[LLVMdev] Preferred alignment of globals > 16bytes

[Chris Lattner]

On Sep 18, 2012, at 3:11 AM, Richard Osborne <richard at xmos.com> wrote:
>>> I want it to be configurable. On my target there is no advantage to aligning anything by more than a 4 bytes - it just wastes space. I'll try to put together a proposal for making it possible to set this per target.
>> Ok, after considering this a bit more, how about we take a different approach: despite its name, TargetData is really mostly for consumption by the mid-level optimizers.  getPreferredAlignment is really something that only TargetLowering[ObjectFile] should be using.  What do you think about sinking getPreferredAlignment down to it? This will make it much easier to target-hookize this.
>> 
>> -Chris
> Hi Chris,
> 
> Sorry for not replying sooner but I've haven't had a chance to look at this until now. I think your suggestion makes a lot of sense. The preferred alignment will be a lot easier to change if it is not embedded in the IR in the form of the target data string. I started trying to moving the functions to get preferred alignment to TargetLowering but found there are a few mid-level optimizers that make use of the preferred alignment. In particular:
> 
> * lib/Analysis/ValueTracking.cpp - uses getPreferredAlignment() to determine number of trailing zeros of the address of a global defined in the current module.

This "works", but isn't safe.  It is relying on an assumption that the backend will provide at least that alignment for a global, and assuming that nothing later than it will change the alignment to 1.  This needs to be fixed.

> * lib/Transforms/Utils/InlineFunction.cpp - Uses getPrefTypeAlignment() to set the alignment of the alloca it introduces when a byval argument is inlined
> * lib/Transforms/IPO/ConstantMerge.cpp - Uses getPreferredAlignment() to determine the alignment of the merged constant.

Can't these just leave the alignment unset and let the code generator do its thing?

> * lib/Transforms/Vectorize/BBVectorize.cpp - Depending on what options are set this pass avoids vectorization if it would introduce loads and store of less than the alignment returned by getPreferredAlignment().

I can't comment on this specifically, but this also sounds wrong.  Hal?

> * lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp - getPrefTypeAlignment() is used to explictly set the alignment of allocas with unspecified alignment. Also instcombine tries to enforce preferred alignment for loads and stores if possible.


This case is more interesting to me - the idea here is that instcombine is trying to expose alignment information beyond the ABI alignment for optimization passes, allowing other passes to benefit from it - and allowing alignment inference of loads and stores to the alloca, which may help codegen.

I tend to think that this is the wrong approach - optimizations that would benefit from higher alignment should themselves force the alloca to be more aligned, not rely on instcombine to do it.  The code generator should wait until the alloca gets a specific alignment (which presumably will be the preferred alignment) and then bump up the alignment of loads and stores based on that alignment, instead of just taking what IR has and blindly going with it.

-Chris