[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
--- Comment #34 from rguenth at gcc dot gnu dot org 2007-04-26 10:02 --- I'll bootstrap regtest that thing and commit it. -- rguenth at gcc dot gnu dot org changed: What|Removed |Added AssignedTo|unassigned at gcc dot gnu |rguenth at gcc dot gnu dot |dot org |org Status|NEW |ASSIGNED Last reconfirmed|2007-02-26 04:38:22 |2007-04-26 10:02:33 date|| http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
--- Comment #35 from rguenth at gcc dot gnu dot org 2007-04-26 12:15 --- Subject: Bug 30567 Author: rguenth Date: Thu Apr 26 12:15:16 2007 New Revision: 124184 URL: http://gcc.gnu.org/viewcvs?root=gccview=revrev=124184 Log: 2007-04-26 Richard Guenther [EMAIL PROTECTED] Daniel Berlin [EMAIL PROTECTED] PR tree-optimization/30567 * tree-ssa-structalias.c (update_alias_info): Record dereference also if ESCAPE_STORED_IN_GLOBAL. * g++.dg/other/pr30567.C: New testcase. Added: branches/gcc-4_2-branch/gcc/testsuite/g++.dg/other/pr30567.C Modified: branches/gcc-4_2-branch/gcc/ChangeLog branches/gcc-4_2-branch/gcc/testsuite/ChangeLog branches/gcc-4_2-branch/gcc/tree-ssa-structalias.c -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
--- Comment #36 from rguenth at gcc dot gnu dot org 2007-04-26 12:16 --- Fixed. -- rguenth at gcc dot gnu dot org changed: What|Removed |Added Status|ASSIGNED|RESOLVED Resolution||FIXED http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
--- Comment #37 from rguenth at gcc dot gnu dot org 2007-04-26 16:50 --- Subject: Bug 30567 Author: rguenth Date: Thu Apr 26 16:50:32 2007 New Revision: 124191 URL: http://gcc.gnu.org/viewcvs?root=gccview=revrev=124191 Log: 2007-04-26 Richard Guenther [EMAIL PROTECTED] Daniel Berlin [EMAIL PROTECTED] PR tree-optimization/30567 * g++.dg/other/pr30567.C: New testcase. Added: trunk/gcc/testsuite/g++.dg/other/pr30567.C Modified: trunk/gcc/testsuite/ChangeLog -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
--- Comment #26 from rguenth at gcc dot gnu dot org 2007-04-25 14:22
---
We fail to add the SFTs to the may_alias set of SMT.11, so add_virtual_operand
sees NULL may_aliases and doesn't add SFTs as clobbered.
I believe compute_flow_insensitive_aliasing is the culprit here as one can
easily
see that if we enter the
for (i = 0; i ai-num_pointers; i++)
{
size_t j;
struct alias_map_d *p_map1 = ai-pointers[i];
tree tag1 = var_ann (p_map1-var)-symbol_mem_tag;
bitmap may_aliases1 = p_map1-may_aliases;
if (PTR_IS_REF_ALL (p_map1-var))
continue;
for (j = i + 1; j ai-num_pointers; j++)
{
struct alias_map_d *p_map2 = ai-pointers[j];
tree tag2 = var_ann (p_map2-var)-symbol_mem_tag;
bitmap may_aliases2 = p_map2-may_aliases;
loop with may_aliases2 empty it will stay so. So as a fix I suggest to
add may aliases symmetrically like with
Index: tree-ssa-alias.c
===
*** tree-ssa-alias.c(revision 124151)
--- tree-ssa-alias.c(working copy)
*** compute_flow_insensitive_aliasing (struc
*** 1287,1292
--- 1287,1294
struct alias_map_d *p_map2 = ai-pointers[j];
tree tag2 = var_ann (p_map2-var)-symbol_mem_tag;
bitmap may_aliases2 = p_map2-may_aliases;
+ unsigned int k;
+ bitmap_iterator bi;
if (PTR_IS_REF_ALL (p_map2-var))
continue;
*** compute_flow_insensitive_aliasing (struc
*** 1301,1323
continue;
if (!bitmap_empty_p (may_aliases2))
! {
! unsigned int k;
! bitmap_iterator bi;
!
! /* Add all the aliases for TAG2 into TAG1's alias set.
!FIXME, update grouping heuristic counters. */
! EXECUTE_IF_SET_IN_BITMAP (may_aliases2, 0, k, bi)
! add_may_alias (tag1, referenced_var (k));
! bitmap_ior_into (may_aliases1, may_aliases2);
! }
else
! {
! /* Since TAG2 does not have any aliases of its own, add
!TAG2 itself to the alias set of TAG1. */
! add_may_alias (tag1, tag2);
! bitmap_set_bit (may_aliases1, DECL_UID (tag2));
! }
}
}
--- 1303,1325
continue;
if (!bitmap_empty_p (may_aliases2))
! /* Add all the aliases for TAG2 into TAG1's alias set.
! FIXME, update grouping heuristic counters. */
! EXECUTE_IF_SET_IN_BITMAP (may_aliases2, 0, k, bi)
! add_may_alias (tag1, referenced_var (k));
! else
! add_may_alias (tag1, tag2);
!
! if (!bitmap_empty_p (may_aliases1))
! /* Add all the aliases for TAG2 into TAG1's alias set.
! FIXME, update grouping heuristic counters. */
! EXECUTE_IF_SET_IN_BITMAP (may_aliases1, 0, k, bi)
! add_may_alias (tag2, referenced_var (k));
else
! add_may_alias (tag2, tag1);
!
! bitmap_ior_into (may_aliases2, may_aliases1);
! bitmap_ior_into (may_aliases1, may_aliases2);
}
}
but I have no clue what I am doing here. 't looks like big mess.
I prepare the workaround as alternative ;)
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
--- Comment #27 from rguenth at gcc dot gnu dot org 2007-04-25 14:42 --- The workaround doesn't work. I'll test the patch in comment #26, otherwise I'm out of ideas and clue on how it should work. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
--- Comment #28 from rguenth at gcc dot gnu dot org 2007-04-25 15:04 --- Created an attachment (id=13439) -- (http://gcc.gnu.org/bugzilla/attachment.cgi?id=13439action=view) prototype patch I'm testing this one. It'll make aliasing slower and more conservative, so I bet it's not the right or the best fix. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
--- Comment #29 from rguenth at gcc dot gnu dot org 2007-04-25 16:32
---
Oh, btw. why the may_aliases bitmap for SMT.11 is empty on entry to the loops
computing the transitive closure (it tries to, right?) is that SFT.0 is not
considered stored to in the first nested loop over pointers and addressable
vars. Shouldn't we add the vars NMT aliases to ai-written_vars in
compute_flow_sensitive_aliasing? This also seems to fix the problem:
Index: tree-ssa-alias.c
===
*** tree-ssa-alias.c(revision 124151)
--- tree-ssa-alias.c(working copy)
*** compute_flow_sensitive_aliasing (struct
*** 1142,1152
one). Note that only pointers that have been dereferenced will
have a name memory tag. */
if (pi-name_mem_tag pi-pt_vars)
! EXECUTE_IF_SET_IN_BITMAP (pi-pt_vars, 0, j, bi)
! {
! add_may_alias (pi-name_mem_tag, referenced_var (j));
! add_may_alias (v_ann-symbol_mem_tag, referenced_var (j));
! }
}
}
--- 1142,1159
one). Note that only pointers that have been dereferenced will
have a name memory tag. */
if (pi-name_mem_tag pi-pt_vars)
! {
! bool stored_to = bitmap_bit_p (ai-dereferenced_ptrs_store,
!DECL_UID (SSA_NAME_VAR (ptr)));
! EXECUTE_IF_SET_IN_BITMAP (pi-pt_vars, 0, j, bi)
! {
! add_may_alias (pi-name_mem_tag, referenced_var (j));
! if (stored_to)
! bitmap_set_bit (ai-written_vars,
! DECL_UID (referenced_var (j)));
! add_may_alias (v_ann-symbol_mem_tag, referenced_var (j));
! }
! }
}
}
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
--- Comment #30 from rguenth at gcc dot gnu dot org 2007-04-25 16:42 --- Created an attachment (id=13441) -- (http://gcc.gnu.org/bugzilla/attachment.cgi?id=13441action=view) a patch I like more this one attached, bootstrap/testing in progress. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
--- Comment #31 from dberlin at gcc dot gnu dot org 2007-04-25 17:01 --- Subject: Re: [4.2 Regression] -O3 optimizer bug On 25 Apr 2007 15:32:41 -, rguenth at gcc dot gnu dot org [EMAIL PROTECTED] wrote: --- Comment #29 from rguenth at gcc dot gnu dot org 2007-04-25 16:32 --- Oh, btw. why the may_aliases bitmap for SMT.11 is empty on entry to the loops computing the transitive closure (it tries to, right?) is that SFT.0 is not considered stored to in the first nested loop over pointers and addressable vars. Shouldn't we add the vars NMT aliases to ai-written_vars in compute_flow_sensitive_aliasing? This is a better fix. The other fix is just a hack. Aliasing is not symmetric, so we shouldn't make the two points-to sets equal. However, I bet if you look at where we initially add written vars, you'll discover we are missing it there for some reason, and don't need to post-add it like you do here. I think i fixed something like this in 4.3 where we didn't consider stores to globals or arguments to be written or somesuch when we were initially generating written vars. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
--- Comment #32 from rguenth at gcc dot gnu dot org 2007-04-25 17:13
---
No idea. The only place I found was setup_pointers_and_addressables, but that
hits the path only if
/* Add pointer variables that have been dereferenced to the POINTERS
array and create a symbol memory tag for them. */
if (POINTER_TYPE_P (TREE_TYPE (var)))
{
but we don't have a pointer through which SMT we do store to SFT.0, and we
don't seem to care about NMTs and their aliases in this loop at all.
(I'm lost here now, I'll post the patch and the result of bootstrap test
tomorrow -- I'd love to see a better approach to fixing this ;)
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
--- Comment #33 from dberlin at gcc dot gnu dot org 2007-04-25 19:45
---
I think richi said on IRC that the following backport from 4.3 will fix it (if
so, it's the correct fix here)
Index: tree-ssa-structalias.c
===
--- tree-ssa-structalias.c (revision 122853)
+++ tree-ssa-structalias.c (working copy)
@@ -3228,7 +3228,8 @@ update_alias_info (tree stmt, struct ali
/* If the statement makes a function call, assume
that pointer OP will be dereferenced in a store
operation inside the called function. */
- if (get_call_expr_in (stmt))
+ if (get_call_expr_in (stmt)
+ || stmt_escape_type == ESCAPE_STORED_IN_GLOBAL)
{
bitmap_set_bit (ai-dereferenced_ptrs_store, DECL_UID (var));
pi-is_dereferenced = 1;
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
-- mmitchel at gcc dot gnu dot org changed: What|Removed |Added Priority|P3 |P1 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
--- Comment #25 from dberlin at gcc dot gnu dot org 2007-04-25 03:14 --- (In reply to comment #23) This is a regression. Danny? It actually should get assigned anything as a points-to set, so the bad constraints are correct. We should also always get correct aliasing even if everything was assigned ANYTHING as a variable. Sadly, we don't in 4.2.0, as this case shows. I'm not going to get a chance to look at this for at least week, I have something taking priority this week. If you want to tackle it before then, I would suggest trying to see why we decide not to give the SFT's to the variable marked with anything. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
--- Comment #23 from rguenth at gcc dot gnu dot org 2007-04-21 18:07 --- This is a regression. Danny? -- rguenth at gcc dot gnu dot org changed: What|Removed |Added CC||rguenth at gcc dot gnu dot ||org Known to work||4.1.3 4.3.0 Summary|-O3 optimizer bug |[4.2 Regression] -O3 ||optimizer bug http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
[Bug tree-optimization/30567] [4.2 Regression] -O3 optimizer bug
--- Comment #24 from rguenth at gcc dot gnu dot org 2007-04-21 18:38
---
Actually the handle_ptr_arith change made a difference as we (luckily?)
for D.2147_17 = D.2144_14 + D.2146_16
D.2144_14 = a_11-begin;
D.2145_15 = i_1 * 4;
D.2146_16 = (const int *) D.2145_15;
D.2147_17 = D.2144_14 + D.2146_16;
add a constraint for offset zero (that's pure luck) for D.2144_14. So, the
difference in constraints good vs. bad is
@@ -26,10 +26,10 @@
D.2144_14 = *a_11
D.2146_16 = ANYTHING
D.2147_17 = D.2144_14
+D.2147_17 = D.2146_16
ESCAPED_VARS = D.2092_25
Collapsing static cycles and doing variable substitution:
-Collapsing D.2147_17 into D.2144_14
Collapsing this_7 into D.2140_6
Collapsing NONLOCAL.8 into ESCAPED_VARS
@@ -54,7 +54,7 @@
a_11 = { D.2142 }
D.2144_14 = { D.2141 D.2141.32 }
D.2146_16 = { ANYTHING }
-D.2147_17 = { D.2141 D.2141.32 }
+D.2147_17 = { ANYTHING D.2141 D.2141.32 }
D.2092_25 = { }
main: Total number of aliased vops: 3
Unrelated to this the following resolves possible problems in handle_ptr_arith:
Index: tree-ssa-structalias.c
===
--- tree-ssa-structalias.c (revision 124018)
+++ tree-ssa-structalias.c (working copy)
@@ -3287,7 +3287,7 @@ handle_ptr_arith (VEC (ce_s, heap) *lhsc
unsigned int i = 0;
unsigned int j = 0;
VEC (ce_s, heap) *temp = NULL;
- unsigned int rhsoffset = 0;
+ unsigned HOST_WIDE_INT rhsoffset = 0;
if (TREE_CODE (expr) != PLUS_EXPR
TREE_CODE (expr) != MINUS_EXPR)
@@ -3298,8 +3298,10 @@ handle_ptr_arith (VEC (ce_s, heap) *lhsc
get_constraint_for (op0, temp);
if (POINTER_TYPE_P (TREE_TYPE (op0))
- TREE_CODE (op1) == INTEGER_CST
- TREE_CODE (expr) == PLUS_EXPR)
+ host_integerp (op1, 0)
+ TREE_CODE (expr) == PLUS_EXPR
+ /* Make sure the pointer offset is positive. */
+ tree_int_cst_msb (op1) == 0)
{
rhsoffset = TREE_INT_CST_LOW (op1) * BITS_PER_UNIT;
}
though that last multiplication may still overflow.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30567
