There's quite some instability in the lattice when match-and-simplify
is mixed with CCP. We deal with part of that already by meeting
with the old value in set_lattice_value, the following deals with this
new case by doing it always.
SSA propagation and match-and-simplify don't play well with each
other, a solution would be to forgo with that and use a simpler
iteration scheme.
Bootstrapped and tested on x86_64-unknown-linux-gnu, applied.
Richard.
2017-05-26 Richard Biener <rguent...@suse.de>
PR tree-optimization/80842
* tree-ssa-ccp.c (set_lattice_value): Always meet with the old
value.
* gcc.dg/torture/pr80842.c: New testcase.
Index: gcc/tree-ssa-ccp.c
===================================================================
--- gcc/tree-ssa-ccp.c (revision 248409)
+++ gcc/tree-ssa-ccp.c (working copy)
@@ -497,9 +497,7 @@ set_lattice_value (tree var, ccp_prop_va
use the meet operator to retain a conservative value.
Missed optimizations like PR65851 makes this necessary.
It also ensures we converge to a stable lattice solution. */
- if (new_val->lattice_val == CONSTANT
- && old_val->lattice_val == CONSTANT
- && TREE_CODE (new_val->value) != SSA_NAME)
+ if (old_val->lattice_val != UNINITIALIZED)
ccp_lattice_meet (new_val, old_val);
gcc_checking_assert (valid_lattice_transition (*old_val, *new_val));
Index: gcc/testsuite/gcc.dg/torture/pr80842.c
===================================================================
--- gcc/testsuite/gcc.dg/torture/pr80842.c (nonexistent)
+++ gcc/testsuite/gcc.dg/torture/pr80842.c (working copy)
@@ -0,0 +1,23 @@
+/* { dg-do compile } */
+
+unsigned a;
+short b;
+char c, d, e;
+void fn1();
+void fn2() {
+ a++;
+ for (; a;)
+ fn1(0, 0);
+}
+void fn3() {
+ fn2();
+l1:;
+ unsigned char f;
+ short g;
+ unsigned char *h = &f;
+ g += &h ? e ? g = 1 : 0 : 0;
+ d = g;
+ c *f;
+ if (d & (b %= *h) < f * d / (d -= 0))
+ goto l1;
+}
