[Bug target/122176] New: -mno-mmx does not behave as documented

[mkretz at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122176

            Bug ID: 122176
           Summary: -mno-mmx does not behave as documented
           Product: gcc
           Version: 16.0
            Status: UNCONFIRMED
          Severity: minor
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: mkretz at gcc dot gnu.org
  Target Milestone: ---
            Target: x86_64-*-*, i?86-*-*

-mno-sse3 overrides -mavx512f (only __MMX__, __SSE__, and __SSE2__ are defined)
whereas -mno-mmx only undefs __MMX__ from the target ISA extensions.

This is in contradiction to config/i386/i386.opt which says that mavx512f (and
basically every other SSE/AVX/AVX512 flag) includes MMX.

Either the documentation or the logic what -mno-mmx does should be adjusted.

Example:

g++ -x c++ -dM -E -o- /dev/null -mavx512f -mno-mmx|grep 'MMX\|SSE\|AVX'

g++ -x c++ -dM -E -o- /dev/null -mavx512f -mno-sse3|grep 'MMX\|SSE\|AVX'

My point is that, according to the documentation, code that sees __SSE__
defined can assume __MMX__ to be defined. But that is not the behavior of GCC.
(And, FWIW, Clang behaves the same as GCC.)

I guess the correct fix is to strike MMX from the documentation in i386.opt?
Because it seems like the goal is to support codegen that ensures SIMD without
any MMX instructions?