Re: [x86 PATCH] Use movss/movsd to implement V4SI/V2DI VEC_PERM.

[Uros Bizjak <ubizjak@gmail.com>]

On Fri, Dec 23, 2022 at 5:46 PM Roger Sayle <roger@nextmovesoftware.com> wrote:
>
>
> This patch tweaks the x86 backend to use the movss and movsd instructions
> to perform some vector permutations on integer vectors (V4SI and V2DI) in
> the same way they are used for floating point vectors (V4SF and V2DF).
>
> As a motivating example, consider:
>
> typedef unsigned int v4si __attribute__((vector_size(16)));
> typedef float v4sf __attribute__((vector_size(16)));
> v4si foo(v4si x,v4si y) { return (v4si){y[0],x[1],x[2],x[3]}; }
> v4sf bar(v4sf x,v4sf y) { return (v4sf){y[0],x[1],x[2],x[3]}; }
>
> which is currently compiled with -O2 to:
>
> foo:    movdqa  %xmm0, %xmm2
>         shufps  $80, %xmm0, %xmm1
>         movdqa  %xmm1, %xmm0
>         shufps  $232, %xmm2, %xmm0
>         ret
>
> bar:    movss   %xmm1, %xmm0
>         ret
>
> with this patch both functions compile to the same form.
> Likewise for the V2DI case:
>
> typedef unsigned long v2di __attribute__((vector_size(16)));
> typedef double v2df __attribute__((vector_size(16)));
>
> v2di foo(v2di x,v2di y) { return (v2di){y[0],x[1]}; }
> v2df bar(v2df x,v2df y) { return (v2df){y[0],x[1]}; }
>
> which is currently generates:
>
> foo:    shufpd  $2, %xmm0, %xmm1
>         movdqa  %xmm1, %xmm0
>         ret
>
> bar:    movsd   %xmm1, %xmm0
>         ret
>
> There are two possible approaches to adding integer vector forms of the
> sse_movss and sse2_movsd instructions.  One is to use a mode iterator
> (VI4F_128 or VI8F_128) on the existing define_insn patterns, but this
> requires renaming the patterns to sse_movss_<mode> which then requires
> changes to i386-builtins.def and through-out the backend to reflect the
> new naming of gen_sse_movss_v4sf.  The alternate approach (taken here)
> is to simply clone and specialize the existing patterns.  Uros, if you'd
> prefer the first approach, I'm happy to make/test/commit those changes.

I would really prefer the variant with VI4F_128/VI8F_128, these two
iterators were introduced specifically for this case (see e.g.
sse_shufps_<mode> and sse2_shufpd_<mode>. The internal name of the
pattern is fairly irrelevant and a trivial search and replace
operation can replace the grand total of 6 occurrences ...)

Also, changing sse2_movsd to use VI8F_128 mode iterator would enable
more alternatives besides movsd, so we give combine pass some more
opportunities with memory operands.

So, the patch with those two iterators is pre-approved.

Uros.

> This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
> and make -k check, both with and without --target_board=unix{-m32},
> with no new failures.  Ok for mainline?
>
> 2022-12-23  Roger Sayle  <roger@nextmovesoftware.com>
>
> gcc/ChangeLog
>         * config/i386/i386-expand.cc (expand_vec_perm_movs): Also allow
>         V4SImode with TARGET_SSE and V2DImode with TARGET_SSE2.
>         * config/i386/sse.md (sse_movss_v4si): New define_insn, a V4SI
>         specialization of sse_movss.
>         (sse2_movsd_v2di): Likewise, a V2DI specialization of sse2_movsd.
>
> gcc/testsuite/ChangeLog
>         * gcc.target/i386/sse-movss-4.c: New test case.
>         * gcc.target/i386/sse2-movsd-3.c: New test case.
>
>
> Thanks in advance,
> Roger
> --
>