[llvm] [SelectionDAG][RISCV] Avoid store merging across function calls (PR #130430)

[Philip Reames via llvm-commits]

================
@@ -1070,6 +1070,13 @@ class RISCVTargetLowering : public TargetLowering {
     return false;
   }
 
+  /// Disables storing and loading vectors by default when there are function
+  /// calls between the load and store, since these are more expensive than just
+  /// using scalars
+  bool shouldMergeStoreOfLoadsOverCall(EVT SrcVT, EVT MergedVT) const override {
+    return SrcVT.isScalarInteger() == MergedVT.isScalarInteger();
----------------
preames wrote:

I glanced at what we do for this today.  Quick summary:
2 x half (no zvfh) -> no merge performed
2 x half (zvfh) -> <2 x half> vector result type
2 x float -> <2 x float> vector result type
2 x float + noimplicitfloat -->  two integer loads, no merge

So basically, I think the patch as it is should be a net improvement (by disabling case 2 and 3).  I'm not opposed to your suggest change, just pointing out it's not needed to avoid a regression.  

https://github.com/llvm/llvm-project/pull/130430