[tex-live] strange discrepancy in running time of etex between TL2015 and TL2017

jfbu jfbu at free.fr
Thu Jul 27 14:44:35 CEST 2017


Thanks to Nelson, Reinhard, Zdenek, 

yesterday I was in the midst of finalizing my maintenance
release of xint. I had not been doing speed benchmarks 
as I was not expecting significant changes;

(I was rather focused on non-regression testing)

so when I observed at
last minute a 40% speed penalty I was surprised, but then
discovered that this was a TL2017 issue (and, I assume
a darwinlegacy issue). So I started this thread.

Later I understood it was actually presumably the same
problem I had already encountered with luatex

Here is a no-package version:

----
\def\A{\B\B\B\B\B\B\B\B\B\B}% 10
\def\B{\C\C\C\C\C\C\C\C\C\C}% 100
\def\C{\noexpand\A\noexpand\A\noexpand\A\noexpand\A
       \noexpand\A\noexpand\A\noexpand\A\noexpand\A
       \noexpand\A\noexpand\A}% 1000
\edef\D{\A}% 1000
\edef\E{\D}% 1000000

\def\A{\edef\z{666}}

\E\message{.}
\E\message{.}
\E\message{.}
\E\message{.}
\E\message{.}
\E\message{.}
\E\message{.}
\E\message{.}
\E\message{.}
\E\message{.}
\bye
----

Hence now no \numexpr, only Knuth tex

When I do 

time tex testspeed.tex

I consistently get on my iMac 10.9.5:

- with TL2015: about 1.62s (last digit non-significant)

- with TL2017: about 2.18s (idem)

Thus, it appears, but somehow I had overlooked it so far,
that darwinlegacy tex (et al.) is at least 33% slower on TL2017
than formerly

I was aware of that with luatex but had not realized it applied to all tex's

If something like this is reproduced on other systems than Mac OS 10.9.5
or more generally the "darwinlegacy", I would be interested to know

Best

(sorry for the xint distraction, but that was my initial focus)

Jean-Francois


More information about the tex-live mailing list