Hi,
the testcase triggers yet another case where we do not update profile
type correctly when inter-proceduraly inlining because of 0 in the
entry_block_ptr count.
Bootstrapped/regtested x86_64-linux, will commit it shortly.
Honza
* ipa-inline-transform.c (update_noncloned_frequencies): Use
profile_count::adjust_for_ipa_scaling.
* tree-inline.c (copy_bb, copy_cfg_body): Likewise.
* profile-count.c (profile_count::adjust_for_ipa_scaling): New member
function.
* profile-count.h (profile_count::adjust_for_ipa_scaling): Declare.
* gcc.c-torture/compile/pr82879.c: New testcase.
Index: ipa-inline-transform.c
===================================================================
--- ipa-inline-transform.c (revision 254567)
+++ ipa-inline-transform.c (working copy)
@@ -60,17 +60,7 @@ update_noncloned_frequencies (struct cgr
{
struct cgraph_edge *e;
- /* We always must scale to be sure counters end up compatible.
- If den is zero, just force it nonzero and hope for reasonable
- approximation.
- When num is forced nonzero, also update den, so we do not scale profile
- to 0. */
- if (!(num == den)
- && !(den.force_nonzero () == den))
- {
- den = den.force_nonzero ();
- num = num.force_nonzero ();
- }
+ profile_count::adjust_for_ipa_scaling (&num, &den);
/* We do not want to ignore high loop nest after freq drops to 0. */
if (!freq_scale)
Index: profile-count.c
===================================================================
--- profile-count.c (revision 254567)
+++ profile-count.c (working copy)
@@ -255,3 +255,29 @@ profile_count::to_cgraph_frequency (prof
return CGRAPH_FREQ_MAX;
return MIN (scale, CGRAPH_FREQ_MAX);
}
+
+/* We want to scale profile across function boundary from NUM to DEN.
+ Take care of the side case when DEN is zeros. We still want to behave
+ sanely here which means
+ - scale to profile_count::zero () if NUM is profile_count::zero
+ - do not affect anything if NUM == DEN
+ - preserve counter value but adjust quality in other cases. */
+
+void
+profile_count::adjust_for_ipa_scaling (profile_count *num,
+ profile_count *den)
+{
+ /* Scaling is no-op if NUM and DEN are the same. */
+ if (*num == *den)
+ return;
+ /* Scaling to zero is always zeor. */
+ if (*num == profile_count::zero ())
+ return;
+ /* If den is non-zero we are safe. */
+ if (den->force_nonzero () == *den)
+ return;
+ /* Force both to non-zero so we do not push profiles to 0 when
+ both num == 0 and den == 0. */
+ *den = den->force_nonzero ();
+ *num = num->force_nonzero ();
+}
Index: profile-count.h
===================================================================
--- profile-count.h (revision 254567)
+++ profile-count.h (working copy)
@@ -1044,6 +1044,11 @@ public:
/* Return true if THIS is known to differ significantly from OTHER. */
bool differs_from_p (profile_count other) const;
+ /* We want to scale profile across function boundary from NUM to DEN.
+ Take care of the side case when NUM and DEN are zeros of incompatible
+ kinds. */
+ static void adjust_for_ipa_scaling (profile_count *num, profile_count *den);
+
/* LTO streaming support. */
static profile_count stream_in (struct lto_input_block *);
void stream_out (struct output_block *);
Index: testsuite/gcc.c-torture/compile/pr82879.c
===================================================================
--- testsuite/gcc.c-torture/compile/pr82879.c (revision 0)
+++ testsuite/gcc.c-torture/compile/pr82879.c (working copy)
@@ -0,0 +1,11 @@
+int a, b;
+static __attribute__((cold)) void fn1() {
+ for (;;)
+ for (; a;)
+ ;
+}
+void fn2() {
+ if (b)
+ fn1();
+}
+
Index: tree-inline.c
===================================================================
--- tree-inline.c (revision 254567)
+++ tree-inline.c (working copy)
@@ -1771,17 +1771,7 @@ copy_bb (copy_body_data *id, basic_block
tree decl;
basic_block prev;
- /* We always must scale to be sure counters end up compatible.
- If den is zero, just force it nonzero and hope for reasonable
- approximation.
- When num is forced nonzero, also update den, so we do not scale profile
- to 0. */
- if (!(num == den)
- && !(den.force_nonzero () == den))
- {
- den = den.force_nonzero ();
- num = num.force_nonzero ();
- }
+ profile_count::adjust_for_ipa_scaling (&num, &den);
/* Search for previous copied basic block. */
prev = bb->prev_bb;
@@ -2698,17 +2688,7 @@ copy_cfg_body (copy_body_data * id, prof
profile_count den = ENTRY_BLOCK_PTR_FOR_FN (src_cfun)->count;
profile_count num = entry_block_map->count;
- /* We always must scale to be sure counters end up compatible.
- If den is zero, just force it nonzero and hope for reasonable
- approximation.
- When num is forced nonzero, also update den, so we do not scale profile
- to 0. */
- if (!(num == den)
- && !(den.force_nonzero () == den))
- {
- den = den.force_nonzero ();
- num = num.force_nonzero ();
- }
+ profile_count::adjust_for_ipa_scaling (&num, &den);
cfun_to_copy = id->src_cfun = DECL_STRUCT_FUNCTION (callee_fndecl);
