[X86] Don't consider functions ABI compatible for ArgumentPromotion pass if they would legalize 512-bit vectors differently.

[craig.topper]

Summary

The use of the -mprefer-vector-width=256 command line option mixed with functions using vector intrinsics can create situations where one function thinks 512 vectors are legal, but another fucntion does not.

If a 512 bit vector is passed between them via a pointer, its possible ArgumentPromotion might try to pass by value instead. This will result in type legalization for the two functions handling the 512 bit vector differently leading to runtime failures.

Had the 512 bit vector been passed by value from clang codegen, both functions would have been tagged with a min-legal-vector-width=512 function attribute. That would make them be legalized the same way.

I observed this issue in 32-bit mode where a union containing a 512 bit vector was being passed by a function that used intrinsics to one that did not. The caller ended up passing in zmm0 and the callee tried to read it from ymm0 and ymm1.

The fix implemented here is just to consider it a mismatch if two functions would handle 512 bit differently without looking at the types that are being considered. This is the easist and safest fix, but it can be improved in the future.

[echristo]

lib/Target/X86/X86TargetTransformInfo.cpp

3085–3089

I'm guessing removing the single use local variables is much less readable than having them?

[craig.topper]

lib/Target/X86/X86TargetTransformInfo.cpp

3085–3089

clang-format gave me this

return static_cast<const X86Subtarget *>(TM.getSubtargetImpl(*Caller))
           ->useAVX512Regs() ==
       static_cast<const X86Subtarget *>(TM.getSubtargetImpl(*Callee))
           ->useAVX512Regs();

[echristo]

lib/Target/X86/X86TargetTransformInfo.cpp

3085–3089

Actually:

class X86TTIImpl : public BasicTTIImplBase<X86TTIImpl> {

typedef BasicTTIImplBase<X86TTIImpl> BaseT;
typedef TargetTransformInfo TTI;
friend BaseT;

const X86Subtarget *ST;
const X86TargetLowering *TLI;

The subtarget is already there :)

[craig.topper]

Use TM.getSubtarget<X86Subtarget> to shorten code.

lib/Target/X86/X86TargetTransformInfo.h

public:
   bool isLegalMaskedLoad(Type *DataType);
   bool isLegalMaskedStore(Type *DataType);
   bool isLegalMaskedGather(Type *DataType);
   bool isLegalMaskedScatter(Type *DataType);
   bool hasDivRemOp(Type *DataType, bool IsSigned);
   bool isFCmpOrdCheaperThanFCmpZero(Type *Ty);
   bool areInlineCompatible(const Function *Caller,
                            const Function *Callee) const;
+  bool areFunctionArgsABICompatible(const Function *Caller,
+                                    const Function *Callee,
+                                    SmallPtrSetImpl<Argument *> &Args) const;
   const TTI::MemCmpExpansionOptions *enableMemCmpExpansion(
       bool IsZeroCmp) const;
   bool enableInterleavedAccessVectorization();
 private:
   int getGSScalarCost(unsigned Opcode, Type *DataTy, bool VariableMask,
                       unsigned Alignment, unsigned AddressSpace);
   int getGSVectorCost(unsigned Opcode, Type *DataTy, Value *Ptr,
                       unsigned Alignment, unsigned AddressSpace);
 
   /// @}
 };
 
 } // end namespace llvm
 
 #endif

lib/Target/X86/X86TargetTransformInfo.cpp

bool X86TTIImpl::areInlineCompatible(const Function *Caller,
   const FeatureBitset &CalleeBits =
       TM.getSubtargetImpl(*Callee)->getFeatureBits();
 
   FeatureBitset RealCallerBits = CallerBits & ~InlineFeatureIgnoreList;
   FeatureBitset RealCalleeBits = CalleeBits & ~InlineFeatureIgnoreList;
   return (RealCallerBits & RealCalleeBits) == RealCalleeBits;
 }
 
+bool X86TTIImpl::areFunctionArgsABICompatible(
+    const Function *Caller, const Function *Callee,
+    SmallPtrSetImpl<Argument *> &Args) const {
+  if (!BaseT::areFunctionArgsABICompatible(Caller, Callee, Args))
+    return false;
+
+  // If we get here, we know the target features match. If one function
+  // considers 512-bit vectors legal and the other does not, consider them
+  // incompatible.
+  // FIXME Look at the arguments and only consider 512 bit or larger vectors?
+  const TargetMachine &TM = getTLI()->getTargetMachine();
+
+  return TM.getSubtarget<X86Subtarget>(*Caller).useAVX512Regs() ==
+         TM.getSubtarget<X86Subtarget>(*Callee).useAVX512Regs();
+}
+
 const X86TTIImpl::TTI::MemCmpExpansionOptions *
 X86TTIImpl::enableMemCmpExpansion(bool IsZeroCmp) const {
   // Only enable vector loads for equality comparison.
   // Right now the vector version is not as fast, see #33329.
   static const auto ThreeWayOptions = [this]() {
     TTI::MemCmpExpansionOptions Options;
     if (ST->is64Bit()) {
       Options.LoadSizes.push_back(8);
     }

test/Transforms/ArgumentPromotion/X86/min-legal-vector-width.ll

+; RUN: opt -S -argpromotion < %s | FileCheck %s
+; RUN: opt -S -passes=argpromotion < %s | FileCheck %s
+; Test that we only promote arguments when the caller/callee have compatible
+; function attrubtes.
+
+target triple = "x86_64-unknown-linux-gnu"
+
+; This should promote
+; CHECK-LABEL: @callee_avx512_legal512_prefer512_call_avx512_legal512_prefer512(<8 x i64>* %arg, <8 x i64> %arg1.val)
+define internal fastcc void @callee_avx512_legal512_prefer512_call_avx512_legal512_prefer512(<8 x i64>* %arg, <8 x i64>* readonly %arg1) #0 {
+bb:
+  %tmp = load <8 x i64>, <8 x i64>* %arg1
+  store <8 x i64> %tmp, <8 x i64>* %arg
+  ret void
+}
+
+define void @avx512_legal512_prefer512_call_avx512_legal512_prefer512(<8 x i64>* %arg) #0 {
+bb:
+  %tmp = alloca <8 x i64>, align 32
+  %tmp2 = alloca <8 x i64>, align 32
+  %tmp3 = bitcast <8 x i64>* %tmp to i8*
+  call void @llvm.memset.p0i8.i64(i8* align 32 %tmp3, i8 0, i64 32, i1 false)
+  call fastcc void @callee_avx512_legal512_prefer512_call_avx512_legal512_prefer512(<8 x i64>* %tmp2, <8 x i64>* %tmp)
+  %tmp4 = load <8 x i64>, <8 x i64>* %tmp2, align 32
+  store <8 x i64> %tmp4, <8 x i64>* %arg, align 2
+  ret void
+}
+
+; This should promote
+; CHECK-LABEL: @callee_avx512_legal512_prefer256_call_avx512_legal512_prefer256(<8 x i64>* %arg, <8 x i64> %arg1.val)
+define internal fastcc void @callee_avx512_legal512_prefer256_call_avx512_legal512_prefer256(<8 x i64>* %arg, <8 x i64>* readonly %arg1) #1 {
+bb:
+  %tmp = load <8 x i64>, <8 x i64>* %arg1
+  store <8 x i64> %tmp, <8 x i64>* %arg
+  ret void
+}
+
+define void @avx512_legal512_prefer256_call_avx512_legal512_prefer256(<8 x i64>* %arg) #1 {
+bb:
+  %tmp = alloca <8 x i64>, align 32
+  %tmp2 = alloca <8 x i64>, align 32
+  %tmp3 = bitcast <8 x i64>* %tmp to i8*
+  call void @llvm.memset.p0i8.i64(i8* align 32 %tmp3, i8 0, i64 32, i1 false)
+  call fastcc void @callee_avx512_legal512_prefer256_call_avx512_legal512_prefer256(<8 x i64>* %tmp2, <8 x i64>* %tmp)
+  %tmp4 = load <8 x i64>, <8 x i64>* %tmp2, align 32
+  store <8 x i64> %tmp4, <8 x i64>* %arg, align 2
+  ret void
+}
+
+; This should promote
+; CHECK-LABEL: @callee_avx512_legal512_prefer512_call_avx512_legal512_prefer256(<8 x i64>* %arg, <8 x i64> %arg1.val)
+define internal fastcc void @callee_avx512_legal512_prefer512_call_avx512_legal512_prefer256(<8 x i64>* %arg, <8 x i64>* readonly %arg1) #1 {
+bb:
+  %tmp = load <8 x i64>, <8 x i64>* %arg1
+  store <8 x i64> %tmp, <8 x i64>* %arg
+  ret void
+}
+
+define void @avx512_legal512_prefer512_call_avx512_legal512_prefer256(<8 x i64>* %arg) #0 {
+bb:
+  %tmp = alloca <8 x i64>, align 32
+  %tmp2 = alloca <8 x i64>, align 32
+  %tmp3 = bitcast <8 x i64>* %tmp to i8*
+  call void @llvm.memset.p0i8.i64(i8* align 32 %tmp3, i8 0, i64 32, i1 false)
+  call fastcc void @callee_avx512_legal512_prefer512_call_avx512_legal512_prefer256(<8 x i64>* %tmp2, <8 x i64>* %tmp)
+  %tmp4 = load <8 x i64>, <8 x i64>* %tmp2, align 32
+  store <8 x i64> %tmp4, <8 x i64>* %arg, align 2
+  ret void
+}
+
+; This should promote
+; CHECK-LABEL: @callee_avx512_legal512_prefer256_call_avx512_legal512_prefer512(<8 x i64>* %arg, <8 x i64> %arg1.val)
+define internal fastcc void @callee_avx512_legal512_prefer256_call_avx512_legal512_prefer512(<8 x i64>* %arg, <8 x i64>* readonly %arg1) #0 {
+bb:
+  %tmp = load <8 x i64>, <8 x i64>* %arg1
+  store <8 x i64> %tmp, <8 x i64>* %arg
+  ret void
+}
+
+define void @avx512_legal512_prefer256_call_avx512_legal512_prefer512(<8 x i64>* %arg) #1 {
+bb:
+  %tmp = alloca <8 x i64>, align 32
+  %tmp2 = alloca <8 x i64>, align 32
+  %tmp3 = bitcast <8 x i64>* %tmp to i8*
+  call void @llvm.memset.p0i8.i64(i8* align 32 %tmp3, i8 0, i64 32, i1 false)
+  call fastcc void @callee_avx512_legal512_prefer256_call_avx512_legal512_prefer512(<8 x i64>* %tmp2, <8 x i64>* %tmp)
+  %tmp4 = load <8 x i64>, <8 x i64>* %tmp2, align 32
+  store <8 x i64> %tmp4, <8 x i64>* %arg, align 2
+  ret void
+}
+
+; This should not promote
+; CHECK-LABEL: @callee_avx512_legal256_prefer256_call_avx512_legal512_prefer256(<8 x i64>* %arg, <8 x i64>* readonly %arg1)
+define internal fastcc void @callee_avx512_legal256_prefer256_call_avx512_legal512_prefer256(<8 x i64>* %arg, <8 x i64>* readonly %arg1) #1 {
+bb:
+  %tmp = load <8 x i64>, <8 x i64>* %arg1
+  store <8 x i64> %tmp, <8 x i64>* %arg
+  ret void
+}
+
+define void @avx512_legal256_prefer256_call_avx512_legal512_prefer256(<8 x i64>* %arg) #2 {
+bb:
+  %tmp = alloca <8 x i64>, align 32
+  %tmp2 = alloca <8 x i64>, align 32
+  %tmp3 = bitcast <8 x i64>* %tmp to i8*
+  call void @llvm.memset.p0i8.i64(i8* align 32 %tmp3, i8 0, i64 32, i1 false)
+  call fastcc void @callee_avx512_legal256_prefer256_call_avx512_legal512_prefer256(<8 x i64>* %tmp2, <8 x i64>* %tmp)
+  %tmp4 = load <8 x i64>, <8 x i64>* %tmp2, align 32
+  store <8 x i64> %tmp4, <8 x i64>* %arg, align 2
+  ret void
+}
+
+; This should not promote
+; CHECK-LABEL: @callee_avx512_legal512_prefer256_call_avx512_legal256_prefer256(<8 x i64>* %arg, <8 x i64>* readonly %arg1)
+define internal fastcc void @callee_avx512_legal512_prefer256_call_avx512_legal256_prefer256(<8 x i64>* %arg, <8 x i64>* readonly %arg1) #2 {
+bb:
+  %tmp = load <8 x i64>, <8 x i64>* %arg1
+  store <8 x i64> %tmp, <8 x i64>* %arg
+  ret void
+}
+
+define void @avx512_legal512_prefer256_call_avx512_legal256_prefer256(<8 x i64>* %arg) #1 {
+bb:
+  %tmp = alloca <8 x i64>, align 32
+  %tmp2 = alloca <8 x i64>, align 32
+  %tmp3 = bitcast <8 x i64>* %tmp to i8*
+  call void @llvm.memset.p0i8.i64(i8* align 32 %tmp3, i8 0, i64 32, i1 false)
+  call fastcc void @callee_avx512_legal512_prefer256_call_avx512_legal256_prefer256(<8 x i64>* %tmp2, <8 x i64>* %tmp)
+  %tmp4 = load <8 x i64>, <8 x i64>* %tmp2, align 32
+  store <8 x i64> %tmp4, <8 x i64>* %arg, align 2
+  ret void
+}
+
+; This should promote
+; CHECK-LABEL: @callee_avx2_legal256_prefer256_call_avx2_legal512_prefer256(<8 x i64>* %arg, <8 x i64> %arg1.val)
+define internal fastcc void @callee_avx2_legal256_prefer256_call_avx2_legal512_prefer256(<8 x i64>* %arg, <8 x i64>* readonly %arg1) #3 {
+bb:
+  %tmp = load <8 x i64>, <8 x i64>* %arg1
+  store <8 x i64> %tmp, <8 x i64>* %arg
+  ret void
+}
+
+define void @avx2_legal256_prefer256_call_avx2_legal512_prefer256(<8 x i64>* %arg) #4 {
+bb:
+  %tmp = alloca <8 x i64>, align 32
+  %tmp2 = alloca <8 x i64>, align 32
+  %tmp3 = bitcast <8 x i64>* %tmp to i8*
+  call void @llvm.memset.p0i8.i64(i8* align 32 %tmp3, i8 0, i64 32, i1 false)
+  call fastcc void @callee_avx2_legal256_prefer256_call_avx2_legal512_prefer256(<8 x i64>* %tmp2, <8 x i64>* %tmp)
+  %tmp4 = load <8 x i64>, <8 x i64>* %tmp2, align 32
+  store <8 x i64> %tmp4, <8 x i64>* %arg, align 2
+  ret void
+}
+
+; This should promote
+; CHECK-LABEL: @callee_avx2_legal512_prefer256_call_avx2_legal256_prefer256(<8 x i64>* %arg, <8 x i64> %arg1.val)
+define internal fastcc void @callee_avx2_legal512_prefer256_call_avx2_legal256_prefer256(<8 x i64>* %arg, <8 x i64>* readonly %arg1) #4 {
+bb:
+  %tmp = load <8 x i64>, <8 x i64>* %arg1
+  store <8 x i64> %tmp, <8 x i64>* %arg
+  ret void
+}
+
+define void @avx2_legal512_prefer256_call_avx2_legal256_prefer256(<8 x i64>* %arg) #3 {
+bb:
+  %tmp = alloca <8 x i64>, align 32
+  %tmp2 = alloca <8 x i64>, align 32
+  %tmp3 = bitcast <8 x i64>* %tmp to i8*
+  call void @llvm.memset.p0i8.i64(i8* align 32 %tmp3, i8 0, i64 32, i1 false)
+  call fastcc void @callee_avx2_legal512_prefer256_call_avx2_legal256_prefer256(<8 x i64>* %tmp2, <8 x i64>* %tmp)
+  %tmp4 = load <8 x i64>, <8 x i64>* %tmp2, align 32
+  store <8 x i64> %tmp4, <8 x i64>* %arg, align 2
+  ret void
+}
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.memset.p0i8.i64(i8* nocapture writeonly, i8, i64, i1) #5
+
+attributes #0 = { inlinehint norecurse nounwind uwtable "target-features"="+avx512vl" "min-legal-vector-width"="512" "prefer-vector-width"="512" }
+attributes #1 = { inlinehint norecurse nounwind uwtable "target-features"="+avx512vl" "min-legal-vector-width"="512" "prefer-vector-width"="256" }
+attributes #2 = { inlinehint norecurse nounwind uwtable "target-features"="+avx512vl" "min-legal-vector-width"="256" "prefer-vector-width"="256" }
+attributes #3 = { inlinehint norecurse nounwind uwtable "target-features"="+avx2" "min-legal-vector-width"="512" "prefer-vector-width"="256" }
+attributes #4 = { inlinehint norecurse nounwind uwtable "target-features"="+avx2" "min-legal-vector-width"="256" "prefer-vector-width"="256" }
+attributes #5 = { argmemonly nounwind }