[ARM] Only change mask if demanded bits says we can optimize

[AZero13]

This prevents overcorrection to the mask, especially if it is valid already.

[llvmbot]

@llvm/pr-subscribers-backend-arm

Author: AZero13 (AZero13)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/165106.diff

1 Files Affected:

• (modified) llvm/lib/Target/ARM/ARMISelLowering.cpp (+1-1)

diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index 313ae3d68fb83..9b7955eda8f8d 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -20178,7 +20178,7 @@ bool ARMTargetLowering::targetShrinkDemandedConstant(
   // code won't do this, so we have to do it explicitly to avoid an infinite
   // loop in obscure cases.)
   if (ExpandedMask == ~0U)
-    return TLO.CombineTo(Op, Op.getOperand(0));
+    return false;
 
   auto IsLegalMask = [ShrunkMask, ExpandedMask](unsigned Mask) -> bool {
     return (ShrunkMask & Mask) == ShrunkMask && (~ExpandedMask & Mask) == 0;

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[AZero13]

This is the best I can do without regressions @davemgreen

To remove the workarounds I have to improve the ubfx code which involves porting getUsefulBits from AArch64, which is a very large effort.

[AZero13]

@rj-jesus Thoughts on this?

[AZero13]

@arsenm Here.