[AVX-512] Fix DecodeVPERMV3Mask to handle cases where the constant pool entry has a different type than the shuffle itself. This is especially important for 32-bit targets with 64-bit shuffle elements.

[craig.topper]

Summary

This is similar to how PSHUFB and VPERMIL handle the same problem.

[RKSimon]

LGTM

lib/Target/X86/X86ISelLowering.cpp

case X86ISD::VPERMV: {
     return false;
   }
   case X86ISD::VPERMV3: {
     IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(2);
     // Unlike most shuffle nodes, VPERMV3's mask operand is the middle one.
     Ops.push_back(N->getOperand(0));
     Ops.push_back(N->getOperand(2));
     SDValue MaskNode = N->getOperand(1);
+    unsigned MaskEltSize = VT.getScalarSizeInBits();
     if (auto *C = getTargetConstantFromNode(MaskNode)) {
-      DecodeVPERMV3Mask(C, VT, Mask);
+      DecodeVPERMV3Mask(C, MaskEltSize, Mask);
       break;
     }
     return false;
   }
   case X86ISD::VPERMIV3: {
     IsUnary = IsFakeUnary = N->getOperand(1) == N->getOperand(2);
     // Unlike most shuffle nodes, VPERMIV3's mask operand is the first one.
     Ops.push_back(N->getOperand(1));
     Ops.push_back(N->getOperand(2));
     SDValue MaskNode = N->getOperand(0);
+    unsigned MaskEltSize = VT.getScalarSizeInBits();
     if (auto *C = getTargetConstantFromNode(MaskNode)) {
-      DecodeVPERMV3Mask(C, VT, Mask);
+      DecodeVPERMV3Mask(C, MaskEltSize, Mask);
       break;
     }
     return false;
   }
   default: llvm_unreachable("unknown target shuffle node");
   }
 
   // Empty mask indicates the decode failed.

lib/Target/X86/X86ShuffleDecodeConstantPool.h

 /// Decode a VPPERM variable mask from an IR-level vector constant.
 void DecodeVPPERMMask(const Constant *C, SmallVectorImpl<int> &ShuffleMask);
 
 /// Decode a VPERM W/D/Q/PS/PD mask from an IR-level vector constant.
 void DecodeVPERMVMask(const Constant *C, MVT VT,
                       SmallVectorImpl<int> &ShuffleMask);
 
 /// Decode a VPERMT2 W/D/Q/PS/PD mask from an IR-level vector constant.
-void DecodeVPERMV3Mask(const Constant *C, MVT VT,
+void DecodeVPERMV3Mask(const Constant *C, unsigned ElSize,
                        SmallVectorImpl<int> &ShuffleMask);
 
 } // llvm namespace
 
 #endif

lib/Target/X86/X86ShuffleDecodeConstantPool.cpp

if (!isa<ConstantInt>(C))
     return;
   uint64_t Element = cast<ConstantInt>(C)->getZExtValue();
   int NumElements = VT.getVectorNumElements();
   Element &= (1 << NumElements) - 1;
   for (int i = 0; i < NumElements; ++i)
     ShuffleMask.push_back(Element);
 }
 
-void DecodeVPERMV3Mask(const Constant *C, MVT VT,
+void DecodeVPERMV3Mask(const Constant *C, unsigned ElSize,
                        SmallVectorImpl<int> &ShuffleMask) {
   Type *MaskTy = C->getType();
-  unsigned NumElements = MaskTy->getVectorNumElements();
-  if (NumElements == VT.getVectorNumElements()) {
-    unsigned EltMaskSize = Log2_64(NumElements * 2);
-    for (unsigned i = 0; i < NumElements; ++i) {
-      Constant *COp = C->getAggregateElement(i);
-      if (!COp) {
-        ShuffleMask.clear();
-        return;
-      }
-      if (isa<UndefValue>(COp))
-        ShuffleMask.push_back(SM_SentinelUndef);
-      else {
-        APInt Element = cast<ConstantInt>(COp)->getValue();
-        Element = Element.getLoBits(EltMaskSize);
-        ShuffleMask.push_back(Element.getZExtValue());
-      }
+  unsigned MaskTySize = MaskTy->getPrimitiveSizeInBits();
+  (void)MaskTySize;
+  assert((MaskTySize == 128 || MaskTySize == 256 || MaskTySize == 512) &&
+         "Unexpected vector size.");
+  assert((ElSize == 8 || ElSize == 16 || ElSize == 32 || ElSize == 64) &&
+         "Unexpected vector element size.");
+
+  // The shuffle mask requires elements the same size as the target.
+  SmallBitVector UndefElts;
+  SmallVector<uint64_t, 8> RawMask;
+  if (!extractConstantMask(C, ElSize, UndefElts, RawMask))
+
+  unsigned NumElts = RawMask.size();
+
+  for (unsigned i = 0; i != NumElts; ++i) {
+    if (UndefElts[i]) {
+      continue;
     }
+    int Index = RawMask[i] & (NumElts*2 - 1);
+    ShuffleMask.push_back(Index);
   }
 }
 } // llvm namespace

test/CodeGen/X86/vector-shuffle-combining-avx512bw.ll

; X64-NEXT:    retq
   %res0 = call <8 x i64> @llvm.x86.avx512.mask.permvar.di.512(<8 x i64> %x0, <8 x i64> <i64 7, i64 6, i64 5, i64 4, i64 3, i64 2, i64 1, i64 0>, <8 x i64> %x1, i8 %m)
   %res1 = call <8 x i64> @llvm.x86.avx512.mask.permvar.di.512(<8 x i64> %res0, <8 x i64> <i64 7, i64 14, i64 5, i64 12, i64 3, i64 10, i64 1, i64 8>, <8 x i64> %res0, i8 %m)
   ret <8 x i64> %res1
 }
 
 define <8 x double> @combine_vpermt2var_8f64_identity(<8 x double> %x0, <8 x double> %x1) {
 ; X32-LABEL: combine_vpermt2var_8f64_identity:
 ; X32:       # BB#0:
-; X32-NEXT:    vmovapd {{.*#+}} zmm2 = [7,0,6,0,5,0,4,0,3,0,2,0,1,0,0,0]
-; X32-NEXT:    vpermt2pd %zmm1, %zmm2, %zmm0
-; X32-NEXT:    vmovapd {{.*#+}} zmm1 = [7,0,14,0,5,0,12,0,3,0,10,0,1,0,8,0]
-; X32-NEXT:    vpermt2pd %zmm0, %zmm1, %zmm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: combine_vpermt2var_8f64_identity:
 ; X64:       # BB#0:
 ; X64-NEXT:    retq
   %res0 = call <8 x double> @llvm.x86.avx512.maskz.vpermt2var.pd.512(<8 x i64> <i64 7, i64 6, i64 5, i64 4, i64 3, i64 2, i64 1, i64 0>, <8 x double> %x0, <8 x double> %x1, i8 -1)
   %res1 = call <8 x double> @llvm.x86.avx512.maskz.vpermt2var.pd.512(<8 x i64> <i64 7, i64 14, i64 5, i64 12, i64 3, i64 10, i64 1, i64 8>, <8 x double> %res0, <8 x double> %res0, i8 -1)
   ret <8 x double> %res1
; X64-NEXT:    retq
   %res0 = call <8 x double> @llvm.x86.avx512.maskz.vpermt2var.pd.512(<8 x i64> <i64 7, i64 6, i64 5, i64 4, i64 3, i64 2, i64 1, i64 0>, <8 x double> %x0, <8 x double> %x1, i8 %m)
   %res1 = call <8 x double> @llvm.x86.avx512.maskz.vpermt2var.pd.512(<8 x i64> <i64 7, i64 14, i64 5, i64 12, i64 3, i64 10, i64 1, i64 8>, <8 x double> %res0, <8 x double> %res0, i8 %m)
   ret <8 x double> %res1
 }
 
 define <8 x double> @combine_vpermt2var_8f64_movddup(<8 x double> %x0, <8 x double> %x1) {
 ; X32-LABEL: combine_vpermt2var_8f64_movddup:
 ; X32:       # BB#0:
-; X32-NEXT:    vmovapd {{.*#+}} zmm2 = <0,0,0,0,2,0,2,0,4,0,4,0,u,u,u,u>
-; X32-NEXT:    vpermt2pd %zmm1, %zmm2, %zmm0
+; X32-NEXT:    vmovddup {{.*#+}} zmm0 = zmm0[0,0,2,2,4,4,6,6]
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: combine_vpermt2var_8f64_movddup:
 ; X64:       # BB#0:
 ; X64-NEXT:    vmovddup {{.*#+}} zmm0 = zmm0[0,0,2,2,4,4,6,6]
 ; X64-NEXT:    retq
   %res0 = call <8 x double> @llvm.x86.avx512.maskz.vpermt2var.pd.512(<8 x i64> <i64 0, i64 0, i64 2, i64 2, i64 4, i64 4, i64 undef, i64 undef>, <8 x double> %x0, <8 x double> %x1, i8 -1)
   ret <8 x double> %res0
 }
 define <8 x double> @combine_vpermt2var_8f64_movddup_load(<8 x double> *%p0, <8 x double> %x1) {
 ; X32-LABEL: combine_vpermt2var_8f64_movddup_load:
 ; X32:       # BB#0:
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-NEXT:    vmovapd (%eax), %zmm1
-; X32-NEXT:    vmovapd {{.*#+}} zmm2 = [0,0,0,0,2,0,2,0,4,0,4,0,6,0,6,0]
-; X32-NEXT:    vpermt2pd %zmm0, %zmm2, %zmm1
-; X32-NEXT:    vmovapd %zmm1, %zmm0
+; X32-NEXT:    vmovddup {{.*#+}} zmm0 = mem[0,0,2,2,4,4,6,6]
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: combine_vpermt2var_8f64_movddup_load:
 ; X64:       # BB#0:
 ; X64-NEXT:    vmovddup {{.*#+}} zmm0 = mem[0,0,2,2,4,4,6,6]
 ; X64-NEXT:    retq
   %x0 = load <8 x double>, <8 x double> *%p0
   %res0 = call <8 x double> @llvm.x86.avx512.maskz.vpermt2var.pd.512(<8 x i64> <i64 0, i64 0, i64 2, i64 2, i64 4, i64 4, i64 6, i64 6>, <8 x double> %x0, <8 x double> %x1, i8 -1)
   ret <8 x double> %res0
 }
 define <8 x double> @combine_vpermt2var_8f64_movddup_mask(<8 x double> %x0, <8 x double> %x1, i8 %m) {
 ; X32-LABEL: combine_vpermt2var_8f64_movddup_mask:
 ; X32:       # BB#0:
 ; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
 ; X32-NEXT:    kmovd %eax, %k1
-; X32-NEXT:    vmovapd {{.*#+}} zmm2 = [0,0,0,0,2,0,2,0,4,0,4,0,6,0,6,0]
-; X32-NEXT:    vpermt2pd %zmm1, %zmm2, %zmm0 {%k1} {z}
+; X32-NEXT:    vmovddup {{.*#+}} zmm0 {%k1} {z} = zmm0[0,0,2,2,4,4,6,6]
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: combine_vpermt2var_8f64_movddup_mask:
 ; X64:       # BB#0:
 ; X64-NEXT:    kmovw %edi, %k1
 ; X64-NEXT:    vmovddup {{.*#+}} zmm0 {%k1} {z} = zmm0[0,0,2,2,4,4,6,6]
 ; X64-NEXT:    retq
   %res0 = call <8 x double> @llvm.x86.avx512.maskz.vpermt2var.pd.512(<8 x i64> <i64 0, i64 0, i64 2, i64 2, i64 4, i64 4, i64 6, i64 6>, <8 x double> %x0, <8 x double> %x1, i8 %m)
; X64-NEXT:    retq
   %res0 = call <32 x i16> @llvm.x86.avx512.mask.permvar.hi.512(<32 x i16> %a0, <32 x i16> <i16 1, i16 0, i16 3, i16 2, i16 4, i16 5, i16 6, i16 7, i16 9, i16 8, i16 11, i16 10, i16 12, i16 13, i16 14, i16 15, i16 17, i16 16, i16 19, i16 18, i16 20, i16 21, i16 22, i16 23, i16 25, i16 24, i16 27, i16 26, i16 28, i16 29, i16 30, i16 31>, <32 x i16> undef, i32 -1)
   %res1 = call <32 x i16> @llvm.x86.avx512.mask.permvar.hi.512(<32 x i16> %res0, <32 x i16> <i16 0, i16 1, i16 2, i16 3, i16 5, i16 4, i16 7, i16 6, i16 8, i16 9, i16 10, i16 11, i16 13, i16 12, i16 15, i16 14, i16 16, i16 17, i16 18, i16 19, i16 21, i16 20, i16 23, i16 22, i16 24, i16 25, i16 26, i16 27, i16 29, i16 28, i16 31, i16 30>, <32 x i16> undef, i32 -1)
   ret <32 x i16> %res1
 }
 
 define <8 x double> @combine_vpermi2var_8f64_identity(<8 x double> %x0, <8 x double> %x1) {
 ; X32-LABEL: combine_vpermi2var_8f64_identity:
 ; X32:       # BB#0:
-; X32-NEXT:    vmovapd {{.*#+}} zmm2 = [7,0,6,0,5,0,4,0,3,0,2,0,1,0,0,0]
-; X32-NEXT:    vpermi2pd %zmm1, %zmm0, %zmm2
-; X32-NEXT:    vmovapd {{.*#+}} zmm0 = [7,0,14,0,5,0,12,0,3,0,10,0,1,0,8,0]
-; X32-NEXT:    vpermi2pd %zmm2, %zmm2, %zmm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: combine_vpermi2var_8f64_identity:
 ; X64:       # BB#0:
 ; X64-NEXT:    retq
   %res0 = call <8 x double> @llvm.x86.avx512.mask.vpermi2var.pd.512(<8 x double> %x0, <8 x i64> <i64 7, i64 6, i64 5, i64 4, i64 3, i64 2, i64 1, i64 0>, <8 x double> %x1, i8 -1)
   %res1 = call <8 x double> @llvm.x86.avx512.mask.vpermi2var.pd.512(<8 x double> %res0, <8 x i64> <i64 7, i64 14, i64 5, i64 12, i64 3, i64 10, i64 1, i64 8>, <8 x double> %res0, i8 -1)
   ret <8 x double> %res1