[Bug c/92436] New: SIMD integer subtract with constant always becomes add

[zingaburga+gcc at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92436

            Bug ID: 92436
           Summary: SIMD integer subtract with constant always becomes add
           Product: gcc
           Version: 9.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: zingaburga+gcc at hotmail dot com
  Target Milestone: ---

Firstly, this isn't a bug, rather a missed optimization opportunity (I presume
this is the place to post these?).


With the optimizer enabled, it seems like SIMD integer subtract, with a
constant, always gets turned into a SIMD integer add with the constant negated.

For a target like x86 SSE, I suppose this may make sense, as the commutative
property of addition gives more flexibility around register placement, but it
isn't always beneficial - for example, if the constant could be re-used
elsewhere.

Example (x86):

        _mm_or_si128(
                _mm_sub_epi8(a, _mm_set1_epi8(99)),
                _mm_set1_epi8(99)
        );

In this case, the '99' constant can be used in both the subtract and or, but
GCC will always convert the first use to a '-99' constant, meaning that it now
has to deal with two constants: https://godbolt.org/z/gaKAkA

This can have a greater effect when the constants are held in registers, as the
negated constant wastes a register, which can sometimes cause otherwise
unnecessary register spilling elsewhere.

The behavior persists with AVX enabled, and I've even seen it for ARM NEON:
https://godbolt.org/z/z3b5mq

---

Perhaps a different issue, but maybe related: I noticed that switching the
order of the arguments for subtract, GCC seems to think the two constants are
different, even though this is not the case: https://godbolt.org/z/6fGhGd

For this second example ((99-a)|99), I'd have thought the more appropriate
assembly to be something like:

        vmovdqa xmm1, XMMWORD PTR .LC0[rip]
        vpsubb  xmm0, xmm1, xmm0
        vpor    xmm0, xmm0, xmm1