I initially ran into this while reviving autoprofiledbootstrap build.
I was able to create a simple reliable test for this bug and captured it in
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108000
I also included the test case in the updated patch below.
Eugene
=
The existing comparison was incorrect for non-PRECISE counts
(e.g., AFDO): we could end up with a 0 base_count, which could
lead to asserts, e.g., in good_cloning_opportunity_p.
Tested on x86_64-pc-linux-gnu.
gcc/ChangeLog:
PR ipa/108000
* ipa-cp.cc (ipcp_propagate_stage): Fix profile count comparison
gcc/testsuite
* gcc.dg/tree-prof/pr108000.c: Regression test
---
gcc/ipa-cp.cc | 2 +-
gcc/testsuite/gcc.dg/tree-prof/pr108000.c | 93 +++
2 files changed, 94 insertions(+), 1 deletion(-)
create mode 100644 gcc/testsuite/gcc.dg/tree-prof/pr108000.c
diff --git a/gcc/ipa-cp.cc b/gcc/ipa-cp.cc
index d5230c7c5e6..cc031ebed0f 100644
--- a/gcc/ipa-cp.cc
+++ b/gcc/ipa-cp.cc
@@ -4225,7 +4225,7 @@ ipcp_propagate_stage (class ipa_topo_info *topo)
for (cgraph_edge *cs = node->callees; cs; cs = cs->next_callee)
{
profile_count count = cs->count.ipa ();
- if (!(count > profile_count::zero ()))
+ if (!count.nonzero_p ())
continue;
enum availability avail;
diff --git a/gcc/testsuite/gcc.dg/tree-prof/pr108000.c
b/gcc/testsuite/gcc.dg/tree-prof/pr108000.c
new file mode 100644
index 000..c59ea799748
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-prof/pr108000.c
@@ -0,0 +1,93 @@
+/* { dg-options "-O2" } */
+
+#include
+
+volatile int flag;
+const int array_size = 10;
+int* array;
+int iterations = 1000;
+
+#define BAR(num) \
+int __attribute__((noinline)) \
+bar##num (int i, int j) \
+{ \
+ if (i == 0) \
+return 2*num - 1; \
+ else \
+return 2*num; \
+}
+
+BAR(1)
+BAR(2)
+BAR(3)
+BAR(4)
+BAR(5)
+BAR(6)
+BAR(7)
+BAR(8)
+BAR(9)
+BAR(10)
+BAR(11)
+BAR(12)
+BAR(13)
+BAR(14)
+BAR(15)
+BAR(16)
+BAR(17)
+BAR(18)
+BAR(19)
+
+int __attribute__((noinline))
+foo ()
+{
+ switch (flag)
+ {
+ case 1:
+ return bar1 (0, 0);
+ case 2:
+ return bar2 (0, 0);
+ case 3:
+ return bar3 (0, 0);
+ case 4:
+ return bar4 (0, 0);
+ case 5:
+ return bar5 (0, 0);
+ case 6:
+ return bar6 (0, 0);
+ case 7:
+ return bar7 (0, 0);
+ case 8:
+ return bar8 (0, 0);
+ case 9:
+ return bar9 (0, 0);
+ case 10:
+ return bar10 (0, 0);
+ case 11:
+ return bar11 (0, 0);
+ case 12:
+ return bar12 (0, 0);
+ case 13:
+ return bar13 (0, 0);
+ case 14:
+ return bar14 (0, 0);
+ case 15:
+ return bar15 (0, 0);
+ case 16:
+ return bar16 (0, 0);
+ case 17:
+ return bar17 (0, 0);
+ case 18:
+ return bar18 (0, 0);
+ default:
+ return bar19(0, 0);
+ }
+}
+
+int
+main ()
+{
+ flag = 0;
+ array = calloc(array_size, sizeof(int));
+ for (int i = 0, j = 0; i < iterations; ++i, j = (j + 1) % 10)
+array[j] = foo ();
+}
--
2.25.1
-Original Message-
From: Jeff Law
Sent: Tuesday, November 22, 2022 12:03 PM
To: Eugene Rozenfeld ; gcc-patches@gcc.gnu.org
Subject: [EXTERNAL] Re: [PATCH] Fix count comparison in ipa-cp
[You don't often get email from jeffreya...@gmail.com. Learn why this is
important at https://aka.ms/LearnAboutSenderIdentification ]
On 11/21/22 14:26, Eugene Rozenfeld via Gcc-patches wrote:
> The existing comparison was incorrect for non-PRECISE counts (e.g.,
> AFDO): we could end up with a 0 base_count, which could lead to
> asserts, e.g., in good_cloning_opportunity_p.
>
> gcc/ChangeLog:
>
> * ipa-cp.cc (ipcp_propagate_stage): Fix profile count comparison.
OK. Probably somewhat painful to pull together a reliable test for this, right?
Jeff
