The new C2 compiler support for Float16, or HalfFloat in the Ideal Graph, is missing some easy optimizations.
- Conversion folding: sequences of ConvertHF2F and ConvertF2HF are not eliminated.
- Float16 constants created with the idiom `Float16.valueOf(3.14f)` end up as a ConF in the ideal graph, preventing 16-bit subtraction in the case of `1 - (2 - f)` (see attached graph).
- Conversion folding: sequences of ConvertHF2F and ConvertF2HF are not eliminated.
- Float16 constants created with the idiom `Float16.valueOf(3.14f)` end up as a ConF in the ideal graph, preventing 16-bit subtraction in the case of `1 - (2 - f)` (see attached graph).
- duplicates
-
JDK-8352635 Improve inferencing of Float16 operations with constant inputs
-
- Resolved
-
- relates to
-
-
- Resolved
-