[llvm] [NVPTX] Improve copy avoidance during lowering. (PR #106423)

[Artem Belevich via llvm-commits]

================
@@ -79,15 +79,15 @@
 //
 //    define void @foo({i32*, i32*}* byval %input) {
 //      %b_param = addrspacecat ptr %input to ptr addrspace(101)
-//      %b_ptr = getelementptr {ptr, ptr}, ptr addrspace(101) %b_param, i64 0, i32 1
-//      %b = load ptr, ptr addrspace(101) %b_ptr
-//      %b_global = addrspacecast ptr %b to ptr addrspace(1)
-//      ; use %b_generic
+//      %b_ptr = getelementptr {ptr, ptr}, ptr addrspace(101) %b_param, i64 0,
+//      i32 1 %b = load ptr, ptr addrspace(101) %b_ptr %b_global = addrspacecast
----------------
Artem-B wrote:

Fixed.

https://github.com/llvm/llvm-project/pull/106423