Front page | perl.perl5.porters |
Postings from October 2017
On Mon, Oct 23, 2017 at 11:34:16PM +0200, Lukas Mai wrote:
> I think it would be a good idea to use compiler intrinsics for overflow
> checks where available.
>
> I've pushed a branch that implements this at mauke/overflow:
> https://perl5.git.perl.org/perl.git/shortlog/refs/heads/mauke/overflow
>
> Dave: I've CC'd you directly because you last worked on this code (commit
> 230ee21f3e366901ce5769d324124c522df7ce8a, "faster add, subtract, multiply").
>
> My changes affect pp_add, pp_subtract, and pp_multiply. I think the new code
> is nicer because it's easier to understand than all the low-level bit
> fiddling, and it passes all tests on my machine. However, I haven't done any
> benchmarks to see how it affects performance (if at all).
>
> Things I need help with:
>
> - code review
Note that Reini did something similar in cperl, although he disabled
my short-cut code in the presence of __builtin_mul_overflow etc, which was
a mistake (and is why the nbody benchmark runs about 30% faster on perl
compared with cperl - at least last time I looked).
At a cursory inspection the code looks good (although I haven't looked
closely at the main body (non-shortcut) part of the code.
Note that your '#ifdefs' probably need indenting with '# ifdef'
in some places since there's already ifdefs surrounding the code.
> - benchmarks (compared to a17768d7c7b82c136fbeacd85db3451973a8007a)
Are you familiar with Porting/bench.pl and t/perf/benchmarks?
I tried running it as follows:
I had 3 executables:
/tmp/perl-a1776 - just before your 3 commits
/tmp/perl-of - the tip of your branch
/tmp/perl-no-of - ditto with this diff applied:
$ diff -u config.h- config.h
+#if 0
#define HAS_BUILTIN_ADD_OVERFLOW /**/
#define HAS_BUILTIN_SUB_OVERFLOW /**/
#define HAS_BUILTIN_MUL_OVERFLOW /**/
+#endif
run the arith subset of the benchmarks against the 3 perls and write the
results to a file: use 8 CPUs in parallel:
$ perl Porting/bench.pl -w /tmp/bm_num -j 8 --tests=/expr::arith::/ -v \
/tmp/perl-a1776 /tmp/perl-no-of /tmp/perl-of
read the results back and display them sorting by number of conditional
branches in the right-most column:
$ perl Porting/bench.pl -r /tmp/bm_num --sort=COND:-1 > /tmp/bm_num.out
You can similarly do --sort=Ir:-1 etc
On my hardware, this shows that the no-builtins build shows no slowdown
(good!) and the builtins build shows a modest improvement in the number
instruction reads and/or conditional branches (again, good).
Here are a couple the best results:
expr::arith::add_lex_ii
add two integers and assign to a lexical var
/tmp/perl-a1776 /tmp/perl-no-of /tmp/perl-of
--------------- --------------- ------------
Ir 100.00 100.00 106.19
Dr 100.00 100.00 100.00
Dw 100.00 100.00 100.00
COND 100.00 100.00 100.00
IND 100.00 100.00 100.00
expr::arith::add_lex_ss
add two short strings and assign to a lexical var
/tmp/perl-a1776 /tmp/perl-no-of /tmp/perl-of
--------------- --------------- ------------
Ir 100.00 100.00 101.96
Dr 100.00 100.00 100.00
Dw 100.00 100.00 100.00
COND 100.00 100.00 104.30
IND 100.00 100.00 100.00
And here's the average. It includes many non add/sub/mult benchmarks, which
dilutes the numbers.
AVERAGE
/tmp/perl-a1776 /tmp/perl-no-of /tmp/perl-of
--------------- --------------- ------------
Ir 100.00 100.00 101.01
Dr 100.00 100.00 100.00
Dw 100.00 100.00 100.00
COND 100.00 100.00 100.22
IND 100.00 100.00 100.00
--
I before E. Except when it isn't.
Thread Previous
|
Thread Next