[llvm] [VPlan] Make VPInstruction::AnyOf poison-safe. (PR #154156)

[David Sherwood via llvm-commits]

david-arm wrote:

Isn't there still a bug here? I think the freezes for any-of are only partially fixing the problem, since there is the issue of outside use of the compares too. Consider

```
vector.body:                                      ; preds = %vector.body, %vector.ph
...
  %6 = icmp ne <4 x i32> %4, zeroinitializer
  %7 = icmp ne <4 x i32> %5, zeroinitializer
...
  %8 = freeze <4 x i1> %6
  %9 = freeze <4 x i1> %7
  %10 = or <4 x i1> %8, %9
  %11 = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> %10)
...

vector.early.exit:                                ; preds = %middle.split
...
  %idx1 = call i64 @llvm.experimental.cttz.elts.i64.v4i1(<4 x i1> %6, i1 true)
  %extract1 = extractelement <4 x i32> %some_val1, i32 %idx1
...
  %idx2 = call i64 @llvm.experimental.cttz.elts.i64.v4i1(<4 x i1> %6, i1 true)
  %extract2 = extractelement <4 x i32> %some_val2, i32 %idx2
```

We're calculating the lane to extract multiple times for each outside use, so if we don't freeze `%6` we're potentially going to get different lane indices each time if `%6` happens to be something like <0, 0, poison, 0>. 

https://github.com/llvm/llvm-project/pull/154156