private:
bool X86FastEmitCompare(const Value *LHS, const Value *RHS, EVT VT, DebugLoc DL);
- bool X86FastEmitLoad(EVT VT, const X86AddressMode &AM, MachineMemOperand *MMO,
- unsigned &ResultReg);
+ bool X86FastEmitLoad(EVT VT, X86AddressMode &AM, MachineMemOperand *MMO,
+ unsigned &ResultReg, unsigned Alignment = 1);
- bool X86FastEmitStore(EVT VT, const Value *Val, const X86AddressMode &AM,
+ bool X86FastEmitStore(EVT VT, const Value *Val, X86AddressMode &AM,
MachineMemOperand *MMO = nullptr, bool Aligned = false);
bool X86FastEmitStore(EVT VT, unsigned ValReg, bool ValIsKill,
- const X86AddressMode &AM,
+ X86AddressMode &AM,
MachineMemOperand *MMO = nullptr, bool Aligned = false);
bool X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
bool foldX86XALUIntrinsic(X86::CondCode &CC, const Instruction *I,
const Value *Cond);
+
+ const MachineInstrBuilder &addFullAddress(const MachineInstrBuilder &MIB,
+ X86AddressMode &AM);
};
} // end anonymous namespace.
return std::make_pair(CC, NeedSwap);
}
+/// \brief Adds a complex addressing mode to the given machine instr builder.
+/// Note, this will constrain the index register. If its not possible to
+/// constrain the given index register, then a new one will be created. The
+/// IndexReg field of the addressing mode will be updated to match in this case.
+const MachineInstrBuilder &
+X86FastISel::addFullAddress(const MachineInstrBuilder &MIB,
+ X86AddressMode &AM) {
+ // First constrain the index register. It needs to be a GR64_NOSP.
+ AM.IndexReg = constrainOperandRegClass(MIB->getDesc(), AM.IndexReg,
+ MIB->getNumOperands() +
+ X86::AddrIndexReg);
+ return ::addFullAddress(MIB, AM);
+}
+
/// \brief Check if it is possible to fold the condition from the XALU intrinsic
/// into the user. The condition code will only be updated on success.
bool X86FastISel::foldX86XALUIntrinsic(X86::CondCode &CC, const Instruction *I,
/// X86FastEmitLoad - Emit a machine instruction to load a value of type VT.
/// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV.
/// Return true and the result register by reference if it is possible.
-bool X86FastISel::X86FastEmitLoad(EVT VT, const X86AddressMode &AM,
- MachineMemOperand *MMO, unsigned &ResultReg) {
+bool X86FastISel::X86FastEmitLoad(EVT VT, X86AddressMode &AM,
+ MachineMemOperand *MMO, unsigned &ResultReg,
+ unsigned Alignment) {
// Get opcode and regclass of the output for the given load instruction.
unsigned Opc = 0;
const TargetRegisterClass *RC = nullptr;
case MVT::f80:
// No f80 support yet.
return false;
+ case MVT::v4f32:
+ if (Alignment >= 16)
+ Opc = Subtarget->hasAVX() ? X86::VMOVAPSrm : X86::MOVAPSrm;
+ else
+ Opc = Subtarget->hasAVX() ? X86::VMOVUPSrm : X86::MOVUPSrm;
+ RC = &X86::VR128RegClass;
+ break;
+ case MVT::v2f64:
+ if (Alignment >= 16)
+ Opc = Subtarget->hasAVX() ? X86::VMOVAPDrm : X86::MOVAPDrm;
+ else
+ Opc = Subtarget->hasAVX() ? X86::VMOVUPDrm : X86::MOVUPDrm;
+ RC = &X86::VR128RegClass;
+ break;
+ case MVT::v4i32:
+ case MVT::v2i64:
+ case MVT::v8i16:
+ case MVT::v16i8:
+ if (Alignment >= 16)
+ Opc = Subtarget->hasAVX() ? X86::VMOVDQArm : X86::MOVDQArm;
+ else
+ Opc = Subtarget->hasAVX() ? X86::VMOVDQUrm : X86::MOVDQUrm;
+ RC = &X86::VR128RegClass;
+ break;
}
ResultReg = createResultReg(RC);
/// and a displacement offset, or a GlobalAddress,
/// i.e. V. Return true if it is possible.
bool X86FastISel::X86FastEmitStore(EVT VT, unsigned ValReg, bool ValIsKill,
- const X86AddressMode &AM,
+ X86AddressMode &AM,
MachineMemOperand *MMO, bool Aligned) {
// Get opcode and regclass of the output for the given store instruction.
unsigned Opc = 0;
}
bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val,
- const X86AddressMode &AM,
+ X86AddressMode &AM,
MachineMemOperand *MMO, bool Aligned) {
// Handle 'null' like i32/i64 0.
if (isa<ConstantPointerNull>(Val))
if (!X86SelectAddress(Ptr, AM))
return false;
+ unsigned Alignment = LI->getAlignment();
+ unsigned ABIAlignment = DL.getABITypeAlignment(LI->getType());
+ if (Alignment == 0) // Ensure that codegen never sees alignment 0
+ Alignment = ABIAlignment;
+
unsigned ResultReg = 0;
- if (!X86FastEmitLoad(VT, AM, createMachineMemOperandFor(LI), ResultReg))
+ if (!X86FastEmitLoad(VT, AM, createMachineMemOperandFor(LI), ResultReg,
+ Alignment))
return false;
updateValueMap(I, ResultReg);
}
}
-/// X86ChooseCmpImmediateOpcode - If we have a comparison with RHS as the RHS
-/// of the comparison, return an opcode that works for the compare (e.g.
-/// CMP32ri) otherwise return 0.
+/// If we have a comparison with RHS as the RHS of the comparison, return an
+/// opcode that works for the compare (e.g. CMP32ri) otherwise return 0.
static unsigned X86ChooseCmpImmediateOpcode(EVT VT, const ConstantInt *RHSC) {
+ int64_t Val = RHSC->getSExtValue();
switch (VT.getSimpleVT().SimpleTy) {
// Otherwise, we can't fold the immediate into this comparison.
- default: return 0;
- case MVT::i8: return X86::CMP8ri;
- case MVT::i16: return X86::CMP16ri;
- case MVT::i32: return X86::CMP32ri;
+ default:
+ return 0;
+ case MVT::i8:
+ return X86::CMP8ri;
+ case MVT::i16:
+ if (isInt<8>(Val))
+ return X86::CMP16ri8;
+ return X86::CMP16ri;
+ case MVT::i32:
+ if (isInt<8>(Val))
+ return X86::CMP32ri8;
+ return X86::CMP32ri;
case MVT::i64:
+ if (isInt<8>(Val))
+ return X86::CMP64ri8;
// 64-bit comparisons are only valid if the immediate fits in a 32-bit sext
// field.
- if ((int)RHSC->getSExtValue() == RHSC->getSExtValue())
+ if (isInt<32>(Val))
return X86::CMP64ri32;
return 0;
}
return true;
}
-/// \brief Emit SSE instructions to lower the select.
+/// \brief Emit SSE or AVX instructions to lower the select.
///
/// Try to use SSE1/SSE2 instructions to simulate a select without branches.
/// This lowers fp selects into a CMP/AND/ANDN/OR sequence when the necessary
-/// SSE instructions are available.
+/// SSE instructions are available. If AVX is available, try to use a VBLENDV.
bool X86FastISel::X86FastEmitSSESelect(MVT RetVT, const Instruction *I) {
// Optimize conditions coming from a compare if both instructions are in the
// same basic block (values defined in other basic blocks may not have
if (NeedSwap)
std::swap(CmpLHS, CmpRHS);
- static unsigned OpcTable[2][2][4] = {
- { { X86::CMPSSrr, X86::FsANDPSrr, X86::FsANDNPSrr, X86::FsORPSrr },
- { X86::VCMPSSrr, X86::VFsANDPSrr, X86::VFsANDNPSrr, X86::VFsORPSrr } },
- { { X86::CMPSDrr, X86::FsANDPDrr, X86::FsANDNPDrr, X86::FsORPDrr },
- { X86::VCMPSDrr, X86::VFsANDPDrr, X86::VFsANDNPDrr, X86::VFsORPDrr } }
+ // Choose the SSE instruction sequence based on data type (float or double).
+ static unsigned OpcTable[2][4] = {
+ { X86::CMPSSrr, X86::FsANDPSrr, X86::FsANDNPSrr, X86::FsORPSrr },
+ { X86::CMPSDrr, X86::FsANDPDrr, X86::FsANDNPDrr, X86::FsORPDrr }
};
- bool HasAVX = Subtarget->hasAVX();
unsigned *Opc = nullptr;
switch (RetVT.SimpleTy) {
default: return false;
- case MVT::f32: Opc = &OpcTable[0][HasAVX][0]; break;
- case MVT::f64: Opc = &OpcTable[1][HasAVX][0]; break;
+ case MVT::f32: Opc = &OpcTable[0][0]; break;
+ case MVT::f64: Opc = &OpcTable[1][0]; break;
}
const Value *LHS = I->getOperand(1);
return false;
const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT);
- unsigned CmpReg = fastEmitInst_rri(Opc[0], RC, CmpLHSReg, CmpLHSIsKill,
- CmpRHSReg, CmpRHSIsKill, CC);
- unsigned AndReg = fastEmitInst_rr(Opc[1], RC, CmpReg, /*IsKill=*/false,
- LHSReg, LHSIsKill);
- unsigned AndNReg = fastEmitInst_rr(Opc[2], RC, CmpReg, /*IsKill=*/true,
- RHSReg, RHSIsKill);
- unsigned ResultReg = fastEmitInst_rr(Opc[3], RC, AndNReg, /*IsKill=*/true,
- AndReg, /*IsKill=*/true);
+ unsigned ResultReg;
+
+ if (Subtarget->hasAVX()) {
+ // If we have AVX, create 1 blendv instead of 3 logic instructions.
+ // Blendv was introduced with SSE 4.1, but the 2 register form implicitly
+ // uses XMM0 as the selection register. That may need just as many
+ // instructions as the AND/ANDN/OR sequence due to register moves, so
+ // don't bother.
+ unsigned CmpOpcode =
+ (RetVT.SimpleTy == MVT::f32) ? X86::VCMPSSrr : X86::VCMPSDrr;
+ unsigned BlendOpcode =
+ (RetVT.SimpleTy == MVT::f32) ? X86::VBLENDVPSrr : X86::VBLENDVPDrr;
+
+ unsigned CmpReg = fastEmitInst_rri(CmpOpcode, RC, CmpLHSReg, CmpLHSIsKill,
+ CmpRHSReg, CmpRHSIsKill, CC);
+ ResultReg = fastEmitInst_rrr(BlendOpcode, RC, RHSReg, RHSIsKill,
+ LHSReg, LHSIsKill, CmpReg, true);
+ } else {
+ unsigned CmpReg = fastEmitInst_rri(Opc[0], RC, CmpLHSReg, CmpLHSIsKill,
+ CmpRHSReg, CmpRHSIsKill, CC);
+ unsigned AndReg = fastEmitInst_rr(Opc[1], RC, CmpReg, /*IsKill=*/false,
+ LHSReg, LHSIsKill);
+ unsigned AndNReg = fastEmitInst_rr(Opc[2], RC, CmpReg, /*IsKill=*/true,
+ RHSReg, RHSIsKill);
+ ResultReg = fastEmitInst_rr(Opc[3], RC, AndNReg, /*IsKill=*/true,
+ AndReg, /*IsKill=*/true);
+ }
updateValueMap(I, ResultReg);
return true;
}
}
bool X86FastISel::X86SelectSIToFP(const Instruction *I) {
+ // The target-independent selection algorithm in FastISel already knows how
+ // to select a SINT_TO_FP if the target is SSE but not AVX.
+ // Early exit if the subtarget doesn't have AVX.
+ if (!Subtarget->hasAVX())
+ return false;
+
if (!I->getOperand(0)->getType()->isIntegerTy(32))
return false;
if (OpReg == 0)
return false;
- bool HasAVX = Subtarget->hasAVX();
const TargetRegisterClass *RC = nullptr;
unsigned Opcode;
- if (I->getType()->isDoubleTy() && X86ScalarSSEf64) {
+ if (I->getType()->isDoubleTy()) {
// sitofp int -> double
- Opcode = HasAVX ? X86::VCVTSI2SDrr : X86::CVTSI2SDrr;
+ Opcode = X86::VCVTSI2SDrr;
RC = &X86::FR64RegClass;
- } else if (I->getType()->isFloatTy() && X86ScalarSSEf32) {
+ } else if (I->getType()->isFloatTy()) {
// sitofp int -> float
- Opcode = HasAVX ? X86::VCVTSI2SSrr : X86::CVTSI2SSrr;
+ Opcode = X86::VCVTSI2SSrr;
RC = &X86::FR32RegClass;
} else
return false;
-
- unsigned ImplicitDefReg = 0;
- if (HasAVX) {
- ImplicitDefReg = createResultReg(RC);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
- TII.get(TargetOpcode::IMPLICIT_DEF), ImplicitDefReg);
- }
-
- const MCInstrDesc &II = TII.get(Opcode);
- OpReg = constrainOperandRegClass(II, OpReg, (HasAVX ? 2 : 1));
-
- unsigned ResultReg = createResultReg(RC);
- MachineInstrBuilder MIB;
- MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg);
- if (ImplicitDefReg)
- MIB.addReg(ImplicitDefReg, RegState::Kill);
- MIB.addReg(OpReg);
+ unsigned ImplicitDefReg = createResultReg(RC);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::IMPLICIT_DEF), ImplicitDefReg);
+ unsigned ResultReg =
+ fastEmitInst_rr(Opcode, RC, ImplicitDefReg, true, OpReg, false);
updateValueMap(I, ResultReg);
return true;
}
return true;
}
+ bool KillInputReg = false;
if (!Subtarget->is64Bit()) {
// If we're on x86-32; we can't extract an i8 from a general register.
// First issue a copy to GR16_ABCD or GR32_ABCD.
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(TargetOpcode::COPY), CopyReg).addReg(InputReg);
InputReg = CopyReg;
+ KillInputReg = true;
}
// Issue an extract_subreg.
unsigned ResultReg = fastEmitInst_extractsubreg(MVT::i8,
- InputReg, /*Kill=*/true,
+ InputReg, KillInputReg,
X86::sub_8bit);
if (!ResultReg)
return false;
default: return false;
case Intrinsic::convert_from_fp16:
case Intrinsic::convert_to_fp16: {
- if (TM.Options.UseSoftFloat || !Subtarget->hasF16C())
+ if (Subtarget->useSoftFloat() || !Subtarget->hasF16C())
return false;
const Value *Op = II->getArgOperand(0);
const MCInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
// FIXME may need to add RegState::Debug to any registers produced,
// although ESP/EBP should be the only ones at the moment.
+ assert(DI->getVariable()->isValidLocationForIntrinsic(DbgLoc) &&
+ "Expected inlined-at fields to agree");
addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II), AM)
.addImm(0)
.addMetadata(DI->getVariable())
// Add a register mask operand representing the call-preserved registers.
// Proper defs for return values will be added by setPhysRegsDeadExcept().
- MIB.addRegMask(TRI.getCallPreservedMask(CC));
+ MIB.addRegMask(TRI.getCallPreservedMask(*FuncInfo.MF, CC));
// Add an implicit use GOT pointer in EBX.
if (Subtarget->isPICStyleGOT())
if (!Result)
return false;
+ // The index register could be in the wrong register class. Unfortunately,
+ // foldMemoryOperandImpl could have commuted the instruction so its not enough
+ // to just look at OpNo + the offset to the index reg. We actually need to
+ // scan the instruction to find the index reg and see if its the correct reg
+ // class.
+ unsigned OperandNo = 0;
+ for (MachineInstr::mop_iterator I = Result->operands_begin(),
+ E = Result->operands_end(); I != E; ++I, ++OperandNo) {
+ MachineOperand &MO = *I;
+ if (!MO.isReg() || MO.isDef() || MO.getReg() != AM.IndexReg)
+ continue;
+ // Found the index reg, now try to rewrite it.
+ unsigned IndexReg = constrainOperandRegClass(Result->getDesc(),
+ MO.getReg(), OperandNo);
+ if (IndexReg == MO.getReg())
+ continue;
+ MO.setReg(IndexReg);
+ }
+
Result->addMemOperand(*FuncInfo.MF, createMachineMemOperandFor(LI));
FuncInfo.MBB->insert(FuncInfo.InsertPt, Result);
MI->eraseFromParent();