drive-by comments.

http://codereview.chromium.org/3169050/diff/15001/16001
File src/ia32/codegen-ia32.cc (right):

http://codereview.chromium.org/3169050/diff/15001/16001#newcode5334
src/ia32/codegen-ia32.cc:5334: __ xor_(Operand(esp, 0),
Immediate(jit_cookie_));
Interleaving stack oprations (e.g. push/pop) with esp relative
addressing causes extra latency between the operations (on at least the
intel core architecture).
If there is any way to do the computation in a register and then push it
afterwards, there is a chance it would still be faster, and parallelize
better if there are more unsafe smi pushes in a row. I needs a scratch
register, though.

http://codereview.chromium.org/3169050/diff/15001/16001#newcode5351
src/ia32/codegen-ia32.cc:5351: __ lea(target, Operand(target,
jit_cookie_));
I don't think we should start depending on this macro not changing
flags. It might change later.
And it doesn't even hold now, since Set might use xor if the value is
zero.

http://codereview.chromium.org/3169050/show

--
v8-dev mailing list
[email protected]
http://groups.google.com/group/v8-dev

Reply via email to