freenode/#clasp - IRC Chatlog
Search
2:16:16
Bike
Could we reasonably special case the situation of a creation or initialization form being only a function call?
2:17:22
drmeister
(with-open-file (*debug-io* "/tmp/output.log" :direction :output) (compile-file "sys:kernel;cleavir;inline.lisp"))
2:20:27
drmeister
make-load-form - you pass it an object as the first argument and it gets substituted in the create and initialization forms that are returned - correct?
2:40:38
drmeister
Yeah - that's still not enough - you need to compile the make-load-form create and initialize forms so that they compile other make-load-form forms.
2:42:02
drmeister
A simple thing I could do is suppress the spilling of register arguments - that's not needed at all.
8:01:04
drmeister
I got inline.ll down to 1.8M lines of llvm-IR by removing the register spilling code.
11:23:17
stassats
just reducing compilation size is just plugging holes, i'd rather make it involve zero compilation for constants
13:51:34
drmeister
I know - just sayin. Well, IIRC it was down to about 28min before Bike switched to SSA for temporary variables.
13:51:54
stassats
i would've tried less compily constants on my own, but 40 minute build times and c++ discourage me
13:52:53
drmeister
Bike: Not you - the SSA for temporary variables might have shifted the burden elsewhere within llvm.
13:54:00
drmeister
Could you confirm my suspicions tomorrow? Pick the commits before the temporaries and after the temporaries and then build those.
13:55:27
drmeister
I'd like to measure the build time, the code size and the number of landingpads due to the changes in how temporaries are handled and stackmaps.
13:57:47
drmeister
I can't be sure if the build time is slower because I could have done something that slowed it down. I've had build times jump around by factors of 2x to 8x due to power management and build configurations.
13:59:27
drmeister
We also need to measure how fast the cleavir compiler is within cclasp. bclasp+cleavir building cclasp is one thing, but cclasp compiling say a quicklisp system is something else and improving that performance is more important than build time.
14:00:18
drmeister
I've just noticed over the last couple of days that the build time was down to about 28min then it jumped up to 1hr and now it's down to 49min.
14:12:32
drmeister
I just can't be sure about my build time observations because they stretch over weeks and my machine isn't carefully set into a calm state before building.