Some changes in preparation for the shrinker ir migration #3903
Conversation
There was a problem hiding this comment.
Very nice! I've just pushed a tiny increment more space-saving by keying off the integer hash of the sorted tuple-key. Unfortunatly WeakValueDictionary doesn't support dicts as values, or I'd have used that instead of the LRUReusedCache...
If you're happy with this, feel free to merge!
| # nans with a sign opposite of both bounds previously gave us trouble | ||
| # trying to use float clampers that didn't exist when drawing. |
There was a problem hiding this comment.
oooooohhhhh, that's what was causing the flakes. nicely cursed!
Zac-HD
force-pushed
the
ir-shrinker-preparation
branch
from
33c65cb
to
63be68e
Compare
|
|
Zac-HD
force-pushed
the
ir-shrinker-preparation
branch
from
63be68e
to
f89db6f
Compare
|
Heh, just noticed that and reverted 😅 |
… into ir-shrinker-preparation
|
d'oh, |
(mostly) split from #3899, as promised.
This is a bunch of small independent changes bundled together, plus a big one in 1e76ce2. I've also included two small bug fixes I ran into (2b05fb8, 251cf43). It may be easiest to review this commit-by-commit.
I'm not super happy with the api design of modifying/copying
IRTrees. I expect to iterate on this pretty quickly as I migrate future shrinking passes, so I haven't put too much effort into trying to get it perfect the first time. That said, if you have strong opinions on how this kind of thing should be designed, feel free to tear the current implementation apart!Conversations of note: