttsd2siq(i.OutputRegister(), i.InputOperand(0));
791 __ AssertZeroExtended(i.OutputRegister());
794 case kSSEInt32ToFloat64:
795 if (instr->InputAt(0)->IsRegister()) {
796 __ cvtlsi2sd(i.OutputDoubleRegister(), i.InputRegister(0));
798 __ cvtlsi2sd(i.OutputDoubleRegister(), i.InputOperand(0));
801 case kSSEUint32ToFloat64:
802 if (instr->InputAt(0)->IsRegister()) {
803 __ movl(kScratchRegister, i.InputRegister(0));
805 __ movl(kScratchRegister, i.InputOperand(0));
807 __ cvtqsi2sd(i.OutputDoubleRegister(), kScratchRegister);
809 case kSSEFloat64ExtractLowWord32:
810 if (instr->InputAt(0)->IsDoubleStackSlot()) {
811 __ movl(i.OutputRegister(), i.InputOperand(0));
813 __ movd(i.OutputRegister(), i.InputDoubleRegister(0));
816 case kSSEFloat64ExtractHighWord32:
817 if (instr->InputAt(0)->IsDoubleStackSlot()) {
818 __ movl(i.OutputRegister(), i.InputOperand(0, kDoubleSize / 2));
820 __ Pextrd(i.OutputRegister(), i.InputDoubleRegister(0), 1);
823 case kSSEFloat64InsertLowWord32:
824 if (instr->InputAt(1)->IsRegister()) {
825 __ Pinsrd(i.OutputDoubleRegister(), i.InputRegister(1), 0);
827 __ Pinsrd(i.OutputDoubleRegister(), i.InputOperand(1), 0);
830 case kSSEFloat64InsertHighWord32:
831 if (instr->InputAt(1)->IsRegister()) {
832 __ Pinsrd(i.OutputDoubleRegister(), i.InputRegister(1), 1);
834 __ Pinsrd(i.OutputDoubleRegister(), i.InputOperand(1), 1);
837 case kSSEFloat64LoadLowWord32:
838 if (instr->InputAt(0)->IsRegister()) {
839 __ movd(i.OutputDoubleRegister(), i.InputRegister(0));
841 __ movd(i.OutputDoubleRegister(), i.InputOperand(0));
845 ASSEMBLE_AVX_DOUBLE_BINOP(vaddsd);
848 ASSEMBLE_AVX_DOUBLE_BINOP(vsubsd);
851 ASSEMBLE_AVX_DOUBLE_BINOP(vmulsd);
854 ASSEMBLE_AVX_DOUBLE_BINOP(vdivsd);
857 ASSEMBLE_AVX_DOUBLE_BINOP(vmaxsd);
860 ASSEMBLE_AVX_DOUBLE_BINOP(vminsd);
863 ASSEMBLE_MOVX(movsxbl);
864 __ AssertZeroExtended(i.OutputRegister());
867 ASSEMBLE_MOVX(movzxbl);
868 __ AssertZeroExtended(i.OutputRegister());
872 Operand operand = i.MemoryOperand(&index);
873 if (HasImmediateInput(instr, index)) {
874 __ movb(operand, Immediate(i.InputInt8(index)));
876 __ movb(operand, i.InputRegister(index));
881 ASSEMBLE_MOVX(movsxwl);
882 __ AssertZeroExtended(i.OutputRegister());
885 ASSEMBLE_MOVX(movzxwl);
886 __ AssertZeroExtended(i.OutputRegister());
890 Operand operand = i.MemoryOperand(&index);
891 if (HasImmediateInput(instr, index)) {
892 __ movw(operand, Immediate(i.InputInt16(index)));
894 __ movw(operand, i.InputRegister(index));
899 if (instr->HasOutput()) {
900 if (instr->addressing_mode() == kMode_None) {
901 if (instr->InputAt(0)->IsRegister()) {
902 __ movl(i.OutputRegister(), i.InputRegister(0));
904 __ movl(i.OutputRegister(), i.InputOperand(0));
907 __ movl(i.OutputRegister(), i.MemoryOperand());
909 __ AssertZeroExtended(i.OutputRegister());
912 Operand operand = i.MemoryOperand(&index);
913 if (HasImmediateInput(instr, index)) {
914 __ movl(operand, i.InputImmediate(index));
916 __ movl(operand, i.InputRegister(index));
921 ASSEMBLE_MOVX(movsxlq);
924 if (instr->HasOutput()) {
925 __ movq(i.OutputRegister(), i.MemoryOperand());
928 Operand operand = i.MemoryOperand(&index);
929 if (HasImmediateInput(instr, index)) {
930 __ movq(operand, i.InputImmediate(index));
932 __ movq(operand, i.InputRegister(index));
937 if (instr->HasOutput()) {
938 __ movss(i.OutputDoubleRegister(), i.MemoryOperand());
941 Operand operand = i.MemoryOperand(&index);
942 __ movss(operand, i.InputDoubleRegister(index));
946 if (instr->HasOutput()) {
947 __ movsd(i.OutputDoubleRegister(), i.MemoryOperand());
950 Operand operand = i.MemoryOperand(&index);
951 __ movsd(operand, i.InputDoubleRegister(index));
955 AddressingMode mode = AddressingModeField::decode(instr->opcode());
956 // Shorten "leal" to "addl", "subl" or "shll" if the register allocation
957 // and addressing mode just happens to work out. The "addl"/"subl" forms
958 // in these cases are faster based on measurements.
959 if (i.InputRegister(0).is(i.OutputRegister())) {
960 if (mode == kMode_MRI) {
961 int32_t constant_summand = i.InputInt32(1);
962 if (constant_summand > 0) {
963 __ addl(i.OutputRegister(), Immediate(constant_summand));
964 } else if (constant_summand < 0) {
965 __ subl(i.OutputRegister(), Immediate(-constant_summand));
967 } else if (mode == kMode_MR1) {
968 if (i.InputRegister(1).is(i.OutputRegister())) {
969 __ shll(i.OutputRegister(), Immediate(1));
971 __ leal(i.OutputRegister(), i.MemoryOperand());
973 } else if (mode == kMode_M2) {
974 __ shll(i.OutputRegister(), Immediate(1));
975 } else if (mode == kMode_M4) {
976 __ shll(i.OutputRegister(), Immediate(2));
977 } else if (mode == kMode_M8) {
978 __ shll(i.OutputRegister(), Immediate(3));
980 __ leal(i.OutputRegister(), i.MemoryOperand());
983 __ leal(i.OutputRegister(), i.MemoryOperand());
985 __ AssertZeroExtended(i.OutputRegister());
989 __ leaq(i.OutputRegister(), i.MemoryOperand());
992 __ decl(i.OutputRegister());
995 __ incl(i.OutputRegister());
998 if (HasImmediateInput(instr, 0)) {
999 __ pushq(i.InputImmediate(0));
1001 if (instr->InputAt(0)->IsRegister()) {
1002 __ pushq(i.InputRegister(0));
1004 __ pushq(i.InputOperand(0));
1008 case kX64StoreWriteBarrier: {
1009 Register object = i.InputRegister(0);
1010 Register index = i.InputRegister(1);
1011 Register value = i.InputRegister(2);
1012 __ movq(Operand(object, index, times_1, 0), value);
1013 __ leaq(index, Operand(object, index, times_1, 0));
1014 SaveFPRegsMode mode =
1015 frame()->DidAllocateDoubleRegisters() ? kSaveFPRegs : kDontSaveFPRegs;
1016 __ RecordWrite(object, index, value, mode);
1019 case kCheckedLoadInt8:
1020 ASSEMBLE_CHECKED_LOAD_INTEGER(movsxbl);
1022 case kCheckedLoadUint8:
1023 ASSEMBLE_CHECKED_LOAD_INTEGER(movzxbl);
1025 case kCheckedLoadInt16:
1026 ASSEMBLE_CHECKED_LOAD_INTEGER(movsxwl);
1028 case kCheckedLoadUint16:
1029 ASSEMBLE_CHECKED_LOAD_INTEGER(movzxwl);
1031 case kCheckedLoadWord32:
1032 ASSEMBLE_CHECKED_LOAD_INTEGER(movl);
1034 case kCheckedLoadFloat32:
1035 ASSEMBLE_CHECKED_LOAD_FLOAT(movss);
1037 case kCheckedLoadFloat64:
1038 ASSEMBLE_CHECKED_LOAD_FLOAT(movsd);
1040 case kCheckedStoreWord8:
1041 ASSEMBLE_CHECKED_STORE_INTEGER(movb);
1043 case kCheckedStoreWord16:
1044 ASSEMBLE_CHECKED_STORE_INTEGER(movw);
1046 case kCheckedStoreWord32:
1047 ASSEMBLE_CHECKED_STORE_INTEGER(movl);
1049 case kCheckedStoreFloat32:
1050 ASSEMBLE_CHECKED_STORE_FLOAT(movss);
1052 case kCheckedStoreFloat64:
1053 ASSEMBLE_CHECKED_STORE_FLOAT(movsd);
1055 case kX64StackCheck:
1056 __ CompareRoot(rsp, Heap::kStackLimitRootIndex);
1059 } // NOLINT(readability/fn_size)
1062 // Assembles branches after this instruction.
1063 void CodeGenerator::AssembleArchBranch(Instruction* instr, BranchInfo* branch) {
1064 X64OperandConverter i(this, instr);
1065 Label::Distance flabel_distance =
1066 branch->fallthru ? Label::kNear : Label::kFar;
1067 Label* tlabel = branch->true_label;
1068 Label* flabel = branch->false_label;
1069 switch (branch->condition) {
1070 case kUnorderedEqual:
1071 __ j(parity_even, flabel, flabel_distance);
1074 __ j(equal, tlabel);
1076 case kUnorderedNotEqual:
1077 __ j(parity_even, tlabel);
1080 __ j(not_equal, tlabel);
1082 case kSignedLessThan:
1085 case kSignedGreaterThanOrEqual:
1086 __ j(greater_equal, tlabel);
1088 case kSignedLessThanOrEqual:
1089 __ j(less_equal, tlabel);
1091 case kSignedGreaterThan:
1092 __ j(greater, tlabel);
1094 case kUnsignedLessThan:
1095 __ j(below, tlabel);
1097 case kUnsignedGreaterThanOrEqual:
1098 __ j(above_equal, tlabel);
1100 case kUnsignedLessThanOrEqual:
1101 __ j(below_equal, tlabel);
1103 case kUnsignedGreaterThan:
1104 __ j(above, tlabel);
1107 __ j(overflow, tlabel);
1110 __ j(no_overflow, tlabel);
1113 if (!branch->fallthru) __ jmp(flabel, flabel_distance);
1117 void CodeGenerator::AssembleArchJump(RpoNumber target) {
1118 if (!IsNextInAssemblyOrder(target)) __ jmp(GetLabel(target));
1122 // Assembles boolean materializations after this instruction.
1123 void CodeGenerator::AssembleArchBoolean(Instruction* instr,
1124 FlagsCondition condition) {
1125 X64OperandConverter i(this, instr);
1128 // Materialize a full 64-bit 1 or 0 value. The result register is always the
1129 // last output of the instruction.
1131 DCHECK_NE(0u, instr->OutputCount());
1132 Register reg = i.OutputRegister(instr->OutputCount() - 1);
1133 Condition cc = no_condition;
1134 switch (condition) {
1135 case kUnorderedEqual:
1136 __ j(parity_odd, &check, Label::kNear);
1137 __ movl(reg, Immediate(0));
1138 __ jmp(&done, Label::kNear);
1143 case kUnorderedNotEqual:
1144 __ j(parity_odd, &check, Label::kNear);
1145 __ movl(reg, Immediate(1));
1146 __ jmp(&done, Label::kNear);
1151 case kSignedLessThan:
1154 case kSignedGreaterThanOrEqual:
1157 case kSignedLessThanOrEqual:
1160 case kSignedGreaterThan:
1163 case kUnsignedLessThan:
1166 case kUnsignedGreaterThanOrEqual:
1169 case kUnsignedLessThanOrEqual:
1172 case kUnsignedGreaterThan:
1184 __ movzxbl(reg, reg);
1189 void CodeGenerator::AssembleArchLookupSwitch(Instruction* instr) {
1190 X64OperandConverter i(this, instr);
1191 Register input = i.InputRegister(0);
1192 for (size_t index = 2; index < instr->InputCount(); index += 2) {
1193 __ cmpl(input, Immediate(i.InputInt32(index + 0)));
1194 __ j(equal, GetLabel(i.InputRpo(index + 1)));
1196 AssembleArchJump(i.InputRpo(1));
1200 void CodeGenerator::AssembleArchTableSwitch(Instruction* instr) {
1201 X64OperandConverter i(this, instr);
1202 Register input = i.InputRegister(0);
1203 int32_t const case_count = static_cast<int32_t>(instr->InputCount() - 2);
1204 Label** cases = zone()->NewArray<Label*>(case_count);
1205 for (int32_t index = 0; index < case_count; ++index) {
1206 cases[index] = GetLabel(i.InputRpo(index + 2));
1208 Label* const table = AddJumpTable(cases, case_count);
1209 __ cmpl(input, Immediate(case_count));
1210 __ j(above_equal, GetLabel(i.InputRpo(1)));
1211 __ leaq(kScratchRegister, Operand(table));
1212 __ jmp(Operand(kScratchRegister, input, times_8, 0));
1216 void CodeGenerator::AssembleDeoptimizerCall(
1217 int deoptimization_id, Deoptimizer::BailoutType bailout_type) {
1218 Address deopt_entry = Deoptimizer::GetDeoptimizationEntry(
1219 isolate(), deoptimization_id, bailout_type);
1220 __ call(deopt_entry, RelocInfo::RUNTIME_ENTRY);
1224 void CodeGenerator::AssemblePrologue() {
1225 CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
1226 int stack_slots = frame()->GetSpillSlotCount();
1227 if (descriptor->kind() == CallDescriptor::kCallAddress) {
1230 const RegList saves = descriptor->CalleeSavedRegisters();
1231 if (saves != 0) { // Save callee-saved registers.
1232 int register_save_area_size = 0;
1233 for (int i = Register::kNumRegisters - 1; i >= 0; i--) {
1234 if (!((1 << i) & saves)) continue;
1235 __ pushq(Register::from_code(i));
1236 register_save_area_size += kPointerSize;
1238 frame()->SetRegisterSaveAreaSize(register_save_area_size);
1240 } else if (descriptor->IsJSFunctionCall()) {
1241 CompilationInfo* info = this->info();
1242 __ Prologue(info->IsCodePreAgingActive());
1243 frame()->SetRegisterSaveAreaSize(
1244 StandardFrameConstants::kFixedFrameSizeFromFp);
1245 } else if (stack_slots > 0) {
1247 frame()->SetRegisterSaveAreaSize(
1248 StandardFrameConstants::kFixedFrameSizeFromFp);
1251 if (info()->is_osr()) {
1252 // TurboFan OSR-compiled functions cannot be entered directly.
1253 __ Abort(kShouldNotDirectlyEnterOsrFunction);
1255 // Unoptimized code jumps directly to this entrypoint while the unoptimized
1256 // frame is still on the stack. Optimized code uses OSR values directly from
1257 // the unoptimized frame. Thus, all that needs to be done is to allocate the
1258 // remaining stack slots.
1259 if (FLAG_code_comments) __ RecordComment("-- OSR entrypoint --");
1260 osr_pc_offset_ = __ pc_offset();
1261 DCHECK(stack_slots >= frame()->GetOsrStackSlotCount());
1262 stack_slots -= frame()->GetOsrStackSlotCount();
1265 if (stack_slots > 0) {
1266 __ subq(rsp, Immediate(stack_slots * kPointerSize));
1271 void CodeGenerator::AssembleReturn() {
1272 CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
1273 int stack_slots = frame()->GetSpillSlotCount();
1274 if (descriptor->kind() == CallDescriptor::kCallAddress) {
1275 if (frame()->GetRegisterSaveAreaSize() > 0) {
1276 // Remove this frame's spill slots first.
1277 if (stack_slots > 0) {
1278 __ addq(rsp, Immediate(stack_slots * kPointerSize));
1280 const RegList saves = descriptor->CalleeSavedRegisters();
1281 // Restore registers.
1283 for (int i = 0; i < Register::kNumRegisters; i++) {
1284 if (!((1 << i) & saves)) continue;
1285 __ popq(Register::from_code(i));
1288 __ popq(rbp); // Pop caller's frame pointer.
1291 // No saved registers.
1292 __ movq(rsp, rbp); // Move stack pointer back to frame pointer.
1293 __ popq(rbp); // Pop caller's frame pointer.
1296 } else if (descriptor->IsJSFunctionCall() || stack_slots > 0) {
1297 __ movq(rsp, rbp); // Move stack pointer back to frame pointer.
1298 __ popq(rbp); // Pop caller's frame pointer.
1299 int pop_count = descriptor->IsJSFunctionCall()
1300 ? static_cast<int>(descriptor->JSParameterCount())
1302 __ ret(pop_count * kPointerSize);
1309 void CodeGenerator::AssembleMove(InstructionOperand* source,
1310 InstructionOperand* destination) {
1311 X64OperandConverter g(this, NULL);
1312 // Dispatch on the source and destination operand kinds. Not all
1313 // combinations are possible.
1314 if (source->IsRegister()) {
1315 DCHECK(destination->IsRegister() || destination->IsStackSlot());
1316 Register src = g.ToRegister(source);
1317 if (destination->IsRegister()) {
1318 __ movq(g.ToRegister(destination), src);
1320 __ movq(g.ToOperand(destination), src);
1322 } else if (source->IsStackSlot()) {
1323 DCHECK(destination->IsRegister() || destination->IsStackSlot());
1324 Operand src = g.ToOperand(source);
1325 if (destination->IsRegister()) {
1326 Register dst = g.ToRegister(destination);
1329 // Spill on demand to use a temporary register for memory-to-memory
1331 Register tmp = kScratchRegister;
1332 Operand dst = g.ToOperand(destination);
1336 } else if (source->IsConstant()) {
1337 ConstantOperand* constant_source = ConstantOperand::cast(source);
1338 Constant src = g.ToConstant(constant_source);
1339 if (destination->IsRegister() || destination->IsStackSlot()) {
1340 Register dst = destination->IsRegister() ? g.ToRegister(destination)
1342 switch (src.type()) {
1343 case Constant::kInt32:
1344 // TODO(dcarney): don't need scratch in this case.
1345 __ Set(dst, src.ToInt32());
1347 case Constant::kInt64:
1348 __ Set(dst, src.ToInt64());
1350 case Constant::kFloat32:
1352 isolate()->factory()->NewNumber(src.ToFloat32(), TENURED));
1354 case Constant::kFloat64:
1356 isolate()->factory()->NewNumber(src.ToFloat64(), TENURED));
1358 case Constant::kExternalReference:
1359 __ Move(dst, src.ToExternalReference());
1361 case Constant::kHeapObject: {
1362 Handle<HeapObject> src_object = src.ToHeapObject();
1363 if (info()->IsOptimizing() &&
1364 src_object.is_identical_to(info()->context())) {
1365 // Loading the context from the frame is way cheaper than
1366 // materializing the actual context heap object address.
1367 __ movp(dst, Operand(rbp, StandardFrameConstants::kContextOffset));
1369 __ Move(dst, src_object);
1373 case Constant::kRpoNumber:
1374 UNREACHABLE(); // TODO(dcarney): load of labels on x64.
1377 if (destination->IsStackSlot()) {
1378 __ movq(g.ToOperand(destination), kScratchRegister);
1380 } else if (src.type() == Constant::kFloat32) {
1381 // TODO(turbofan): Can we do better here?
1382 uint32_t src_const = bit_cast<uint32_t>(src.ToFloat32());
1383 if (destination->IsDoubleRegister()) {
1384 __ Move(g.ToDoubleRegister(destination), src_const);
1386 DCHECK(destination->IsDoubleStackSlot());
1387 Operand dst = g.ToOperand(destination);
1388 __ movl(dst, Immediate(src_const));
1391 DCHECK_EQ(Constant::kFloat64, src.type());
1392 uint64_t src_const = bit_cast<uint64_t>(src.ToFloat64());
1393 if (destination->IsDoubleRegister()) {
1394 __ Move(g.ToDoubleRegister(destination), src_const);
1396 DCHECK(destination->IsDoubleStackSlot());
1397 __ movq(kScratchRegister, src_const);
1398 __ movq(g.ToOperand(destination), kScratchRegister);
1401 } else if (source->IsDoubleRegister()) {
1402 XMMRegister src = g.ToDoubleRegister(source);
1403 if (destination->IsDoubleRegister()) {
1404 XMMRegister dst = g.ToDoubleRegister(destination);
1405 __ movaps(dst, src);
1407 DCHECK(destination->IsDoubleStackSlot());
1408 Operand dst = g.ToOperand(destination);
1411 } else if (source->IsDoubleStackSlot()) {
1412 DCHECK(destination->IsDoubleRegister() || destination->IsDoubleStackSlot());
1413 Operand src = g.ToOperand(source);
1414 if (destination->IsDoubleRegister()) {
1415 XMMRegister dst = g.ToDoubleRegister(destination);
1418 // We rely on having xmm0 available as a fixed scratch register.
1419 Operand dst = g.ToOperand(destination);
1420 __ movsd(xmm0, src);
1421 __ movsd(dst, xmm0);
1429 void CodeGenerator::AssembleSwap(InstructionOperand* source,
1430 InstructionOperand* destination) {
1431 X64OperandConverter g(this, NULL);
1432 // Dispatch on the source and destination operand kinds. Not all
1433 // combinations are possible.
1434 if (source->IsRegister() && destination->IsRegister()) {
1435 // Register-register.
1436 __ xchgq(g.ToRegister(source), g.ToRegister(destination));
1437 } else if (source->IsRegister() && destination->IsStackSlot()) {
1438 Register src = g.ToRegister(source);
1439 Operand dst = g.ToOperand(destination);
1441 } else if ((source->IsStackSlot() && destination->IsStackSlot()) ||
1442 (source->IsDoubleStackSlot() &&
1443 destination->IsDoubleStackSlot())) {
1445 Register tmp = kScratchRegister;
1446 Operand src = g.ToOperand(source);
1447 Operand dst = g.ToOperand(destination);
1451 } else if (source->IsDoubleRegister() && destination->IsDoubleRegister()) {
1452 // XMM register-register swap. We rely on having xmm0
1453 // available as a fixed scratch register.
1454 XMMRegister src = g.ToDoubleRegister(source);
1455 XMMRegister dst = g.ToDoubleRegister(destination);
1456 __ movaps(xmm0, src);
1457 __ movaps(src, dst);
1458 __ movaps(dst, xmm0);
1459 } else if (source->IsDoubleRegister() && destination->IsDoubleStackSlot()) {
1460 // XMM register-memory swap. We rely on having xmm0
1461 // available as a fixed scratch register.
1462 XMMRegister src = g.ToDoubleRegister(source);
1463 Operand dst = g.ToOperand(destination);
1464 __ movsd(xmm0, src);
1466 __ movsd(dst, xmm0);
1468 // No other combinations are possible.
1474 void CodeGenerator::AssembleJumpTable(Label** targets, size_t target_count) {
1475 for (size_t index = 0; index < target_count; ++index) {
1476 __ dq(targets[index]);
1481 void CodeGenerator::AddNopForSmiCodeInlining() { __ nop(); }
1484 void CodeGenerator::EnsureSpaceForLazyDeopt() {
1485 int space_needed = Deoptimizer::patch_size();
1486 if (!info()->IsStub()) {
1487 // Ensure that we have enough space after the previous lazy-bailout
1488 // instruction for patching the code here.
1489 int current_pc = masm()->pc_offset();
1490 if (current_pc < last_lazy_deopt_pc_ + space_needed) {
1491 int padding_size = last_lazy_deopt_pc_ + space_needed - current_pc;
1492 __ Nop(padding_size);
1495 MarkLazyDeoptSite();
1500 } // namespace internal
1501 } // namespace compiler