new liveness pass
Stephen Weeks
MLton@sourcelight.com
Mon, 4 Dec 2000 12:02:22 -0800 (PST)
> Also, here is the log for a self compile with the new pass. The self compile
> time is down to 765 seconds. This is the first time we're under 800! The funny
> thing is, the last self compile was at 948 seconds, so the difference can not be
> explained by this pass alone. The (unchanged) x86 codegen sped up by 59 seconds!
> My only conjecture is that the backend is leaving around some property lists on
> labels, which is slowing down the codegen. The new liveness pass doesn't
> leave quite as many around, so maybe that was helping. I'll look into making
> sure the backend cleans up after itself.
I figured out this discrepancy. I mistakenly ran the second self compile with a
larger heap, which cut down dramatically on the gc time in the codegen. I don't
think the backend was leaving around any property lists.
Anyways, the time breakdown for the latest self compile is
generate machine: 240
x86 code gen: 480
compile .s: 37
---
total 757