562857 - TM: use LIR_d2i instead of js_DoubleToInt32() when possible

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Nicholas Nethercote [inactive]]

Attachment: patch

In makeNumberInt32(), we convert a double to an int, then convert the int back to a double and compare it against the first double;  if not equal, we exit.  In other words, we exit if the double wasn't integral.

We sometimes use js_DoubleToInt32() for the double-to-int conversion, but that's overkill, we can instead use the LIR_d2i instruction.  LIR_d2i uses platform rounding rules for inexact cases, but it gives the same results as js_DoubleToInt32() for integral cases, which are the ones that will avoid side-exits.  (At least, that's my understanding;  the main part of reviewing this patch should be confirming that this is true.)

This improves instruction counts on a few SS benchmarks (i386):

3d-morph:
  total          74,361,061      67,229,750     1.106x better
  on-trace       26,384,213      25,520,222     1.034x better
3d-raytrace:
  total         141,205,085     138,777,675     1.017x better
  on-trace       26,154,743      25,661,153     1.019x better
access-fannkuch:
  total         241,803,476     227,917,250     1.061x better
  on-trace      117,430,228     113,871,377     1.031x better
string-validate-input:
  total          83,040,900      82,153,955     1.011x better
  on-trace        8,045,440       7,853,415     1.024x better
-------
all:
  total       2,522,688,742   2,498,272,094     1.010x better
  on-trace      627,236,604     622,128,042     1.008x better

And one V8 benchmark:

v8-crypto:
  total       2,592,495,616   2,519,708,387     1.029x better
  on-trace    1,488,287,032   1,472,364,445     1.011x better

Timewise, fannkuch is about 3--5% faster, 3d-morph is 1--10%, others have smaller improvements.  (I timed on 32-bit and 64-bit, hence the big variations.)
V8 timing differences were below noise level.

Comment 1

[Andreas Gal :gal]

Comment on attachment 442598 [details] [diff] [review]
patch

I think splitting that up and then using separate entry points instead of the boolean flag would be nicer. Up to you. Nice work. I did the work for F2I a while but never got around to actually wire it up. This could be used for other stuff too (BoxDouble i.e.).

Comment 2

[Nicholas Nethercote [inactive]]

(In reply to comment #1)
> This could be used for other stuff too (BoxDouble i.e.).

I discussed this with Andreas on IRC and I think the conclusion was that it can't be used for BoxDouble, that the use done in the patch is the only suitable one.

Comment 3

[Nicholas Nethercote [inactive]]

http://hg.mozilla.org/tracemonkey/rev/cc661d5e1ce3

Comment 4

[Robert Sayre]

http://hg.mozilla.org/mozilla-central/rev/cc661d5e1ce3