[PATCH] [TTI/CostModel] improve TTI::getGEPCost and use it in CostModel::getInstructionCost

[Jingyue Wu]

Makes sense. Thank you!

On Wed, May 20, 2015 at 9:25 AM, hfinkel at anl.gov <hfinkel at anl.gov> wrote:

> ================
> Comment at: include/llvm/Analysis/TargetTransformInfoImpl.h:392
> @@ +391,3 @@
> +  unsigned getGEPCost(const Value *Ptr, ArrayRef<Value *> Operands) {
> +    const GlobalValue *BaseGV = dyn_cast<GlobalValue>(Ptr);
> +    bool HasBaseReg = (BaseGV == nullptr);
> ----------------
> jingyue wrote:
> > hfinkel wrote:
> > > hfinkel wrote:
> > > > You should call stripPointerCasts before doing the dyn_cast.
> > > Should be dyn_cast_or_null b/c the cost model should not require the
> base pointer be known when estimating costs.
> > Interesting -- when Ptr == nullptr, what is a good estimation? Fall back
> to using `getAddressComputationCost`?
> No, but there is another issue here. We're currently moving to give
> pointer an opaque type, and so you'll need to separately require a Type* on
> this interface. Please add that. Once you have the Type *, you should:
>
>  1. assert that the type is equal to the pointer element type of the
> provided pointer (if it is provided) -- we'll remove this once we actually
> get rid of pointer element types, but for now we need the asserts
>  2. use that type in the code below (there is a getGEPReturnType overload
> that takes the element type instead of the pointer itself)
>  3. Make some default assumption for HasBaseReg (I think you should assume
> true).
>
> http://reviews.llvm.org/D9819
>
> EMAIL PREFERENCES
>   http://reviews.llvm.org/settings/panel/emailpreferences/
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150520/1d56726b/attachment.html>