boostorg/math

The future of fpclassify.hpp

Opened this issue · 1 comments

jzmaddock commented

I'm opening this as a place for discussion really.

We currently have boost::math::fpclassify/isnan/isinf/isnormal etc which began life as workarounds for C++03 compilers, but they have another role: for IEEE conforming types they work even in the presence of -ffast-math.

Now, I have some dim recollection, that gcc/clang used to replace say std::isnan(x) with a literal false even with just -O3. And that we had some very inscrutable bug reports caused by this. However, that appears not to be the case with current releases, so perhaps this has been fixed/changed?

So... I'm undecided whether the boost functions should become synonyms for the std ones, but just wanted to flag the issue.

NAThompson commented

Yeah replacing std::isnan(x) with false is the worst!

Here's a godbolt to play around with the compiler flags.

Summarizing:

  • gcc 10.2: -O3 -ffast-math -> std::isnan(x) always is false. However, the flag is -O3 -ffast-math -fno-finite-math-only gets the correct behavior.
  • icc appears to get it right no matter what the flag is.
  • clang 11 does the "wrong" thing with -O3 -ffast-math, and does the "right" thing with -O3 as well as -O3 -ffast-math -fno-finite-math-only .