[llvm-commits] [PATCH] Bug in APCS

[Eric Christopher]

On Oct 12, 2010, at 2:11 PM, Bob Wilson wrote:

> 
> On Oct 12, 2010, at 1:55 PM, Chris Lattner wrote:
> 
>> 
>> On Oct 12, 2010, at 1:51 PM, Eric Christopher wrote:
>> 
>>>>> Sure, but what do you think should happen with the return value? Seems like whatever generated the arg passing in the first place should have done the right thing?
>>>> 
>>>> I'm not sure what you mean.  The code generator shouldn't crash on IR that returns a <3 x double>.  Codegen should do something sensible even though it isn't constrained to match a specific ABI.
>>> 
>>> Sure, I'd prefer it raise an error if it's a type that it can't handle conformantly rather than generate code that's unexpected.
>> 
>> What frontend/language are you assuming?  Not all frontends have types that are defined by the ABI.  The code generator should still do something useful with them.
> 
> Are we just talking about raising the size limit for supported vector return values from 4 to 6 registers?  Why stop there?  By your logic, the code generator should do something useful with every possible return type.  I don't see how that's possible in general for ARM.
> 
> If there's a specific project that requires 6 registers only, that's fine and I have no objections to the patch.  If this is just one instance of a larger problem, i.e., if someone is going to come back next month and want 8 or 10 register return values, then we might as well draw the line at 4.

This will, of course, allow your C front end to decide to mistakenly return 3xdouble in registers and the backend won't notice... if we want to have a custom ABI for it I'm fine with that, just saying it's going to be different.

-eric