[clang] [Serialization] Load Specializations Lazily (1/2) (PR #76774)

[Vassil Vassilev via cfe-commits]

vgvassilev wrote:

> > > > This is a great way to start a new year ;)
> > > > The phab link is https://reviews.llvm.org/D41416.
> > > > In general I was wondering could we simplify the implementation by loading the specialization hash table upon module load. That should be relatively cheap as we will read 2 integers per specialization.
> > > > Perhaps we should put both patches together and that'd allow us to test them if they are on par with https://reviews.llvm.org/D41416 which we use downstream.
> > > > Thanks for working on this!
> > > 
> > > 
> > > Hi Vassilev, for testing purpose I sent https://github.com/ChuanqiXu9/llvm-project/tree/LoadSpecializationUpdatesLazily. I didn't create stacked review since I feel a standalone branch may be sufficient.
> > 
> > 
> > @ChuanqiXu9, I'd prefer to review both patches at the same time. Otherwise we risk of missing some important details.
> 
> Got it. I can try to create a stacked review. But from I know about the status quo stacked review now, it will require us to lost the current contexnt...
> 
> And it will still be pretty valuable if you can test this with your internal workloads, then may be we can find something pretty important in the high level before going into the details. I've tested this in our local workloads, and it looks good and the performance improvements remains. But I know our uses about modules may be not so complex like yours.

I would just push the second commit here. It should be good enough. 

https://github.com/llvm/llvm-project/pull/76774