[Polly][PATCH 6/8] [Fix] Expression builder loads isl access expressions

Tobias Grosser tobias at grosser.es
Sun Aug 10 23:43:35 PDT 2014


I think this one really needs a better explanation. Specifically, access 
expressions can be both loads/writes. Why did you choose to use a load 
instruction/not a write instruction? Why is this not a bug? Why did we 
not create loads/stores before.

We can really give more explanation in the commit messages to allow 
people who are not following your patches. And for the ones who review 
these patches it is also helpful to understand if you thought about 
those questions.

To me it seems this works for now, as we never do a pure load/store, but
only do the address_of + access computation.

What about moving the address computation into a separate function that 
is called directly from address_of, and putting an assert in case we
detect an actual op_access as we can not really decide if a load or a 
store would be needed.

Tobias



More information about the llvm-commits mailing list