[X86] EltsFromConsecutiveLoads - only attempt to recognise reverse patterns from all load element patterns

[RKSimon]

Noticed while investigating a regression reported after #168706 - don't attempt this if any elements are zero/undef - there is a potential issue with reverse(insert_subvector(undef,load(),0)) patterns - but I haven't managed to create a fail case.

[llvmbot]

@llvm/pr-subscribers-backend-x86

Author: Simon Pilgrim (RKSimon)

Changes

Noticed while investigating a regression reported after #168706 - don't attempt this if any elements are zero/undef - there is a potential issue with reverse(insert_subvector(undef,load(),0)) patterns - but I haven't managed to create a fail case.

---

Full diff: https://github.com/llvm/llvm-project/pull/170852.diff

1 Files Affected:

• (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+2-1)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 2722321a03e2e..d0ae75b2e6785 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -7571,7 +7571,8 @@ static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
 
   // REVERSE - attempt to match the loads in reverse and then shuffle back.
   // TODO: Do this for any permute or mismatching element counts.
-  if (Depth == 0 && !ZeroMask && TLI.isTypeLegal(VT) && VT.isVector() &&
+  if (Depth == 0 && ZeroMask.isZero() && UndefMask.isZero() &&
+      TLI.isTypeLegal(VT) && VT.isVector() &&
       NumElems == VT.getVectorNumElements()) {
     SmallVector<SDValue, 16> ReverseElts(Elts.rbegin(), Elts.rend());
     if (SDValue RevLd = EltsFromConsecutiveLoads(

[llvm-ci]

LLVM Buildbot has detected a new failure on builder ppc64le-mlir-rhel-clang running on ppc64le-mlir-rhel-test while building llvm at step 6 "test-build-check-mlir-build-only-check-mlir".

Full details are available at: https://lab.llvm.org/buildbot/#/builders/129/builds/34451

Here is the relevant piece of the build log for the reference

Step 6 (test-build-check-mlir-build-only-check-mlir) failure: 1200 seconds without output running [b'ninja', b'check-mlir'], attempting to kill
...
PASS: MLIR-Unit :: IR/./MLIRIRTests/101/130 (3713 of 3724)
PASS: MLIR :: mlir-tblgen/llvm-intrinsics.td (3714 of 3724)
PASS: MLIR-Unit :: IR/./MLIRIRTests/0/130 (3715 of 3724)
PASS: MLIR-Unit :: IR/./MLIRIRTests/39/130 (3716 of 3724)
PASS: MLIR-Unit :: IR/./MLIRIRTests/38/130 (3717 of 3724)
PASS: MLIR-Unit :: Interfaces/./MLIRInterfacesTests/11/22 (3718 of 3724)
PASS: MLIR-Unit :: Interfaces/./MLIRInterfacesTests/13/22 (3719 of 3724)
PASS: MLIR-Unit :: Pass/./MLIRPassTests/10/13 (3720 of 3724)
PASS: MLIR-Unit :: Interfaces/./MLIRInterfacesTests/12/22 (3721 of 3724)
PASS: MLIR :: mlir-reduce/dce-test.mlir (3722 of 3724)
command timed out: 1200 seconds without output running [b'ninja', b'check-mlir'], attempting to kill
process killed by signal 9
program finished with exit code -1
elapsedTime=3026.984528