| Issue |
242941
|
| Summary |
[KnownFPClass] Refine fma, fmuladd, and unfused_fma
|
| Labels |
floating-point
|
| Assignees |
|
| Reporter |
ZERICO2005
|
Tracking issue for the types of refinements we can make for `fma`/`fmuladd`/`unfused_fma` without the magnitude refinement for `KnownFPClass` https://github.com/llvm/llvm-project/issues/240038
## fma
- [ ] Rule out infinity for common cases
- [ ] Rule out infinity for subnormal special cases for `Dynamic`
- [ ] Rule out infinity for subnormal special cases for `NearestTiesToEven`
- [ ] Rule out infinity for `TowardZero`, `TowardPositive`, `TowardNegative`
- [ ] Rule out a result of zero or subnormal.
- [ ] Rule out exact zero for `NearestTiesToEven`.
- [ ] Rule out exact zero for directed rounding modes.
- [ ] Handle special cases where at least one argument is non-finite (result is always inf or NaN, but I don't think it is worth doing this).
## fmuladd/unfused_fma
- [ ] `subnormal * normal + normal` won't overflow under `NearestTiesToEven`. There will be a more useful form in the future where `fmuladd(<=1.0, finite, <=1.0)` won't overflow for `NearestTiesToEven` (or use `ConstantFPRange`).
- [ ] `subnormal * normal + subnormal` won't overflow under `Dynamic`.
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs