David Anderson [:dvander] - inactive, e-mail if emergency Reporter
	Description • 16 years ago

Brendan and I were debugging a patch with an inner-tree abort for this reason:

checking nested types 0x318a30: callee0=O/O this0=O/O vars0=I/I vars1=I/I vars2=I/I vars3=I/I vars4=D/I vars5=I/I vars6=O/O vars7=I/I vars8=I/I vars9=I/I vars10=I/I  (demotes 4)
adjusting will fail, vars7, slot 9
...etc

It turns out these failed slots are integers, for example jsval(1), and the tracker has a LIR_fcall(js_UnboxDouble).  It seems that unbox_jsval() doesn't specialize for integer GETELEM results like these.  The LIR filter changes to an UnboxInt32 when a DoubleToInt32 is present, but there is none here.

	Andreas Gal :gal
	Comment 1 • 16 years ago

Improvement. Should be done as part of the speculative demotion patch.

	Andreas Gal :gal
	Comment 2 • 15 years ago

David, this is the bug. The problem is really unbox_jsval, which is used by getprop/getelem and native calls. It should speculatively unpack as integer if the oracle doesn't say otherwise.

	David Anderson [:dvander] - inactive, e-mail if emergency Reporter
	Comment 3 • 15 years ago

This is a quick mockup to see perf gains, so I'm not sure if it's correct. Passes trace-tests at least.

3d-raytrace got 30ms slower, 3d-cube got 5ms slower. both build more traces and get more exits.

fannkuch got 15ms faster though.

	Andreas Gal :gal
	Comment 4 • 15 years ago

I think the key is better oracle feedback here. I will look at what extra side exits we get here.

	David Anderson [:dvander] - inactive, e-mail if emergency Reporter
	Comment 6 • 15 years ago

perma-paste: http://pastebin.mozilla.org/690129

	David Anderson [:dvander] - inactive, e-mail if emergency Reporter
	Comment 7 • 15 years ago

According to Julian's float-ops-that-are-useless profiler, this basically eliminated the millions of useless float ops on v8-richards. It's about a 70ms perf win on the SS version of the benchmark, and ~80 "perf points" or whatever the v8 harness measures.

So it seems we want to figure out a way to make this work on less friendlier cases.

	Ryan VanderMeulen [:RyanVM]
	Comment 8 • 13 years ago

Obsolete with the removal of tracejit.