Re: [PATCH] Optimize away x86 mem stores of what the mem contains already (PR rtl-optimization/79593)

[Uros Bizjak <ubizjak@gmail.com>]

On Mon, Jan 7, 2019 at 11:51 PM Jakub Jelinek <jakub@redhat.com> wrote:
>
> Hi!
>
> As mentioned in that PR, we have a SI->DImode zero extension and RA happens
> to choose to zero extend from a SImode memory slot which is the low part of
> the DImode memory slot into which the zero extension is to be stored.
> Unfortunately, the RTL DSE part really doesn't have infrastructure to
> remember and, if needed, invalidate loads, it just remembers stores, so
> handling this generically is quite unlikely at least for GCC9.
>
> This patch just handles that through a peephole2 (other option would be to
> handle it in the define_split for the zero extension, but the peephole2 is
> likely to catch more things).
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

Is there a reason stack registers are excluded? Before stackreg pass,
these registers are just like other hard registers.

Other that that, there is no need for REG_P predicate; after reload we
don't have subregs and register_operand will match only hard regs.
Also, please put peep2_reg_dead_p predicate in the pattern predicate.

Uros.

> 2019-01-07  Jakub Jelinek  <jakub@redhat.com>
>
>         PR rtl-optimization/79593
>         * config/i386/i386.md (reg = mem; mem = reg): New define_peephole2.
>
> --- gcc/config/i386/i386.md.jj  2019-01-01 12:37:31.564738571 +0100
> +++ gcc/config/i386/i386.md     2019-01-07 17:11:21.056392168 +0100
> @@ -18740,6 +18740,21 @@ (define_peephole2
>                        const0_rtx);
>  })
>
> +;; Attempt to optimize away memory stores of values the memory already
> +;; has.  See PR79593.
> +(define_peephole2
> +  [(set (match_operand 0 "register_operand")
> +        (match_operand 1 "memory_operand"))
> +   (set (match_dup 1) (match_dup 0))]
> +  "REG_P (operands[0])
> +   && !STACK_REGNO_P (operands[0])
> +   && !MEM_VOLATILE_P (operands[1])"
> +  [(set (match_dup 0) (match_dup 1))]
> +{
> +  if (peep2_reg_dead_p (1, operands[0]))
> +    DONE;
> +})
> +
>  ;; Attempt to always use XOR for zeroing registers (including FP modes).
>  (define_peephole2
>    [(set (match_operand 0 "general_reg_operand")
>
>         Jakub