Re: [PATCH] Speedup cse_insn

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

On Fri, Feb 3, 2023 at 2:06 PM Richard Biener via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> When cse_insn prunes src{,_folded,_eqv_here,_related} with the
> equivalence set in the *_same_value chain it also searches for
> an equivalence to the destination of the instruction with
>
>           /* This is the same as the destination of the insns, we want
>              to prefer it.  Copy it to src_related.  The code below will
>              then give it a negative cost.  */
>           if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
>             src_related = p->exp;
>
> this picks up the last such equivalence and in particular any
> later duplicate will be pruned by the preceeding
>
>           else if (src_related && GET_CODE (src_related) == code
>                    && rtx_equal_p (src_related, p->exp))
>             src_related = 0;
>
> first.  This wastes cycles doing extra rtx_equal_p checks.  The
> following instead searches for the first destination equivalence
> separately in this loop and delays using src_related for it until
> we are about to process that, avoiding another redundant rtx_equal_p
> check.
>
> I've came here because of a testcase with very large equivalence
> lists and compile-time of cse_insn.  The patch below doesn't speed
> it up significantly since there's no equivalence on the destination.
>
> In theory this opens the possibility to track dest_related
> separately, avoiding the implicit pruning of any previous
> value in src_related.  As is the change should be a no-op for
> code generation.
>
> Bootstrapped and tested on x86_64-unknown-linux-gnu, queued for
> stage1.

I have pushed this now after re-bootstrapping and testing on
x86_64-unknown-linux-gnu.

Richard.

>         * cse.cc (cse_insn): Track an equivalence to the destination
>         separately and delay using src_related for it.
> ---
>  gcc/cse.cc | 51 +++++++++++++++++++++++++++------------------------
>  1 file changed, 27 insertions(+), 24 deletions(-)
>
> diff --git a/gcc/cse.cc b/gcc/cse.cc
> index 8fbda4ecc86..543cb1fe36f 100644
> --- a/gcc/cse.cc
> +++ b/gcc/cse.cc
> @@ -4614,6 +4614,7 @@ cse_insn (rtx_insn *insn)
>        rtx src_eqv_here;
>        rtx src_const = 0;
>        rtx src_related = 0;
> +      rtx dest_related = 0;
>        bool src_related_is_const_anchor = false;
>        struct table_elt *src_const_elt = 0;
>        int src_cost = MAX_COST;
> @@ -5085,10 +5086,11 @@ cse_insn (rtx_insn *insn)
>             src_related = 0;
>
>           /* This is the same as the destination of the insns, we want
> -            to prefer it.  Copy it to src_related.  The code below will
> -            then give it a negative cost.  */
> -         if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
> -           src_related = p->exp;
> +            to prefer it.  The code below will then give it a negative
> +            cost.  */
> +         if (!dest_related
> +             && GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
> +           dest_related = p->exp;
>         }
>
>        /* Find the cheapest valid equivalent, trying all the available
> @@ -5130,27 +5132,28 @@ cse_insn (rtx_insn *insn)
>             }
>         }
>
> -      if (src_related)
> +      if (dest_related)
>         {
> -         if (rtx_equal_p (src_related, dest))
> -           src_related_cost = src_related_regcost = -1;
> -         else
> -           {
> -             src_related_cost = COST (src_related, mode);
> -             src_related_regcost = approx_reg_cost (src_related);
> -
> -             /* If a const-anchor is used to synthesize a constant that
> -                normally requires multiple instructions then slightly prefer
> -                it over the original sequence.  These instructions are likely
> -                to become redundant now.  We can't compare against the cost
> -                of src_eqv_here because, on MIPS for example, multi-insn
> -                constants have zero cost; they are assumed to be hoisted from
> -                loops.  */
> -             if (src_related_is_const_anchor
> -                 && src_related_cost == src_cost
> -                 && src_eqv_here)
> -               src_related_cost--;
> -           }
> +         src_related_cost = src_related_regcost = -1;
> +         /* Handle it as src_related.  */
> +         src_related = dest_related;
> +       }
> +      else if (src_related)
> +       {
> +         src_related_cost = COST (src_related, mode);
> +         src_related_regcost = approx_reg_cost (src_related);
> +
> +         /* If a const-anchor is used to synthesize a constant that
> +            normally requires multiple instructions then slightly prefer
> +            it over the original sequence.  These instructions are likely
> +            to become redundant now.  We can't compare against the cost
> +            of src_eqv_here because, on MIPS for example, multi-insn
> +            constants have zero cost; they are assumed to be hoisted from
> +            loops.  */
> +         if (src_related_is_const_anchor
> +             && src_related_cost == src_cost
> +             && src_eqv_here)
> +           src_related_cost--;
>         }
>
>        /* If this was an indirect jump insn, a known label will really be
> --
> 2.35.3