On Sat, Jun 8, 2013 at 2:26 AM, Xinliang David Li <davi...@google.com> wrote: > On Fri, Jun 7, 2013 at 6:47 AM, Richard Biener > <richard.guent...@gmail.com> wrote: >> On Fri, Jun 7, 2013 at 3:30 PM, Xinliang David Li <davi...@google.com> wrote: >>> On Fri, Jun 7, 2013 at 2:05 AM, Richard Biener >>> <richard.guent...@gmail.com> wrote: >>>> On Thu, Jun 6, 2013 at 5:10 PM, Dehao Chen <de...@google.com> wrote: >>>>> Hi, Martin, >>>>> >>>>> Yes, your patch can fix my case. Thanks a lot for the fix. >>>>> >>>>> With the fix, value profiling will still promote the wrong indirect >>>>> call target. Though it will not be inlining, but it results in an >>>>> additional check. How about in check_ic_target, after calling >>>>> gimple_check_call_matching_types, we also check if number of args >>>>> match number of params in target->symbol.decl? >>>> >>>> I wonder what's the point in the gimple_check_call_matching_types check >>>> in the profiling case. It's at least no longer >>>> >>>> /* Perform sanity check on the indirect call target. Due to race >>>> conditions, >>>> false function target may be attributed to an indirect call site. If the >>>> call expression type mismatches with the target function's type, >>>> expand_call >>>> may ICE. >>>> >>>> because since the introduction of gimple_call_fntype we will _not_ ICE. >>>> >>>> Thus I argue that check_ic_target should be even removed instead of >>>> enhancing it! >>>> >>> >>> Another reason is what Dehao had mentioned -- wrong target leads to >>> useless transformation. >> >> Sure, but a not wrong in the sense of the predicate does not guarantee >> a useful transformation either. > > The case in reality is very rare -- most of the cases, the > transformation is good. > >> >>>> How does IC profiling determine the called target? That is, what does it >>>> do when the target is not always the same? (because the checking code >>>> talks about race conditions for example) >>> >>> >>> The race condition is the happening at instrumentation time -- the >>> indirect call counters are not thread local. We have seen this a lot >>> in the past that a totally bogus target is attributed to a indirect >>> callsite. >> >> So it simply uses whatever function was called last? Instead of >> using the function that was called most of the time? > > It uses the most frequent target -- but the target id recorded for the > most frequent target might be corrupted and got mapped to a false > target during profile-use.
But that's not due to "race conditions" but rather AutoFDO which isn't even in trunk, right? Anyway, the discussion is probably moot - the patch is ok with me and my argument would be we should use the function in less places. Thanks, Richard. > David > >> >> Richard. >> >>> thanks, >>> >>> David >>>> >>>> Richard. >>>> >>>> >>>>> Thanks, >>>>> Dehao >>>>> >>>>> >>>>> On Thu, Jun 6, 2013 at 7:11 AM, Martin Jambor <mjam...@suse.cz> wrote: >>>>>> >>>>>> Hi, >>>>>> >>>>>> On Tue, Jun 04, 2013 at 05:19:02PM -0700, Dehao Chen wrote: >>>>>> > attached is a testcase that would cause problem when source has >>>>>> > changed: >>>>>> > >>>>>> > $ g++ test.cc -O2 -fprofile-generate -DOLD >>>>>> > $ ./a.out >>>>>> > $ g++ test.cc -O2 -fprofile-use >>>>>> > test.cc:34:1: internal compiler error: in operator[], at vec.h:815 >>>>>> > } >>>>>> > ^ >>>>>> > 0x512740 vec<tree_node*, va_heap, vl_embed>::operator[](unsigned int) >>>>>> > ../../gcc/vec.h:815 >>>>>> > 0x512740 vec<tree_node*, va_heap, vl_ptr>::operator[](unsigned int) >>>>>> > ../../gcc/vec.h:1244 >>>>>> > 0xf24464 vec<tree_node*, va_heap, vl_embed>::operator[](unsigned int) >>>>>> > ../../gcc/vec.h:815 >>>>>> > 0xf24464 vec<tree_node*, va_heap, vl_ptr>::operator[](unsigned int) >>>>>> > ../../gcc/vec.h:1244 >>>>>> > 0xf24464 ipa_get_indirect_edge_target_1 >>>>>> > ../../gcc/ipa-cp.c:1535 >>>>>> > 0x971b9a estimate_edge_devirt_benefit >>>>>> > ../../gcc/ipa-inline-analysis.c:2757 >>>>>> >>>>>> Hm, this seems rather like an omission in ipa_get_indirect_edge_target_1. >>>>>> Since it is called also from inlining, we can have parameter count >>>>>> mismatches... and in fact in non-virtual paths of that function we do >>>>>> check that we don't. Because all callers have to pass known_vals >>>>>> describing all formal parameters of the inline tree root, we should >>>>>> apply the fix below (I've only just started running a bootstrap and >>>>>> testsuite on x86_64, though). >>>>>> >>>>>> OTOH, while I understand that FDO can change inlining sufficiently so >>>>>> that this error occurs, IMHO this should not be caused by outdated >>>>>> profiles but there is somewhere a parameter mismatch in the source. >>>>>> >>>>>> Dehao, can you please check that this patch helps? >>>>>> >>>>>> Richi, if it does and the patch passes bootstrap and tests, is it OK >>>>>> for trunk and 4.8 branch? >>>>>> >>>>>> Thanks and sorry for the trouble, >>>>>> >>>>>> Martin >>>>>> >>>>>> >>>>>> 2013-06-06 Martin Jambor <mjam...@suse.cz> >>>>>> >>>>>> * ipa-cp.c (ipa_get_indirect_edge_target_1): Check that >>>>>> param_index is >>>>>> within bounds at the beginning of the function. >>>>>> >>>>>> Index: src/gcc/ipa-cp.c >>>>>> =================================================================== >>>>>> --- src.orig/gcc/ipa-cp.c >>>>>> +++ src/gcc/ipa-cp.c >>>>>> @@ -1481,7 +1481,8 @@ ipa_get_indirect_edge_target_1 (struct c >>>>>> tree otr_type; >>>>>> tree t; >>>>>> >>>>>> - if (param_index == -1) >>>>>> + if (param_index == -1 >>>>>> + || known_vals.length () <= (unsigned int) param_index) >>>>>> return NULL_TREE; >>>>>> >>>>>> if (!ie->indirect_info->polymorphic) >>>>>> @@ -1516,8 +1517,7 @@ ipa_get_indirect_edge_target_1 (struct c >>>>>> t = NULL; >>>>>> } >>>>>> else >>>>>> - t = (known_vals.length () > (unsigned int) param_index >>>>>> - ? known_vals[param_index] : NULL); >>>>>> + t = NULL; >>>>>> >>>>>> if (t && >>>>>> TREE_CODE (t) == ADDR_EXPR
