From 92f4a2bf0a9d3962a25d15c1df5b6824950c74dc Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Fri, 4 Oct 2024 18:09:53 +0100
Subject: [PATCH] [x86] combineMUL - when looking for a vector multiply by
 splat constant, ensure we're only accepting ConstantInt splat scalars.

Fixes #111170

(cherry picked from commit 9459d729d22b7bfedad9d3a4237162077c6984a4)
---
 llvm/lib/Target/X86/X86ISelLowering.cpp |  3 ++-
 llvm/test/CodeGen/X86/pr111170.ll       | 33 +++++++++++++++++++++++++
 2 files changed, 35 insertions(+), 1 deletion(-)
 create mode 100644 llvm/test/CodeGen/X86/pr111170.ll

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 10f269f803778..4e3a181f9a3ae 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -47899,7 +47899,8 @@ static SDValue combineMul(SDNode *N, SelectionDAG &DAG,
     if (VT.isVector())
       if (auto *RawC = getTargetConstantFromNode(N->getOperand(1)))
         if (auto *SplatC = RawC->getSplatValue())
-          C = &(SplatC->getUniqueInteger());
+          if (auto *SplatCI = dyn_cast<ConstantInt>(SplatC))
+            C = &(SplatCI->getValue());
 
     if (!C || C->getBitWidth() != VT.getScalarSizeInBits())
       return SDValue();
diff --git a/llvm/test/CodeGen/X86/pr111170.ll b/llvm/test/CodeGen/X86/pr111170.ll
new file mode 100644
index 0000000000000..145bf7119edcb
--- /dev/null
+++ b/llvm/test/CodeGen/X86/pr111170.ll
@@ -0,0 +1,33 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -mtriple=i686-pc-windows-msvc -mcpu=corei7-avx | FileCheck %s
+
+define void @PR111170(<16 x i32> %x_load, ptr %offsetsPtr.i) {
+; CHECK-LABEL: PR111170:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; CHECK-NEXT:    vbroadcastss {{.*#+}} xmm2 = [2.80259693E-44,2.80259693E-44,2.80259693E-44,2.80259693E-44]
+; CHECK-NEXT:    vpmulld %xmm2, %xmm1, %xmm3
+; CHECK-NEXT:    vextractf128 $1, %ymm1, %xmm1
+; CHECK-NEXT:    vpmulld %xmm2, %xmm1, %xmm1
+; CHECK-NEXT:    vpmulld %xmm2, %xmm0, %xmm4
+; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; CHECK-NEXT:    vpmulld %xmm2, %xmm0, %xmm0
+; CHECK-NEXT:    vmovdqu %xmm0, 16(%eax)
+; CHECK-NEXT:    vmovdqu %xmm4, (%eax)
+; CHECK-NEXT:    vmovdqu %xmm1, 48(%eax)
+; CHECK-NEXT:    vmovdqu %xmm3, 32(%eax)
+; CHECK-NEXT:    vpxor %xmm0, %xmm0, %xmm0
+; CHECK-NEXT:    vmovdqu %xmm0, 16
+; CHECK-NEXT:    vmovdqu %xmm0, 0
+; CHECK-NEXT:    vmovdqu %xmm0, 48
+; CHECK-NEXT:    vmovdqu %xmm0, 32
+; CHECK-NEXT:    vzeroupper
+; CHECK-NEXT:    retl
+  %mul__x_load = mul <16 x i32> <i32 20, i32 20, i32 20, i32 20, i32 20, i32 20, i32 20, i32 20, i32 20, i32 20, i32 20, i32 20, i32 20, i32 20, i32 20, i32 20>, %x_load
+  store <16 x i32> %mul__x_load, ptr %offsetsPtr.i, align 4
+  %blend1.i12.i = call <8 x float> @llvm.x86.avx.blendv.ps.256(<8 x float> zeroinitializer, <8 x float> <float 0x36E4000000000000, float 0x36E4000000000000, float 0x36E4000000000000, float 0x36E4000000000000, float 0x36E4000000000000, float 0x36E4000000000000, float 0x36E4000000000000, float 0x36E4000000000000>, <8 x float> zeroinitializer)
+  %blend.i13.i = shufflevector <8 x float> zeroinitializer, <8 x float> %blend1.i12.i, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %blendAsInt.i14.i = bitcast <16 x float> %blend.i13.i to <16 x i32>
+  store <16 x i32> %blendAsInt.i14.i, ptr null, align 4
+  ret void
+}