On Thu, 26 Mar 2015, Jakub Jelinek wrote:
> On Thu, Mar 26, 2015 at 10:32:18AM +0100, Richard Biener wrote:
> > I think I simply didn't want to change more testcases at that point,
> > but I can't see how followup passes at the very point wouldn't
> > clean things up very quickly (via fre or copyprop). Eventually
> > it was -Og and pass_fold_builtins interaction (thinking of
> > __builtin_constant_p (__builtin_object_size (...)) or so).
> > If so I'm sure I added a testcase.
>
> Can't find such a testcase.
> Anyway, the shorter version of the patch regresses:
>
> struct S { char a[10]; char b[10]; };
> int
> foo (int x)
> {
> struct S s;
> char *p = &s.a[4];
> if (x)
> p += 1;
> return __builtin_constant_p (__builtin_object_size (p, 1));
> }
>
> with -Og, while it works with the longer version.
>
> Though, say for
> return __builtin_constant_p (__builtin_object_size (p, 1) + 1);
>
> we already return 0 for -Og (and not -O2) already in 4.9.
Yeah, I think for -Og ideally both pass_object_sizes and
pass_fold_builtins should run from pass_ccp. We do have
some pass ordering issue in -Og otherwise. It's pipeline
would then also simplify to
Index: gcc/passes.def
===================================================================
--- gcc/passes.def (revision 221689)
+++ gcc/passes.def (working copy)
@@ -314,13 +314,7 @@ along with GCC; see the file COPYING3.
NEXT_PASS (pass_lower_complex);
NEXT_PASS (pass_lower_vector_ssa);
/* Perform simple scalar cleanup which is constant/copy
propagation. */
- NEXT_PASS (pass_ccp);
- NEXT_PASS (pass_object_sizes);
- /* Fold remaining builtins. */
- NEXT_PASS (pass_fold_builtins);
- /* Copy propagation also copy-propagates constants, this is
necessary
- to forward object-size and builtin folding results properly. */
- NEXT_PASS (pass_copy_prop);
+ NEXT_PASS (pass_ccp_with_objsize_and_fab);
NEXT_PASS (pass_dce);
NEXT_PASS (pass_asan);
NEXT_PASS (pass_tsan);
well, at least if ccp would also propagate copies... (I'd really
like to unify CCP and copyprop and have some flags to control
what it performs). So the above would probably need a copyprop
before that pass_ccp_with_objsize_and_fab, or teaching ccp
to also propagate copies (lattice value CONSTANT, value == SSA_NAME
and mask == ~0).
I wouldn't care about this for GCC 5.
Richard.
