[Bug tree-optimization/52272] [4.7/4.8/4.9 regression] Performance regression of 410.bwaves on x86.

["amker.cheng at gmail dot com" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52272

bin.cheng <amker.cheng at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |amker.cheng at gmail dot com

--- Comment #21 from bin.cheng <amker.cheng at gmail dot com> ---
Hi Richard,
I looked into PR50955 for which the mentioned commit causing this PR is
applied:

Commit 
2012-02-06  Richard Guenther  <rguenther@suse.de>

        PR tree-optimization/50955
        * tree-ssa-loop-ivopts.c (get_computation_cost_at): Artificially
        raise cost of expressions that replace an address with an
        expression based on a different pointer.

I noticed that the offending non-linear use in PR50955 is actually from memory
reference.  If I understand the issue correct, the whole alias issue is
introduced by rewriting iv use with one base_object through candidate with
another incompatible base_object, and it is related to memory reference.  An
genuine non-linear iv use (the pointer never de-referenced, like in this PR)
won't have this issue.

So I come up this idea to relax the condition:

-  if (address_p)
+  if (address_p
+      || (use->iv->base_object
+         && cand->iv->base_object
+         && POINTER_TYPE_P (TREE_TYPE (use->iv->base_object))
+         && POINTER_TYPE_P (TREE_TYPE (cand->iv->base_object))))
     {
       /* Do not try to express address of an object with computation based
         on address of a different object.  This may cause problems in rtl

to non-linear uses which truly occurred in memory reference, something like:

-  if (address_p)
+  if (address_p
+      || (use->in_mem_ref_p
+         && use->iv->base_object
+         && cand->iv->base_object
+         && POINTER_TYPE_P (TREE_TYPE (use->iv->base_object))
+         && POINTER_TYPE_P (TREE_TYPE (cand->iv->base_object))))
     {
       /* Do not try to express address of an object with computation based
         on address of a different object.  This may cause problems in rtl

The flag in_mem_ref_p can be set for appropriate uses when finding interesting
address uses.

With this change, this PR should be resolved while not violating PR50955.

I am not very much into 50955, so how does this sound? I can send a patch for
review if the idea is in right direction.

BTW, I cannot reproduce 50955 with the reported revision of GCC.  The store
isn't deleted by pass_cd_dce, though it is re-written just as the PR reported. 
So maybe I just misunderstood something.

Any words?

Thanks,
bin