Implement Math.floor stub on ARM. Uses VFP when available....

Implement Math.floor stub on ARM. Uses VFP when available.

TEST=none
BUG=none

[Mads Ager (chromium), 2010-11-09 15:22:32]

This looks like something for Erik to review. :)

[Erik Corry, 2010-11-10 13:57:19]

Almost looks good to me.  A few comments below.

http://codereview.chromium.org/4735003/diff/1/src/arm/stub-cache-arm.cc
File src/arm/stub-cache-arm.cc (right):

http://codereview.chromium.org/4735003/diff/1/src/arm/stub-cache-arm.cc#newco...
src/arm/stub-cache-arm.cc:1734: __ Ldrd(r4, r5, MemOperand(r0,
HeapNumber::kValueOffset - kHeapObjectTag));
Use FieldMemOperand for this.

http://codereview.chromium.org/4735003/diff/1/src/arm/stub-cache-arm.cc#newco...
src/arm/stub-cache-arm.cc:1744: __ bic(r9, r3, Operand((1 << 22) | 0xf | (1 <<
24)));
This should be with named constants.  For example
bic(r9, r3, Operand(kVFPExceptionMask | kVFPRoundingModeMask |
kWhateverBit24Means)
orr(r9, r9, Operand(kVFPRoundToZeroBits))

http://codereview.chromium.org/4735003/diff/1/src/arm/stub-cache-arm.cc#newco...
src/arm/stub-cache-arm.cc:1758: __ tst(r9, Operand(0xf));
Named constant here too.

http://codereview.chromium.org/4735003/diff/1/src/arm/stub-cache-arm.cc#newco...
src/arm/stub-cache-arm.cc:1805: 
This non-VFP code makes testing difficult and it only catches trivial inputs.  I
think it is actually preferable to just return undefined if there is no VFP. 
This has the advantage of being 'obviously correct'.

[m.m.capewell, 2010-11-12 11:59:44]

I've addressed your comments, but we're currently investigating a performance
regression caused by this patch. Will reupload when we have resolved this.

[m.m.capewell, 2010-11-16 15:05:16]

http://codereview.chromium.org/4735003/diff/1/src/arm/stub-cache-arm.cc
File src/arm/stub-cache-arm.cc (right):

http://codereview.chromium.org/4735003/diff/1/src/arm/stub-cache-arm.cc#newco...
src/arm/stub-cache-arm.cc:1700: __ BranchOnNotSmi(r1, &miss);
Fixed this - should be BranchOnSmi().

http://codereview.chromium.org/4735003/diff/1/src/arm/stub-cache-arm.cc#newco...
src/arm/stub-cache-arm.cc:1734: __ Ldrd(r4, r5, MemOperand(r0,
HeapNumber::kValueOffset - kHeapObjectTag));
On 2010/11/10 13:57:19, Erik Corry wrote:
> Use FieldMemOperand for this.

Done.

http://codereview.chromium.org/4735003/diff/1/src/arm/stub-cache-arm.cc#newco...
src/arm/stub-cache-arm.cc:1744: __ bic(r9, r3, Operand((1 << 22) | 0xf | (1 <<
24)));
On 2010/11/10 13:57:19, Erik Corry wrote:
> This should be with named constants.  For example
> bic(r9, r3, Operand(kVFPExceptionMask | kVFPRoundingModeMask |
> kWhateverBit24Means)
> orr(r9, r9, Operand(kVFPRoundToZeroBits))

Done.

http://codereview.chromium.org/4735003/diff/1/src/arm/stub-cache-arm.cc#newco...
src/arm/stub-cache-arm.cc:1758: __ tst(r9, Operand(0xf));
On 2010/11/10 13:57:19, Erik Corry wrote:
> Named constant here too.

Done.

http://codereview.chromium.org/4735003/diff/1/src/arm/stub-cache-arm.cc#newco...
src/arm/stub-cache-arm.cc:1805: 
On 2010/11/10 13:57:19, Erik Corry wrote:
> This non-VFP code makes testing difficult and it only catches trivial inputs. 
I
> think it is actually preferable to just return undefined if there is no VFP. 
> This has the advantage of being 'obviously correct'.

Done.

[m.m.capewell, 2010-11-16 15:07:48]

Fixed the performance regression - BranchOnNotSmi vs BranchOnSmi.

However, uploading the changed file has created a new issue:
http://codereview.chromium.org/5075002/ . Sorry.