Fix MathFloorOfDiv optimization on non-SDIV ARM

Fix MathFloorOfDiv optimization on non-SDIV ARM

R=verwaest@chromium.org

Committed: http://code.google.com/p/v8/source/detail?r=15174

[Jakob Kummerow, 2013-06-14 18:14:38]

PTAL.

[Toon Verwaest, 2013-06-14 18:15:32]

lgtm

[Jakob Kummerow, 2013-06-14 18:15:54]

Committed patchset #1 manually as r15174 (presubmit successful).

[Sven Panne, 2013-06-14 20:05:24]

https://codereview.chromium.org/17069006/diff/1/src/hydrogen-instructions.cc
File src/hydrogen-instructions.cc (right):

https://codereview.chromium.org/17069006/diff/1/src/hydrogen-instructions.cc#...
src/hydrogen-instructions.cc:1553: #ifdef V8_TARGET_ARCH_ARM
o_O I think we're doing something fundamentally wrong here if we use
target-specific #ifdefs. It's OK for a temporary quick-n-dirty (performance?
can't even tell...) fix to get things landed, but this should be removed as
quickly as possible at all cost (no idea how, let's talk next week).

We already have the ugly #ifdef wart because of the non-uniform handling of the
context register here, we should improve things and not make it worse.

[Jakob Kummerow, 2013-06-14 20:20:03]

Yes, of course that fix is ugly.

The real fix would be to beef up code generation on ARM for MathFloorOfDiv to
support arbitrary divisors even when no SDIV is available. That's certainly
possible using VFP instructions (just like we do for plain div-i), but it's not
something I'll implement on a Friday night at 8pm.

[Sven Panne, 2013-06-16 13:45:35]

On 2013/06/14 20:20:03, Jakob wrote:
> Yes, of course that fix is ugly.
> 
> The real fix would be to beef up code generation on ARM for MathFloorOfDiv to
> support arbitrary divisors even when no SDIV is available. That's certainly
> possible using VFP instructions (just like we do for plain div-i), but it's
not
> something I'll implement on a Friday night at 8pm.

Ah, OK, a tiny hint or a TODO in the code would have been nice, nevertheless.
The "why" is much more important than the "what"... (OK, I'm repeating myself.
:-)