On Mon, 2011-08-01 at 12:21 +0200, Andy Wingo wrote:
> The numbers do fluctuate a bit; the only real constant here is that with
> my patch we get minus zero bailouts being set for some unary math ops.
> I'm going to see if I can whittle this down to one test in particular.
Got one :)
On bleeding_edge:
d8> function roundsum(i,n) { var ret = Math.round(n); while (--i > 0)
ret += Math.round(n); return ret; }
d8> roundsum (1000000, 2.2) // warm up
2000000
d8> 1/roundsum (1000000, -0.0)
Infinity
d8>
Change::MinusZeroBailoutCounter = 2
HValue::InferRange::RangeAllocCounter[0] = 0
HValue::InferRange::RangeAllocCounter[1] = 44
HValue::InferRange::RangeAllocCounter[2] = 5
HValue::InferRange::RangeAllocCounter[3] = 2
HValue::InferRange::RangeAllocCounter[4] = 0
With the patch from r8753:
d8> function roundsum(i,n) { var ret = Math.round(n); while (--i > 0)
ret += Math.round(n); return ret; }
d8> roundsum (1000000, 2.2) // warm up
2000000
d8> 1/roundsum (1000000, -0.0)
-Infinity
d8>
UnaryMathOp::MinusZeroBailoutCounter = 10
Change::MinusZeroBailoutCounter = 2
HValue::InferRange::RangeAllocCounter[0] = 0
HValue::InferRange::RangeAllocCounter[1] = 257
HValue::InferRange::RangeAllocCounter[2] = 41
HValue::InferRange::RangeAllocCounter[3] = 11
HValue::InferRange::RangeAllocCounter[4] = 0
As you can see bleeding_edge produces the wrong result for this
calculation. The reason is that it doesn't deoptimize on -0.0, because
the default HValue::InferRange returns an empty range for double values,
but one which assumes that -0.0 is not in the range -- and thus `floor'
does not do a deopt on -0.0.
You can see this in the MinusZeroBailoutCounter, occurring for the
`round'.
Good one, no? :) I'll submit another patch to the code review, with a
regression test.
Andy
--
v8-dev mailing list
[email protected]
http://groups.google.com/group/v8-dev
