Truncate rather than round to nearest when performing float-to-integer...

Truncate rather than round to nearest when performing float-to-integer
conversions for external array types, which implement the Typed Array
spec. (Revision of http://codereview.chromium.org/6315004 .)

Prefer SSE2 code path on x86 processors. Non-SSE2 processors now make
a slow runtime call for float-to-int conversions. Use SSE3 for 32-bit
signed and unsigned int array types where possible.

The movement of code from ic-arm.cc to stub-cache-arm.cc caused the
VFP3 code path to be tested for the first time. Fixed bugs in the
register usage and in the constant value stored into integer arrays
for NaN and +/-Infinity.

Added new truncation test to test-api.cc. Storage of NaN and +/-Inf
was already covered. Ran unit tests on x86, x64 and ARM simulator.
Tested ia32 and x64 code in Chromium on Mac and Linux respectively
with Typed Array unit tests and WebGL content.

BUG=http://code.google.com/p/chromium/issues/detail?id=50972
TEST=test-api/ExternalArrays


Committed: http://code.google.com/p/v8/source/detail?r=6431

[Erik Corry, 2011-01-20 14:41:23]

This needs tests in our own v8/test directory that test rounding, infinities,
NaNs, large numbers.

Is it really right that writing 1e100 to a signed byte array will write -128
while Infinity will write 0?  Any clarification from the standards committee on
this?

Also it really needs to work on non-SSE2 machines.  It's OK to just call the
runtime in this case, it just has to work.  x64 always has SSE2.

http://codereview.chromium.org/6303012/diff/1/src/arm/stub-cache-arm.cc
File src/arm/stub-cache-arm.cc (right):

http://codereview.chromium.org/6303012/diff/1/src/arm/stub-cache-arm.cc#newco...
src/arm/stub-cache-arm.cc:3866: // If exponent is too big than result is minimal
value.
than -> then

[Ken Russell (switch to Gerrit), 2011-01-20 15:51:10]

On 2011/01/20 14:41:23, Erik Corry wrote:
> This needs tests in our own v8/test directory that test rounding, infinities,
> NaNs, large numbers.

I will be happy to add more tests for out of range values in a follow-on CL, but
I am looking for an LGTM on this CL as is. The previous version which fixed the
x86 and x64 code paths was already LGTM'd.

> Is it really right that writing 1e100 to a signed byte array will write -128
> while Infinity will write 0?  Any clarification from the standards committee
on
> this?

I am pretty sure this is the behavior defined by the combination of the Web IDL
and ECMA-262 specifications.

> Also it really needs to work on non-SSE2 machines.  It's OK to just call the
> runtime in this case, it just has to work.  x64 always has SSE2.

I've filed a follow-on bug already about fixing this case.
http://code.google.com/p/v8/issues/detail?id=1056

> http://codereview.chromium.org/6303012/diff/1/src/arm/stub-cache-arm.cc
> File src/arm/stub-cache-arm.cc (right):
> 
>
http://codereview.chromium.org/6303012/diff/1/src/arm/stub-cache-arm.cc#newco...
> src/arm/stub-cache-arm.cc:3866: // If exponent is too big than result is
minimal
> value.
> than -> then

I'll fix this preexisting typo upon landing.

[Ken Russell (switch to Gerrit), 2011-01-21 07:45:37]

Added truncation test cases to test-api.cc. NaN, Infinity were already tested.

Rewrote ia32 stub. Now calls slow runtime stub on non-SSE2 processors for
float-to-int conversions for correctness. SSE2 is now the preferred code path.
SSE3 is used if available to handle int and unsigned int arrays.

Ran all external array tests on x86, x64, ARM simulator.

[Erik Corry, 2011-01-21 08:20:29]

LGTM

http://codereview.chromium.org/6303012/diff/42002/src/ia32/stub-cache-ia32.cc
File src/ia32/stub-cache-ia32.cc (right):

http://codereview.chromium.org/6303012/diff/42002/src/ia32/stub-cache-ia32.cc...
src/ia32/stub-cache-ia32.cc:3590: __ jmp(&slow);
This jmp is not needed.

[Ken Russell (switch to Gerrit), 2011-01-22 00:00:35]

On 2011/01/21 08:20:29, Erik Corry wrote:
>
http://codereview.chromium.org/6303012/diff/42002/src/ia32/stub-cache-ia32.cc...
> src/ia32/stub-cache-ia32.cc:3590: __ jmp(&slow);
> This jmp is not needed.

Removed in committed version.