This is the mail archive of the
mailing list for the GCC project.
Re: [google gcc-4_8] alternate hirate for builtin_expert
- From: Xinliang David Li <davidxl at google dot com>
- To: Jan Hubicka <hubicka at ucw dot cz>
- Cc: Rong Xu <xur at google dot com>, GCC Patches <gcc-patches at gcc dot gnu dot org>
- Date: Thu, 26 Sep 2013 15:52:57 -0700
- Subject: Re: [google gcc-4_8] alternate hirate for builtin_expert
- Authentication-results: sourceware.org; auth=none
- References: <CAF1bQ=RnCQGRrdUxV0Mpj9wAiB=fUy+fzouGOf0Rj_YWt2TMoA at mail dot gmail dot com> <20130926222744 dot GC21484 at kam dot mff dot cuni dot cz>
it might worth extend __builtin_expect to take more parameters:
1) argument to specify actual probability: __builtin_expect (x, 10, 0.6)
2) a more general way of doing this is to allow specifying multiple
values, and the probability is determined by # of occurances:
__builtin_expect (x, 10, 10, 20) --> tells compiler x is expected to
be 10 66% of the time, and 33% of time with value twenty.
3) a special value can be reserved to indicate if the branch is
predictable or not.
On Thu, Sep 26, 2013 at 3:27 PM, Jan Hubicka <email@example.com> wrote:
>> Current default probably for builtin_expect is 0.9996.
>> This makes the freq of unlikely bb very low (4), which
>> suppresses the inlining of any calls within those bb.
>> We used FDO data to measure the branch probably for
>> the branch annotated with builtin_expert.
>> For google
>> internal benchmarks, the weight average
>> (the profile count value as the weight) is 0.9081.
>> Linux kernel is another program that is heavily annotated
>> with builtin-expert. We measured its weight average as 0.8717,
>> using google search as
>> the workload.
> This is interesting. I was always unsure if programmers use builtin_expect
> more often to mark an impossible paths (as those leading to crash) or just
> unlikely paths. Your data seems to suggest the second.
> We probably also ought to get analyze_brprob working again. It was always
> useful to get such a data.
>> This patch sets the alternate hirate probability for
>> to 90%. With the alternate hirate, we measured performance
>> improvement for google
>> benchmarks and Linux kernel.
>> 2013-09-26 Rong Xu <firstname.lastname@example.org>
>> * params.def (DEFPARAM): New.
>> * params.def: New.
>> * predict.c (tree_predict_by_opcode): Alternate
>> probablity hirate for builtin_expect.
> This also seems resonable for mainline. Please add a comment
> to the parameter explaining how the value was determined.
> Please also add invoke.texi documentation.
> For patches that seems resonable for mainline in FDO/IPA area,
> i would apprechiate if you just added me to CC, so I do not lose
> track of them.