[PATCH] Caller/Calllee unsafe-fp-math attribute fixup prior to inlining.
chandlerc at google.com
Tue May 6 11:46:06 PDT 2014
On Tue, May 6, 2014 at 11:43 AM, Philip Reames <listmail at philipreames.com>wrote:
> On 05/05/2014 03:12 PM, Chandler Carruth wrote:
>> 3) add some more noisy guidance that textual attributes should only be
>> used for *target* attributes, not ones for which IR-optimizers are
>> semantically aware
> Point 3 is the only part of this I disagree with. One important secondary
> use for string attributes is while prototyping attributes. Adding a string
> attributes is _much_ easier than adding enum ones. I've got several such
> string attributes in out of tree branches at the moment. Most are planned
> for cleanup and submission (as full attributes), but being able to
> prototype them rapidly was very useful.
Oh, sure. Anything experimental seems fine here. But I wouldn't really
expect to start threading such an attribute's semantics through the trunk
optimizer... If there was some need to keep an attribute clearly marked
"experimental" for a time in trunk until it settled, that also seems not
unreasonable. But I don't think that's the situation we're in here.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-commits