llvm-beanz wrote: Quick sanity check question for @bogner:
In DXC we apply the `fast` flag, which in LLVM 3.7 implies: `nnan ninf nsz arcp` and includes the modern interpretation of `reassoc` With this change we are applying: `reassoc nnan ninf nsz arcp afn` (no `fast` because its meaning is redefined. The main addition is `afn`. > Approximate functions - Allow substitution of approximate calculations for > functions (sin, log, sqrt, etc). See floating-point intrinsic definitions for > places where this can apply to LLVM’s intrinsic math functions. [source](https://llvm.org/docs/LangRef.html#fast-math-flags) Two questions: 1) Do we foresee any issue adding `afn` (I suspect not) 2) Do we also need to apply `contract`, which I believe was also implied by `fast` in 3.7 although the docs are unclear? Also, I think we may need to put implement the backend support for translating the fast math flags to 3.7-equivalents before we merge this. I'm sorry I hadn't thought of that until now, but I expect this change breaks DXIL generation. https://github.com/llvm/llvm-project/pull/119820 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
