Improve performance of esnstrument

[msridhar]

Right now, instrumenting code is a bit slow. I've put the v8 profiling output for instrumenting pdf.js here. One concerning thing is that ~20% of time is spent in GC; would have to dig in further to see why.

Otherwise, I don't see any quick fixes. The major change I can think of would be to not transform the AST and use escodegen, but instead generate the instrumented code string directly during the AST pass. Not sure how nightmare-ish this would be, though.

Other thoughts welcome :-)

[msridhar]

Another thought was to generate code strings where convenient, cache them in the AST, and then modify escodegen to use cached strings when available. But, this would involve using a hacked version of escodegen.

[msridhar]

One other tidbit: running with --trace_gc shows that the escodegen phase allocates a massive amount of memory: heap goes from ~200MB to between 500MB and 800MB for pdfjs. So somewhere it's allocating a ton of temporary objects.

[ksen007]

any idea if my traversals are also using massive amount of memory. 

Sent from Samsung Mobile

-------- Original message --------
From: Manu Sridharan [email protected]
Date: 23/11/2013 11:39 AM (GMT-08:00)
To: SRA-SiliconValley/jalangi [email protected]
Subject: Re: [jalangi] Improve performance of esnstrument (#8)

One other tidbit: running with --trace_gc shows that the escodegen phase allocates a massive amount of memory: heap goes from ~200MB to between 500MB and 800MB for pdfjs. So somewhere it's allocating a ton of temporary objects.

—
Reply to this email directly or view it on GitHub.

[msridhar]

Doesn't look too bad from a GC trace. Will have to profile in the browser
and look at some heap snapshots to know better. But the mem usage from
escodegen is ridiculous.
On Nov 23, 2013 12:14 PM, "ksen007" [email protected] wrote:

any idea if my traversals are also using massive amount of memory.

Sent from Samsung Mobile

-------- Original message --------
From: Manu Sridharan [email protected]
Date: 23/11/2013 11:39 AM (GMT-08:00)
To: SRA-SiliconValley/jalangi [email protected]
Subject: Re: [jalangi] Improve performance of esnstrument (#8)

One other tidbit: running with --trace_gc shows that the escodegen phase
allocates a massive amount of memory: heap goes from ~200MB to between
500MB and 800MB for pdfjs. So somewhere it's allocating a ton of temporary
objects.

—
Reply to this email directly or view it on GitHub.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/8#issuecomment-29140615
.

[msridhar]

We got significant improvements by switching to acorn for parsing and patching a performance bug in escodegen. Will leave this open, as instrumentation is still a bit slow.