[Bug tree-optimization/67191] [6 Regression] ICE: in before_dom_children, at tree-ssa-sccvn.c:4372
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67191 --- Comment #6 from Richard Biener rguenth at gcc dot gnu.org --- Ok, reproduced with a cross. The assert means the DOM walk didn't visit the definition of an operand before its use which shouldn't happen. The use is in a loop nested in a loop which contains the definition dominating the nested loop entry. Ah, so we are detecting the outer loop as not executable. But then the DOM walk goes to a CFG part looking like bb10 /\ / bb11 bb28 |\ | bb20 \ |/loop exit bb13 |\ | other loop exit loop latch and for some reason it visits this in order 10, 28, 13(!) instead of what I would expect (10, 28, 11, 20, 13). So we don't see bb13 as not executable and thus the assert triggers. In domwalk the sort using the postorder number is supposed to make sure that CFG merges inside the dominator children are visited last. Somehow this doesn't work here and thus results in non-optimal order for propagation. Of course due to the CFG the children are only 13, 28 and 11 (not 20) so there isn't a merge within the children :/ In the end the assert wasn't meant to stay - it's fine if it triggers for code that ends up being unreachable. So the immediate fix is to remove it but we still want to fix that dominator walk ordering issue IMHO, which of course would make it a non-dominator walk in some sense (walking bb11 children before completing bb10 children). PRE uses a worklist approach and defers blocks which have unvisited predecessors exactly to overcome this ordering issue which is bad for propagation passes (if only for compile-time reasons).
[Bug tree-optimization/67191] [6 Regression] ICE: in before_dom_children, at tree-ssa-sccvn.c:4372
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67191 --- Comment #7 from Richard Biener rguenth at gcc dot gnu.org --- Author: rguenth Date: Thu Aug 13 09:39:50 2015 New Revision: 226854 URL: https://gcc.gnu.org/viewcvs?rev=226854root=gccview=rev Log: 2015-08-13 Richard Biener rguent...@suse.de PR tree-optimization/67191 * tree-ssa-sccvn.c (sccvn_dom_walker::before_dom_children): Remove assert we value-numbered last stmts operand because it can validly trigger for unreachable code. * gcc.dg/torture/pr67191.c: New testcase. * g++.dg/torture/pr67191.C: Likewise. Added: trunk/gcc/testsuite/g++.dg/torture/pr67191.C trunk/gcc/testsuite/gcc.dg/torture/pr67191.c Modified: trunk/gcc/ChangeLog trunk/gcc/testsuite/ChangeLog trunk/gcc/tree-ssa-sccvn.c
[Bug tree-optimization/67191] [6 Regression] ICE: in before_dom_children, at tree-ssa-sccvn.c:4372
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67191 Richard Biener rguenth at gcc dot gnu.org changed: What|Removed |Added Status|ASSIGNED|RESOLVED Resolution|--- |FIXED --- Comment #8 from Richard Biener rguenth at gcc dot gnu.org --- Fixed.
[Bug tree-optimization/67191] [6 Regression] ICE: in before_dom_children, at tree-ssa-sccvn.c:4372
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67191 --- Comment #3 from Richard Biener rguenth at gcc dot gnu.org --- Oh, maybe just the reduced testcase no longer fails after r226814.
[Bug tree-optimization/67191] [6 Regression] ICE: in before_dom_children, at tree-ssa-sccvn.c:4372
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67191 Richard Biener rguenth at gcc dot gnu.org changed: What|Removed |Added Status|NEW |ASSIGNED Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot gnu.org Target Milestone|--- |6.0 --- Comment #2 from Richard Biener rguenth at gcc dot gnu.org --- Hmm, doesn't reproduce on x86_64 for me. Anyway mine, hopefully we'll get a x86_64 testcase as well.
[Bug tree-optimization/67191] [6 Regression] ICE: in before_dom_children, at tree-ssa-sccvn.c:4372
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67191 --- Comment #4 from Markus Trippelsdorf trippels at gcc dot gnu.org --- (In reply to Richard Biener from comment #3) Oh, maybe just the reduced testcase no longer fails after r226814. Still fails on ppc64le after r226814. So a cross should reproduce this.
[Bug tree-optimization/67191] [6 Regression] ICE: in before_dom_children, at tree-ssa-sccvn.c:4372
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67191
Antoine Balestrat antoine.balestrat at gmail dot com changed:
What|Removed |Added
CC||antoine.balestrat at gmail dot
com
--- Comment #5 from Antoine Balestrat antoine.balestrat at gmail dot com ---
Hello ! This simpler C testcase seemingly triggers the exact same ICE at -O1 on
my x86_64 machine :
$ cat tested.c
int a;
void f(void)
{
int b;
for(a=1; a;);
for(; b; b++)
lbl:
b || a;
if(a)
goto lbl;
}
$ xgcc -w -O1 tested.c
tested.c: In function ‘f’:
tested.c:11:1: internal compiler error: in before_dom_children, at
tree-ssa-sccvn.c:4372
}
^
0xc66ffc sccvn_dom_walker::before_dom_children(basic_block_def*)
../../srcdir/gcc/tree-ssa-sccvn.c:4371
0x10c46a1 dom_walker::walk(basic_block_def*)
../../srcdir/gcc/domwalk.c:177
0xc681d6 run_scc_vn(vn_lookup_kind)
../../srcdir/gcc/tree-ssa-sccvn.c:4459
0xc3a3b2 execute
../../srcdir/gcc/tree-ssa-pre.c:4959
Please submit a full bug report,
with preprocessed source if appropriate.
Please include the complete backtrace with any bug report.
See http://gcc.gnu.org/bugs.html for instructions.
[Bug tree-optimization/67191] [6 Regression] ICE: in before_dom_children, at tree-ssa-sccvn.c:4372
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67191
Markus Trippelsdorf trippels at gcc dot gnu.org changed:
What|Removed |Added
Status|UNCONFIRMED |NEW
Last reconfirmed||2015-08-12
Ever confirmed|0 |1
--- Comment #1 from Markus Trippelsdorf trippels at gcc dot gnu.org ---
trippels@gcc2-power8 tools % cat fwd_set_test.ii
template typename class A;
template typename _Tp using __allocator_base = _Tp;
template class T, class = T, class = int, class = __allocator_baseint
class B;
template class T, class H, class P, class A
bool operator==(BT, H, P, A const , BT, H, P, A const );
template class T, class H, class P, class A
bool operator!=(BT, H, P, A const , BT, H, P, A const );
typedef Bint int_multiset;
int a;
template typename struct C {
C(int) {}
};
template typename struct D;
template typename struct K;
struct L : CADint, CAKint {
template typename First, typename Second
L(First, Second)
: CADint(0), CAKint(0) {}
};
template typename Node struct F {
typedef typename Node::node_pointer node_pointer;
node_pointer node_;
F();
F(typename Node::link_pointer p1) : node_(static_castnode_pointer(p1)) {}
void operator++() { node_ = 0; }
int operator!=(F p1) { return node_ != p1.node_; }
};
struct G {
typedef G *link_pointer;
};
struct H {
static int new_bucket_count(int) {
int b;
int *c = 0;
if (a)
b = *c;
return b;
}
};
class functions {
public:
functions(int, int) {}
~functions();
};
template typename Types struct table : functions {
typedef typename Types::policy policy;
typedef Types node_allocator;
typedef Ftypename Types::node iterator;
L allocators_;
int bucket_count_;
int size_;
typename Types::link_pointer get_previous_start() const;
iterator begin() const { return size_ ? get_previous_start() : 0; }
table(int, typename Types::hasher, typename Types::key_equal, node_allocator)
: functions(0, 0), allocators_(0, 0),
bucket_count_(policy::new_bucket_count(0)), size_() {}
};
template typename struct K : G { typedef K *node_pointer; };
struct I {
typedef G *link_pointer;
};
struct J {
typedef I::link_pointer link_pointer;
};
template typename struct D {
typedef int hasher;
typedef int key_equal;
typedef Kint node;
typedef J::link_pointer link_pointer;
typedef H policy;
};
struct M : tableDint {
node_allocator grouped_table_impl_a;
M(int, int) : table(0, 0, 0, grouped_table_impl_a) {}
void equals(M const ) const {
for (iterator d = begin(); d.node_;) {
iterator e;
group_equals(e);
}
}
static int group_equals(iterator p1) {
int f;
iterator g;
for (; g != p1; ++g)
if (find())
if (f)
return 0;
}
static int find();
};
template class, class, class, class class B {
M table_;
public:
B(unsigned long = 0);
friend bool operator==(B const , B const );
friend bool operator!=(B const , B const );
};
template class T, class H, class P, class A
BT, H, P, A::B(unsigned long)
: table_(0, 0) {}
template class T, class H, class P, class A
bool operator==(BT, H, P, A const p1, BT, H, P, A const p2) {
p1.table_.equals(p2.table_);
}
template class T, class H, class P, class A
bool operator!=(BT, H, P, A const p1, BT, H, P, A const p2) {
p1.table_.equals(p2.table_);
}
void use_multiset_fwd_declared_function_typerun() {
int_multiset x, y;
x == y;
x != y;
}
trippels@gcc2-power8 tools % g++ -O2 -c fwd_set_test.ii
fwd_set_test.ii: In function ‘void
use_multiset_fwd_declared_function_typerun()’:
fwd_set_test.ii:111:6: internal compiler error: in before_dom_children, at
tree-ssa-sccvn.c:4372
void use_multiset_fwd_declared_function_typerun() {
^
