Implement Math.pow using FPU instructions and inline it in crankshaft (ia32).

Implement Math.pow using FPU instructions and inline it in crankshaft (ia32).

Committed: http://code.google.com/p/v8/source/detail?r=10133

[Yang, 2011-11-30 15:53:03]

Please take a look.

[Bill Hesse, 2011-12-01 13:54:24]

This is great!  I worked on the previous code for this, and didn't think that
all the corner cases could be detected, but you do it by testing the status word
after the log and exp calculations.  I don't see where you clear the status
word, before these calculations, though.

http://codereview.chromium.org/8749002/diff/4013/src/ia32/code-stubs-ia32.cc
File src/ia32/code-stubs-ia32.cc (right):

http://codereview.chromium.org/8749002/diff/4013/src/ia32/code-stubs-ia32.cc#...
src/ia32/code-stubs-ia32.cc:3078: __ fnstsw_ax();
Where does the status word get cleared?  How do we accumulate exceptions from
just the above code, and not earlier?  Do you use the fnclex instruction?

[ulan, 2011-12-01 18:11:26]

LGTM with comments.

http://codereview.chromium.org/8749002/diff/5005/src/ia32/code-stubs-ia32.cc
File src/ia32/code-stubs-ia32.cc (right):

http://codereview.chromium.org/8749002/diff/5005/src/ia32/code-stubs-ia32.cc#...
src/ia32/code-stubs-ia32.cc:2965: __ j(greater_equal, &generic_runtime);
j(equal) seems to be sufficient as ecx cannot be greater than kExponentMask.

http://codereview.chromium.org/8749002/diff/5005/src/ia32/code-stubs-ia32.cc#...
src/ia32/code-stubs-ia32.cc:2968: __ jmp(&unpack_exponent, Label::kNear);
Move the jump before the empty line?

http://codereview.chromium.org/8749002/diff/5005/src/ia32/code-stubs-ia32.cc#...
src/ia32/code-stubs-ia32.cc:2981: if (exponent_type_ == ON_STACK) {
I wonder if copy-pasting and keeping the ON_STACK and TAGGED cases disjoint
would be better.

http://codereview.chromium.org/8749002/diff/5005/src/ia32/code-stubs-ia32.cc#...
src/ia32/code-stubs-ia32.cc:2998: __ extractps(ecx, xmm1, 4 * kBitsPerByte);
Since we already use HeapNumber::kExponentMask below, maybe replace 4 by
(HeapNumber::kExponentOffset - HeapNumber::kMantissaOffset)?

http://codereview.chromium.org/8749002/diff/5005/src/ia32/code-stubs-ia32.cc#...
src/ia32/code-stubs-ia32.cc:3029: // xmm3 now has -0.5.
xmm4 now has -0.5.

http://codereview.chromium.org/8749002/diff/5005/src/ia32/code-stubs-ia32.cc#...
src/ia32/code-stubs-ia32.cc:3040: __ extractps(ecx, xmm1, 4 * kBitsPerByte);
Redundant instruction.

http://codereview.chromium.org/8749002/diff/5005/src/ia32/code-stubs-ia32.cc#...
src/ia32/code-stubs-ia32.cc:3098: // xmm0: base as double that is not +/-
Infinity or NaN
xmm0 -> xmm1

http://codereview.chromium.org/8749002/diff/5005/src/ia32/code-stubs-ia32.cc#...
src/ia32/code-stubs-ia32.cc:3100: // Save exponent in base as we need to check
if exponent is negative later.
base -> ecx in the comment

http://codereview.chromium.org/8749002/diff/5005/src/ia32/code-stubs-ia32.cc#...
src/ia32/code-stubs-ia32.cc:3121: // base has the original value of the exponent
- if the exponent  is
base -> ecx in the comment

http://codereview.chromium.org/8749002/diff/5005/src/ia32/code-stubs-ia32.cc#...
src/ia32/code-stubs-ia32.cc:3138: // xmm1: result
xmm1 -> xmm3 in the comment.

http://codereview.chromium.org/8749002/diff/5005/test/mjsunit/math-pow.js
File test/mjsunit/math-pow.js (right):

http://codereview.chromium.org/8749002/diff/5005/test/mjsunit/math-pow.js#new...
test/mjsunit/math-pow.js:140:
assertTrue((1*((Math.pow(2,53))-1)*(Math.pow(2,-1074))) ===
4.4501477170144023e-308);
Lines are too long.

http://codereview.chromium.org/8749002/diff/9030/src/ia32/code-stubs-ia32.cc
File src/ia32/code-stubs-ia32.cc (right):

http://codereview.chromium.org/8749002/diff/9030/src/ia32/code-stubs-ia32.cc#...
src/ia32/code-stubs-ia32.cc:2998: UNREACHABLE();
Since you listed all the cases above, the default case can be removed. This will
enable static checks by compiler instead of run-time checks by UNREACHABLE().