[Bug tree-optimization/64541] New: .fre1 pass optimization failure

[Bug tree-optimization/64541] New: .fre1 pass optimization failure

[skvadrik at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64541

            Bug ID: 64541
           Summary: .fre1 pass optimization failure
           Product: gcc
           Version: 5.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: skvadrik at gmail dot com

Created attachment 34403
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=34403&action=edit
1.c 2.c 1.c.028t.esra 2.c.028t.esra 1.c.030t.fre1 2.c.030t.fre1 1.s 2.s
compile_gcc.sh

Files 1.c and 2.c in attach are equivalent from C/C++ standpoint:

$ diff 1.c 2.c
3c3,5
<       return *(*q = ++*p);
---
>       ++*p;
>       *q = *p;
>       return **p;

Compiled with gcc-5.0.0, disassembled with objdump (GNU Binutils) 2.24
(see full script compile_gcc.sh in attach).

For 1.c gcc generates better code than for 2.c
(compare 1.s vs 2.s in attach).

I looked at GIMPLE optimization dumps (-fdump-tree-all) and found that
up to *.029t.ealias pass dumps differ insignificantly (attached 
*.028t.esra dumps as they are shorter), but .030t.fre1 pass fails to
reduce intermediate variable '_9' in the second case (attached *.030t.fre
dumps).

Looks like a bug in full redundancy elimination.

[Bug tree-optimization/64541] FRE pass optimization failure

[skvadrik at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64541

--- Comment #2 from Ulya <skvadrik at gmail dot com> ---
> we have to assume that p == q and thus the store to *q invalidates the previously load *p

I see.
It seemed to me from GIMPLE dumps that both cases are equally easy to optimize,
perhaps I'm missing something.

[Bug tree-optimization/64541] FRE pass optimization failure

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64541

--- Comment #3 from rguenther at suse dot de <rguenther at suse dot de> ---
On Wed, 14 Jan 2015, skvadrik at gmail dot com wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64541
> 
> --- Comment #2 from Ulya <skvadrik at gmail dot com> ---
> > we have to assume that p == q and thus the store to *q invalidates the previously load *p
> 
> I see.
> It seemed to me from GIMPLE dumps that both cases are equally easy to optimize,
> perhaps I'm missing something.

1:

  _6 = *p_2(D);
  *q_7(D) = _6;
  _9 = *q_7(D);
  _10 = *_9;

it's easy to see that the load _9 = *q_7(D) results in _6.

2:

  *p_2(D) = _4;
  _6 = *p_2(D);
  *q_7(D) = _6;
  _9 = *p_2(D);
  _10 = *_9;

not so much here for the load _9 = *p_2(D) as I explained
(the store *q_7(D) = _6 aliases it)

[Bug tree-optimization/64541] FRE pass optimization failure

[skvadrik at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64541

--- Comment #4 from Ulya <skvadrik at gmail dot com> ---
Ah! Now I see the problem, thanks.

[Bug tree-optimization/64541] FRE pass optimization failure

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64541

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement