qt6-build.eclass: workaround mismatching cpu flags sets

qsimd_p.h tries determine if x86-64-v3 or v4 sets supported by the
compiler+flags seem complete, and errors out if not. This works out
rather badly on Gentoo, even -march=native can fail with some hard.

With qsimd_p.h being a private header this only affects dev-qt/*
packages, but fwics still need to fix all of dev-qt/* at once given
they are going to be using private bits from qtbase.

Debated a few options for this:

1. Patch headers (like old fix from bug #898644), but just not setting
   e.g. __haswell__ is not enough when __AVX2__ still confuses some
   code. Then unsetting these without going through the compiler leads
   to left over macros and more confusion. Besides... changing system
   headers behavior (why is __AVX2__ undef with -mavx2!?) only in
   distros tend to end in disaster.

2. Detect issues in qtbase ebuild, then disable x86intrin altogether
   if needed (carries over to other modules) + warn users (aka figure
   it out yourself). Not really great, users may also end up
   mismatching flags between dev-qt/ to fix this and then run into
   issues anyway.

3. Backport the next (wip/unmerged) upstream fix[1] but wait, looks
   like we are currently still playing whack-a-mole:
#  if defined(__AVX2__)
// List of features present with -march=x86-64-v3 and not architecturally
// implied by __AVX2__
#    define ARCH_HASWELL_MACROS     \
    (__AVX2__ && __BMI__ && __BMI2__ && __F16C__ \
     && __FMA__ && __LZCNT__ && __POPCNT__)
#    if ARCH_HASWELL_MACROS == 0
#      error "Please enable all x86-64-v3 extensions; <snip>
    ...so -mno-avx2 -mfma (bug #908420) is fine, older (bug #898644)
    is too, but if for any reason (VMs, buggy hardware, or the machine
    without F16C from bug #910419) anything else is disabled, then the
    issue is still there and may in fact trigger more than before.

4. Similarly to #2, detect issues but in the eclass. Then append
   -mno-* as needed to flags. Not too bad but passing -mno-*
   leads us to bug #913400.

5. Based on users flags, pick highest usable -march=x86-64-v* and
   strip everything else (or strip -march too...). Messy and think
   users wouldn't be happy about this. It would be what upstream
   likely wants us to do though, and causes no further problems.

Ultimately went with #4 for now, bug #913400 needs fixing either way.
- missing from set in #3 += -mno-avx2 (until #3, also -mno-fma)
- missing the AVX512* checked in qsimd_h += -mno-avx512*
(then let compiler disable features that depend on these)

Not great but better than doing nothing, no-op for non-affected users.

[1] https://codereview.qt-project.org/c/qt/qtbase/+/498799

Bug: https://bugs.gentoo.org/898644
Bug: https://bugs.gentoo.org/913400
Closes: https://bugs.gentoo.org/908420
Signed-off-by: Ionen Wolkens <ionen@gentoo.org>

0 files changed, 0 insertions, 0 deletions