[PATCH] Resolve issues with fmuladd intrinsic handling across multiple backends

[Stephen Lin]

Thanks, r185956.

On Tue, Jul 9, 2013 at 10:44 AM, Owen Anderson <resistor at mac.com> wrote:
>
> On Jul 9, 2013, at 9:50 AM, Stephen Lin <swlin at post.harvard.edu> wrote:
>
>> On Tue, Jul 9, 2013 at 9:44 AM, Stephen Lin <swlin at post.harvard.edu> wrote:
>>>> I'm not sure this is what you want.  Graphics-oriented users sometimes have weird vector sizes (<3 x float>, <3 x half>, etc.) that should still get fused if possible.  I think you should be able to implement the checks you need without VT having to be simple.
>>>>
>>>> Otherwise, your SelectionDAG changes look fine.
>>>>
>>>
>>> Hi Owen, in addition to allowing extended types I changed DAGCombiner
>>> to use the LegalOperations boolean to detect whether or not it's safe
>>> to introduce in illegal operation, rather than introducing my own.
>>>
>>> Can you take a look at the new patch and make sure I didn't do
>>> anything silly, if you get a chance? Basically, as I understand it,
>>> LegalOperations is true if we've already legalized all operations and
>>> false otherwise.
>>>
>>> Thanks,
>>> Stephen
>>
>> Sorry, that seems to be the old version. Here's the right one.
>> <fma-faster-check-types.patch>
>
> LGTM.
>
> --Owen