[PATCH] D142093: [InstCombine] trunc (fptoui|fptosi)

[Eli Friedman via Phabricator via llvm-commits]

efriedma added inline comments.


================
Comment at: llvm/test/Transforms/InstCombine/trunc-fp-to-int.ll:357
+; CHECK-NEXT:    [[CONV:%.*]] = fptosi float [[X:%.*]] to i33
+; CHECK-NEXT:    ret i33 [[CONV]]
 ;
----------------
>From alive2:

```
define i33 @src(float noundef %x) {
%0:
  %conv = fptosi float noundef %x to i64
  %conv.1 = trunc i64 %conv to i33
  ret i33 %conv.1
}
=>
define i33 @tgt(float noundef %x) {
%0:
  %conv = fptosi float noundef %x to i33
  ret i33 %conv
}
Transformation doesn't verify!

ERROR: Target is more poisonous than source
```

The final result type must be able to hold the largest finite number representable in the floating-point type; otherwise, the transform isn't legal.  For float, that's an i128 or i129, I think?

AArch64 does have an instruction FJCVTZS you could theoretically use for this kind of thing, but that seems unlikely to be worthwhile.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D142093/new/

https://reviews.llvm.org/D142093