Use SSE2 instructions when available on ia32 platform.

Use SSE2 instructions when available on ia32 platform.

Committed: http://code.google.com/p/v8/source/detail?r=2868

[Erik Corry, 2009-09-09 12:11:43]

LGTM

We should ensure that SSE2 and CMOV are always enabled on MacOS, even if we are
doing a snapshot.

http://codereview.chromium.org/197057/diff/1/2
File src/ia32/codegen-ia32.cc (right):

http://codereview.chromium.org/197057/diff/1/2#newcode7428
Line 7428: // Jump to builtin for NaN.
Since you have the cc_ comparison you can actually generate the right answer for
NaN without going into the runtime system.  The ARM port does this.

http://codereview.chromium.org/197057/diff/1/2#newcode7445
Line 7445: // use edx, eax to convert unsigned to signed comparison
All comments should be sentences incl. punctuation.  Also, I can't see that we
are using edx here.

[William Hesse, 2009-09-09 14:43:41]

http://codereview.chromium.org/197057/diff/1/2
File src/ia32/codegen-ia32.cc (right):

http://codereview.chromium.org/197057/diff/1/2#newcode7428
Line 7428: // Jump to builtin for NaN.
On 2009/09/09 12:11:43, Erik Corry wrote:
> Since you have the cc_ comparison you can actually generate the right answer
for
> NaN without going into the runtime system.  The ARM port does this.

Done.

http://codereview.chromium.org/197057/diff/1/2#newcode7445
Line 7445: // use edx, eax to convert unsigned to signed comparison
On 2009/09/09 12:11:43, Erik Corry wrote:
> All comments should be sentences incl. punctuation.  Also, I can't see that we
> are using edx here.

Done.

[Mark Mentovai, 2009-09-09 15:09:55]

I'm familiar with x86 but much less so with V8 internals, so these questions may
be off base.

http://codereview.chromium.org/197057/diff/4001/4002
File src/ia32/codegen-ia32.cc (right):

http://codereview.chromium.org/197057/diff/4001/4002#newcode6745
Line 6745: FloatingPointHelper::CheckFloatOperands(masm, &call_runtime, ebx);
Shouldn't this be done for both SSE2 and x87, or should there be a
CheckSse2Operands?

http://codereview.chromium.org/197057/diff/4001/4002#newcode6746
Line 6746: // Allocate a heap number, if needed.
Is there a reason that this is done before the x87 operations in this branch,
but after the SSE2 operations in the branch above?

Is there a way to share this duplicated code between the two branches?

[William Hesse, 2009-09-10 07:27:36]

http://codereview.chromium.org/197057/diff/4001/4002
File src/ia32/codegen-ia32.cc (right):

http://codereview.chromium.org/197057/diff/4001/4002#newcode6745
Line 6745: FloatingPointHelper::CheckFloatOperands(masm, &call_runtime, ebx);
On 2009/09/09 15:09:55, Mark Mentovai wrote:
> Shouldn't this be done for both SSE2 and x87, or should there be a
> CheckSse2Operands?

The checking and the loading are combined into one function in LoadSse2Operands,
so that we can be more efficient.  The Smi tag is only checked once.  This is
documented in the declaration of LoadSse2Operands().

http://codereview.chromium.org/197057/diff/4001/4002#newcode6746
Line 6746: // Allocate a heap number, if needed.
On 2009/09/09 15:09:55, Mark Mentovai wrote:
> Is there a reason that this is done before the x87 operations in this branch,
> but after the SSE2 operations in the branch above?
> 
This code has to go after the check, but before the store of the result, so it
is enclosed by code that depends on whether we use SSE2.  It also needs to have
the values still in the eax and edx registers.  Most of the common code has been
factored out into AllocateHeapNumber() already.

So either we need to put it in a separate function, or we need to end the
use_sse2 scope, run the common code, and the n run a test and potentially enter
a new sse2 scope again.  This doesn't seem worth it.
> Is there a way to share this duplicated code between the two branches?