(this report covers two months)
This is my monthly report on work done during November-December 2023
covered by my TPF perl core maintenance grant.
I mainly continued my work on making the perl stack reference counted.
As well as "unwrapping" a few more ops, I also took the opportunity
to introduce some basic optimisations to get the speed of a PERL_RC_STACK
perl interpreter build back closer to a vanilla build.
On my most recent branch (rc7, pushed today, 3rd Jan), the average of the
500 or so benchmarks in t/perf/benchmarks under PERL_RC_STACK has the
following instruction read, conditional branches, etc results as compared
to a vanilla perl build. 100% is unchanged, higher is better.
Ir 93.31
Dr 94.54
Dw 93.90
COND 92.50
IND 98.12
Note that the benchmark coverage is not yet comprehensive, and my
optimising efforts will likely have been biased to make these numbers look
better, rather than fixing the ops which aren't yet covered.
I've also made a start into looking at how XS code can be made (at least
sometimes) to work directly under a ref-counted stack, rather than each
XS call needing to be wrapped.
SUMMARY:
86:42 make stack reference counted
12:22 make stack reference counted - XS
9:33 process p5p mailbox
------
108:37 TOTAL (HH::MM)
--
Nothing ventured, nothing lost.
This is my monthly report on work done during November-December 2023
covered by my TPF perl core maintenance grant.
I mainly continued my work on making the perl stack reference counted.
As well as "unwrapping" a few more ops, I also took the opportunity
to introduce some basic optimisations to get the speed of a PERL_RC_STACK
perl interpreter build back closer to a vanilla build.
On my most recent branch (rc7, pushed today, 3rd Jan), the average of the
500 or so benchmarks in t/perf/benchmarks under PERL_RC_STACK has the
following instruction read, conditional branches, etc results as compared
to a vanilla perl build. 100% is unchanged, higher is better.
Ir 93.31
Dr 94.54
Dw 93.90
COND 92.50
IND 98.12
Note that the benchmark coverage is not yet comprehensive, and my
optimising efforts will likely have been biased to make these numbers look
better, rather than fixing the ops which aren't yet covered.
I've also made a start into looking at how XS code can be made (at least
sometimes) to work directly under a ref-counted stack, rather than each
XS call needing to be wrapped.
SUMMARY:
86:42 make stack reference counted
12:22 make stack reference counted - XS
9:33 process p5p mailbox
------
108:37 TOTAL (HH::MM)
--
Nothing ventured, nothing lost.