[llvm] r297161 - Relax the conflicting function arg verifier to allow for inlined debug

[Adrian Prantl via llvm-commits]

> On Mar 7, 2017, at 11:19 AM, David Blaikie <dblaikie at gmail.com> wrote:
> 
> Ah, I see - might be worth a comment in the test case to describe a bit more about what it's aiming for?
> 
> & would this suffice?
> 
> void sink();
> inline __attribute__((always_inline)) void f1(int) {
>   sink();
> }
> void f2() {
>   f1(0);
>   f1(0);
> }
> 
> (if that's not sufficient, then perhaps the verifier needs to be made stronger to catch this case where they're actually different variables even though they're represented by the same DILocalVariable? (two different inlinings) - or I guess this case it got correct for free/accidentally because it was the same variable object)

The verifier is not smart enough to catch this, no. It also can't catch conflicts for inlined variables. It's a trade-off for speed. Currently it uses constant time lookups, if we support looking into inlined functions we'd have to use DenseMap lookups instead, and I was afraid this might blow up on, e.g., boost code.

-- adrian

> 
> Anyway, all good (:
> 
> On Tue, Mar 7, 2017 at 11:00 AM Adrian Prantl <aprantl at apple.com <mailto:aprantl at apple.com>> wrote:
>> On Mar 7, 2017, at 10:58 AM, David Blaikie <dblaikie at gmail.com <mailto:dblaikie at gmail.com>> wrote:
>> 
>> Ah, it might be more obvious if one of 'f' or 'g' was left standalone/non-inline, then? Or perhaps I'm not fully understanding the conflict being tested for? Two args with the same name but different arg objects in the same function?
> 
> The name is not relevant. The condition is two args with different variables that both claim to be argNo: 1 and are not inlined.
> 
> -- adrian
> 
>> 
>> OK, so I can see why that would require two inlinings, if that's the case. Thanks for explaining!
>> 
>> On Tue, Mar 7, 2017 at 10:49 AM Adrian Prantl <aprantl at apple.com <mailto:aprantl at apple.com>> wrote:
>> Because the outermost function is nodebug and we need at least two function args to create a conflict.
>> 
>> -- adrian
>> 
>>> On Mar 7, 2017, at 10:47 AM, David Blaikie <dblaikie at gmail.com <mailto:dblaikie at gmail.com>> wrote:
>>> 
>>> Thanks! Any idea why it requires both 'f' and 'g' to reproduce this?
>>> 
>>> (& I imagine the return values probably aren't needed?)
>>> 
>>> On Tue, Mar 7, 2017 at 10:39 AM Adrian Prantl <aprantl at apple.com <mailto:aprantl at apple.com>> wrote:
>>> I replaced it with a smaller, hand-crafted testcase in r297176.
>>> 
>>> -- adrian
>>> 
>> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170307/6ec641e6/attachment.html>