Lowered kMaxVirtualRegisters (fixes v8 2139 and chrome 123822 and 128252).

Lowered kMaxVirtualRegisters (fixes v8 2139 and chrome 123822 and 128252).

BUG=128252
Committed: https://code.google.com/p/v8/source/detail?r=12613

[Vyacheslav Egorov (Google), 2012-09-21 11:42:08]

dbc

[I would say that big functions should be sent to background thread for
compilation or split into chunks for chunked compilation; but this is too big a
change]

https://chromiumcodereview.appspot.com/10966031/diff/1/src/lithium.h
File src/lithium.h (right):

https://chromiumcodereview.appspot.com/10966031/diff/1/src/lithium.h#newcode159
src/lithium.h:159: static const int kMaxVirtualRegisters = 1 << 13;
can you give some of this bits to other fields?

if you increase kMax/MinFixedIndex it will help Issue 2223.

[Massi, 2012-09-21 13:53:21]

I totally agree on the separate thread approach, apart from the fact that we
also had OOM issues with too many vregs so the limit had to be lowered anyway
(reaching the OOM condition is painfully slow on most machines).

This last patch settles on 15 bits after testing on many benchmarks trying to
find a reasonable compromise, and it gives all remaining bits to
kFixedIndexWidth.

[Jakob Kummerow, 2012-09-21 13:58:14]

lgtm

[Massi, 2012-09-21 14:01:23]

To clarify, with "I tried multiple benchmarks" I meant these golem4 runs:

http://70.32.156.88:8080/Comparison#targetA%3Dv8%3BmachineTypeA%3Dlinux-ia32%...

(the 4 "Fix-128252..." patches are with 13, 14, 15, and 16 bits to the vregs
limit)

[Sven Panne, 2012-10-01 09:40:41]

Just another DBC (late, but it's the 1st day after my vacation :-) :

When I looked into Alon Zakai's Emscripten version of Box2D (taken from his
blog), I found out that we actually need to *increase* these limits a lot,
otherwise we suck heavily because we don't crankshaft the hot functions. So this
CL is just hiding a symptom and makes things worse for other benchmarks.

I am not sure if we really need the background compilation approach for this
(although we should really try it), just deopting less  and speeding up
Crankshaft itself might be a viable direction, too...