| commit | 416a31796e6a34a134d251760c6055bacc39878c | |
| author | Egor Bogatov <egorbo@gmail.com> | |
| Tue, 3 Sep 2019 12:19:03 +0000 (3 15:19 +0300) | ||
| committer | Steve Pfister <steveisok@users.noreply.github.com> | |
| Tue, 3 Sep 2019 12:19:03 +0000 (3 08:19 -0400) | ||
| tree | 6448f86c916ae59e11566383c28cf7e704706ed5 | treesnapshot (tar.gz zip) |
| parent | 5692180ea21e596a08120090dd804872be3df1aa | commitdiff |
Improve Fast-Math mode for LLVM (#16599)
It turns out it's not enough to set global fast math parameters, it's also required to mark some instructions as "fast" (see https://llvm.org/docs/LangRef.html#fast-math-flags) like, for instance, clang does, see https://godbolt.org/z/U8TFST
E.g. all optimizations inside SimplifyLibCalls pass (it's a subpass of InstCombine) need it.
It turns out it's not enough to set global fast math parameters, it's also required to mark some instructions as "fast" (see https://llvm.org/docs/LangRef.html#fast-math-flags) like, for instance, clang does, see https://godbolt.org/z/U8TFST
E.g. all optimizations inside SimplifyLibCalls pass (it's a subpass of InstCombine) need it.