[Bug middle-end/67553] New: Saturating SSE/AVX instructions do not get optimized

[Bug middle-end/67553] New: Saturating SSE/AVX instructions do not get optimized

[tmb99 at gmx dot net]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67553

            Bug ID: 67553
           Summary: Saturating SSE/AVX instructions do not get optimized
           Product: gcc
           Version: 5.2.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: tmb99 at gmx dot net
  Target Milestone: ---

Compiling this code with -O3 -mavx

__m128i v0 = _mm_setzero_si128();
__m128i v2 = _mm_setzero_si128();
__m128i sum = _mm_adds_epi16(v0,v2);
__m128i dif = _mm_subs_epi16(v0,v2);

results in the following badly optimized assembly code:

vpxor   %xmm0, %xmm0, %xmm0
vpsubsw %xmm0, %xmm0, %xmm1
vpaddsw %xmm0, %xmm0, %xmm0


IMHO the adds and subs instructions should be eliminated by the optimizer

[Bug rtl-optimization/67553] Saturating SSE/AVX instructions do not get optimized

[tmb99 at gmx dot net]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67553

--- Comment #2 from tmb99 at gmx dot net ---
seems to be the same for most saturating instructions:

__m128i v0 = _mm_setzero_si128();
__m128i v2 = _mm_setzero_si128();
__m128i sum = _mm_adds_epi16(v0,v2);
__m128i dif = _mm_subs_epi8(v0,v2);
__m128i hsum = _mm_hadds_epi16(v0,v2);
__m128i hdif = _mm_hsubs_epi16(v0,v2);
__m128i pacu = _mm_packus_epi16(v0,v2);
__m128i pacs = _mm_packs_epi32(v0,v2);

compiles to:

vpxor   %xmm0, %xmm0, %xmm0
vpxor   %xmm2, %xmm2, %xmm2
vphsubsw        %xmm0, %xmm0, %xmm4
vpackuswb       %xmm0, %xmm0, %xmm3
vphaddsw        %xmm0, %xmm0, %xmm5
vpsubsb %xmm2, %xmm2, %xmm2
vpxor   %xmm1, %xmm1, %xmm1
vpaddsw %xmm0, %xmm0, %xmm0
vpackssdw       %xmm1, %xmm1, %xmm1

also: 3 setzero/vpxor instructions instead of just one.

[Bug rtl-optimization/67553] Saturating SSE/AVX instructions do not get optimized

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67553

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
             Target|                            |x86_64-*-*, i?86-*-*

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
The backend probably uses unspecs for this (the intrinsics use builtins).  I
wonder if the backend could enable fixed-point types/modes more generally.