Intrinsify Uint8ClampedArray's store and load indexed, inline load indexed.

Intrinsify Uint8ClampedArray's store and load indexed, inline load indexed.
Committed: https://code.google.com/p/dart/source/detail?r=16666

[regis, 2013-01-05 00:53:36]

LGTM

[sra1, 2013-01-05 01:34:49]

https://codereview.chromium.org/11783006/diff/8001/runtime/vm/intrinsifier_ia...
File runtime/vm/intrinsifier_ia32.cc (right):

https://codereview.chromium.org/11783006/diff/8001/runtime/vm/intrinsifier_ia...
runtime/vm/intrinsifier_ia32.cc:614: // Free EBX for the value since we want a
byte register.
Could you free EAX and EBX here by generating the address with lea(EDX, ...)?

https://codereview.chromium.org/11783006/diff/8001/runtime/vm/intrinsifier_ia...
runtime/vm/intrinsifier_ia32.cc:623: __ j(GREATER, &load_0xff, 
Assembler::kNearJump);
Instead of jumping on carry with data dependent branch mispredicts, we could
compute the value based on the carry flag:

__ sbbl(EBX, EBX);
__ notl(EBX);
// store

[srdjan, 2013-01-08 00:34:06]

Thanks for comments, will do it in a next CL.

https://codereview.chromium.org/11783006/diff/8001/runtime/vm/intrinsifier_ia...
File runtime/vm/intrinsifier_ia32.cc (right):

https://codereview.chromium.org/11783006/diff/8001/runtime/vm/intrinsifier_ia...
runtime/vm/intrinsifier_ia32.cc:614: // Free EBX for the value since we want a
byte register.
On 2013/01/05 01:34:50, sra1 wrote:
> Could you free EAX and EBX here by generating the address with lea(EDX, ...)?

I cannot use EDX (must be preserved, see comment on the top), using EAX instead.

https://codereview.chromium.org/11783006/diff/8001/runtime/vm/intrinsifier_ia...
runtime/vm/intrinsifier_ia32.cc:623: __ j(GREATER, &load_0xff, 
Assembler::kNearJump);
On 2013/01/05 01:34:50, sra1 wrote:
> Instead of jumping on carry with data dependent branch mispredicts, we could
> compute the value based on the carry flag:
> 
> __ sbbl(EBX, EBX);
> __ notl(EBX);
> // store
> 
> 

I do not think that can work. If compare sets CF we will always jump to
'store_value' label [BELOW_EQUAL = (CF OR ZF) = 1]/

[sra1, 2013-01-08 02:31:14]

These ideas might be more important if you are open-coding the operation.

https://codereview.chromium.org/11783006/diff/8001/runtime/vm/intrinsifier_ia...
File runtime/vm/intrinsifier_ia32.cc (right):

https://codereview.chromium.org/11783006/diff/8001/runtime/vm/intrinsifier_ia...
runtime/vm/intrinsifier_ia32.cc:614: // Free EBX for the value since we want a
byte register.
On 2013/01/08 00:34:06, srdjan wrote:
> On 2013/01/05 01:34:50, sra1 wrote:
> > Could you free EAX and EBX here by generating the address with lea(EDX,
...)?
> 
> I cannot use EDX (must be preserved, see comment on the top), using EAX
instead.

Sorry, I mis-wrote - EDI.  EDI is used below.

https://codereview.chromium.org/11783006/diff/8001/runtime/vm/intrinsifier_ia...
runtime/vm/intrinsifier_ia32.cc:623: __ j(GREATER, &load_0xff, 
Assembler::kNearJump);
On 2013/01/08 00:34:06, srdjan wrote:
> On 2013/01/05 01:34:50, sra1 wrote:
> > Instead of jumping on carry with data dependent branch mispredicts, we could
> > compute the value based on the carry flag:
> > 
> > __ sbbl(EBX, EBX);
> > __ notl(EBX);
> > // store
> > 
> > 
> 
> I do not think that can work. If compare sets CF we will always jump to
> 'store_value' label [BELOW_EQUAL = (CF OR ZF) = 1]/

You are right!  I'm getting my signed and unsigned comparisons mixed up.
I think at this point the value is actually a function of sign bit:

__ sarl(EBX, 31);
__ notl(EBX);
// store.

Some other thoughts:

If it is worth doing this, it is probably worth making the in-range case be
fall-through.

I wonder of going completely branch free would work?  I guess it would depend on
having data that needs clamping.

lea edi,[eax+ebx*1+k]   // free EAX,EBX
mov ebx, [esp+4]
test ebx,1
jnz fall_through

SmiUntag(ebx) 
mov eax,ebx
sar eax,31
not eax
cmp ebx,0xff
cmova ebx,eax
mov [edi],bl
ret