RKSimon wrote: > > fp reductions are a nightmare - every time I thought we were getting > > somewhere, something else fastmath related causes more headaches. > > @RKSimon should I not proceed with this PR? I was really hoping to use this > in HLSL as it makes implementing many of our runtime apis easy in the > frontend.
I'd definitely suggest you reduce the scope - maybe always expand the reduction serially in the frontend and not create the llvm intrinsic at all? Or maybe an alternative `__builtin_reduce_fastmath_fadd / fmul` that makes it clear whats happening, and have them always emit the reassoc variants of the reductions? There was a lot of resistance to reductions that change behaviour depending on which fast math flags were enabled. CC @andykaylor @jcranmer-intel https://github.com/llvm/llvm-project/pull/120367 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
