diff options
| author | Xinliang David Li <davidxl@google.com> | 2011-04-08 21:03:12 +0000 |
|---|---|---|
| committer | Xinliang David Li <davidxl@gcc.gnu.org> | 2011-04-08 21:03:12 +0000 |
| commit | 041e059f264744313dc38f2223098409f8d86c29 (patch) | |
| tree | 82219d1423d58d8f8c24a68e0c7c5cf69f26f652 | |
| parent | 6a801cf261e01c7b3bb6e51dc7e34fa0cf61651a (diff) | |
| download | gcc-041e059f264744313dc38f2223098409f8d86c29.zip gcc-041e059f264744313dc38f2223098409f8d86c29.tar.gz gcc-041e059f264744313dc38f2223098409f8d86c29.tar.bz2 | |
FDO insane profile
From-SVN: r172213
| -rw-r--r-- | gcc/ChangeLog | 5 | ||||
| -rw-r--r-- | gcc/ipa-cp.c | 23 |
2 files changed, 28 insertions, 0 deletions
diff --git a/gcc/ChangeLog b/gcc/ChangeLog index 3abeed0..6d7b907 100644 --- a/gcc/ChangeLog +++ b/gcc/ChangeLog @@ -1,5 +1,10 @@ 2011-04-08 Xinliang David Li <davidxl@google.com> + * ipa-cp.c (ipcp_update_profiling): Correct + negative scale factor due to insane profile data. + +2011-04-08 Xinliang David Li <davidxl@google.com> + * final.c (dump_basic_block_info): New function. (final): Dump basic block. (final_scan_insn): Remove old dump. diff --git a/gcc/ipa-cp.c b/gcc/ipa-cp.c index db0941c..2b1e50dc 100644 --- a/gcc/ipa-cp.c +++ b/gcc/ipa-cp.c @@ -1113,6 +1113,29 @@ ipcp_update_profiling (void) scale = ipcp_get_node_scale (orig_node); node->count = orig_node->count * scale / REG_BR_PROB_BASE; scale_complement = REG_BR_PROB_BASE - scale; + + /* Negative scale complement can result from insane profile data + in which the total incoming edge counts in this module is + larger than the callee's entry count. The insane profile data + usually gets generated due to the following reasons: + + 1) in multithreaded programs, when profile data is dumped + to gcda files in gcov_exit, some other threads are still running. + The profile counters are dumped in bottom up order (call graph). + The caller's BB counters may still be updated while the callee's + counter data is already saved to disk. + + 2) Comdat functions: comdat functions' profile data are not + allocated in comdat. When a comdat callee function gets inlined + at some callsites after instrumentation, and the remaining calls + to this function resolves to a comdat copy in another module, + the profile counters for this function are split. This can + result in sum of incoming edge counts from this module being + larger than callee instance's entry count. */ + + if (scale_complement < 0 && flag_profile_correction) + scale_complement = 0; + orig_node->count = orig_node->count * scale_complement / REG_BR_PROB_BASE; for (cs = node->callees; cs; cs = cs->next_callee) |