[Bug tree-optimization/80620] [8 Regression] gcc produces wrong code with -O3

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80620

Richard Biener  changed:

   What|Removed |Added

 Status|ASSIGNED|RESOLVED
 Resolution|--- |FIXED

--- Comment #9 from Richard Biener  ---
Now really fixed.

[Bug tree-optimization/80620] [8 Regression] gcc produces wrong code with -O3

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80620

--- Comment #8 from Richard Biener  ---
Author: rguenth
Date: Tue Jul 18 07:35:40 2017
New Revision: 250297

URL: https://gcc.gnu.org/viewcvs?rev=250297=gcc=rev
Log:
2017-07-18  Richard Biener  

PR tree-optimization/80620
PR tree-optimization/81403
* tree-ssa-pre.c (phi_translate_1): Clear range and points-to
info when re-using a VN table entry.

* gcc.dg/torture/pr80620.c: New testcase.
* gcc.dg/torture/pr81403.c: Likewise.

Added:
trunk/gcc/testsuite/gcc.dg/torture/pr80620.c
trunk/gcc/testsuite/gcc.dg/torture/pr81403.c
Modified:
trunk/gcc/ChangeLog
trunk/gcc/testsuite/ChangeLog
trunk/gcc/tree-ssa-pre.c

[Bug tree-optimization/80620] [8 Regression] gcc produces wrong code with -O3

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80620

--- Comment #7 from Richard Biener  ---
Now needs sth before r247596 to reproduce.

The whole

PRE_EXPR_NARY (expr) = newnary;
constant = fully_constant_expression (expr);
PRE_EXPR_NARY (expr) = nary;
if (constant != expr)
  {
/* For non-CONSTANTs we have to make sure we can eventually
   insert the expression.  Which means we need to have a
   leader for it.  */
if (constant->kind != CONSTANT)
  {
/* Do not allow simplifications to non-constants over
   backedges as this will likely result in a loop PHI node
   to be inserted and increased register pressure.
   See PR77498 - this avoids doing predcoms work in
   a less efficient way.  */
if (find_edge (pred, phiblock)->flags & EDGE_DFS_BACK)
  ;
else
  {
unsigned value_id = get_expr_value_id (constant);
constant = find_leader_in_sets (value_id, set1, set2,
AVAIL_OUT (pred));
if (constant)
  return constant;

mitigates the fact that

tree result = vn_nary_op_lookup_pieces (newnary->length,
newnary->opcode,
newnary->type,
>op[0],
);
if (result && is_gimple_min_invariant (result))
  return get_or_alloc_expr_for_constant (result);

doesn't allow SSA name result to prevail here.  That is vn_nary_simplify
also just looks up the expr now even if not simplified.  It also gets
fed leaders which are not necessarily final values and thus do not have
range-info cleared.

The SCCVN band-aid to clear SSA_NAME_RANGE_INFO doesn't apply here because
nothing is value-numbered to a.4_4 besides itself.  But phi-translation
effectively value-numbers an expression to it.

[Bug tree-optimization/80620] [8 Regression] gcc produces wrong code with -O3

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80620

--- Comment #6 from Richard Biener  ---
Ugh.  This is really another latent issue with PRE and range-info.  When
PHI-translating {bit_and_expr,a.12_17,4081116982543369} when valueizing
a.12_17 we end up with a leader for a.12_17 as {nop_expr,a.6_10} (as this is
in antic-in).  We then translate that as {nop_expr,a.0_1} for which we have
a.4_4 as "representative" -- but this has range-info that isn't valid in
the context we are looking at, when we the simplify using match.pd we use
that range-info and BOOM!.

The rev. in question just made this more likely to trigger (previously
only match.pd simplifications to constants were considered).

[Bug tree-optimization/80620] [8 Regression] gcc produces wrong code with -O3

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80620

Richard Biener  changed:

   What|Removed |Added

 Status|NEW |ASSIGNED
Version|7.0.1   |8.0
   Assignee|unassigned at gcc dot gnu.org  |rguenth at gcc dot 
gnu.org
   Target Milestone|7.2 |8.0
Summary|[7/8 Regression] gcc|[8 Regression] gcc produces
   |produces wrong code with|wrong code with -O3
   |-O3 |

--- Comment #5 from Richard Biener  ---
The rev. is not on the GCC 7 branch.  I will have a look.