V8 team,

I'm trying to fix http://crbug.com/50972 , where the intrinsics I
added some time ago for Typed Arrays are using the wrong rounding
semantics. The associated inline caches are built into the snapshot,
and because the Serializer is active at the time the CPU features are
probed, in 32-bit mode the system thinks that there is neither SSE2
nor SSE3 support when the ICs are being generated. My current fix only
works completely properly on SSE3 capable CPUs, and mostly properly
with SSE2.

Is it possible to cause some of the builtins to be generated at run
time rather than into the snapshot? I realize that doing so could
negatively impact startup performance, but these ICs are small and
currently rarely used, so lazy generation would be appropriate.

The CL in progress is at http://codereview.chromium.org/6158005 . I'd
appreciate any advice or feedback, especially on fixing the TODOs.
IntegerConvert in code-stubs-ia32.cc seems overly complex as the
fallback for this particular use case, since these external array
types don't need to handle conversion failures.

Thanks,

-Ken

-- 
v8-dev mailing list
[email protected]
http://groups.google.com/group/v8-dev

Reply via email to