Author: Armin Rigo <[email protected]> Branch: extradoc Changeset: r5636:d80eeb4f9ad1 Date: 2016-04-12 16:05 +0200 http://bitbucket.org/pypy/extradoc/changeset/d80eeb4f9ad1/
Log: more diff --git a/planning/misc.txt b/planning/misc.txt --- a/planning/misc.txt +++ b/planning/misc.txt @@ -3,6 +3,8 @@ minor/major collections: order the visit to the objects by... address? +using two lists of pointers, popping from one and appending to the +other, and when the first one is empty, sort the other and swap them make the resizing of dict/lists more GC-aware @@ -56,3 +58,12 @@ avoid MOVSD/MOVSS between registers; do a full copy with MOVAPD or MOVDQA + +on CPUs with 'emsr' in /proc/cpuinfo, a memcpy is documented as best +implemented as simply REP MOVSB. But measures on my CPU show it's not +the case. E.g. sizes <= 128 are three times faster if done by a call +to memcpy than by REP MOVSB, for any size. + +still, look at replacing CALLs to memcpy by, say, up to 3 pairs of +MOVs of 1, 2, 4, 8 or 16 bytes each. it is a win also because it +doesn't force specific registers _______________________________________________ pypy-commit mailing list [email protected] https://mail.python.org/mailman/listinfo/pypy-commit
