1020494 - 10% slower AOT compilation speed on sqlite

Status: RESOLVED WONTFIX

(Core :: JavaScript Engine, defect)

Description

[Alon Zakai (:azakai)]

See

https://wiki.mozilla.org/Platform/Games/Testing/32v64bit

One result from there is that 64-bit windows builds compile sqlite (sqlite-cold-preparation) 10% or so slower than 32-bit builds on the same machine, which is a little surprising, so perhaps worth profiling.

Comment 1

[Alon Zakai (:azakai)]

sunfish had the very plausible suggestion that on 64-bit, we have many more registers, and on sqlite regalloc is a hard problem - so the two builds are solving different problems. E.g. perhaps 32-bit just keeps spilling all the time, while 64-bit can do better but needs to work more. Profiling that shows the time difference is all in regalloc would support that theory.

Comment 2

[Luke Wagner [:luke]]

I profiled this a bit:
 Parse time:                ~20ms *faster* on x64
 Validate and generate MIR: ~120ms slower on x64
 Optimize MIR:              ~100ms slower on x64
 Generate LIR and regalloc: ~100ms *faster* on x64
Overall compile time on x64 is ~800ms.

So my theory on the two cases where we're slower is the same as it was earlier: I think the slowdown is from touching 2x more data; parse, MIR and LIR nodes are all primarily pointers.  Validation and MIR-generation in particular touches both the AST and entire MIR graph at the same time.  I'm guessing regalloc is faster since more registers means less pressure so less work.

To test this I ran cachegrind on compilation and this shows:

               x64   x86
I refs        6.2B   6.7B
D refs        2.8B   3.8B
LL refs       136M    98M
LL misses      20M    11M

(LL = Last-level cache)  So this shows x64 executing less instructions and memory accesses, but falling back on the last-level cache a lot more and missing that a lot more.  This would seem to support my hypothesis, although it'd take more fine-grained testing to be sure.

Overall, I'm thinking of resolving wontfix, but I'd be happy to hear other opinions.