[PATCH] D91709: [X86][AVX] Only share broadcasts of different widths from the same SDValue of the same SDNode (PR48215)

[Simon Pilgrim via Phabricator via llvm-commits]

RKSimon created this revision.
RKSimon added reviewers: craig.topper, spatel, pengfei, wristow.
Herald added a subscriber: hiraditya.
Herald added a project: LLVM.
RKSimon requested review of this revision.

D57663 <https://reviews.llvm.org/D57663> allowed us to reuse broadcasts of the same scalar value by extracting low subvectors from the widest type.

Unfortunately we weren't ensuring the broadcasts were from the same SDValue, just the same SDNode - which failed on multiple-value nodes like ISD::SDIVREM

FYI: I intend to request this be merged into the 11.x release branch.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D91709

Files:
  llvm/lib/Target/X86/X86ISelLowering.cpp
  llvm/test/CodeGen/X86/pr48215.ll


Index: llvm/test/CodeGen/X86/pr48215.ll
===================================================================
--- llvm/test/CodeGen/X86/pr48215.ll
+++ llvm/test/CodeGen/X86/pr48215.ll
@@ -33,12 +33,14 @@
 ; AVX2-NEXT:    idivl %esi
 ; AVX2-NEXT:    vmovd %eax, %xmm0
 ; AVX2-NEXT:    vpbroadcastd %xmm0, %ymm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [0,1,2,3,4,5,6,7]
+; AVX2-NEXT:    vmovd %edx, %xmm1
+; AVX2-NEXT:    vpbroadcastd %xmm1, %xmm1
+; AVX2-NEXT:    vmovdqa {{.*#+}} ymm2 = [0,1,2,3,4,5,6,7]
+; AVX2-NEXT:    vpcmpgtd %ymm0, %ymm2, %ymm0
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm2 = [4,5,6,7]
-; AVX2-NEXT:    vpcmpgtd %ymm0, %ymm1, %ymm1
-; AVX2-NEXT:    vmovmskps %ymm1, %ecx
-; AVX2-NEXT:    vpcmpgtd %xmm0, %xmm2, %xmm0
-; AVX2-NEXT:    vmovmskps %xmm0, %eax
+; AVX2-NEXT:    vpcmpgtd %xmm1, %xmm2, %xmm1
+; AVX2-NEXT:    vmovmskps %ymm0, %ecx
+; AVX2-NEXT:    vmovmskps %xmm1, %eax
 ; AVX2-NEXT:    addl %ecx, %eax
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
@@ -49,8 +51,9 @@
 ; AVX512-NEXT:    cltd
 ; AVX512-NEXT:    idivl %esi
 ; AVX512-NEXT:    vpbroadcastd %eax, %ymm0
+; AVX512-NEXT:    vpbroadcastd %edx, %xmm1
 ; AVX512-NEXT:    vpcmpltd {{.*}}(%rip), %ymm0, %k0
-; AVX512-NEXT:    vpcmpltd {{.*}}(%rip), %xmm0, %k1
+; AVX512-NEXT:    vpcmpltd {{.*}}(%rip), %xmm1, %k1
 ; AVX512-NEXT:    kmovw %k0, %eax
 ; AVX512-NEXT:    movzbl %al, %ecx
 ; AVX512-NEXT:    kmovw %k1, %eax
Index: llvm/lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- llvm/lib/Target/X86/X86ISelLowering.cpp
+++ llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -36715,8 +36715,10 @@
       return DAG.getNode(X86ISD::VBROADCAST, DL, VT, Src.getOperand(0));
 
     // Share broadcast with the longest vector and extract low subvector (free).
+    // Ensure the same SDValue from the SDNode use is being used.
     for (SDNode *User : Src->uses())
       if (User != N.getNode() && User->getOpcode() == X86ISD::VBROADCAST &&
+          Src == User->getOperand(0) &&
           User->getValueSizeInBits(0).getFixedSize() >
               VT.getFixedSizeInBits()) {
         return extractSubVector(SDValue(User, 0), 0, DAG, DL,


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D91709.306098.patch
Type: text/x-patch
Size: 2173 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20201118/94b8f542/attachment.bin>