quinnp added inline comments.
================ Comment at: llvm/test/CodeGen/PowerPC/LowerCheckedFPArith.ll:36 +; CHECK-NEXT: %2 = fdiv fast float %0, %1 +; CHECK-NEXT: %3 = fcmp une float %2, %2 +; CHECK-NEXT: br i1 %3, label %swdiv_HWDIV, label %swdiv_MERGE ---------------- efriedma wrote: > quinnp wrote: > > efriedma wrote: > > > A "fast" fdiv never produces NaN, per LangRef. Using fcmp like this is > > > fragile at best. > > > > > > (Maybe you want "fdiv arcp"?) > > Thank you, I see what you mean. I have changed it to emit a `fdiv ninf > > arcp` instead of a `fdiv fast`. I included the `ninf` flag because without > > it the compiler doesn't produce the software div estimate. > ninf is also an issue, although maybe less likely to bite in practice. > Consider what happens if someone passes infinity to swdivs: the fdiv reduces > to poison, so the branch is undefined behavior. We've decided to emit an `fdiv` without any fast math flags for these builtins. This will be safe and will emit the software estimate for `-Ofast`. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D106959/new/ https://reviews.llvm.org/D106959 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
