[llvm] [NVPTX] Add TLI hook for load slice cost and implement it (PR #131847)

[Alex MacLean via llvm-commits]

================
@@ -3133,6 +3133,20 @@ class TargetLoweringBase {
     return false;
   }
 
+  virtual unsigned getLoadSliceCost(bool ForCodeSize, unsigned Loads,
+                                    unsigned CrossRegisterBanksCopies,
+                                    unsigned Truncates, unsigned ZExts,
+                                    unsigned Shifts) const {
+    // Assume cross register banks copies are as expensive as loads.
+    unsigned ExpensiveOps = Loads + CrossRegisterBanksCopies;
+
+    // Unless we are optimizing for code size, prioritize expensive operations.
+    if (!ForCodeSize)
+      ExpensiveOps = ExpensiveOps * 20;
----------------
AlexMaclean wrote:

The 20 is quite arbitrary. In the previous implementation we first compared expensive ops and only checked inexpensive ops if expensive ops were equal. This meant that expensive ops always trumped inexpensive ops. I don't think that previous logic really made sense as there must be some number of inexpensive ops that is more costly than an expensive op. I chose 20 because in practice I don't think there are going to be any cases where there are 20 inexpensive ops, so this should essentially get us the same behavior. I agree this is a bit weird though, I'm happy to try out any alternatives you can suggest.

https://github.com/llvm/llvm-project/pull/131847