AMDGPU: Remove a fixme for ptrrtoint handling

[arsenm]

Summary

We should not attempt to handle ptrtoint and just give up.

Also use cast<> instead of dyn_cast It should be
impossible to have an alloca with a user which is
not an instruction.

[• tstellarAMD]

LGTM.

[arsenm]

r262854

lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp

if (CallInst *CI = dyn_cast<CallInst>(User)) {
       // constantexpr bitcast of a function.
       if (!CI->getCalledFunction())
         return false;
 
       WorkList.push_back(User);
       continue;
     }
 
-    // FIXME: Correctly handle ptrtoint instructions.
-    Instruction *UseInst = dyn_cast<Instruction>(User);
-    if (UseInst && UseInst->getOpcode() == Instruction::PtrToInt)
+    Instruction *UseInst = cast<Instruction>(User);
+    if (UseInst->getOpcode() == Instruction::PtrToInt)
       return false;
 
     if (StoreInst *SI = dyn_cast_or_null<StoreInst>(UseInst)) {
       // Reject if the stored value is not the pointer operand.
       if (SI->getPointerOperand() != Val)
         return false;
     }