On Thu, Jan 04, 2018 at 10:31:16PM +0100, Jakub Jelinek wrote:
> While debugging PR target/83554, I first thought the bug is that
> regrename creates a new raw REG each time and when the insn uses
> match_dup that it would be a problem.
>
> While that is not the case, for match_dup we actually check it with
> rtx_equal_p, the patch I wrote can be IMHO still useful to cheaply
> decrease compile time memory a little bit, the original registers are
> often pointer equal and seems they can be replaced by the same new
> REG too.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2018-01-04 Jakub Jelinek <ja...@redhat.com>
>
> * regrename.c (regrename_do_replace): If replacing the same
> reg multiple times, try to reuse last created gen_raw_REG.
I'd like to ping this patch (for some reason it didn't make it into
gcc-patches archive).
> --- gcc/regrename.c.jj 2018-01-04 00:43:15.077702454 +0100
> +++ gcc/regrename.c 2018-01-04 11:19:36.975798717 +0100
> @@ -963,6 +963,7 @@ regrename_do_replace (struct du_head *he
> struct du_chain *chain;
> unsigned int base_regno = head->regno;
> machine_mode mode;
> + rtx last_reg = NULL_RTX, last_repl = NULL_RTX;
>
> for (chain = head->first; chain; chain = chain->next_use)
> {
> @@ -975,12 +976,16 @@ regrename_do_replace (struct du_head *he
> gen_rtx_UNKNOWN_VAR_LOC (), true);
> else
> {
> - validate_change (chain->insn, chain->loc,
> - gen_raw_REG (GET_MODE (*chain->loc), reg), true);
> - if (regno >= FIRST_PSEUDO_REGISTER)
> - ORIGINAL_REGNO (*chain->loc) = regno;
> - REG_ATTRS (*chain->loc) = attr;
> - REG_POINTER (*chain->loc) = reg_ptr;
> + if (*chain->loc != last_reg)
> + {
> + last_repl = gen_raw_REG (GET_MODE (*chain->loc), reg);
> + if (regno >= FIRST_PSEUDO_REGISTER)
> + ORIGINAL_REGNO (last_repl) = regno;
> + REG_ATTRS (last_repl) = attr;
> + REG_POINTER (last_repl) = reg_ptr;
> + last_reg = *chain->loc;
> + }
> + validate_change (chain->insn, chain->loc, last_repl, true);
> }
> }
>
>
Jakub
