Restore saved caller FP registers on stub failure

Restore saved caller FP registers on stub failure
and preserve FP registers on NotifyStubFailure.

In debug mode, clobber FP registers on each runtime call to increase
chances of catching such bugs.

R=danno@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=18000

[ulan, 2013-11-20 13:29:07]

PTAL

[danno, 2013-11-20 13:55:51]

Definitely heading in the right direction, I like this a lot.

https://codereview.chromium.org/78283002/diff/60001/src/arguments.cc
File src/arguments.cc (right):

https://codereview.chromium.org/78283002/diff/60001/src/arguments.cc#newcode120
src/arguments.cc:120: double ClobberDoubleRegisters(double x1, double x2, double
x3, double x4) {
This is pretty fragile. How about making this a platform-specific function that
is written in assembly? It also doesn't need to be #if DEBUG, either, the linker
will strip it if it's not used, and I prefer to compile all code if possible.

[ulan, 2013-11-20 16:20:05]

Ported to other architectures. Please take a look.

https://codereview.chromium.org/78283002/diff/60001/src/arguments.cc
File src/arguments.cc (right):

https://codereview.chromium.org/78283002/diff/60001/src/arguments.cc#newcode120
src/arguments.cc:120: double ClobberDoubleRegisters(double x1, double x2, double
x3, double x4) {
Removed DEBUG and add TODO for platform specific assembly implementation as
discussed.

On 2013/11/20 13:55:52, danno wrote:
> This is pretty fragile. How about making this a platform-specific function
that
> is written in assembly? It also doesn't need to be #if DEBUG, either, the
linker
> will strip it if it's not used, and I prefer to compile all code if possible.

[danno, 2013-11-21 20:24:29]

lgtm

[ulan, 2013-11-22 10:22:07]

Committed patchset #2 manually as r18000 (presubmit successful).