[LLVMdev] Type strengthening and type weakening

[Chris Lattner]

On Sep 16, 2009, at 12:12 AM, Sebastian Redl wrote:

> Talin wrote:
>> For example, the GWT compiler does type strengthening - that is, if  
>> you
>> are calling a method on an interface or abstract type, and the  
>> compiler
>> determines through live variable analysis what the concrete type is,
>> then it goes ahead and re-writes the type information to be the  
>> stronger
>> type. The advantage is that it may then be able to do additional
>> optimizations, such as inlining the method or avoiding indirect  
>> dispatch.
>>
>> Conversely, Java's "type erasure" is an extreme case of type  
>> weakening -
>> that is, if you have a bunch of functions which operate on similar  
>> types
>> (which might be produced through something like C++ template
>> instantiation), you can throw away all of the information about those
>> types that aren't relevant to those functions, and you may find as a
>> result that many of them become identical and can be merged into a
>> single function.
>>
>> In the context of LLVM, it would be interesting to have a pass that
>> could go through a function and transform the types of the input
>> arguments to be opaque except where the function actually needs the  
>> type
>> information - so for example if a pointer argument is never  
>> dereferenced
>> it can be converted into a void *.
>>
> Given that LLVM doesn't have much information about the relationships
> between types, would this be worth it? Such an optimization should IMO
> be done by the front-end, which has higher-level information about
> types, such as class hierarchies.

Getting this information into the IR is exactly what the extensible  
metadata proposal is all about :)
http://nondot.org/sabre/LLVMNotes/ExtensibleMetadata.txt

-Chris