[cfe-dev] Notes in Clang diagnostics and LLVM DiagnosticInfo

[Tobias Grosser]

On 02/25/2014 12:07 AM, Richard Smith wrote:
> On Sun, Feb 23, 2014 at 1:21 AM, Tobias Grosser <tobias at grosser.es> wrote:
>
>> On 02/23/2014 10:08 AM, tsett wrote:
>>
>>> Hi Tobias,
>>> I got the problem.
>>> I use the DiagnostigBuilder::setForceEmit() for this.
>>> An instance of DiagnostigBuilder will be returned if you use the
>>> report-function.
>>>
>>
>> Good point. The following patch makes my messages appear again:
>>
>> --- a/lib/CodeGen/CodeGenAction.cpp
>> +++ b/lib/CodeGen/CodeGenAction.cpp
>> @@ -384,7 +384,9 @@ void BackendConsumer::DiagnosticHandlerImpl(const
>> DiagnosticInfo &DI) {
>>
>>     // Report the backend message using the usual diagnostic mechanism.
>>     FullSourceLoc Loc;
>> -  Diags.Report(Loc, DiagID).AddString(MsgStorage);
>> +  DiagnosticBuilder DB = Diags.Report(Loc, DiagID);
>> +  DB.setForceEmit();
>> +  DB.AddString(MsgStorage);
>>
>> We could use this to either force printing of all backend diagnostics or
>> only the notes. However, I am afraid that would break existing flags to
>> suppress diagnostics.
>>
>
> Please don't do this. Notes are designed to attach additional information
> to the prior non-note diagnostic, and this assumption is pervasive within
> the diagnostics mechanism inside clang, as well as in the model we present
> to our users and in our API.

That's what I assumed. Thanks for verifying.

> Adding an 'info' or 'remark' level for diagnostics seems much more in line
> with our existing design. It may be of somewhat limited utility, but the
> cost of having this support seems rather low. (In the long term, I assume
> that end users of polly will not care about these diagnostics, and that
> they're primarily to aid people developing or testing polly -- and
> certainly this does not match how other optimization passes in LLVM report
> debug or progress information to their users.) FWIW, some other compilers
> have a 'remark' diagnostic level, for messages that don't rise to the level
> of 'warning'.

The intention is to support something similar to -vec-report in icc. 
Information about loops that have been vectorized, transformed, ...

To my knowledge, at the moment LLVM is not really providing such kind of 
diagnostics to the users. The majority of the users probably would not 
look at them, but people working on performance or trying to understand 
what happens will be interested in such kind of diagnostics.

I submitted two patches to LLVM and clang that introduce new 'remark' 
level diagnostics.

Cheers,
Tobias