X64: Made bit-fiddling fallback for double-to-int32 conversion.

X64: Made bit-fiddling fallback for double-to-int32 conversion.

[Lasse Reichstein, 2010-05-11 09:10:12]

Short but convoluted bit fiddling.

[William Hesse, 2010-05-11 10:38:16]

LGTM if it improves performance.

http://codereview.chromium.org/2048007/diff/7001/8001
File src/x64/codegen-x64.cc (right):

http://codereview.chromium.org/2048007/diff/7001/8001#newcode8061
src/x64/codegen-x64.cc:8061: // Trashes rdi, rbx and rcx. Result cannot be rdi
or rbx.
Destroys the contents of

http://codereview.chromium.org/2048007/diff/7001/8001#newcode8065
src/x64/codegen-x64.cc:8065: // Result may be rcx. If result and source is same
register, source will
are the same register, source will be overwritten.

http://codereview.chromium.org/2048007/diff/7001/8001#newcode8102
src/x64/codegen-x64.cc:8102: __ xorl(result, result);  // Ensure that the CPU
knows that result holds zero.
What proof do you have that this does anything?  Explain.

http://codereview.chromium.org/2048007/diff/7001/8001#newcode8125
src/x64/codegen-x64.cc:8125: 
Does this actually have a performance impact?

[Lasse Reichstein, 2010-05-11 11:24:51]

http://codereview.chromium.org/2048007/diff/7001/8001
File src/x64/codegen-x64.cc (right):

http://codereview.chromium.org/2048007/diff/7001/8001#newcode8061
src/x64/codegen-x64.cc:8061: // Trashes rdi, rbx and rcx. Result cannot be rdi
or rbx.
Used "Overwrites" to be consistent with the next comment.
Can't I use "clobber"? I really like to say "clobber"!

http://codereview.chromium.org/2048007/diff/7001/8001#newcode8065
src/x64/codegen-x64.cc:8065: // Result may be rcx. If result and source is same
register, source will
Done.

http://codereview.chromium.org/2048007/diff/7001/8001#newcode8102
src/x64/codegen-x64.cc:8102: __ xorl(result, result);  // Ensure that the CPU
knows that result holds zero.
Hmm, none, really. I seem to remember that zeroing the register immediately
before the partial write is better, because the zero is in the register Reorder
Buffer (ROB), and doesn't have to be read from the permanent register file. But
I can't find anybody saying it.
 
I have a fix that doesn't use setcc, leaving no risk of partial register stalls.
Will change and reupload.

http://codereview.chromium.org/2048007/diff/7001/8001#newcode8125
src/x64/codegen-x64.cc:8125: 
Good question. I'll check.

[Lasse Reichstein, 2010-05-12 11:06:25]

http://codereview.chromium.org/2048007/diff/7001/8001
File src/x64/codegen-x64.cc (right):

http://codereview.chromium.org/2048007/diff/7001/8001#newcode8125
src/x64/codegen-x64.cc:8125: 
It does seem to have a (slight) positive performance impact, so I'll commit it.