[X86][FP16] Replace vXi16 to vXf16 instead of v8f16

[pengfei]

Summary

Fixes pr52561

[RKSimon]

Which of the patterns that you changed is actually causing the crash and is it possible to create a test case for the other?

[LuoYuanke]

LGTM, pls address Simon's comments.

[pengfei]

In D114304#3144642, @RKSimon wrote:

Which of the patterns that you changed is actually causing the crash and is it possible to create a test case for the other?

The pattern is MaskVT == MVT::v16i16 and Mask = {0, -2, -2, ..., -2}. It matches the conditions in line 36273 and 36274. I tried several times but failed to generate Mask = {0, -1, -1, ..., -1} to match the condition in line 36220.
The conditions in line 36221 and 36222 are only avaliable for MaskVT == MVT::v8i16. So I didn't create another test cast.

In D114304#3156949, @LuoYuanke wrote:

LGTM, pls address Simon's comments.

Thanks!

llvm/lib/Target/X86/X86ISelLowering.cpp

static bool matchUnaryShuffle(MVT MaskVT, ArrayRef<int> Mask,
 
   // Match against a VZEXT_MOVL vXi32 and vXi16 zero-extending instruction.
   if (Mask[0] == 0 &&
       (MaskEltSize == 32 || (MaskEltSize == 16 && Subtarget.hasFP16()))) {
     if ((isUndefOrZero(Mask[1]) && isUndefInRange(Mask, 2, NumMaskElts - 2)) ||
         (V1.getOpcode() == ISD::SCALAR_TO_VECTOR &&
          isUndefOrZeroInRange(Mask, 1, NumMaskElts - 1))) {
       Shuffle = X86ISD::VZEXT_MOVL;
-      SrcVT = DstVT = MaskEltSize == 16      ? MVT::v8f16
-                      : !Subtarget.hasSSE2() ? MVT::v4f32
-                                             : MaskVT;
+      if (MaskEltSize == 16)
+        SrcVT = DstVT = MaskVT.changeVectorElementType(MVT::f16);
+      else
+        SrcVT = DstVT = !Subtarget.hasSSE2() ? MVT::v4f32 : MaskVT;
       return true;
     }
   }
 
   // Match against a ANY/ZERO_EXTEND_VECTOR_INREG instruction.
   // TODO: Add 512-bit vector support (split AVX512F and AVX512BW).
   if (AllowIntDomain && ((MaskVT.is128BitVector() && Subtarget.hasSSE41()) ||
                          (MaskVT.is256BitVector() && Subtarget.hasInt256()))) {
static bool matchUnaryShuffle(MVT MaskVT, ArrayRef<int> Mask,
   }
 
   // Match against a VZEXT_MOVL instruction, SSE1 only supports 32-bits (MOVSS).
   if (((MaskEltSize == 32) || (MaskEltSize == 64 && Subtarget.hasSSE2()) ||
        (MaskEltSize == 16 && Subtarget.hasFP16())) &&
       isUndefOrEqual(Mask[0], 0) &&
       isUndefOrZeroInRange(Mask, 1, NumMaskElts - 1)) {
     Shuffle = X86ISD::VZEXT_MOVL;
-    SrcVT = DstVT = MaskEltSize == 16      ? MVT::v8f16
-                    : !Subtarget.hasSSE2() ? MVT::v4f32
-                                           : MaskVT;
+    if (MaskEltSize == 16)
+      SrcVT = DstVT = MaskVT.changeVectorElementType(MVT::f16);
+    else
+      SrcVT = DstVT = !Subtarget.hasSSE2() ? MVT::v4f32 : MaskVT;
     return true;
   }
 
   // Check if we have SSE3 which will let us use MOVDDUP etc. The
   // instructions are no slower than UNPCKLPD but has the option to
   // fold the input operand into even an unaligned memory load.
   if (MaskVT.is128BitVector() && Subtarget.hasSSE3() && AllowFloatDomain) {
     if (isTargetShuffleEquivalent(MaskVT, Mask, {0, 0}, V1)) {

llvm/test/CodeGen/X86/avx512fp16-mov.ll

 
 for.body.preheader:                               ; preds = %entry
   store i8 0, i8* %c, align 1
   br label %for.end
 
 for.end:                                          ; preds = %for.body.preheader, %entry
   ret void
 }
+
+define <16 x i32> @pr52561(<16 x i32> %a, <16 x i32> %b) "min-legal-vector-width"="256" "prefer-vector-width"="256" nounwind {
+; X64-LABEL: pr52561:
+; X64:       # %bb.0:
+; X64-NEXT:    vpbroadcastd {{.*#+}} ymm4 = [112,112,112,112,112,112,112,112]
+; X64-NEXT:    vpaddd %ymm4, %ymm2, %ymm2
+; X64-NEXT:    vpaddd %ymm2, %ymm0, %ymm0
+; X64-NEXT:    vpaddd %ymm4, %ymm3, %ymm2
+; X64-NEXT:    vpaddd %ymm2, %ymm1, %ymm1
+; X64-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
+; X64-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; X64-NEXT:    vmovsh %xmm0, %xmm2, %xmm0
+; X64-NEXT:    retq
+;
+; X86-LABEL: pr52561:
+; X86:       # %bb.0:
+; X86-NEXT:    pushl %ebp
+; X86-NEXT:    movl %esp, %ebp
+; X86-NEXT:    andl $-32, %esp
+; X86-NEXT:    subl $32, %esp
+; X86-NEXT:    vpaddd 8(%ebp), %ymm1, %ymm1
+; X86-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [112,112,112,112,112,112,112,112]
+; X86-NEXT:    vpaddd %ymm3, %ymm2, %ymm2
+; X86-NEXT:    vpaddd %ymm2, %ymm0, %ymm0
+; X86-NEXT:    vpaddd %ymm3, %ymm1, %ymm1
+; X86-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}, %ymm1, %ymm1
+; X86-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; X86-NEXT:    vmovsh %xmm0, %xmm2, %xmm0
+; X86-NEXT:    movl %ebp, %esp
+; X86-NEXT:    popl %ebp
+; X86-NEXT:    retl
+  %1 = add <16 x i32> %a, <i32 112, i32 112, i32 112, i32 112, i32 112, i32 112, i32 112, i32 112, i32 112, i32 112, i32 112, i32 112, i32 112, i32 112, i32 112, i32 112>
+  %2 = add <16 x i32> %1, %b
+  %3 = and <16 x i32> %2, <i32 65535, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 65535>
+  ret <16 x i32> %3
+}