search

trofi / trofi.github.io.gen

source for blog generator
6 stars 5 forks source link

now --ffast-math doesn't mess with FTZ/DOZ. Time to edit or make new article? #16

Closed NHOrus closed 4 months ago

NHOrus commented 4 months ago

now with bug https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55522 closed and shared libraries not tainting stuff with subnormals, is it time to make amends for existing article or make a new one, expanding on current situation?

trofi commented 4 months ago

Updating sounds fine. Would updating Tl;DR wording to mention compiler version for library case be enough? Or you see more places in need to update?

The initial article contained this verbiage in the middle:

""" This effect was seen as too problematic by many and gcc-13 stopped enabling crtfastmath.o for libraries (-shared option) in https://gcc.gnu.org/PR55522.

While the fix limits the impact to final programs that enable -Ofast it is still able to break all the libraries linked into the program linked with -Ofast.

As the library test suites are usually not ran with -Ofast the breakage might not be noticed until data corruption starts happening on real data in final applications. """

NHOrus commented 4 months ago

Ah, I missed this verbage. I think mentioning compiler version in TL;DR sounds fine. And the're a Clang bug that follows along and stops enabling crtfastmath.o for libraries, too. https://github.com/llvm/llvm-project/issues/57589 It's not yet in any Clang releases, I think? But should be in Clang-19. Hopefully. I don't know.

trofi commented 4 months ago
trofi commented 4 months ago

I assume the above is enough. Declaring it done.