[x86] make sure horizontal op and broadcast types match to simplify (PR41414)

If the types don't match, we can't just remove the shuffle.
There may be some other opportunity for optimization here,
but this should prevent the crashing seen in:
https://bugs.llvm.org/show_bug.cgi?id=41414

llvm-svn: 359095

2 files changed, 71 insertions, 3 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index d1e0f690913..f7c91ba0e65 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -32820,10 +32820,13 @@ static SDValue foldShuffleOfHorizOp(SDNode *N) {
 
   // For a broadcast, peek through an extract element of index 0 to find the
   // horizontal op: broadcast (ext_vec_elt HOp, 0)
+  EVT VT = N->getValueType(0);
   if (Opcode == X86ISD::VBROADCAST) {
     SDValue SrcOp = N->getOperand(0);
     if (SrcOp.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
-        SrcOp.getValueType() == MVT::f64 && isNullConstant(SrcOp.getOperand(1)))
+        SrcOp.getValueType() == MVT::f64 &&
+        SrcOp.getOperand(0).getValueType() == VT &&
+        isNullConstant(SrcOp.getOperand(1)))
       N = SrcOp.getNode();
   }
 
@@ -32847,7 +32850,8 @@ static SDValue foldShuffleOfHorizOp(SDNode *N) {
     // movddup (hadd X, X) --> hadd X, X
     // broadcast (extract_vec_elt (hadd X, X), 0) --> hadd X, X
     assert((HOp.getValueType() == MVT::v2f64 ||
-            HOp.getValueType() == MVT::v4f64) && "Unexpected type for h-op");
+            HOp.getValueType() == MVT::v4f64) && HOp.getValueType() == VT &&
+           "Unexpected type for h-op");
     return HOp;
   }
 
diff --git a/llvm/test/CodeGen/X86/haddsub-3.ll b/llvm/test/CodeGen/X86/haddsub-3.ll
index b1406fc3417..29b9a626dd2 100644
--- a/llvm/test/CodeGen/X86/haddsub-3.ll
+++ b/llvm/test/CodeGen/X86/haddsub-3.ll
@@ -1,7 +1,8 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse2 | FileCheck %s --check-prefixes=SSE,SSE2
 ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+ssse3 | FileCheck %s --check-prefixes=SSE,SSSE3
-; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX
+; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1
+; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2
 
 define float @pr26491(<4 x float> %a0) {
 ; SSE2-LABEL: pr26491:
@@ -37,3 +38,66 @@ define float @pr26491(<4 x float> %a0) {
   %5 = fadd float %3, %4
   ret float %5
 }
+
+; When simplifying away a splat (broadcast), the hop type must match the shuffle type.
+
+define <4 x double> @PR41414(i64 %x, <4 x double> %y) {
+; SSE2-LABEL: PR41414:
+; SSE2:       # %bb.0:
+; SSE2-NEXT:    movq %rdi, %xmm2
+; SSE2-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1]
+; SSE2-NEXT:    subpd {{.*}}(%rip), %xmm2
+; SSE2-NEXT:    movapd %xmm2, %xmm3
+; SSE2-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm2[1]
+; SSE2-NEXT:    addpd %xmm2, %xmm3
+; SSE2-NEXT:    unpcklpd {{.*#+}} xmm3 = xmm3[0,0]
+; SSE2-NEXT:    divpd %xmm3, %xmm1
+; SSE2-NEXT:    divpd %xmm3, %xmm0
+; SSE2-NEXT:    xorpd %xmm2, %xmm2
+; SSE2-NEXT:    addpd %xmm2, %xmm0
+; SSE2-NEXT:    addpd %xmm2, %xmm1
+; SSE2-NEXT:    retq
+;
+; SSSE3-LABEL: PR41414:
+; SSSE3:       # %bb.0:
+; SSSE3-NEXT:    movq %rdi, %xmm2
+; SSSE3-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1]
+; SSSE3-NEXT:    subpd {{.*}}(%rip), %xmm2
+; SSSE3-NEXT:    haddpd %xmm2, %xmm2
+; SSSE3-NEXT:    divpd %xmm2, %xmm1
+; SSSE3-NEXT:    divpd %xmm2, %xmm0
+; SSSE3-NEXT:    xorpd %xmm2, %xmm2
+; SSSE3-NEXT:    addpd %xmm2, %xmm0
+; SSSE3-NEXT:    addpd %xmm2, %xmm1
+; SSSE3-NEXT:    retq
+;
+; AVX1-LABEL: PR41414:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vmovq %rdi, %xmm1
+; AVX1-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
+; AVX1-NEXT:    vsubpd {{.*}}(%rip), %xmm1, %xmm1
+; AVX1-NEXT:    vhaddpd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm1, %ymm1
+; AVX1-NEXT:    vdivpd %ymm1, %ymm0, %ymm0
+; AVX1-NEXT:    vxorpd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vaddpd %ymm1, %ymm0, %ymm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: PR41414:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vmovq %rdi, %xmm1
+; AVX2-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
+; AVX2-NEXT:    vsubpd {{.*}}(%rip), %xmm1, %xmm1
+; AVX2-NEXT:    vhaddpd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vbroadcastsd %xmm1, %ymm1
+; AVX2-NEXT:    vdivpd %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vxorpd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vaddpd %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    retq
+  %conv = uitofp i64 %x to double
+  %t0 = insertelement <4 x double> undef, double %conv, i32 0
+  %t1 = shufflevector <4 x double> %t0, <4 x double> undef, <4 x i32> zeroinitializer
+  %t2 = fdiv <4 x double> %y, %t1
+  %t3 = fadd <4 x double> zeroinitializer, %t2
+  ret <4 x double> %t3
+}