Fix bug with range of shift operations.

Fix bug with range of shift operations.

A shift can only deoptimize if the rhs is negative: This is the
case if the range of the rhs is not contained within [0, +inf].

!range->IsPositive() is _not_ the same as range->IsNegative(). I find
these names confusing and we should find better alternatives.

R=vegorov@google.com

Committed: https://code.google.com/p/dart/source/detail?r=37176

[kasperl, 2014-06-10 10:55:15]

Regression test?

[Vyacheslav Egorov (Google), 2014-06-10 11:41:38]

lgtm

[Vyacheslav Egorov (Google), 2014-06-10 11:48:35]

I suspect there is a test that fails, if there is none, please add one. 

I don't think name IsPositive() and IsNegative() are too confusing. We can
rename them to IsOnlyPositive()/IsOnlyNegative or define CanBeNegative() =
!IsPositive() and CanBePositive()  instead of them. 

[Sorry for not spotting this quite obvious bug, I should not do reviews on the
road. Also the discrepancy between ia32 and other architectures for
right_may_be_negative should have screamed at me, but I just glanced over the
code assuming textual replacement with no semantics changes.]

[Florian Schneider, 2014-06-10 12:05:23]

On 2014/06/10 11:48:35, Vyacheslav Egorov (Google) wrote:
> I suspect there is a test that fails, if there is none, please add one. 
> 
> I don't think name IsPositive() and IsNegative() are too confusing. We can
> rename them to IsOnlyPositive()/IsOnlyNegative or define CanBeNegative() =
> !IsPositive() and CanBePositive()  instead of them. 
> 
> [Sorry for not spotting this quite obvious bug, I should not do reviews on the
> road. Also the discrepancy between ia32 and other architectures for
> right_may_be_negative should have screamed at me, but I just glanced over the
> code assuming textual replacement with no semantics changes.]

I'll add a regression test.

[Florian Schneider, 2014-06-10 12:58:40]

Committed patchset #3 manually as r37176 (presubmit successful).