[llvm] [NVPTX] Add truncate and zero-extend free type conversions (PR #125580)

[Justin Fargnoli via llvm-commits]

================
@@ -150,6 +150,28 @@ class NVPTXTargetLowering : public TargetLowering {
            DstTy->getPrimitiveSizeInBits() == 32;
   }
 
+  bool isTruncateFree(EVT SrcVT, EVT DstVT) const override {
+    // Truncating from i64 to i32 is free
+    if (SrcVT.isInteger() && DstVT.isInteger())
+      return SrcVT.getSizeInBits() == 64 && DstVT.getSizeInBits() == 32;
+    return false;
+  }
+
+  bool isZExtFree(EVT FromVT, EVT ToVT) const override {
+    // Zero-extending from i32 to i64 is free
+    if (FromVT.isInteger() && ToVT.isInteger())
+      return FromVT.getSizeInBits() == 32 && ToVT.getSizeInBits() == 64;
+    return false;
+  }
+
+  bool isZExtFree(Type *SrcTy, Type *DstTy) const override {
+    // Zero-extending from i32 to i64 is free
+    if (SrcTy->isIntegerTy() && DstTy->isIntegerTy())
+      return SrcTy->getPrimitiveSizeInBits() == 32 &&
+             DstTy->getPrimitiveSizeInBits() == 64;
+    return false;
+  }
----------------
justinfargnoli wrote:

I'm no longer confident that `zext` is free in this case. 

In [this simple example](https://godbolt.org/z/b77hd3f5s) an instruction is issued to zero out `R3`. Are there counter-examples that show why it is free? 

https://github.com/llvm/llvm-project/pull/125580