Add 'remark' diagnostic type in LLVM

[Quentin Colombet]

Hi Tobias,

On Feb 25, 2014, at 4:19 AM, Tobias Grosser <tobias at grosser.es> wrote:

> On 02/25/2014 11:48 AM, Alp Toker wrote:
>> 
>> On 25/02/2014 10:18, Tobias Grosser wrote:
>>> Hi,
>>> 
>>> I would like to add the following patch.
>>> 
>>> --------------
>>> Add 'remark' diagnostic type in LLVM
>>> 
>>> A 'remark' is information that is not an error or a warning, but
>>> rather some additional information provided to the user. In contrast
>>> to a 'note' a 'remark' is an independent diagnostic, whereas a 'note'
>>> always depends on another diagnostic.
>>> 
>>> A typical use case for remark nodes is information provided to the
>>> user, e.g. information provided by the vectorizer about loops that
>>> have been vectorized.
>> 
>> 
>> One nice thing about warnings is that they can be suppressed with -Wno
>> or upgraded to errors with -Werror= and effort has been made to ensure
>> each warning has an associated option (mostly useful for built-in
>> diagnostics).
>> 
>> I'm guessing this remark capability is intended to support Quentin's
>> backend diagnostics -- could you provide some background on the planned
>> mechanism to enable or suppress them?
> 
> Hi Alp,
> 
> yes, this is intended for Quentin's backend diagnostics and the question you ask is very valid.
> 
> The current diagnostics emitted in the backend are all warnings or errors and are controlled by explicit flags (-Wbackend-plugin, -Winline-asm,..). This is already taken care of.
> 
> For the 'remark' diagnostics I just proposed I did not yet add any flags. I also do not think it makes sense to use -Werror or something to promote those remarks to errors.
Agreed.

> 
> However, enabling them using command line options is necessary. I see two approaches here:
> 
> 1) Model them similar to the -W flags
> 
> We could use something like '-Rvector '-Rno-vector' to enable these diagnostics. And then allow more fine-grained control with '-Rloop-vector' or '-Rslp-vector'.
> 
> -R does not yet have any meaning for clang and gcc, so using it for this purpose may be fine. This may also give us another data-point for the 'info' vs. 'remark' discussion currently going on as "-I" is already taken.
> 
> To later enable different verbosity modes, we could use a formulation such as '-Rvector=3' with '-Rno-vector' equal to '-Rvector=0’.
I have a preference for this approach. This will be easier to present the information in a consistent way to the user.
On the other hand, we may want to be compatible with gcc flags. But I guess we can produce some aliases if this is a requirement.

Regarding the patch itself.
Assuming we pursue in that direction, this is breaking the API compatibility:
   LTO_DS_WARNING,
+  LTO_DS_REMARK,
   LTO_DS_NOTE
 } lto_codegen_diagnostic_severity_t;

More sure to bump the LTO API version with that change.

Thanks,
-Quentin

> 
> 2) Have separate flags
> 
> We could control those also with flags such as -vec-report=3 (icc) or
> -ftree-vectorizer-verbose=n (gcc)
> 
> This is less systematic but possibly more straightforward. And as those 'remarks' are off by default we may not really need a systematic way to identify the flags needed to turn off the diagnostics.
> 
> Do we have any preferences here?
> 
> Cheers,
> Tobias

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140225/0db1f733/attachment.html>