__builtin_expect vs. PGO

__builtin_expect vs. PGO

[Scott Meyers]

What has more weight during codegen in a compilation with -fprofile-use 
when the code uses __builtin_expect to favor one branch while the PGO 
data says the other branch is more likely?

Thanks,

Scott

Re: __builtin_expect vs. PGO

[Ian Lance Taylor]

Scott Meyers <NeverRead@aristeia.com> writes:

> What has more weight during codegen in a compilation with
> -fprofile-use when the code uses __builtin_expect to favor one branch
> while the PGO data says the other branch is more likely?

I believe the probabilities will be combined in that case.  The
weighting of __builtin_expect is such that it will most likely be the
predicted case, unless the profiling data says that the unexpected path
is always taken.

Ian

Re: __builtin_expect vs. PGO

[Scott Meyers]

On 12/21/2011 8:49 AM, Ian Lance Taylor wrote:
> I believe the probabilities will be combined in that case.  The
> weighting of __builtin_expect is such that it will most likely be the
> predicted case, unless the profiling data says that the unexpected path
> is always taken.

Thanks for the comment.  The documentation for 
-fno-guess-branch-probability says:

   GCC will use heuristics to guess branch probabilities if they are
   not provided by profiling feedback (-fprofile-arcs). These
   heuristics are based on the control flow graph. If some branch
   probabilities are specified by `__builtin_expect', then the
   heuristics will be used to guess branch probabilities for the
   rest of the control flow graph, taking the `__builtin_expect'
   info into account. The interactions between the heuristics and
   `__builtin_expect' can be complex, and in some cases, it may be
   useful to disable the heuristics so that the effects of
   `__builtin_expect' are easier to understand.

Is this is any way relevant? On the one hand, it's talking about 
heuristics to be used in the absence of profiling feedback, and I'm 
assuming we have profiling information.  On the other hand, it's talking 
about the interaction of probabilities and __builtin_expect, and those 
probabilities presumably could come from PGO data.

Scott

Re: __builtin_expect vs. PGO

[Ian Lance Taylor]

Scott Meyers <NeverRead@aristeia.com> writes:

> On 12/21/2011 8:49 AM, Ian Lance Taylor wrote:
>> I believe the probabilities will be combined in that case.  The
>> weighting of __builtin_expect is such that it will most likely be the
>> predicted case, unless the profiling data says that the unexpected path
>> is always taken.
>
> Thanks for the comment.  The documentation for
> -fno-guess-branch-probability says:
>
>   GCC will use heuristics to guess branch probabilities if they are
>   not provided by profiling feedback (-fprofile-arcs). These
>   heuristics are based on the control flow graph. If some branch
>   probabilities are specified by `__builtin_expect', then the
>   heuristics will be used to guess branch probabilities for the
>   rest of the control flow graph, taking the `__builtin_expect'
>   info into account. The interactions between the heuristics and
>   `__builtin_expect' can be complex, and in some cases, it may be
>   useful to disable the heuristics so that the effects of
>   `__builtin_expect' are easier to understand.
>
> Is this is any way relevant? On the one hand, it's talking about
> heuristics to be used in the absence of profiling feedback, and I'm
> assuming we have profiling information.  On the other hand, it's
> talking about the interaction of probabilities and __builtin_expect,
> and those probabilities presumably could come from PGO data.

Yes, by the time it matters in the compiler, probabilities are
probabilities wherever they come from.  The above paragraph is basically
saying that in the absence of profiling information, the guessed
probabilities are combined with the __builtin_expect probabilities.  For
a simple two-way branch I think __builtin_expect will always beat a
guessed probabilities, just as it will generally beat a profiled
probability.  For something like
  switch (__builtin_expect (x, 1))
    {
      case 0: ...
      case 1: ...
      case 2: ...
    }
I'm not really sure what will happen.

Ian

Re: __builtin_expect vs. PGO

[Scott Meyers]

On 12/21/2011 5:02 PM, Ian Lance Taylor wrote:
> Yes, by the time it matters in the compiler, probabilities are
> probabilities wherever they come from.

Makes sense, thanks.

Scott