[PATCH] D26855: New unsafe-fp-math implementation for X86 target

[Andrew V. Tischenko via Phabricator via llvm-commits]

avt77 added a comment.

Hi All, 
I found a really "stress" test for div operations (see the attachment)F3025626: spill_fdiv.ll <https://reviews.llvm.org/F3025626> (tnx to Sanjay Patel). The test shows maybe the worst case of the possible degradation because of this patch. I used the following command with 2 different compilers:

cmake -E time ./llc  spill_div.ll -o /dev/null -enable-unsafe-fp-math

For "pure" trunk compiler I got:   Elapsed time 2 s.
For compiler with patch I got:      Elapsed time 18 s.

(I launched the test several times with the same results.)
What's now? Is it acceptable? Should I try to optimize the patch? Should I try other benchmarks?
(I tried both LNT and Eigen but unfortunately they don't work for me at the moment because of unpredictable runtime issues.)


https://reviews.llvm.org/D26855