From 24ad2b5e7d8ceecb045bfc499107b9ed90439adb Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Wed, 31 Jul 2019 12:55:39 +0000
Subject: [PATCH] [X86][AVX] Ensure chained subvector insertions are the same
 size (PR42833)

Before combining insert_subvector(insert_subvector(vec, sub0, c0), sub1, c1) patterns, ensure that the subvectors are all the same type. On AVX512 targets especially we might have a mixture of 128/256 subvector insertions.

llvm-svn: 367429
---
 llvm/lib/Target/X86/X86ISelLowering.cpp |  2 ++
 llvm/test/CodeGen/X86/oddsubvector.ll   | 20 +++++++++++---------
 2 files changed, 13 insertions(+), 9 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 300fc3e..352f2ed 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -5538,6 +5538,7 @@ static bool collectConcatOps(SDNode *N, SmallVectorImpl<SDValue> &Ops) {
     if (VT.getSizeInBits() == (SubVT.getSizeInBits() * 2) &&
         Idx == (VT.getVectorNumElements() / 2) &&
         Src.getOpcode() == ISD::INSERT_SUBVECTOR &&
+        Src.getOperand(1).getValueType() == SubVT &&
         isNullConstant(Src.getOperand(2))) {
       Ops.push_back(Src.getOperand(1));
       Ops.push_back(Sub);
@@ -43978,6 +43979,7 @@ static SDValue combineInsertSubvector(SDNode *N, SelectionDAG &DAG,
       Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
       OpVT.getSizeInBits() == SubVecVT.getSizeInBits() * 2 &&
       isNullConstant(Vec.getOperand(2)) && !Vec.getOperand(0).isUndef() &&
+      Vec.getOperand(1).getValueSizeInBits() == SubVecVT.getSizeInBits() &&
       Vec.hasOneUse()) {
     Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, DAG.getUNDEF(OpVT),
                       Vec.getOperand(1), Vec.getOperand(2));
diff --git a/llvm/test/CodeGen/X86/oddsubvector.ll b/llvm/test/CodeGen/X86/oddsubvector.ll
index d38f21e..5bcfbe6 100644
--- a/llvm/test/CodeGen/X86/oddsubvector.ll
+++ b/llvm/test/CodeGen/X86/oddsubvector.ll
@@ -349,19 +349,21 @@ define void @PR42833() {
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    movl {{.*}}(%rip), %eax
 ; AVX512-NEXT:    vmovdqu c+{{.*}}(%rip), %ymm0
+; AVX512-NEXT:    vmovdqu64 c+{{.*}}(%rip), %zmm1
 ; AVX512-NEXT:    addl c+{{.*}}(%rip), %eax
-; AVX512-NEXT:    vmovd %eax, %xmm1
-; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],mem[1,2,3,4,5,6,7]
-; AVX512-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
-; AVX512-NEXT:    vpsllvd %ymm1, %ymm0, %ymm0
-; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1,2,3,4,5,6,7]
-; AVX512-NEXT:    vmovdqa c+{{.*}}(%rip), %xmm1
+; AVX512-NEXT:    vmovd %eax, %xmm2
+; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],mem[1,2,3,4,5,6,7]
+; AVX512-NEXT:    vpaddd %ymm2, %ymm0, %ymm3
+; AVX512-NEXT:    vpsllvd %ymm2, %ymm0, %ymm0
+; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0],ymm0[1,2,3,4,5,6,7]
+; AVX512-NEXT:    vmovdqa c+{{.*}}(%rip), %xmm2
 ; AVX512-NEXT:    vmovdqu %ymm0, c+{{.*}}(%rip)
 ; AVX512-NEXT:    vmovdqu c+{{.*}}(%rip), %ymm0
-; AVX512-NEXT:    vmovdqu64 d+{{.*}}(%rip), %zmm2
-; AVX512-NEXT:    vpinsrd $0, %eax, %xmm1, %xmm1
+; AVX512-NEXT:    vmovdqu64 d+{{.*}}(%rip), %zmm3
+; AVX512-NEXT:    vpinsrd $0, %eax, %xmm2, %xmm2
+; AVX512-NEXT:    vinserti32x4 $0, %xmm2, %zmm1, %zmm1
 ; AVX512-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm1
-; AVX512-NEXT:    vpsubd %zmm1, %zmm2, %zmm1
+; AVX512-NEXT:    vpsubd %zmm1, %zmm3, %zmm1
 ; AVX512-NEXT:    vmovdqu64 %zmm1, d+{{.*}}(%rip)
 ; AVX512-NEXT:    vpaddd %ymm0, %ymm0, %ymm0
 ; AVX512-NEXT:    vmovdqu %ymm0, c+{{.*}}(%rip)
-- 
2.7.4