During WPA we merge 2 functions where one is built with -O3
-fprofile-use while the other one uses -O0. We end up with:
prevailing node:
BitwiseCast (const double aFrom)
{
long unsigned int temp;
long unsigned int D.4528;
<bb 2> :
BitwiseCast (aFrom_2(D), &temp);
_4 = temp;
temp ={v} {CLOBBER};
<bb 3> :
<L0>:
return _4;
}
merged node:
BitwiseCast (const double aFrom)
{
<bb 2> [count: 1509]:
_3 = VIEW_CONVERT_EXPR<long unsigned int>(aFrom_2(D));
return _3;
}
What happens is that we copy dst->count = src->count.ipa() but
we skip merging of BBs (a different CFG). Then we end up in situation
where profile of a ENTRY_BB_FOR_FN == uninitialized while
dst->count.quality is precise.
Patch can bootstrap on x86_64-linux-gnu and survives regression tests.
Ready to be installed?
Thanks,
Martin
gcc/ChangeLog:
PR ipa/97295
* ipa-utils.c (ipa_merge_profiles):
---
gcc/ipa-utils.c | 8 ++++----
1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/gcc/ipa-utils.c b/gcc/ipa-utils.c
index 23e7f714306..fe2149069be 100644
--- a/gcc/ipa-utils.c
+++ b/gcc/ipa-utils.c
@@ -437,10 +437,7 @@ ipa_merge_profiles (struct cgraph_node *dst,
else if (dst->count.ipa ().initialized_p ())
;
else if (src->count.ipa ().initialized_p ())
- {
- copy_counts = true;
- dst->count = src->count.ipa ();
- }
+ copy_counts = true;
/* If no updating needed return early. */
if (dst->count == orig_count)
@@ -620,6 +617,9 @@ ipa_merge_profiles (struct cgraph_node *dst,
bool dstscale = !copy_counts
&& dstnum.initialized_p () && !(dstnum == dstden);
+ if (copy_counts)
+ dst->count = src->count.ipa ();
+
/* TODO: merge also statement histograms. */
FOR_ALL_BB_FN (srcbb, srccfun)
{
--
2.28.0
