Fix crash in optimized code caused by missing IC data.

Fix crash in optimized code caused by missing IC data.

When inling intrinsified methods like String.isEmpty that are
not native, we have to emit the full unoptimized code so that
those method can be correctly optimized and inlined later.

Also, remove an invalid assertion in the type propagator: ToNullableCid
may be called on values from the callee graph in the inliner where
type propagation has not yet been performed.

BUG=https://code.google.com/p/dart/issues/detail?id=12691
TEST=tests/language/optimized_isempty_test.dart
R=srdjan@google.com

Committed: https://code.google.com/p/dart/source/detail?r=26629

[srdjan, 2013-08-23 14:54:04]

LGTM, although it seems a little brittle. Every time we inline an intrinsified
method we must manually check that it generates full unoptimized code. Should we
byte the bullet and always generate unoptimized code aka 'slow' path?

https://codereview.chromium.org/23123012/diff/1/runtime/vm/intrinsifier_arm.cc
File runtime/vm/intrinsifier_arm.cc (right):

https://codereview.chromium.org/23123012/diff/1/runtime/vm/intrinsifier_arm.c...
runtime/vm/intrinsifier_arm.cc:1483: return false;
Please add a comment why we need to generate 'slow' path. Ditto for other
architectures.

[Florian Schneider, 2013-08-26 09:01:34]

On 2013/08/23 14:54:04, srdjan wrote:
> LGTM, although it seems a little brittle. Every time we inline an intrinsified
> method we must manually check that it generates full unoptimized code. Should
we
> byte the bullet and always generate unoptimized code aka 'slow' path?
> 
> https://codereview.chromium.org/23123012/diff/1/runtime/vm/intrinsifier_arm.cc
> File runtime/vm/intrinsifier_arm.cc (right):
> 
>
https://codereview.chromium.org/23123012/diff/1/runtime/vm/intrinsifier_arm.c...
> runtime/vm/intrinsifier_arm.cc:1483: return false;
> Please add a comment why we need to generate 'slow' path. Ditto for other
> architectures.

I agree it's tricky to decide case-by-case. I think we can just always generate
the full code without any noticable cost because those are small methods,
anyway. I'll do that
as a separate CL though.

[Florian Schneider, 2013-08-26 09:46:43]

Committed patchset #1 manually as r26629 (presubmit successful).