Prevent deopt when assigning double values to typed arrays

Prevent deopt when assigning double values to typed arrays

BUG=1515
TEST=

[danno, 2011-04-19 12:31:43]

Work in progress, but please take a look to make sure I'm not completely
off-base. Thanks -- Danno

[Kevin Millikin (Chromium), 2011-04-19 13:29:29]

+fschneider

[fschneider, 2011-05-10 10:44:52]

Here are some high-level comments. My worry is that it gets too complicated with
too many different representations.

http://codereview.chromium.org/6881003/diff/4001/src/hydrogen.cc
File src/hydrogen.cc (left):

http://codereview.chromium.org/6881003/diff/4001/src/hydrogen.cc#oldcode1837
src/hydrogen.cc:1837: CompareConversionUses(to_convert->at(index),
Thanks for cleaning this up!

http://codereview.chromium.org/6881003/diff/4001/src/hydrogen.cc
File src/hydrogen.cc (right):

http://codereview.chromium.org/6881003/diff/4001/src/hydrogen.cc#newcode1867
src/hydrogen.cc:1867: HPhase phase("Insert representation changes", this);
Not sure that the representation computation does the right thing. Why can you
just remove this part?

Maybe we can do something a little simpler without introducing completely
separate representations for ClampedUint8 and TruncatedInt32.

All values for ClampedUint8, int8, int16, int32, uint8 and uint16 can be
precisely represented as our current Representation::Integer32.

How about attaching the additional information to the range-information and let
the HStoreKeyedSpecialized decide whether to generate conversion code based on
the range and the representation of the input value.

This would avoid the complications with having phi-instructions and phi-uses of
even more different representations.

We still have a small problem with Uint32 values since they are outside the
int32 range - but probably we should deal with this one in a separate change
anyway.