https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79067
Bug ID: 79067
Summary: gcc.dg/tree-prof/cold_partition_label.c runs a million
times longer than it used to and times out
Product: gcc
Version: 7.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: testsuite
Assignee: unassigned at gcc dot gnu.org
Reporter: sandra at gcc dot gnu.org
Target Milestone: ---
The change committed as part of r238325 made
gcc.dg/tree-prof/cold_partition_label.c run 1000000 times as long as it used
to, causing a timeout on a slow embedded target (nios2-linux-gnu):
Index: gcc/testsuite/gcc.dg/tree-prof/cold_partition_label.c
===================================================================
--- gcc/testsuite/gcc.dg/tree-prof/cold_partition_label.c (revision
238324)
+++ gcc/testsuite/gcc.dg/tree-prof/cold_partition_label.c (revision
238325)
@@ -29,9 +29,11 @@ foo (int path)
int
main (int argc, char *argv[])
{
+ int i;
buf_hot = "hello";
buf_cold = "world";
- foo (argc);
+ for (i = 0; i < 1000000; i++)
+ foo (argc);
return 0;
}
Other tests modified in the same commit have similar changes.
I did some experiments with different loop counts, measuring the total clock
time to run the one test as reported by the test harness:
1000 :12
10000 :19
100000 1:32
1000000 timed out after 5:11
I'm not sure how to fix this. I looked at the discussion on the mailing list
about the patch that caused the regression and saw a comment that, even with
the high counts, these tests don't run long enough to get good auto-fdo results
on the targets that support that feature, so cutting down the count across the
board doesn't seem like a good idea. Nor does simply bumping up
dg-timeout-factor across the board-- why waste 10+ minutes running this test on
a slow target like nios2-linux-gnu that doesn't even support the auto-fdo
feature it's trying to test? Maybe add something to target-supports.exp to
test for the presence of fdo support and use it to either skip the tests or
control a -D option on the command-line to set the number of iterations?