On Fri, 2 Jun 2017, Jakub Jelinek wrote:
> Hi!
>
> As has been mentioned earlier, when the doloop code uses the expander after
> expansion, we need to unshare stuff because the expander relies on unsharing
> being done after it.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
Ok.
Richard.
> 2017-06-02 Jakub Jelinek <ja...@redhat.com>
>
> PR rtl-optimization/80903
> * loop-doloop.c (add_test): Unshare sequence.
>
> * gcc.dg/pr80903.c: New test.
>
> --- gcc/loop-doloop.c.jj 2017-03-24 14:36:35.000000000 +0100
> +++ gcc/loop-doloop.c 2017-05-29 12:40:39.062107631 +0200
> @@ -367,6 +367,7 @@ add_test (rtx cond, edge *e, basic_block
> }
>
> seq = get_insns ();
> + unshare_all_rtl_in_chain (seq);
> end_sequence ();
>
> /* There always is at least the jump insn in the sequence. */
> --- gcc/testsuite/gcc.dg/pr80903.c.jj 2017-05-29 12:40:27.988242239 +0200
> +++ gcc/testsuite/gcc.dg/pr80903.c 2017-05-29 12:37:39.000000000 +0200
> @@ -0,0 +1,34 @@
> +/* PR rtl-optimization/80903 */
> +/* { dg-do compile } */
> +/* { dg-options "-O3 -funroll-loops" } */
> +
> +short int a;
> +
> +void
> +foo (int x, short int y, short int z)
> +{
> + if (y != 0)
> + {
> + const short int b = 37;
> + y = 0;
> + while (y < b)
> + {
> + while (y < b)
> + {
> + lab:
> + ++y;
> + }
> + for (y = 0; y < b - 1; ++y)
> + ;
> + if (z != 0)
> + {
> + --a;
> + y *= a;
> + }
> + z = x;
> + }
> + x = 0;
> + }
> +
> + goto lab;
> +}
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB
21284 (AG Nuernberg)
