[Chicken-hackers] Add branch prediction for C_demand checks [was: Re: [PATCH] Statically determine if argvector can be reused]

[Peter Bex]

On Thu, Dec 15, 2016 at 11:55:01PM +0100, Peter Bex wrote:
> Hi all,
> 
> I've been playing around a little bit with "perf" and Valgrind's
> cachegrind tool, and I noticed that the number of branch prediction
> misses can be reduced if the argvector reusing check can be hardcoded
> for cases where we know the size of the calling function's argvector.
> 
> Here's a simple patch to make this happen (works on both the master
> and chicken-5 branches).

And here's another one, that adds C_likely() and C_unlikely() macros,
a la the Linux kernel's likely() and unlikely().  These are simple
wrappers for __builtin_expect() which tell the compiler which branches
in a conditional expression are likely to be taken.

These are now used in generated code, when we do a C_demand() check to
allocate memory.  This is noticably faster in some benchmarks, and
can bring down compilation time a bit as well.  Attached also are
benchmark results for unpatched CHICKEN 5, the previous patch which
adds static argvector checks and results for both this patch and the
static argvector.

Question: Should this be:

# define C_unlikely(x)             __builtin_expect((x), 0)

or, like Linux:

# define C_unlikely(x)             __builtin_expect(!!(x), 0)

The latter seems to me like it would add more instructions due to
the double negation, and the exact value of the expression in an if()
is never important anyway, as long as it's zero or nonzero.  Right?

Cheers,
Peter

0001-Add-__builtin_expect-to-help-branch-prediction.chicken-5.patch
Description: Text Data

0001-Add-__builtin_expect-to-help-branch-prediction.master.patch
Description: Text Data

benchmarks.comparison
Description: Text document

signature.asc
Description: Digital signature