[Bug middle-end/48590] New: Alias analysis confused by builtin-stack-save/restore

[Bug middle-end/48590] New: Alias analysis confused by builtin-stack-save/restore

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48590

           Summary: Alias analysis confused by builtin-stack-save/restore
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: rguenth@gcc.gnu.org


Lame testcase for now:

float foo (int i, double *d)
{
  float f;
  double dd;
  int j;
  dd = *d;
  for (j = 0; j<i; ++j)
    {
      float a[i];
    f+=a[j] + *d + dd;
    }
  return f;
}

we fail to CSE *d because we think that __builtin_stack_save () clobbers it.
__builtin_stack_restore () has similar effects on DSE (it appears to use
all aliased memory).

We have to be careful when improving this to not allow sinking operations
on alloca'd memory over a __builtin_stack_restore ().

[Bug middle-end/48590] Alias analysis confused by builtin-stack-save/restore

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48590

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2011.04.13 11:04:23
         AssignedTo|unassigned at gcc dot       |rguenth at gcc dot gnu.org
                   |gnu.org                     |
     Ever Confirmed|0                           |1

--- Comment #1 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-04-13 11:04:23 UTC ---
Mine anyway.

[Bug middle-end/48590] Alias analysis confused by builtin-stack-save/restore

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48590

--- Comment #2 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-04-13 11:13:14 UTC ---
The following should be safe (but doesn't try to address the dead code
elimination issue).


Index: gcc/tree-ssa-alias.c
===================================================================
--- gcc/tree-ssa-alias.c    (revision 170321)
+++ gcc/tree-ssa-alias.c    (working copy)
@@ -1432,6 +1432,9 @@ call_may_clobber_ref_p_1 (gimple call, a
         tree ptr = gimple_call_arg (call, 0);
         return ptr_deref_may_alias_ref_p_1 (ptr, ref);
       }
+    case BUILT_IN_STACK_SAVE:
+    case BUILT_IN_STACK_RESTORE:
+      return false;
     case BUILT_IN_GAMMA_R:
     case BUILT_IN_GAMMAF_R:
     case BUILT_IN_GAMMAL_R:

[Bug middle-end/48590] Alias analysis confused by builtin-stack-save/restore

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48590

--- Comment #3 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-04-14 12:19:35 UTC ---
Author: rguenth
Date: Thu Apr 14 12:19:28 2011
New Revision: 172427

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=172427
Log:
2011-04-14  Richard Guenther  <rguenther@suse.de>

    PR tree-optimization/48590
    * tree-ssa-alias.c (ref_maybe_used_by_call_p_1): Handle
    BUILT_IN_ALLOCA, BUILT_IN_STACK_SAVE and BUILT_IN_STACK_RESTORE.
    (call_may_clobber_ref_p_1): Handle BUILT_IN_ALLOCA and
    BUILT_IN_STACK_SAVE.
    * tree-ssa-dce.c (propagate_necessity): Handle
    BUILT_IN_ALLOCA, BUILT_IN_STACK_SAVE and BUILT_IN_STACK_RESTORE.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/tree-ssa-alias.c
    trunk/gcc/tree-ssa-dce.c

[Bug middle-end/48590] Alias analysis confused by builtin-stack-save/restore

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48590

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED

--- Comment #4 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-04-14 12:37:40 UTC ---
Fixed.