Re: [PATCH] [x86] reenable dword MOVE_MAX for better memmove inlining

[Richard Biener <richard.guenther@gmail.com>]

On Wed, May 24, 2023 at 7:47 AM Alexandre Oliva via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
>
> MOVE_MAX on x86* used to accept up to 16 bytes, even without SSE,
> which enabled inlining of small memmove by loading and then storing
> the entire range.  After the "x86: Update piecewise move and store"
> r12-2666 change, memmove of more than 4 bytes would not be inlined in
> gimple_fold_bultin_memory_op, failing the expectations of a few tests.
>
> I can see how lowering it for MOVE_MAX_PIECES can get us better
> codegen decisions overall, but surely inlining memmove with 2 32-bit
> loads and stores is better than an outline call that requires setting
> up 3 arguments.  I suppose even 3 or 4 could do better.  But maybe it
> is gimple_fold_builtin_memory_op that needs tweaking?

gimple_fold_builtin_memory_op tries to expand the call to a single
load plus a single store so we can handle overlaps by first loading
everything to registers and then storing:

      /* If we can perform the copy efficiently with first doing all loads
         and then all stores inline it that way.  Currently efficiently
         means that we can load all the memory into a single integer
         register which is what MOVE_MAX gives us.  */

using DImode on i?86 without SSE means we eventually perform two
loads and two stores which means we need two registers available.
That might not be an issue on x86_64 doing 16 bytes with two DImode
ops (and -mno-sse) since there's plenty of regs available.

So I think if we want to expand this further at the GIMPLE level we
should still honor MOVE_MAX but eventually emit multiple loads/stores
honoring the MOVE_MAX_PIECES set of constraints there and avoid
expanding to sequences where we cannot interleave the loads/stores
(aka for the memmove case).

> Anyhow, this patch raises MOVE_MAX back a little for non-SSE targets,
> while preserving the new value for MOVE_MAX_PIECES.
>
> Bootstrapped on x86_64-linux-gnu.  Also tested on ppc- and x86-vx7r2
> with gcc-12.
>
> for gcc/ChangeLog
>
>         * config/i386/i386.h (MOVE_MAX): Rename to...
>         (MOVE_MAX_VEC): ... this.  Add NONVEC parameter, and use it as
>         the last resort, instead of UNITS_PER_WORD.
>         (MOVE_MAX): Reintroduce in terms of MOVE_MAX_VEC, with
>         2*UNITS_PER_WORD.
>         (MOVE_MAX_PIECES): Likewise, but with UNITS_PER_WORD.
> ---
>  gcc/config/i386/i386.h |    6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
>
> diff --git a/gcc/config/i386/i386.h b/gcc/config/i386/i386.h
> index c7439f89bdf92..5293a332a969a 100644
> --- a/gcc/config/i386/i386.h
> +++ b/gcc/config/i386/i386.h
> @@ -1801,7 +1801,9 @@ typedef struct ix86_args {
>     is the number of bytes at a time which we can move efficiently.
>     MOVE_MAX_PIECES defaults to MOVE_MAX.  */
>
> -#define MOVE_MAX \
> +#define MOVE_MAX MOVE_MAX_VEC (2 * UNITS_PER_WORD)
> +#define MOVE_MAX_PIECES MOVE_MAX_VEC (UNITS_PER_WORD)
> +#define MOVE_MAX_VEC(NONVEC) \
>    ((TARGET_AVX512F \
>      && (ix86_move_max == PVW_AVX512 \
>         || ix86_store_max == PVW_AVX512)) \
> @@ -1813,7 +1815,7 @@ typedef struct ix86_args {
>        : ((TARGET_SSE2 \
>           && TARGET_SSE_UNALIGNED_LOAD_OPTIMAL \
>           && TARGET_SSE_UNALIGNED_STORE_OPTIMAL) \
> -        ? 16 : UNITS_PER_WORD)))
> +        ? 16 : (NONVEC))))
>
>  /* STORE_MAX_PIECES is the number of bytes at a time that we can store
>     efficiently.  Allow 16/32/64 bytes only if inter-unit move is enabled
>
> --
> Alexandre Oliva, happy hacker                https://FSFLA.org/blogs/lxo/
>    Free Software Activist                       GNU Toolchain Engineer
> Disinformation flourishes because many people care deeply about injustice
> but very few check the facts.  Ask me about <https://stallmansupport.org>