diff options
Diffstat (limited to 'js/src/jit/x64')
-rw-r--r-- | js/src/jit/x64/Assembler-x64.cpp | 246 | ||||
-rw-r--r-- | js/src/jit/x64/Assembler-x64.h | 1249 | ||||
-rw-r--r-- | js/src/jit/x64/BaseAssembler-x64.h | 1373 | ||||
-rw-r--r-- | js/src/jit/x64/CodeGenerator-x64.cpp | 984 | ||||
-rw-r--r-- | js/src/jit/x64/CodeGenerator-x64.h | 41 | ||||
-rw-r--r-- | js/src/jit/x64/LIR-x64.h | 170 | ||||
-rw-r--r-- | js/src/jit/x64/Lowering-x64.cpp | 565 | ||||
-rw-r--r-- | js/src/jit/x64/Lowering-x64.h | 70 | ||||
-rw-r--r-- | js/src/jit/x64/MacroAssembler-x64-inl.h | 1099 | ||||
-rw-r--r-- | js/src/jit/x64/MacroAssembler-x64.cpp | 1747 | ||||
-rw-r--r-- | js/src/jit/x64/MacroAssembler-x64.h | 1218 | ||||
-rw-r--r-- | js/src/jit/x64/SharedICHelpers-x64-inl.h | 80 | ||||
-rw-r--r-- | js/src/jit/x64/SharedICHelpers-x64.h | 70 | ||||
-rw-r--r-- | js/src/jit/x64/SharedICRegisters-x64.h | 33 | ||||
-rw-r--r-- | js/src/jit/x64/Trampoline-x64.cpp | 888 |
15 files changed, 9833 insertions, 0 deletions
diff --git a/js/src/jit/x64/Assembler-x64.cpp b/js/src/jit/x64/Assembler-x64.cpp new file mode 100644 index 0000000000..e00e34e830 --- /dev/null +++ b/js/src/jit/x64/Assembler-x64.cpp @@ -0,0 +1,246 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#include "jit/x64/Assembler-x64.h" + +#include "gc/Tracer.h" +#include "util/Memory.h" + +using namespace js; +using namespace js::jit; + +ABIArgGenerator::ABIArgGenerator() + : +#if defined(XP_WIN) + regIndex_(0), + stackOffset_(ShadowStackSpace), +#else + intRegIndex_(0), + floatRegIndex_(0), + stackOffset_(0), +#endif + current_() { +} + +ABIArg ABIArgGenerator::next(MIRType type) { +#if defined(XP_WIN) + static_assert(NumIntArgRegs == NumFloatArgRegs); + if (regIndex_ == NumIntArgRegs) { + if (type == MIRType::Simd128) { + // On Win64, >64 bit args need to be passed by reference. However, wasm + // doesn't allow passing SIMD values to JS, so the only way to reach this + // is wasm to wasm calls. Ergo we can break the native ABI here and use + // the Wasm ABI instead. + stackOffset_ = AlignBytes(stackOffset_, SimdMemoryAlignment); + current_ = ABIArg(stackOffset_); + stackOffset_ += Simd128DataSize; + } else { + current_ = ABIArg(stackOffset_); + stackOffset_ += sizeof(uint64_t); + } + return current_; + } + switch (type) { + case MIRType::Int32: + case MIRType::Int64: + case MIRType::Pointer: + case MIRType::RefOrNull: + case MIRType::StackResults: + current_ = ABIArg(IntArgRegs[regIndex_++]); + break; + case MIRType::Float32: + current_ = ABIArg(FloatArgRegs[regIndex_++].asSingle()); + break; + case MIRType::Double: + current_ = ABIArg(FloatArgRegs[regIndex_++]); + break; + case MIRType::Simd128: + // On Win64, >64 bit args need to be passed by reference, but wasm + // doesn't allow passing SIMD values to FFIs. The only way to reach + // here is asm to asm calls, so we can break the ABI here. + current_ = ABIArg(FloatArgRegs[regIndex_++].asSimd128()); + break; + default: + MOZ_CRASH("Unexpected argument type"); + } + return current_; +#else + switch (type) { + case MIRType::Int32: + case MIRType::Int64: + case MIRType::Pointer: + case MIRType::RefOrNull: + case MIRType::StackResults: + if (intRegIndex_ == NumIntArgRegs) { + current_ = ABIArg(stackOffset_); + stackOffset_ += sizeof(uint64_t); + break; + } + current_ = ABIArg(IntArgRegs[intRegIndex_++]); + break; + case MIRType::Double: + case MIRType::Float32: + if (floatRegIndex_ == NumFloatArgRegs) { + current_ = ABIArg(stackOffset_); + stackOffset_ += sizeof(uint64_t); + break; + } + if (type == MIRType::Float32) { + current_ = ABIArg(FloatArgRegs[floatRegIndex_++].asSingle()); + } else { + current_ = ABIArg(FloatArgRegs[floatRegIndex_++]); + } + break; + case MIRType::Simd128: + if (floatRegIndex_ == NumFloatArgRegs) { + stackOffset_ = AlignBytes(stackOffset_, SimdMemoryAlignment); + current_ = ABIArg(stackOffset_); + stackOffset_ += Simd128DataSize; + break; + } + current_ = ABIArg(FloatArgRegs[floatRegIndex_++].asSimd128()); + break; + default: + MOZ_CRASH("Unexpected argument type"); + } + return current_; +#endif +} + +void Assembler::addPendingJump(JmpSrc src, ImmPtr target, + RelocationKind reloc) { + MOZ_ASSERT(target.value != nullptr); + + // Emit reloc before modifying the jump table, since it computes a 0-based + // index. This jump is not patchable at runtime. + if (reloc == RelocationKind::JITCODE) { + jumpRelocations_.writeUnsigned(src.offset()); + } + + static_assert(MaxCodeBytesPerProcess <= uint64_t(2) * 1024 * 1024 * 1024, + "Code depends on using int32_t for cross-JitCode jump offsets"); + + MOZ_ASSERT_IF(reloc == RelocationKind::JITCODE, + AddressIsInExecutableMemory(target.value)); + + RelativePatch patch(src.offset(), target.value, reloc); + if (reloc == RelocationKind::JITCODE || + AddressIsInExecutableMemory(target.value)) { + enoughMemory_ &= codeJumps_.append(patch); + } else { + enoughMemory_ &= extendedJumps_.append(patch); + } +} + +void Assembler::finish() { + if (oom()) { + return; + } + + AutoCreatedBy acb(*this, "Assembler::finish"); + + if (!extendedJumps_.length()) { + // Since we may be folowed by non-executable data, eagerly insert an + // undefined instruction byte to prevent processors from decoding + // gibberish into their pipelines. See Intel performance guides. + masm.ud2(); + return; + } + + // Emit the jump table. + masm.haltingAlign(SizeOfJumpTableEntry); + extendedJumpTable_ = masm.size(); + + // Zero the extended jumps table. + for (size_t i = 0; i < extendedJumps_.length(); i++) { +#ifdef DEBUG + size_t oldSize = masm.size(); +#endif + MOZ_ASSERT(hasCreator()); + masm.jmp_rip(2); + MOZ_ASSERT_IF(!masm.oom(), masm.size() - oldSize == 6); + // Following an indirect branch with ud2 hints to the hardware that + // there's no fall-through. This also aligns the 64-bit immediate. + masm.ud2(); + MOZ_ASSERT_IF(!masm.oom(), masm.size() - oldSize == 8); + masm.immediate64(0); + MOZ_ASSERT_IF(!masm.oom(), masm.size() - oldSize == SizeOfExtendedJump); + MOZ_ASSERT_IF(!masm.oom(), masm.size() - oldSize == SizeOfJumpTableEntry); + } +} + +void Assembler::executableCopy(uint8_t* buffer) { + AssemblerX86Shared::executableCopy(buffer); + + for (RelativePatch& rp : codeJumps_) { + uint8_t* src = buffer + rp.offset; + MOZ_ASSERT(rp.target); + + MOZ_RELEASE_ASSERT(X86Encoding::CanRelinkJump(src, rp.target)); + X86Encoding::SetRel32(src, rp.target); + } + + for (size_t i = 0; i < extendedJumps_.length(); i++) { + RelativePatch& rp = extendedJumps_[i]; + uint8_t* src = buffer + rp.offset; + MOZ_ASSERT(rp.target); + + if (X86Encoding::CanRelinkJump(src, rp.target)) { + X86Encoding::SetRel32(src, rp.target); + } else { + // An extended jump table must exist, and its offset must be in + // range. + MOZ_ASSERT(extendedJumpTable_); + MOZ_ASSERT((extendedJumpTable_ + i * SizeOfJumpTableEntry) <= + size() - SizeOfJumpTableEntry); + + // Patch the jump to go to the extended jump entry. + uint8_t* entry = buffer + extendedJumpTable_ + i * SizeOfJumpTableEntry; + X86Encoding::SetRel32(src, entry); + + // Now patch the pointer, note that we need to align it to + // *after* the extended jump, i.e. after the 64-bit immedate. + X86Encoding::SetPointer(entry + SizeOfExtendedJump, rp.target); + } + } +} + +class RelocationIterator { + CompactBufferReader reader_; + uint32_t offset_ = 0; + + public: + explicit RelocationIterator(CompactBufferReader& reader) : reader_(reader) {} + + bool read() { + if (!reader_.more()) { + return false; + } + offset_ = reader_.readUnsigned(); + return true; + } + + uint32_t offset() const { return offset_; } +}; + +JitCode* Assembler::CodeFromJump(JitCode* code, uint8_t* jump) { + uint8_t* target = (uint8_t*)X86Encoding::GetRel32Target(jump); + + MOZ_ASSERT(!code->containsNativePC(target), + "Extended jump table not used for cross-JitCode jumps"); + + return JitCode::FromExecutable(target); +} + +void Assembler::TraceJumpRelocations(JSTracer* trc, JitCode* code, + CompactBufferReader& reader) { + RelocationIterator iter(reader); + while (iter.read()) { + JitCode* child = CodeFromJump(code, code->raw() + iter.offset()); + TraceManuallyBarrieredEdge(trc, &child, "rel32"); + MOZ_ASSERT(child == CodeFromJump(code, code->raw() + iter.offset())); + } +} diff --git a/js/src/jit/x64/Assembler-x64.h b/js/src/jit/x64/Assembler-x64.h new file mode 100644 index 0000000000..2f44d9f3e4 --- /dev/null +++ b/js/src/jit/x64/Assembler-x64.h @@ -0,0 +1,1249 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#ifndef jit_x64_Assembler_x64_h +#define jit_x64_Assembler_x64_h + +#include <iterator> + +#include "jit/JitCode.h" +#include "jit/shared/Assembler-shared.h" + +namespace js { +namespace jit { + +static constexpr Register rax{X86Encoding::rax}; +static constexpr Register rbx{X86Encoding::rbx}; +static constexpr Register rcx{X86Encoding::rcx}; +static constexpr Register rdx{X86Encoding::rdx}; +static constexpr Register rsi{X86Encoding::rsi}; +static constexpr Register rdi{X86Encoding::rdi}; +static constexpr Register rbp{X86Encoding::rbp}; +static constexpr Register r8{X86Encoding::r8}; +static constexpr Register r9{X86Encoding::r9}; +static constexpr Register r10{X86Encoding::r10}; +static constexpr Register r11{X86Encoding::r11}; +static constexpr Register r12{X86Encoding::r12}; +static constexpr Register r13{X86Encoding::r13}; +static constexpr Register r14{X86Encoding::r14}; +static constexpr Register r15{X86Encoding::r15}; +static constexpr Register rsp{X86Encoding::rsp}; + +static constexpr FloatRegister xmm0 = + FloatRegister(X86Encoding::xmm0, FloatRegisters::Double); +static constexpr FloatRegister xmm1 = + FloatRegister(X86Encoding::xmm1, FloatRegisters::Double); +static constexpr FloatRegister xmm2 = + FloatRegister(X86Encoding::xmm2, FloatRegisters::Double); +static constexpr FloatRegister xmm3 = + FloatRegister(X86Encoding::xmm3, FloatRegisters::Double); +static constexpr FloatRegister xmm4 = + FloatRegister(X86Encoding::xmm4, FloatRegisters::Double); +static constexpr FloatRegister xmm5 = + FloatRegister(X86Encoding::xmm5, FloatRegisters::Double); +static constexpr FloatRegister xmm6 = + FloatRegister(X86Encoding::xmm6, FloatRegisters::Double); +static constexpr FloatRegister xmm7 = + FloatRegister(X86Encoding::xmm7, FloatRegisters::Double); +static constexpr FloatRegister xmm8 = + FloatRegister(X86Encoding::xmm8, FloatRegisters::Double); +static constexpr FloatRegister xmm9 = + FloatRegister(X86Encoding::xmm9, FloatRegisters::Double); +static constexpr FloatRegister xmm10 = + FloatRegister(X86Encoding::xmm10, FloatRegisters::Double); +static constexpr FloatRegister xmm11 = + FloatRegister(X86Encoding::xmm11, FloatRegisters::Double); +static constexpr FloatRegister xmm12 = + FloatRegister(X86Encoding::xmm12, FloatRegisters::Double); +static constexpr FloatRegister xmm13 = + FloatRegister(X86Encoding::xmm13, FloatRegisters::Double); +static constexpr FloatRegister xmm14 = + FloatRegister(X86Encoding::xmm14, FloatRegisters::Double); +static constexpr FloatRegister xmm15 = + FloatRegister(X86Encoding::xmm15, FloatRegisters::Double); + +// Vector registers fixed for use with some instructions, e.g. PBLENDVB. +static constexpr FloatRegister vmm0 = + FloatRegister(X86Encoding::xmm0, FloatRegisters::Simd128); + +// X86-common synonyms. +static constexpr Register eax = rax; +static constexpr Register ebx = rbx; +static constexpr Register ecx = rcx; +static constexpr Register edx = rdx; +static constexpr Register esi = rsi; +static constexpr Register edi = rdi; +static constexpr Register ebp = rbp; +static constexpr Register esp = rsp; + +static constexpr Register InvalidReg{X86Encoding::invalid_reg}; +static constexpr FloatRegister InvalidFloatReg = FloatRegister(); + +static constexpr Register StackPointer = rsp; +static constexpr Register FramePointer = rbp; +static constexpr Register JSReturnReg = rcx; +// Avoid, except for assertions. +static constexpr Register JSReturnReg_Type = JSReturnReg; +static constexpr Register JSReturnReg_Data = JSReturnReg; + +static constexpr Register ScratchReg = r11; + +// Helper class for ScratchRegister usage. Asserts that only one piece +// of code thinks it has exclusive ownership of the scratch register. +struct ScratchRegisterScope : public AutoRegisterScope { + explicit ScratchRegisterScope(MacroAssembler& masm) + : AutoRegisterScope(masm, ScratchReg) {} +}; + +static constexpr Register ReturnReg = rax; +static constexpr Register HeapReg = r15; +static constexpr Register64 ReturnReg64(rax); +static constexpr FloatRegister ReturnFloat32Reg = + FloatRegister(X86Encoding::xmm0, FloatRegisters::Single); +static constexpr FloatRegister ReturnDoubleReg = + FloatRegister(X86Encoding::xmm0, FloatRegisters::Double); +static constexpr FloatRegister ReturnSimd128Reg = + FloatRegister(X86Encoding::xmm0, FloatRegisters::Simd128); +static constexpr FloatRegister ScratchFloat32Reg_ = + FloatRegister(X86Encoding::xmm15, FloatRegisters::Single); +static constexpr FloatRegister ScratchDoubleReg_ = + FloatRegister(X86Encoding::xmm15, FloatRegisters::Double); +static constexpr FloatRegister ScratchSimd128Reg = + FloatRegister(X86Encoding::xmm15, FloatRegisters::Simd128); + +// Avoid rbp, which is the FramePointer, which is unavailable in some modes. +static constexpr Register CallTempReg0 = rax; +static constexpr Register CallTempReg1 = rdi; +static constexpr Register CallTempReg2 = rbx; +static constexpr Register CallTempReg3 = rcx; +static constexpr Register CallTempReg4 = rsi; +static constexpr Register CallTempReg5 = rdx; + +// Different argument registers for WIN64 +#if defined(_WIN64) +static constexpr Register IntArgReg0 = rcx; +static constexpr Register IntArgReg1 = rdx; +static constexpr Register IntArgReg2 = r8; +static constexpr Register IntArgReg3 = r9; +static constexpr uint32_t NumIntArgRegs = 4; +static constexpr Register IntArgRegs[NumIntArgRegs] = {rcx, rdx, r8, r9}; + +static constexpr Register CallTempNonArgRegs[] = {rax, rdi, rbx, rsi}; +static constexpr uint32_t NumCallTempNonArgRegs = std::size(CallTempNonArgRegs); + +static constexpr FloatRegister FloatArgReg0 = xmm0; +static constexpr FloatRegister FloatArgReg1 = xmm1; +static constexpr FloatRegister FloatArgReg2 = xmm2; +static constexpr FloatRegister FloatArgReg3 = xmm3; +static constexpr uint32_t NumFloatArgRegs = 4; +static constexpr FloatRegister FloatArgRegs[NumFloatArgRegs] = {xmm0, xmm1, + xmm2, xmm3}; +#else +static constexpr Register IntArgReg0 = rdi; +static constexpr Register IntArgReg1 = rsi; +static constexpr Register IntArgReg2 = rdx; +static constexpr Register IntArgReg3 = rcx; +static constexpr Register IntArgReg4 = r8; +static constexpr Register IntArgReg5 = r9; +static constexpr uint32_t NumIntArgRegs = 6; +static constexpr Register IntArgRegs[NumIntArgRegs] = {rdi, rsi, rdx, + rcx, r8, r9}; + +static constexpr Register CallTempNonArgRegs[] = {rax, rbx}; +static constexpr uint32_t NumCallTempNonArgRegs = std::size(CallTempNonArgRegs); + +static constexpr FloatRegister FloatArgReg0 = xmm0; +static constexpr FloatRegister FloatArgReg1 = xmm1; +static constexpr FloatRegister FloatArgReg2 = xmm2; +static constexpr FloatRegister FloatArgReg3 = xmm3; +static constexpr FloatRegister FloatArgReg4 = xmm4; +static constexpr FloatRegister FloatArgReg5 = xmm5; +static constexpr FloatRegister FloatArgReg6 = xmm6; +static constexpr FloatRegister FloatArgReg7 = xmm7; +static constexpr uint32_t NumFloatArgRegs = 8; +static constexpr FloatRegister FloatArgRegs[NumFloatArgRegs] = { + xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7}; +#endif + +// Registers used by RegExpMatcher and RegExpExecMatch stubs (do not use +// JSReturnOperand). +static constexpr Register RegExpMatcherRegExpReg = CallTempReg0; +static constexpr Register RegExpMatcherStringReg = CallTempReg1; +static constexpr Register RegExpMatcherLastIndexReg = CallTempReg2; + +// Registers used by RegExpExecTest stub (do not use ReturnReg). +static constexpr Register RegExpExecTestRegExpReg = CallTempReg1; +static constexpr Register RegExpExecTestStringReg = CallTempReg2; + +// Registers used by RegExpSearcher stub (do not use ReturnReg). +static constexpr Register RegExpSearcherRegExpReg = CallTempReg1; +static constexpr Register RegExpSearcherStringReg = CallTempReg2; +static constexpr Register RegExpSearcherLastIndexReg = CallTempReg3; + +class ABIArgGenerator { +#if defined(XP_WIN) + unsigned regIndex_; +#else + unsigned intRegIndex_; + unsigned floatRegIndex_; +#endif + uint32_t stackOffset_; + ABIArg current_; + + public: + ABIArgGenerator(); + ABIArg next(MIRType argType); + ABIArg& current() { return current_; } + uint32_t stackBytesConsumedSoFar() const { return stackOffset_; } + void increaseStackOffset(uint32_t bytes) { stackOffset_ += bytes; } +}; + +// These registers may be volatile or nonvolatile. +// Avoid r11, which is the MacroAssembler's ScratchReg. +static constexpr Register ABINonArgReg0 = rax; +static constexpr Register ABINonArgReg1 = rbx; +static constexpr Register ABINonArgReg2 = r10; +static constexpr Register ABINonArgReg3 = r12; + +// This register may be volatile or nonvolatile. Avoid xmm15 which is the +// ScratchDoubleReg. +static constexpr FloatRegister ABINonArgDoubleReg = + FloatRegister(X86Encoding::xmm8, FloatRegisters::Double); + +// These registers may be volatile or nonvolatile. +// Note: these three registers are all guaranteed to be different +static constexpr Register ABINonArgReturnReg0 = r10; +static constexpr Register ABINonArgReturnReg1 = r12; +static constexpr Register ABINonVolatileReg = r13; + +// This register is guaranteed to be clobberable during the prologue and +// epilogue of an ABI call which must preserve both ABI argument, return +// and non-volatile registers. +static constexpr Register ABINonArgReturnVolatileReg = r10; + +// Instance pointer argument register for WebAssembly functions. This must not +// alias any other register used for passing function arguments or return +// values. Preserved by WebAssembly functions. +static constexpr Register InstanceReg = r14; + +// Registers used for asm.js/wasm table calls. These registers must be disjoint +// from the ABI argument registers, InstanceReg and each other. +static constexpr Register WasmTableCallScratchReg0 = ABINonArgReg0; +static constexpr Register WasmTableCallScratchReg1 = ABINonArgReg1; +static constexpr Register WasmTableCallSigReg = ABINonArgReg2; +static constexpr Register WasmTableCallIndexReg = ABINonArgReg3; + +// Registers used for ref calls. +static constexpr Register WasmCallRefCallScratchReg0 = ABINonArgReg0; +static constexpr Register WasmCallRefCallScratchReg1 = ABINonArgReg1; +static constexpr Register WasmCallRefReg = ABINonArgReg3; + +// Register used as a scratch along the return path in the fast js -> wasm stub +// code. This must not overlap ReturnReg, JSReturnOperand, or InstanceReg. +// It must be a volatile register. +static constexpr Register WasmJitEntryReturnScratch = rbx; + +static constexpr Register OsrFrameReg = IntArgReg3; + +static constexpr Register PreBarrierReg = rdx; + +static constexpr Register InterpreterPCReg = r14; + +static constexpr uint32_t ABIStackAlignment = 16; +static constexpr uint32_t CodeAlignment = 16; +static constexpr uint32_t JitStackAlignment = 16; + +static constexpr uint32_t JitStackValueAlignment = + JitStackAlignment / sizeof(Value); +static_assert(JitStackAlignment % sizeof(Value) == 0 && + JitStackValueAlignment >= 1, + "Stack alignment should be a non-zero multiple of sizeof(Value)"); + +static constexpr uint32_t SimdMemoryAlignment = 16; + +static_assert(CodeAlignment % SimdMemoryAlignment == 0, + "Code alignment should be larger than any of the alignments " + "which are used for " + "the constant sections of the code buffer. Thus it should be " + "larger than the " + "alignment for SIMD constants."); + +static_assert(JitStackAlignment % SimdMemoryAlignment == 0, + "Stack alignment should be larger than any of the alignments " + "which are used for " + "spilled values. Thus it should be larger than the alignment " + "for SIMD accesses."); + +static constexpr uint32_t WasmStackAlignment = SimdMemoryAlignment; +static constexpr uint32_t WasmTrapInstructionLength = 2; + +// See comments in wasm::GenerateFunctionPrologue. The difference between these +// is the size of the largest callable prologue on the platform. +static constexpr uint32_t WasmCheckedCallEntryOffset = 0u; + +static constexpr Scale ScalePointer = TimesEight; + +} // namespace jit +} // namespace js + +#include "jit/x86-shared/Assembler-x86-shared.h" + +namespace js { +namespace jit { + +// Return operand from a JS -> JS call. +static constexpr ValueOperand JSReturnOperand = ValueOperand(JSReturnReg); + +class Assembler : public AssemblerX86Shared { + // x64 jumps may need extra bits of relocation, because a jump may extend + // beyond the signed 32-bit range. To account for this we add an extended + // jump table at the bottom of the instruction stream, and if a jump + // overflows its range, it will redirect here. + // + // Each entry in this table is a jmp [rip], followed by a ud2 to hint to the + // hardware branch predictor that there is no fallthrough, followed by the + // eight bytes containing an immediate address. This comes out to 16 bytes. + // +1 byte for opcode + // +1 byte for mod r/m + // +4 bytes for rip-relative offset (2) + // +2 bytes for ud2 instruction + // +8 bytes for 64-bit address + // + static const uint32_t SizeOfExtendedJump = 1 + 1 + 4 + 2 + 8; + static const uint32_t SizeOfJumpTableEntry = 16; + + // Two kinds of jumps on x64: + // + // * codeJumps_ tracks jumps with target within the executable code region + // for the process. These jumps don't need entries in the extended jump + // table because source and target must be within 2 GB of each other. + // + // * extendedJumps_ tracks jumps with target outside the executable code + // region. These jumps need entries in the extended jump table described + // above. + using PendingJumpVector = Vector<RelativePatch, 8, SystemAllocPolicy>; + PendingJumpVector codeJumps_; + PendingJumpVector extendedJumps_; + + uint32_t extendedJumpTable_; + + static JitCode* CodeFromJump(JitCode* code, uint8_t* jump); + + private: + void addPendingJump(JmpSrc src, ImmPtr target, RelocationKind reloc); + + public: + using AssemblerX86Shared::j; + using AssemblerX86Shared::jmp; + using AssemblerX86Shared::pop; + using AssemblerX86Shared::push; + using AssemblerX86Shared::vmovq; + + Assembler() : extendedJumpTable_(0) {} + + static void TraceJumpRelocations(JSTracer* trc, JitCode* code, + CompactBufferReader& reader); + + // The buffer is about to be linked, make sure any constant pools or excess + // bookkeeping has been flushed to the instruction stream. + void finish(); + + // Copy the assembly code to the given buffer, and perform any pending + // relocations relying on the target address. + void executableCopy(uint8_t* buffer); + + void assertNoGCThings() const { +#ifdef DEBUG + MOZ_ASSERT(dataRelocations_.length() == 0); + for (auto& j : codeJumps_) { + MOZ_ASSERT(j.kind == RelocationKind::HARDCODED); + } + for (auto& j : extendedJumps_) { + MOZ_ASSERT(j.kind == RelocationKind::HARDCODED); + } +#endif + } + + // Actual assembly emitting functions. + + void push(const ImmGCPtr ptr) { + movq(ptr, ScratchReg); + push(ScratchReg); + } + void push(const ImmWord ptr) { + // We often end up with ImmWords that actually fit into int32. + // Be aware of the sign extension behavior. + if (ptr.value <= INT32_MAX) { + push(Imm32(ptr.value)); + } else { + movq(ptr, ScratchReg); + push(ScratchReg); + } + } + void push(ImmPtr imm) { push(ImmWord(uintptr_t(imm.value))); } + void push(FloatRegister src) { + subq(Imm32(sizeof(double)), StackPointer); + vmovsd(src, Address(StackPointer, 0)); + } + CodeOffset pushWithPatch(ImmWord word) { + CodeOffset label = movWithPatch(word, ScratchReg); + push(ScratchReg); + return label; + } + + void pop(FloatRegister src) { + vmovsd(Address(StackPointer, 0), src); + addq(Imm32(sizeof(double)), StackPointer); + } + + CodeOffset movWithPatch(ImmWord word, Register dest) { + masm.movq_i64r(word.value, dest.encoding()); + return CodeOffset(masm.currentOffset()); + } + CodeOffset movWithPatch(ImmPtr imm, Register dest) { + return movWithPatch(ImmWord(uintptr_t(imm.value)), dest); + } + + // This is for patching during code generation, not after. + void patchAddq(CodeOffset offset, int32_t n) { + unsigned char* code = masm.data(); + X86Encoding::SetInt32(code + offset.offset(), n); + } + + // Load an ImmWord value into a register. Note that this instruction will + // attempt to optimize its immediate field size. When a full 64-bit + // immediate is needed for a relocation, use movWithPatch. + void movq(ImmWord word, Register dest) { + // Load a 64-bit immediate into a register. If the value falls into + // certain ranges, we can use specialized instructions which have + // smaller encodings. + if (word.value <= UINT32_MAX) { + // movl has a 32-bit unsigned (effectively) immediate field. + masm.movl_i32r((uint32_t)word.value, dest.encoding()); + } else if ((intptr_t)word.value >= INT32_MIN && + (intptr_t)word.value <= INT32_MAX) { + // movq has a 32-bit signed immediate field. + masm.movq_i32r((int32_t)(intptr_t)word.value, dest.encoding()); + } else { + // Otherwise use movabs. + masm.movq_i64r(word.value, dest.encoding()); + } + } + void movq(ImmPtr imm, Register dest) { + movq(ImmWord(uintptr_t(imm.value)), dest); + } + void movq(ImmGCPtr ptr, Register dest) { + masm.movq_i64r(uintptr_t(ptr.value), dest.encoding()); + writeDataRelocation(ptr); + } + void movq(const Operand& src, Register dest) { + switch (src.kind()) { + case Operand::REG: + masm.movq_rr(src.reg(), dest.encoding()); + break; + case Operand::MEM_REG_DISP: + masm.movq_mr(src.disp(), src.base(), dest.encoding()); + break; + case Operand::MEM_SCALE: + masm.movq_mr(src.disp(), src.base(), src.index(), src.scale(), + dest.encoding()); + break; + case Operand::MEM_ADDRESS32: + masm.movq_mr(src.address(), dest.encoding()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + void movq(Register src, const Operand& dest) { + switch (dest.kind()) { + case Operand::REG: + masm.movq_rr(src.encoding(), dest.reg()); + break; + case Operand::MEM_REG_DISP: + masm.movq_rm(src.encoding(), dest.disp(), dest.base()); + break; + case Operand::MEM_SCALE: + masm.movq_rm(src.encoding(), dest.disp(), dest.base(), dest.index(), + dest.scale()); + break; + case Operand::MEM_ADDRESS32: + masm.movq_rm(src.encoding(), dest.address()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + void movq(Imm32 imm32, const Operand& dest) { + switch (dest.kind()) { + case Operand::REG: + masm.movl_i32r(imm32.value, dest.reg()); + break; + case Operand::MEM_REG_DISP: + masm.movq_i32m(imm32.value, dest.disp(), dest.base()); + break; + case Operand::MEM_SCALE: + masm.movq_i32m(imm32.value, dest.disp(), dest.base(), dest.index(), + dest.scale()); + break; + case Operand::MEM_ADDRESS32: + masm.movq_i32m(imm32.value, dest.address()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + void vmovq(Register src, FloatRegister dest) { + masm.vmovq_rr(src.encoding(), dest.encoding()); + } + void vmovq(FloatRegister src, Register dest) { + masm.vmovq_rr(src.encoding(), dest.encoding()); + } + void movq(Register src, Register dest) { + masm.movq_rr(src.encoding(), dest.encoding()); + } + + void cmovCCq(Condition cond, const Operand& src, Register dest) { + X86Encoding::Condition cc = static_cast<X86Encoding::Condition>(cond); + switch (src.kind()) { + case Operand::REG: + masm.cmovCCq_rr(cc, src.reg(), dest.encoding()); + break; + case Operand::MEM_REG_DISP: + masm.cmovCCq_mr(cc, src.disp(), src.base(), dest.encoding()); + break; + case Operand::MEM_SCALE: + masm.cmovCCq_mr(cc, src.disp(), src.base(), src.index(), src.scale(), + dest.encoding()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + void cmovCCq(Condition cond, Register src, Register dest) { + X86Encoding::Condition cc = static_cast<X86Encoding::Condition>(cond); + masm.cmovCCq_rr(cc, src.encoding(), dest.encoding()); + } + + void cmovzq(const Operand& src, Register dest) { + cmovCCq(Condition::Zero, src, dest); + } + void cmovnzq(const Operand& src, Register dest) { + cmovCCq(Condition::NonZero, src, dest); + } + + template <typename T> + void lock_addq(T src, const Operand& op) { + masm.prefix_lock(); + addq(src, op); + } + template <typename T> + void lock_subq(T src, const Operand& op) { + masm.prefix_lock(); + subq(src, op); + } + template <typename T> + void lock_andq(T src, const Operand& op) { + masm.prefix_lock(); + andq(src, op); + } + template <typename T> + void lock_orq(T src, const Operand& op) { + masm.prefix_lock(); + orq(src, op); + } + template <typename T> + void lock_xorq(T src, const Operand& op) { + masm.prefix_lock(); + xorq(src, op); + } + + void lock_cmpxchgq(Register src, const Operand& mem) { + masm.prefix_lock(); + switch (mem.kind()) { + case Operand::MEM_REG_DISP: + masm.cmpxchgq(src.encoding(), mem.disp(), mem.base()); + break; + case Operand::MEM_SCALE: + masm.cmpxchgq(src.encoding(), mem.disp(), mem.base(), mem.index(), + mem.scale()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + + void xchgq(Register src, Register dest) { + masm.xchgq_rr(src.encoding(), dest.encoding()); + } + + void xchgq(Register src, const Operand& mem) { + switch (mem.kind()) { + case Operand::MEM_REG_DISP: + masm.xchgq_rm(src.encoding(), mem.disp(), mem.base()); + break; + case Operand::MEM_SCALE: + masm.xchgq_rm(src.encoding(), mem.disp(), mem.base(), mem.index(), + mem.scale()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + + void lock_xaddq(Register srcdest, const Operand& mem) { + switch (mem.kind()) { + case Operand::MEM_REG_DISP: + masm.lock_xaddq_rm(srcdest.encoding(), mem.disp(), mem.base()); + break; + case Operand::MEM_SCALE: + masm.lock_xaddq_rm(srcdest.encoding(), mem.disp(), mem.base(), + mem.index(), mem.scale()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + + void movsbq(const Operand& src, Register dest) { + switch (src.kind()) { + case Operand::REG: + masm.movsbq_rr(src.reg(), dest.encoding()); + break; + case Operand::MEM_REG_DISP: + masm.movsbq_mr(src.disp(), src.base(), dest.encoding()); + break; + case Operand::MEM_SCALE: + masm.movsbq_mr(src.disp(), src.base(), src.index(), src.scale(), + dest.encoding()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + + void movzbq(const Operand& src, Register dest) { + // movzbl zero-extends to 64 bits and is one byte smaller, so use that + // instead. + movzbl(src, dest); + } + + void movswq(const Operand& src, Register dest) { + switch (src.kind()) { + case Operand::REG: + masm.movswq_rr(src.reg(), dest.encoding()); + break; + case Operand::MEM_REG_DISP: + masm.movswq_mr(src.disp(), src.base(), dest.encoding()); + break; + case Operand::MEM_SCALE: + masm.movswq_mr(src.disp(), src.base(), src.index(), src.scale(), + dest.encoding()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + + void movzwq(const Operand& src, Register dest) { + // movzwl zero-extends to 64 bits and is one byte smaller, so use that + // instead. + movzwl(src, dest); + } + + void movslq(Register src, Register dest) { + masm.movslq_rr(src.encoding(), dest.encoding()); + } + void movslq(const Operand& src, Register dest) { + switch (src.kind()) { + case Operand::REG: + masm.movslq_rr(src.reg(), dest.encoding()); + break; + case Operand::MEM_REG_DISP: + masm.movslq_mr(src.disp(), src.base(), dest.encoding()); + break; + case Operand::MEM_SCALE: + masm.movslq_mr(src.disp(), src.base(), src.index(), src.scale(), + dest.encoding()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + + void andq(Register src, Register dest) { + masm.andq_rr(src.encoding(), dest.encoding()); + } + void andq(Imm32 imm, Register dest) { + masm.andq_ir(imm.value, dest.encoding()); + } + void andq(const Operand& src, Register dest) { + switch (src.kind()) { + case Operand::REG: + masm.andq_rr(src.reg(), dest.encoding()); + break; + case Operand::MEM_REG_DISP: + masm.andq_mr(src.disp(), src.base(), dest.encoding()); + break; + case Operand::MEM_SCALE: + masm.andq_mr(src.disp(), src.base(), src.index(), src.scale(), + dest.encoding()); + break; + case Operand::MEM_ADDRESS32: + masm.andq_mr(src.address(), dest.encoding()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + void andq(Register src, const Operand& dest) { + switch (dest.kind()) { + case Operand::REG: + masm.andq_rr(src.encoding(), dest.reg()); + break; + case Operand::MEM_REG_DISP: + masm.andq_rm(src.encoding(), dest.disp(), dest.base()); + break; + case Operand::MEM_SCALE: + masm.andq_rm(src.encoding(), dest.disp(), dest.base(), dest.index(), + dest.scale()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + + void addq(Imm32 imm, Register dest) { + masm.addq_ir(imm.value, dest.encoding()); + } + CodeOffset addqWithPatch(Imm32 imm, Register dest) { + masm.addq_i32r(imm.value, dest.encoding()); + return CodeOffset(masm.currentOffset()); + } + void addq(Imm32 imm, const Operand& dest) { + switch (dest.kind()) { + case Operand::REG: + masm.addq_ir(imm.value, dest.reg()); + break; + case Operand::MEM_REG_DISP: + masm.addq_im(imm.value, dest.disp(), dest.base()); + break; + case Operand::MEM_ADDRESS32: + masm.addq_im(imm.value, dest.address()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + void addq(Register src, Register dest) { + masm.addq_rr(src.encoding(), dest.encoding()); + } + void addq(const Operand& src, Register dest) { + switch (src.kind()) { + case Operand::REG: + masm.addq_rr(src.reg(), dest.encoding()); + break; + case Operand::MEM_REG_DISP: + masm.addq_mr(src.disp(), src.base(), dest.encoding()); + break; + case Operand::MEM_ADDRESS32: + masm.addq_mr(src.address(), dest.encoding()); + break; + case Operand::MEM_SCALE: + masm.addq_mr(src.disp(), src.base(), src.index(), src.scale(), + dest.encoding()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + void addq(Register src, const Operand& dest) { + switch (dest.kind()) { + case Operand::REG: + masm.addq_rr(src.encoding(), dest.reg()); + break; + case Operand::MEM_REG_DISP: + masm.addq_rm(src.encoding(), dest.disp(), dest.base()); + break; + case Operand::MEM_SCALE: + masm.addq_rm(src.encoding(), dest.disp(), dest.base(), dest.index(), + dest.scale()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + + void subq(Imm32 imm, Register dest) { + masm.subq_ir(imm.value, dest.encoding()); + } + void subq(Register src, Register dest) { + masm.subq_rr(src.encoding(), dest.encoding()); + } + void subq(const Operand& src, Register dest) { + switch (src.kind()) { + case Operand::REG: + masm.subq_rr(src.reg(), dest.encoding()); + break; + case Operand::MEM_REG_DISP: + masm.subq_mr(src.disp(), src.base(), dest.encoding()); + break; + case Operand::MEM_ADDRESS32: + masm.subq_mr(src.address(), dest.encoding()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + void subq(Register src, const Operand& dest) { + switch (dest.kind()) { + case Operand::REG: + masm.subq_rr(src.encoding(), dest.reg()); + break; + case Operand::MEM_REG_DISP: + masm.subq_rm(src.encoding(), dest.disp(), dest.base()); + break; + case Operand::MEM_SCALE: + masm.subq_rm(src.encoding(), dest.disp(), dest.base(), dest.index(), + dest.scale()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + void shlq(Imm32 imm, Register dest) { + masm.shlq_ir(imm.value, dest.encoding()); + } + void shrq(Imm32 imm, Register dest) { + masm.shrq_ir(imm.value, dest.encoding()); + } + void sarq(Imm32 imm, Register dest) { + masm.sarq_ir(imm.value, dest.encoding()); + } + void shlq_cl(Register dest) { masm.shlq_CLr(dest.encoding()); } + void shrq_cl(Register dest) { masm.shrq_CLr(dest.encoding()); } + void sarq_cl(Register dest) { masm.sarq_CLr(dest.encoding()); } + void sarxq(Register src, Register shift, Register dest) { + MOZ_ASSERT(HasBMI2()); + masm.sarxq_rrr(src.encoding(), shift.encoding(), dest.encoding()); + } + void shlxq(Register src, Register shift, Register dest) { + MOZ_ASSERT(HasBMI2()); + masm.shlxq_rrr(src.encoding(), shift.encoding(), dest.encoding()); + } + void shrxq(Register src, Register shift, Register dest) { + MOZ_ASSERT(HasBMI2()); + masm.shrxq_rrr(src.encoding(), shift.encoding(), dest.encoding()); + } + void rolq(Imm32 imm, Register dest) { + masm.rolq_ir(imm.value, dest.encoding()); + } + void rolq_cl(Register dest) { masm.rolq_CLr(dest.encoding()); } + void rorq(Imm32 imm, Register dest) { + masm.rorq_ir(imm.value, dest.encoding()); + } + void rorq_cl(Register dest) { masm.rorq_CLr(dest.encoding()); } + void orq(Imm32 imm, Register dest) { + masm.orq_ir(imm.value, dest.encoding()); + } + void orq(Register src, Register dest) { + masm.orq_rr(src.encoding(), dest.encoding()); + } + void orq(const Operand& src, Register dest) { + switch (src.kind()) { + case Operand::REG: + masm.orq_rr(src.reg(), dest.encoding()); + break; + case Operand::MEM_REG_DISP: + masm.orq_mr(src.disp(), src.base(), dest.encoding()); + break; + case Operand::MEM_ADDRESS32: + masm.orq_mr(src.address(), dest.encoding()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + void orq(Register src, const Operand& dest) { + switch (dest.kind()) { + case Operand::REG: + masm.orq_rr(src.encoding(), dest.reg()); + break; + case Operand::MEM_REG_DISP: + masm.orq_rm(src.encoding(), dest.disp(), dest.base()); + break; + case Operand::MEM_SCALE: + masm.orq_rm(src.encoding(), dest.disp(), dest.base(), dest.index(), + dest.scale()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + void xorq(Register src, Register dest) { + masm.xorq_rr(src.encoding(), dest.encoding()); + } + void xorq(Imm32 imm, Register dest) { + masm.xorq_ir(imm.value, dest.encoding()); + } + void xorq(const Operand& src, Register dest) { + switch (src.kind()) { + case Operand::REG: + masm.xorq_rr(src.reg(), dest.encoding()); + break; + case Operand::MEM_REG_DISP: + masm.xorq_mr(src.disp(), src.base(), dest.encoding()); + break; + case Operand::MEM_SCALE: + masm.xorq_mr(src.disp(), src.base(), src.index(), src.scale(), + dest.encoding()); + break; + case Operand::MEM_ADDRESS32: + masm.xorq_mr(src.address(), dest.encoding()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + void xorq(Register src, const Operand& dest) { + switch (dest.kind()) { + case Operand::REG: + masm.xorq_rr(src.encoding(), dest.reg()); + break; + case Operand::MEM_REG_DISP: + masm.xorq_rm(src.encoding(), dest.disp(), dest.base()); + break; + case Operand::MEM_SCALE: + masm.xorq_rm(src.encoding(), dest.disp(), dest.base(), dest.index(), + dest.scale()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + + void bsrq(const Register& src, const Register& dest) { + masm.bsrq_rr(src.encoding(), dest.encoding()); + } + void bsfq(const Register& src, const Register& dest) { + masm.bsfq_rr(src.encoding(), dest.encoding()); + } + void bswapq(const Register& reg) { masm.bswapq_r(reg.encoding()); } + void lzcntq(const Register& src, const Register& dest) { + masm.lzcntq_rr(src.encoding(), dest.encoding()); + } + void tzcntq(const Register& src, const Register& dest) { + masm.tzcntq_rr(src.encoding(), dest.encoding()); + } + void popcntq(const Register& src, const Register& dest) { + masm.popcntq_rr(src.encoding(), dest.encoding()); + } + + void imulq(Imm32 imm, Register src, Register dest) { + masm.imulq_ir(imm.value, src.encoding(), dest.encoding()); + } + void imulq(Register src, Register dest) { + masm.imulq_rr(src.encoding(), dest.encoding()); + } + void imulq(const Operand& src, Register dest) { + switch (src.kind()) { + case Operand::REG: + masm.imulq_rr(src.reg(), dest.encoding()); + break; + case Operand::MEM_REG_DISP: + masm.imulq_mr(src.disp(), src.base(), dest.encoding()); + break; + case Operand::MEM_ADDRESS32: + MOZ_CRASH("NYI"); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + + void cqo() { masm.cqo(); } + void idivq(Register divisor) { masm.idivq_r(divisor.encoding()); } + void udivq(Register divisor) { masm.divq_r(divisor.encoding()); } + + void vcvtsi2sdq(Register src, FloatRegister dest) { + masm.vcvtsi2sdq_rr(src.encoding(), dest.encoding()); + } + + void vpextrq(unsigned lane, FloatRegister src, Register dest) { + MOZ_ASSERT(HasSSE41()); + masm.vpextrq_irr(lane, src.encoding(), dest.encoding()); + } + + void vpinsrq(unsigned lane, Register src1, FloatRegister src0, + FloatRegister dest) { + MOZ_ASSERT(HasSSE41()); + masm.vpinsrq_irr(lane, src1.encoding(), src0.encoding(), dest.encoding()); + } + + void negq(Register reg) { masm.negq_r(reg.encoding()); } + + void notq(Register reg) { masm.notq_r(reg.encoding()); } + + void mov(ImmWord word, Register dest) { + // Use xor for setting registers to zero, as it is specially optimized + // for this purpose on modern hardware. Note that it does clobber FLAGS + // though. Use xorl instead of xorq since they are functionally + // equivalent (32-bit instructions zero-extend their results to 64 bits) + // and xorl has a smaller encoding. + if (word.value == 0) { + xorl(dest, dest); + } else { + movq(word, dest); + } + } + void mov(ImmPtr imm, Register dest) { movq(imm, dest); } + void mov(wasm::SymbolicAddress imm, Register dest) { + masm.movq_i64r(-1, dest.encoding()); + append(wasm::SymbolicAccess(CodeOffset(masm.currentOffset()), imm)); + } + void mov(const Operand& src, Register dest) { movq(src, dest); } + void mov(Register src, const Operand& dest) { movq(src, dest); } + void mov(Imm32 imm32, const Operand& dest) { movq(imm32, dest); } + void mov(Register src, Register dest) { movq(src, dest); } + void mov(CodeLabel* label, Register dest) { + masm.movq_i64r(/* placeholder */ 0, dest.encoding()); + label->patchAt()->bind(masm.size()); + } + void xchg(Register src, Register dest) { xchgq(src, dest); } + + void lea(const Operand& src, Register dest) { + switch (src.kind()) { + case Operand::MEM_REG_DISP: + masm.leaq_mr(src.disp(), src.base(), dest.encoding()); + break; + case Operand::MEM_SCALE: + masm.leaq_mr(src.disp(), src.base(), src.index(), src.scale(), + dest.encoding()); + break; + default: + MOZ_CRASH("unexepcted operand kind"); + } + } + + void cmovz32(const Operand& src, Register dest) { return cmovzl(src, dest); } + void cmovzPtr(const Operand& src, Register dest) { return cmovzq(src, dest); } + + CodeOffset loadRipRelativeInt32(Register dest) { + return CodeOffset(masm.movl_ripr(dest.encoding()).offset()); + } + CodeOffset loadRipRelativeInt64(Register dest) { + return CodeOffset(masm.movq_ripr(dest.encoding()).offset()); + } + CodeOffset loadRipRelativeDouble(FloatRegister dest) { + return CodeOffset(masm.vmovsd_ripr(dest.encoding()).offset()); + } + CodeOffset loadRipRelativeFloat32(FloatRegister dest) { + return CodeOffset(masm.vmovss_ripr(dest.encoding()).offset()); + } + CodeOffset loadRipRelativeInt32x4(FloatRegister dest) { + return CodeOffset(masm.vmovdqa_ripr(dest.encoding()).offset()); + } + CodeOffset loadRipRelativeFloat32x4(FloatRegister dest) { + return CodeOffset(masm.vmovaps_ripr(dest.encoding()).offset()); + } + CodeOffset leaRipRelative(Register dest) { + return CodeOffset(masm.leaq_rip(dest.encoding()).offset()); + } + + void cmpq(Register rhs, Register lhs) { + masm.cmpq_rr(rhs.encoding(), lhs.encoding()); + } + void cmpq(Register rhs, const Operand& lhs) { + switch (lhs.kind()) { + case Operand::REG: + masm.cmpq_rr(rhs.encoding(), lhs.reg()); + break; + case Operand::MEM_REG_DISP: + masm.cmpq_rm(rhs.encoding(), lhs.disp(), lhs.base()); + break; + case Operand::MEM_SCALE: + masm.cmpq_rm(rhs.encoding(), lhs.disp(), lhs.base(), lhs.index(), + lhs.scale()); + break; + case Operand::MEM_ADDRESS32: + masm.cmpq_rm(rhs.encoding(), lhs.address()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + void cmpq(Imm32 rhs, Register lhs) { + masm.cmpq_ir(rhs.value, lhs.encoding()); + } + void cmpq(Imm32 rhs, const Operand& lhs) { + switch (lhs.kind()) { + case Operand::REG: + masm.cmpq_ir(rhs.value, lhs.reg()); + break; + case Operand::MEM_REG_DISP: + masm.cmpq_im(rhs.value, lhs.disp(), lhs.base()); + break; + case Operand::MEM_SCALE: + masm.cmpq_im(rhs.value, lhs.disp(), lhs.base(), lhs.index(), + lhs.scale()); + break; + case Operand::MEM_ADDRESS32: + masm.cmpq_im(rhs.value, lhs.address()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + void cmpq(const Operand& rhs, Register lhs) { + switch (rhs.kind()) { + case Operand::REG: + masm.cmpq_rr(rhs.reg(), lhs.encoding()); + break; + case Operand::MEM_REG_DISP: + masm.cmpq_mr(rhs.disp(), rhs.base(), lhs.encoding()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + } + } + + void testq(Imm32 rhs, Register lhs) { + masm.testq_ir(rhs.value, lhs.encoding()); + } + void testq(Register rhs, Register lhs) { + masm.testq_rr(rhs.encoding(), lhs.encoding()); + } + void testq(Imm32 rhs, const Operand& lhs) { + switch (lhs.kind()) { + case Operand::REG: + masm.testq_ir(rhs.value, lhs.reg()); + break; + case Operand::MEM_REG_DISP: + masm.testq_i32m(rhs.value, lhs.disp(), lhs.base()); + break; + default: + MOZ_CRASH("unexpected operand kind"); + break; + } + } + + void jmp(ImmPtr target, RelocationKind reloc = RelocationKind::HARDCODED) { + MOZ_ASSERT(hasCreator()); + JmpSrc src = masm.jmp(); + addPendingJump(src, target, reloc); + } + void j(Condition cond, ImmPtr target, + RelocationKind reloc = RelocationKind::HARDCODED) { + JmpSrc src = masm.jCC(static_cast<X86Encoding::Condition>(cond)); + addPendingJump(src, target, reloc); + } + + void jmp(JitCode* target) { + jmp(ImmPtr(target->raw()), RelocationKind::JITCODE); + } + void j(Condition cond, JitCode* target) { + j(cond, ImmPtr(target->raw()), RelocationKind::JITCODE); + } + void call(JitCode* target) { + JmpSrc src = masm.call(); + addPendingJump(src, ImmPtr(target->raw()), RelocationKind::JITCODE); + } + void call(ImmWord target) { call(ImmPtr((void*)target.value)); } + void call(ImmPtr target) { + JmpSrc src = masm.call(); + addPendingJump(src, target, RelocationKind::HARDCODED); + } + + // Emit a CALL or CMP (nop) instruction. ToggleCall can be used to patch + // this instruction. + CodeOffset toggledCall(JitCode* target, bool enabled) { + CodeOffset offset(size()); + JmpSrc src = enabled ? masm.call() : masm.cmp_eax(); + addPendingJump(src, ImmPtr(target->raw()), RelocationKind::JITCODE); + MOZ_ASSERT_IF(!oom(), size() - offset.offset() == ToggledCallSize(nullptr)); + return offset; + } + + static size_t ToggledCallSize(uint8_t* code) { + // Size of a call instruction. + return 5; + } + + // Do not mask shared implementations. + using AssemblerX86Shared::call; + + void vcvttsd2sq(FloatRegister src, Register dest) { + masm.vcvttsd2sq_rr(src.encoding(), dest.encoding()); + } + void vcvttss2sq(FloatRegister src, Register dest) { + masm.vcvttss2sq_rr(src.encoding(), dest.encoding()); + } + void vcvtsq2sd(Register src1, FloatRegister src0, FloatRegister dest) { + masm.vcvtsq2sd_rr(src1.encoding(), src0.encoding(), dest.encoding()); + } + void vcvtsq2ss(Register src1, FloatRegister src0, FloatRegister dest) { + masm.vcvtsq2ss_rr(src1.encoding(), src0.encoding(), dest.encoding()); + } +}; + +static inline bool GetIntArgReg(uint32_t intArg, uint32_t floatArg, + Register* out) { +#if defined(_WIN64) + uint32_t arg = intArg + floatArg; +#else + uint32_t arg = intArg; +#endif + if (arg >= NumIntArgRegs) { + return false; + } + *out = IntArgRegs[arg]; + return true; +} + +// Get a register in which we plan to put a quantity that will be used as an +// integer argument. This differs from GetIntArgReg in that if we have no more +// actual argument registers to use we will fall back on using whatever +// CallTempReg* don't overlap the argument registers, and only fail once those +// run out too. +static inline bool GetTempRegForIntArg(uint32_t usedIntArgs, + uint32_t usedFloatArgs, Register* out) { + if (GetIntArgReg(usedIntArgs, usedFloatArgs, out)) { + return true; + } + // Unfortunately, we have to assume things about the point at which + // GetIntArgReg returns false, because we need to know how many registers it + // can allocate. +#if defined(_WIN64) + uint32_t arg = usedIntArgs + usedFloatArgs; +#else + uint32_t arg = usedIntArgs; +#endif + arg -= NumIntArgRegs; + if (arg >= NumCallTempNonArgRegs) { + return false; + } + *out = CallTempNonArgRegs[arg]; + return true; +} + +static inline bool GetFloatArgReg(uint32_t intArg, uint32_t floatArg, + FloatRegister* out) { +#if defined(_WIN64) + uint32_t arg = intArg + floatArg; +#else + uint32_t arg = floatArg; +#endif + if (floatArg >= NumFloatArgRegs) { + return false; + } + *out = FloatArgRegs[arg]; + return true; +} + +} // namespace jit +} // namespace js + +#endif /* jit_x64_Assembler_x64_h */ diff --git a/js/src/jit/x64/BaseAssembler-x64.h b/js/src/jit/x64/BaseAssembler-x64.h new file mode 100644 index 0000000000..f5a9bb99f9 --- /dev/null +++ b/js/src/jit/x64/BaseAssembler-x64.h @@ -0,0 +1,1373 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#ifndef jit_x64_BaseAssembler_x64_h +#define jit_x64_BaseAssembler_x64_h + +#include "jit/x86-shared/BaseAssembler-x86-shared.h" + +namespace js { +namespace jit { + +namespace X86Encoding { + +class BaseAssemblerX64 : public BaseAssembler { + public: + // Arithmetic operations: + + void addq_rr(RegisterID src, RegisterID dst) { + spew("addq %s, %s", GPReg64Name(src), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_ADD_GvEv, src, dst); + } + + void addq_mr(int32_t offset, RegisterID base, RegisterID dst) { + spew("addq " MEM_ob ", %s", ADDR_ob(offset, base), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_ADD_GvEv, offset, base, dst); + } + + void addq_mr(const void* addr, RegisterID dst) { + spew("addq %p, %s", addr, GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_ADD_GvEv, addr, dst); + } + + void addq_mr(int32_t offset, RegisterID base, RegisterID index, int scale, + RegisterID dst) { + spew("addq " MEM_obs ", %s", ADDR_obs(offset, base, index, scale), + GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_ADD_GvEv, offset, base, index, scale, dst); + } + + void addq_rm(RegisterID src, int32_t offset, RegisterID base) { + spew("addq %s, " MEM_ob, GPReg64Name(src), ADDR_ob(offset, base)); + m_formatter.oneByteOp64(OP_ADD_EvGv, offset, base, src); + } + + void addq_rm(RegisterID src, int32_t offset, RegisterID base, + RegisterID index, int scale) { + spew("addq %s, " MEM_obs, GPReg64Name(src), + ADDR_obs(offset, base, index, scale)); + m_formatter.oneByteOp64(OP_ADD_EvGv, offset, base, index, scale, src); + } + + void addq_ir(int32_t imm, RegisterID dst) { + spew("addq $%d, %s", imm, GPReg64Name(dst)); + if (CAN_SIGN_EXTEND_8_32(imm)) { + m_formatter.oneByteOp64(OP_GROUP1_EvIb, dst, GROUP1_OP_ADD); + m_formatter.immediate8s(imm); + } else { + if (dst == rax) { + m_formatter.oneByteOp64(OP_ADD_EAXIv); + } else { + m_formatter.oneByteOp64(OP_GROUP1_EvIz, dst, GROUP1_OP_ADD); + } + m_formatter.immediate32(imm); + } + } + + void addq_i32r(int32_t imm, RegisterID dst) { + // 32-bit immediate always, for patching. + spew("addq $0x%04x, %s", uint32_t(imm), GPReg64Name(dst)); + if (dst == rax) { + m_formatter.oneByteOp64(OP_ADD_EAXIv); + } else { + m_formatter.oneByteOp64(OP_GROUP1_EvIz, dst, GROUP1_OP_ADD); + } + m_formatter.immediate32(imm); + } + + void addq_im(int32_t imm, int32_t offset, RegisterID base) { + spew("addq $%d, " MEM_ob, imm, ADDR_ob(offset, base)); + if (CAN_SIGN_EXTEND_8_32(imm)) { + m_formatter.oneByteOp64(OP_GROUP1_EvIb, offset, base, GROUP1_OP_ADD); + m_formatter.immediate8s(imm); + } else { + m_formatter.oneByteOp64(OP_GROUP1_EvIz, offset, base, GROUP1_OP_ADD); + m_formatter.immediate32(imm); + } + } + + void addq_im(int32_t imm, const void* addr) { + spew("addq $%d, %p", imm, addr); + if (CAN_SIGN_EXTEND_8_32(imm)) { + m_formatter.oneByteOp64(OP_GROUP1_EvIb, addr, GROUP1_OP_ADD); + m_formatter.immediate8s(imm); + } else { + m_formatter.oneByteOp64(OP_GROUP1_EvIz, addr, GROUP1_OP_ADD); + m_formatter.immediate32(imm); + } + } + + void andq_rr(RegisterID src, RegisterID dst) { + spew("andq %s, %s", GPReg64Name(src), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_AND_GvEv, src, dst); + } + + void andq_mr(int32_t offset, RegisterID base, RegisterID dst) { + spew("andq " MEM_ob ", %s", ADDR_ob(offset, base), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_AND_GvEv, offset, base, dst); + } + + void andq_mr(int32_t offset, RegisterID base, RegisterID index, int scale, + RegisterID dst) { + spew("andq " MEM_obs ", %s", ADDR_obs(offset, base, index, scale), + GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_AND_GvEv, offset, base, index, scale, dst); + } + + void andq_mr(const void* addr, RegisterID dst) { + spew("andq %p, %s", addr, GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_AND_GvEv, addr, dst); + } + + void andq_rm(RegisterID src, int32_t offset, RegisterID base) { + spew("andq %s, " MEM_ob, GPReg64Name(src), ADDR_ob(offset, base)); + m_formatter.oneByteOp64(OP_AND_EvGv, offset, base, src); + } + + void andq_rm(RegisterID src, int32_t offset, RegisterID base, + RegisterID index, int scale) { + spew("andq %s, " MEM_obs, GPReg64Name(src), + ADDR_obs(offset, base, index, scale)); + m_formatter.oneByteOp64(OP_AND_EvGv, offset, base, index, scale, src); + } + + void orq_mr(int32_t offset, RegisterID base, RegisterID dst) { + spew("orq " MEM_ob ", %s", ADDR_ob(offset, base), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_OR_GvEv, offset, base, dst); + } + + void orq_mr(const void* addr, RegisterID dst) { + spew("orq %p, %s", addr, GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_OR_GvEv, addr, dst); + } + + void orq_rm(RegisterID src, int32_t offset, RegisterID base) { + spew("orq %s, " MEM_ob, GPReg64Name(src), ADDR_ob(offset, base)); + m_formatter.oneByteOp64(OP_OR_EvGv, offset, base, src); + } + + void orq_rm(RegisterID src, int32_t offset, RegisterID base, RegisterID index, + int scale) { + spew("orq %s, " MEM_obs, GPReg64Name(src), + ADDR_obs(offset, base, index, scale)); + m_formatter.oneByteOp64(OP_OR_EvGv, offset, base, index, scale, src); + } + + void xorq_mr(int32_t offset, RegisterID base, RegisterID dst) { + spew("xorq " MEM_ob ", %s", ADDR_ob(offset, base), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_XOR_GvEv, offset, base, dst); + } + + void xorq_mr(int32_t offset, RegisterID base, RegisterID index, int scale, + RegisterID dst) { + spew("xorq " MEM_obs ", %s", ADDR_obs(offset, base, index, scale), + GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_XOR_GvEv, offset, base, index, scale, dst); + } + + void xorq_mr(const void* addr, RegisterID dst) { + spew("xorq %p, %s", addr, GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_XOR_GvEv, addr, dst); + } + + void xorq_rm(RegisterID src, int32_t offset, RegisterID base) { + spew("xorq %s, " MEM_ob, GPReg64Name(src), ADDR_ob(offset, base)); + m_formatter.oneByteOp64(OP_XOR_EvGv, offset, base, src); + } + + void xorq_rm(RegisterID src, int32_t offset, RegisterID base, + RegisterID index, int scale) { + spew("xorq %s, " MEM_obs, GPReg64Name(src), + ADDR_obs(offset, base, index, scale)); + m_formatter.oneByteOp64(OP_XOR_EvGv, offset, base, index, scale, src); + } + + void bswapq_r(RegisterID dst) { + spew("bswapq %s", GPReg64Name(dst)); + m_formatter.twoByteOp64(OP2_BSWAP, dst); + } + + void bsrq_rr(RegisterID src, RegisterID dst) { + spew("bsrq %s, %s", GPReg64Name(src), GPReg64Name(dst)); + m_formatter.twoByteOp64(OP2_BSR_GvEv, src, dst); + } + + void bsfq_rr(RegisterID src, RegisterID dst) { + spew("bsfq %s, %s", GPReg64Name(src), GPReg64Name(dst)); + m_formatter.twoByteOp64(OP2_BSF_GvEv, src, dst); + } + + void lzcntq_rr(RegisterID src, RegisterID dst) { + spew("lzcntq %s, %s", GPReg64Name(src), GPReg64Name(dst)); + m_formatter.legacySSEPrefix(VEX_SS); + m_formatter.twoByteOp64(OP2_LZCNT_GvEv, src, dst); + } + + void tzcntq_rr(RegisterID src, RegisterID dst) { + spew("tzcntq %s, %s", GPReg64Name(src), GPReg64Name(dst)); + m_formatter.legacySSEPrefix(VEX_SS); + m_formatter.twoByteOp64(OP2_TZCNT_GvEv, src, dst); + } + + void popcntq_rr(RegisterID src, RegisterID dst) { + spew("popcntq %s, %s", GPReg64Name(src), GPReg64Name(dst)); + m_formatter.legacySSEPrefix(VEX_SS); + m_formatter.twoByteOp64(OP2_POPCNT_GvEv, src, dst); + } + + void andq_ir(int32_t imm, RegisterID dst) { + spew("andq $0x%" PRIx64 ", %s", uint64_t(imm), GPReg64Name(dst)); + if (CAN_SIGN_EXTEND_8_32(imm)) { + m_formatter.oneByteOp64(OP_GROUP1_EvIb, dst, GROUP1_OP_AND); + m_formatter.immediate8s(imm); + } else { + if (dst == rax) { + m_formatter.oneByteOp64(OP_AND_EAXIv); + } else { + m_formatter.oneByteOp64(OP_GROUP1_EvIz, dst, GROUP1_OP_AND); + } + m_formatter.immediate32(imm); + } + } + + void negq_r(RegisterID dst) { + spew("negq %s", GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_GROUP3_Ev, dst, GROUP3_OP_NEG); + } + + void orq_rr(RegisterID src, RegisterID dst) { + spew("orq %s, %s", GPReg64Name(src), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_OR_GvEv, src, dst); + } + + void orq_ir(int32_t imm, RegisterID dst) { + spew("orq $0x%" PRIx64 ", %s", uint64_t(imm), GPReg64Name(dst)); + if (CAN_SIGN_EXTEND_8_32(imm)) { + m_formatter.oneByteOp64(OP_GROUP1_EvIb, dst, GROUP1_OP_OR); + m_formatter.immediate8s(imm); + } else { + if (dst == rax) { + m_formatter.oneByteOp64(OP_OR_EAXIv); + } else { + m_formatter.oneByteOp64(OP_GROUP1_EvIz, dst, GROUP1_OP_OR); + } + m_formatter.immediate32(imm); + } + } + + void notq_r(RegisterID dst) { + spew("notq %s", GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_GROUP3_Ev, dst, GROUP3_OP_NOT); + } + + void subq_rr(RegisterID src, RegisterID dst) { + spew("subq %s, %s", GPReg64Name(src), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_SUB_GvEv, src, dst); + } + + void subq_rm(RegisterID src, int32_t offset, RegisterID base) { + spew("subq %s, " MEM_ob, GPReg64Name(src), ADDR_ob(offset, base)); + m_formatter.oneByteOp64(OP_SUB_EvGv, offset, base, src); + } + + void subq_rm(RegisterID src, int32_t offset, RegisterID base, + RegisterID index, int scale) { + spew("subq %s, " MEM_obs, GPReg64Name(src), + ADDR_obs(offset, base, index, scale)); + m_formatter.oneByteOp64(OP_SUB_EvGv, offset, base, index, scale, src); + } + + void subq_mr(int32_t offset, RegisterID base, RegisterID dst) { + spew("subq " MEM_ob ", %s", ADDR_ob(offset, base), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_SUB_GvEv, offset, base, dst); + } + + void subq_mr(const void* addr, RegisterID dst) { + spew("subq %p, %s", addr, GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_SUB_GvEv, addr, dst); + } + + void subq_ir(int32_t imm, RegisterID dst) { + spew("subq $%d, %s", imm, GPReg64Name(dst)); + if (CAN_SIGN_EXTEND_8_32(imm)) { + m_formatter.oneByteOp64(OP_GROUP1_EvIb, dst, GROUP1_OP_SUB); + m_formatter.immediate8s(imm); + } else { + if (dst == rax) { + m_formatter.oneByteOp64(OP_SUB_EAXIv); + } else { + m_formatter.oneByteOp64(OP_GROUP1_EvIz, dst, GROUP1_OP_SUB); + } + m_formatter.immediate32(imm); + } + } + + void xorq_rr(RegisterID src, RegisterID dst) { + spew("xorq %s, %s", GPReg64Name(src), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_XOR_GvEv, src, dst); + } + + void xorq_ir(int32_t imm, RegisterID dst) { + spew("xorq $0x%" PRIx64 ", %s", uint64_t(imm), GPReg64Name(dst)); + if (CAN_SIGN_EXTEND_8_32(imm)) { + m_formatter.oneByteOp64(OP_GROUP1_EvIb, dst, GROUP1_OP_XOR); + m_formatter.immediate8s(imm); + } else { + if (dst == rax) { + m_formatter.oneByteOp64(OP_XOR_EAXIv); + } else { + m_formatter.oneByteOp64(OP_GROUP1_EvIz, dst, GROUP1_OP_XOR); + } + m_formatter.immediate32(imm); + } + } + + void sarq_CLr(RegisterID dst) { + spew("sarq %%cl, %s", GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_GROUP2_EvCL, dst, GROUP2_OP_SAR); + } + + void shlq_CLr(RegisterID dst) { + spew("shlq %%cl, %s", GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_GROUP2_EvCL, dst, GROUP2_OP_SHL); + } + + void shrq_CLr(RegisterID dst) { + spew("shrq %%cl, %s", GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_GROUP2_EvCL, dst, GROUP2_OP_SHR); + } + + void sarq_ir(int32_t imm, RegisterID dst) { + MOZ_ASSERT(imm < 64); + spew("sarq $%d, %s", imm, GPReg64Name(dst)); + if (imm == 1) { + m_formatter.oneByteOp64(OP_GROUP2_Ev1, dst, GROUP2_OP_SAR); + } else { + m_formatter.oneByteOp64(OP_GROUP2_EvIb, dst, GROUP2_OP_SAR); + m_formatter.immediate8u(imm); + } + } + + void shlq_ir(int32_t imm, RegisterID dst) { + MOZ_ASSERT(imm < 64); + spew("shlq $%d, %s", imm, GPReg64Name(dst)); + if (imm == 1) { + m_formatter.oneByteOp64(OP_GROUP2_Ev1, dst, GROUP2_OP_SHL); + } else { + m_formatter.oneByteOp64(OP_GROUP2_EvIb, dst, GROUP2_OP_SHL); + m_formatter.immediate8u(imm); + } + } + + void shrq_ir(int32_t imm, RegisterID dst) { + MOZ_ASSERT(imm < 64); + spew("shrq $%d, %s", imm, GPReg64Name(dst)); + if (imm == 1) { + m_formatter.oneByteOp64(OP_GROUP2_Ev1, dst, GROUP2_OP_SHR); + } else { + m_formatter.oneByteOp64(OP_GROUP2_EvIb, dst, GROUP2_OP_SHR); + m_formatter.immediate8u(imm); + } + } + + void rolq_ir(int32_t imm, RegisterID dst) { + MOZ_ASSERT(imm < 64); + spew("rolq $%d, %s", imm, GPReg64Name(dst)); + if (imm == 1) { + m_formatter.oneByteOp64(OP_GROUP2_Ev1, dst, GROUP2_OP_ROL); + } else { + m_formatter.oneByteOp64(OP_GROUP2_EvIb, dst, GROUP2_OP_ROL); + m_formatter.immediate8u(imm); + } + } + void rolq_CLr(RegisterID dst) { + spew("rolq %%cl, %s", GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_GROUP2_EvCL, dst, GROUP2_OP_ROL); + } + + void rorq_ir(int32_t imm, RegisterID dst) { + MOZ_ASSERT(imm < 64); + spew("rorq $%d, %s", imm, GPReg64Name(dst)); + if (imm == 1) { + m_formatter.oneByteOp64(OP_GROUP2_Ev1, dst, GROUP2_OP_ROR); + } else { + m_formatter.oneByteOp64(OP_GROUP2_EvIb, dst, GROUP2_OP_ROR); + m_formatter.immediate8u(imm); + } + } + void rorq_CLr(RegisterID dst) { + spew("rorq %%cl, %s", GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_GROUP2_EvCL, dst, GROUP2_OP_ROR); + } + + void imulq_rr(RegisterID src, RegisterID dst) { + spew("imulq %s, %s", GPReg64Name(src), GPReg64Name(dst)); + m_formatter.twoByteOp64(OP2_IMUL_GvEv, src, dst); + } + + void imulq_mr(int32_t offset, RegisterID base, RegisterID dst) { + spew("imulq " MEM_ob ", %s", ADDR_ob(offset, base), GPReg64Name(dst)); + m_formatter.twoByteOp64(OP2_IMUL_GvEv, offset, base, dst); + } + + void imulq_ir(int32_t value, RegisterID src, RegisterID dst) { + spew("imulq $%d, %s, %s", value, GPReg64Name(src), GPReg64Name(dst)); + if (CAN_SIGN_EXTEND_8_32(value)) { + m_formatter.oneByteOp64(OP_IMUL_GvEvIb, src, dst); + m_formatter.immediate8s(value); + } else { + m_formatter.oneByteOp64(OP_IMUL_GvEvIz, src, dst); + m_formatter.immediate32(value); + } + } + + void cqo() { + spew("cqo "); + m_formatter.oneByteOp64(OP_CDQ); + } + + void idivq_r(RegisterID divisor) { + spew("idivq %s", GPReg64Name(divisor)); + m_formatter.oneByteOp64(OP_GROUP3_Ev, divisor, GROUP3_OP_IDIV); + } + + void divq_r(RegisterID divisor) { + spew("divq %s", GPReg64Name(divisor)); + m_formatter.oneByteOp64(OP_GROUP3_Ev, divisor, GROUP3_OP_DIV); + } + + // Comparisons: + + void cmpq_rr(RegisterID rhs, RegisterID lhs) { + spew("cmpq %s, %s", GPReg64Name(rhs), GPReg64Name(lhs)); + m_formatter.oneByteOp64(OP_CMP_GvEv, rhs, lhs); + } + + void cmpq_rm(RegisterID rhs, int32_t offset, RegisterID base) { + spew("cmpq %s, " MEM_ob, GPReg64Name(rhs), ADDR_ob(offset, base)); + m_formatter.oneByteOp64(OP_CMP_EvGv, offset, base, rhs); + } + + void cmpq_rm(RegisterID rhs, int32_t offset, RegisterID base, + RegisterID index, int scale) { + spew("cmpq %s, " MEM_obs, GPReg64Name(rhs), + ADDR_obs(offset, base, index, scale)); + m_formatter.oneByteOp64(OP_CMP_EvGv, offset, base, index, scale, rhs); + } + + void cmpq_mr(int32_t offset, RegisterID base, RegisterID lhs) { + spew("cmpq " MEM_ob ", %s", ADDR_ob(offset, base), GPReg64Name(lhs)); + m_formatter.oneByteOp64(OP_CMP_GvEv, offset, base, lhs); + } + + void cmpq_ir(int32_t rhs, RegisterID lhs) { + if (rhs == 0) { + testq_rr(lhs, lhs); + return; + } + + spew("cmpq $0x%" PRIx64 ", %s", uint64_t(rhs), GPReg64Name(lhs)); + if (CAN_SIGN_EXTEND_8_32(rhs)) { + m_formatter.oneByteOp64(OP_GROUP1_EvIb, lhs, GROUP1_OP_CMP); + m_formatter.immediate8s(rhs); + } else { + if (lhs == rax) { + m_formatter.oneByteOp64(OP_CMP_EAXIv); + } else { + m_formatter.oneByteOp64(OP_GROUP1_EvIz, lhs, GROUP1_OP_CMP); + } + m_formatter.immediate32(rhs); + } + } + + void cmpq_im(int32_t rhs, int32_t offset, RegisterID base) { + spew("cmpq $0x%" PRIx64 ", " MEM_ob, uint64_t(rhs), + ADDR_ob(offset, base)); + if (CAN_SIGN_EXTEND_8_32(rhs)) { + m_formatter.oneByteOp64(OP_GROUP1_EvIb, offset, base, GROUP1_OP_CMP); + m_formatter.immediate8s(rhs); + } else { + m_formatter.oneByteOp64(OP_GROUP1_EvIz, offset, base, GROUP1_OP_CMP); + m_formatter.immediate32(rhs); + } + } + + void cmpq_im(int32_t rhs, int32_t offset, RegisterID base, RegisterID index, + int scale) { + spew("cmpq $0x%x, " MEM_obs, uint32_t(rhs), + ADDR_obs(offset, base, index, scale)); + if (CAN_SIGN_EXTEND_8_32(rhs)) { + m_formatter.oneByteOp64(OP_GROUP1_EvIb, offset, base, index, scale, + GROUP1_OP_CMP); + m_formatter.immediate8s(rhs); + } else { + m_formatter.oneByteOp64(OP_GROUP1_EvIz, offset, base, index, scale, + GROUP1_OP_CMP); + m_formatter.immediate32(rhs); + } + } + void cmpq_im(int32_t rhs, const void* addr) { + spew("cmpq $0x%" PRIx64 ", %p", uint64_t(rhs), addr); + if (CAN_SIGN_EXTEND_8_32(rhs)) { + m_formatter.oneByteOp64(OP_GROUP1_EvIb, addr, GROUP1_OP_CMP); + m_formatter.immediate8s(rhs); + } else { + m_formatter.oneByteOp64(OP_GROUP1_EvIz, addr, GROUP1_OP_CMP); + m_formatter.immediate32(rhs); + } + } + void cmpq_rm(RegisterID rhs, const void* addr) { + spew("cmpq %s, %p", GPReg64Name(rhs), addr); + m_formatter.oneByteOp64(OP_CMP_EvGv, addr, rhs); + } + + void testq_rr(RegisterID rhs, RegisterID lhs) { + spew("testq %s, %s", GPReg64Name(rhs), GPReg64Name(lhs)); + m_formatter.oneByteOp64(OP_TEST_EvGv, lhs, rhs); + } + + void testq_ir(int32_t rhs, RegisterID lhs) { + // If the mask fits in a 32-bit immediate, we can use testl with a + // 32-bit subreg. + if (CAN_ZERO_EXTEND_32_64(rhs)) { + testl_ir(rhs, lhs); + return; + } + spew("testq $0x%" PRIx64 ", %s", uint64_t(rhs), GPReg64Name(lhs)); + if (lhs == rax) { + m_formatter.oneByteOp64(OP_TEST_EAXIv); + } else { + m_formatter.oneByteOp64(OP_GROUP3_EvIz, lhs, GROUP3_OP_TEST); + } + m_formatter.immediate32(rhs); + } + + void testq_i32m(int32_t rhs, int32_t offset, RegisterID base) { + spew("testq $0x%" PRIx64 ", " MEM_ob, uint64_t(rhs), + ADDR_ob(offset, base)); + m_formatter.oneByteOp64(OP_GROUP3_EvIz, offset, base, GROUP3_OP_TEST); + m_formatter.immediate32(rhs); + } + + void testq_i32m(int32_t rhs, int32_t offset, RegisterID base, + RegisterID index, int scale) { + spew("testq $0x%4x, " MEM_obs, uint32_t(rhs), + ADDR_obs(offset, base, index, scale)); + m_formatter.oneByteOp64(OP_GROUP3_EvIz, offset, base, index, scale, + GROUP3_OP_TEST); + m_formatter.immediate32(rhs); + } + + // Various move ops: + + void cmovCCq_rr(Condition cond, RegisterID src, RegisterID dst) { + spew("cmov%s %s, %s", CCName(cond), GPReg64Name(src), GPReg64Name(dst)); + m_formatter.twoByteOp64(cmovccOpcode(cond), src, dst); + } + void cmovCCq_mr(Condition cond, int32_t offset, RegisterID base, + RegisterID dst) { + spew("cmov%s " MEM_ob ", %s", CCName(cond), ADDR_ob(offset, base), + GPReg64Name(dst)); + m_formatter.twoByteOp64(cmovccOpcode(cond), offset, base, dst); + } + void cmovCCq_mr(Condition cond, int32_t offset, RegisterID base, + RegisterID index, int scale, RegisterID dst) { + spew("cmov%s " MEM_obs ", %s", CCName(cond), + ADDR_obs(offset, base, index, scale), GPReg64Name(dst)); + m_formatter.twoByteOp64(cmovccOpcode(cond), offset, base, index, scale, + dst); + } + + void cmpxchgq(RegisterID src, int32_t offset, RegisterID base) { + spew("cmpxchgq %s, " MEM_ob, GPReg64Name(src), ADDR_ob(offset, base)); + m_formatter.twoByteOp64(OP2_CMPXCHG_GvEw, offset, base, src); + } + + void cmpxchgq(RegisterID src, int32_t offset, RegisterID base, + RegisterID index, int scale) { + spew("cmpxchgq %s, " MEM_obs, GPReg64Name(src), + ADDR_obs(offset, base, index, scale)); + m_formatter.twoByteOp64(OP2_CMPXCHG_GvEw, offset, base, index, scale, src); + } + + void lock_xaddq_rm(RegisterID srcdest, int32_t offset, RegisterID base) { + spew("lock xaddq %s, " MEM_ob, GPReg64Name(srcdest), ADDR_ob(offset, base)); + m_formatter.oneByteOp(PRE_LOCK); + m_formatter.twoByteOp64(OP2_XADD_EvGv, offset, base, srcdest); + } + + void lock_xaddq_rm(RegisterID srcdest, int32_t offset, RegisterID base, + RegisterID index, int scale) { + spew("lock xaddq %s, " MEM_obs, GPReg64Name(srcdest), + ADDR_obs(offset, base, index, scale)); + m_formatter.oneByteOp(PRE_LOCK); + m_formatter.twoByteOp64(OP2_XADD_EvGv, offset, base, index, scale, srcdest); + } + + void xchgq_rr(RegisterID src, RegisterID dst) { + spew("xchgq %s, %s", GPReg64Name(src), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_XCHG_GvEv, src, dst); + } + void xchgq_rm(RegisterID src, int32_t offset, RegisterID base) { + spew("xchgq %s, " MEM_ob, GPReg64Name(src), ADDR_ob(offset, base)); + m_formatter.oneByteOp64(OP_XCHG_GvEv, offset, base, src); + } + void xchgq_rm(RegisterID src, int32_t offset, RegisterID base, + RegisterID index, int scale) { + spew("xchgq %s, " MEM_obs, GPReg64Name(src), + ADDR_obs(offset, base, index, scale)); + m_formatter.oneByteOp64(OP_XCHG_GvEv, offset, base, index, scale, src); + } + + void movq_rr(RegisterID src, RegisterID dst) { + spew("movq %s, %s", GPReg64Name(src), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_MOV_EvGv, dst, src); + } + + void movq_rm(RegisterID src, int32_t offset, RegisterID base) { + spew("movq %s, " MEM_ob, GPReg64Name(src), ADDR_ob(offset, base)); + m_formatter.oneByteOp64(OP_MOV_EvGv, offset, base, src); + } + + void movq_rm_disp32(RegisterID src, int32_t offset, RegisterID base) { + spew("movq %s, " MEM_o32b, GPReg64Name(src), ADDR_o32b(offset, base)); + m_formatter.oneByteOp64_disp32(OP_MOV_EvGv, offset, base, src); + } + + void movq_rm(RegisterID src, int32_t offset, RegisterID base, + RegisterID index, int scale) { + spew("movq %s, " MEM_obs, GPReg64Name(src), + ADDR_obs(offset, base, index, scale)); + m_formatter.oneByteOp64(OP_MOV_EvGv, offset, base, index, scale, src); + } + + void movq_rm(RegisterID src, const void* addr) { + if (src == rax && !IsAddressImmediate(addr)) { + movq_EAXm(addr); + return; + } + + spew("movq %s, %p", GPReg64Name(src), addr); + m_formatter.oneByteOp64(OP_MOV_EvGv, addr, src); + } + + void movq_mEAX(const void* addr) { + if (IsAddressImmediate(addr)) { + movq_mr(addr, rax); + return; + } + + spew("movq %p, %%rax", addr); + m_formatter.oneByteOp64(OP_MOV_EAXOv); + m_formatter.immediate64(reinterpret_cast<int64_t>(addr)); + } + + void movq_EAXm(const void* addr) { + if (IsAddressImmediate(addr)) { + movq_rm(rax, addr); + return; + } + + spew("movq %%rax, %p", addr); + m_formatter.oneByteOp64(OP_MOV_OvEAX); + m_formatter.immediate64(reinterpret_cast<int64_t>(addr)); + } + + void movq_mr(int32_t offset, RegisterID base, RegisterID dst) { + spew("movq " MEM_ob ", %s", ADDR_ob(offset, base), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_MOV_GvEv, offset, base, dst); + } + + void movq_mr_disp32(int32_t offset, RegisterID base, RegisterID dst) { + spew("movq " MEM_o32b ", %s", ADDR_o32b(offset, base), + GPReg64Name(dst)); + m_formatter.oneByteOp64_disp32(OP_MOV_GvEv, offset, base, dst); + } + + void movq_mr(int32_t offset, RegisterID base, RegisterID index, int scale, + RegisterID dst) { + spew("movq " MEM_obs ", %s", ADDR_obs(offset, base, index, scale), + GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_MOV_GvEv, offset, base, index, scale, dst); + } + + void movq_mr(const void* addr, RegisterID dst) { + if (dst == rax && !IsAddressImmediate(addr)) { + movq_mEAX(addr); + return; + } + + spew("movq %p, %s", addr, GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_MOV_GvEv, addr, dst); + } + + void leaq_mr(int32_t offset, RegisterID base, RegisterID index, int scale, + RegisterID dst) { + spew("leaq " MEM_obs ", %s", ADDR_obs(offset, base, index, scale), + GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_LEA, offset, base, index, scale, dst); + } + + void movq_i32m(int32_t imm, int32_t offset, RegisterID base) { + spew("movq $%d, " MEM_ob, imm, ADDR_ob(offset, base)); + m_formatter.oneByteOp64(OP_GROUP11_EvIz, offset, base, GROUP11_MOV); + m_formatter.immediate32(imm); + } + void movq_i32m(int32_t imm, int32_t offset, RegisterID base, RegisterID index, + int scale) { + spew("movq $%d, " MEM_obs, imm, ADDR_obs(offset, base, index, scale)); + m_formatter.oneByteOp64(OP_GROUP11_EvIz, offset, base, index, scale, + GROUP11_MOV); + m_formatter.immediate32(imm); + } + void movq_i32m(int32_t imm, const void* addr) { + spew("movq $%d, %p", imm, addr); + m_formatter.oneByteOp64(OP_GROUP11_EvIz, addr, GROUP11_MOV); + m_formatter.immediate32(imm); + } + + // Note that this instruction sign-extends its 32-bit immediate field to 64 + // bits and loads the 64-bit value into a 64-bit register. + // + // Note also that this is similar to the movl_i32r instruction, except that + // movl_i32r *zero*-extends its 32-bit immediate, and it has smaller code + // size, so it's preferred for values which could use either. + void movq_i32r(int32_t imm, RegisterID dst) { + spew("movq $%d, %s", imm, GPRegName(dst)); + m_formatter.oneByteOp64(OP_GROUP11_EvIz, dst, GROUP11_MOV); + m_formatter.immediate32(imm); + } + + void movq_i64r(int64_t imm, RegisterID dst) { + spew("movabsq $0x%" PRIx64 ", %s", uint64_t(imm), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_MOV_EAXIv, dst); + m_formatter.immediate64(imm); + } + + void movsbq_rr(RegisterID src, RegisterID dst) { + spew("movsbq %s, %s", GPReg32Name(src), GPReg64Name(dst)); + m_formatter.twoByteOp64(OP2_MOVSX_GvEb, src, dst); + } + void movsbq_mr(int32_t offset, RegisterID base, RegisterID dst) { + spew("movsbq " MEM_ob ", %s", ADDR_ob(offset, base), GPReg64Name(dst)); + m_formatter.twoByteOp64(OP2_MOVSX_GvEb, offset, base, dst); + } + void movsbq_mr(int32_t offset, RegisterID base, RegisterID index, int scale, + RegisterID dst) { + spew("movsbq " MEM_obs ", %s", ADDR_obs(offset, base, index, scale), + GPReg64Name(dst)); + m_formatter.twoByteOp64(OP2_MOVSX_GvEb, offset, base, index, scale, dst); + } + + void movswq_rr(RegisterID src, RegisterID dst) { + spew("movswq %s, %s", GPReg32Name(src), GPReg64Name(dst)); + m_formatter.twoByteOp64(OP2_MOVSX_GvEw, src, dst); + } + void movswq_mr(int32_t offset, RegisterID base, RegisterID dst) { + spew("movswq " MEM_ob ", %s", ADDR_ob(offset, base), GPReg64Name(dst)); + m_formatter.twoByteOp64(OP2_MOVSX_GvEw, offset, base, dst); + } + void movswq_mr(int32_t offset, RegisterID base, RegisterID index, int scale, + RegisterID dst) { + spew("movswq " MEM_obs ", %s", ADDR_obs(offset, base, index, scale), + GPReg64Name(dst)); + m_formatter.twoByteOp64(OP2_MOVSX_GvEw, offset, base, index, scale, dst); + } + + void movslq_rr(RegisterID src, RegisterID dst) { + spew("movslq %s, %s", GPReg32Name(src), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_MOVSXD_GvEv, src, dst); + } + void movslq_mr(int32_t offset, RegisterID base, RegisterID dst) { + spew("movslq " MEM_ob ", %s", ADDR_ob(offset, base), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_MOVSXD_GvEv, offset, base, dst); + } + void movslq_mr(int32_t offset, RegisterID base, RegisterID index, int scale, + RegisterID dst) { + spew("movslq " MEM_obs ", %s", ADDR_obs(offset, base, index, scale), + GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_MOVSXD_GvEv, offset, base, index, scale, dst); + } + + [[nodiscard]] JmpSrc movl_ripr(RegisterID dst) { + m_formatter.oneByteRipOp(OP_MOV_GvEv, 0, (RegisterID)dst); + JmpSrc label(m_formatter.size()); + spew("movl " MEM_o32r ", %s", ADDR_o32r(label.offset()), + GPReg32Name(dst)); + return label; + } + + [[nodiscard]] JmpSrc movl_rrip(RegisterID src) { + m_formatter.oneByteRipOp(OP_MOV_EvGv, 0, (RegisterID)src); + JmpSrc label(m_formatter.size()); + spew("movl %s, " MEM_o32r "", GPReg32Name(src), + ADDR_o32r(label.offset())); + return label; + } + + [[nodiscard]] JmpSrc movq_ripr(RegisterID dst) { + m_formatter.oneByteRipOp64(OP_MOV_GvEv, 0, dst); + JmpSrc label(m_formatter.size()); + spew("movq " MEM_o32r ", %s", ADDR_o32r(label.offset()), + GPRegName(dst)); + return label; + } + + [[nodiscard]] JmpSrc movq_rrip(RegisterID src) { + m_formatter.oneByteRipOp64(OP_MOV_EvGv, 0, (RegisterID)src); + JmpSrc label(m_formatter.size()); + spew("movq %s, " MEM_o32r "", GPRegName(src), + ADDR_o32r(label.offset())); + return label; + } + + void leaq_mr(int32_t offset, RegisterID base, RegisterID dst) { + spew("leaq " MEM_ob ", %s", ADDR_ob(offset, base), GPReg64Name(dst)); + m_formatter.oneByteOp64(OP_LEA, offset, base, dst); + } + + [[nodiscard]] JmpSrc leaq_rip(RegisterID dst) { + m_formatter.oneByteRipOp64(OP_LEA, 0, dst); + JmpSrc label(m_formatter.size()); + spew("leaq " MEM_o32r ", %s", ADDR_o32r(label.offset()), + GPRegName(dst)); + return label; + } + + // Flow control: + + void jmp_rip(int ripOffset) { + // rip-relative addressing. + spew("jmp *%d(%%rip)", ripOffset); + m_formatter.oneByteRipOp(OP_GROUP5_Ev, ripOffset, GROUP5_OP_JMPN); + } + + void immediate64(int64_t imm) { + spew(".quad %lld", (long long)imm); + m_formatter.immediate64(imm); + } + + // SSE operations: + + void vcvtsq2sd_rr(RegisterID src1, XMMRegisterID src0, XMMRegisterID dst) { + twoByteOpInt64Simd("vcvtsi2sd", VEX_SD, OP2_CVTSI2SD_VsdEd, src1, src0, + dst); + } + void vcvtsq2ss_rr(RegisterID src1, XMMRegisterID src0, XMMRegisterID dst) { + twoByteOpInt64Simd("vcvtsi2ss", VEX_SS, OP2_CVTSI2SD_VsdEd, src1, src0, + dst); + } + + void vcvtsi2sdq_rr(RegisterID src, XMMRegisterID dst) { + twoByteOpInt64Simd("vcvtsi2sdq", VEX_SD, OP2_CVTSI2SD_VsdEd, src, + invalid_xmm, dst); + } + + void vcvttsd2sq_rr(XMMRegisterID src, RegisterID dst) { + twoByteOpSimdInt64("vcvttsd2si", VEX_SD, OP2_CVTTSD2SI_GdWsd, src, dst); + } + + void vcvttss2sq_rr(XMMRegisterID src, RegisterID dst) { + twoByteOpSimdInt64("vcvttss2si", VEX_SS, OP2_CVTTSD2SI_GdWsd, src, dst); + } + + void vmovq_rr(XMMRegisterID src, RegisterID dst) { + // While this is called "vmovq", it actually uses the vmovd encoding + // with a REX prefix modifying it to be 64-bit. + twoByteOpSimdInt64("vmovq", VEX_PD, OP2_MOVD_EdVd, (XMMRegisterID)dst, + (RegisterID)src); + } + + void vpextrq_irr(unsigned lane, XMMRegisterID src, RegisterID dst) { + MOZ_ASSERT(lane < 2); + threeByteOpImmSimdInt64("vpextrq", VEX_PD, OP3_PEXTRQ_EvVdqIb, ESCAPE_3A, + lane, src, dst); + } + + void vpinsrq_irr(unsigned lane, RegisterID src1, XMMRegisterID src0, + XMMRegisterID dst) { + MOZ_ASSERT(lane < 2); + threeByteOpImmInt64Simd("vpinsrq", VEX_PD, OP3_PINSRQ_VdqEvIb, ESCAPE_3A, + lane, src1, src0, dst); + } + + void vmovq_rr(RegisterID src, XMMRegisterID dst) { + // While this is called "vmovq", it actually uses the vmovd encoding + // with a REX prefix modifying it to be 64-bit. + twoByteOpInt64Simd("vmovq", VEX_PD, OP2_MOVD_VdEd, src, invalid_xmm, dst); + } + + [[nodiscard]] JmpSrc vmovsd_ripr(XMMRegisterID dst) { + return twoByteRipOpSimd("vmovsd", VEX_SD, OP2_MOVSD_VsdWsd, dst); + } + [[nodiscard]] JmpSrc vmovss_ripr(XMMRegisterID dst) { + return twoByteRipOpSimd("vmovss", VEX_SS, OP2_MOVSD_VsdWsd, dst); + } + [[nodiscard]] JmpSrc vmovaps_ripr(XMMRegisterID dst) { + return twoByteRipOpSimd("vmovaps", VEX_PS, OP2_MOVAPS_VsdWsd, dst); + } + [[nodiscard]] JmpSrc vmovdqa_ripr(XMMRegisterID dst) { + return twoByteRipOpSimd("vmovdqa", VEX_PD, OP2_MOVDQ_VdqWdq, dst); + } + + [[nodiscard]] JmpSrc vpaddb_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpaddb", VEX_PD, OP2_PADDB_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpaddw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpaddw", VEX_PD, OP2_PADDW_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpaddd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpaddd", VEX_PD, OP2_PADDD_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpaddq_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpaddq", VEX_PD, OP2_PADDQ_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpsubb_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpsubb", VEX_PD, OP2_PSUBB_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpsubw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpsubw", VEX_PD, OP2_PSUBW_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpsubd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpsubd", VEX_PD, OP2_PSUBD_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpsubq_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpsubq", VEX_PD, OP2_PSUBQ_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpmullw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpmullw", VEX_PD, OP2_PMULLW_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpmulld_ripr(XMMRegisterID src, XMMRegisterID dst) { + return threeByteRipOpSimd("vpmulld", VEX_PD, OP3_PMULLD_VdqWdq, ESCAPE_38, + src, dst); + } + [[nodiscard]] JmpSrc vpaddsb_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpaddsb", VEX_PD, OP2_PADDSB_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpaddusb_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpaddusb", VEX_PD, OP2_PADDUSB_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpaddsw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpaddsw", VEX_PD, OP2_PADDSW_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpaddusw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpaddusw", VEX_PD, OP2_PADDUSW_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpsubsb_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpsubsb", VEX_PD, OP2_PSUBSB_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpsubusb_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpsubusb", VEX_PD, OP2_PSUBUSB_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpsubsw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpsubsw", VEX_PD, OP2_PSUBSW_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpsubusw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpsubusw", VEX_PD, OP2_PSUBUSW_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpminsb_ripr(XMMRegisterID src, XMMRegisterID dst) { + return threeByteRipOpSimd("vpminsb", VEX_PD, OP3_PMINSB_VdqWdq, ESCAPE_38, + src, dst); + } + [[nodiscard]] JmpSrc vpminub_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpminub", VEX_PD, OP2_PMINUB_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpminsw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpminsw", VEX_PD, OP2_PMINSW_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpminuw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return threeByteRipOpSimd("vpminuw", VEX_PD, OP3_PMINUW_VdqWdq, ESCAPE_38, + src, dst); + } + [[nodiscard]] JmpSrc vpminsd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return threeByteRipOpSimd("vpminsd", VEX_PD, OP3_PMINSD_VdqWdq, ESCAPE_38, + src, dst); + } + [[nodiscard]] JmpSrc vpminud_ripr(XMMRegisterID src, XMMRegisterID dst) { + return threeByteRipOpSimd("vpminud", VEX_PD, OP3_PMINUD_VdqWdq, ESCAPE_38, + src, dst); + } + [[nodiscard]] JmpSrc vpmaxsb_ripr(XMMRegisterID src, XMMRegisterID dst) { + return threeByteRipOpSimd("vpmaxsb", VEX_PD, OP3_PMAXSB_VdqWdq, ESCAPE_38, + src, dst); + } + [[nodiscard]] JmpSrc vpmaxub_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpmaxub", VEX_PD, OP2_PMAXUB_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpmaxsw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpmaxsw", VEX_PD, OP2_PMAXSW_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpmaxuw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return threeByteRipOpSimd("vpmaxuw", VEX_PD, OP3_PMAXUW_VdqWdq, ESCAPE_38, + src, dst); + } + [[nodiscard]] JmpSrc vpmaxsd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return threeByteRipOpSimd("vpmaxsd", VEX_PD, OP3_PMAXSD_VdqWdq, ESCAPE_38, + src, dst); + } + [[nodiscard]] JmpSrc vpmaxud_ripr(XMMRegisterID src, XMMRegisterID dst) { + return threeByteRipOpSimd("vpmaxud", VEX_PD, OP3_PMAXUD_VdqWdq, ESCAPE_38, + src, dst); + } + [[nodiscard]] JmpSrc vpand_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpand", VEX_PD, OP2_PANDDQ_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpxor_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpxor", VEX_PD, OP2_PXORDQ_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpor_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpor", VEX_PD, OP2_PORDQ_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vaddps_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vaddps", VEX_PS, OP2_ADDPS_VpsWps, src, dst); + } + [[nodiscard]] JmpSrc vaddpd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vaddpd", VEX_PD, OP2_ADDPD_VpdWpd, src, dst); + } + [[nodiscard]] JmpSrc vsubps_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vsubps", VEX_PS, OP2_SUBPS_VpsWps, src, dst); + } + [[nodiscard]] JmpSrc vsubpd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vsubpd", VEX_PD, OP2_SUBPD_VpdWpd, src, dst); + } + [[nodiscard]] JmpSrc vdivps_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vdivps", VEX_PS, OP2_DIVPS_VpsWps, src, dst); + } + [[nodiscard]] JmpSrc vdivpd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vdivpd", VEX_PD, OP2_DIVPD_VpdWpd, src, dst); + } + [[nodiscard]] JmpSrc vmulps_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vmulps", VEX_PS, OP2_MULPS_VpsWps, src, dst); + } + [[nodiscard]] JmpSrc vmulpd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vmulpd", VEX_PD, OP2_MULPD_VpdWpd, src, dst); + } + [[nodiscard]] JmpSrc vandpd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vandpd", VEX_PD, OP2_ANDPD_VpdWpd, src, dst); + } + [[nodiscard]] JmpSrc vminpd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vminpd", VEX_PD, OP2_MINPD_VpdWpd, src, dst); + } + [[nodiscard]] JmpSrc vpacksswb_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpacksswb", VEX_PD, OP2_PACKSSWB_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpackuswb_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpackuswb", VEX_PD, OP2_PACKUSWB_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpackssdw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpackssdw", VEX_PD, OP2_PACKSSDW_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpackusdw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return threeByteRipOpSimd("vpackusdw", VEX_PD, OP3_PACKUSDW_VdqWdq, + ESCAPE_38, src, dst); + } + [[nodiscard]] JmpSrc vpunpckldq_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpunpckldq", VEX_PD, OP2_PUNPCKLDQ_VdqWdq, src, + dst); + } + [[nodiscard]] JmpSrc vunpcklps_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vunpcklps", VEX_PS, OP2_UNPCKLPS_VsdWsd, src, dst); + } + [[nodiscard]] JmpSrc vptest_ripr(XMMRegisterID lhs) { + return threeByteRipOpSimd("vptest", VEX_PD, OP3_PTEST_VdVd, ESCAPE_38, lhs); + } + [[nodiscard]] JmpSrc vpshufb_ripr(XMMRegisterID src, XMMRegisterID dst) { + return threeByteRipOpSimd("vpshufb", VEX_PD, OP3_PSHUFB_VdqWdq, ESCAPE_38, + src, dst); + } + [[nodiscard]] JmpSrc vpmaddwd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpmaddwd", VEX_PD, OP2_PMADDWD_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpcmpeqb_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpcmpeqb", VEX_PD, OP2_PCMPEQB_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpcmpgtb_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpcmpgtb", VEX_PD, OP2_PCMPGTB_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpcmpeqw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpcmpeqw", VEX_PD, OP2_PCMPEQW_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpcmpgtw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpcmpgtw", VEX_PD, OP2_PCMPGTW_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpcmpeqd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpcmpeqd", VEX_PD, OP2_PCMPEQD_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vpcmpgtd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpcmpgtd", VEX_PD, OP2_PCMPGTD_VdqWdq, src, dst); + } + [[nodiscard]] JmpSrc vcmpeqps_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpImmSimd("vcmpps", VEX_PS, OP2_CMPPS_VpsWps, + X86Encoding::ConditionCmp_EQ, src, dst); + } + [[nodiscard]] JmpSrc vcmpneqps_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpImmSimd("vcmpps", VEX_PS, OP2_CMPPS_VpsWps, + X86Encoding::ConditionCmp_NEQ, src, dst); + } + [[nodiscard]] JmpSrc vcmpltps_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpImmSimd("vcmpps", VEX_PS, OP2_CMPPS_VpsWps, + X86Encoding::ConditionCmp_LT, src, dst); + } + [[nodiscard]] JmpSrc vcmpleps_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpImmSimd("vcmpps", VEX_PS, OP2_CMPPS_VpsWps, + X86Encoding::ConditionCmp_LE, src, dst); + } + [[nodiscard]] JmpSrc vcmpgeps_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpImmSimd("vcmpps", VEX_PS, OP2_CMPPS_VpsWps, + X86Encoding::ConditionCmp_GE, src, dst); + } + [[nodiscard]] JmpSrc vcmpeqpd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpImmSimd("vcmppd", VEX_PD, OP2_CMPPD_VpdWpd, + X86Encoding::ConditionCmp_EQ, src, dst); + } + [[nodiscard]] JmpSrc vcmpneqpd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpImmSimd("vcmppd", VEX_PD, OP2_CMPPD_VpdWpd, + X86Encoding::ConditionCmp_NEQ, src, dst); + } + [[nodiscard]] JmpSrc vcmpltpd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpImmSimd("vcmppd", VEX_PD, OP2_CMPPD_VpdWpd, + X86Encoding::ConditionCmp_LT, src, dst); + } + [[nodiscard]] JmpSrc vcmplepd_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpImmSimd("vcmppd", VEX_PD, OP2_CMPPD_VpdWpd, + X86Encoding::ConditionCmp_LE, src, dst); + } + [[nodiscard]] JmpSrc vpmaddubsw_ripr(XMMRegisterID src, XMMRegisterID dst) { + return threeByteRipOpSimd("vpmaddubsw", VEX_PD, OP3_PMADDUBSW_VdqWdq, + ESCAPE_38, src, dst); + } + [[nodiscard]] JmpSrc vpmuludq_ripr(XMMRegisterID src, XMMRegisterID dst) { + return twoByteRipOpSimd("vpmuludq", VEX_PD, OP2_PMULUDQ_VdqWdq, src, dst); + } + + // BMI instructions: + + void sarxq_rrr(RegisterID src, RegisterID shift, RegisterID dst) { + spew("sarxq %s, %s, %s", GPReg64Name(src), GPReg64Name(shift), + GPReg64Name(dst)); + + RegisterID rm = src; + XMMRegisterID src0 = static_cast<XMMRegisterID>(shift); + int reg = dst; + m_formatter.threeByteOpVex64(VEX_SS /* = F3 */, OP3_SARX_GyEyBy, ESCAPE_38, + rm, src0, reg); + } + + void shlxq_rrr(RegisterID src, RegisterID shift, RegisterID dst) { + spew("shlxq %s, %s, %s", GPReg64Name(src), GPReg64Name(shift), + GPReg64Name(dst)); + + RegisterID rm = src; + XMMRegisterID src0 = static_cast<XMMRegisterID>(shift); + int reg = dst; + m_formatter.threeByteOpVex64(VEX_PD /* = 66 */, OP3_SHLX_GyEyBy, ESCAPE_38, + rm, src0, reg); + } + + void shrxq_rrr(RegisterID src, RegisterID shift, RegisterID dst) { + spew("shrxq %s, %s, %s", GPReg64Name(src), GPReg64Name(shift), + GPReg64Name(dst)); + + RegisterID rm = src; + XMMRegisterID src0 = static_cast<XMMRegisterID>(shift); + int reg = dst; + m_formatter.threeByteOpVex64(VEX_SD /* = F2 */, OP3_SHRX_GyEyBy, ESCAPE_38, + rm, src0, reg); + } + + private: + [[nodiscard]] JmpSrc twoByteRipOpSimd(const char* name, VexOperandType ty, + TwoByteOpcodeID opcode, + XMMRegisterID reg) { + MOZ_ASSERT(!IsXMMReversedOperands(opcode)); + m_formatter.legacySSEPrefix(ty); + m_formatter.twoByteRipOp(opcode, 0, reg); + JmpSrc label(m_formatter.size()); + spew("%-11s " MEM_o32r ", %s", legacySSEOpName(name), + ADDR_o32r(label.offset()), XMMRegName(reg)); + return label; + } + + [[nodiscard]] JmpSrc twoByteRipOpSimd(const char* name, VexOperandType ty, + TwoByteOpcodeID opcode, + XMMRegisterID src0, XMMRegisterID dst) { + MOZ_ASSERT(src0 != invalid_xmm && !IsXMMReversedOperands(opcode)); + if (useLegacySSEEncoding(src0, dst)) { + m_formatter.legacySSEPrefix(ty); + m_formatter.twoByteRipOp(opcode, 0, dst); + JmpSrc label(m_formatter.size()); + spew("%-11s" MEM_o32r ", %s", legacySSEOpName(name), + ADDR_o32r(label.offset()), XMMRegName(dst)); + return label; + } + + m_formatter.twoByteRipOpVex(ty, opcode, 0, src0, dst); + JmpSrc label(m_formatter.size()); + spew("%-11s, " MEM_o32r ", %s, %s", name, ADDR_o32r(label.offset()), + XMMRegName(src0), XMMRegName(dst)); + return label; + } + + [[nodiscard]] JmpSrc twoByteRipOpImmSimd(const char* name, VexOperandType ty, + TwoByteOpcodeID opcode, uint32_t imm, + XMMRegisterID src0, + XMMRegisterID dst) { + MOZ_ASSERT(src0 != invalid_xmm && !IsXMMReversedOperands(opcode)); + if (useLegacySSEEncoding(src0, dst)) { + m_formatter.legacySSEPrefix(ty); + m_formatter.twoByteRipOp(opcode, 0, dst); + m_formatter.immediate8u(imm); + JmpSrc label(m_formatter.size(), + /* bytes trailing the patch field = */ 1); + spew("%-11s$0x%x, " MEM_o32r ", %s", legacySSEOpName(name), imm, + ADDR_o32r(label.offset()), XMMRegName(dst)); + return label; + } + + m_formatter.twoByteRipOpVex(ty, opcode, 0, src0, dst); + m_formatter.immediate8u(imm); + JmpSrc label(m_formatter.size(), + /* bytes trailing the patch field = */ 1); + spew("%-11s$0x%x, " MEM_o32r ", %s, %s", name, imm, + ADDR_o32r(label.offset()), XMMRegName(src0), XMMRegName(dst)); + return label; + } + + void twoByteOpInt64Simd(const char* name, VexOperandType ty, + TwoByteOpcodeID opcode, RegisterID rm, + XMMRegisterID src0, XMMRegisterID dst) { + if (useLegacySSEEncoding(src0, dst)) { + if (IsXMMReversedOperands(opcode)) { + spew("%-11s%s, %s", legacySSEOpName(name), XMMRegName(dst), + GPRegName(rm)); + } else { + spew("%-11s%s, %s", legacySSEOpName(name), GPRegName(rm), + XMMRegName(dst)); + } + m_formatter.legacySSEPrefix(ty); + m_formatter.twoByteOp64(opcode, rm, dst); + return; + } + + if (src0 == invalid_xmm) { + if (IsXMMReversedOperands(opcode)) { + spew("%-11s%s, %s", name, XMMRegName(dst), GPRegName(rm)); + } else { + spew("%-11s%s, %s", name, GPRegName(rm), XMMRegName(dst)); + } + } else { + spew("%-11s%s, %s, %s", name, GPRegName(rm), XMMRegName(src0), + XMMRegName(dst)); + } + m_formatter.twoByteOpVex64(ty, opcode, rm, src0, dst); + } + + void twoByteOpSimdInt64(const char* name, VexOperandType ty, + TwoByteOpcodeID opcode, XMMRegisterID rm, + RegisterID dst) { + if (useLegacySSEEncodingAlways()) { + if (IsXMMReversedOperands(opcode)) { + spew("%-11s%s, %s", legacySSEOpName(name), GPRegName(dst), + XMMRegName(rm)); + } else if (opcode == OP2_MOVD_EdVd) { + spew("%-11s%s, %s", legacySSEOpName(name), + XMMRegName((XMMRegisterID)dst), GPRegName((RegisterID)rm)); + } else { + spew("%-11s%s, %s", legacySSEOpName(name), XMMRegName(rm), + GPRegName(dst)); + } + m_formatter.legacySSEPrefix(ty); + m_formatter.twoByteOp64(opcode, (RegisterID)rm, dst); + return; + } + + if (IsXMMReversedOperands(opcode)) { + spew("%-11s%s, %s", name, GPRegName(dst), XMMRegName(rm)); + } else if (opcode == OP2_MOVD_EdVd) { + spew("%-11s%s, %s", name, XMMRegName((XMMRegisterID)dst), + GPRegName((RegisterID)rm)); + } else { + spew("%-11s%s, %s", name, XMMRegName(rm), GPRegName(dst)); + } + m_formatter.twoByteOpVex64(ty, opcode, (RegisterID)rm, invalid_xmm, + (XMMRegisterID)dst); + } + + [[nodiscard]] JmpSrc threeByteRipOpSimd(const char* name, VexOperandType ty, + ThreeByteOpcodeID opcode, + ThreeByteEscape escape, + XMMRegisterID dst) { + m_formatter.legacySSEPrefix(ty); + m_formatter.threeByteRipOp(opcode, escape, 0, dst); + JmpSrc label(m_formatter.size()); + spew("%-11s" MEM_o32r ", %s", legacySSEOpName(name), + ADDR_o32r(label.offset()), XMMRegName(dst)); + return label; + } + + [[nodiscard]] JmpSrc threeByteRipOpSimd(const char* name, VexOperandType ty, + ThreeByteOpcodeID opcode, + ThreeByteEscape escape, + XMMRegisterID src0, + XMMRegisterID dst) { + MOZ_ASSERT(src0 != invalid_xmm); + if (useLegacySSEEncoding(src0, dst)) { + m_formatter.legacySSEPrefix(ty); + m_formatter.threeByteRipOp(opcode, escape, 0, dst); + JmpSrc label(m_formatter.size()); + spew("%-11s" MEM_o32r ", %s", legacySSEOpName(name), + ADDR_o32r(label.offset()), XMMRegName(dst)); + return label; + } + + m_formatter.threeByteRipOpVex(ty, opcode, escape, 0, src0, dst); + JmpSrc label(m_formatter.size()); + spew("%-11s" MEM_o32r ", %s, %s", name, ADDR_o32r(label.offset()), + XMMRegName(src0), XMMRegName(dst)); + return label; + } + + void threeByteOpImmSimdInt64(const char* name, VexOperandType ty, + ThreeByteOpcodeID opcode, ThreeByteEscape escape, + uint32_t imm, XMMRegisterID src, + RegisterID dst) { + spew("%-11s$0x%x, %s, %s", legacySSEOpName(name), imm, GPReg64Name(dst), + XMMRegName(src)); + m_formatter.legacySSEPrefix(ty); + m_formatter.threeByteOp64(opcode, escape, dst, (RegisterID)src); + m_formatter.immediate8u(imm); + } + + void threeByteOpImmInt64Simd(const char* name, VexOperandType ty, + ThreeByteOpcodeID opcode, ThreeByteEscape escape, + uint32_t imm, RegisterID src1, + XMMRegisterID src0, XMMRegisterID dst) { + if (useLegacySSEEncoding(src0, dst)) { + spew("%-11s$0x%x, %s, %s", legacySSEOpName(name), imm, GPReg64Name(src1), + XMMRegName(dst)); + m_formatter.legacySSEPrefix(ty); + m_formatter.threeByteOp64(opcode, escape, src1, (RegisterID)dst); + m_formatter.immediate8u(imm); + return; + } + + MOZ_ASSERT(src0 != invalid_xmm); + spew("%-11s$0x%x, %s, %s, %s", name, imm, GPReg64Name(src1), + XMMRegName(src0), XMMRegName(dst)); + m_formatter.threeByteOpVex64(ty, opcode, escape, src1, src0, + (RegisterID)dst); + m_formatter.immediate8u(imm); + } +}; + +using BaseAssemblerSpecific = BaseAssemblerX64; + +} // namespace X86Encoding + +} // namespace jit +} // namespace js + +#endif /* jit_x64_BaseAssembler_x64_h */ diff --git a/js/src/jit/x64/CodeGenerator-x64.cpp b/js/src/jit/x64/CodeGenerator-x64.cpp new file mode 100644 index 0000000000..9bd7e9b253 --- /dev/null +++ b/js/src/jit/x64/CodeGenerator-x64.cpp @@ -0,0 +1,984 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#include "jit/x64/CodeGenerator-x64.h" + +#include "jit/CodeGenerator.h" +#include "jit/MIR.h" +#include "js/ScalarType.h" // js::Scalar::Type + +#include "jit/MacroAssembler-inl.h" +#include "jit/shared/CodeGenerator-shared-inl.h" + +using namespace js; +using namespace js::jit; + +using mozilla::DebugOnly; + +CodeGeneratorX64::CodeGeneratorX64(MIRGenerator* gen, LIRGraph* graph, + MacroAssembler* masm) + : CodeGeneratorX86Shared(gen, graph, masm) {} + +ValueOperand CodeGeneratorX64::ToValue(LInstruction* ins, size_t pos) { + return ValueOperand(ToRegister(ins->getOperand(pos))); +} + +ValueOperand CodeGeneratorX64::ToTempValue(LInstruction* ins, size_t pos) { + return ValueOperand(ToRegister(ins->getTemp(pos))); +} + +Operand CodeGeneratorX64::ToOperand64(const LInt64Allocation& a64) { + const LAllocation& a = a64.value(); + MOZ_ASSERT(!a.isFloatReg()); + if (a.isGeneralReg()) { + return Operand(a.toGeneralReg()->reg()); + } + return Operand(ToAddress(a)); +} + +void CodeGenerator::visitValue(LValue* value) { + ValueOperand result = ToOutValue(value); + masm.moveValue(value->value(), result); +} + +void CodeGenerator::visitBox(LBox* box) { + const LAllocation* in = box->getOperand(0); + ValueOperand result = ToOutValue(box); + + masm.moveValue(TypedOrValueRegister(box->type(), ToAnyRegister(in)), result); + + if (JitOptions.spectreValueMasking && IsFloatingPointType(box->type())) { + ScratchRegisterScope scratch(masm); + masm.movePtr(ImmWord(JSVAL_SHIFTED_TAG_MAX_DOUBLE), scratch); + masm.cmpPtrMovePtr(Assembler::Below, scratch, result.valueReg(), scratch, + result.valueReg()); + } +} + +void CodeGenerator::visitUnbox(LUnbox* unbox) { + MUnbox* mir = unbox->mir(); + + Register result = ToRegister(unbox->output()); + + if (mir->fallible()) { + const ValueOperand value = ToValue(unbox, LUnbox::Input); + Label bail; + switch (mir->type()) { + case MIRType::Int32: + masm.fallibleUnboxInt32(value, result, &bail); + break; + case MIRType::Boolean: + masm.fallibleUnboxBoolean(value, result, &bail); + break; + case MIRType::Object: + masm.fallibleUnboxObject(value, result, &bail); + break; + case MIRType::String: + masm.fallibleUnboxString(value, result, &bail); + break; + case MIRType::Symbol: + masm.fallibleUnboxSymbol(value, result, &bail); + break; + case MIRType::BigInt: + masm.fallibleUnboxBigInt(value, result, &bail); + break; + default: + MOZ_CRASH("Given MIRType cannot be unboxed."); + } + bailoutFrom(&bail, unbox->snapshot()); + return; + } + + // Infallible unbox. + + Operand input = ToOperand(unbox->getOperand(LUnbox::Input)); + +#ifdef DEBUG + // Assert the types match. + JSValueTag tag = MIRTypeToTag(mir->type()); + Label ok; + masm.splitTag(input, ScratchReg); + masm.branch32(Assembler::Equal, ScratchReg, Imm32(tag), &ok); + masm.assumeUnreachable("Infallible unbox type mismatch"); + masm.bind(&ok); +#endif + + switch (mir->type()) { + case MIRType::Int32: + masm.unboxInt32(input, result); + break; + case MIRType::Boolean: + masm.unboxBoolean(input, result); + break; + case MIRType::Object: + masm.unboxObject(input, result); + break; + case MIRType::String: + masm.unboxString(input, result); + break; + case MIRType::Symbol: + masm.unboxSymbol(input, result); + break; + case MIRType::BigInt: + masm.unboxBigInt(input, result); + break; + default: + MOZ_CRASH("Given MIRType cannot be unboxed."); + } +} + +void CodeGenerator::visitCompareI64(LCompareI64* lir) { + MCompare* mir = lir->mir(); + MOZ_ASSERT(mir->compareType() == MCompare::Compare_Int64 || + mir->compareType() == MCompare::Compare_UInt64); + + const LInt64Allocation lhs = lir->getInt64Operand(LCompareI64::Lhs); + const LInt64Allocation rhs = lir->getInt64Operand(LCompareI64::Rhs); + Register lhsReg = ToRegister64(lhs).reg; + Register output = ToRegister(lir->output()); + + if (IsConstant(rhs)) { + masm.cmpPtr(lhsReg, ImmWord(ToInt64(rhs))); + } else { + masm.cmpPtr(lhsReg, ToOperand64(rhs)); + } + + bool isSigned = mir->compareType() == MCompare::Compare_Int64; + masm.emitSet(JSOpToCondition(lir->jsop(), isSigned), output); +} + +void CodeGenerator::visitCompareI64AndBranch(LCompareI64AndBranch* lir) { + MCompare* mir = lir->cmpMir(); + MOZ_ASSERT(mir->compareType() == MCompare::Compare_Int64 || + mir->compareType() == MCompare::Compare_UInt64); + + LInt64Allocation lhs = lir->getInt64Operand(LCompareI64::Lhs); + LInt64Allocation rhs = lir->getInt64Operand(LCompareI64::Rhs); + Register lhsReg = ToRegister64(lhs).reg; + + if (IsConstant(rhs)) { + masm.cmpPtr(lhsReg, ImmWord(ToInt64(rhs))); + } else { + masm.cmpPtr(lhsReg, ToOperand64(rhs)); + } + + bool isSigned = mir->compareType() == MCompare::Compare_Int64; + emitBranch(JSOpToCondition(lir->jsop(), isSigned), lir->ifTrue(), + lir->ifFalse()); +} + +void CodeGenerator::visitBitAndAndBranch(LBitAndAndBranch* baab) { + Register regL = ToRegister(baab->left()); + if (baab->is64()) { + if (baab->right()->isConstant()) { + masm.test64(regL, Imm64(ToInt64(baab->right()))); + } else { + masm.test64(regL, ToRegister(baab->right())); + } + } else { + if (baab->right()->isConstant()) { + masm.test32(regL, Imm32(ToInt32(baab->right()))); + } else { + masm.test32(regL, ToRegister(baab->right())); + } + } + emitBranch(baab->cond(), baab->ifTrue(), baab->ifFalse()); +} + +void CodeGenerator::visitDivOrModI64(LDivOrModI64* lir) { + Register lhs = ToRegister(lir->lhs()); + Register rhs = ToRegister(lir->rhs()); + Register output = ToRegister(lir->output()); + + MOZ_ASSERT_IF(lhs != rhs, rhs != rax); + MOZ_ASSERT(rhs != rdx); + MOZ_ASSERT_IF(output == rax, ToRegister(lir->remainder()) == rdx); + MOZ_ASSERT_IF(output == rdx, ToRegister(lir->remainder()) == rax); + + Label done; + + // Put the lhs in rax. + if (lhs != rax) { + masm.mov(lhs, rax); + } + + // Handle divide by zero. + if (lir->canBeDivideByZero()) { + Label nonZero; + masm.branchTestPtr(Assembler::NonZero, rhs, rhs, &nonZero); + masm.wasmTrap(wasm::Trap::IntegerDivideByZero, lir->bytecodeOffset()); + masm.bind(&nonZero); + } + + // Handle an integer overflow exception from INT64_MIN / -1. + if (lir->canBeNegativeOverflow()) { + Label notOverflow; + masm.branchPtr(Assembler::NotEqual, lhs, ImmWord(INT64_MIN), ¬Overflow); + masm.branchPtr(Assembler::NotEqual, rhs, ImmWord(-1), ¬Overflow); + if (lir->mir()->isMod()) { + masm.xorl(output, output); + } else { + masm.wasmTrap(wasm::Trap::IntegerOverflow, lir->bytecodeOffset()); + } + masm.jump(&done); + masm.bind(¬Overflow); + } + + // Sign extend the lhs into rdx to make rdx:rax. + masm.cqo(); + masm.idivq(rhs); + + masm.bind(&done); +} + +void CodeGenerator::visitUDivOrModI64(LUDivOrModI64* lir) { + Register lhs = ToRegister(lir->lhs()); + Register rhs = ToRegister(lir->rhs()); + + DebugOnly<Register> output = ToRegister(lir->output()); + MOZ_ASSERT_IF(lhs != rhs, rhs != rax); + MOZ_ASSERT(rhs != rdx); + MOZ_ASSERT_IF(output.value == rax, ToRegister(lir->remainder()) == rdx); + MOZ_ASSERT_IF(output.value == rdx, ToRegister(lir->remainder()) == rax); + + // Put the lhs in rax. + if (lhs != rax) { + masm.mov(lhs, rax); + } + + Label done; + + // Prevent divide by zero. + if (lir->canBeDivideByZero()) { + Label nonZero; + masm.branchTestPtr(Assembler::NonZero, rhs, rhs, &nonZero); + masm.wasmTrap(wasm::Trap::IntegerDivideByZero, lir->bytecodeOffset()); + masm.bind(&nonZero); + } + + // Zero extend the lhs into rdx to make (rdx:rax). + masm.xorl(rdx, rdx); + masm.udivq(rhs); + + masm.bind(&done); +} + +void CodeGeneratorX64::emitBigIntDiv(LBigIntDiv* ins, Register dividend, + Register divisor, Register output, + Label* fail) { + // Callers handle division by zero and integer overflow. + + MOZ_ASSERT(dividend == rax); + MOZ_ASSERT(output == rdx); + + // Sign extend the lhs into rdx to make rdx:rax. + masm.cqo(); + + masm.idivq(divisor); + + // Create and return the result. + masm.newGCBigInt(output, divisor, initialBigIntHeap(), fail); + masm.initializeBigInt(output, dividend); +} + +void CodeGeneratorX64::emitBigIntMod(LBigIntMod* ins, Register dividend, + Register divisor, Register output, + Label* fail) { + // Callers handle division by zero and integer overflow. + + MOZ_ASSERT(dividend == rax); + MOZ_ASSERT(output == rdx); + + // Sign extend the lhs into rdx to make rdx:rax. + masm.cqo(); + + masm.idivq(divisor); + + // Move the remainder from rdx. + masm.movq(output, dividend); + + // Create and return the result. + masm.newGCBigInt(output, divisor, initialBigIntHeap(), fail); + masm.initializeBigInt(output, dividend); +} + +void CodeGenerator::visitAtomicLoad64(LAtomicLoad64* lir) { + Register elements = ToRegister(lir->elements()); + Register temp = ToRegister(lir->temp()); + Register64 temp64 = ToRegister64(lir->temp64()); + Register out = ToRegister(lir->output()); + + const MLoadUnboxedScalar* mir = lir->mir(); + + Scalar::Type storageType = mir->storageType(); + + // NOTE: the generated code must match the assembly code in gen_load in + // GenerateAtomicOperations.py + auto sync = Synchronization::Load(); + + masm.memoryBarrierBefore(sync); + if (lir->index()->isConstant()) { + Address source = + ToAddress(elements, lir->index(), storageType, mir->offsetAdjustment()); + masm.load64(source, temp64); + } else { + BaseIndex source(elements, ToRegister(lir->index()), + ScaleFromScalarType(storageType), mir->offsetAdjustment()); + masm.load64(source, temp64); + } + masm.memoryBarrierAfter(sync); + + emitCreateBigInt(lir, storageType, temp64, out, temp); +} + +void CodeGenerator::visitAtomicStore64(LAtomicStore64* lir) { + Register elements = ToRegister(lir->elements()); + Register value = ToRegister(lir->value()); + Register64 temp1 = ToRegister64(lir->temp1()); + + Scalar::Type writeType = lir->mir()->writeType(); + + masm.loadBigInt64(value, temp1); + + // NOTE: the generated code must match the assembly code in gen_store in + // GenerateAtomicOperations.py + auto sync = Synchronization::Store(); + + masm.memoryBarrierBefore(sync); + if (lir->index()->isConstant()) { + Address dest = ToAddress(elements, lir->index(), writeType); + masm.store64(temp1, dest); + } else { + BaseIndex dest(elements, ToRegister(lir->index()), + ScaleFromScalarType(writeType)); + masm.store64(temp1, dest); + } + masm.memoryBarrierAfter(sync); +} + +void CodeGenerator::visitCompareExchangeTypedArrayElement64( + LCompareExchangeTypedArrayElement64* lir) { + Register elements = ToRegister(lir->elements()); + Register oldval = ToRegister(lir->oldval()); + Register newval = ToRegister(lir->newval()); + Register64 temp1 = ToRegister64(lir->temp1()); + Register64 temp2 = ToRegister64(lir->temp2()); + Register out = ToRegister(lir->output()); + + MOZ_ASSERT(temp1.reg == rax); + + Scalar::Type arrayType = lir->mir()->arrayType(); + + masm.loadBigInt64(oldval, temp1); + masm.loadBigInt64(newval, temp2); + + if (lir->index()->isConstant()) { + Address dest = ToAddress(elements, lir->index(), arrayType); + masm.compareExchange64(Synchronization::Full(), dest, temp1, temp2, temp1); + } else { + BaseIndex dest(elements, ToRegister(lir->index()), + ScaleFromScalarType(arrayType)); + masm.compareExchange64(Synchronization::Full(), dest, temp1, temp2, temp1); + } + + emitCreateBigInt(lir, arrayType, temp1, out, temp2.reg); +} + +void CodeGenerator::visitAtomicExchangeTypedArrayElement64( + LAtomicExchangeTypedArrayElement64* lir) { + Register elements = ToRegister(lir->elements()); + Register value = ToRegister(lir->value()); + Register64 temp1 = ToRegister64(lir->temp1()); + Register temp2 = ToRegister(lir->temp2()); + Register out = ToRegister(lir->output()); + + Scalar::Type arrayType = lir->mir()->arrayType(); + + masm.loadBigInt64(value, temp1); + + if (lir->index()->isConstant()) { + Address dest = ToAddress(elements, lir->index(), arrayType); + masm.atomicExchange64(Synchronization::Full(), dest, temp1, temp1); + } else { + BaseIndex dest(elements, ToRegister(lir->index()), + ScaleFromScalarType(arrayType)); + masm.atomicExchange64(Synchronization::Full(), dest, temp1, temp1); + } + + emitCreateBigInt(lir, arrayType, temp1, out, temp2); +} + +void CodeGenerator::visitAtomicTypedArrayElementBinop64( + LAtomicTypedArrayElementBinop64* lir) { + MOZ_ASSERT(!lir->mir()->isForEffect()); + + Register elements = ToRegister(lir->elements()); + Register value = ToRegister(lir->value()); + Register64 temp1 = ToRegister64(lir->temp1()); + Register64 temp2 = ToRegister64(lir->temp2()); + Register out = ToRegister(lir->output()); + + Scalar::Type arrayType = lir->mir()->arrayType(); + AtomicOp atomicOp = lir->mir()->operation(); + + masm.loadBigInt64(value, temp1); + + Register64 fetchTemp = Register64(out); + Register64 fetchOut = temp2; + Register createTemp = temp1.reg; + + // Add and Sub don't need |fetchTemp| and can save a `mov` when the value and + // output register are equal to each other. + if (atomicOp == AtomicFetchAddOp || atomicOp == AtomicFetchSubOp) { + fetchTemp = Register64::Invalid(); + fetchOut = temp1; + createTemp = temp2.reg; + } else { + MOZ_ASSERT(temp2.reg == rax); + } + + if (lir->index()->isConstant()) { + Address dest = ToAddress(elements, lir->index(), arrayType); + masm.atomicFetchOp64(Synchronization::Full(), atomicOp, temp1, dest, + fetchTemp, fetchOut); + } else { + BaseIndex dest(elements, ToRegister(lir->index()), + ScaleFromScalarType(arrayType)); + masm.atomicFetchOp64(Synchronization::Full(), atomicOp, temp1, dest, + fetchTemp, fetchOut); + } + + emitCreateBigInt(lir, arrayType, fetchOut, out, createTemp); +} + +void CodeGenerator::visitAtomicTypedArrayElementBinopForEffect64( + LAtomicTypedArrayElementBinopForEffect64* lir) { + MOZ_ASSERT(lir->mir()->isForEffect()); + + Register elements = ToRegister(lir->elements()); + Register value = ToRegister(lir->value()); + Register64 temp1 = ToRegister64(lir->temp1()); + + Scalar::Type arrayType = lir->mir()->arrayType(); + AtomicOp atomicOp = lir->mir()->operation(); + + masm.loadBigInt64(value, temp1); + + if (lir->index()->isConstant()) { + Address dest = ToAddress(elements, lir->index(), arrayType); + masm.atomicEffectOp64(Synchronization::Full(), atomicOp, temp1, dest); + } else { + BaseIndex dest(elements, ToRegister(lir->index()), + ScaleFromScalarType(arrayType)); + masm.atomicEffectOp64(Synchronization::Full(), atomicOp, temp1, dest); + } +} + +void CodeGenerator::visitWasmSelectI64(LWasmSelectI64* lir) { + MOZ_ASSERT(lir->mir()->type() == MIRType::Int64); + + Register cond = ToRegister(lir->condExpr()); + + Operand falseExpr = ToOperandOrRegister64(lir->falseExpr()); + + Register64 out = ToOutRegister64(lir); + MOZ_ASSERT(ToRegister64(lir->trueExpr()) == out, + "true expr is reused for input"); + + masm.test32(cond, cond); + masm.cmovzq(falseExpr, out.reg); +} + +// We expect to handle only the cases: compare is {U,}Int{32,64}, and select +// is {U,}Int{32,64}, independently. Some values may be stack allocated, and +// the "true" input is reused for the output. +void CodeGenerator::visitWasmCompareAndSelect(LWasmCompareAndSelect* ins) { + bool cmpIs32bit = ins->compareType() == MCompare::Compare_Int32 || + ins->compareType() == MCompare::Compare_UInt32; + bool cmpIs64bit = ins->compareType() == MCompare::Compare_Int64 || + ins->compareType() == MCompare::Compare_UInt64; + bool selIs32bit = ins->mir()->type() == MIRType::Int32; + bool selIs64bit = ins->mir()->type() == MIRType::Int64; + + // Throw out unhandled cases + MOZ_RELEASE_ASSERT( + cmpIs32bit != cmpIs64bit && selIs32bit != selIs64bit, + "CodeGenerator::visitWasmCompareAndSelect: unexpected types"); + + using C = Assembler::Condition; + using R = Register; + using A = const Address&; + + // Identify macroassembler methods to generate instructions, based on the + // type of the comparison and the select. This avoids having to duplicate + // the code-generation tree below 4 times. These assignments to + // `cmpMove_CRRRR` et al are unambiguous as a result of the combination of + // the template parameters and the 5 argument types ((C, R, R, R, R) etc). + void (MacroAssembler::*cmpMove_CRRRR)(C, R, R, R, R) = nullptr; + void (MacroAssembler::*cmpMove_CRARR)(C, R, A, R, R) = nullptr; + void (MacroAssembler::*cmpLoad_CRRAR)(C, R, R, A, R) = nullptr; + void (MacroAssembler::*cmpLoad_CRAAR)(C, R, A, A, R) = nullptr; + + if (cmpIs32bit) { + if (selIs32bit) { + cmpMove_CRRRR = &MacroAssemblerX64::cmpMove<32, 32>; + cmpMove_CRARR = &MacroAssemblerX64::cmpMove<32, 32>; + cmpLoad_CRRAR = &MacroAssemblerX64::cmpLoad<32, 32>; + cmpLoad_CRAAR = &MacroAssemblerX64::cmpLoad<32, 32>; + } else { + cmpMove_CRRRR = &MacroAssemblerX64::cmpMove<32, 64>; + cmpMove_CRARR = &MacroAssemblerX64::cmpMove<32, 64>; + cmpLoad_CRRAR = &MacroAssemblerX64::cmpLoad<32, 64>; + cmpLoad_CRAAR = &MacroAssemblerX64::cmpLoad<32, 64>; + } + } else { + if (selIs32bit) { + cmpMove_CRRRR = &MacroAssemblerX64::cmpMove<64, 32>; + cmpMove_CRARR = &MacroAssemblerX64::cmpMove<64, 32>; + cmpLoad_CRRAR = &MacroAssemblerX64::cmpLoad<64, 32>; + cmpLoad_CRAAR = &MacroAssemblerX64::cmpLoad<64, 32>; + } else { + cmpMove_CRRRR = &MacroAssemblerX64::cmpMove<64, 64>; + cmpMove_CRARR = &MacroAssemblerX64::cmpMove<64, 64>; + cmpLoad_CRRAR = &MacroAssemblerX64::cmpLoad<64, 64>; + cmpLoad_CRAAR = &MacroAssemblerX64::cmpLoad<64, 64>; + } + } + + Register trueExprAndDest = ToRegister(ins->output()); + MOZ_ASSERT(ToRegister(ins->ifTrueExpr()) == trueExprAndDest, + "true expr input is reused for output"); + + Assembler::Condition cond = Assembler::InvertCondition( + JSOpToCondition(ins->compareType(), ins->jsop())); + const LAllocation* rhs = ins->rightExpr(); + const LAllocation* falseExpr = ins->ifFalseExpr(); + Register lhs = ToRegister(ins->leftExpr()); + + // We generate one of four cmp+cmov pairings, depending on whether one of + // the cmp args and one of the cmov args is in memory or a register. + if (rhs->isRegister()) { + if (falseExpr->isRegister()) { + (masm.*cmpMove_CRRRR)(cond, lhs, ToRegister(rhs), ToRegister(falseExpr), + trueExprAndDest); + } else { + (masm.*cmpLoad_CRRAR)(cond, lhs, ToRegister(rhs), ToAddress(falseExpr), + trueExprAndDest); + } + } else { + if (falseExpr->isRegister()) { + (masm.*cmpMove_CRARR)(cond, lhs, ToAddress(rhs), ToRegister(falseExpr), + trueExprAndDest); + } else { + (masm.*cmpLoad_CRAAR)(cond, lhs, ToAddress(rhs), ToAddress(falseExpr), + trueExprAndDest); + } + } +} + +void CodeGenerator::visitWasmReinterpretFromI64(LWasmReinterpretFromI64* lir) { + MOZ_ASSERT(lir->mir()->type() == MIRType::Double); + MOZ_ASSERT(lir->mir()->input()->type() == MIRType::Int64); + masm.vmovq(ToRegister(lir->input()), ToFloatRegister(lir->output())); +} + +void CodeGenerator::visitWasmReinterpretToI64(LWasmReinterpretToI64* lir) { + MOZ_ASSERT(lir->mir()->type() == MIRType::Int64); + MOZ_ASSERT(lir->mir()->input()->type() == MIRType::Double); + masm.vmovq(ToFloatRegister(lir->input()), ToRegister(lir->output())); +} + +void CodeGenerator::visitWasmUint32ToDouble(LWasmUint32ToDouble* lir) { + masm.convertUInt32ToDouble(ToRegister(lir->input()), + ToFloatRegister(lir->output())); +} + +void CodeGenerator::visitWasmUint32ToFloat32(LWasmUint32ToFloat32* lir) { + masm.convertUInt32ToFloat32(ToRegister(lir->input()), + ToFloatRegister(lir->output())); +} + +void CodeGeneratorX64::wasmStore(const wasm::MemoryAccessDesc& access, + const LAllocation* value, Operand dstAddr) { + if (value->isConstant()) { + masm.memoryBarrierBefore(access.sync()); + + const MConstant* mir = value->toConstant(); + Imm32 cst = + Imm32(mir->type() == MIRType::Int32 ? mir->toInt32() : mir->toInt64()); + + masm.append(access, masm.size()); + switch (access.type()) { + case Scalar::Int8: + case Scalar::Uint8: + masm.movb(cst, dstAddr); + break; + case Scalar::Int16: + case Scalar::Uint16: + masm.movw(cst, dstAddr); + break; + case Scalar::Int32: + case Scalar::Uint32: + masm.movl(cst, dstAddr); + break; + case Scalar::Int64: + case Scalar::Simd128: + case Scalar::Float32: + case Scalar::Float64: + case Scalar::Uint8Clamped: + case Scalar::BigInt64: + case Scalar::BigUint64: + case Scalar::MaxTypedArrayViewType: + MOZ_CRASH("unexpected array type"); + } + + masm.memoryBarrierAfter(access.sync()); + } else { + masm.wasmStore(access, ToAnyRegister(value), dstAddr); + } +} + +void CodeGenerator::visitWasmHeapBase(LWasmHeapBase* ins) { + MOZ_ASSERT(ins->instance()->isBogus()); + masm.movePtr(HeapReg, ToRegister(ins->output())); +} + +template <typename T> +void CodeGeneratorX64::emitWasmLoad(T* ins) { + const MWasmLoad* mir = ins->mir(); + + uint32_t offset = mir->access().offset(); + MOZ_ASSERT(offset < masm.wasmMaxOffsetGuardLimit()); + + // ptr is a GPR and is either a 32-bit value zero-extended to 64-bit, or a + // true 64-bit value. + const LAllocation* ptr = ins->ptr(); + Operand srcAddr = ptr->isBogus() + ? Operand(HeapReg, offset) + : Operand(HeapReg, ToRegister(ptr), TimesOne, offset); + + if (mir->type() == MIRType::Int64) { + masm.wasmLoadI64(mir->access(), srcAddr, ToOutRegister64(ins)); + } else { + masm.wasmLoad(mir->access(), srcAddr, ToAnyRegister(ins->output())); + } +} + +void CodeGenerator::visitWasmLoad(LWasmLoad* ins) { emitWasmLoad(ins); } + +void CodeGenerator::visitWasmLoadI64(LWasmLoadI64* ins) { emitWasmLoad(ins); } + +template <typename T> +void CodeGeneratorX64::emitWasmStore(T* ins) { + const MWasmStore* mir = ins->mir(); + const wasm::MemoryAccessDesc& access = mir->access(); + + uint32_t offset = access.offset(); + MOZ_ASSERT(offset < masm.wasmMaxOffsetGuardLimit()); + + const LAllocation* value = ins->getOperand(ins->ValueIndex); + const LAllocation* ptr = ins->ptr(); + Operand dstAddr = ptr->isBogus() + ? Operand(HeapReg, offset) + : Operand(HeapReg, ToRegister(ptr), TimesOne, offset); + + wasmStore(access, value, dstAddr); +} + +void CodeGenerator::visitWasmStore(LWasmStore* ins) { emitWasmStore(ins); } + +void CodeGenerator::visitWasmStoreI64(LWasmStoreI64* ins) { + MOZ_CRASH("Unused on this platform"); +} + +void CodeGenerator::visitWasmCompareExchangeHeap( + LWasmCompareExchangeHeap* ins) { + MWasmCompareExchangeHeap* mir = ins->mir(); + + Register ptr = ToRegister(ins->ptr()); + Register oldval = ToRegister(ins->oldValue()); + Register newval = ToRegister(ins->newValue()); + MOZ_ASSERT(ins->addrTemp()->isBogusTemp()); + + Scalar::Type accessType = mir->access().type(); + BaseIndex srcAddr(HeapReg, ptr, TimesOne, mir->access().offset()); + + if (accessType == Scalar::Int64) { + masm.wasmCompareExchange64(mir->access(), srcAddr, Register64(oldval), + Register64(newval), ToOutRegister64(ins)); + } else { + masm.wasmCompareExchange(mir->access(), srcAddr, oldval, newval, + ToRegister(ins->output())); + } +} + +void CodeGenerator::visitWasmAtomicExchangeHeap(LWasmAtomicExchangeHeap* ins) { + MWasmAtomicExchangeHeap* mir = ins->mir(); + + Register ptr = ToRegister(ins->ptr()); + Register value = ToRegister(ins->value()); + MOZ_ASSERT(ins->addrTemp()->isBogusTemp()); + + Scalar::Type accessType = mir->access().type(); + + BaseIndex srcAddr(HeapReg, ptr, TimesOne, mir->access().offset()); + + if (accessType == Scalar::Int64) { + masm.wasmAtomicExchange64(mir->access(), srcAddr, Register64(value), + ToOutRegister64(ins)); + } else { + masm.wasmAtomicExchange(mir->access(), srcAddr, value, + ToRegister(ins->output())); + } +} + +void CodeGenerator::visitWasmAtomicBinopHeap(LWasmAtomicBinopHeap* ins) { + MWasmAtomicBinopHeap* mir = ins->mir(); + MOZ_ASSERT(mir->hasUses()); + + Register ptr = ToRegister(ins->ptr()); + const LAllocation* value = ins->value(); + Register temp = + ins->temp()->isBogusTemp() ? InvalidReg : ToRegister(ins->temp()); + Register output = ToRegister(ins->output()); + MOZ_ASSERT(ins->addrTemp()->isBogusTemp()); + + Scalar::Type accessType = mir->access().type(); + if (accessType == Scalar::Uint32) { + accessType = Scalar::Int32; + } + + AtomicOp op = mir->operation(); + BaseIndex srcAddr(HeapReg, ptr, TimesOne, mir->access().offset()); + + if (accessType == Scalar::Int64) { + Register64 val = Register64(ToRegister(value)); + Register64 out = Register64(output); + Register64 tmp = Register64(temp); + masm.wasmAtomicFetchOp64(mir->access(), op, val, srcAddr, tmp, out); + } else if (value->isConstant()) { + masm.wasmAtomicFetchOp(mir->access(), op, Imm32(ToInt32(value)), srcAddr, + temp, output); + } else { + masm.wasmAtomicFetchOp(mir->access(), op, ToRegister(value), srcAddr, temp, + output); + } +} + +void CodeGenerator::visitWasmAtomicBinopHeapForEffect( + LWasmAtomicBinopHeapForEffect* ins) { + MWasmAtomicBinopHeap* mir = ins->mir(); + MOZ_ASSERT(!mir->hasUses()); + + Register ptr = ToRegister(ins->ptr()); + const LAllocation* value = ins->value(); + MOZ_ASSERT(ins->addrTemp()->isBogusTemp()); + + Scalar::Type accessType = mir->access().type(); + AtomicOp op = mir->operation(); + + BaseIndex srcAddr(HeapReg, ptr, TimesOne, mir->access().offset()); + + if (accessType == Scalar::Int64) { + Register64 val = Register64(ToRegister(value)); + masm.wasmAtomicEffectOp64(mir->access(), op, val, srcAddr); + } else if (value->isConstant()) { + Imm32 c(0); + if (value->toConstant()->type() == MIRType::Int64) { + c = Imm32(ToInt64(value)); + } else { + c = Imm32(ToInt32(value)); + } + masm.wasmAtomicEffectOp(mir->access(), op, c, srcAddr, InvalidReg); + } else { + masm.wasmAtomicEffectOp(mir->access(), op, ToRegister(value), srcAddr, + InvalidReg); + } +} + +void CodeGenerator::visitTruncateDToInt32(LTruncateDToInt32* ins) { + FloatRegister input = ToFloatRegister(ins->input()); + Register output = ToRegister(ins->output()); + + // On x64, branchTruncateDouble uses vcvttsd2sq. Unlike the x86 + // implementation, this should handle most doubles and we can just + // call a stub if it fails. + emitTruncateDouble(input, output, ins->mir()); +} + +void CodeGenerator::visitWasmBuiltinTruncateDToInt32( + LWasmBuiltinTruncateDToInt32* lir) { + FloatRegister input = ToFloatRegister(lir->getOperand(0)); + Register output = ToRegister(lir->getDef(0)); + + emitTruncateDouble(input, output, lir->mir()); +} + +void CodeGenerator::visitWasmBuiltinTruncateFToInt32( + LWasmBuiltinTruncateFToInt32* lir) { + FloatRegister input = ToFloatRegister(lir->getOperand(0)); + Register output = ToRegister(lir->getDef(0)); + + emitTruncateFloat32(input, output, lir->mir()); +} + +void CodeGenerator::visitTruncateFToInt32(LTruncateFToInt32* ins) { + FloatRegister input = ToFloatRegister(ins->input()); + Register output = ToRegister(ins->output()); + + // On x64, branchTruncateFloat32 uses vcvttss2sq. Unlike the x86 + // implementation, this should handle most floats and we can just + // call a stub if it fails. + emitTruncateFloat32(input, output, ins->mir()); +} + +void CodeGenerator::visitWrapInt64ToInt32(LWrapInt64ToInt32* lir) { + const LAllocation* input = lir->getOperand(0); + Register output = ToRegister(lir->output()); + + if (lir->mir()->bottomHalf()) { + masm.movl(ToOperand(input), output); + } else { + MOZ_CRASH("Not implemented."); + } +} + +void CodeGenerator::visitExtendInt32ToInt64(LExtendInt32ToInt64* lir) { + const LAllocation* input = lir->getOperand(0); + Register output = ToRegister(lir->output()); + + if (lir->mir()->isUnsigned()) { + masm.movl(ToOperand(input), output); + } else { + masm.movslq(ToOperand(input), output); + } +} + +void CodeGenerator::visitWasmExtendU32Index(LWasmExtendU32Index* lir) { + // Generates no code on this platform because the input is assumed to have + // canonical form. + Register output = ToRegister(lir->output()); + MOZ_ASSERT(ToRegister(lir->input()) == output); + masm.debugAssertCanonicalInt32(output); +} + +void CodeGenerator::visitWasmWrapU32Index(LWasmWrapU32Index* lir) { + // Generates no code on this platform because the input is assumed to have + // canonical form. + Register output = ToRegister(lir->output()); + MOZ_ASSERT(ToRegister(lir->input()) == output); + masm.debugAssertCanonicalInt32(output); +} + +void CodeGenerator::visitSignExtendInt64(LSignExtendInt64* ins) { + Register64 input = ToRegister64(ins->getInt64Operand(0)); + Register64 output = ToOutRegister64(ins); + switch (ins->mode()) { + case MSignExtendInt64::Byte: + masm.movsbq(Operand(input.reg), output.reg); + break; + case MSignExtendInt64::Half: + masm.movswq(Operand(input.reg), output.reg); + break; + case MSignExtendInt64::Word: + masm.movslq(Operand(input.reg), output.reg); + break; + } +} + +void CodeGenerator::visitWasmTruncateToInt64(LWasmTruncateToInt64* lir) { + FloatRegister input = ToFloatRegister(lir->input()); + Register64 output = ToOutRegister64(lir); + + MWasmTruncateToInt64* mir = lir->mir(); + MIRType inputType = mir->input()->type(); + + MOZ_ASSERT(inputType == MIRType::Double || inputType == MIRType::Float32); + + auto* ool = new (alloc()) OutOfLineWasmTruncateCheck(mir, input, output); + addOutOfLineCode(ool, mir); + + FloatRegister temp = + mir->isUnsigned() ? ToFloatRegister(lir->temp()) : InvalidFloatReg; + + Label* oolEntry = ool->entry(); + Label* oolRejoin = ool->rejoin(); + bool isSaturating = mir->isSaturating(); + if (inputType == MIRType::Double) { + if (mir->isUnsigned()) { + masm.wasmTruncateDoubleToUInt64(input, output, isSaturating, oolEntry, + oolRejoin, temp); + } else { + masm.wasmTruncateDoubleToInt64(input, output, isSaturating, oolEntry, + oolRejoin, temp); + } + } else { + if (mir->isUnsigned()) { + masm.wasmTruncateFloat32ToUInt64(input, output, isSaturating, oolEntry, + oolRejoin, temp); + } else { + masm.wasmTruncateFloat32ToInt64(input, output, isSaturating, oolEntry, + oolRejoin, temp); + } + } +} + +void CodeGenerator::visitInt64ToFloatingPoint(LInt64ToFloatingPoint* lir) { + Register64 input = ToRegister64(lir->getInt64Operand(0)); + FloatRegister output = ToFloatRegister(lir->output()); + + MInt64ToFloatingPoint* mir = lir->mir(); + bool isUnsigned = mir->isUnsigned(); + + MIRType outputType = mir->type(); + MOZ_ASSERT(outputType == MIRType::Double || outputType == MIRType::Float32); + MOZ_ASSERT(isUnsigned == !lir->getTemp(0)->isBogusTemp()); + + if (outputType == MIRType::Double) { + if (isUnsigned) { + masm.convertUInt64ToDouble(input, output, ToRegister(lir->getTemp(0))); + } else { + masm.convertInt64ToDouble(input, output); + } + } else { + if (isUnsigned) { + masm.convertUInt64ToFloat32(input, output, ToRegister(lir->getTemp(0))); + } else { + masm.convertInt64ToFloat32(input, output); + } + } +} + +void CodeGenerator::visitNotI64(LNotI64* lir) { + masm.cmpq(Imm32(0), ToRegister(lir->input())); + masm.emitSet(Assembler::Equal, ToRegister(lir->output())); +} + +void CodeGenerator::visitClzI64(LClzI64* lir) { + Register64 input = ToRegister64(lir->getInt64Operand(0)); + Register64 output = ToOutRegister64(lir); + masm.clz64(input, output.reg); +} + +void CodeGenerator::visitCtzI64(LCtzI64* lir) { + Register64 input = ToRegister64(lir->getInt64Operand(0)); + Register64 output = ToOutRegister64(lir); + masm.ctz64(input, output.reg); +} + +void CodeGenerator::visitBitNotI64(LBitNotI64* ins) { + const LAllocation* input = ins->getOperand(0); + MOZ_ASSERT(!input->isConstant()); + Register inputR = ToRegister(input); + MOZ_ASSERT(inputR == ToRegister(ins->output())); + masm.notq(inputR); +} + +void CodeGenerator::visitTestI64AndBranch(LTestI64AndBranch* lir) { + Register input = ToRegister(lir->input()); + masm.testq(input, input); + emitBranch(Assembler::NonZero, lir->ifTrue(), lir->ifFalse()); +} diff --git a/js/src/jit/x64/CodeGenerator-x64.h b/js/src/jit/x64/CodeGenerator-x64.h new file mode 100644 index 0000000000..c3359d0190 --- /dev/null +++ b/js/src/jit/x64/CodeGenerator-x64.h @@ -0,0 +1,41 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#ifndef jit_x64_CodeGenerator_x64_h +#define jit_x64_CodeGenerator_x64_h + +#include "jit/x86-shared/CodeGenerator-x86-shared.h" + +namespace js { +namespace jit { + +class CodeGeneratorX64 : public CodeGeneratorX86Shared { + protected: + CodeGeneratorX64(MIRGenerator* gen, LIRGraph* graph, MacroAssembler* masm); + + Operand ToOperand64(const LInt64Allocation& a); + ValueOperand ToValue(LInstruction* ins, size_t pos); + ValueOperand ToTempValue(LInstruction* ins, size_t pos); + + void emitBigIntDiv(LBigIntDiv* ins, Register dividend, Register divisor, + Register output, Label* fail); + void emitBigIntMod(LBigIntMod* ins, Register dividend, Register divisor, + Register output, Label* fail); + + void wasmStore(const wasm::MemoryAccessDesc& access, const LAllocation* value, + Operand dstAddr); + template <typename T> + void emitWasmLoad(T* ins); + template <typename T> + void emitWasmStore(T* ins); +}; + +using CodeGeneratorSpecific = CodeGeneratorX64; + +} // namespace jit +} // namespace js + +#endif /* jit_x64_CodeGenerator_x64_h */ diff --git a/js/src/jit/x64/LIR-x64.h b/js/src/jit/x64/LIR-x64.h new file mode 100644 index 0000000000..efaedc4499 --- /dev/null +++ b/js/src/jit/x64/LIR-x64.h @@ -0,0 +1,170 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#ifndef jit_x64_LIR_x64_h +#define jit_x64_LIR_x64_h + +namespace js { +namespace jit { + +// Given an untyped input, guards on whether it's a specific type and returns +// the unboxed payload. +class LUnboxBase : public LInstructionHelper<1, 1, 0> { + public: + LUnboxBase(LNode::Opcode op, const LAllocation& input) + : LInstructionHelper<1, 1, 0>(op) { + setOperand(0, input); + } + + static const size_t Input = 0; + + MUnbox* mir() const { return mir_->toUnbox(); } +}; + +class LUnbox : public LUnboxBase { + public: + LIR_HEADER(Unbox) + + explicit LUnbox(const LAllocation& input) : LUnboxBase(classOpcode, input) {} + + const char* extraName() const { return StringFromMIRType(mir()->type()); } +}; + +class LUnboxFloatingPoint : public LUnboxBase { + MIRType type_; + + public: + LIR_HEADER(UnboxFloatingPoint) + + LUnboxFloatingPoint(const LAllocation& input, MIRType type) + : LUnboxBase(classOpcode, input), type_(type) {} + + MIRType type() const { return type_; } + const char* extraName() const { return StringFromMIRType(type_); } +}; + +// Convert a 32-bit unsigned integer to a double. +class LWasmUint32ToDouble : public LInstructionHelper<1, 1, 0> { + public: + LIR_HEADER(WasmUint32ToDouble) + + explicit LWasmUint32ToDouble(const LAllocation& input) + : LInstructionHelper(classOpcode) { + setOperand(0, input); + } +}; + +// Convert a 32-bit unsigned integer to a float32. +class LWasmUint32ToFloat32 : public LInstructionHelper<1, 1, 0> { + public: + LIR_HEADER(WasmUint32ToFloat32) + + explicit LWasmUint32ToFloat32(const LAllocation& input) + : LInstructionHelper(classOpcode) { + setOperand(0, input); + } +}; + +class LDivOrModI64 : public LBinaryMath<1> { + public: + LIR_HEADER(DivOrModI64) + + LDivOrModI64(const LAllocation& lhs, const LAllocation& rhs, + const LDefinition& temp) + : LBinaryMath(classOpcode) { + setOperand(0, lhs); + setOperand(1, rhs); + setTemp(0, temp); + } + + const LDefinition* remainder() { return getTemp(0); } + + MBinaryArithInstruction* mir() const { + MOZ_ASSERT(mir_->isDiv() || mir_->isMod()); + return static_cast<MBinaryArithInstruction*>(mir_); + } + bool canBeDivideByZero() const { + if (mir_->isMod()) { + return mir_->toMod()->canBeDivideByZero(); + } + return mir_->toDiv()->canBeDivideByZero(); + } + bool canBeNegativeOverflow() const { + if (mir_->isMod()) { + return mir_->toMod()->canBeNegativeDividend(); + } + return mir_->toDiv()->canBeNegativeOverflow(); + } + wasm::BytecodeOffset bytecodeOffset() const { + MOZ_ASSERT(mir_->isDiv() || mir_->isMod()); + if (mir_->isMod()) { + return mir_->toMod()->bytecodeOffset(); + } + return mir_->toDiv()->bytecodeOffset(); + } +}; + +// This class performs a simple x86 'div', yielding either a quotient or +// remainder depending on whether this instruction is defined to output +// rax (quotient) or rdx (remainder). +class LUDivOrModI64 : public LBinaryMath<1> { + public: + LIR_HEADER(UDivOrModI64); + + LUDivOrModI64(const LAllocation& lhs, const LAllocation& rhs, + const LDefinition& temp) + : LBinaryMath(classOpcode) { + setOperand(0, lhs); + setOperand(1, rhs); + setTemp(0, temp); + } + + const LDefinition* remainder() { return getTemp(0); } + + const char* extraName() const { + return mir()->isTruncated() ? "Truncated" : nullptr; + } + + MBinaryArithInstruction* mir() const { + MOZ_ASSERT(mir_->isDiv() || mir_->isMod()); + return static_cast<MBinaryArithInstruction*>(mir_); + } + + bool canBeDivideByZero() const { + if (mir_->isMod()) { + return mir_->toMod()->canBeDivideByZero(); + } + return mir_->toDiv()->canBeDivideByZero(); + } + + wasm::BytecodeOffset bytecodeOffset() const { + MOZ_ASSERT(mir_->isDiv() || mir_->isMod()); + if (mir_->isMod()) { + return mir_->toMod()->bytecodeOffset(); + } + return mir_->toDiv()->bytecodeOffset(); + } +}; + +class LWasmTruncateToInt64 : public LInstructionHelper<1, 1, 1> { + public: + LIR_HEADER(WasmTruncateToInt64); + + LWasmTruncateToInt64(const LAllocation& in, const LDefinition& temp) + : LInstructionHelper(classOpcode) { + setOperand(0, in); + setTemp(0, temp); + } + + MWasmTruncateToInt64* mir() const { return mir_->toWasmTruncateToInt64(); } + + const LDefinition* temp() { return getTemp(0); } +}; + +} // namespace jit +} // namespace js + +#endif /* jit_x64_LIR_x64_h */ diff --git a/js/src/jit/x64/Lowering-x64.cpp b/js/src/jit/x64/Lowering-x64.cpp new file mode 100644 index 0000000000..d2e328930c --- /dev/null +++ b/js/src/jit/x64/Lowering-x64.cpp @@ -0,0 +1,565 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#include "jit/x64/Lowering-x64.h" + +#include "jit/Lowering.h" +#include "jit/MIR.h" +#include "jit/x64/Assembler-x64.h" + +#include "jit/shared/Lowering-shared-inl.h" + +using namespace js; +using namespace js::jit; + +LBoxAllocation LIRGeneratorX64::useBoxFixed(MDefinition* mir, Register reg1, + Register, bool useAtStart) { + MOZ_ASSERT(mir->type() == MIRType::Value); + + ensureDefined(mir); + return LBoxAllocation(LUse(reg1, mir->virtualRegister(), useAtStart)); +} + +LAllocation LIRGeneratorX64::useByteOpRegister(MDefinition* mir) { + return useRegister(mir); +} + +LAllocation LIRGeneratorX64::useByteOpRegisterAtStart(MDefinition* mir) { + return useRegisterAtStart(mir); +} + +LAllocation LIRGeneratorX64::useByteOpRegisterOrNonDoubleConstant( + MDefinition* mir) { + return useRegisterOrNonDoubleConstant(mir); +} + +LDefinition LIRGeneratorX64::tempByteOpRegister() { return temp(); } + +LDefinition LIRGeneratorX64::tempToUnbox() { return temp(); } + +void LIRGeneratorX64::lowerForALUInt64( + LInstructionHelper<INT64_PIECES, INT64_PIECES, 0>* ins, MDefinition* mir, + MDefinition* input) { + ins->setInt64Operand(0, useInt64RegisterAtStart(input)); + defineInt64ReuseInput(ins, mir, 0); +} + +void LIRGeneratorX64::lowerForALUInt64( + LInstructionHelper<INT64_PIECES, 2 * INT64_PIECES, 0>* ins, + MDefinition* mir, MDefinition* lhs, MDefinition* rhs) { + ins->setInt64Operand(0, useInt64RegisterAtStart(lhs)); + ins->setInt64Operand(INT64_PIECES, willHaveDifferentLIRNodes(lhs, rhs) + ? useInt64OrConstant(rhs) + : useInt64OrConstantAtStart(rhs)); + defineInt64ReuseInput(ins, mir, 0); +} + +void LIRGeneratorX64::lowerForMulInt64(LMulI64* ins, MMul* mir, + MDefinition* lhs, MDefinition* rhs) { + // X64 doesn't need a temp for 64bit multiplication. + ins->setInt64Operand(0, useInt64RegisterAtStart(lhs)); + ins->setInt64Operand(INT64_PIECES, willHaveDifferentLIRNodes(lhs, rhs) + ? useInt64OrConstant(rhs) + : useInt64OrConstantAtStart(rhs)); + defineInt64ReuseInput(ins, mir, 0); +} + +void LIRGenerator::visitBox(MBox* box) { + MDefinition* opd = box->getOperand(0); + + // If the operand is a constant, emit near its uses. + if (opd->isConstant() && box->canEmitAtUses()) { + emitAtUses(box); + return; + } + + if (opd->isConstant()) { + define(new (alloc()) LValue(opd->toConstant()->toJSValue()), box, + LDefinition(LDefinition::BOX)); + } else { + LBox* ins = new (alloc()) LBox(useRegister(opd), opd->type()); + define(ins, box, LDefinition(LDefinition::BOX)); + } +} + +void LIRGenerator::visitUnbox(MUnbox* unbox) { + MDefinition* box = unbox->getOperand(0); + MOZ_ASSERT(box->type() == MIRType::Value); + + LUnboxBase* lir; + if (IsFloatingPointType(unbox->type())) { + lir = new (alloc()) + LUnboxFloatingPoint(useRegisterAtStart(box), unbox->type()); + } else if (unbox->fallible()) { + // If the unbox is fallible, load the Value in a register first to + // avoid multiple loads. + lir = new (alloc()) LUnbox(useRegisterAtStart(box)); + } else { + lir = new (alloc()) LUnbox(useAtStart(box)); + } + + if (unbox->fallible()) { + assignSnapshot(lir, unbox->bailoutKind()); + } + + define(lir, unbox); +} + +void LIRGenerator::visitReturnImpl(MDefinition* opd, bool isGenerator) { + MOZ_ASSERT(opd->type() == MIRType::Value); + + LReturn* ins = new (alloc()) LReturn(isGenerator); + ins->setOperand(0, useFixed(opd, JSReturnReg)); + add(ins); +} + +void LIRGeneratorX64::lowerUntypedPhiInput(MPhi* phi, uint32_t inputPosition, + LBlock* block, size_t lirIndex) { + lowerTypedPhiInput(phi, inputPosition, block, lirIndex); +} + +void LIRGeneratorX64::defineInt64Phi(MPhi* phi, size_t lirIndex) { + defineTypedPhi(phi, lirIndex); +} + +void LIRGeneratorX64::lowerInt64PhiInput(MPhi* phi, uint32_t inputPosition, + LBlock* block, size_t lirIndex) { + lowerTypedPhiInput(phi, inputPosition, block, lirIndex); +} + +void LIRGenerator::visitCompareExchangeTypedArrayElement( + MCompareExchangeTypedArrayElement* ins) { + MOZ_ASSERT(ins->elements()->type() == MIRType::Elements); + MOZ_ASSERT(ins->index()->type() == MIRType::IntPtr); + + if (Scalar::isBigIntType(ins->arrayType())) { + LUse elements = useRegister(ins->elements()); + LAllocation index = + useRegisterOrIndexConstant(ins->index(), ins->arrayType()); + LUse oldval = useRegister(ins->oldval()); + LUse newval = useRegister(ins->newval()); + LInt64Definition temp1 = tempInt64Fixed(Register64(rax)); + LInt64Definition temp2 = tempInt64(); + + auto* lir = new (alloc()) LCompareExchangeTypedArrayElement64( + elements, index, oldval, newval, temp1, temp2); + define(lir, ins); + assignSafepoint(lir, ins); + return; + } + + lowerCompareExchangeTypedArrayElement(ins, + /* useI386ByteRegisters = */ false); +} + +void LIRGenerator::visitAtomicExchangeTypedArrayElement( + MAtomicExchangeTypedArrayElement* ins) { + MOZ_ASSERT(ins->elements()->type() == MIRType::Elements); + MOZ_ASSERT(ins->index()->type() == MIRType::IntPtr); + + if (Scalar::isBigIntType(ins->arrayType())) { + LUse elements = useRegister(ins->elements()); + LAllocation index = + useRegisterOrIndexConstant(ins->index(), ins->arrayType()); + LAllocation value = useRegister(ins->value()); + LInt64Definition temp1 = tempInt64(); + LDefinition temp2 = temp(); + + auto* lir = new (alloc()) LAtomicExchangeTypedArrayElement64( + elements, index, value, temp1, temp2); + define(lir, ins); + assignSafepoint(lir, ins); + return; + } + + lowerAtomicExchangeTypedArrayElement(ins, /* useI386ByteRegisters = */ false); +} + +void LIRGenerator::visitAtomicTypedArrayElementBinop( + MAtomicTypedArrayElementBinop* ins) { + MOZ_ASSERT(ins->elements()->type() == MIRType::Elements); + MOZ_ASSERT(ins->index()->type() == MIRType::IntPtr); + + if (Scalar::isBigIntType(ins->arrayType())) { + LUse elements = useRegister(ins->elements()); + LAllocation index = + useRegisterOrIndexConstant(ins->index(), ins->arrayType()); + LAllocation value = useRegister(ins->value()); + + // Case 1: the result of the operation is not used. + // + // We can omit allocating the result BigInt. + + if (ins->isForEffect()) { + LInt64Definition temp = tempInt64(); + + auto* lir = new (alloc()) LAtomicTypedArrayElementBinopForEffect64( + elements, index, value, temp); + add(lir, ins); + return; + } + + // Case 2: the result of the operation is used. + // + // For ADD and SUB we'll use XADD. + // + // For AND/OR/XOR we need to use a CMPXCHG loop with rax as a temp register. + + bool bitOp = !(ins->operation() == AtomicFetchAddOp || + ins->operation() == AtomicFetchSubOp); + + LInt64Definition temp1 = tempInt64(); + LInt64Definition temp2; + if (bitOp) { + temp2 = tempInt64Fixed(Register64(rax)); + } else { + temp2 = tempInt64(); + } + + auto* lir = new (alloc()) + LAtomicTypedArrayElementBinop64(elements, index, value, temp1, temp2); + define(lir, ins); + assignSafepoint(lir, ins); + return; + } + + lowerAtomicTypedArrayElementBinop(ins, /* useI386ByteRegisters = */ false); +} + +void LIRGeneratorX64::lowerAtomicLoad64(MLoadUnboxedScalar* ins) { + const LUse elements = useRegister(ins->elements()); + const LAllocation index = + useRegisterOrIndexConstant(ins->index(), ins->storageType()); + + auto* lir = new (alloc()) LAtomicLoad64(elements, index, temp(), tempInt64()); + define(lir, ins); + assignSafepoint(lir, ins); +} + +void LIRGeneratorX64::lowerAtomicStore64(MStoreUnboxedScalar* ins) { + LUse elements = useRegister(ins->elements()); + LAllocation index = + useRegisterOrIndexConstant(ins->index(), ins->writeType()); + LAllocation value = useRegister(ins->value()); + + add(new (alloc()) LAtomicStore64(elements, index, value, tempInt64()), ins); +} + +void LIRGenerator::visitWasmUnsignedToDouble(MWasmUnsignedToDouble* ins) { + MOZ_ASSERT(ins->input()->type() == MIRType::Int32); + LWasmUint32ToDouble* lir = + new (alloc()) LWasmUint32ToDouble(useRegisterAtStart(ins->input())); + define(lir, ins); +} + +void LIRGenerator::visitWasmUnsignedToFloat32(MWasmUnsignedToFloat32* ins) { + MOZ_ASSERT(ins->input()->type() == MIRType::Int32); + LWasmUint32ToFloat32* lir = + new (alloc()) LWasmUint32ToFloat32(useRegisterAtStart(ins->input())); + define(lir, ins); +} + +void LIRGenerator::visitWasmHeapBase(MWasmHeapBase* ins) { + auto* lir = new (alloc()) LWasmHeapBase(LAllocation()); + define(lir, ins); +} + +void LIRGenerator::visitWasmLoad(MWasmLoad* ins) { + MDefinition* base = ins->base(); + // 'base' is a GPR but may be of either type. If it is 32-bit it is + // zero-extended and can act as 64-bit. + MOZ_ASSERT(base->type() == MIRType::Int32 || base->type() == MIRType::Int64); + + if (ins->type() != MIRType::Int64) { + auto* lir = new (alloc()) LWasmLoad(useRegisterOrZeroAtStart(base)); + define(lir, ins); + return; + } + + auto* lir = new (alloc()) LWasmLoadI64(useRegisterOrZeroAtStart(base)); + defineInt64(lir, ins); +} + +void LIRGenerator::visitWasmStore(MWasmStore* ins) { + MDefinition* base = ins->base(); + // See comment in visitWasmLoad re the type of 'base'. + MOZ_ASSERT(base->type() == MIRType::Int32 || base->type() == MIRType::Int64); + + MDefinition* value = ins->value(); + LAllocation valueAlloc; + switch (ins->access().type()) { + case Scalar::Int8: + case Scalar::Uint8: + case Scalar::Int16: + case Scalar::Uint16: + case Scalar::Int32: + case Scalar::Uint32: + valueAlloc = useRegisterOrConstantAtStart(value); + break; + case Scalar::Int64: + // No way to encode an int64-to-memory move on x64. + if (value->isConstant() && value->type() != MIRType::Int64) { + valueAlloc = useOrConstantAtStart(value); + } else { + valueAlloc = useRegisterAtStart(value); + } + break; + case Scalar::Float32: + case Scalar::Float64: + valueAlloc = useRegisterAtStart(value); + break; + case Scalar::Simd128: +#ifdef ENABLE_WASM_SIMD + valueAlloc = useRegisterAtStart(value); + break; +#else + MOZ_CRASH("unexpected array type"); +#endif + case Scalar::BigInt64: + case Scalar::BigUint64: + case Scalar::Uint8Clamped: + case Scalar::MaxTypedArrayViewType: + MOZ_CRASH("unexpected array type"); + } + + LAllocation baseAlloc = useRegisterOrZeroAtStart(base); + auto* lir = new (alloc()) LWasmStore(baseAlloc, valueAlloc); + add(lir, ins); +} + +void LIRGenerator::visitWasmCompareExchangeHeap(MWasmCompareExchangeHeap* ins) { + MDefinition* base = ins->base(); + // See comment in visitWasmLoad re the type of 'base'. + MOZ_ASSERT(base->type() == MIRType::Int32 || base->type() == MIRType::Int64); + + // The output may not be used but will be clobbered regardless, so + // pin the output to eax. + // + // The input values must both be in registers. + + const LAllocation oldval = useRegister(ins->oldValue()); + const LAllocation newval = useRegister(ins->newValue()); + + LWasmCompareExchangeHeap* lir = + new (alloc()) LWasmCompareExchangeHeap(useRegister(base), oldval, newval); + + defineFixed(lir, ins, LAllocation(AnyRegister(eax))); +} + +void LIRGenerator::visitWasmAtomicExchangeHeap(MWasmAtomicExchangeHeap* ins) { + // See comment in visitWasmLoad re the type of 'base'. + MOZ_ASSERT(ins->base()->type() == MIRType::Int32 || + ins->base()->type() == MIRType::Int64); + + const LAllocation base = useRegister(ins->base()); + const LAllocation value = useRegister(ins->value()); + + // The output may not be used but will be clobbered regardless, + // so ignore the case where we're not using the value and just + // use the output register as a temp. + + LWasmAtomicExchangeHeap* lir = + new (alloc()) LWasmAtomicExchangeHeap(base, value); + define(lir, ins); +} + +void LIRGenerator::visitWasmAtomicBinopHeap(MWasmAtomicBinopHeap* ins) { + MDefinition* base = ins->base(); + // See comment in visitWasmLoad re the type of 'base'. + MOZ_ASSERT(base->type() == MIRType::Int32 || base->type() == MIRType::Int64); + + // No support for 64-bit operations with constants at the masm level. + + bool canTakeConstant = ins->access().type() != Scalar::Int64; + + // Case 1: the result of the operation is not used. + // + // We'll emit a single instruction: LOCK ADD, LOCK SUB, LOCK AND, + // LOCK OR, or LOCK XOR. + + if (!ins->hasUses()) { + LAllocation value = canTakeConstant ? useRegisterOrConstant(ins->value()) + : useRegister(ins->value()); + LWasmAtomicBinopHeapForEffect* lir = + new (alloc()) LWasmAtomicBinopHeapForEffect(useRegister(base), value); + add(lir, ins); + return; + } + + // Case 2: the result of the operation is used. + // + // For ADD and SUB we'll use XADD with word and byte ops as + // appropriate. Any output register can be used and if value is a + // register it's best if it's the same as output: + // + // movl value, output ; if value != output + // lock xaddl output, mem + // + // For AND/OR/XOR we need to use a CMPXCHG loop, and the output is + // always in rax: + // + // movl *mem, rax + // L: mov rax, temp + // andl value, temp + // lock cmpxchg temp, mem ; reads rax also + // jnz L + // ; result in rax + // + // Note the placement of L, cmpxchg will update rax with *mem if + // *mem does not have the expected value, so reloading it at the + // top of the loop would be redundant. + + bool bitOp = !(ins->operation() == AtomicFetchAddOp || + ins->operation() == AtomicFetchSubOp); + bool reuseInput = false; + LAllocation value; + + if (bitOp || ins->value()->isConstant()) { + value = canTakeConstant ? useRegisterOrConstant(ins->value()) + : useRegister(ins->value()); + } else { + reuseInput = true; + value = useRegisterAtStart(ins->value()); + } + + auto* lir = new (alloc()) LWasmAtomicBinopHeap( + useRegister(base), value, bitOp ? temp() : LDefinition::BogusTemp()); + + if (reuseInput) { + defineReuseInput(lir, ins, LWasmAtomicBinopHeap::valueOp); + } else if (bitOp) { + defineFixed(lir, ins, LAllocation(AnyRegister(rax))); + } else { + define(lir, ins); + } +} + +void LIRGenerator::visitSubstr(MSubstr* ins) { + LSubstr* lir = new (alloc()) + LSubstr(useRegister(ins->string()), useRegister(ins->begin()), + useRegister(ins->length()), temp(), temp(), tempByteOpRegister()); + define(lir, ins); + assignSafepoint(lir, ins); +} + +void LIRGeneratorX64::lowerDivI64(MDiv* div) { + if (div->isUnsigned()) { + lowerUDivI64(div); + return; + } + + LDivOrModI64* lir = new (alloc()) LDivOrModI64( + useRegister(div->lhs()), useRegister(div->rhs()), tempFixed(rdx)); + defineInt64Fixed(lir, div, LInt64Allocation(LAllocation(AnyRegister(rax)))); +} + +void LIRGeneratorX64::lowerWasmBuiltinDivI64(MWasmBuiltinDivI64* div) { + MOZ_CRASH("We don't use runtime div for this architecture"); +} + +void LIRGeneratorX64::lowerModI64(MMod* mod) { + if (mod->isUnsigned()) { + lowerUModI64(mod); + return; + } + + LDivOrModI64* lir = new (alloc()) LDivOrModI64( + useRegister(mod->lhs()), useRegister(mod->rhs()), tempFixed(rax)); + defineInt64Fixed(lir, mod, LInt64Allocation(LAllocation(AnyRegister(rdx)))); +} + +void LIRGeneratorX64::lowerWasmBuiltinModI64(MWasmBuiltinModI64* mod) { + MOZ_CRASH("We don't use runtime mod for this architecture"); +} + +void LIRGeneratorX64::lowerUDivI64(MDiv* div) { + LUDivOrModI64* lir = new (alloc()) LUDivOrModI64( + useRegister(div->lhs()), useRegister(div->rhs()), tempFixed(rdx)); + defineInt64Fixed(lir, div, LInt64Allocation(LAllocation(AnyRegister(rax)))); +} + +void LIRGeneratorX64::lowerUModI64(MMod* mod) { + LUDivOrModI64* lir = new (alloc()) LUDivOrModI64( + useRegister(mod->lhs()), useRegister(mod->rhs()), tempFixed(rax)); + defineInt64Fixed(lir, mod, LInt64Allocation(LAllocation(AnyRegister(rdx)))); +} + +void LIRGeneratorX64::lowerBigIntDiv(MBigIntDiv* ins) { + auto* lir = new (alloc()) LBigIntDiv( + useRegister(ins->lhs()), useRegister(ins->rhs()), tempFixed(rax), temp()); + defineFixed(lir, ins, LAllocation(AnyRegister(rdx))); + assignSafepoint(lir, ins); +} + +void LIRGeneratorX64::lowerBigIntMod(MBigIntMod* ins) { + auto* lir = new (alloc()) LBigIntMod( + useRegister(ins->lhs()), useRegister(ins->rhs()), tempFixed(rax), temp()); + defineFixed(lir, ins, LAllocation(AnyRegister(rdx))); + assignSafepoint(lir, ins); +} + +void LIRGenerator::visitWasmTruncateToInt64(MWasmTruncateToInt64* ins) { + MDefinition* opd = ins->input(); + MOZ_ASSERT(opd->type() == MIRType::Double || opd->type() == MIRType::Float32); + + LDefinition maybeTemp = + ins->isUnsigned() ? tempDouble() : LDefinition::BogusTemp(); + defineInt64(new (alloc()) LWasmTruncateToInt64(useRegister(opd), maybeTemp), + ins); +} + +void LIRGeneratorX64::lowerWasmBuiltinTruncateToInt64( + MWasmBuiltinTruncateToInt64* ins) { + MOZ_CRASH("We don't use it for this architecture"); +} + +void LIRGenerator::visitInt64ToFloatingPoint(MInt64ToFloatingPoint* ins) { + MDefinition* opd = ins->input(); + MOZ_ASSERT(opd->type() == MIRType::Int64); + MOZ_ASSERT(IsFloatingPointType(ins->type())); + + LDefinition maybeTemp = ins->isUnsigned() ? temp() : LDefinition::BogusTemp(); + define(new (alloc()) LInt64ToFloatingPoint(useInt64Register(opd), maybeTemp), + ins); +} + +void LIRGeneratorX64::lowerBuiltinInt64ToFloatingPoint( + MBuiltinInt64ToFloatingPoint* ins) { + MOZ_CRASH("We don't use it for this architecture"); +} + +void LIRGenerator::visitExtendInt32ToInt64(MExtendInt32ToInt64* ins) { + defineInt64(new (alloc()) LExtendInt32ToInt64(useAtStart(ins->input())), ins); +} + +void LIRGenerator::visitSignExtendInt64(MSignExtendInt64* ins) { + defineInt64(new (alloc()) + LSignExtendInt64(useInt64RegisterAtStart(ins->input())), + ins); +} + +// On x64 we specialize the cases: compare is {U,}Int{32,64}, and select is +// {U,}Int{32,64}, independently. +bool LIRGeneratorShared::canSpecializeWasmCompareAndSelect( + MCompare::CompareType compTy, MIRType insTy) { + return (insTy == MIRType::Int32 || insTy == MIRType::Int64) && + (compTy == MCompare::Compare_Int32 || + compTy == MCompare::Compare_UInt32 || + compTy == MCompare::Compare_Int64 || + compTy == MCompare::Compare_UInt64); +} + +void LIRGeneratorShared::lowerWasmCompareAndSelect(MWasmSelect* ins, + MDefinition* lhs, + MDefinition* rhs, + MCompare::CompareType compTy, + JSOp jsop) { + MOZ_ASSERT(canSpecializeWasmCompareAndSelect(compTy, ins->type())); + auto* lir = new (alloc()) LWasmCompareAndSelect( + useRegister(lhs), useAny(rhs), compTy, jsop, + useRegisterAtStart(ins->trueExpr()), useAny(ins->falseExpr())); + defineReuseInput(lir, ins, LWasmCompareAndSelect::IfTrueExprIndex); +} diff --git a/js/src/jit/x64/Lowering-x64.h b/js/src/jit/x64/Lowering-x64.h new file mode 100644 index 0000000000..1c34ea8693 --- /dev/null +++ b/js/src/jit/x64/Lowering-x64.h @@ -0,0 +1,70 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#ifndef jit_x64_Lowering_x64_h +#define jit_x64_Lowering_x64_h + +#include "jit/x86-shared/Lowering-x86-shared.h" + +namespace js { +namespace jit { + +class LIRGeneratorX64 : public LIRGeneratorX86Shared { + protected: + LIRGeneratorX64(MIRGenerator* gen, MIRGraph& graph, LIRGraph& lirGraph) + : LIRGeneratorX86Shared(gen, graph, lirGraph) {} + + void lowerUntypedPhiInput(MPhi* phi, uint32_t inputPosition, LBlock* block, + size_t lirIndex); + void lowerInt64PhiInput(MPhi* phi, uint32_t inputPosition, LBlock* block, + size_t lirIndex); + void defineInt64Phi(MPhi* phi, size_t lirIndex); + + void lowerForALUInt64(LInstructionHelper<INT64_PIECES, INT64_PIECES, 0>* ins, + MDefinition* mir, MDefinition* input); + void lowerForALUInt64( + LInstructionHelper<INT64_PIECES, 2 * INT64_PIECES, 0>* ins, + MDefinition* mir, MDefinition* lhs, MDefinition* rhs); + void lowerForMulInt64(LMulI64* ins, MMul* mir, MDefinition* lhs, + MDefinition* rhs); + + // Returns a box allocation. reg2 is ignored on 64-bit platforms. + LBoxAllocation useBoxFixed(MDefinition* mir, Register reg1, Register, + bool useAtStart = false); + + // x86 has constraints on what registers can be formatted for 1-byte + // stores and loads; on x64 all registers are okay. + LAllocation useByteOpRegister(MDefinition* mir); + LAllocation useByteOpRegisterAtStart(MDefinition* mir); + LAllocation useByteOpRegisterOrNonDoubleConstant(MDefinition* mir); + LDefinition tempByteOpRegister(); + + LDefinition tempToUnbox(); + + bool needTempForPostBarrier() { return true; } + + void lowerBuiltinInt64ToFloatingPoint(MBuiltinInt64ToFloatingPoint* ins); + void lowerWasmBuiltinTruncateToInt64(MWasmBuiltinTruncateToInt64* ins); + void lowerDivI64(MDiv* div); + void lowerWasmBuiltinDivI64(MWasmBuiltinDivI64* div); + void lowerModI64(MMod* mod); + void lowerWasmBuiltinModI64(MWasmBuiltinModI64* mod); + void lowerUDivI64(MDiv* div); + void lowerUModI64(MMod* mod); + + void lowerBigIntDiv(MBigIntDiv* ins); + void lowerBigIntMod(MBigIntMod* ins); + + void lowerAtomicLoad64(MLoadUnboxedScalar* ins); + void lowerAtomicStore64(MStoreUnboxedScalar* ins); +}; + +using LIRGeneratorSpecific = LIRGeneratorX64; + +} // namespace jit +} // namespace js + +#endif /* jit_x64_Lowering_x64_h */ diff --git a/js/src/jit/x64/MacroAssembler-x64-inl.h b/js/src/jit/x64/MacroAssembler-x64-inl.h new file mode 100644 index 0000000000..6869e6c4b6 --- /dev/null +++ b/js/src/jit/x64/MacroAssembler-x64-inl.h @@ -0,0 +1,1099 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#ifndef jit_x64_MacroAssembler_x64_inl_h +#define jit_x64_MacroAssembler_x64_inl_h + +#include "jit/x64/MacroAssembler-x64.h" + +#include "jit/x86-shared/MacroAssembler-x86-shared-inl.h" + +namespace js { +namespace jit { + +//{{{ check_macroassembler_style +// =============================================================== + +void MacroAssembler::move64(Imm64 imm, Register64 dest) { + // Use mov instead of movq because it has special optimizations for imm == 0. + mov(ImmWord(imm.value), dest.reg); +} + +void MacroAssembler::move64(Register64 src, Register64 dest) { + movq(src.reg, dest.reg); +} + +void MacroAssembler::moveDoubleToGPR64(FloatRegister src, Register64 dest) { + vmovq(src, dest.reg); +} + +void MacroAssembler::moveGPR64ToDouble(Register64 src, FloatRegister dest) { + vmovq(src.reg, dest); +} + +void MacroAssembler::move64To32(Register64 src, Register dest) { + movl(src.reg, dest); +} + +void MacroAssembler::move32To64ZeroExtend(Register src, Register64 dest) { + movl(src, dest.reg); +} + +void MacroAssembler::move8To64SignExtend(Register src, Register64 dest) { + movsbq(Operand(src), dest.reg); +} + +void MacroAssembler::move16To64SignExtend(Register src, Register64 dest) { + movswq(Operand(src), dest.reg); +} + +void MacroAssembler::move32To64SignExtend(Register src, Register64 dest) { + movslq(src, dest.reg); +} + +void MacroAssembler::move32SignExtendToPtr(Register src, Register dest) { + movslq(src, dest); +} + +void MacroAssembler::move32ZeroExtendToPtr(Register src, Register dest) { + movl(src, dest); +} + +// =============================================================== +// Load instructions + +void MacroAssembler::load32SignExtendToPtr(const Address& src, Register dest) { + movslq(Operand(src), dest); +} + +// =============================================================== +// Logical instructions + +void MacroAssembler::notPtr(Register reg) { notq(reg); } + +void MacroAssembler::andPtr(Register src, Register dest) { andq(src, dest); } + +void MacroAssembler::andPtr(Imm32 imm, Register dest) { andq(imm, dest); } + +void MacroAssembler::and64(Imm64 imm, Register64 dest) { + if (INT32_MIN <= int64_t(imm.value) && int64_t(imm.value) <= INT32_MAX) { + andq(Imm32(imm.value), dest.reg); + } else { + ScratchRegisterScope scratch(*this); + movq(ImmWord(uintptr_t(imm.value)), scratch); + andq(scratch, dest.reg); + } +} + +void MacroAssembler::or64(Imm64 imm, Register64 dest) { + if (INT32_MIN <= int64_t(imm.value) && int64_t(imm.value) <= INT32_MAX) { + orq(Imm32(imm.value), dest.reg); + } else { + ScratchRegisterScope scratch(*this); + movq(ImmWord(uintptr_t(imm.value)), scratch); + orq(scratch, dest.reg); + } +} + +void MacroAssembler::xor64(Imm64 imm, Register64 dest) { + if (INT32_MIN <= int64_t(imm.value) && int64_t(imm.value) <= INT32_MAX) { + xorq(Imm32(imm.value), dest.reg); + } else { + ScratchRegisterScope scratch(*this); + movq(ImmWord(uintptr_t(imm.value)), scratch); + xorq(scratch, dest.reg); + } +} + +void MacroAssembler::orPtr(Register src, Register dest) { orq(src, dest); } + +void MacroAssembler::orPtr(Imm32 imm, Register dest) { orq(imm, dest); } + +void MacroAssembler::and64(Register64 src, Register64 dest) { + andq(src.reg, dest.reg); +} + +void MacroAssembler::or64(Register64 src, Register64 dest) { + orq(src.reg, dest.reg); +} + +void MacroAssembler::xor64(Register64 src, Register64 dest) { + xorq(src.reg, dest.reg); +} + +void MacroAssembler::xorPtr(Register src, Register dest) { xorq(src, dest); } + +void MacroAssembler::xorPtr(Imm32 imm, Register dest) { xorq(imm, dest); } + +void MacroAssembler::and64(const Operand& src, Register64 dest) { + andq(src, dest.reg); +} + +void MacroAssembler::or64(const Operand& src, Register64 dest) { + orq(src, dest.reg); +} + +void MacroAssembler::xor64(const Operand& src, Register64 dest) { + xorq(src, dest.reg); +} + +// =============================================================== +// Swap instructions + +void MacroAssembler::byteSwap64(Register64 reg) { bswapq(reg.reg); } + +// =============================================================== +// Arithmetic functions + +void MacroAssembler::addPtr(Register src, Register dest) { addq(src, dest); } + +void MacroAssembler::addPtr(Imm32 imm, Register dest) { addq(imm, dest); } + +void MacroAssembler::addPtr(ImmWord imm, Register dest) { + ScratchRegisterScope scratch(*this); + MOZ_ASSERT(dest != scratch); + if ((intptr_t)imm.value <= INT32_MAX && (intptr_t)imm.value >= INT32_MIN) { + addq(Imm32((int32_t)imm.value), dest); + } else { + mov(imm, scratch); + addq(scratch, dest); + } +} + +void MacroAssembler::addPtr(Imm32 imm, const Address& dest) { + addq(imm, Operand(dest)); +} + +void MacroAssembler::addPtr(Imm32 imm, const AbsoluteAddress& dest) { + addq(imm, Operand(dest)); +} + +void MacroAssembler::addPtr(const Address& src, Register dest) { + addq(Operand(src), dest); +} + +void MacroAssembler::add64(const Operand& src, Register64 dest) { + addq(src, dest.reg); +} + +void MacroAssembler::add64(Register64 src, Register64 dest) { + addq(src.reg, dest.reg); +} + +void MacroAssembler::add64(Imm32 imm, Register64 dest) { addq(imm, dest.reg); } + +void MacroAssembler::add64(Imm64 imm, Register64 dest) { + addPtr(ImmWord(imm.value), dest.reg); +} + +CodeOffset MacroAssembler::sub32FromStackPtrWithPatch(Register dest) { + moveStackPtrTo(dest); + addqWithPatch(Imm32(0), dest); + return CodeOffset(currentOffset()); +} + +void MacroAssembler::patchSub32FromStackPtr(CodeOffset offset, Imm32 imm) { + patchAddq(offset, -imm.value); +} + +void MacroAssembler::subPtr(Register src, Register dest) { subq(src, dest); } + +void MacroAssembler::subPtr(Register src, const Address& dest) { + subq(src, Operand(dest)); +} + +void MacroAssembler::subPtr(Imm32 imm, Register dest) { subq(imm, dest); } + +void MacroAssembler::subPtr(ImmWord imm, Register dest) { + ScratchRegisterScope scratch(*this); + MOZ_ASSERT(dest != scratch); + if ((intptr_t)imm.value <= INT32_MAX && (intptr_t)imm.value >= INT32_MIN) { + subq(Imm32((int32_t)imm.value), dest); + } else { + mov(imm, scratch); + subq(scratch, dest); + } +} + +void MacroAssembler::subPtr(const Address& addr, Register dest) { + subq(Operand(addr), dest); +} + +void MacroAssembler::sub64(const Operand& src, Register64 dest) { + subq(src, dest.reg); +} + +void MacroAssembler::sub64(Register64 src, Register64 dest) { + subq(src.reg, dest.reg); +} + +void MacroAssembler::sub64(Imm64 imm, Register64 dest) { + subPtr(ImmWord(imm.value), dest.reg); +} + +void MacroAssembler::mulHighUnsigned32(Imm32 imm, Register src, Register dest) { + // To compute the unsigned multiplication using imulq, we have to ensure both + // operands don't have any bits set in the high word. + + if (imm.value >= 0) { + // Clear the high word of |src|. + movl(src, src); + + // |imm| and |src| are both positive, so directly perform imulq. + imulq(imm, src, dest); + } else { + // Store the low word of |src| into |dest|. + movl(src, dest); + + // Compute the unsigned value of |imm| before performing imulq. + movl(imm, ScratchReg); + imulq(ScratchReg, dest); + } + + // Move the high word into |dest|. + shrq(Imm32(32), dest); +} + +void MacroAssembler::mulPtr(Register rhs, Register srcDest) { + imulq(rhs, srcDest); +} + +void MacroAssembler::mul64(Imm64 imm, const Register64& dest, + const Register temp) { + MOZ_ASSERT(temp == InvalidReg); + mul64(imm, dest); +} + +void MacroAssembler::mul64(Imm64 imm, const Register64& dest) { + if (INT32_MIN <= int64_t(imm.value) && int64_t(imm.value) <= INT32_MAX) { + imulq(Imm32((int32_t)imm.value), dest.reg, dest.reg); + } else { + movq(ImmWord(uintptr_t(imm.value)), ScratchReg); + imulq(ScratchReg, dest.reg); + } +} + +void MacroAssembler::mul64(const Register64& src, const Register64& dest, + const Register temp) { + MOZ_ASSERT(temp == InvalidReg); + mul64(Operand(src.reg), dest); +} + +void MacroAssembler::mul64(const Operand& src, const Register64& dest) { + imulq(src, dest.reg); +} + +void MacroAssembler::mul64(const Operand& src, const Register64& dest, + const Register temp) { + MOZ_ASSERT(temp == InvalidReg); + mul64(src, dest); +} + +void MacroAssembler::mulBy3(Register src, Register dest) { + lea(Operand(src, src, TimesTwo), dest); +} + +void MacroAssembler::mulDoublePtr(ImmPtr imm, Register temp, + FloatRegister dest) { + movq(imm, ScratchReg); + vmulsd(Operand(ScratchReg, 0), dest, dest); +} + +void MacroAssembler::inc64(AbsoluteAddress dest) { + if (X86Encoding::IsAddressImmediate(dest.addr)) { + addPtr(Imm32(1), dest); + } else { + ScratchRegisterScope scratch(*this); + mov(ImmPtr(dest.addr), scratch); + addPtr(Imm32(1), Address(scratch, 0)); + } +} + +void MacroAssembler::neg64(Register64 reg) { negq(reg.reg); } + +void MacroAssembler::negPtr(Register reg) { negq(reg); } + +// =============================================================== +// Shift functions + +void MacroAssembler::lshiftPtr(Imm32 imm, Register dest) { + MOZ_ASSERT(0 <= imm.value && imm.value < 64); + shlq(imm, dest); +} + +void MacroAssembler::lshiftPtr(Register shift, Register srcDest) { + if (Assembler::HasBMI2()) { + shlxq(srcDest, shift, srcDest); + return; + } + MOZ_ASSERT(shift == rcx); + shlq_cl(srcDest); +} + +void MacroAssembler::lshift64(Imm32 imm, Register64 dest) { + MOZ_ASSERT(0 <= imm.value && imm.value < 64); + lshiftPtr(imm, dest.reg); +} + +void MacroAssembler::lshift64(Register shift, Register64 srcDest) { + if (Assembler::HasBMI2()) { + shlxq(srcDest.reg, shift, srcDest.reg); + return; + } + MOZ_ASSERT(shift == rcx); + shlq_cl(srcDest.reg); +} + +void MacroAssembler::rshiftPtr(Imm32 imm, Register dest) { + MOZ_ASSERT(0 <= imm.value && imm.value < 64); + shrq(imm, dest); +} + +void MacroAssembler::rshiftPtr(Register shift, Register srcDest) { + if (Assembler::HasBMI2()) { + shrxq(srcDest, shift, srcDest); + return; + } + MOZ_ASSERT(shift == rcx); + shrq_cl(srcDest); +} + +void MacroAssembler::rshift64(Imm32 imm, Register64 dest) { + rshiftPtr(imm, dest.reg); +} + +void MacroAssembler::rshift64(Register shift, Register64 srcDest) { + if (Assembler::HasBMI2()) { + shrxq(srcDest.reg, shift, srcDest.reg); + return; + } + MOZ_ASSERT(shift == rcx); + shrq_cl(srcDest.reg); +} + +void MacroAssembler::rshiftPtrArithmetic(Imm32 imm, Register dest) { + MOZ_ASSERT(0 <= imm.value && imm.value < 64); + sarq(imm, dest); +} + +void MacroAssembler::rshift64Arithmetic(Imm32 imm, Register64 dest) { + MOZ_ASSERT(0 <= imm.value && imm.value < 64); + rshiftPtrArithmetic(imm, dest.reg); +} + +void MacroAssembler::rshift64Arithmetic(Register shift, Register64 srcDest) { + if (Assembler::HasBMI2()) { + sarxq(srcDest.reg, shift, srcDest.reg); + return; + } + MOZ_ASSERT(shift == rcx); + sarq_cl(srcDest.reg); +} + +// =============================================================== +// Rotation functions + +void MacroAssembler::rotateLeft64(Register count, Register64 src, + Register64 dest) { + MOZ_ASSERT(src == dest, "defineReuseInput"); + MOZ_ASSERT(count == ecx, "defineFixed(ecx)"); + + rolq_cl(dest.reg); +} + +void MacroAssembler::rotateLeft64(Register count, Register64 src, + Register64 dest, Register temp) { + MOZ_ASSERT(temp == InvalidReg); + rotateLeft64(count, src, dest); +} + +void MacroAssembler::rotateRight64(Register count, Register64 src, + Register64 dest) { + MOZ_ASSERT(src == dest, "defineReuseInput"); + MOZ_ASSERT(count == ecx, "defineFixed(ecx)"); + + rorq_cl(dest.reg); +} + +void MacroAssembler::rotateRight64(Register count, Register64 src, + Register64 dest, Register temp) { + MOZ_ASSERT(temp == InvalidReg); + rotateRight64(count, src, dest); +} + +void MacroAssembler::rotateLeft64(Imm32 count, Register64 src, + Register64 dest) { + MOZ_ASSERT(src == dest, "defineReuseInput"); + rolq(count, dest.reg); +} + +void MacroAssembler::rotateLeft64(Imm32 count, Register64 src, Register64 dest, + Register temp) { + MOZ_ASSERT(temp == InvalidReg); + rotateLeft64(count, src, dest); +} + +void MacroAssembler::rotateRight64(Imm32 count, Register64 src, + Register64 dest) { + MOZ_ASSERT(src == dest, "defineReuseInput"); + rorq(count, dest.reg); +} + +void MacroAssembler::rotateRight64(Imm32 count, Register64 src, Register64 dest, + Register temp) { + MOZ_ASSERT(temp == InvalidReg); + rotateRight64(count, src, dest); +} + +// =============================================================== +// Condition functions + +void MacroAssembler::cmp64Set(Condition cond, Address lhs, Imm64 rhs, + Register dest) { + cmpPtrSet(cond, lhs, ImmWord(static_cast<uintptr_t>(rhs.value)), dest); +} + +template <typename T1, typename T2> +void MacroAssembler::cmpPtrSet(Condition cond, T1 lhs, T2 rhs, Register dest) { + cmpPtr(lhs, rhs); + emitSet(cond, dest); +} + +// =============================================================== +// Bit counting functions + +void MacroAssembler::clz64(Register64 src, Register dest) { + if (AssemblerX86Shared::HasLZCNT()) { + lzcntq(src.reg, dest); + return; + } + + Label nonzero; + bsrq(src.reg, dest); + j(Assembler::NonZero, &nonzero); + movq(ImmWord(0x7F), dest); + bind(&nonzero); + xorq(Imm32(0x3F), dest); +} + +void MacroAssembler::ctz64(Register64 src, Register dest) { + if (AssemblerX86Shared::HasBMI1()) { + tzcntq(src.reg, dest); + return; + } + + Label nonzero; + bsfq(src.reg, dest); + j(Assembler::NonZero, &nonzero); + movq(ImmWord(64), dest); + bind(&nonzero); +} + +void MacroAssembler::popcnt64(Register64 src64, Register64 dest64, + Register tmp) { + Register src = src64.reg; + Register dest = dest64.reg; + + if (AssemblerX86Shared::HasPOPCNT()) { + MOZ_ASSERT(tmp == InvalidReg); + popcntq(src, dest); + return; + } + + if (src != dest) { + movq(src, dest); + } + + MOZ_ASSERT(tmp != dest); + + ScratchRegisterScope scratch(*this); + + // Equivalent to mozilla::CountPopulation32, adapted for 64 bits. + // x -= (x >> 1) & m1; + movq(src, tmp); + movq(ImmWord(0x5555555555555555), scratch); + shrq(Imm32(1), tmp); + andq(scratch, tmp); + subq(tmp, dest); + + // x = (x & m2) + ((x >> 2) & m2); + movq(dest, tmp); + movq(ImmWord(0x3333333333333333), scratch); + andq(scratch, dest); + shrq(Imm32(2), tmp); + andq(scratch, tmp); + addq(tmp, dest); + + // x = (x + (x >> 4)) & m4; + movq(dest, tmp); + movq(ImmWord(0x0f0f0f0f0f0f0f0f), scratch); + shrq(Imm32(4), tmp); + addq(tmp, dest); + andq(scratch, dest); + + // (x * h01) >> 56 + movq(ImmWord(0x0101010101010101), scratch); + imulq(scratch, dest); + shrq(Imm32(56), dest); +} + +// =============================================================== +// Branch functions + +void MacroAssembler::branch32(Condition cond, const AbsoluteAddress& lhs, + Register rhs, Label* label) { + if (X86Encoding::IsAddressImmediate(lhs.addr)) { + branch32(cond, Operand(lhs), rhs, label); + } else { + ScratchRegisterScope scratch(*this); + mov(ImmPtr(lhs.addr), scratch); + branch32(cond, Address(scratch, 0), rhs, label); + } +} +void MacroAssembler::branch32(Condition cond, const AbsoluteAddress& lhs, + Imm32 rhs, Label* label) { + if (X86Encoding::IsAddressImmediate(lhs.addr)) { + branch32(cond, Operand(lhs), rhs, label); + } else { + ScratchRegisterScope scratch(*this); + mov(ImmPtr(lhs.addr), scratch); + branch32(cond, Address(scratch, 0), rhs, label); + } +} + +void MacroAssembler::branch32(Condition cond, wasm::SymbolicAddress lhs, + Imm32 rhs, Label* label) { + ScratchRegisterScope scratch(*this); + mov(lhs, scratch); + branch32(cond, Address(scratch, 0), rhs, label); +} + +void MacroAssembler::branch64(Condition cond, Register64 lhs, Imm64 val, + Label* success, Label* fail) { + MOZ_ASSERT(cond == Assembler::NotEqual || cond == Assembler::Equal || + cond == Assembler::LessThan || + cond == Assembler::LessThanOrEqual || + cond == Assembler::GreaterThan || + cond == Assembler::GreaterThanOrEqual || + cond == Assembler::Below || cond == Assembler::BelowOrEqual || + cond == Assembler::Above || cond == Assembler::AboveOrEqual, + "other condition codes not supported"); + + branchPtr(cond, lhs.reg, ImmWord(val.value), success); + if (fail) { + jump(fail); + } +} + +void MacroAssembler::branch64(Condition cond, Register64 lhs, Register64 rhs, + Label* success, Label* fail) { + MOZ_ASSERT(cond == Assembler::NotEqual || cond == Assembler::Equal || + cond == Assembler::LessThan || + cond == Assembler::LessThanOrEqual || + cond == Assembler::GreaterThan || + cond == Assembler::GreaterThanOrEqual || + cond == Assembler::Below || cond == Assembler::BelowOrEqual || + cond == Assembler::Above || cond == Assembler::AboveOrEqual, + "other condition codes not supported"); + + branchPtr(cond, lhs.reg, rhs.reg, success); + if (fail) { + jump(fail); + } +} + +void MacroAssembler::branch64(Condition cond, const Address& lhs, Imm64 val, + Label* label) { + MOZ_ASSERT(cond == Assembler::NotEqual || cond == Assembler::Equal, + "other condition codes not supported"); + + branchPtr(cond, lhs, ImmWord(val.value), label); +} + +void MacroAssembler::branch64(Condition cond, const Address& lhs, + Register64 rhs, Label* label) { + MOZ_ASSERT(cond == Assembler::NotEqual || cond == Assembler::Equal, + "other condition codes not supported"); + + branchPtr(cond, lhs, rhs.reg, label); +} + +void MacroAssembler::branch64(Condition cond, const Address& lhs, + const Address& rhs, Register scratch, + Label* label) { + MOZ_ASSERT(cond == Assembler::NotEqual || cond == Assembler::Equal, + "other condition codes not supported"); + MOZ_ASSERT(lhs.base != scratch); + MOZ_ASSERT(rhs.base != scratch); + + loadPtr(rhs, scratch); + branchPtr(cond, lhs, scratch, label); +} + +void MacroAssembler::branchPtr(Condition cond, const AbsoluteAddress& lhs, + Register rhs, Label* label) { + ScratchRegisterScope scratch(*this); + MOZ_ASSERT(rhs != scratch); + if (X86Encoding::IsAddressImmediate(lhs.addr)) { + branchPtrImpl(cond, Operand(lhs), rhs, label); + } else { + mov(ImmPtr(lhs.addr), scratch); + branchPtrImpl(cond, Operand(scratch, 0x0), rhs, label); + } +} + +void MacroAssembler::branchPtr(Condition cond, const AbsoluteAddress& lhs, + ImmWord rhs, Label* label) { + if (X86Encoding::IsAddressImmediate(lhs.addr)) { + branchPtrImpl(cond, Operand(lhs), rhs, label); + } else { + ScratchRegisterScope scratch(*this); + mov(ImmPtr(lhs.addr), scratch); + branchPtrImpl(cond, Operand(scratch, 0x0), rhs, label); + } +} + +void MacroAssembler::branchPtr(Condition cond, wasm::SymbolicAddress lhs, + Register rhs, Label* label) { + ScratchRegisterScope scratch(*this); + MOZ_ASSERT(rhs != scratch); + mov(lhs, scratch); + branchPtrImpl(cond, Operand(scratch, 0x0), rhs, label); +} + +void MacroAssembler::branchPrivatePtr(Condition cond, const Address& lhs, + Register rhs, Label* label) { + branchPtr(cond, lhs, rhs, label); +} + +void MacroAssembler::branchTruncateFloat32ToPtr(FloatRegister src, + Register dest, Label* fail) { + vcvttss2sq(src, dest); + + // Same trick as for Doubles + cmpPtr(dest, Imm32(1)); + j(Assembler::Overflow, fail); +} + +void MacroAssembler::branchTruncateFloat32MaybeModUint32(FloatRegister src, + Register dest, + Label* fail) { + branchTruncateFloat32ToPtr(src, dest, fail); + movl(dest, dest); // Zero upper 32-bits. +} + +void MacroAssembler::branchTruncateFloat32ToInt32(FloatRegister src, + Register dest, Label* fail) { + branchTruncateFloat32ToPtr(src, dest, fail); + + // Check that the result is in the int32_t range. + ScratchRegisterScope scratch(*this); + move32To64SignExtend(dest, Register64(scratch)); + cmpPtr(dest, scratch); + j(Assembler::NotEqual, fail); + + movl(dest, dest); // Zero upper 32-bits. +} + +void MacroAssembler::branchTruncateDoubleToPtr(FloatRegister src, Register dest, + Label* fail) { + vcvttsd2sq(src, dest); + + // vcvttsd2sq returns 0x8000000000000000 on failure. Test for it by + // subtracting 1 and testing overflow (this avoids the need to + // materialize that value in a register). + cmpPtr(dest, Imm32(1)); + j(Assembler::Overflow, fail); +} + +void MacroAssembler::branchTruncateDoubleMaybeModUint32(FloatRegister src, + Register dest, + Label* fail) { + branchTruncateDoubleToPtr(src, dest, fail); + movl(dest, dest); // Zero upper 32-bits. +} + +void MacroAssembler::branchTruncateDoubleToInt32(FloatRegister src, + Register dest, Label* fail) { + branchTruncateDoubleToPtr(src, dest, fail); + + // Check that the result is in the int32_t range. + ScratchRegisterScope scratch(*this); + move32To64SignExtend(dest, Register64(scratch)); + cmpPtr(dest, scratch); + j(Assembler::NotEqual, fail); + + movl(dest, dest); // Zero upper 32-bits. +} + +void MacroAssembler::branchTest32(Condition cond, const AbsoluteAddress& lhs, + Imm32 rhs, Label* label) { + if (X86Encoding::IsAddressImmediate(lhs.addr)) { + test32(Operand(lhs), rhs); + } else { + ScratchRegisterScope scratch(*this); + mov(ImmPtr(lhs.addr), scratch); + test32(Operand(scratch, 0), rhs); + } + j(cond, label); +} + +template <class L> +void MacroAssembler::branchTest64(Condition cond, Register64 lhs, + Register64 rhs, Register temp, L label) { + branchTestPtr(cond, lhs.reg, rhs.reg, label); +} + +void MacroAssembler::branchTestBooleanTruthy(bool truthy, + const ValueOperand& value, + Label* label) { + test32(value.valueReg(), value.valueReg()); + j(truthy ? NonZero : Zero, label); +} + +void MacroAssembler::branchTestMagic(Condition cond, const Address& valaddr, + JSWhyMagic why, Label* label) { + uint64_t magic = MagicValue(why).asRawBits(); + cmpPtr(valaddr, ImmWord(magic)); + j(cond, label); +} + +void MacroAssembler::branchTestValue(Condition cond, const BaseIndex& lhs, + const ValueOperand& rhs, Label* label) { + MOZ_ASSERT(cond == Assembler::Equal || cond == Assembler::NotEqual); + branchPtr(cond, lhs, rhs.valueReg(), label); +} + +void MacroAssembler::branchToComputedAddress(const BaseIndex& address) { + jmp(Operand(address)); +} + +void MacroAssembler::cmpPtrMovePtr(Condition cond, Register lhs, Register rhs, + Register src, Register dest) { + cmpPtr(lhs, rhs); + cmovCCq(cond, src, dest); +} + +void MacroAssembler::cmpPtrMovePtr(Condition cond, Register lhs, + const Address& rhs, Register src, + Register dest) { + cmpPtr(lhs, Operand(rhs)); + cmovCCq(cond, src, dest); +} + +void MacroAssembler::cmp32MovePtr(Condition cond, Register lhs, Imm32 rhs, + Register src, Register dest) { + cmp32(lhs, rhs); + cmovCCq(cond, Operand(src), dest); +} + +void MacroAssembler::cmp32LoadPtr(Condition cond, const Address& lhs, Imm32 rhs, + const Address& src, Register dest) { + cmp32(lhs, rhs); + cmovCCq(cond, Operand(src), dest); +} + +void MacroAssembler::test32LoadPtr(Condition cond, const Address& addr, + Imm32 mask, const Address& src, + Register dest) { + MOZ_ASSERT(cond == Assembler::Zero || cond == Assembler::NonZero); + test32(addr, mask); + cmovCCq(cond, Operand(src), dest); +} + +void MacroAssembler::test32MovePtr(Condition cond, const Address& addr, + Imm32 mask, Register src, Register dest) { + MOZ_ASSERT(cond == Assembler::Zero || cond == Assembler::NonZero); + test32(addr, mask); + cmovCCq(cond, Operand(src), dest); +} + +void MacroAssembler::spectreMovePtr(Condition cond, Register src, + Register dest) { + cmovCCq(cond, Operand(src), dest); +} + +void MacroAssembler::spectreBoundsCheck32(Register index, Register length, + Register maybeScratch, + Label* failure) { + MOZ_ASSERT(length != maybeScratch); + MOZ_ASSERT(index != maybeScratch); + + ScratchRegisterScope scratch(*this); + MOZ_ASSERT(index != scratch); + MOZ_ASSERT(length != scratch); + + if (JitOptions.spectreIndexMasking) { + move32(Imm32(0), scratch); + } + + cmp32(index, length); + j(Assembler::AboveOrEqual, failure); + + if (JitOptions.spectreIndexMasking) { + cmovCCl(Assembler::AboveOrEqual, scratch, index); + } +} + +void MacroAssembler::spectreBoundsCheck32(Register index, const Address& length, + Register maybeScratch, + Label* failure) { + MOZ_ASSERT(index != length.base); + MOZ_ASSERT(length.base != maybeScratch); + MOZ_ASSERT(index != maybeScratch); + + ScratchRegisterScope scratch(*this); + MOZ_ASSERT(index != scratch); + MOZ_ASSERT(length.base != scratch); + + if (JitOptions.spectreIndexMasking) { + move32(Imm32(0), scratch); + } + + cmp32(index, Operand(length)); + j(Assembler::AboveOrEqual, failure); + + if (JitOptions.spectreIndexMasking) { + cmovCCl(Assembler::AboveOrEqual, scratch, index); + } +} + +void MacroAssembler::spectreBoundsCheckPtr(Register index, Register length, + Register maybeScratch, + Label* failure) { + MOZ_ASSERT(length != maybeScratch); + MOZ_ASSERT(index != maybeScratch); + + ScratchRegisterScope scratch(*this); + MOZ_ASSERT(index != scratch); + MOZ_ASSERT(length != scratch); + + if (JitOptions.spectreIndexMasking) { + movePtr(ImmWord(0), scratch); + } + + cmpPtr(index, length); + j(Assembler::AboveOrEqual, failure); + + if (JitOptions.spectreIndexMasking) { + cmovCCq(Assembler::AboveOrEqual, scratch, index); + } +} + +void MacroAssembler::spectreBoundsCheckPtr(Register index, + const Address& length, + Register maybeScratch, + Label* failure) { + MOZ_ASSERT(index != length.base); + MOZ_ASSERT(length.base != maybeScratch); + MOZ_ASSERT(index != maybeScratch); + + ScratchRegisterScope scratch(*this); + MOZ_ASSERT(index != scratch); + MOZ_ASSERT(length.base != scratch); + + if (JitOptions.spectreIndexMasking) { + movePtr(ImmWord(0), scratch); + } + + cmpPtr(index, Operand(length)); + j(Assembler::AboveOrEqual, failure); + + if (JitOptions.spectreIndexMasking) { + cmovCCq(Assembler::AboveOrEqual, scratch, index); + } +} + +// ======================================================================== +// SIMD. + +// Extract lane as scalar + +void MacroAssembler::extractLaneInt64x2(uint32_t lane, FloatRegister src, + Register64 dest) { + if (lane == 0) { + vmovq(src, dest.reg); + } else { + vpextrq(lane, src, dest.reg); + } +} + +// Replace lane value + +void MacroAssembler::replaceLaneInt64x2(unsigned lane, Register64 rhs, + FloatRegister lhsDest) { + vpinsrq(lane, rhs.reg, lhsDest, lhsDest); +} + +void MacroAssembler::replaceLaneInt64x2(unsigned lane, FloatRegister lhs, + Register64 rhs, FloatRegister dest) { + vpinsrq(lane, rhs.reg, lhs, dest); +} + +// Splat + +void MacroAssembler::splatX2(Register64 src, FloatRegister dest) { + vmovq(src.reg, dest); + if (HasAVX2()) { + vbroadcastq(Operand(dest), dest); + } else { + vpunpcklqdq(dest, dest, dest); + } +} + +// ======================================================================== +// Truncate floating point. + +void MacroAssembler::truncateFloat32ToUInt64(Address src, Address dest, + Register temp, + FloatRegister floatTemp) { + Label done; + + loadFloat32(src, floatTemp); + + truncateFloat32ToInt64(src, dest, temp); + + // For unsigned conversion the case of [INT64, UINT64] needs to get handled + // separately. + loadPtr(dest, temp); + branchPtr(Assembler::Condition::NotSigned, temp, Imm32(0), &done); + + // Move the value inside INT64 range. + storeFloat32(floatTemp, dest); + loadConstantFloat32(double(int64_t(0x8000000000000000)), floatTemp); + vaddss(Operand(dest), floatTemp, floatTemp); + storeFloat32(floatTemp, dest); + truncateFloat32ToInt64(dest, dest, temp); + + loadPtr(dest, temp); + or64(Imm64(0x8000000000000000), Register64(temp)); + storePtr(temp, dest); + + bind(&done); +} + +void MacroAssembler::truncateDoubleToUInt64(Address src, Address dest, + Register temp, + FloatRegister floatTemp) { + Label done; + + loadDouble(src, floatTemp); + + truncateDoubleToInt64(src, dest, temp); + + // For unsigned conversion the case of [INT64, UINT64] needs to get handle + // seperately. + loadPtr(dest, temp); + branchPtr(Assembler::Condition::NotSigned, temp, Imm32(0), &done); + + // Move the value inside INT64 range. + storeDouble(floatTemp, dest); + loadConstantDouble(double(int64_t(0x8000000000000000)), floatTemp); + vaddsd(Operand(dest), floatTemp, floatTemp); + storeDouble(floatTemp, dest); + truncateDoubleToInt64(dest, dest, temp); + + loadPtr(dest, temp); + or64(Imm64(0x8000000000000000), Register64(temp)); + storePtr(temp, dest); + + bind(&done); +} + +void MacroAssemblerX64::fallibleUnboxPtrImpl(const Operand& src, Register dest, + JSValueType type, Label* fail) { + MOZ_ASSERT(type == JSVAL_TYPE_OBJECT || type == JSVAL_TYPE_STRING || + type == JSVAL_TYPE_SYMBOL || type == JSVAL_TYPE_BIGINT); + // dest := src XOR mask + // scratch := dest >> JSVAL_TAG_SHIFT + // fail if scratch != 0 + // + // Note: src and dest can be the same register. + ScratchRegisterScope scratch(asMasm()); + mov(ImmWord(JSVAL_TYPE_TO_SHIFTED_TAG(type)), scratch); + xorq(src, scratch); + mov(scratch, dest); + shrq(Imm32(JSVAL_TAG_SHIFT), scratch); + j(Assembler::NonZero, fail); +} + +void MacroAssembler::fallibleUnboxPtr(const ValueOperand& src, Register dest, + JSValueType type, Label* fail) { + fallibleUnboxPtrImpl(Operand(src.valueReg()), dest, type, fail); +} + +void MacroAssembler::fallibleUnboxPtr(const Address& src, Register dest, + JSValueType type, Label* fail) { + fallibleUnboxPtrImpl(Operand(src), dest, type, fail); +} + +void MacroAssembler::fallibleUnboxPtr(const BaseIndex& src, Register dest, + JSValueType type, Label* fail) { + fallibleUnboxPtrImpl(Operand(src), dest, type, fail); +} + +//}}} check_macroassembler_style +// =============================================================== + +void MacroAssemblerX64::incrementInt32Value(const Address& addr) { + asMasm().addPtr(Imm32(1), addr); +} + +void MacroAssemblerX64::unboxValue(const ValueOperand& src, AnyRegister dest, + JSValueType type) { + if (dest.isFloat()) { + Label notInt32, end; + asMasm().branchTestInt32(Assembler::NotEqual, src, ¬Int32); + convertInt32ToDouble(src.valueReg(), dest.fpu()); + jump(&end); + bind(¬Int32); + unboxDouble(src, dest.fpu()); + bind(&end); + } else { + unboxNonDouble(src, dest.gpr(), type); + } +} + +template <typename T> +void MacroAssemblerX64::loadInt32OrDouble(const T& src, FloatRegister dest) { + Label notInt32, end; + asMasm().branchTestInt32(Assembler::NotEqual, src, ¬Int32); + convertInt32ToDouble(src, dest); + jump(&end); + bind(¬Int32); + unboxDouble(src, dest); + bind(&end); +} + +// If source is a double, load it into dest. If source is int32, +// convert it to double. Else, branch to failure. +void MacroAssemblerX64::ensureDouble(const ValueOperand& source, + FloatRegister dest, Label* failure) { + Label isDouble, done; + { + ScratchTagScope tag(asMasm(), source); + splitTagForTest(source, tag); + asMasm().branchTestDouble(Assembler::Equal, tag, &isDouble); + asMasm().branchTestInt32(Assembler::NotEqual, tag, failure); + } + + { + ScratchRegisterScope scratch(asMasm()); + unboxInt32(source, scratch); + convertInt32ToDouble(scratch, dest); + } + jump(&done); + + bind(&isDouble); + unboxDouble(source, dest); + + bind(&done); +} + +} // namespace jit +} // namespace js + +#endif /* jit_x64_MacroAssembler_x64_inl_h */ diff --git a/js/src/jit/x64/MacroAssembler-x64.cpp b/js/src/jit/x64/MacroAssembler-x64.cpp new file mode 100644 index 0000000000..c6ab1fe935 --- /dev/null +++ b/js/src/jit/x64/MacroAssembler-x64.cpp @@ -0,0 +1,1747 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#include "jit/x64/MacroAssembler-x64.h" + +#include "jit/BaselineFrame.h" +#include "jit/JitFrames.h" +#include "jit/JitRuntime.h" +#include "jit/MacroAssembler.h" +#include "jit/MoveEmitter.h" +#include "util/Memory.h" +#include "vm/BigIntType.h" +#include "vm/JitActivation.h" // js::jit::JitActivation +#include "vm/JSContext.h" +#include "vm/StringType.h" + +#include "jit/MacroAssembler-inl.h" + +using namespace js; +using namespace js::jit; + +void MacroAssemblerX64::loadConstantDouble(double d, FloatRegister dest) { + if (maybeInlineDouble(d, dest)) { + return; + } + Double* dbl = getDouble(d); + if (!dbl) { + return; + } + // The constants will be stored in a pool appended to the text (see + // finish()), so they will always be a fixed distance from the + // instructions which reference them. This allows the instructions to use + // PC-relative addressing. Use "jump" label support code, because we need + // the same PC-relative address patching that jumps use. + JmpSrc j = masm.vmovsd_ripr(dest.encoding()); + propagateOOM(dbl->uses.append(j)); +} + +void MacroAssemblerX64::loadConstantFloat32(float f, FloatRegister dest) { + if (maybeInlineFloat(f, dest)) { + return; + } + Float* flt = getFloat(f); + if (!flt) { + return; + } + // See comment in loadConstantDouble + JmpSrc j = masm.vmovss_ripr(dest.encoding()); + propagateOOM(flt->uses.append(j)); +} + +void MacroAssemblerX64::vpRiprOpSimd128( + const SimdConstant& v, FloatRegister reg, + JmpSrc (X86Encoding::BaseAssemblerX64::*op)( + X86Encoding::XMMRegisterID id)) { + SimdData* val = getSimdData(v); + if (!val) { + return; + } + JmpSrc j = (masm.*op)(reg.encoding()); + propagateOOM(val->uses.append(j)); +} + +void MacroAssemblerX64::vpRiprOpSimd128( + const SimdConstant& v, FloatRegister src, FloatRegister dest, + JmpSrc (X86Encoding::BaseAssemblerX64::*op)( + X86Encoding::XMMRegisterID srcId, X86Encoding::XMMRegisterID destId)) { + SimdData* val = getSimdData(v); + if (!val) { + return; + } + JmpSrc j = (masm.*op)(src.encoding(), dest.encoding()); + propagateOOM(val->uses.append(j)); +} + +void MacroAssemblerX64::loadConstantSimd128Int(const SimdConstant& v, + FloatRegister dest) { + if (maybeInlineSimd128Int(v, dest)) { + return; + } + vpRiprOpSimd128(v, dest, &X86Encoding::BaseAssemblerX64::vmovdqa_ripr); +} + +void MacroAssemblerX64::loadConstantSimd128Float(const SimdConstant& v, + FloatRegister dest) { + if (maybeInlineSimd128Float(v, dest)) { + return; + } + vpRiprOpSimd128(v, dest, &X86Encoding::BaseAssemblerX64::vmovaps_ripr); +} + +void MacroAssemblerX64::vpaddbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpaddb_ripr); +} + +void MacroAssemblerX64::vpaddwSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpaddw_ripr); +} + +void MacroAssemblerX64::vpadddSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpaddd_ripr); +} + +void MacroAssemblerX64::vpaddqSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpaddq_ripr); +} + +void MacroAssemblerX64::vpsubbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpsubb_ripr); +} + +void MacroAssemblerX64::vpsubwSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpsubw_ripr); +} + +void MacroAssemblerX64::vpsubdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpsubd_ripr); +} + +void MacroAssemblerX64::vpsubqSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpsubq_ripr); +} + +void MacroAssemblerX64::vpmullwSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpmullw_ripr); +} + +void MacroAssemblerX64::vpmulldSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpmulld_ripr); +} + +void MacroAssemblerX64::vpaddsbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpaddsb_ripr); +} + +void MacroAssemblerX64::vpaddusbSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpaddusb_ripr); +} + +void MacroAssemblerX64::vpaddswSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpaddsw_ripr); +} + +void MacroAssemblerX64::vpadduswSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpaddusw_ripr); +} + +void MacroAssemblerX64::vpsubsbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpsubsb_ripr); +} + +void MacroAssemblerX64::vpsubusbSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpsubusb_ripr); +} + +void MacroAssemblerX64::vpsubswSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpsubsw_ripr); +} + +void MacroAssemblerX64::vpsubuswSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpsubusw_ripr); +} + +void MacroAssemblerX64::vpminsbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpminsb_ripr); +} + +void MacroAssemblerX64::vpminubSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpminub_ripr); +} + +void MacroAssemblerX64::vpminswSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpminsw_ripr); +} + +void MacroAssemblerX64::vpminuwSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpminuw_ripr); +} + +void MacroAssemblerX64::vpminsdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpminsd_ripr); +} + +void MacroAssemblerX64::vpminudSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpminud_ripr); +} + +void MacroAssemblerX64::vpmaxsbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpmaxsb_ripr); +} + +void MacroAssemblerX64::vpmaxubSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpmaxub_ripr); +} + +void MacroAssemblerX64::vpmaxswSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpmaxsw_ripr); +} + +void MacroAssemblerX64::vpmaxuwSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpmaxuw_ripr); +} + +void MacroAssemblerX64::vpmaxsdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpmaxsd_ripr); +} + +void MacroAssemblerX64::vpmaxudSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpmaxud_ripr); +} + +void MacroAssemblerX64::vpandSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpand_ripr); +} + +void MacroAssemblerX64::vpxorSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpxor_ripr); +} + +void MacroAssemblerX64::vporSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpor_ripr); +} + +void MacroAssemblerX64::vaddpsSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vaddps_ripr); +} + +void MacroAssemblerX64::vaddpdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vaddpd_ripr); +} + +void MacroAssemblerX64::vsubpsSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vsubps_ripr); +} + +void MacroAssemblerX64::vsubpdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vsubpd_ripr); +} + +void MacroAssemblerX64::vdivpsSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vdivps_ripr); +} + +void MacroAssemblerX64::vdivpdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vdivpd_ripr); +} + +void MacroAssemblerX64::vmulpsSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vmulps_ripr); +} + +void MacroAssemblerX64::vmulpdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vmulpd_ripr); +} + +void MacroAssemblerX64::vandpdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vandpd_ripr); +} + +void MacroAssemblerX64::vminpdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vminpd_ripr); +} + +void MacroAssemblerX64::vpacksswbSimd128(const SimdConstant& v, + FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpacksswb_ripr); +} + +void MacroAssemblerX64::vpackuswbSimd128(const SimdConstant& v, + FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpackuswb_ripr); +} + +void MacroAssemblerX64::vpackssdwSimd128(const SimdConstant& v, + FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpackssdw_ripr); +} + +void MacroAssemblerX64::vpackusdwSimd128(const SimdConstant& v, + FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpackusdw_ripr); +} + +void MacroAssemblerX64::vpunpckldqSimd128(const SimdConstant& v, + FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, + &X86Encoding::BaseAssemblerX64::vpunpckldq_ripr); +} + +void MacroAssemblerX64::vunpcklpsSimd128(const SimdConstant& v, + FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vunpcklps_ripr); +} + +void MacroAssemblerX64::vpshufbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpshufb_ripr); +} + +void MacroAssemblerX64::vptestSimd128(const SimdConstant& v, + FloatRegister lhs) { + vpRiprOpSimd128(v, lhs, &X86Encoding::BaseAssemblerX64::vptest_ripr); +} + +void MacroAssemblerX64::vpmaddwdSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpmaddwd_ripr); +} + +void MacroAssemblerX64::vpcmpeqbSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpcmpeqb_ripr); +} + +void MacroAssemblerX64::vpcmpgtbSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpcmpgtb_ripr); +} + +void MacroAssemblerX64::vpcmpeqwSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpcmpeqw_ripr); +} + +void MacroAssemblerX64::vpcmpgtwSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpcmpgtw_ripr); +} + +void MacroAssemblerX64::vpcmpeqdSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpcmpeqd_ripr); +} + +void MacroAssemblerX64::vpcmpgtdSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpcmpgtd_ripr); +} + +void MacroAssemblerX64::vcmpeqpsSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vcmpeqps_ripr); +} + +void MacroAssemblerX64::vcmpneqpsSimd128(const SimdConstant& v, + FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vcmpneqps_ripr); +} + +void MacroAssemblerX64::vcmpltpsSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vcmpltps_ripr); +} + +void MacroAssemblerX64::vcmplepsSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vcmpleps_ripr); +} + +void MacroAssemblerX64::vcmpgepsSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vcmpgeps_ripr); +} + +void MacroAssemblerX64::vcmpeqpdSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vcmpeqpd_ripr); +} + +void MacroAssemblerX64::vcmpneqpdSimd128(const SimdConstant& v, + FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vcmpneqpd_ripr); +} + +void MacroAssemblerX64::vcmpltpdSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vcmpltpd_ripr); +} + +void MacroAssemblerX64::vcmplepdSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vcmplepd_ripr); +} + +void MacroAssemblerX64::vpmaddubswSimd128(const SimdConstant& v, + FloatRegister lhs, + FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, + &X86Encoding::BaseAssemblerX64::vpmaddubsw_ripr); +} + +void MacroAssemblerX64::vpmuludqSimd128(const SimdConstant& v, + FloatRegister lhs, FloatRegister dest) { + vpRiprOpSimd128(v, lhs, dest, &X86Encoding::BaseAssemblerX64::vpmuludq_ripr); +} + +void MacroAssemblerX64::bindOffsets( + const MacroAssemblerX86Shared::UsesVector& uses) { + for (JmpSrc src : uses) { + JmpDst dst(currentOffset()); + // Using linkJump here is safe, as explained in the comment in + // loadConstantDouble. + masm.linkJump(src, dst); + } +} + +void MacroAssemblerX64::finish() { + if (!doubles_.empty()) { + masm.haltingAlign(sizeof(double)); + } + for (const Double& d : doubles_) { + bindOffsets(d.uses); + masm.doubleConstant(d.value); + } + + if (!floats_.empty()) { + masm.haltingAlign(sizeof(float)); + } + for (const Float& f : floats_) { + bindOffsets(f.uses); + masm.floatConstant(f.value); + } + + // SIMD memory values must be suitably aligned. + if (!simds_.empty()) { + masm.haltingAlign(SimdMemoryAlignment); + } + for (const SimdData& v : simds_) { + bindOffsets(v.uses); + masm.simd128Constant(v.value.bytes()); + } + + MacroAssemblerX86Shared::finish(); +} + +void MacroAssemblerX64::boxValue(JSValueType type, Register src, + Register dest) { + MOZ_ASSERT(src != dest); + + JSValueShiftedTag tag = (JSValueShiftedTag)JSVAL_TYPE_TO_SHIFTED_TAG(type); +#ifdef DEBUG + if (type == JSVAL_TYPE_INT32 || type == JSVAL_TYPE_BOOLEAN) { + Label upper32BitsZeroed; + movePtr(ImmWord(UINT32_MAX), dest); + asMasm().branchPtr(Assembler::BelowOrEqual, src, dest, &upper32BitsZeroed); + breakpoint(); + bind(&upper32BitsZeroed); + } +#endif + mov(ImmShiftedTag(tag), dest); + orq(src, dest); +} + +void MacroAssemblerX64::handleFailureWithHandlerTail(Label* profilerExitTail, + Label* bailoutTail) { + // Reserve space for exception information. + subq(Imm32(sizeof(ResumeFromException)), rsp); + movq(rsp, rax); + + // Call the handler. + using Fn = void (*)(ResumeFromException * rfe); + asMasm().setupUnalignedABICall(rcx); + asMasm().passABIArg(rax); + asMasm().callWithABI<Fn, HandleException>( + MoveOp::GENERAL, CheckUnsafeCallWithABI::DontCheckHasExitFrame); + + Label entryFrame; + Label catch_; + Label finally; + Label returnBaseline; + Label returnIon; + Label bailout; + Label wasm; + Label wasmCatch; + + load32(Address(rsp, ResumeFromException::offsetOfKind()), rax); + asMasm().branch32(Assembler::Equal, rax, + Imm32(ExceptionResumeKind::EntryFrame), &entryFrame); + asMasm().branch32(Assembler::Equal, rax, Imm32(ExceptionResumeKind::Catch), + &catch_); + asMasm().branch32(Assembler::Equal, rax, Imm32(ExceptionResumeKind::Finally), + &finally); + asMasm().branch32(Assembler::Equal, rax, + Imm32(ExceptionResumeKind::ForcedReturnBaseline), + &returnBaseline); + asMasm().branch32(Assembler::Equal, rax, + Imm32(ExceptionResumeKind::ForcedReturnIon), &returnIon); + asMasm().branch32(Assembler::Equal, rax, Imm32(ExceptionResumeKind::Bailout), + &bailout); + asMasm().branch32(Assembler::Equal, rax, Imm32(ExceptionResumeKind::Wasm), + &wasm); + asMasm().branch32(Assembler::Equal, rax, + Imm32(ExceptionResumeKind::WasmCatch), &wasmCatch); + + breakpoint(); // Invalid kind. + + // No exception handler. Load the error value, restore state and return from + // the entry frame. + bind(&entryFrame); + asMasm().moveValue(MagicValue(JS_ION_ERROR), JSReturnOperand); + loadPtr(Address(rsp, ResumeFromException::offsetOfFramePointer()), rbp); + loadPtr(Address(rsp, ResumeFromException::offsetOfStackPointer()), rsp); + ret(); + + // If we found a catch handler, this must be a baseline frame. Restore state + // and jump to the catch block. + bind(&catch_); + loadPtr(Address(rsp, ResumeFromException::offsetOfTarget()), rax); + loadPtr(Address(rsp, ResumeFromException::offsetOfFramePointer()), rbp); + loadPtr(Address(rsp, ResumeFromException::offsetOfStackPointer()), rsp); + jmp(Operand(rax)); + + // If we found a finally block, this must be a baseline frame. Push two + // values expected by the finally block: the exception and BooleanValue(true). + bind(&finally); + ValueOperand exception = ValueOperand(rcx); + loadValue(Address(esp, ResumeFromException::offsetOfException()), exception); + + loadPtr(Address(rsp, ResumeFromException::offsetOfTarget()), rax); + loadPtr(Address(rsp, ResumeFromException::offsetOfFramePointer()), rbp); + loadPtr(Address(rsp, ResumeFromException::offsetOfStackPointer()), rsp); + + pushValue(exception); + pushValue(BooleanValue(true)); + jmp(Operand(rax)); + + // Return BaselineFrame->returnValue() to the caller. + // Used in debug mode and for GeneratorReturn. + Label profilingInstrumentation; + bind(&returnBaseline); + loadPtr(Address(rsp, ResumeFromException::offsetOfFramePointer()), rbp); + loadPtr(Address(rsp, ResumeFromException::offsetOfStackPointer()), rsp); + loadValue(Address(rbp, BaselineFrame::reverseOffsetOfReturnValue()), + JSReturnOperand); + jmp(&profilingInstrumentation); + + // Return the given value to the caller. + bind(&returnIon); + loadValue(Address(rsp, ResumeFromException::offsetOfException()), + JSReturnOperand); + loadPtr(Address(rsp, ResumeFromException::offsetOfFramePointer()), rbp); + loadPtr(Address(rsp, ResumeFromException::offsetOfStackPointer()), rsp); + + // If profiling is enabled, then update the lastProfilingFrame to refer to + // caller frame before returning. This code is shared by ForcedReturnIon + // and ForcedReturnBaseline. + bind(&profilingInstrumentation); + { + Label skipProfilingInstrumentation; + AbsoluteAddress addressOfEnabled( + asMasm().runtime()->geckoProfiler().addressOfEnabled()); + asMasm().branch32(Assembler::Equal, addressOfEnabled, Imm32(0), + &skipProfilingInstrumentation); + jump(profilerExitTail); + bind(&skipProfilingInstrumentation); + } + + movq(rbp, rsp); + pop(rbp); + ret(); + + // If we are bailing out to baseline to handle an exception, jump to the + // bailout tail stub. Load 1 (true) in ReturnReg to indicate success. + bind(&bailout); + loadPtr(Address(rsp, ResumeFromException::offsetOfBailoutInfo()), r9); + loadPtr(Address(rsp, ResumeFromException::offsetOfStackPointer()), rsp); + move32(Imm32(1), ReturnReg); + jump(bailoutTail); + + // If we are throwing and the innermost frame was a wasm frame, reset SP and + // FP; SP is pointing to the unwound return address to the wasm entry, so + // we can just ret(). + bind(&wasm); + loadPtr(Address(rsp, ResumeFromException::offsetOfFramePointer()), rbp); + loadPtr(Address(rsp, ResumeFromException::offsetOfStackPointer()), rsp); + movePtr(ImmPtr((const void*)wasm::FailInstanceReg), InstanceReg); + masm.ret(); + + // Found a wasm catch handler, restore state and jump to it. + bind(&wasmCatch); + loadPtr(Address(rsp, ResumeFromException::offsetOfTarget()), rax); + loadPtr(Address(rsp, ResumeFromException::offsetOfFramePointer()), rbp); + loadPtr(Address(rsp, ResumeFromException::offsetOfStackPointer()), rsp); + jmp(Operand(rax)); +} + +void MacroAssemblerX64::profilerEnterFrame(Register framePtr, + Register scratch) { + asMasm().loadJSContext(scratch); + loadPtr(Address(scratch, offsetof(JSContext, profilingActivation_)), scratch); + storePtr(framePtr, + Address(scratch, JitActivation::offsetOfLastProfilingFrame())); + storePtr(ImmPtr(nullptr), + Address(scratch, JitActivation::offsetOfLastProfilingCallSite())); +} + +void MacroAssemblerX64::profilerExitFrame() { + jump(asMasm().runtime()->jitRuntime()->getProfilerExitFrameTail()); +} + +Assembler::Condition MacroAssemblerX64::testStringTruthy( + bool truthy, const ValueOperand& value) { + ScratchRegisterScope scratch(asMasm()); + unboxString(value, scratch); + cmp32(Operand(scratch, JSString::offsetOfLength()), Imm32(0)); + return truthy ? Assembler::NotEqual : Assembler::Equal; +} + +Assembler::Condition MacroAssemblerX64::testBigIntTruthy( + bool truthy, const ValueOperand& value) { + ScratchRegisterScope scratch(asMasm()); + unboxBigInt(value, scratch); + cmp32(Operand(scratch, JS::BigInt::offsetOfDigitLength()), Imm32(0)); + return truthy ? Assembler::NotEqual : Assembler::Equal; +} + +MacroAssembler& MacroAssemblerX64::asMasm() { + return *static_cast<MacroAssembler*>(this); +} + +const MacroAssembler& MacroAssemblerX64::asMasm() const { + return *static_cast<const MacroAssembler*>(this); +} + +void MacroAssembler::subFromStackPtr(Imm32 imm32) { + if (imm32.value) { + // On windows, we cannot skip very far down the stack without touching the + // memory pages in-between. This is a corner-case code for situations where + // the Ion frame data for a piece of code is very large. To handle this + // special case, for frames over 4k in size we allocate memory on the stack + // incrementally, touching it as we go. + // + // When the amount is quite large, which it can be, we emit an actual loop, + // in order to keep the function prologue compact. Compactness is a + // requirement for eg Wasm's CodeRange data structure, which can encode only + // 8-bit offsets. + uint32_t amountLeft = imm32.value; + uint32_t fullPages = amountLeft / 4096; + if (fullPages <= 8) { + while (amountLeft > 4096) { + subq(Imm32(4096), StackPointer); + store32(Imm32(0), Address(StackPointer, 0)); + amountLeft -= 4096; + } + subq(Imm32(amountLeft), StackPointer); + } else { + ScratchRegisterScope scratch(*this); + Label top; + move32(Imm32(fullPages), scratch); + bind(&top); + subq(Imm32(4096), StackPointer); + store32(Imm32(0), Address(StackPointer, 0)); + subl(Imm32(1), scratch); + j(Assembler::NonZero, &top); + amountLeft -= fullPages * 4096; + if (amountLeft) { + subq(Imm32(amountLeft), StackPointer); + } + } + } +} + +void MacroAssemblerX64::convertDoubleToPtr(FloatRegister src, Register dest, + Label* fail, + bool negativeZeroCheck) { + // Check for -0.0 + if (negativeZeroCheck) { + branchNegativeZero(src, dest, fail); + } + + ScratchDoubleScope scratch(asMasm()); + vcvttsd2sq(src, dest); + asMasm().convertInt64ToDouble(Register64(dest), scratch); + vucomisd(scratch, src); + j(Assembler::Parity, fail); + j(Assembler::NotEqual, fail); +} + +//{{{ check_macroassembler_style +// =============================================================== +// ABI function calls. + +void MacroAssembler::setupUnalignedABICall(Register scratch) { + setupNativeABICall(); + dynamicAlignment_ = true; + + movq(rsp, scratch); + andq(Imm32(~(ABIStackAlignment - 1)), rsp); + push(scratch); +} + +void MacroAssembler::callWithABIPre(uint32_t* stackAdjust, bool callFromWasm) { + MOZ_ASSERT(inCall_); + uint32_t stackForCall = abiArgs_.stackBytesConsumedSoFar(); + + if (dynamicAlignment_) { + // sizeof(intptr_t) accounts for the saved stack pointer pushed by + // setupUnalignedABICall. + stackForCall += ComputeByteAlignment(stackForCall + sizeof(intptr_t), + ABIStackAlignment); + } else { + uint32_t alignmentAtPrologue = callFromWasm ? sizeof(wasm::Frame) : 0; + stackForCall += ComputeByteAlignment( + stackForCall + framePushed() + alignmentAtPrologue, ABIStackAlignment); + } + + *stackAdjust = stackForCall; + reserveStack(stackForCall); + + // Position all arguments. + { + enoughMemory_ &= moveResolver_.resolve(); + if (!enoughMemory_) { + return; + } + + MoveEmitter emitter(*this); + emitter.emit(moveResolver_); + emitter.finish(); + } + + assertStackAlignment(ABIStackAlignment); +} + +void MacroAssembler::callWithABIPost(uint32_t stackAdjust, MoveOp::Type result, + bool cleanupArg) { + freeStack(stackAdjust); + if (dynamicAlignment_) { + pop(rsp); + } + +#ifdef DEBUG + MOZ_ASSERT(inCall_); + inCall_ = false; +#endif +} + +static bool IsIntArgReg(Register reg) { + for (uint32_t i = 0; i < NumIntArgRegs; i++) { + if (IntArgRegs[i] == reg) { + return true; + } + } + + return false; +} + +void MacroAssembler::callWithABINoProfiler(Register fun, MoveOp::Type result) { + if (IsIntArgReg(fun)) { + // Callee register may be clobbered for an argument. Move the callee to + // r10, a volatile, non-argument register. + propagateOOM(moveResolver_.addMove(MoveOperand(fun), MoveOperand(r10), + MoveOp::GENERAL)); + fun = r10; + } + + MOZ_ASSERT(!IsIntArgReg(fun)); + + uint32_t stackAdjust; + callWithABIPre(&stackAdjust); + call(fun); + callWithABIPost(stackAdjust, result); +} + +void MacroAssembler::callWithABINoProfiler(const Address& fun, + MoveOp::Type result) { + Address safeFun = fun; + if (IsIntArgReg(safeFun.base)) { + // Callee register may be clobbered for an argument. Move the callee to + // r10, a volatile, non-argument register. + propagateOOM(moveResolver_.addMove(MoveOperand(fun.base), MoveOperand(r10), + MoveOp::GENERAL)); + safeFun.base = r10; + } + + MOZ_ASSERT(!IsIntArgReg(safeFun.base)); + + uint32_t stackAdjust; + callWithABIPre(&stackAdjust); + call(safeFun); + callWithABIPost(stackAdjust, result); +} + +// =============================================================== +// Move instructions + +void MacroAssembler::moveValue(const TypedOrValueRegister& src, + const ValueOperand& dest) { + if (src.hasValue()) { + moveValue(src.valueReg(), dest); + return; + } + + MIRType type = src.type(); + AnyRegister reg = src.typedReg(); + + if (!IsFloatingPointType(type)) { + boxValue(ValueTypeFromMIRType(type), reg.gpr(), dest.valueReg()); + return; + } + + ScratchDoubleScope scratch(*this); + FloatRegister freg = reg.fpu(); + if (type == MIRType::Float32) { + convertFloat32ToDouble(freg, scratch); + freg = scratch; + } + boxDouble(freg, dest, freg); +} + +void MacroAssembler::moveValue(const ValueOperand& src, + const ValueOperand& dest) { + if (src == dest) { + return; + } + movq(src.valueReg(), dest.valueReg()); +} + +void MacroAssembler::moveValue(const Value& src, const ValueOperand& dest) { + movWithPatch(ImmWord(src.asRawBits()), dest.valueReg()); + writeDataRelocation(src); +} + +// =============================================================== +// Branch functions + +void MacroAssembler::loadStoreBuffer(Register ptr, Register buffer) { + if (ptr != buffer) { + movePtr(ptr, buffer); + } + andPtr(Imm32(int32_t(~gc::ChunkMask)), buffer); + loadPtr(Address(buffer, gc::ChunkStoreBufferOffset), buffer); +} + +void MacroAssembler::branchPtrInNurseryChunk(Condition cond, Register ptr, + Register temp, Label* label) { + MOZ_ASSERT(cond == Assembler::Equal || cond == Assembler::NotEqual); + + ScratchRegisterScope scratch(*this); + MOZ_ASSERT(ptr != temp); + MOZ_ASSERT(ptr != scratch); + + movePtr(ptr, scratch); + andPtr(Imm32(int32_t(~gc::ChunkMask)), scratch); + branchPtr(InvertCondition(cond), Address(scratch, gc::ChunkStoreBufferOffset), + ImmWord(0), label); +} + +template <typename T> +void MacroAssembler::branchValueIsNurseryCellImpl(Condition cond, + const T& value, Register temp, + Label* label) { + MOZ_ASSERT(cond == Assembler::Equal || cond == Assembler::NotEqual); + MOZ_ASSERT(temp != InvalidReg); + + Label done; + branchTestGCThing(Assembler::NotEqual, value, + cond == Assembler::Equal ? &done : label); + + getGCThingValueChunk(value, temp); + branchPtr(InvertCondition(cond), Address(temp, gc::ChunkStoreBufferOffset), + ImmWord(0), label); + + bind(&done); +} + +void MacroAssembler::branchValueIsNurseryCell(Condition cond, + const Address& address, + Register temp, Label* label) { + branchValueIsNurseryCellImpl(cond, address, temp, label); +} + +void MacroAssembler::branchValueIsNurseryCell(Condition cond, + ValueOperand value, Register temp, + Label* label) { + branchValueIsNurseryCellImpl(cond, value, temp, label); +} + +void MacroAssembler::branchTestValue(Condition cond, const ValueOperand& lhs, + const Value& rhs, Label* label) { + MOZ_ASSERT(cond == Equal || cond == NotEqual); + ScratchRegisterScope scratch(*this); + MOZ_ASSERT(lhs.valueReg() != scratch); + moveValue(rhs, ValueOperand(scratch)); + cmpPtr(lhs.valueReg(), scratch); + j(cond, label); +} + +// ======================================================================== +// Memory access primitives. +template <typename T> +void MacroAssembler::storeUnboxedValue(const ConstantOrRegister& value, + MIRType valueType, const T& dest) { + MOZ_ASSERT(valueType < MIRType::Value); + + if (valueType == MIRType::Double) { + boxDouble(value.reg().typedReg().fpu(), dest); + return; + } + + if (value.constant()) { + storeValue(value.value(), dest); + } else { + storeValue(ValueTypeFromMIRType(valueType), value.reg().typedReg().gpr(), + dest); + } +} + +template void MacroAssembler::storeUnboxedValue(const ConstantOrRegister& value, + MIRType valueType, + const Address& dest); +template void MacroAssembler::storeUnboxedValue( + const ConstantOrRegister& value, MIRType valueType, + const BaseObjectElementIndex& dest); + +void MacroAssembler::PushBoxed(FloatRegister reg) { + subq(Imm32(sizeof(double)), StackPointer); + boxDouble(reg, Address(StackPointer, 0)); + adjustFrame(sizeof(double)); +} + +// ======================================================================== +// wasm support + +void MacroAssembler::wasmLoad(const wasm::MemoryAccessDesc& access, + Operand srcAddr, AnyRegister out) { + // NOTE: the generated code must match the assembly code in gen_load in + // GenerateAtomicOperations.py + memoryBarrierBefore(access.sync()); + + MOZ_ASSERT_IF( + access.isZeroExtendSimd128Load(), + access.type() == Scalar::Float32 || access.type() == Scalar::Float64); + MOZ_ASSERT_IF( + access.isSplatSimd128Load(), + access.type() == Scalar::Uint8 || access.type() == Scalar::Uint16 || + access.type() == Scalar::Float32 || access.type() == Scalar::Float64); + MOZ_ASSERT_IF(access.isWidenSimd128Load(), access.type() == Scalar::Float64); + + append(access, size()); + switch (access.type()) { + case Scalar::Int8: + movsbl(srcAddr, out.gpr()); + break; + case Scalar::Uint8: + if (access.isSplatSimd128Load()) { + vbroadcastb(srcAddr, out.fpu()); + } else { + movzbl(srcAddr, out.gpr()); + } + break; + case Scalar::Int16: + movswl(srcAddr, out.gpr()); + break; + case Scalar::Uint16: + if (access.isSplatSimd128Load()) { + vbroadcastw(srcAddr, out.fpu()); + } else { + movzwl(srcAddr, out.gpr()); + } + break; + case Scalar::Int32: + case Scalar::Uint32: + movl(srcAddr, out.gpr()); + break; + case Scalar::Float32: + if (access.isSplatSimd128Load()) { + vbroadcastss(srcAddr, out.fpu()); + } else { + // vmovss does the right thing also for access.isZeroExtendSimd128Load() + vmovss(srcAddr, out.fpu()); + } + break; + case Scalar::Float64: + if (access.isSplatSimd128Load()) { + vmovddup(srcAddr, out.fpu()); + } else if (access.isWidenSimd128Load()) { + switch (access.widenSimdOp()) { + case wasm::SimdOp::V128Load8x8S: + vpmovsxbw(srcAddr, out.fpu()); + break; + case wasm::SimdOp::V128Load8x8U: + vpmovzxbw(srcAddr, out.fpu()); + break; + case wasm::SimdOp::V128Load16x4S: + vpmovsxwd(srcAddr, out.fpu()); + break; + case wasm::SimdOp::V128Load16x4U: + vpmovzxwd(srcAddr, out.fpu()); + break; + case wasm::SimdOp::V128Load32x2S: + vpmovsxdq(srcAddr, out.fpu()); + break; + case wasm::SimdOp::V128Load32x2U: + vpmovzxdq(srcAddr, out.fpu()); + break; + default: + MOZ_CRASH("Unexpected widening op for wasmLoad"); + } + } else { + // vmovsd does the right thing also for access.isZeroExtendSimd128Load() + vmovsd(srcAddr, out.fpu()); + } + break; + case Scalar::Simd128: + MacroAssemblerX64::loadUnalignedSimd128(srcAddr, out.fpu()); + break; + case Scalar::Int64: + MOZ_CRASH("int64 loads must use load64"); + case Scalar::BigInt64: + case Scalar::BigUint64: + case Scalar::Uint8Clamped: + case Scalar::MaxTypedArrayViewType: + MOZ_CRASH("unexpected scalar type for wasmLoad"); + } + + memoryBarrierAfter(access.sync()); +} + +void MacroAssembler::wasmLoadI64(const wasm::MemoryAccessDesc& access, + Operand srcAddr, Register64 out) { + // NOTE: the generated code must match the assembly code in gen_load in + // GenerateAtomicOperations.py + memoryBarrierBefore(access.sync()); + + append(access, size()); + switch (access.type()) { + case Scalar::Int8: + movsbq(srcAddr, out.reg); + break; + case Scalar::Uint8: + movzbq(srcAddr, out.reg); + break; + case Scalar::Int16: + movswq(srcAddr, out.reg); + break; + case Scalar::Uint16: + movzwq(srcAddr, out.reg); + break; + case Scalar::Int32: + movslq(srcAddr, out.reg); + break; + // Int32 to int64 moves zero-extend by default. + case Scalar::Uint32: + movl(srcAddr, out.reg); + break; + case Scalar::Int64: + movq(srcAddr, out.reg); + break; + case Scalar::Float32: + case Scalar::Float64: + case Scalar::Simd128: + MOZ_CRASH("float loads must use wasmLoad"); + case Scalar::Uint8Clamped: + case Scalar::BigInt64: + case Scalar::BigUint64: + case Scalar::MaxTypedArrayViewType: + MOZ_CRASH("unexpected scalar type for wasmLoadI64"); + } + + memoryBarrierAfter(access.sync()); +} + +void MacroAssembler::wasmStore(const wasm::MemoryAccessDesc& access, + AnyRegister value, Operand dstAddr) { + // NOTE: the generated code must match the assembly code in gen_store in + // GenerateAtomicOperations.py + memoryBarrierBefore(access.sync()); + + append(access, masm.size()); + switch (access.type()) { + case Scalar::Int8: + case Scalar::Uint8: + movb(value.gpr(), dstAddr); + break; + case Scalar::Int16: + case Scalar::Uint16: + movw(value.gpr(), dstAddr); + break; + case Scalar::Int32: + case Scalar::Uint32: + movl(value.gpr(), dstAddr); + break; + case Scalar::Int64: + movq(value.gpr(), dstAddr); + break; + case Scalar::Float32: + storeUncanonicalizedFloat32(value.fpu(), dstAddr); + break; + case Scalar::Float64: + storeUncanonicalizedDouble(value.fpu(), dstAddr); + break; + case Scalar::Simd128: + MacroAssemblerX64::storeUnalignedSimd128(value.fpu(), dstAddr); + break; + case Scalar::Uint8Clamped: + case Scalar::BigInt64: + case Scalar::BigUint64: + case Scalar::MaxTypedArrayViewType: + MOZ_CRASH("unexpected array type"); + } + + memoryBarrierAfter(access.sync()); +} + +void MacroAssembler::wasmTruncateDoubleToUInt32(FloatRegister input, + Register output, + bool isSaturating, + Label* oolEntry) { + vcvttsd2sq(input, output); + + // Check that the result is in the uint32_t range. + ScratchRegisterScope scratch(*this); + move32(Imm32(0xffffffff), scratch); + cmpq(scratch, output); + j(Assembler::Above, oolEntry); +} + +void MacroAssembler::wasmTruncateFloat32ToUInt32(FloatRegister input, + Register output, + bool isSaturating, + Label* oolEntry) { + vcvttss2sq(input, output); + + // Check that the result is in the uint32_t range. + ScratchRegisterScope scratch(*this); + move32(Imm32(0xffffffff), scratch); + cmpq(scratch, output); + j(Assembler::Above, oolEntry); +} + +void MacroAssembler::wasmTruncateDoubleToInt64( + FloatRegister input, Register64 output, bool isSaturating, Label* oolEntry, + Label* oolRejoin, FloatRegister tempReg) { + vcvttsd2sq(input, output.reg); + cmpq(Imm32(1), output.reg); + j(Assembler::Overflow, oolEntry); + bind(oolRejoin); +} + +void MacroAssembler::wasmTruncateFloat32ToInt64( + FloatRegister input, Register64 output, bool isSaturating, Label* oolEntry, + Label* oolRejoin, FloatRegister tempReg) { + vcvttss2sq(input, output.reg); + cmpq(Imm32(1), output.reg); + j(Assembler::Overflow, oolEntry); + bind(oolRejoin); +} + +void MacroAssembler::wasmTruncateDoubleToUInt64( + FloatRegister input, Register64 output, bool isSaturating, Label* oolEntry, + Label* oolRejoin, FloatRegister tempReg) { + // If the input < INT64_MAX, vcvttsd2sq will do the right thing, so + // we use it directly. Else, we subtract INT64_MAX, convert to int64, + // and then add INT64_MAX to the result. + + Label isLarge; + + ScratchDoubleScope scratch(*this); + loadConstantDouble(double(0x8000000000000000), scratch); + branchDouble(Assembler::DoubleGreaterThanOrEqual, input, scratch, &isLarge); + vcvttsd2sq(input, output.reg); + testq(output.reg, output.reg); + j(Assembler::Signed, oolEntry); + jump(oolRejoin); + + bind(&isLarge); + + moveDouble(input, tempReg); + vsubsd(scratch, tempReg, tempReg); + vcvttsd2sq(tempReg, output.reg); + testq(output.reg, output.reg); + j(Assembler::Signed, oolEntry); + or64(Imm64(0x8000000000000000), output); + + bind(oolRejoin); +} + +void MacroAssembler::wasmTruncateFloat32ToUInt64( + FloatRegister input, Register64 output, bool isSaturating, Label* oolEntry, + Label* oolRejoin, FloatRegister tempReg) { + // If the input < INT64_MAX, vcvttss2sq will do the right thing, so + // we use it directly. Else, we subtract INT64_MAX, convert to int64, + // and then add INT64_MAX to the result. + + Label isLarge; + + ScratchFloat32Scope scratch(*this); + loadConstantFloat32(float(0x8000000000000000), scratch); + branchFloat(Assembler::DoubleGreaterThanOrEqual, input, scratch, &isLarge); + vcvttss2sq(input, output.reg); + testq(output.reg, output.reg); + j(Assembler::Signed, oolEntry); + jump(oolRejoin); + + bind(&isLarge); + + moveFloat32(input, tempReg); + vsubss(scratch, tempReg, tempReg); + vcvttss2sq(tempReg, output.reg); + testq(output.reg, output.reg); + j(Assembler::Signed, oolEntry); + or64(Imm64(0x8000000000000000), output); + + bind(oolRejoin); +} + +void MacroAssembler::widenInt32(Register r) { + move32To64ZeroExtend(r, Register64(r)); +} + +// ======================================================================== +// Convert floating point. + +void MacroAssembler::convertInt64ToDouble(Register64 input, + FloatRegister output) { + // Zero the output register to break dependencies, see convertInt32ToDouble. + zeroDouble(output); + + vcvtsq2sd(input.reg, output, output); +} + +void MacroAssembler::convertInt64ToFloat32(Register64 input, + FloatRegister output) { + // Zero the output register to break dependencies, see convertInt32ToDouble. + zeroFloat32(output); + + vcvtsq2ss(input.reg, output, output); +} + +bool MacroAssembler::convertUInt64ToDoubleNeedsTemp() { return true; } + +void MacroAssembler::convertUInt64ToDouble(Register64 input, + FloatRegister output, + Register temp) { + // Zero the output register to break dependencies, see convertInt32ToDouble. + zeroDouble(output); + + // If the input's sign bit is not set we use vcvtsq2sd directly. + // Else, we divide by 2 and keep the LSB, convert to double, and multiply + // the result by 2. + Label done; + Label isSigned; + + testq(input.reg, input.reg); + j(Assembler::Signed, &isSigned); + vcvtsq2sd(input.reg, output, output); + jump(&done); + + bind(&isSigned); + + ScratchRegisterScope scratch(*this); + mov(input.reg, scratch); + mov(input.reg, temp); + shrq(Imm32(1), scratch); + andq(Imm32(1), temp); + orq(temp, scratch); + + vcvtsq2sd(scratch, output, output); + vaddsd(output, output, output); + + bind(&done); +} + +void MacroAssembler::convertUInt64ToFloat32(Register64 input, + FloatRegister output, + Register temp) { + // Zero the output register to break dependencies, see convertInt32ToDouble. + zeroFloat32(output); + + // See comment in convertUInt64ToDouble. + Label done; + Label isSigned; + + testq(input.reg, input.reg); + j(Assembler::Signed, &isSigned); + vcvtsq2ss(input.reg, output, output); + jump(&done); + + bind(&isSigned); + + ScratchRegisterScope scratch(*this); + mov(input.reg, scratch); + mov(input.reg, temp); + shrq(Imm32(1), scratch); + andq(Imm32(1), temp); + orq(temp, scratch); + + vcvtsq2ss(scratch, output, output); + vaddss(output, output, output); + + bind(&done); +} + +void MacroAssembler::convertIntPtrToDouble(Register src, FloatRegister dest) { + convertInt64ToDouble(Register64(src), dest); +} + +// ======================================================================== +// Primitive atomic operations. + +void MacroAssembler::wasmCompareExchange64(const wasm::MemoryAccessDesc& access, + const Address& mem, + Register64 expected, + Register64 replacement, + Register64 output) { + MOZ_ASSERT(output.reg == rax); + if (expected != output) { + movq(expected.reg, output.reg); + } + append(access, size()); + lock_cmpxchgq(replacement.reg, Operand(mem)); +} + +void MacroAssembler::wasmCompareExchange64(const wasm::MemoryAccessDesc& access, + const BaseIndex& mem, + Register64 expected, + Register64 replacement, + Register64 output) { + MOZ_ASSERT(output.reg == rax); + if (expected != output) { + movq(expected.reg, output.reg); + } + append(access, size()); + lock_cmpxchgq(replacement.reg, Operand(mem)); +} + +void MacroAssembler::wasmAtomicExchange64(const wasm::MemoryAccessDesc& access, + const Address& mem, Register64 value, + Register64 output) { + if (value != output) { + movq(value.reg, output.reg); + } + append(access, masm.size()); + xchgq(output.reg, Operand(mem)); +} + +void MacroAssembler::wasmAtomicExchange64(const wasm::MemoryAccessDesc& access, + const BaseIndex& mem, + Register64 value, Register64 output) { + if (value != output) { + movq(value.reg, output.reg); + } + append(access, masm.size()); + xchgq(output.reg, Operand(mem)); +} + +template <typename T> +static void AtomicFetchOp64(MacroAssembler& masm, + const wasm::MemoryAccessDesc* access, AtomicOp op, + Register value, const T& mem, Register temp, + Register output) { + // NOTE: the generated code must match the assembly code in gen_fetchop in + // GenerateAtomicOperations.py + if (op == AtomicFetchAddOp) { + if (value != output) { + masm.movq(value, output); + } + if (access) { + masm.append(*access, masm.size()); + } + masm.lock_xaddq(output, Operand(mem)); + } else if (op == AtomicFetchSubOp) { + if (value != output) { + masm.movq(value, output); + } + masm.negq(output); + if (access) { + masm.append(*access, masm.size()); + } + masm.lock_xaddq(output, Operand(mem)); + } else { + Label again; + MOZ_ASSERT(output == rax); + MOZ_ASSERT(value != output); + MOZ_ASSERT(value != temp); + MOZ_ASSERT(temp != output); + if (access) { + masm.append(*access, masm.size()); + } + masm.movq(Operand(mem), rax); + masm.bind(&again); + masm.movq(rax, temp); + switch (op) { + case AtomicFetchAndOp: + masm.andq(value, temp); + break; + case AtomicFetchOrOp: + masm.orq(value, temp); + break; + case AtomicFetchXorOp: + masm.xorq(value, temp); + break; + default: + MOZ_CRASH(); + } + masm.lock_cmpxchgq(temp, Operand(mem)); + masm.j(MacroAssembler::NonZero, &again); + } +} + +void MacroAssembler::wasmAtomicFetchOp64(const wasm::MemoryAccessDesc& access, + AtomicOp op, Register64 value, + const Address& mem, Register64 temp, + Register64 output) { + AtomicFetchOp64(*this, &access, op, value.reg, mem, temp.reg, output.reg); +} + +void MacroAssembler::wasmAtomicFetchOp64(const wasm::MemoryAccessDesc& access, + AtomicOp op, Register64 value, + const BaseIndex& mem, Register64 temp, + Register64 output) { + AtomicFetchOp64(*this, &access, op, value.reg, mem, temp.reg, output.reg); +} + +template <typename T> +static void AtomicEffectOp64(MacroAssembler& masm, + const wasm::MemoryAccessDesc* access, AtomicOp op, + Register value, const T& mem) { + if (access) { + masm.append(*access, masm.size()); + } + switch (op) { + case AtomicFetchAddOp: + masm.lock_addq(value, Operand(mem)); + break; + case AtomicFetchSubOp: + masm.lock_subq(value, Operand(mem)); + break; + case AtomicFetchAndOp: + masm.lock_andq(value, Operand(mem)); + break; + case AtomicFetchOrOp: + masm.lock_orq(value, Operand(mem)); + break; + case AtomicFetchXorOp: + masm.lock_xorq(value, Operand(mem)); + break; + default: + MOZ_CRASH(); + } +} + +void MacroAssembler::wasmAtomicEffectOp64(const wasm::MemoryAccessDesc& access, + AtomicOp op, Register64 value, + const BaseIndex& mem) { + AtomicEffectOp64(*this, &access, op, value.reg, mem); +} + +void MacroAssembler::compareExchange64(const Synchronization&, + const Address& mem, Register64 expected, + Register64 replacement, + Register64 output) { + // NOTE: the generated code must match the assembly code in gen_cmpxchg in + // GenerateAtomicOperations.py + MOZ_ASSERT(output.reg == rax); + if (expected != output) { + movq(expected.reg, output.reg); + } + lock_cmpxchgq(replacement.reg, Operand(mem)); +} + +void MacroAssembler::compareExchange64(const Synchronization&, + const BaseIndex& mem, + Register64 expected, + Register64 replacement, + Register64 output) { + MOZ_ASSERT(output.reg == rax); + if (expected != output) { + movq(expected.reg, output.reg); + } + lock_cmpxchgq(replacement.reg, Operand(mem)); +} + +void MacroAssembler::atomicExchange64(const Synchronization&, + const Address& mem, Register64 value, + Register64 output) { + // NOTE: the generated code must match the assembly code in gen_exchange in + // GenerateAtomicOperations.py + if (value != output) { + movq(value.reg, output.reg); + } + xchgq(output.reg, Operand(mem)); +} + +void MacroAssembler::atomicExchange64(const Synchronization&, + const BaseIndex& mem, Register64 value, + Register64 output) { + if (value != output) { + movq(value.reg, output.reg); + } + xchgq(output.reg, Operand(mem)); +} + +void MacroAssembler::atomicFetchOp64(const Synchronization& sync, AtomicOp op, + Register64 value, const Address& mem, + Register64 temp, Register64 output) { + AtomicFetchOp64(*this, nullptr, op, value.reg, mem, temp.reg, output.reg); +} + +void MacroAssembler::atomicFetchOp64(const Synchronization& sync, AtomicOp op, + Register64 value, const BaseIndex& mem, + Register64 temp, Register64 output) { + AtomicFetchOp64(*this, nullptr, op, value.reg, mem, temp.reg, output.reg); +} + +void MacroAssembler::atomicEffectOp64(const Synchronization& sync, AtomicOp op, + Register64 value, const Address& mem) { + AtomicEffectOp64(*this, nullptr, op, value.reg, mem); +} + +void MacroAssembler::atomicEffectOp64(const Synchronization& sync, AtomicOp op, + Register64 value, const BaseIndex& mem) { + AtomicEffectOp64(*this, nullptr, op, value.reg, mem); +} + +CodeOffset MacroAssembler::moveNearAddressWithPatch(Register dest) { + return leaRipRelative(dest); +} + +void MacroAssembler::patchNearAddressMove(CodeLocationLabel loc, + CodeLocationLabel target) { + ptrdiff_t off = target - loc; + MOZ_ASSERT(off > ptrdiff_t(INT32_MIN)); + MOZ_ASSERT(off < ptrdiff_t(INT32_MAX)); + PatchWrite_Imm32(loc, Imm32(off)); +} + +void MacroAssembler::wasmBoundsCheck64(Condition cond, Register64 index, + Register64 boundsCheckLimit, Label* ok) { + cmpPtr(index.reg, boundsCheckLimit.reg); + j(cond, ok); + if (JitOptions.spectreIndexMasking) { + cmovCCq(cond, Operand(boundsCheckLimit.reg), index.reg); + } +} + +void MacroAssembler::wasmBoundsCheck64(Condition cond, Register64 index, + Address boundsCheckLimit, Label* ok) { + cmpPtr(index.reg, Operand(boundsCheckLimit)); + j(cond, ok); + if (JitOptions.spectreIndexMasking) { + cmovCCq(cond, Operand(boundsCheckLimit), index.reg); + } +} + +// ======================================================================== +// Integer compare-then-conditionally-load/move operations. + +// cmpMove, Cond-Reg-Reg-Reg-Reg cases + +template <size_t CmpSize, size_t MoveSize> +void MacroAssemblerX64::cmpMove(Condition cond, Register lhs, Register rhs, + Register falseVal, Register trueValAndDest) { + if constexpr (CmpSize == 32) { + cmp32(lhs, rhs); + } else { + static_assert(CmpSize == 64); + cmpPtr(lhs, rhs); + } + if constexpr (MoveSize == 32) { + cmovCCl(cond, Operand(falseVal), trueValAndDest); + } else { + static_assert(MoveSize == 64); + cmovCCq(cond, Operand(falseVal), trueValAndDest); + } +} +template void MacroAssemblerX64::cmpMove<32, 32>(Condition cond, Register lhs, + Register rhs, + Register falseVal, + Register trueValAndDest); +template void MacroAssemblerX64::cmpMove<32, 64>(Condition cond, Register lhs, + Register rhs, + Register falseVal, + Register trueValAndDest); +template void MacroAssemblerX64::cmpMove<64, 32>(Condition cond, Register lhs, + Register rhs, + Register falseVal, + Register trueValAndDest); +template void MacroAssemblerX64::cmpMove<64, 64>(Condition cond, Register lhs, + Register rhs, + Register falseVal, + Register trueValAndDest); + +// cmpMove, Cond-Reg-Addr-Reg-Reg cases + +template <size_t CmpSize, size_t MoveSize> +void MacroAssemblerX64::cmpMove(Condition cond, Register lhs, + const Address& rhs, Register falseVal, + Register trueValAndDest) { + if constexpr (CmpSize == 32) { + cmp32(lhs, Operand(rhs)); + } else { + static_assert(CmpSize == 64); + cmpPtr(lhs, Operand(rhs)); + } + if constexpr (MoveSize == 32) { + cmovCCl(cond, Operand(falseVal), trueValAndDest); + } else { + static_assert(MoveSize == 64); + cmovCCq(cond, Operand(falseVal), trueValAndDest); + } +} +template void MacroAssemblerX64::cmpMove<32, 32>(Condition cond, Register lhs, + const Address& rhs, + Register falseVal, + Register trueValAndDest); +template void MacroAssemblerX64::cmpMove<32, 64>(Condition cond, Register lhs, + const Address& rhs, + Register falseVal, + Register trueValAndDest); +template void MacroAssemblerX64::cmpMove<64, 32>(Condition cond, Register lhs, + const Address& rhs, + Register falseVal, + Register trueValAndDest); +template void MacroAssemblerX64::cmpMove<64, 64>(Condition cond, Register lhs, + const Address& rhs, + Register falseVal, + Register trueValAndDest); + +// cmpLoad, Cond-Reg-Reg-Addr-Reg cases + +template <size_t CmpSize, size_t LoadSize> +void MacroAssemblerX64::cmpLoad(Condition cond, Register lhs, Register rhs, + const Address& falseVal, + Register trueValAndDest) { + if constexpr (CmpSize == 32) { + cmp32(lhs, rhs); + } else { + static_assert(CmpSize == 64); + cmpPtr(lhs, rhs); + } + if constexpr (LoadSize == 32) { + cmovCCl(cond, Operand(falseVal), trueValAndDest); + } else { + static_assert(LoadSize == 64); + cmovCCq(cond, Operand(falseVal), trueValAndDest); + } +} +template void MacroAssemblerX64::cmpLoad<32, 32>(Condition cond, Register lhs, + Register rhs, + const Address& falseVal, + Register trueValAndDest); +template void MacroAssemblerX64::cmpLoad<32, 64>(Condition cond, Register lhs, + Register rhs, + const Address& falseVal, + Register trueValAndDest); +template void MacroAssemblerX64::cmpLoad<64, 32>(Condition cond, Register lhs, + Register rhs, + const Address& falseVal, + Register trueValAndDest); +template void MacroAssemblerX64::cmpLoad<64, 64>(Condition cond, Register lhs, + Register rhs, + const Address& falseVal, + Register trueValAndDest); + +// cmpLoad, Cond-Reg-Addr-Addr-Reg cases + +template <size_t CmpSize, size_t LoadSize> +void MacroAssemblerX64::cmpLoad(Condition cond, Register lhs, + const Address& rhs, const Address& falseVal, + Register trueValAndDest) { + if constexpr (CmpSize == 32) { + cmp32(lhs, Operand(rhs)); + } else { + static_assert(CmpSize == 64); + cmpPtr(lhs, Operand(rhs)); + } + if constexpr (LoadSize == 32) { + cmovCCl(cond, Operand(falseVal), trueValAndDest); + } else { + static_assert(LoadSize == 64); + cmovCCq(cond, Operand(falseVal), trueValAndDest); + } +} +template void MacroAssemblerX64::cmpLoad<32, 32>(Condition cond, Register lhs, + const Address& rhs, + const Address& falseVal, + Register trueValAndDest); +template void MacroAssemblerX64::cmpLoad<32, 64>(Condition cond, Register lhs, + const Address& rhs, + const Address& falseVal, + Register trueValAndDest); +template void MacroAssemblerX64::cmpLoad<64, 32>(Condition cond, Register lhs, + const Address& rhs, + const Address& falseVal, + Register trueValAndDest); +template void MacroAssemblerX64::cmpLoad<64, 64>(Condition cond, Register lhs, + const Address& rhs, + const Address& falseVal, + Register trueValAndDest); + +//}}} check_macroassembler_style diff --git a/js/src/jit/x64/MacroAssembler-x64.h b/js/src/jit/x64/MacroAssembler-x64.h new file mode 100644 index 0000000000..80e2dfed28 --- /dev/null +++ b/js/src/jit/x64/MacroAssembler-x64.h @@ -0,0 +1,1218 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#ifndef jit_x64_MacroAssembler_x64_h +#define jit_x64_MacroAssembler_x64_h + +#include "jit/x86-shared/MacroAssembler-x86-shared.h" +#include "js/HeapAPI.h" +#include "wasm/WasmBuiltins.h" + +namespace js { +namespace jit { + +struct ImmShiftedTag : public ImmWord { + explicit ImmShiftedTag(JSValueShiftedTag shtag) : ImmWord((uintptr_t)shtag) {} + + explicit ImmShiftedTag(JSValueType type) + : ImmWord(uintptr_t(JSVAL_TYPE_TO_SHIFTED_TAG(type))) {} +}; + +struct ImmTag : public Imm32 { + explicit ImmTag(JSValueTag tag) : Imm32(tag) {} +}; + +// ScratchTagScope and ScratchTagScopeRelease are used to manage the tag +// register for splitTagForTest(), which has different register management on +// different platforms. On 64-bit platforms it requires a scratch register that +// does not interfere with other operations; on 32-bit platforms it uses a +// register that is already part of the Value. +// +// The ScratchTagScope RAII type acquires the appropriate register; a reference +// to a variable of this type is then passed to splitTagForTest(). +// +// On 64-bit platforms ScratchTagScopeRelease makes the owned scratch register +// available in a dynamic scope during compilation. However it is important to +// remember that that does not preserve the register value in any way, so this +// RAII type should only be used along paths that eventually branch past further +// uses of the extracted tag value. +// +// On 32-bit platforms ScratchTagScopeRelease has no effect, since it does not +// manage a register, it only aliases a register in the ValueOperand. + +class ScratchTagScope : public ScratchRegisterScope { + public: + ScratchTagScope(MacroAssembler& masm, const ValueOperand&) + : ScratchRegisterScope(masm) {} +}; + +class ScratchTagScopeRelease { + ScratchTagScope* ts_; + + public: + explicit ScratchTagScopeRelease(ScratchTagScope* ts) : ts_(ts) { + ts_->release(); + } + ~ScratchTagScopeRelease() { ts_->reacquire(); } +}; + +class MacroAssemblerX64 : public MacroAssemblerX86Shared { + private: + // Perform a downcast. Should be removed by Bug 996602. + MacroAssembler& asMasm(); + const MacroAssembler& asMasm() const; + + void bindOffsets(const MacroAssemblerX86Shared::UsesVector&); + + void vpRiprOpSimd128(const SimdConstant& v, FloatRegister reg, + JmpSrc (X86Encoding::BaseAssemblerX64::*op)( + X86Encoding::XMMRegisterID id)); + + void vpRiprOpSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest, + JmpSrc (X86Encoding::BaseAssemblerX64::*op)( + X86Encoding::XMMRegisterID srcId, + X86Encoding::XMMRegisterID destId)); + + public: + using MacroAssemblerX86Shared::load32; + using MacroAssemblerX86Shared::store16; + using MacroAssemblerX86Shared::store32; + + MacroAssemblerX64() = default; + + // The buffer is about to be linked, make sure any constant pools or excess + // bookkeeping has been flushed to the instruction stream. + void finish(); + + ///////////////////////////////////////////////////////////////// + // X64 helpers. + ///////////////////////////////////////////////////////////////// + void writeDataRelocation(const Value& val) { + // Raw GC pointer relocations and Value relocations both end up in + // Assembler::TraceDataRelocations. + if (val.isGCThing()) { + gc::Cell* cell = val.toGCThing(); + if (cell && gc::IsInsideNursery(cell)) { + embedsNurseryPointers_ = true; + } + dataRelocations_.writeUnsigned(masm.currentOffset()); + } + } + + // Refers to the upper 32 bits of a 64-bit Value operand. + // On x86_64, the upper 32 bits do not necessarily only contain the type. + Operand ToUpper32(Operand base) { + switch (base.kind()) { + case Operand::MEM_REG_DISP: + return Operand(Register::FromCode(base.base()), base.disp() + 4); + + case Operand::MEM_SCALE: + return Operand(Register::FromCode(base.base()), + Register::FromCode(base.index()), base.scale(), + base.disp() + 4); + + default: + MOZ_CRASH("unexpected operand kind"); + } + } + static inline Operand ToUpper32(const Address& address) { + return Operand(address.base, address.offset + 4); + } + static inline Operand ToUpper32(const BaseIndex& address) { + return Operand(address.base, address.index, address.scale, + address.offset + 4); + } + + uint32_t Upper32Of(JSValueShiftedTag tag) { return uint32_t(tag >> 32); } + + JSValueShiftedTag GetShiftedTag(JSValueType type) { + return (JSValueShiftedTag)JSVAL_TYPE_TO_SHIFTED_TAG(type); + } + + ///////////////////////////////////////////////////////////////// + // X86/X64-common interface. + ///////////////////////////////////////////////////////////////// + + void storeValue(ValueOperand val, Operand dest) { + movq(val.valueReg(), dest); + } + void storeValue(ValueOperand val, const Address& dest) { + storeValue(val, Operand(dest)); + } + template <typename T> + void storeValue(JSValueType type, Register reg, const T& dest) { + // Value types with 32-bit payloads can be emitted as two 32-bit moves. + if (type == JSVAL_TYPE_INT32 || type == JSVAL_TYPE_BOOLEAN) { + movl(reg, Operand(dest)); + movl(Imm32(Upper32Of(GetShiftedTag(type))), ToUpper32(Operand(dest))); + } else { + ScratchRegisterScope scratch(asMasm()); + boxValue(type, reg, scratch); + movq(scratch, Operand(dest)); + } + } + template <typename T> + void storeValue(const Value& val, const T& dest) { + ScratchRegisterScope scratch(asMasm()); + if (val.isGCThing()) { + movWithPatch(ImmWord(val.asRawBits()), scratch); + writeDataRelocation(val); + } else { + mov(ImmWord(val.asRawBits()), scratch); + } + movq(scratch, Operand(dest)); + } + void storeValue(ValueOperand val, BaseIndex dest) { + storeValue(val, Operand(dest)); + } + void storeValue(const Address& src, const Address& dest, Register temp) { + loadPtr(src, temp); + storePtr(temp, dest); + } + void storePrivateValue(Register src, const Address& dest) { + storePtr(src, dest); + } + void storePrivateValue(ImmGCPtr imm, const Address& dest) { + storePtr(imm, dest); + } + void loadValue(Operand src, ValueOperand val) { movq(src, val.valueReg()); } + void loadValue(Address src, ValueOperand val) { + loadValue(Operand(src), val); + } + void loadValue(const BaseIndex& src, ValueOperand val) { + loadValue(Operand(src), val); + } + void loadUnalignedValue(const Address& src, ValueOperand dest) { + loadValue(src, dest); + } + void tagValue(JSValueType type, Register payload, ValueOperand dest) { + ScratchRegisterScope scratch(asMasm()); + MOZ_ASSERT(dest.valueReg() != scratch); + if (payload != dest.valueReg()) { + movq(payload, dest.valueReg()); + } + mov(ImmShiftedTag(type), scratch); + orq(scratch, dest.valueReg()); + } + void pushValue(ValueOperand val) { push(val.valueReg()); } + void popValue(ValueOperand val) { pop(val.valueReg()); } + void pushValue(const Value& val) { + if (val.isGCThing()) { + ScratchRegisterScope scratch(asMasm()); + movWithPatch(ImmWord(val.asRawBits()), scratch); + writeDataRelocation(val); + push(scratch); + } else { + push(ImmWord(val.asRawBits())); + } + } + void pushValue(JSValueType type, Register reg) { + ScratchRegisterScope scratch(asMasm()); + boxValue(type, reg, scratch); + push(scratch); + } + void pushValue(const Address& addr) { push(Operand(addr)); } + + void pushValue(const BaseIndex& addr, Register scratch) { + push(Operand(addr)); + } + + void boxValue(JSValueType type, Register src, Register dest); + + Condition testUndefined(Condition cond, Register tag) { + MOZ_ASSERT(cond == Equal || cond == NotEqual); + cmp32(tag, ImmTag(JSVAL_TAG_UNDEFINED)); + return cond; + } + Condition testInt32(Condition cond, Register tag) { + MOZ_ASSERT(cond == Equal || cond == NotEqual); + cmp32(tag, ImmTag(JSVAL_TAG_INT32)); + return cond; + } + Condition testBoolean(Condition cond, Register tag) { + MOZ_ASSERT(cond == Equal || cond == NotEqual); + cmp32(tag, ImmTag(JSVAL_TAG_BOOLEAN)); + return cond; + } + Condition testNull(Condition cond, Register tag) { + MOZ_ASSERT(cond == Equal || cond == NotEqual); + cmp32(tag, ImmTag(JSVAL_TAG_NULL)); + return cond; + } + Condition testString(Condition cond, Register tag) { + MOZ_ASSERT(cond == Equal || cond == NotEqual); + cmp32(tag, ImmTag(JSVAL_TAG_STRING)); + return cond; + } + Condition testSymbol(Condition cond, Register tag) { + MOZ_ASSERT(cond == Equal || cond == NotEqual); + cmp32(tag, ImmTag(JSVAL_TAG_SYMBOL)); + return cond; + } + Condition testBigInt(Condition cond, Register tag) { + MOZ_ASSERT(cond == Equal || cond == NotEqual); + cmp32(tag, ImmTag(JSVAL_TAG_BIGINT)); + return cond; + } + Condition testObject(Condition cond, Register tag) { + MOZ_ASSERT(cond == Equal || cond == NotEqual); + cmp32(tag, ImmTag(JSVAL_TAG_OBJECT)); + return cond; + } + Condition testDouble(Condition cond, Register tag) { + MOZ_ASSERT(cond == Equal || cond == NotEqual); + cmp32(tag, Imm32(JSVAL_TAG_MAX_DOUBLE)); + return cond == Equal ? BelowOrEqual : Above; + } + Condition testNumber(Condition cond, Register tag) { + MOZ_ASSERT(cond == Equal || cond == NotEqual); + cmp32(tag, Imm32(JS::detail::ValueUpperInclNumberTag)); + return cond == Equal ? BelowOrEqual : Above; + } + Condition testGCThing(Condition cond, Register tag) { + MOZ_ASSERT(cond == Equal || cond == NotEqual); + cmp32(tag, Imm32(JS::detail::ValueLowerInclGCThingTag)); + return cond == Equal ? AboveOrEqual : Below; + } + + Condition testMagic(Condition cond, Register tag) { + MOZ_ASSERT(cond == Equal || cond == NotEqual); + cmp32(tag, ImmTag(JSVAL_TAG_MAGIC)); + return cond; + } + Condition testError(Condition cond, Register tag) { + return testMagic(cond, tag); + } + Condition testPrimitive(Condition cond, Register tag) { + MOZ_ASSERT(cond == Equal || cond == NotEqual); + cmp32(tag, ImmTag(JS::detail::ValueUpperExclPrimitiveTag)); + return cond == Equal ? Below : AboveOrEqual; + } + + Condition testUndefined(Condition cond, const ValueOperand& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testUndefined(cond, scratch); + } + Condition testInt32(Condition cond, const ValueOperand& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testInt32(cond, scratch); + } + Condition testBoolean(Condition cond, const ValueOperand& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testBoolean(cond, scratch); + } + Condition testDouble(Condition cond, const ValueOperand& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testDouble(cond, scratch); + } + Condition testNumber(Condition cond, const ValueOperand& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testNumber(cond, scratch); + } + Condition testNull(Condition cond, const ValueOperand& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testNull(cond, scratch); + } + Condition testString(Condition cond, const ValueOperand& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testString(cond, scratch); + } + Condition testSymbol(Condition cond, const ValueOperand& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testSymbol(cond, scratch); + } + Condition testBigInt(Condition cond, const ValueOperand& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testBigInt(cond, scratch); + } + Condition testObject(Condition cond, const ValueOperand& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testObject(cond, scratch); + } + Condition testGCThing(Condition cond, const ValueOperand& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testGCThing(cond, scratch); + } + Condition testPrimitive(Condition cond, const ValueOperand& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testPrimitive(cond, scratch); + } + + Condition testUndefined(Condition cond, const Address& src) { + cmp32(ToUpper32(src), + Imm32(Upper32Of(GetShiftedTag(JSVAL_TYPE_UNDEFINED)))); + return cond; + } + Condition testInt32(Condition cond, const Address& src) { + cmp32(ToUpper32(src), Imm32(Upper32Of(GetShiftedTag(JSVAL_TYPE_INT32)))); + return cond; + } + Condition testBoolean(Condition cond, const Address& src) { + cmp32(ToUpper32(src), Imm32(Upper32Of(GetShiftedTag(JSVAL_TYPE_BOOLEAN)))); + return cond; + } + Condition testDouble(Condition cond, const Address& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testDouble(cond, scratch); + } + Condition testNumber(Condition cond, const Address& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testNumber(cond, scratch); + } + Condition testNull(Condition cond, const Address& src) { + cmp32(ToUpper32(src), Imm32(Upper32Of(GetShiftedTag(JSVAL_TYPE_NULL)))); + return cond; + } + Condition testString(Condition cond, const Address& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testString(cond, scratch); + } + Condition testSymbol(Condition cond, const Address& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testSymbol(cond, scratch); + } + Condition testBigInt(Condition cond, const Address& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testBigInt(cond, scratch); + } + Condition testObject(Condition cond, const Address& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testObject(cond, scratch); + } + Condition testPrimitive(Condition cond, const Address& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testPrimitive(cond, scratch); + } + Condition testGCThing(Condition cond, const Address& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testGCThing(cond, scratch); + } + Condition testMagic(Condition cond, const Address& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testMagic(cond, scratch); + } + + Condition testUndefined(Condition cond, const BaseIndex& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testUndefined(cond, scratch); + } + Condition testNull(Condition cond, const BaseIndex& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testNull(cond, scratch); + } + Condition testBoolean(Condition cond, const BaseIndex& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testBoolean(cond, scratch); + } + Condition testString(Condition cond, const BaseIndex& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testString(cond, scratch); + } + Condition testSymbol(Condition cond, const BaseIndex& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testSymbol(cond, scratch); + } + Condition testBigInt(Condition cond, const BaseIndex& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testBigInt(cond, scratch); + } + Condition testInt32(Condition cond, const BaseIndex& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testInt32(cond, scratch); + } + Condition testObject(Condition cond, const BaseIndex& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testObject(cond, scratch); + } + Condition testDouble(Condition cond, const BaseIndex& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testDouble(cond, scratch); + } + Condition testMagic(Condition cond, const BaseIndex& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testMagic(cond, scratch); + } + Condition testGCThing(Condition cond, const BaseIndex& src) { + ScratchRegisterScope scratch(asMasm()); + splitTag(src, scratch); + return testGCThing(cond, scratch); + } + + Condition isMagic(Condition cond, const ValueOperand& src, JSWhyMagic why) { + uint64_t magic = MagicValue(why).asRawBits(); + cmpPtr(src.valueReg(), ImmWord(magic)); + return cond; + } + + void cmpPtr(Register lhs, const ImmWord rhs) { + ScratchRegisterScope scratch(asMasm()); + MOZ_ASSERT(lhs != scratch); + if (intptr_t(rhs.value) <= INT32_MAX && intptr_t(rhs.value) >= INT32_MIN) { + cmpPtr(lhs, Imm32(int32_t(rhs.value))); + } else { + movePtr(rhs, scratch); + cmpPtr(lhs, scratch); + } + } + void cmpPtr(Register lhs, const ImmPtr rhs) { + cmpPtr(lhs, ImmWord(uintptr_t(rhs.value))); + } + void cmpPtr(Register lhs, const ImmGCPtr rhs) { + ScratchRegisterScope scratch(asMasm()); + MOZ_ASSERT(lhs != scratch); + movePtr(rhs, scratch); + cmpPtr(lhs, scratch); + } + void cmpPtr(Register lhs, const Imm32 rhs) { cmpq(rhs, lhs); } + void cmpPtr(const Operand& lhs, const ImmGCPtr rhs) { + ScratchRegisterScope scratch(asMasm()); + MOZ_ASSERT(!lhs.containsReg(scratch)); + movePtr(rhs, scratch); + cmpPtr(lhs, scratch); + } + void cmpPtr(const Operand& lhs, const ImmWord rhs) { + if ((intptr_t)rhs.value <= INT32_MAX && (intptr_t)rhs.value >= INT32_MIN) { + cmpPtr(lhs, Imm32((int32_t)rhs.value)); + } else { + ScratchRegisterScope scratch(asMasm()); + movePtr(rhs, scratch); + cmpPtr(lhs, scratch); + } + } + void cmpPtr(const Operand& lhs, const ImmPtr rhs) { + cmpPtr(lhs, ImmWord(uintptr_t(rhs.value))); + } + void cmpPtr(const Address& lhs, const ImmGCPtr rhs) { + cmpPtr(Operand(lhs), rhs); + } + void cmpPtr(const Address& lhs, const ImmWord rhs) { + cmpPtr(Operand(lhs), rhs); + } + void cmpPtr(const Address& lhs, const ImmPtr rhs) { + cmpPtr(lhs, ImmWord(uintptr_t(rhs.value))); + } + void cmpPtr(const Operand& lhs, Register rhs) { cmpq(rhs, lhs); } + void cmpPtr(Register lhs, const Operand& rhs) { cmpq(rhs, lhs); } + void cmpPtr(const Operand& lhs, const Imm32 rhs) { cmpq(rhs, lhs); } + void cmpPtr(const Address& lhs, Register rhs) { cmpPtr(Operand(lhs), rhs); } + void cmpPtr(Register lhs, Register rhs) { cmpq(rhs, lhs); } + void testPtr(Register lhs, Register rhs) { testq(rhs, lhs); } + void testPtr(Register lhs, Imm32 rhs) { testq(rhs, lhs); } + void testPtr(const Operand& lhs, Imm32 rhs) { testq(rhs, lhs); } + void test64(Register lhs, Register rhs) { testq(rhs, lhs); } + void test64(Register lhs, const Imm64 rhs) { + if ((intptr_t)rhs.value <= INT32_MAX && (intptr_t)rhs.value >= INT32_MIN) { + testq(Imm32((int32_t)rhs.value), lhs); + } else { + ScratchRegisterScope scratch(asMasm()); + movq(ImmWord(rhs.value), scratch); + testq(scratch, lhs); + } + } + + // Compare-then-conditionally-move/load, for integer types + template <size_t CmpSize, size_t MoveSize> + void cmpMove(Condition cond, Register lhs, Register rhs, Register falseVal, + Register trueValAndDest); + + template <size_t CmpSize, size_t MoveSize> + void cmpMove(Condition cond, Register lhs, const Address& rhs, + Register falseVal, Register trueValAndDest); + + template <size_t CmpSize, size_t LoadSize> + void cmpLoad(Condition cond, Register lhs, Register rhs, + const Address& falseVal, Register trueValAndDest); + + template <size_t CmpSize, size_t LoadSize> + void cmpLoad(Condition cond, Register lhs, const Address& rhs, + const Address& falseVal, Register trueValAndDest); + + ///////////////////////////////////////////////////////////////// + // Common interface. + ///////////////////////////////////////////////////////////////// + + void movePtr(Register src, Register dest) { movq(src, dest); } + void movePtr(Register src, const Operand& dest) { movq(src, dest); } + void movePtr(ImmWord imm, Register dest) { mov(imm, dest); } + void movePtr(ImmPtr imm, Register dest) { mov(imm, dest); } + void movePtr(wasm::SymbolicAddress imm, Register dest) { mov(imm, dest); } + void movePtr(ImmGCPtr imm, Register dest) { movq(imm, dest); } + void loadPtr(AbsoluteAddress address, Register dest) { + if (X86Encoding::IsAddressImmediate(address.addr)) { + movq(Operand(address), dest); + } else { + ScratchRegisterScope scratch(asMasm()); + mov(ImmPtr(address.addr), scratch); + loadPtr(Address(scratch, 0x0), dest); + } + } + void loadPtr(const Address& address, Register dest) { + movq(Operand(address), dest); + } + void load64(const Address& address, Register dest) { + movq(Operand(address), dest); + } + void loadPtr(const Operand& src, Register dest) { movq(src, dest); } + void loadPtr(const BaseIndex& src, Register dest) { + movq(Operand(src), dest); + } + void loadPrivate(const Address& src, Register dest) { loadPtr(src, dest); } + void load32(AbsoluteAddress address, Register dest) { + if (X86Encoding::IsAddressImmediate(address.addr)) { + movl(Operand(address), dest); + } else { + ScratchRegisterScope scratch(asMasm()); + mov(ImmPtr(address.addr), scratch); + load32(Address(scratch, 0x0), dest); + } + } + void load64(const Operand& address, Register64 dest) { + movq(address, dest.reg); + } + void load64(const Address& address, Register64 dest) { + movq(Operand(address), dest.reg); + } + void load64(const BaseIndex& address, Register64 dest) { + movq(Operand(address), dest.reg); + } + template <typename S> + void load64Unaligned(const S& src, Register64 dest) { + load64(src, dest); + } + template <typename T> + void storePtr(ImmWord imm, T address) { + if ((intptr_t)imm.value <= INT32_MAX && (intptr_t)imm.value >= INT32_MIN) { + movq(Imm32((int32_t)imm.value), Operand(address)); + } else { + ScratchRegisterScope scratch(asMasm()); + mov(imm, scratch); + movq(scratch, Operand(address)); + } + } + template <typename T> + void storePtr(ImmPtr imm, T address) { + storePtr(ImmWord(uintptr_t(imm.value)), address); + } + template <typename T> + void storePtr(ImmGCPtr imm, T address) { + ScratchRegisterScope scratch(asMasm()); + movq(imm, scratch); + movq(scratch, Operand(address)); + } + void storePtr(Register src, const Address& address) { + movq(src, Operand(address)); + } + void store64(Register src, const Address& address) { + movq(src, Operand(address)); + } + void store64(Register64 src, const Operand& address) { + movq(src.reg, address); + } + void storePtr(Register src, const BaseIndex& address) { + movq(src, Operand(address)); + } + void storePtr(Register src, const Operand& dest) { movq(src, dest); } + void storePtr(Register src, AbsoluteAddress address) { + if (X86Encoding::IsAddressImmediate(address.addr)) { + movq(src, Operand(address)); + } else { + ScratchRegisterScope scratch(asMasm()); + mov(ImmPtr(address.addr), scratch); + storePtr(src, Address(scratch, 0x0)); + } + } + void store32(Register src, AbsoluteAddress address) { + if (X86Encoding::IsAddressImmediate(address.addr)) { + movl(src, Operand(address)); + } else { + ScratchRegisterScope scratch(asMasm()); + mov(ImmPtr(address.addr), scratch); + store32(src, Address(scratch, 0x0)); + } + } + void store16(Register src, AbsoluteAddress address) { + if (X86Encoding::IsAddressImmediate(address.addr)) { + movw(src, Operand(address)); + } else { + ScratchRegisterScope scratch(asMasm()); + mov(ImmPtr(address.addr), scratch); + store16(src, Address(scratch, 0x0)); + } + } + void store64(Register64 src, Address address) { storePtr(src.reg, address); } + void store64(Register64 src, const BaseIndex& address) { + storePtr(src.reg, address); + } + void store64(Imm64 imm, Address address) { + storePtr(ImmWord(imm.value), address); + } + void store64(Imm64 imm, const BaseIndex& address) { + storePtr(ImmWord(imm.value), address); + } + template <typename S, typename T> + void store64Unaligned(const S& src, const T& dest) { + store64(src, dest); + } + + void splitTag(Register src, Register dest) { + if (src != dest) { + movq(src, dest); + } + shrq(Imm32(JSVAL_TAG_SHIFT), dest); + } + void splitTag(const ValueOperand& operand, Register dest) { + splitTag(operand.valueReg(), dest); + } + void splitTag(const Operand& operand, Register dest) { + movq(operand, dest); + shrq(Imm32(JSVAL_TAG_SHIFT), dest); + } + void splitTag(const Address& operand, Register dest) { + splitTag(Operand(operand), dest); + } + void splitTag(const BaseIndex& operand, Register dest) { + splitTag(Operand(operand), dest); + } + + // Extracts the tag of a value and places it in tag. + void splitTagForTest(const ValueOperand& value, ScratchTagScope& tag) { + splitTag(value, tag); + } + void cmpTag(const ValueOperand& operand, ImmTag tag) { + ScratchTagScope reg(asMasm(), operand); + splitTagForTest(operand, reg); + cmp32(reg, tag); + } + + Condition testMagic(Condition cond, const ValueOperand& src) { + ScratchTagScope scratch(asMasm(), src); + splitTagForTest(src, scratch); + return testMagic(cond, scratch); + } + Condition testError(Condition cond, const ValueOperand& src) { + return testMagic(cond, src); + } + + void testNullSet(Condition cond, const ValueOperand& value, Register dest) { + cond = testNull(cond, value); + emitSet(cond, dest); + } + + void testObjectSet(Condition cond, const ValueOperand& value, Register dest) { + cond = testObject(cond, value); + emitSet(cond, dest); + } + + void testUndefinedSet(Condition cond, const ValueOperand& value, + Register dest) { + cond = testUndefined(cond, value); + emitSet(cond, dest); + } + + void boxDouble(FloatRegister src, const ValueOperand& dest, FloatRegister) { + vmovq(src, dest.valueReg()); + } + void boxNonDouble(JSValueType type, Register src, const ValueOperand& dest) { + MOZ_ASSERT(src != dest.valueReg()); + boxValue(type, src, dest.valueReg()); + } + + // Note that the |dest| register here may be ScratchReg, so we shouldn't + // use it. + void unboxInt32(const ValueOperand& src, Register dest) { + movl(src.valueReg(), dest); + } + void unboxInt32(const Operand& src, Register dest) { movl(src, dest); } + void unboxInt32(const Address& src, Register dest) { + unboxInt32(Operand(src), dest); + } + void unboxInt32(const BaseIndex& src, Register dest) { + unboxInt32(Operand(src), dest); + } + template <typename T> + void unboxDouble(const T& src, FloatRegister dest) { + loadDouble(Operand(src), dest); + } + + void unboxArgObjMagic(const ValueOperand& src, Register dest) { + unboxArgObjMagic(Operand(src.valueReg()), dest); + } + void unboxArgObjMagic(const Operand& src, Register dest) { + mov(ImmWord(0), dest); + } + void unboxArgObjMagic(const Address& src, Register dest) { + unboxArgObjMagic(Operand(src), dest); + } + + void unboxBoolean(const ValueOperand& src, Register dest) { + movl(src.valueReg(), dest); + } + void unboxBoolean(const Operand& src, Register dest) { movl(src, dest); } + void unboxBoolean(const Address& src, Register dest) { + unboxBoolean(Operand(src), dest); + } + void unboxBoolean(const BaseIndex& src, Register dest) { + unboxBoolean(Operand(src), dest); + } + + void unboxMagic(const ValueOperand& src, Register dest) { + movl(src.valueReg(), dest); + } + + void unboxDouble(const ValueOperand& src, FloatRegister dest) { + vmovq(src.valueReg(), dest); + } + + void notBoolean(const ValueOperand& val) { xorq(Imm32(1), val.valueReg()); } + + void unboxNonDouble(const ValueOperand& src, Register dest, + JSValueType type) { + MOZ_ASSERT(type != JSVAL_TYPE_DOUBLE); + if (type == JSVAL_TYPE_INT32 || type == JSVAL_TYPE_BOOLEAN) { + movl(src.valueReg(), dest); + return; + } + if (src.valueReg() == dest) { + ScratchRegisterScope scratch(asMasm()); + mov(ImmWord(JSVAL_TYPE_TO_SHIFTED_TAG(type)), scratch); + xorq(scratch, dest); + } else { + mov(ImmWord(JSVAL_TYPE_TO_SHIFTED_TAG(type)), dest); + xorq(src.valueReg(), dest); + } + } + void unboxNonDouble(const Operand& src, Register dest, JSValueType type) { + MOZ_ASSERT(type != JSVAL_TYPE_DOUBLE); + if (type == JSVAL_TYPE_INT32 || type == JSVAL_TYPE_BOOLEAN) { + movl(src, dest); + return; + } + // Explicitly permits |dest| to be used in |src|. + ScratchRegisterScope scratch(asMasm()); + MOZ_ASSERT(dest != scratch); + if (src.containsReg(dest)) { + mov(ImmWord(JSVAL_TYPE_TO_SHIFTED_TAG(type)), scratch); + // If src is already a register, then src and dest are the same + // thing and we don't need to move anything into dest. + if (src.kind() != Operand::REG) { + movq(src, dest); + } + xorq(scratch, dest); + } else { + mov(ImmWord(JSVAL_TYPE_TO_SHIFTED_TAG(type)), dest); + xorq(src, dest); + } + } + void unboxNonDouble(const Address& src, Register dest, JSValueType type) { + unboxNonDouble(Operand(src), dest, type); + } + void unboxNonDouble(const BaseIndex& src, Register dest, JSValueType type) { + unboxNonDouble(Operand(src), dest, type); + } + + void unboxString(const ValueOperand& src, Register dest) { + unboxNonDouble(src, dest, JSVAL_TYPE_STRING); + } + void unboxString(const Operand& src, Register dest) { + unboxNonDouble(src, dest, JSVAL_TYPE_STRING); + } + void unboxString(const Address& src, Register dest) { + unboxNonDouble(src, dest, JSVAL_TYPE_STRING); + } + + void unboxSymbol(const ValueOperand& src, Register dest) { + unboxNonDouble(src, dest, JSVAL_TYPE_SYMBOL); + } + void unboxSymbol(const Operand& src, Register dest) { + unboxNonDouble(src, dest, JSVAL_TYPE_SYMBOL); + } + + void unboxBigInt(const ValueOperand& src, Register dest) { + unboxNonDouble(src, dest, JSVAL_TYPE_BIGINT); + } + void unboxBigInt(const Operand& src, Register dest) { + unboxNonDouble(src, dest, JSVAL_TYPE_BIGINT); + } + void unboxBigInt(const Address& src, Register dest) { + unboxNonDouble(src, dest, JSVAL_TYPE_BIGINT); + } + + void unboxObject(const ValueOperand& src, Register dest) { + unboxNonDouble(src, dest, JSVAL_TYPE_OBJECT); + } + void unboxObject(const Operand& src, Register dest) { + unboxNonDouble(src, dest, JSVAL_TYPE_OBJECT); + } + void unboxObject(const Address& src, Register dest) { + unboxNonDouble(Operand(src), dest, JSVAL_TYPE_OBJECT); + } + void unboxObject(const BaseIndex& src, Register dest) { + unboxNonDouble(Operand(src), dest, JSVAL_TYPE_OBJECT); + } + + template <typename T> + void unboxObjectOrNull(const T& src, Register dest) { + unboxNonDouble(src, dest, JSVAL_TYPE_OBJECT); + ScratchRegisterScope scratch(asMasm()); + mov(ImmWord(~JS::detail::ValueObjectOrNullBit), scratch); + andq(scratch, dest); + } + + // This should only be used for GC barrier code, to unbox a GC thing Value. + // It's fine there because we don't depend on the actual Value type (all Cells + // are treated the same way). In almost all other cases this would be + // Spectre-unsafe - use unboxNonDouble and friends instead. + void unboxGCThingForGCBarrier(const Address& src, Register dest) { + movq(ImmWord(JS::detail::ValueGCThingPayloadMask), dest); + andq(Operand(src), dest); + } + void unboxGCThingForGCBarrier(const ValueOperand& src, Register dest) { + MOZ_ASSERT(src.valueReg() != dest); + movq(ImmWord(JS::detail::ValueGCThingPayloadMask), dest); + andq(src.valueReg(), dest); + } + + // Like unboxGCThingForGCBarrier, but loads the GC thing's chunk base. + void getGCThingValueChunk(const Address& src, Register dest) { + movq(ImmWord(JS::detail::ValueGCThingPayloadChunkMask), dest); + andq(Operand(src), dest); + } + void getGCThingValueChunk(const ValueOperand& src, Register dest) { + MOZ_ASSERT(src.valueReg() != dest); + movq(ImmWord(JS::detail::ValueGCThingPayloadChunkMask), dest); + andq(src.valueReg(), dest); + } + + inline void fallibleUnboxPtrImpl(const Operand& src, Register dest, + JSValueType type, Label* fail); + + // Extended unboxing API. If the payload is already in a register, returns + // that register. Otherwise, provides a move to the given scratch register, + // and returns that. + [[nodiscard]] Register extractObject(const Address& address, + Register scratch) { + MOZ_ASSERT(scratch != ScratchReg); + unboxObject(address, scratch); + return scratch; + } + [[nodiscard]] Register extractObject(const ValueOperand& value, + Register scratch) { + MOZ_ASSERT(scratch != ScratchReg); + unboxObject(value, scratch); + return scratch; + } + [[nodiscard]] Register extractSymbol(const ValueOperand& value, + Register scratch) { + MOZ_ASSERT(scratch != ScratchReg); + unboxSymbol(value, scratch); + return scratch; + } + [[nodiscard]] Register extractInt32(const ValueOperand& value, + Register scratch) { + MOZ_ASSERT(scratch != ScratchReg); + unboxInt32(value, scratch); + return scratch; + } + [[nodiscard]] Register extractBoolean(const ValueOperand& value, + Register scratch) { + MOZ_ASSERT(scratch != ScratchReg); + unboxBoolean(value, scratch); + return scratch; + } + [[nodiscard]] Register extractTag(const Address& address, Register scratch) { + MOZ_ASSERT(scratch != ScratchReg); + loadPtr(address, scratch); + splitTag(scratch, scratch); + return scratch; + } + [[nodiscard]] Register extractTag(const ValueOperand& value, + Register scratch) { + MOZ_ASSERT(scratch != ScratchReg); + splitTag(value, scratch); + return scratch; + } + + inline void unboxValue(const ValueOperand& src, AnyRegister dest, + JSValueType type); + + // These two functions use the low 32-bits of the full value register. + void boolValueToDouble(const ValueOperand& operand, FloatRegister dest) { + convertInt32ToDouble(operand.valueReg(), dest); + } + void int32ValueToDouble(const ValueOperand& operand, FloatRegister dest) { + convertInt32ToDouble(operand.valueReg(), dest); + } + + void boolValueToFloat32(const ValueOperand& operand, FloatRegister dest) { + convertInt32ToFloat32(operand.valueReg(), dest); + } + void int32ValueToFloat32(const ValueOperand& operand, FloatRegister dest) { + convertInt32ToFloat32(operand.valueReg(), dest); + } + + void loadConstantDouble(double d, FloatRegister dest); + void loadConstantFloat32(float f, FloatRegister dest); + + void loadConstantSimd128Int(const SimdConstant& v, FloatRegister dest); + void loadConstantSimd128Float(const SimdConstant& v, FloatRegister dest); + void vpaddbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpaddwSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpadddSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpaddqSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpsubbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpsubwSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpsubdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpsubqSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpmullwSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpmulldSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpaddsbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpaddusbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpaddswSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpadduswSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpsubsbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpsubusbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpsubswSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpsubuswSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpminsbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpminubSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpminswSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpminuwSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpminsdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpminudSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpmaxsbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpmaxubSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpmaxswSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpmaxuwSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpmaxsdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpmaxudSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpandSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpxorSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vporSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vaddpsSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vaddpdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vsubpsSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vsubpdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vdivpsSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vdivpdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vmulpsSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vmulpdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vandpdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vminpdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpacksswbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpackuswbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpackssdwSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpackusdwSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpunpckldqSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vunpcklpsSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpshufbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vptestSimd128(const SimdConstant& v, FloatRegister lhs); + void vpmaddwdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpcmpeqbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpcmpgtbSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpcmpeqwSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpcmpgtwSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpcmpeqdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpcmpgtdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vcmpeqpsSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vcmpneqpsSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vcmpltpsSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vcmplepsSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vcmpgepsSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vcmpeqpdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vcmpneqpdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vcmpltpdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vcmplepdSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpmaddubswSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + void vpmuludqSimd128(const SimdConstant& v, FloatRegister lhs, + FloatRegister dest); + + public: + Condition testInt32Truthy(bool truthy, const ValueOperand& operand) { + test32(operand.valueReg(), operand.valueReg()); + return truthy ? NonZero : Zero; + } + Condition testStringTruthy(bool truthy, const ValueOperand& value); + Condition testBigIntTruthy(bool truthy, const ValueOperand& value); + + template <typename T> + inline void loadInt32OrDouble(const T& src, FloatRegister dest); + + template <typename T> + void loadUnboxedValue(const T& src, MIRType type, AnyRegister dest) { + if (dest.isFloat()) { + loadInt32OrDouble(src, dest.fpu()); + } else { + unboxNonDouble(Operand(src), dest.gpr(), ValueTypeFromMIRType(type)); + } + } + + template <typename T> + void storeUnboxedPayload(ValueOperand value, T address, size_t nbytes, + JSValueType type) { + switch (nbytes) { + case 8: { + ScratchRegisterScope scratch(asMasm()); + unboxNonDouble(value, scratch, type); + storePtr(scratch, address); + if (type == JSVAL_TYPE_OBJECT) { + // Ideally we would call unboxObjectOrNull, but we need an extra + // scratch register for that. So unbox as object, then clear the + // object-or-null bit. + mov(ImmWord(~JS::detail::ValueObjectOrNullBit), scratch); + andq(scratch, Operand(address)); + } + return; + } + case 4: + store32(value.valueReg(), address); + return; + case 1: + store8(value.valueReg(), address); + return; + default: + MOZ_CRASH("Bad payload width"); + } + } + + // Checks whether a double is representable as a 64-bit integer. If so, the + // integer is written to the output register. Otherwise, a bailout is taken to + // the given snapshot. This function overwrites the scratch float register. + void convertDoubleToPtr(FloatRegister src, Register dest, Label* fail, + bool negativeZeroCheck = true); + + void convertUInt32ToDouble(Register src, FloatRegister dest) { + // Zero the output register to break dependencies, see convertInt32ToDouble. + zeroDouble(dest); + + vcvtsq2sd(src, dest, dest); + } + + void convertUInt32ToFloat32(Register src, FloatRegister dest) { + // Zero the output register to break dependencies, see convertInt32ToDouble. + zeroDouble(dest); + + vcvtsq2ss(src, dest, dest); + } + + inline void incrementInt32Value(const Address& addr); + + inline void ensureDouble(const ValueOperand& source, FloatRegister dest, + Label* failure); + + public: + void handleFailureWithHandlerTail(Label* profilerExitTail, + Label* bailoutTail); + + // Instrumentation for entering and leaving the profiler. + void profilerEnterFrame(Register framePtr, Register scratch); + void profilerExitFrame(); +}; + +using MacroAssemblerSpecific = MacroAssemblerX64; + +} // namespace jit +} // namespace js + +#endif /* jit_x64_MacroAssembler_x64_h */ diff --git a/js/src/jit/x64/SharedICHelpers-x64-inl.h b/js/src/jit/x64/SharedICHelpers-x64-inl.h new file mode 100644 index 0000000000..c40eb45c74 --- /dev/null +++ b/js/src/jit/x64/SharedICHelpers-x64-inl.h @@ -0,0 +1,80 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#ifndef jit_x64_SharedICHelpers_x64_inl_h +#define jit_x64_SharedICHelpers_x64_inl_h + +#include "jit/BaselineFrame.h" +#include "jit/SharedICHelpers.h" + +#include "jit/MacroAssembler-inl.h" + +namespace js { +namespace jit { + +inline void EmitBaselineTailCallVM(TrampolinePtr target, MacroAssembler& masm, + uint32_t argSize) { +#ifdef DEBUG + ScratchRegisterScope scratch(masm); + + // We can assume during this that R0 and R1 have been pushed. + // Store frame size without VMFunction arguments for debug assertions. + masm.movq(FramePointer, scratch); + masm.subq(StackPointer, scratch); + masm.subq(Imm32(argSize), scratch); + Address frameSizeAddr(FramePointer, + BaselineFrame::reverseOffsetOfDebugFrameSize()); + masm.store32(scratch, frameSizeAddr); +#endif + + // Push frame descriptor and perform the tail call. + masm.pushFrameDescriptor(FrameType::BaselineJS); + masm.push(ICTailCallReg); + masm.jump(target); +} + +inline void EmitBaselineCallVM(TrampolinePtr target, MacroAssembler& masm) { + masm.pushFrameDescriptor(FrameType::BaselineStub); + masm.call(target); +} + +inline void EmitBaselineEnterStubFrame(MacroAssembler& masm, Register) { +#ifdef DEBUG + // Compute frame size. Because the return address is still on the stack, + // this is: + // + // FramePointer + // - StackPointer + // - sizeof(return address) + + ScratchRegisterScope scratch(masm); + masm.movq(FramePointer, scratch); + masm.subq(StackPointer, scratch); + masm.subq(Imm32(sizeof(void*)), scratch); // Return address. + + Address frameSizeAddr(FramePointer, + BaselineFrame::reverseOffsetOfDebugFrameSize()); + masm.store32(scratch, frameSizeAddr); +#endif + + // Push the return address that's currently on top of the stack. + masm.Push(Operand(StackPointer, 0)); + + // Replace the original return address with the frame descriptor. + masm.storePtr(ImmWord(MakeFrameDescriptor(FrameType::BaselineJS)), + Address(StackPointer, sizeof(uintptr_t))); + + // Save old frame pointer, stack pointer and stub reg. + masm.Push(FramePointer); + masm.mov(StackPointer, FramePointer); + + masm.Push(ICStubReg); +} + +} // namespace jit +} // namespace js + +#endif /* jit_x64_SharedICHelpers_x64_inl_h */ diff --git a/js/src/jit/x64/SharedICHelpers-x64.h b/js/src/jit/x64/SharedICHelpers-x64.h new file mode 100644 index 0000000000..8233db5735 --- /dev/null +++ b/js/src/jit/x64/SharedICHelpers-x64.h @@ -0,0 +1,70 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#ifndef jit_x64_SharedICHelpers_x64_h +#define jit_x64_SharedICHelpers_x64_h + +#include "jit/BaselineIC.h" +#include "jit/JitFrames.h" +#include "jit/MacroAssembler.h" +#include "jit/SharedICRegisters.h" + +namespace js { +namespace jit { + +// Distance from Stack top to the top Value inside an IC stub (this is the +// return address). +static const size_t ICStackValueOffset = sizeof(void*); + +inline void EmitRestoreTailCallReg(MacroAssembler& masm) { + masm.Pop(ICTailCallReg); +} + +inline void EmitRepushTailCallReg(MacroAssembler& masm) { + masm.Push(ICTailCallReg); +} + +inline void EmitCallIC(MacroAssembler& masm, CodeOffset* callOffset) { + // The stub pointer must already be in ICStubReg. + // Call the stubcode. + masm.call(Address(ICStubReg, ICStub::offsetOfStubCode())); + *callOffset = CodeOffset(masm.currentOffset()); +} + +inline void EmitReturnFromIC(MacroAssembler& masm) { masm.ret(); } + +inline void EmitBaselineLeaveStubFrame(MacroAssembler& masm) { + Address stubAddr(FramePointer, BaselineStubFrameLayout::ICStubOffsetFromFP); + masm.loadPtr(stubAddr, ICStubReg); + + masm.mov(FramePointer, StackPointer); + masm.Pop(FramePointer); + + // The return address is on top of the stack, followed by the frame + // descriptor. Use a pop instruction to overwrite the frame descriptor + // with the return address. Note that pop increments the stack pointer + // before computing the address. + masm.Pop(Operand(StackPointer, 0)); +} + +template <typename AddrType> +inline void EmitPreBarrier(MacroAssembler& masm, const AddrType& addr, + MIRType type) { + masm.guardedCallPreBarrier(addr, type); +} + +inline void EmitStubGuardFailure(MacroAssembler& masm) { + // Load next stub into ICStubReg + masm.loadPtr(Address(ICStubReg, ICCacheIRStub::offsetOfNext()), ICStubReg); + + // Return address is already loaded, just jump to the next stubcode. + masm.jmp(Operand(ICStubReg, ICStub::offsetOfStubCode())); +} + +} // namespace jit +} // namespace js + +#endif /* jit_x64_SharedICHelpers_x64_h */ diff --git a/js/src/jit/x64/SharedICRegisters-x64.h b/js/src/jit/x64/SharedICRegisters-x64.h new file mode 100644 index 0000000000..8e52e5f3c9 --- /dev/null +++ b/js/src/jit/x64/SharedICRegisters-x64.h @@ -0,0 +1,33 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#ifndef jit_x64_SharedICRegisters_x64_h +#define jit_x64_SharedICRegisters_x64_h + +#include "jit/Registers.h" +#include "jit/RegisterSets.h" +#include "jit/x64/Assembler-x64.h" + +namespace js { +namespace jit { + +static constexpr ValueOperand R0(rcx); +static constexpr ValueOperand R1(rbx); +static constexpr ValueOperand R2(rax); + +static constexpr Register ICTailCallReg = rsi; +static constexpr Register ICStubReg = rdi; + +// FloatReg0 must be equal to ReturnFloatReg. +static constexpr FloatRegister FloatReg0 = xmm0; +static constexpr FloatRegister FloatReg1 = xmm1; +static constexpr FloatRegister FloatReg2 = xmm2; +static constexpr FloatRegister FloatReg3 = xmm3; + +} // namespace jit +} // namespace js + +#endif /* jit_x64_SharedICRegisters_x64_h */ diff --git a/js/src/jit/x64/Trampoline-x64.cpp b/js/src/jit/x64/Trampoline-x64.cpp new file mode 100644 index 0000000000..dcc9291299 --- /dev/null +++ b/js/src/jit/x64/Trampoline-x64.cpp @@ -0,0 +1,888 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#include "jit/Bailouts.h" +#include "jit/BaselineFrame.h" +#include "jit/CalleeToken.h" +#include "jit/JitFrames.h" +#include "jit/JitRuntime.h" +#include "jit/PerfSpewer.h" +#include "jit/VMFunctions.h" +#include "jit/x64/SharedICRegisters-x64.h" +#include "vm/JitActivation.h" // js::jit::JitActivation +#include "vm/JSContext.h" + +#include "jit/MacroAssembler-inl.h" + +using namespace js; +using namespace js::jit; + +using mozilla::IsPowerOfTwo; + +// This struct reflects the contents of the stack entry. +// Given a `CommonFrameLayout* frame`: +// - `frame->prevType()` should be `FrameType::CppToJSJit`. +// - Then EnterJITStackEntry starts at: +// frame->callerFramePtr() + EnterJITStackEntry::offsetFromFP() +// (the offset is negative, so this subtracts from the frame pointer) +struct EnterJITStackEntry { + // Offset from frame pointer to EnterJITStackEntry*. + static constexpr int32_t offsetFromFP() { + return -int32_t(offsetof(EnterJITStackEntry, rbp)); + } + + void* result; + +#if defined(_WIN64) + struct XMM { + using XMM128 = char[16]; + XMM128 xmm6; + XMM128 xmm7; + XMM128 xmm8; + XMM128 xmm9; + XMM128 xmm10; + XMM128 xmm11; + XMM128 xmm12; + XMM128 xmm13; + XMM128 xmm14; + XMM128 xmm15; + } xmm; + + // 16-byte aligment for xmm registers above. + uint64_t xmmPadding; + + void* rsi; + void* rdi; +#endif + + void* r15; + void* r14; + void* r13; + void* r12; + void* rbx; + void* rbp; + + // Pushed by CALL. + void* rip; +}; + +// All registers to save and restore. This includes the stack pointer, since we +// use the ability to reference register values on the stack by index. +static const LiveRegisterSet AllRegs = + LiveRegisterSet(GeneralRegisterSet(Registers::AllMask), + FloatRegisterSet(FloatRegisters::AllMask)); + +// Generates a trampoline for calling Jit compiled code from a C++ function. +// The trampoline use the EnterJitCode signature, with the standard x64 fastcall +// calling convention. +void JitRuntime::generateEnterJIT(JSContext* cx, MacroAssembler& masm) { + AutoCreatedBy acb(masm, "JitRuntime::generateEnterJIT"); + + enterJITOffset_ = startTrampolineCode(masm); + + masm.assertStackAlignment(ABIStackAlignment, + -int32_t(sizeof(uintptr_t)) /* return address */); + + const Register reg_code = IntArgReg0; + const Register reg_argc = IntArgReg1; + const Register reg_argv = IntArgReg2; + static_assert(OsrFrameReg == IntArgReg3); + +#if defined(_WIN64) + const Address token = Address(rbp, 16 + ShadowStackSpace); + const Operand scopeChain = Operand(rbp, 24 + ShadowStackSpace); + const Operand numStackValuesAddr = Operand(rbp, 32 + ShadowStackSpace); + const Operand result = Operand(rbp, 40 + ShadowStackSpace); +#else + const Register token = IntArgReg4; + const Register scopeChain = IntArgReg5; + const Operand numStackValuesAddr = Operand(rbp, 16 + ShadowStackSpace); + const Operand result = Operand(rbp, 24 + ShadowStackSpace); +#endif + + // Note: the stack pushes below must match the fields in EnterJITStackEntry. + + // Save old stack frame pointer, set new stack frame pointer. + masm.push(rbp); + masm.mov(rsp, rbp); + + // Save non-volatile registers. These must be saved by the trampoline, rather + // than by the JIT'd code, because they are scanned by the conservative + // scanner. + masm.push(rbx); + masm.push(r12); + masm.push(r13); + masm.push(r14); + masm.push(r15); +#if defined(_WIN64) + masm.push(rdi); + masm.push(rsi); + + // 16-byte aligment for vmovdqa + masm.subq(Imm32(sizeof(EnterJITStackEntry::XMM) + 8), rsp); + + masm.vmovdqa(xmm6, Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm6))); + masm.vmovdqa(xmm7, Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm7))); + masm.vmovdqa(xmm8, Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm8))); + masm.vmovdqa(xmm9, Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm9))); + masm.vmovdqa(xmm10, Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm10))); + masm.vmovdqa(xmm11, Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm11))); + masm.vmovdqa(xmm12, Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm12))); + masm.vmovdqa(xmm13, Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm13))); + masm.vmovdqa(xmm14, Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm14))); + masm.vmovdqa(xmm15, Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm15))); +#endif + + // Save arguments passed in registers needed after function call. + masm.push(result); + + // End of pushes reflected in EnterJITStackEntry, i.e. EnterJITStackEntry + // starts at this rsp. + + // Remember number of bytes occupied by argument vector + masm.mov(reg_argc, r13); + + // if we are constructing, that also needs to include newTarget + { + Label noNewTarget; + masm.branchTest32(Assembler::Zero, token, + Imm32(CalleeToken_FunctionConstructing), &noNewTarget); + + masm.addq(Imm32(1), r13); + + masm.bind(&noNewTarget); + } + + masm.shll(Imm32(3), r13); // r13 = argc * sizeof(Value) + static_assert(sizeof(Value) == 1 << 3, "Constant is baked in assembly code"); + + // Guarantee stack alignment of Jit frames. + // + // This code compensates for the offset created by the copy of the vector of + // arguments, such that the jit frame will be aligned once the return + // address is pushed on the stack. + // + // In the computation of the offset, we omit the size of the JitFrameLayout + // which is pushed on the stack, as the JitFrameLayout size is a multiple of + // the JitStackAlignment. + masm.mov(rsp, r12); + masm.subq(r13, r12); + static_assert( + sizeof(JitFrameLayout) % JitStackAlignment == 0, + "No need to consider the JitFrameLayout for aligning the stack"); + masm.andl(Imm32(JitStackAlignment - 1), r12); + masm.subq(r12, rsp); + + /*************************************************************** + Loop over argv vector, push arguments onto stack in reverse order + ***************************************************************/ + + // r13 still stores the number of bytes in the argument vector. + masm.addq(reg_argv, r13); // r13 points above last argument or newTarget + + // while r13 > rdx, push arguments. + { + Label header, footer; + masm.bind(&header); + + masm.cmpPtr(r13, reg_argv); + masm.j(AssemblerX86Shared::BelowOrEqual, &footer); + + masm.subq(Imm32(8), r13); + masm.push(Operand(r13, 0)); + masm.jmp(&header); + + masm.bind(&footer); + } + + // Load the number of actual arguments. |result| is used to store the + // actual number of arguments without adding an extra argument to the enter + // JIT. + masm.movq(result, reg_argc); + masm.unboxInt32(Operand(reg_argc, 0), reg_argc); + + // Push the callee token. + masm.push(token); + + // Push the descriptor. + masm.pushFrameDescriptorForJitCall(FrameType::CppToJSJit, reg_argc, reg_argc); + + CodeLabel returnLabel; + Label oomReturnLabel; + { + // Handle Interpreter -> Baseline OSR. + AllocatableGeneralRegisterSet regs(GeneralRegisterSet::All()); + MOZ_ASSERT(!regs.has(rbp)); + regs.take(OsrFrameReg); + regs.take(reg_code); + + Register scratch = regs.takeAny(); + + Label notOsr; + masm.branchTestPtr(Assembler::Zero, OsrFrameReg, OsrFrameReg, ¬Osr); + + Register numStackValues = regs.takeAny(); + masm.movq(numStackValuesAddr, numStackValues); + + // Push return address + masm.mov(&returnLabel, scratch); + masm.push(scratch); + + // Frame prologue. + masm.push(rbp); + masm.mov(rsp, rbp); + + // Reserve frame. + masm.subPtr(Imm32(BaselineFrame::Size()), rsp); + + Register framePtrScratch = regs.takeAny(); + masm.touchFrameValues(numStackValues, scratch, framePtrScratch); + masm.mov(rsp, framePtrScratch); + + // Reserve space for locals and stack values. + Register valuesSize = regs.takeAny(); + masm.mov(numStackValues, valuesSize); + masm.shll(Imm32(3), valuesSize); + masm.subPtr(valuesSize, rsp); + + // Enter exit frame. + masm.pushFrameDescriptor(FrameType::BaselineJS); + masm.push(Imm32(0)); // Fake return address. + masm.push(FramePointer); + // No GC things to mark, push a bare token. + masm.loadJSContext(scratch); + masm.enterFakeExitFrame(scratch, scratch, ExitFrameType::Bare); + + regs.add(valuesSize); + + masm.push(reg_code); + + using Fn = bool (*)(BaselineFrame * frame, InterpreterFrame * interpFrame, + uint32_t numStackValues); + masm.setupUnalignedABICall(scratch); + masm.passABIArg(framePtrScratch); // BaselineFrame + masm.passABIArg(OsrFrameReg); // InterpreterFrame + masm.passABIArg(numStackValues); + masm.callWithABI<Fn, jit::InitBaselineFrameForOsr>( + MoveOp::GENERAL, CheckUnsafeCallWithABI::DontCheckHasExitFrame); + + masm.pop(reg_code); + + MOZ_ASSERT(reg_code != ReturnReg); + + Label error; + masm.addPtr(Imm32(ExitFrameLayout::SizeWithFooter()), rsp); + masm.branchIfFalseBool(ReturnReg, &error); + + // If OSR-ing, then emit instrumentation for setting lastProfilerFrame + // if profiler instrumentation is enabled. + { + Label skipProfilingInstrumentation; + AbsoluteAddress addressOfEnabled( + cx->runtime()->geckoProfiler().addressOfEnabled()); + masm.branch32(Assembler::Equal, addressOfEnabled, Imm32(0), + &skipProfilingInstrumentation); + masm.profilerEnterFrame(rbp, scratch); + masm.bind(&skipProfilingInstrumentation); + } + + masm.jump(reg_code); + + // OOM: frame epilogue, load error value, discard return address and return. + masm.bind(&error); + masm.mov(rbp, rsp); + masm.pop(rbp); + masm.addPtr(Imm32(sizeof(uintptr_t)), rsp); // Return address. + masm.moveValue(MagicValue(JS_ION_ERROR), JSReturnOperand); + masm.jump(&oomReturnLabel); + + masm.bind(¬Osr); + masm.movq(scopeChain, R1.scratchReg()); + } + + // The call will push the return address and frame pointer on the stack, thus + // we check that the stack would be aligned once the call is complete. + masm.assertStackAlignment(JitStackAlignment, 2 * sizeof(uintptr_t)); + + // Call function. + masm.callJitNoProfiler(reg_code); + + { + // Interpreter -> Baseline OSR will return here. + masm.bind(&returnLabel); + masm.addCodeLabel(returnLabel); + masm.bind(&oomReturnLabel); + } + + // Discard arguments and padding. Set rsp to the address of the + // EnterJITStackEntry on the stack. + masm.lea(Operand(rbp, EnterJITStackEntry::offsetFromFP()), rsp); + + /***************************************************************** + Place return value where it belongs, pop all saved registers + *****************************************************************/ + masm.pop(r12); // vp + masm.storeValue(JSReturnOperand, Operand(r12, 0)); + + // Restore non-volatile registers. +#if defined(_WIN64) + masm.vmovdqa(Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm6)), xmm6); + masm.vmovdqa(Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm7)), xmm7); + masm.vmovdqa(Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm8)), xmm8); + masm.vmovdqa(Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm9)), xmm9); + masm.vmovdqa(Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm10)), xmm10); + masm.vmovdqa(Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm11)), xmm11); + masm.vmovdqa(Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm12)), xmm12); + masm.vmovdqa(Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm13)), xmm13); + masm.vmovdqa(Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm14)), xmm14); + masm.vmovdqa(Operand(rsp, offsetof(EnterJITStackEntry::XMM, xmm15)), xmm15); + + masm.addq(Imm32(sizeof(EnterJITStackEntry::XMM) + 8), rsp); + + masm.pop(rsi); + masm.pop(rdi); +#endif + masm.pop(r15); + masm.pop(r14); + masm.pop(r13); + masm.pop(r12); + masm.pop(rbx); + + // Restore frame pointer and return. + masm.pop(rbp); + masm.ret(); +} + +// static +mozilla::Maybe<::JS::ProfilingFrameIterator::RegisterState> +JitRuntime::getCppEntryRegisters(JitFrameLayout* frameStackAddress) { + if (frameStackAddress->prevType() != FrameType::CppToJSJit) { + // This is not a CppToJSJit frame, there are no C++ registers here. + return mozilla::Nothing{}; + } + + // Compute pointer to start of EnterJITStackEntry on the stack. + uint8_t* fp = frameStackAddress->callerFramePtr(); + auto* enterJITStackEntry = reinterpret_cast<EnterJITStackEntry*>( + fp + EnterJITStackEntry::offsetFromFP()); + + // Extract native function call registers. + ::JS::ProfilingFrameIterator::RegisterState registerState; + registerState.fp = enterJITStackEntry->rbp; + registerState.pc = enterJITStackEntry->rip; + // sp should be inside the caller's frame, so set sp to the value of the stack + // pointer before the call to the EnterJit trampoline. + registerState.sp = &enterJITStackEntry->rip + 1; + // No lr in this world. + registerState.lr = nullptr; + return mozilla::Some(registerState); +} + +// Push AllRegs in a way that is compatible with RegisterDump, regardless of +// what PushRegsInMask might do to reduce the set size. +static void DumpAllRegs(MacroAssembler& masm) { +#ifdef ENABLE_WASM_SIMD + masm.PushRegsInMask(AllRegs); +#else + // When SIMD isn't supported, PushRegsInMask reduces the set of float + // registers to be double-sized, while the RegisterDump expects each of + // the float registers to have the maximal possible size + // (Simd128DataSize). To work around this, we just spill the double + // registers by hand here, using the register dump offset directly. + for (GeneralRegisterBackwardIterator iter(AllRegs.gprs()); iter.more(); + ++iter) { + masm.Push(*iter); + } + + masm.reserveStack(sizeof(RegisterDump::FPUArray)); + for (FloatRegisterBackwardIterator iter(AllRegs.fpus()); iter.more(); + ++iter) { + FloatRegister reg = *iter; + Address spillAddress(StackPointer, reg.getRegisterDumpOffsetInBytes()); + masm.storeDouble(reg, spillAddress); + } +#endif +} + +void JitRuntime::generateInvalidator(MacroAssembler& masm, Label* bailoutTail) { + AutoCreatedBy acb(masm, "JitRuntime::generateInvalidator"); + + // See explanatory comment in x86's JitRuntime::generateInvalidator. + + invalidatorOffset_ = startTrampolineCode(masm); + + // Push registers such that we can access them from [base + code]. + DumpAllRegs(masm); + + masm.movq(rsp, rax); // Argument to jit::InvalidationBailout. + + // Make space for InvalidationBailout's bailoutInfo outparam. + masm.reserveStack(sizeof(void*)); + masm.movq(rsp, rbx); + + using Fn = + bool (*)(InvalidationBailoutStack * sp, BaselineBailoutInfo * *info); + masm.setupUnalignedABICall(rdx); + masm.passABIArg(rax); + masm.passABIArg(rbx); + masm.callWithABI<Fn, InvalidationBailout>( + MoveOp::GENERAL, CheckUnsafeCallWithABI::DontCheckOther); + + masm.pop(r9); // Get the bailoutInfo outparam. + + // Pop the machine state and the dead frame. + masm.moveToStackPtr(FramePointer); + + // Jump to shared bailout tail. The BailoutInfo pointer has to be in r9. + masm.jmp(bailoutTail); +} + +void JitRuntime::generateArgumentsRectifier(MacroAssembler& masm, + ArgumentsRectifierKind kind) { + // Do not erase the frame pointer in this function. + + AutoCreatedBy acb(masm, "JitRuntime::generateArgumentsRectifier"); + + switch (kind) { + case ArgumentsRectifierKind::Normal: + argumentsRectifierOffset_ = startTrampolineCode(masm); + break; + case ArgumentsRectifierKind::TrialInlining: + trialInliningArgumentsRectifierOffset_ = startTrampolineCode(masm); + break; + } + + // Caller: + // [arg2] [arg1] [this] [[argc] [callee] [descr] [raddr]] <- rsp + + // Frame prologue. + // + // NOTE: if this changes, fix the Baseline bailout code too! + // See BaselineStackBuilder::calculatePrevFramePtr and + // BaselineStackBuilder::buildRectifierFrame (in BaselineBailouts.cpp). + masm.push(FramePointer); + masm.movq(rsp, FramePointer); + + // Load argc. + masm.loadNumActualArgs(FramePointer, r8); + + // Load |nformals| into %rcx. + masm.loadPtr(Address(rbp, RectifierFrameLayout::offsetOfCalleeToken()), rax); + masm.mov(rax, rcx); + masm.andq(Imm32(uint32_t(CalleeTokenMask)), rcx); + masm.loadFunctionArgCount(rcx, rcx); + + // Stash another copy in r11, since we are going to do destructive operations + // on rcx + masm.mov(rcx, r11); + + static_assert( + CalleeToken_FunctionConstructing == 1, + "Ensure that we can use the constructing bit to count the value"); + masm.mov(rax, rdx); + masm.andq(Imm32(uint32_t(CalleeToken_FunctionConstructing)), rdx); + + // Including |this|, and |new.target|, there are (|nformals| + 1 + + // isConstructing) arguments to push to the stack. Then we push a + // JitFrameLayout. We compute the padding expressed in the number of extra + // |undefined| values to push on the stack. + static_assert( + sizeof(JitFrameLayout) % JitStackAlignment == 0, + "No need to consider the JitFrameLayout for aligning the stack"); + static_assert( + JitStackAlignment % sizeof(Value) == 0, + "Ensure that we can pad the stack by pushing extra UndefinedValue"); + static_assert(IsPowerOfTwo(JitStackValueAlignment), + "must have power of two for masm.andl to do its job"); + + masm.addl( + Imm32(JitStackValueAlignment - 1 /* for padding */ + 1 /* for |this| */), + rcx); + masm.addl(rdx, rcx); + masm.andl(Imm32(~(JitStackValueAlignment - 1)), rcx); + + // Load the number of |undefined|s to push into %rcx. Subtract 1 for |this|. + masm.subl(r8, rcx); + masm.subl(Imm32(1), rcx); + + // Caller: + // [arg2] [arg1] [this] [ [argc] [callee] [descr] [raddr] ] <- rsp + // '--- #r8 ---' + // + // Rectifier frame: + // [rbp'] [undef] [undef] [undef] [arg2] [arg1] [this] [ [argc] [callee] + // [descr] [raddr] ] + // '------- #rcx --------' '--- #r8 ---' + + // Copy the number of actual arguments into rdx. + masm.mov(r8, rdx); + + masm.moveValue(UndefinedValue(), ValueOperand(r10)); + + // Push undefined. (including the padding) + { + Label undefLoopTop; + masm.bind(&undefLoopTop); + + masm.push(r10); + masm.subl(Imm32(1), rcx); + masm.j(Assembler::NonZero, &undefLoopTop); + } + + // Get the topmost argument. + static_assert(sizeof(Value) == 8, "TimesEight is used to skip arguments"); + + // Get the topmost argument. + BaseIndex b(FramePointer, r8, TimesEight, sizeof(RectifierFrameLayout)); + masm.lea(Operand(b), rcx); + + // Push arguments, |nargs| + 1 times (to include |this|). + masm.addl(Imm32(1), r8); + { + Label copyLoopTop; + + masm.bind(©LoopTop); + masm.push(Operand(rcx, 0x0)); + masm.subq(Imm32(sizeof(Value)), rcx); + masm.subl(Imm32(1), r8); + masm.j(Assembler::NonZero, ©LoopTop); + } + + // if constructing, copy newTarget + { + Label notConstructing; + + masm.branchTest32(Assembler::Zero, rax, + Imm32(CalleeToken_FunctionConstructing), + ¬Constructing); + + // thisFrame[numFormals] = prevFrame[argc] + ValueOperand newTarget(r10); + + // Load vp[argc]. Add sizeof(Value) for |this|. + BaseIndex newTargetSrc(FramePointer, rdx, TimesEight, + sizeof(RectifierFrameLayout) + sizeof(Value)); + masm.loadValue(newTargetSrc, newTarget); + + // Again, 1 for |this| + BaseIndex newTargetDest(rsp, r11, TimesEight, sizeof(Value)); + masm.storeValue(newTarget, newTargetDest); + + masm.bind(¬Constructing); + } + + // Caller: + // [arg2] [arg1] [this] [ [argc] [callee] [descr] [raddr] ] + // + // + // Rectifier frame: + // [rbp'] <- rbp [undef] [undef] [undef] [arg2] [arg1] [this] <- rsp [ [argc] + // [callee] [descr] [raddr] ] + // + + // Construct JitFrameLayout. + masm.push(rax); // callee token + masm.pushFrameDescriptorForJitCall(FrameType::Rectifier, rdx, rdx); + + // Call the target function. + masm.andq(Imm32(uint32_t(CalleeTokenMask)), rax); + switch (kind) { + case ArgumentsRectifierKind::Normal: + masm.loadJitCodeRaw(rax, rax); + argumentsRectifierReturnOffset_ = masm.callJitNoProfiler(rax); + break; + case ArgumentsRectifierKind::TrialInlining: + Label noBaselineScript, done; + masm.loadBaselineJitCodeRaw(rax, rbx, &noBaselineScript); + masm.callJitNoProfiler(rbx); + masm.jump(&done); + + // See BaselineCacheIRCompiler::emitCallInlinedFunction. + masm.bind(&noBaselineScript); + masm.loadJitCodeRaw(rax, rax); + masm.callJitNoProfiler(rax); + masm.bind(&done); + break; + } + + masm.mov(FramePointer, StackPointer); + masm.pop(FramePointer); + masm.ret(); +} + +static void PushBailoutFrame(MacroAssembler& masm, Register spArg) { + // Push registers such that we can access them from [base + code]. + DumpAllRegs(masm); + + // Get the stack pointer into a register, pre-alignment. + masm.movq(rsp, spArg); +} + +static void GenerateBailoutThunk(MacroAssembler& masm, Label* bailoutTail) { + PushBailoutFrame(masm, r8); + + // Make space for Bailout's bailoutInfo outparam. + masm.reserveStack(sizeof(void*)); + masm.movq(rsp, r9); + + // Call the bailout function. + using Fn = bool (*)(BailoutStack * sp, BaselineBailoutInfo * *info); + masm.setupUnalignedABICall(rax); + masm.passABIArg(r8); + masm.passABIArg(r9); + masm.callWithABI<Fn, Bailout>(MoveOp::GENERAL, + CheckUnsafeCallWithABI::DontCheckOther); + + masm.pop(r9); // Get the bailoutInfo outparam. + + // Remove both the bailout frame and the topmost Ion frame's stack. + masm.moveToStackPtr(FramePointer); + + // Jump to shared bailout tail. The BailoutInfo pointer has to be in r9. + masm.jmp(bailoutTail); +} + +void JitRuntime::generateBailoutHandler(MacroAssembler& masm, + Label* bailoutTail) { + AutoCreatedBy acb(masm, "JitRuntime::generateBailoutHandler"); + + bailoutHandlerOffset_ = startTrampolineCode(masm); + + GenerateBailoutThunk(masm, bailoutTail); +} + +bool JitRuntime::generateVMWrapper(JSContext* cx, MacroAssembler& masm, + const VMFunctionData& f, DynFn nativeFun, + uint32_t* wrapperOffset) { + AutoCreatedBy acb(masm, "JitRuntime::generateVMWrapper"); + + *wrapperOffset = startTrampolineCode(masm); + + // Avoid conflicts with argument registers while discarding the result after + // the function call. + AllocatableGeneralRegisterSet regs(Register::Codes::WrapperMask); + + static_assert( + (Register::Codes::VolatileMask & ~Register::Codes::WrapperMask) == 0, + "Wrapper register set must be a superset of Volatile register set"); + + // The context is the first argument. + Register cxreg = IntArgReg0; + regs.take(cxreg); + + // Stack is: + // ... frame ... + // +12 [args] + // +8 descriptor + // +0 returnAddress + // + // Push the frame pointer to finish the exit frame, then link it up. + masm.Push(FramePointer); + masm.moveStackPtrTo(FramePointer); + masm.loadJSContext(cxreg); + masm.enterExitFrame(cxreg, regs.getAny(), &f); + + // Save the current stack pointer as the base for copying arguments. + Register argsBase = InvalidReg; + if (f.explicitArgs) { + argsBase = r10; + regs.take(argsBase); + masm.lea(Operand(rsp, ExitFrameLayout::SizeWithFooter()), argsBase); + } + + // Reserve space for the outparameter. + Register outReg = InvalidReg; + switch (f.outParam) { + case Type_Value: + outReg = regs.takeAny(); + masm.reserveStack(sizeof(Value)); + masm.movq(esp, outReg); + break; + + case Type_Handle: + outReg = regs.takeAny(); + masm.PushEmptyRooted(f.outParamRootType); + masm.movq(esp, outReg); + break; + + case Type_Int32: + case Type_Bool: + case Type_Pointer: + outReg = regs.takeAny(); + masm.reserveStack(sizeof(uintptr_t)); + masm.movq(esp, outReg); + break; + + case Type_Double: + outReg = regs.takeAny(); + masm.reserveStack(sizeof(double)); + masm.movq(esp, outReg); + break; + + default: + MOZ_ASSERT(f.outParam == Type_Void); + break; + } + + masm.setupUnalignedABICall(regs.getAny()); + masm.passABIArg(cxreg); + + size_t argDisp = 0; + + // Copy arguments. + for (uint32_t explicitArg = 0; explicitArg < f.explicitArgs; explicitArg++) { + switch (f.argProperties(explicitArg)) { + case VMFunctionData::WordByValue: + if (f.argPassedInFloatReg(explicitArg)) { + masm.passABIArg(MoveOperand(argsBase, argDisp), MoveOp::DOUBLE); + } else { + masm.passABIArg(MoveOperand(argsBase, argDisp), MoveOp::GENERAL); + } + argDisp += sizeof(void*); + break; + case VMFunctionData::WordByRef: + masm.passABIArg( + MoveOperand(argsBase, argDisp, MoveOperand::Kind::EffectiveAddress), + MoveOp::GENERAL); + argDisp += sizeof(void*); + break; + case VMFunctionData::DoubleByValue: + case VMFunctionData::DoubleByRef: + MOZ_CRASH("NYI: x64 callVM should not be used with 128bits values."); + } + } + + // Copy the implicit outparam, if any. + if (outReg != InvalidReg) { + masm.passABIArg(outReg); + } + + masm.callWithABI(nativeFun, MoveOp::GENERAL, + CheckUnsafeCallWithABI::DontCheckHasExitFrame); + + // Test for failure. + switch (f.failType()) { + case Type_Cell: + masm.branchTestPtr(Assembler::Zero, rax, rax, masm.failureLabel()); + break; + case Type_Bool: + masm.testb(rax, rax); + masm.j(Assembler::Zero, masm.failureLabel()); + break; + case Type_Void: + break; + default: + MOZ_CRASH("unknown failure kind"); + } + + // Load the outparam and free any allocated stack. + switch (f.outParam) { + case Type_Handle: + masm.popRooted(f.outParamRootType, ReturnReg, JSReturnOperand); + break; + + case Type_Value: + masm.loadValue(Address(esp, 0), JSReturnOperand); + masm.freeStack(sizeof(Value)); + break; + + case Type_Int32: + masm.load32(Address(esp, 0), ReturnReg); + masm.freeStack(sizeof(uintptr_t)); + break; + + case Type_Bool: + masm.load8ZeroExtend(Address(esp, 0), ReturnReg); + masm.freeStack(sizeof(uintptr_t)); + break; + + case Type_Double: + masm.loadDouble(Address(esp, 0), ReturnDoubleReg); + masm.freeStack(sizeof(double)); + break; + + case Type_Pointer: + masm.loadPtr(Address(esp, 0), ReturnReg); + masm.freeStack(sizeof(uintptr_t)); + break; + + default: + MOZ_ASSERT(f.outParam == Type_Void); + break; + } + + // Until C++ code is instrumented against Spectre, prevent speculative + // execution from returning any private data. + if (f.returnsData() && JitOptions.spectreJitToCxxCalls) { + masm.speculationBarrier(); + } + + // Pop ExitFooterFrame and the frame pointer. + masm.leaveExitFrame(0); + masm.pop(FramePointer); + + // Return. Subtract sizeof(void*) for the frame pointer. + masm.retn(Imm32(sizeof(ExitFrameLayout) - sizeof(void*) + + f.explicitStackSlots() * sizeof(void*) + + f.extraValuesToPop * sizeof(Value))); + + return true; +} + +uint32_t JitRuntime::generatePreBarrier(JSContext* cx, MacroAssembler& masm, + MIRType type) { + AutoCreatedBy acb(masm, "JitRuntime::generatePreBarrier"); + + uint32_t offset = startTrampolineCode(masm); + + static_assert(PreBarrierReg == rdx); + Register temp1 = rax; + Register temp2 = rbx; + Register temp3 = rcx; + masm.push(temp1); + masm.push(temp2); + masm.push(temp3); + + Label noBarrier; + masm.emitPreBarrierFastPath(cx->runtime(), type, temp1, temp2, temp3, + &noBarrier); + + // Call into C++ to mark this GC thing. + masm.pop(temp3); + masm.pop(temp2); + masm.pop(temp1); + + LiveRegisterSet regs = + LiveRegisterSet(GeneralRegisterSet(Registers::VolatileMask), + FloatRegisterSet(FloatRegisters::VolatileMask)); + masm.PushRegsInMask(regs); + + masm.mov(ImmPtr(cx->runtime()), rcx); + + masm.setupUnalignedABICall(rax); + masm.passABIArg(rcx); + masm.passABIArg(rdx); + masm.callWithABI(JitPreWriteBarrier(type)); + + masm.PopRegsInMask(regs); + masm.ret(); + + masm.bind(&noBarrier); + masm.pop(temp3); + masm.pop(temp2); + masm.pop(temp1); + masm.ret(); + + return offset; +} + +void JitRuntime::generateBailoutTailStub(MacroAssembler& masm, + Label* bailoutTail) { + AutoCreatedBy acb(masm, "JitRuntime::generateBailoutTailStub"); + + masm.bind(bailoutTail); + masm.generateBailoutTail(rdx, r9); +} |