aboutsummaryrefslogtreecommitdiff
path: root/lib/Target/X86/X86ISelLowering.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'lib/Target/X86/X86ISelLowering.cpp')
-rw-r--r--lib/Target/X86/X86ISelLowering.cpp100
1 files changed, 93 insertions, 7 deletions
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index bdfe245027..ffaf04cea7 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -522,6 +522,10 @@ X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
+ setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i64, Custom);
+ setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i64, Custom);
+ setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i64, Custom);
+ setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i64, Custom);
}
if (Subtarget->hasCmpxchg16b()) {
@@ -1357,7 +1361,7 @@ X86TargetLowering::getOptimalMemOpType(uint64_t Size,
// cases like PR2962. This should be removed when PR2962 is fixed.
const Function *F = MF.getFunction();
if (IsZeroVal &&
- !F->hasFnAttr(Attribute::NoImplicitFloat)) {
+ !F->getFnAttributes().hasNoImplicitFloatAttr()) {
if (Size >= 16 &&
(Subtarget->isUnalignedMemAccessFast() ||
((DstAlign == 0 || DstAlign >= 16) &&
@@ -2048,7 +2052,7 @@ X86TargetLowering::LowerFormalArguments(SDValue Chain,
unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
TotalNumIntRegs);
- bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
+ bool NoImplicitFloatOps = Fn->getFnAttributes().hasNoImplicitFloatAttr();
assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
"SSE register cannot be used when SSE is disabled!");
assert(!(NumXMMRegs && MF.getTarget().Options.UseSoftFloat &&
@@ -2240,7 +2244,7 @@ X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
// Check if it's really possible to do a tail call.
isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
isVarArg, SR != NotStructReturn,
- MF.getFunction()->hasStructRetAttr(),
+ MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
Outs, OutVals, Ins, DAG);
// Sibcalls are automatically detected tailcalls which do not require
@@ -2524,7 +2528,7 @@ X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
OpFlags = X86II::MO_DARWIN_STUB;
} else if (Subtarget->isPICStyleRIPRel() &&
isa<Function>(GV) &&
- cast<Function>(GV)->hasFnAttr(Attribute::NonLazyBind)) {
+ cast<Function>(GV)->getFnAttributes().hasNonLazyBindAttr()) {
// If the function is marked as non-lazy, generate an indirect call
// which loads from the GOT directly. This avoids runtime overhead
// at the cost of eager binding (and one extra byte of encoding).
@@ -2761,6 +2765,7 @@ X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
bool isVarArg,
bool isCalleeStructRet,
bool isCallerStructRet,
+ Type *RetTy,
const SmallVectorImpl<ISD::OutputArg> &Outs,
const SmallVectorImpl<SDValue> &OutVals,
const SmallVectorImpl<ISD::InputArg> &Ins,
@@ -2772,6 +2777,13 @@ X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
// If -tailcallopt is specified, make fastcc functions tail-callable.
const MachineFunction &MF = DAG.getMachineFunction();
const Function *CallerF = DAG.getMachineFunction().getFunction();
+
+ // If the function return type is x86_fp80 and the callee return type is not,
+ // then the FP_EXTEND of the call result is not a nop. It's not safe to
+ // perform a tailcall optimization here.
+ if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
+ return false;
+
CallingConv::ID CallerCC = CallerF->getCallingConv();
bool CCMatch = CallerCC == CalleeCC;
@@ -6661,7 +6673,7 @@ X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
bool HasAVX = Subtarget->hasAVX();
bool HasAVX2 = Subtarget->hasAVX2();
MachineFunction &MF = DAG.getMachineFunction();
- bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
+ bool OptForSize = MF.getFunction()->getFnAttributes().hasOptimizeForSizeAttr();
assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
@@ -9783,7 +9795,7 @@ SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
// Sanity Check: Make sure using fp_offset makes sense.
assert(!getTargetMachine().Options.UseSoftFloat &&
!(DAG.getMachineFunction()
- .getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
+ .getFunction()->getFnAttributes().hasNoImplicitFloatAttr()) &&
Subtarget->hasSSE1());
}
@@ -11769,6 +11781,10 @@ void X86TargetLowering::ReplaceNodeResults(SDNode *N,
case ISD::ATOMIC_LOAD_OR:
case ISD::ATOMIC_LOAD_SUB:
case ISD::ATOMIC_LOAD_XOR:
+ case ISD::ATOMIC_LOAD_MAX:
+ case ISD::ATOMIC_LOAD_MIN:
+ case ISD::ATOMIC_LOAD_UMAX:
+ case ISD::ATOMIC_LOAD_UMIN:
case ISD::ATOMIC_SWAP: {
unsigned Opc;
switch (N->getOpcode()) {
@@ -11791,6 +11807,18 @@ void X86TargetLowering::ReplaceNodeResults(SDNode *N,
case ISD::ATOMIC_LOAD_XOR:
Opc = X86ISD::ATOMXOR64_DAG;
break;
+ case ISD::ATOMIC_LOAD_MAX:
+ Opc = X86ISD::ATOMMAX64_DAG;
+ break;
+ case ISD::ATOMIC_LOAD_MIN:
+ Opc = X86ISD::ATOMMIN64_DAG;
+ break;
+ case ISD::ATOMIC_LOAD_UMAX:
+ Opc = X86ISD::ATOMUMAX64_DAG;
+ break;
+ case ISD::ATOMIC_LOAD_UMIN:
+ Opc = X86ISD::ATOMUMIN64_DAG;
+ break;
case ISD::ATOMIC_SWAP:
Opc = X86ISD::ATOMSWAP64_DAG;
break;
@@ -12182,6 +12210,10 @@ static unsigned getNonAtomic6432Opcode(unsigned Opc, unsigned &HiOpc) {
case X86::ATOMADD6432: HiOpc = X86::ADC32rr; return X86::ADD32rr;
case X86::ATOMSUB6432: HiOpc = X86::SBB32rr; return X86::SUB32rr;
case X86::ATOMSWAP6432: HiOpc = X86::MOV32rr; return X86::MOV32rr;
+ case X86::ATOMMAX6432: HiOpc = X86::SETLr; return X86::SETLr;
+ case X86::ATOMMIN6432: HiOpc = X86::SETGr; return X86::SETGr;
+ case X86::ATOMUMAX6432: HiOpc = X86::SETBr; return X86::SETBr;
+ case X86::ATOMUMIN6432: HiOpc = X86::SETAr; return X86::SETAr;
}
llvm_unreachable("Unhandled atomic-load-op opcode!");
}
@@ -12499,6 +12531,7 @@ X86TargetLowering::EmitAtomicLoadArith6432(MachineInstr *MI,
SrcHiReg = MI->getOperand(CurOp++).getReg();
const TargetRegisterClass *RC = &X86::GR32RegClass;
+ const TargetRegisterClass *RC8 = &X86::GR8RegClass;
unsigned LCMPXCHGOpc = X86::LCMPXCHG8B;
unsigned LOADOpc = X86::MOV32rm;
@@ -12586,6 +12619,55 @@ X86TargetLowering::EmitAtomicLoadArith6432(MachineInstr *MI,
BuildMI(mainMBB, DL, TII->get(NOTOpc), t1H).addReg(t2H);
break;
}
+ case X86::ATOMMAX6432:
+ case X86::ATOMMIN6432:
+ case X86::ATOMUMAX6432:
+ case X86::ATOMUMIN6432: {
+ unsigned HiOpc;
+ unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
+ unsigned cL = MRI.createVirtualRegister(RC8);
+ unsigned cH = MRI.createVirtualRegister(RC8);
+ unsigned cL32 = MRI.createVirtualRegister(RC);
+ unsigned cH32 = MRI.createVirtualRegister(RC);
+ unsigned cc = MRI.createVirtualRegister(RC);
+ // cl := cmp src_lo, lo
+ BuildMI(mainMBB, DL, TII->get(X86::CMP32rr))
+ .addReg(SrcLoReg).addReg(LoReg);
+ BuildMI(mainMBB, DL, TII->get(LoOpc), cL);
+ BuildMI(mainMBB, DL, TII->get(X86::MOVZX32rr8), cL32).addReg(cL);
+ // ch := cmp src_hi, hi
+ BuildMI(mainMBB, DL, TII->get(X86::CMP32rr))
+ .addReg(SrcHiReg).addReg(HiReg);
+ BuildMI(mainMBB, DL, TII->get(HiOpc), cH);
+ BuildMI(mainMBB, DL, TII->get(X86::MOVZX32rr8), cH32).addReg(cH);
+ // cc := if (src_hi == hi) ? cl : ch;
+ if (Subtarget->hasCMov()) {
+ BuildMI(mainMBB, DL, TII->get(X86::CMOVE32rr), cc)
+ .addReg(cH32).addReg(cL32);
+ } else {
+ MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), cc)
+ .addReg(cH32).addReg(cL32)
+ .addImm(X86::COND_E);
+ mainMBB = EmitLoweredSelect(MIB, mainMBB);
+ }
+ BuildMI(mainMBB, DL, TII->get(X86::TEST32rr)).addReg(cc).addReg(cc);
+ if (Subtarget->hasCMov()) {
+ BuildMI(mainMBB, DL, TII->get(X86::CMOVNE32rr), t1L)
+ .addReg(SrcLoReg).addReg(LoReg);
+ BuildMI(mainMBB, DL, TII->get(X86::CMOVNE32rr), t1H)
+ .addReg(SrcHiReg).addReg(HiReg);
+ } else {
+ MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), t1L)
+ .addReg(SrcLoReg).addReg(LoReg)
+ .addImm(X86::COND_NE);
+ mainMBB = EmitLoweredSelect(MIB, mainMBB);
+ MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), t1H)
+ .addReg(SrcHiReg).addReg(HiReg)
+ .addImm(X86::COND_NE);
+ mainMBB = EmitLoweredSelect(MIB, mainMBB);
+ }
+ break;
+ }
case X86::ATOMSWAP6432: {
unsigned HiOpc;
unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
@@ -13576,6 +13658,10 @@ X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
case X86::ATOMNAND6432:
case X86::ATOMADD6432:
case X86::ATOMSUB6432:
+ case X86::ATOMMAX6432:
+ case X86::ATOMMIN6432:
+ case X86::ATOMUMAX6432:
+ case X86::ATOMUMIN6432:
case X86::ATOMSWAP6432:
return EmitAtomicLoadArith6432(MI, BB);
@@ -15562,7 +15648,7 @@ static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
return SDValue();
const Function *F = DAG.getMachineFunction().getFunction();
- bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
+ bool NoImplicitFloatOps = F->getFnAttributes().hasNoImplicitFloatAttr();
bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
&& Subtarget->hasSSE2();
if ((VT.isVector() ||
generated by cgit v1.2.3-18-g5258 (git 2.32.0) at 2025-07-23 03:01:53 +0000