https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78847
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |amker at gcc dot gnu.org,
| |rguenth at gcc dot gnu.org
--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
And the real inefficiency is niter analysis returning unsimplified
(unsigned long) &MEM[(void *)&foo + 9B] - (unsigned long) ((const char *)
&foo.ascii_ + 1)
from here:
#1 0x000000000134f088 in number_of_iterations_ne (loop=0x7ffff5534550,
type=<pointer_type 0x7ffff569f3f0>, iv=0x7fffffffd4b0,
final=<addr_expr 0x7ffff51b4620>, niter=0x7fffffffd5e0,
exit_must_be_taken=true, bnds=0x7fffffffd3d0)
at /space/rguenther/src/svn/gcc-7-branch/gcc/tree-ssa-loop-niter.c:991
991 fold_convert (niter_type, iv->base));
(gdb) l
986 else
987 {
988 s = fold_convert (niter_type, iv->step);
989 c = fold_build2 (MINUS_EXPR, niter_type,
990 fold_convert (niter_type, final),
991 fold_convert (niter_type, iv->base));
where
final: &MEM[(void *)&foo + 9B]
iv->base: (const char *) &foo.ascii_ + 1
one could either try to enhance the associate: case of fold_binary or use
tree-affine above, like with
Index: gcc/tree-ssa-loop-niter.c
===================================================================
--- gcc/tree-ssa-loop-niter.c (revision 245276)
+++ gcc/tree-ssa-loop-niter.c (working copy)
@@ -42,6 +42,7 @@ along with GCC; see the file COPYING3.
#include "tree-chrec.h"
#include "tree-scalar-evolution.h"
#include "params.h"
+#include "tree-affine.h"
/* The maximum number of dominator BBs we search for conditions
@@ -986,9 +987,12 @@ number_of_iterations_ne (struct loop *lo
else
{
s = fold_convert (niter_type, iv->step);
- c = fold_build2 (MINUS_EXPR, niter_type,
- fold_convert (niter_type, final),
- fold_convert (niter_type, iv->base));
+ aff_tree aff_ivbase, aff_final;
+ tree_to_aff_combination (iv->base, niter_type, &aff_ivbase);
+ tree_to_aff_combination (final, niter_type, &aff_final);
+ aff_combination_scale (&aff_ivbase, -1);
+ aff_combination_add (&aff_ivbase, &aff_final);
+ c = aff_combination_to_tree (&aff_ivbase);
}
mpz_init (max);
though this is horribly incomplete (fixes the testcase though) and if, then
niter analysis should use aff_trees throughout (not re-build trees for the
intermediate stuff). With that patch we immediately get '9' as the
number of iterations and thus ldist produces
<bb 2> [10.00%]:
_18 = buf__9(D) + ptr_6(D);
__builtin_memcpy (_18, &foo, 9);
