This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: Make nonoverlapping_component_refs_since_match_p work with non-trivial MEM_REFs and TMRs
On Thu, 11 Jul 2019, Jan Hubicka wrote:
> > > On Thu, 11 Jul 2019, Jan Hubicka wrote:
> > >
> > > > Hi,
> > > > this patch makes nonoverlapping_component_refs_since_match_p to accept
> > > > paths with non-trivial MEM_REFs and TMRs assuming that they have same
> > > > semantics.
> > >
> > > Hmm. We'll never get any TARGET_MEM_REFs wrapped with
> > > handled-components so I wonder if it makes sense to handle it in
> > > nonoverlapping_component_refs_since_match_p at all.
> >
> > OK, that makes my life easier. Here is updated patch.
> Hi,
> the patch finished testing on x86_64-linux so here is with Changelog and
> testcase. OK?
OK.
Richard.
>
>
> * tree-ssa-alias.c (same_tmr_indexing_p): Break out from ...
> (indirect_refs_may_alias_p): ... here.
> (nonoverlapping_component_refs_since_match_p): Support also non-trivial
> mem refs in the access paths.
> Index: testsuite/gcc.dg/tree-ssa/alias-access-path-9.c
> ===================================================================
> --- testsuite/gcc.dg/tree-ssa/alias-access-path-9.c (nonexistent)
> +++ testsuite/gcc.dg/tree-ssa/alias-access-path-9.c (working copy)
> @@ -0,0 +1,44 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -fdump-tree-fre1" } */
> +
> +/* This testcase tests nonoverlapping_component_refs_since_match_p in presence
> + of non-trivial mem-refs. */
> +struct a {int a,b;};
> +struct b {struct a a[10];};
> +struct c {int c; struct b b;} c, *cptr;
> +
> +void
> +set_a(struct a *a, int p)
> +{
> + a->a=p;
> +}
> +void
> +set_b(struct a *a, int p)
> +{
> + a->b=p;
> +}
> +int
> +get_a(struct a *a)
> +{
> + return a->a;
> +}
> +
> +int
> +test(int i, int j)
> +{
> + struct b *bptr = &c.b;
> + set_a (&bptr->a[i], 123);
> + set_b (&bptr->a[j], 124);
> + return get_a (&bptr->a[i]);
> +}
> +
> +int
> +test2(int i, int j)
> +{
> + struct b *bptr = &cptr->b;
> + set_a (&bptr->a[i], 125);
> + set_b (&bptr->a[j], 126);
> + return get_a (&bptr->a[i]);
> +}
> +/* { dg-final { scan-tree-dump-times "return 123" 1 "fre1"} } */
> +/* { dg-final { scan-tree-dump-times "return 125" 1 "fre1"} } */
> Index: tree-ssa-alias.c
> ===================================================================
> --- tree-ssa-alias.c (revision 273322)
> +++ tree-ssa-alias.c (working copy)
> @@ -1265,20 +1265,6 @@ nonoverlapping_component_refs_since_matc
> component_refs1.safe_push (ref1);
> ref1 = TREE_OPERAND (ref1, 0);
> }
> - if (TREE_CODE (ref1) == MEM_REF && ref1 != match1)
> - {
> - if (!integer_zerop (TREE_OPERAND (ref1, 1)))
> - {
> - ++alias_stats.nonoverlapping_component_refs_since_match_p_may_alias;
> - return -1;
> - }
> - }
> - /* TODO: Handle TARGET_MEM_REF later. */
> - if (TREE_CODE (ref1) == TARGET_MEM_REF && ref1 != match1)
> - {
> - ++alias_stats.nonoverlapping_component_refs_since_match_p_may_alias;
> - return -1;
> - }
>
> /* Create the stack of handled components for REF2. */
> while (handled_component_p (ref2) && ref2 != match2)
> @@ -1290,20 +1276,31 @@ nonoverlapping_component_refs_since_matc
> component_refs2.safe_push (ref2);
> ref2 = TREE_OPERAND (ref2, 0);
> }
> - if (TREE_CODE (ref2) == MEM_REF && ref2 != match2)
> - {
> - if (!integer_zerop (TREE_OPERAND (ref2, 1)))
> - {
> - ++alias_stats.nonoverlapping_component_refs_since_match_p_may_alias;
> - return -1;
> - }
> - }
> - if (TREE_CODE (ref2) == TARGET_MEM_REF && ref2 != match2)
> +
> + bool mem_ref1 = TREE_CODE (ref1) == MEM_REF && ref1 != match1;
> + bool mem_ref2 = TREE_CODE (ref2) == MEM_REF && ref2 != match2;
> +
> + /* If only one of access paths starts with MEM_REF check that offset is 0
> + so the addresses stays the same after stripping it.
> + TODO: In this case we may walk the other access path until we get same
> + offset.
> +
> + If both starts with MEM_REF, offset has to be same. */
> + if ((mem_ref1 && !mem_ref2 && !integer_zerop (TREE_OPERAND (ref1, 1)))
> + || (mem_ref2 && !mem_ref1 && !integer_zerop (TREE_OPERAND (ref2, 1)))
> + || (mem_ref1 && mem_ref2
> + && !tree_int_cst_equal (TREE_OPERAND (ref1, 1),
> + TREE_OPERAND (ref2, 1))))
> {
> ++alias_stats.nonoverlapping_component_refs_since_match_p_may_alias;
> return -1;
> }
>
> + /* TARGET_MEM_REF are never wrapped in handled components, so we do not need
> + to handle them here at all. */
> + gcc_checking_assert (TREE_CODE (ref1) != TARGET_MEM_REF
> + && TREE_CODE (ref2) != TARGET_MEM_REF);
> +
> /* Pop the stacks in parallel and examine the COMPONENT_REFs of the same
> rank. This is sufficient because we start from the same DECL and you
> cannot reference several fields at a time with COMPONENT_REFs (unlike
>
--
Richard Biener <rguenther@suse.de>
SUSE Linux GmbH, Maxfeldstrasse 5, 90409 Nuernberg, Germany;
GF: Felix Imendörffer, Mary Higgins, Sri Rasiah; HRB 21284 (AG Nürnberg)