-
Notifications
You must be signed in to change notification settings - Fork 0
Note Rust performance fixes
This is a page of random performance ideas, mostly from pcwalton.
pcwalton: graydon: there are a bunch of unnecessary allocations in trans
pcwalton: I fixed a bunch of 'em but there are certainly more
pcwalton: one thing that would very much help is a SmallVector in libstd
pcwalton: fail() really needs to be a lang_item and it needs to be parameterized over ToStr
pcwalton: the idea being that if you're failing with a constant string
pcwalton: it should not call upcall_new_str_uniq() at the site of failure
pcwalton: rather it should pass a string slice (&static/str) to a function which itself copies the string
pcwalton: move the copy out of line
pcwalton: likewise fmt! generates a ton of code. most of the time the unique string it produces is not really being used, except to be written once and thrown away
graydon: mhm
pcwalton: (1) stop zeroing stuff out to revoke cleanups, use a flag instead to indicate whether the dtor needs to be run (this is potentially a quite large win)
pcwalton: (2) try to coalesce failure landing pads instead of duplicating the cleanups all the time
pcwalton: also ty::sty comparison in the ty hashtable is calling into the shape glue. I'm going to try to fix that this week
pcwalton: oh yeah, another thing I was thinking: stop making so many basic blocks
pcwalton: especially when optimization is off, LLVM goes basic-block-by-basic-block in codegen
pcwalton: oh yeah, we should also see if we can get away with not zeroing out memory when we allocate it
pcwalton: that's like 5% of our performance right there
pcwalton: in some profiles anyway
If you see large sequences of "mov" instructions, use call_memmove
or memzero
in trans, as appropriate. This is usually a symptom of code like Store(bcx, Load(bcx, foo, bar), baz);
, which is not an efficient way to move large structural types around.