https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121074
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Well, so it's coverage_init opening the gcno file which is indeed never closed
and then afdo trying to open it's own in the afdo_offline pass.
#0 gcov_open (name=0x4eb7b60 "a-t.gcno", mode=-1)
at /space/rguenther/src/gcc-clean/gcc/gcov-io.cc:166
#1 0x000000000130a038 in coverage_init (filename=0x4f10940 "a-t")
at /space/rguenther/src/gcc-clean/gcc/coverage.cc:1334
#2 0x0000000001a994ff in do_compile ()
at /space/rguenther/src/gcc-clean/gcc/toplev.cc:2218
#3 0x0000000001a9999b in toplev::main (this=0x7fffffffdb9a, argc=24,
argv=0x7fffffffdcc8) at /space/rguenther/src/gcc-clean/gcc/toplev.cc:2389
#4 0x0000000003821d5a in main (argc=24, argv=0x7fffffffdcc8)
at /space/rguenther/src/gcc-clean/gcc/main.cc:39
it seems that AFDO and -ftest-coverage should simply conflict? Or the
single-file restriction of the gcov API needs to be lifted.
There's an overlap I guess - calling coverage_finish from ipa_tree_profile
would not help:
NEXT_PASS (pass_ipa_auto_profile);
PUSH_INSERT_PASSES_WITHIN (pass_ipa_auto_profile)
NEXT_PASS (pass_feedback_split_functions);
POP_INSERT_PASSES ()
NEXT_PASS (pass_ipa_tree_profile);
PUSH_INSERT_PASSES_WITHIN (pass_ipa_tree_profile)
NEXT_PASS (pass_feedback_split_functions);
POP_INSERT_PASSES ()
NEXT_PASS (pass_ipa_free_fn_summary, true /* small_p */);
NEXT_PASS (pass_ipa_increase_alignment);
NEXT_PASS (pass_ipa_tm);
NEXT_PASS (pass_ipa_lower_emutls);
NEXT_PASS (pass_ipa_asm_lgen);
TERMINATE_PASS_LIST (all_small_ipa_passes)
we're currently calling coverage_finish after all_small_ipa_passes.
IMO a sorry() from finish_options would be appropriate?
On the 15 branch we did either coverage or AFDO with reading the data at
coverage_init time:
if (flag_auto_profile)
read_autofdo_file ();
else if (flag_branch_probabilities)
read_counts_file ();
but now we consume both? Part of a fix might possibly want to include
diff --git a/gcc/coverage.cc b/gcc/coverage.cc
index f2a095a1279..0f7fe09d865 100644
--- a/gcc/coverage.cc
+++ b/gcc/coverage.cc
@@ -1316,7 +1316,7 @@ coverage_init (const char *filename)
strcpy (da_file_name + prefix_len + len, GCOV_DATA_SUFFIX);
bbg_file_stamp = local_tick;
- if (flag_branch_probabilities)
+ if (flag_branch_probabilities && !flag_auto_profile)
read_counts_file ();
/* Name of bbg file. */
but the problematical part is the delayed read_autofdo_file which happens
when the coverage file is still open. So it's going to be a functional
regression. I'm not sure gcov supports closing & re-opening easily.
