Hi,
IPA-CP currently assumes that cloning estimates always have some
positive size cost. However, there are apparently situations in which
estimate_ipcp_clone_size_and_time does return zero size and which then
mostly lead to divisions by zero or failed asserts. This patch avoids
that by simply bumping the sizes to 1 in those cases.
Bootstrapped and tested on x86_64-linux. OK for trunk?
Thanks,
Martin
2011-12-12 Martin Jambor <mjam...@suse.cz>
PR tree-optimization/51362
* ipa-cp.c (estimate_local_effects): When estimated size of a
specialized clone is zero, bump it to one.
* testsuite/gcc.dg/ipa/pr51362.c: New test.
Index: src/gcc/ipa-cp.c
===================================================================
--- src.orig/gcc/ipa-cp.c
+++ src/gcc/ipa-cp.c
@@ -1409,6 +1409,14 @@ estimate_local_effects (struct cgraph_no
+ devirtualization_time_bonus (node, known_csts, known_binfos)
+ removable_params_cost + emc;
+ gcc_checking_assert (size >=0);
+ /* The inliner-heuristics based estimates may think that in certain
+ contexts some functions do not have any size at all but we want
+ all specializations to have at least a tiny cost, not least not to
+ divide by zero. */
+ if (size == 0)
+ size = 1;
+
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, " - estimates for value ");
Index: src/gcc/testsuite/gcc.dg/ipa/pr51362.c
===================================================================
--- /dev/null
+++ src/gcc/testsuite/gcc.dg/ipa/pr51362.c
@@ -0,0 +1,22 @@
+/* { dg-do compile } */
+/* { dg-options "-O -fipa-cp -fipa-cp-clone" } */
+
+int
+baz (void)
+{
+ return 0;
+}
+
+int make_mess;
+
+__attribute__ ((noinline))
+int bar (int x, int (*f) (void))
+{
+ return f ();
+}
+
+int
+foo (void)
+{
+ return bar (1, baz);
+}
