| Message ID | 791bc219-eba0-ee14-9007-7a000c5786ff@suse.cz |
|---|---|
| State | New |
| Headers | show |
| Series | Fix missing profiles with PGO (PR tree-optimization/83051). | expand |
> Hello. > > Following ICE can be seen when we have -fprofile-generate or -fprofile-use w/ missing > gcda file. I hope the proper fix is to check for reliable profile. > > Patch can bootstrap on ppc64le-redhat-linux and survives regression tests. > > Ready to be installed? > Martin > > gcc/ChangeLog: > > 2018-01-19 Martin Liska <mliska@suse.cz> > > PR tree-optimization/83051 > * predict.c (handle_missing_profiles): Consider profile only > if it's reliable. > > gcc/testsuite/ChangeLog: > > 2018-01-19 Martin Liska <mliska@suse.cz> > > PR tree-optimization/83051 > * gcc.dg/torture/pr83055.c: New test. Hi, the problem here is that zero profiles are detected incorrectly because they should be done at IPA level. While doing that I also noticed that drop_profile has bug where it does not set up node's profile that eventually leads to ICE in IRA if the logic for dropping profiles is bad. Bootstrapped/regtested x86_64-linux, will commit it shortly. PR middle-end/83055 * predict.c (drop_profile): Do not push/pop cfun; update also node->count. (handle_missing_profiles): Fix logic looking for zero profiles. * gcc.dg/torture/pr83055.c: New testcase. Index: predict.c =================================================================== --- predict.c (revision 257010) +++ predict.c (working copy) @@ -3311,32 +3311,28 @@ drop_profile (struct cgraph_node *node, } basic_block bb; - push_cfun (DECL_STRUCT_FUNCTION (node->decl)); - if (flag_guess_branch_prob) + if (opt_for_fn (node->decl, flag_guess_branch_prob)) { bool clear_zeros - = ENTRY_BLOCK_PTR_FOR_FN - (DECL_STRUCT_FUNCTION (node->decl))->count.nonzero_p (); + = !ENTRY_BLOCK_PTR_FOR_FN (fn)->count.nonzero_p (); FOR_ALL_BB_FN (bb, fn) if (clear_zeros || !(bb->count == profile_count::zero ())) bb->count = bb->count.guessed_local (); - DECL_STRUCT_FUNCTION (node->decl)->cfg->count_max = - DECL_STRUCT_FUNCTION (node->decl)->cfg->count_max.guessed_local (); + fn->cfg->count_max = fn->cfg->count_max.guessed_local (); } else { FOR_ALL_BB_FN (bb, fn) bb->count = profile_count::uninitialized (); - DECL_STRUCT_FUNCTION (node->decl)->cfg->count_max - = profile_count::uninitialized (); + fn->cfg->count_max = profile_count::uninitialized (); } - pop_cfun (); struct cgraph_edge *e; for (e = node->callees; e; e = e->next_callee) e->count = gimple_bb (e->call_stmt)->count; for (e = node->indirect_calls; e; e = e->next_callee) e->count = gimple_bb (e->call_stmt)->count; + node->count = ENTRY_BLOCK_PTR_FOR_FN (fn)->count; profile_status_for_fn (fn) = (flag_guess_branch_prob ? PROFILE_GUESSED : PROFILE_ABSENT); @@ -3373,12 +3369,12 @@ handle_missing_profiles (void) gcov_type max_tp_first_run = 0; struct function *fn = DECL_STRUCT_FUNCTION (node->decl); - if (!(node->count == profile_count::zero ())) + if (node->count.ipa ().nonzero_p ()) continue; for (e = node->callers; e; e = e->next_caller) - if (e->count.initialized_p () && e->count > 0) + if (e->count.ipa ().initialized_p () && e->count.ipa () > 0) { - call_count = call_count + e->count; + call_count = call_count + e->count.ipa (); if (e->caller->tp_first_run > max_tp_first_run) max_tp_first_run = e->caller->tp_first_run; @@ -3411,7 +3407,8 @@ handle_missing_profiles (void) struct cgraph_node *callee = e->callee; struct function *fn = DECL_STRUCT_FUNCTION (callee->decl); - if (callee->count > 0) + if (!(e->count.ipa () == profile_count::zero ()) + && callee->count.ipa ().nonzero_p ()) continue; if ((DECL_COMDAT (callee->decl) || DECL_EXTERNAL (callee->decl)) && fn && fn->cfg Index: testsuite/gcc.dg/torture/pr83055.c =================================================================== --- testsuite/gcc.dg/torture/pr83055.c (revision 0) +++ testsuite/gcc.dg/torture/pr83055.c (working copy) @@ -0,0 +1,13 @@ +/* { dg-do compile } */ +/* { dg-additional-options "-fprofile-generate" } */ + +void __attribute__ ((__cold__)) a (void); +void b (void); +void __attribute__ ((noinline)) c (void) { a (); } + +void +d (void) +{ + b (); + c (); +}
diff --git a/gcc/predict.c b/gcc/predict.c index 4c1e4489b55..c5144e49c8d 100644 --- a/gcc/predict.c +++ b/gcc/predict.c @@ -3357,7 +3357,7 @@ handle_missing_profiles (void) if (!(node->count == profile_count::zero ())) continue; for (e = node->callers; e; e = e->next_caller) - if (e->count.initialized_p () && e->count > 0) + if (e->count.initialized_p () && e->count.reliable_p ()) { call_count = call_count + e->count; diff --git a/gcc/testsuite/gcc.dg/torture/pr83055.c b/gcc/testsuite/gcc.dg/torture/pr83055.c new file mode 100644 index 00000000000..9bc71c6cddf --- /dev/null +++ b/gcc/testsuite/gcc.dg/torture/pr83055.c @@ -0,0 +1,13 @@ +/* { dg-do compile } */ +/* { dg-additional-options "-fprofile-generate" } */ + +void __attribute__ ((__cold__)) a (void); +void b (void); +void __attribute__ ((noinline)) c (void) { a (); } + +void +d (void) +{ + b (); + c (); +}