Am Samstag, den 01.02.2020, 09:55 -0500 schrieb Dan Eble:
I've noticed a subjective increase in profiling differences in regtests since
updating to the latest master this morning
(bc8a3fa7e4f12bf5ac1eb0293bfab658d52c4ae8), even from one run to another on
unchanged code.
There are also log differences that might be correlated, with some test cases
having
-
-Writing timing to ./lily-run-total.profile...
-Success: compilation successfully completed
and others having
+
+Writing timing to ./lily-run-total.profile...
+Success: compilation successfully completed
I've only run regtests a couple of times so far, so I'm not sure how repeatable
this is. Were there recent changes that might contribute to this? Nothing
stands out to me in the headlines.
I fear this might be the switch to Python 3 (yay, finally there!). Due
to hash() not being stable across invocations, it's now more likely
that snippets are processed in a different order. The solution is to
use a real hash function (like md5), see
https://sourceforge.net/p/testlilyissues/issues/5721/.
Unfortunately that change would not have worked with the previous
requirement of Python 2.4, so I could only post it today. I always got
some profiling differences since running 'make check' regularly, so I
didn't notice that it got significantly worse.
How distracting are the profiling differences for your work? I'm
obviously not a big fan of reverting bc8a3fa7e4, but I also don't want
to block anybody else from making progress.