[PATCH] Implement ADRP CSE for global symbols

[Jiangning Liu]

Hi Quentin,

> > But if we set MergedGlobal's alignment to RoundUpToPowerOfTwo(sizeof(MergedGlobal)), both syntax A and B should be correct.
>
> Yes, but my point is, we do not want to have this big alignment. Especially because it may make sense *only* for ARM64. Moreover, the fact that this alignment works is because both ADD and LDR have the same encoding space for the immediate. In my opinion this is a lot of assumptions that would be hard to match for all targets to make that a desirable change.h for all targets to make that a desirable change.

I see your point now. I think you are correct and there is an assumption here. What about we create a target specific hook and allow target/back-end to choose the alignment between RoundupToPowerOfTwo(sizeof(MergedGlobal)) and NaturalAlignment(MergedGlobal)?

Thanks,
-Jiangning

http://reviews.llvm.org/D3432