On Tue, Sep 19, 2017 at 5:54 PM, Tamar Christina
<tamar.christ...@arm.com> wrote:
> -- sorry for the duplicate, forgot to post to list as well first time --
>
> Hi Richard,
>
> The testcase seems to fail on aarch64-none-elf when -O1 or -O2,
>
> -O0, -Os and -O3 seem to work fine.
>
> dc[0] ends up being 0 for the cases that fail.
What ISL version are you using?
Richard.
> Kind regards,
> Tamar
> ________________________________________
> From: gcc-patches-ow...@gcc.gnu.org <gcc-patches-ow...@gcc.gnu.org> on behalf
> of Richard Biener <rguent...@suse.de>
> Sent: Monday, September 18, 2017 8:31 AM
> To: gcc-patches@gcc.gnu.org
> Subject: [PATCH] Fix PR79622
>
> The following patch fixes the other known wrong-code bug in GRAPHITE
> which shows we're mishandling PHIs in not properly considering the
> edge copies they represent as living outside of the black-box we're
> analyzing.
>
> Bootstrapped and tested on x86_64-unknown-linux-gnu, applied to trunk.
>
> Note the testcase still fails with ISL 0.16.1 but passes with 0.18
> for me. I'll update the version in download_prerequesites to 0.18.
>
> Richard.
>
> 2017-09-18 Richard Biener <rguent...@suse.de>
>
> PR tree-optimization/79622
> * graphite-scop-detection.c (build_cross_bb_scalars_def): Properly
> handle PHIs.
> (build_cross_bb_scalars_use): Likewise.
>
> * gcc.dg/graphite/pr79622.c: New testcase.
>
> Index: gcc/graphite-scop-detection.c
> ===================================================================
> --- gcc/graphite-scop-detection.c (revision 252806)
> +++ gcc/graphite-scop-detection.c (working copy)
> @@ -1744,7 +1744,9 @@ build_cross_bb_scalars_def (scop_p scop,
> gimple *use_stmt;
> imm_use_iterator imm_iter;
> FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
> - if (def_bb != gimple_bb (use_stmt) && !is_gimple_debug (use_stmt))
> + if ((def_bb != gimple_bb (use_stmt) && !is_gimple_debug (use_stmt))
> + /* PHIs have their effect at "BBs" on the edges. See PR79622. */
> + || gimple_code (SSA_NAME_DEF_STMT (def)) == GIMPLE_PHI)
> {
> writes->safe_push (def);
> DEBUG_PRINT (dp << "Adding scalar write: ";
> @@ -1758,7 +1760,8 @@ build_cross_bb_scalars_def (scop_p scop,
> }
> }
>
> -/* Record DEF if it is used in other bbs different than DEF_BB in the SCOP.
> */
> +/* Record USE if it is defined in other bbs different than USE_STMT
> + in the SCOP. */
>
> static void
> build_cross_bb_scalars_use (scop_p scop, tree use, gimple *use_stmt,
> @@ -1774,7 +1777,9 @@ build_cross_bb_scalars_use (scop_p scop,
> return;
>
> gimple *def_stmt = SSA_NAME_DEF_STMT (use);
> - if (gimple_bb (def_stmt) != gimple_bb (use_stmt))
> + if (gimple_bb (def_stmt) != gimple_bb (use_stmt)
> + /* PHIs have their effect at "BBs" on the edges. See PR79622. */
> + || gimple_code (def_stmt) == GIMPLE_PHI)
> {
> DEBUG_PRINT (dp << "Adding scalar read: ";
> print_generic_expr (dump_file, use);
> Index: gcc/testsuite/gcc.dg/graphite/pr79622.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/graphite/pr79622.c (nonexistent)
> +++ gcc/testsuite/gcc.dg/graphite/pr79622.c (working copy)
> @@ -0,0 +1,26 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -floop-nest-optimize" } */
> +
> +int bf;
> +
> +int
> +main (void)
> +{
> + int dc[5];
> +
> + for (bf = 0; bf < 2; ++bf)
> + {
> + int l9, g5 = -1;
> +
> + for (l9 = 0; l9 < 5; ++l9)
> + {
> + dc[l9] = g5;
> + g5 = (dc[l9] > 0);
> + }
> + }
> +
> + if (dc[0] != -1)
> + __builtin_abort ();
> +
> + return 0;
> +}
