[PATCH] D103378: [VectorCombine] Freeze index unless it is known to be non-poison.

[Eli Friedman via Phabricator via llvm-commits]

efriedma added inline comments.


================
Comment at: llvm/test/Transforms/VectorCombine/AArch64/load-extractelement-scalarization.ll:153
+; CHECK-NEXT:    [[R:%.*]] = load i32, i32* [[TMP1]], align 4
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
----------------
fhahn wrote:
> efriedma wrote:
> > This is still not right.  See https://alive2.llvm.org/ce/z/dqg8HF
> Thanks Eli, I thought I ran Alive2 on the test after the changes but something went wrong. I limited the transform to non-poison indices only for now in 96ca03493ae5.
> 
> Not sure if there's much else we can do to make this save using freeze, if we cannot rely on the constrained range.
The following is legal, in general:

```
define i32 @src(<4 x i32>* %x, i4 %idx) {
entry:
  %lv = load <4 x i32>, <4 x i32>* %x
  %r = extractelement <4 x i32> %lv, i4 %idx
  ret i32 %r
}

define i32 @tgt(<4 x i32>* %x, i4 %idx) {
entry:
  %idx.freeze = freeze i4 %idx
  %idx.clamped = and i4 %idx.freeze, 3
  %gep = getelementptr inbounds <4 x i32>, <4 x i32>* %x, i32 0, i4 %idx.clamped
  %r = load i32, i32* %gep
  ret i32 %r
}
```

I haven't thought about when it's profitable.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D103378/new/

https://reviews.llvm.org/D103378