https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121652
--- Comment #17 from Aurelien Jarno <aurelien at aurel32 dot net> --- (In reply to Jeffrey A. Law from comment #16) > WRT c#13, yes, it's worth it, the performance gains were huge. Of course if > we look out Zfa should become ubiquitious and that other path won't matter. > But we're not at that point yet IMHO. Yep, but they were huge for the -ffast-math case used by SPEC2017. The code works fine for that case, and changing the condition of the pattern as suggested does not change the code generated with -ffast-math and thus the SPEC2017 performance. My point is do we want to inline the case that handles both inexact and NaN? I guess we are talking about going from 7 to 13 instructions.
