[llvm-commits] RFC fix for the MCContext doInitialization/doFinalization issues

[Jim Grosbach]

On Dec 6, 2012, at 3:24 PM, Chandler Carruth <chandlerc at google.com> wrote:

> On Thu, Dec 6, 2012 at 3:11 PM, Pedro Artigas <partigas at apple.com> wrote:
> 
> Answer inline.
> 
> On Dec 6, 2012, at 2:51 PM, Chandler Carruth <chandlerc at google.com> wrote:
> 
>> On Thu, Dec 6, 2012 at 2:46 PM, Pedro Artigas <partigas at apple.com> wrote:
>> Sure, no problem.
>> 
>> The MCContext is used inside MachineModuleInfo which is a ImmutablePass under the pass manager model, it can also be used standalone. the change added a boolean that indicates if someone will call doInitialization / doFinalization to reuse the MCContext data structure across multiple modules or not, today only MachineModuleInfo does that, and, since it is a Pass it gets it's doInitialization/doFinalization called automagically by the pass manager and calls the MCContext doInitialization / doFinalization methods there.
>> 
>> The issue was that since there are some uses of the MCContext that are outside of the PassManager, those need to have doInitialization and doFinalization called automatically by constructor/destructor as otherwise nobody would call doInitialization/doFinalization (which were the root cause of valgrind bugs in my prior check in)
>> 
>> Is that sufficient context?
>> 
>> This is a fantastic explanation. In the future, I would love to see this type of explanation in the commit log, it helps when reviewing the patch.
>> 
>> As a side-note, I feel it might be a slightly better design to have an MCContext, and an MCContextPass, where the MCContext doesn't have doInitialization / doFinalization (and in fact, doesn't derive from the Pass machinery), but the latter does. The latter can then construct and destroy the MCContext for each module. This design would make the bool flag unnecessary by encoding the two interfaces in the type system itself.
>> 
>> Would that make sense? Or is it really important performance wise to just clear the data structures inside MCContext rather than destroying and recreating them?
> 
> It does make sense to have a MCContextManual and MCContextPass or some other name like this (the non Manual/Pass version, being the more specialized, should derive from the regular/Manual  one). 
> 
> This would save the bool inside the class, (which is insignificant as it is a single bool in a class that has just a single instanced object) but I believe would make the code a little cleaner. Both approaches achieve pretty much the same thing with regards to compile time savings.
> 
> It is important to just clear the data structures, that is the source of compile time savings as there is no extra allocate/deallocate.
> 
> I would structure this as:
> 
> MCContextImpl which exposes non-virtual 'clear' or 'reset' method to clean up the datastructures.
> 
> MCContext, derived from MCContextImpl, which just provides a normal, non-pass instance interface.
> 
> MCContextPass, derived from MCContextImpl and ImmutablePass, and redirects the virtual ImmutablePass interface into the MCContextImpl's reset.
> 
> 
> If for some reason the multiple inheritance is problematic (which would be reasonable) I might go for:
> 
> MCContext, which works much like it used to, but adds a 'reset' method which resets the datastructures.
> 
> MCContextPass, which *contains* an MCContext, exposes it via a 'get' method, and resets it on the doInitialization / doFinalization call (whichever is more appropriate).
> 
> 
> Thoughts Jim? Other parties interested in the MC design here?


I like the second construction (an MCContextPass which contains an MCContext). It feels more intuitively correct to me. Plus I have an allergy to multiple inheritance.

-Jim

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20121206/56940c59/attachment.html>