void
CodeGeneratorX64::emitSimdLoad(LAsmJSLoadHeap *ins)
{
MAsmJSLoadHeap *mir = ins->mir();
Scalar::Type type = mir->accessType();
const LAllocation *ptr = ins->ptr();
FloatRegister out = ToFloatRegister(ins->output());
Operand srcAddr(HeapReg);
if (ptr->isConstant()) {
int32_t ptrImm = ptr->toConstant()->toInt32();
MOZ_ASSERT(ptrImm >= 0);
srcAddr = Operand(HeapReg, ptrImm);
} else {
srcAddr = Operand(HeapReg, ToRegister(ptr), TimesOne);
}
uint32_t maybeCmpOffset = AsmJSHeapAccess::NoLengthCheck;
if (mir->needsBoundsCheck()) {
maybeCmpOffset = masm.cmp32WithPatch(ToRegister(ptr), Imm32(0)).offset();
masm.j(Assembler::AboveOrEqual, mir->outOfBoundsLabel()); // Throws RangeError
}
unsigned numElems = mir->numSimdElems();
if (numElems == 3) {
MOZ_ASSERT(type == Scalar::Int32x4 || type == Scalar::Float32x4);
Operand shiftedOffset(HeapReg);
if (ptr->isConstant())
shiftedOffset = Operand(HeapReg, ptr->toConstant()->toInt32() + 2 * sizeof(float));
else
shiftedOffset = Operand(HeapReg, ToRegister(ptr), TimesOne, 2 * sizeof(float));
// Load XY
uint32_t before = masm.size();
loadSimd(type, 2, srcAddr, out);
uint32_t after = masm.size();
// We're noting a load of 3 elements, so that the bounds check checks
// for 3 elements.
masm.append(AsmJSHeapAccess(before, after, 3, type, maybeCmpOffset));
// Load Z (W is zeroed)
before = after;
loadSimd(type, 1, shiftedOffset, ScratchSimdReg);
after = masm.size();
masm.append(AsmJSHeapAccess(before, after, 1, type));
// Move ZW atop XY
masm.vmovlhps(ScratchSimdReg, out, out);
return;
}
uint32_t before = masm.size();
loadSimd(type, numElems, srcAddr, out);
uint32_t after = masm.size();
masm.append(AsmJSHeapAccess(before, after, numElems, type, maybeCmpOffset));
}
