void JIT::compileOpCall(OpcodeID opcodeID, Instruction* instruction, unsigned callLinkInfoIndex)
{
int callee = instruction[1].u.operand;
int argCount = instruction[2].u.operand;
int registerOffset = instruction[3].u.operand;
Jump wasEval;
if (opcodeID == op_call_eval) {
JITStubCall stubCall(this, cti_op_call_eval);
stubCall.addArgument(callee);
stubCall.addArgument(JIT::Imm32(registerOffset));
stubCall.addArgument(JIT::Imm32(argCount));
stubCall.call();
wasEval = branch32(NotEqual, regT1, Imm32(JSValue::EmptyValueTag));
}
emitLoad(callee, regT1, regT0);
DataLabelPtr addressOfLinkedFunctionCheck;
BEGIN_UNINTERRUPTED_SEQUENCE(sequenceOpCall);
Jump jumpToSlow = branchPtrWithPatch(NotEqual, regT0, addressOfLinkedFunctionCheck, ImmPtr(0));
END_UNINTERRUPTED_SEQUENCE(sequenceOpCall);
addSlowCase(jumpToSlow);
ASSERT_JIT_OFFSET(differenceBetween(addressOfLinkedFunctionCheck, jumpToSlow), patchOffsetOpCallCompareToJump);
m_callStructureStubCompilationInfo[callLinkInfoIndex].hotPathBegin = addressOfLinkedFunctionCheck;
addSlowCase(branch32(NotEqual, regT1, Imm32(JSValue::CellTag)));
// The following is the fast case, only used whan a callee can be linked.
// Fast version of stack frame initialization, directly relative to edi.
// Note that this omits to set up RegisterFile::CodeBlock, which is set in the callee
loadPtr(Address(regT0, OBJECT_OFFSETOF(JSFunction, m_scopeChain)), regT2);
store32(Imm32(JSValue::Int32Tag), tagFor(registerOffset + RegisterFile::ArgumentCount));
store32(Imm32(argCount), payloadFor(registerOffset + RegisterFile::ArgumentCount));
storePtr(callFrameRegister, payloadFor(RegisterFile::CallerFrame + registerOffset, callFrameRegister));
emitStore(registerOffset + RegisterFile::Callee, regT1, regT0);
store32(Imm32(JSValue::CellTag), tagFor(registerOffset + RegisterFile::ScopeChain));
store32(regT2, payloadFor(registerOffset + RegisterFile::ScopeChain));
addPtr(Imm32(registerOffset * sizeof(Register)), callFrameRegister);
// Call to the callee
m_callStructureStubCompilationInfo[callLinkInfoIndex].hotPathOther = emitNakedCall();
if (opcodeID == op_call_eval)
wasEval.link(this);
sampleCodeBlock(m_codeBlock);
}
void JIT::compileOpCall(OpcodeID opcodeID, Instruction* instruction, unsigned callLinkInfoIndex)
{
int dst = instruction[1].u.operand;
int callee = instruction[2].u.operand;
int argCount = instruction[3].u.operand;
int registerOffset = instruction[4].u.operand;
// Handle eval
Jump wasEval;
if (opcodeID == op_call_eval) {
emitGetVirtualRegister(callee, X86::ecx);
compileOpCallEvalSetupArgs(instruction);
emitCTICall(Interpreter::cti_op_call_eval);
wasEval = jnePtr(X86::eax, ImmPtr(JSValuePtr::encode(JSImmediate::impossibleValue())));
}
// This plants a check for a cached JSFunction value, so we can plant a fast link to the callee.
// This deliberately leaves the callee in ecx, used when setting up the stack frame below
emitGetVirtualRegister(callee, X86::ecx);
DataLabelPtr addressOfLinkedFunctionCheck;
Jump jumpToSlow = jnePtrWithPatch(X86::ecx, addressOfLinkedFunctionCheck, ImmPtr(JSValuePtr::encode(JSImmediate::impossibleValue())));
addSlowCase(jumpToSlow);
ASSERT(differenceBetween(addressOfLinkedFunctionCheck, jumpToSlow) == patchOffsetOpCallCompareToJump);
m_callStructureStubCompilationInfo[callLinkInfoIndex].hotPathBegin = addressOfLinkedFunctionCheck;
// The following is the fast case, only used whan a callee can be linked.
// In the case of OpConstruct, call out to a cti_ function to create the new object.
if (opcodeID == op_construct) {
int proto = instruction[5].u.operand;
int thisRegister = instruction[6].u.operand;
emitPutJITStubArg(X86::ecx, 1);
emitPutJITStubArgFromVirtualRegister(proto, 4, X86::eax);
emitCTICall(Interpreter::cti_op_construct_JSConstruct);
emitPutVirtualRegister(thisRegister);
emitGetVirtualRegister(callee, X86::ecx);
}
// Fast version of stack frame initialization, directly relative to edi.
// Note that this omits to set up RegisterFile::CodeBlock, which is set in the callee
storePtr(ImmPtr(JSValuePtr::encode(noValue())), Address(callFrameRegister, (registerOffset + RegisterFile::OptionalCalleeArguments) * static_cast<int>(sizeof(Register))));
storePtr(X86::ecx, Address(callFrameRegister, (registerOffset + RegisterFile::Callee) * static_cast<int>(sizeof(Register))));
loadPtr(Address(X86::ecx, FIELD_OFFSET(JSFunction, m_scopeChain) + FIELD_OFFSET(ScopeChain, m_node)), X86::edx); // newScopeChain
store32(Imm32(argCount), Address(callFrameRegister, (registerOffset + RegisterFile::ArgumentCount) * static_cast<int>(sizeof(Register))));
storePtr(callFrameRegister, Address(callFrameRegister, (registerOffset + RegisterFile::CallerFrame) * static_cast<int>(sizeof(Register))));
storePtr(X86::edx, Address(callFrameRegister, (registerOffset + RegisterFile::ScopeChain) * static_cast<int>(sizeof(Register))));
addPtr(Imm32(registerOffset * sizeof(Register)), callFrameRegister);
// Call to the callee
m_callStructureStubCompilationInfo[callLinkInfoIndex].hotPathOther = emitNakedCall(reinterpret_cast<void*>(unreachable));
if (opcodeID == op_call_eval)
wasEval.link(this);
// Put the return value in dst. In the interpreter, op_ret does this.
emitPutVirtualRegister(dst);
#if ENABLE(CODEBLOCK_SAMPLING)
storePtr(ImmPtr(m_codeBlock), m_interpreter->sampler()->codeBlockSlot());
#endif
}
void Jit::Comp_FPU2op(MIPSOpcode op) {
CONDITIONAL_DISABLE(FPU);
int fs = _FS;
int fd = _FD;
auto execRounding = [&](void (XEmitter::*conv)(X64Reg, OpArg), int setMXCSR) {
fpr.SpillLock(fd, fs);
fpr.MapReg(fd, fs == fd, true);
// Small optimization: 0 is our default mode anyway.
if (setMXCSR == 0 && !js.hasSetRounding) {
setMXCSR = -1;
}
if (setMXCSR != -1) {
STMXCSR(MIPSSTATE_VAR(mxcsrTemp));
MOV(32, R(TEMPREG), MIPSSTATE_VAR(mxcsrTemp));
AND(32, R(TEMPREG), Imm32(~(3 << 13)));
OR(32, R(TEMPREG), Imm32(setMXCSR << 13));
MOV(32, MIPSSTATE_VAR(temp), R(TEMPREG));
LDMXCSR(MIPSSTATE_VAR(temp));
}
(this->*conv)(TEMPREG, fpr.R(fs));
// Did we get an indefinite integer value?
CMP(32, R(TEMPREG), Imm32(0x80000000));
FixupBranch skip = J_CC(CC_NE);
if (fd != fs) {
CopyFPReg(fpr.RX(fd), fpr.R(fs));
}
XORPS(XMM1, R(XMM1));
CMPSS(fpr.RX(fd), R(XMM1), CMP_LT);
// At this point, -inf = 0xffffffff, inf/nan = 0x00000000.
// We want -inf to be 0x80000000 inf/nan to be 0x7fffffff, so we flip those bits.
MOVD_xmm(R(TEMPREG), fpr.RX(fd));
XOR(32, R(TEMPREG), Imm32(0x7fffffff));
SetJumpTarget(skip);
MOVD_xmm(fpr.RX(fd), R(TEMPREG));
if (setMXCSR != -1) {
LDMXCSR(MIPSSTATE_VAR(mxcsrTemp));
}
};
switch (op & 0x3f) {
case 5: //F(fd) = fabsf(F(fs)); break; //abs
fpr.SpillLock(fd, fs);
fpr.MapReg(fd, fd == fs, true);
MOV(PTRBITS, R(TEMPREG), ImmPtr(&ssNoSignMask[0]));
if (fd != fs && fpr.IsMapped(fs)) {
MOVAPS(fpr.RX(fd), MatR(TEMPREG));
ANDPS(fpr.RX(fd), fpr.R(fs));
} else {
if (fd != fs) {
MOVSS(fpr.RX(fd), fpr.R(fs));
}
ANDPS(fpr.RX(fd), MatR(TEMPREG));
}
break;
case 6: //F(fd) = F(fs); break; //mov
if (fd != fs) {
fpr.SpillLock(fd, fs);
fpr.MapReg(fd, fd == fs, true);
CopyFPReg(fpr.RX(fd), fpr.R(fs));
}
break;
case 7: //F(fd) = -F(fs); break; //neg
fpr.SpillLock(fd, fs);
fpr.MapReg(fd, fd == fs, true);
MOV(PTRBITS, R(TEMPREG), ImmPtr(&ssSignBits2[0]));
if (fd != fs && fpr.IsMapped(fs)) {
MOVAPS(fpr.RX(fd), MatR(TEMPREG));
XORPS(fpr.RX(fd), fpr.R(fs));
} else {
if (fd != fs) {
MOVSS(fpr.RX(fd), fpr.R(fs));
}
XORPS(fpr.RX(fd), MatR(TEMPREG));
}
break;
case 4: //F(fd) = sqrtf(F(fs)); break; //sqrt
fpr.SpillLock(fd, fs);
fpr.MapReg(fd, fd == fs, true);
SQRTSS(fpr.RX(fd), fpr.R(fs));
break;
case 13: //FsI(fd) = F(fs)>=0 ? (int)floorf(F(fs)) : (int)ceilf(F(fs)); break; //trunc.w.s
execRounding(&XEmitter::CVTTSS2SI, -1);
break;
case 32: //F(fd) = (float)FsI(fs); break; //cvt.s.w
fpr.SpillLock(fd, fs);
fpr.MapReg(fd, fs == fd, true);
if (fpr.IsMapped(fs)) {
CVTDQ2PS(fpr.RX(fd), fpr.R(fs));
} else {
// If fs was fd, we'd be in the case above since we mapped fd.
MOVSS(fpr.RX(fd), fpr.R(fs));
CVTDQ2PS(fpr.RX(fd), fpr.R(fd));
}
break;
case 36: //FsI(fd) = (int) F(fs); break; //cvt.w.s
// Uses the current rounding mode.
execRounding(&XEmitter::CVTSS2SI, -1);
break;
case 12: //FsI(fd) = (int)floorf(F(fs)+0.5f); break; //round.w.s
execRounding(&XEmitter::CVTSS2SI, 0);
break;
case 14: //FsI(fd) = (int)ceilf (F(fs)); break; //ceil.w.s
execRounding(&XEmitter::CVTSS2SI, 2);
break;
case 15: //FsI(fd) = (int)floorf(F(fs)); break; //floor.w.s
execRounding(&XEmitter::CVTSS2SI, 1);
break;
default:
DISABLE;
return;
}
fpr.ReleaseSpillLocks();
}
