[llvm] [NVPTX] Fixup v2i8 parameter and return lowering (PR #145585)

[Alex MacLean via llvm-commits]

================
@@ -247,7 +243,14 @@ define <4 x i8> @test_v4i8(<4 x i8> %a) {
 ; CHECK:      call.uni (retval0), test_v5i8,
 ; CHECK-DAG:  ld.param.v4.b8  {[[RE0:%rs[0-9]+]], [[RE1:%rs[0-9]+]], [[RE2:%rs[0-9]+]], [[RE3:%rs[0-9]+]]}, [retval0];
 ; CHECK-DAG:  ld.param.b8     [[RE4:%rs[0-9]+]], [retval0+4];
-; CHECK-DAG:  st.param.v4.b8  [func_retval0], {[[RE0]], [[RE1]], [[RE2]], [[RE3]]}
+; CHECK-DAG:  cvt.u32.u16     [[R3:%r[0-9]+]], [[RE3]];
+; CHECK-DAG:  cvt.u32.u16     [[R2:%r[0-9]+]], [[RE2]];
+; CHECK-DAG:  prmt.b32        [[P0:%r[0-9]+]], [[R2]], [[R3]], 0x3340U;
+; CHECK-DAG:  cvt.u32.u16     [[R1:%r[0-9]+]], [[RE1]];
+; CHECK-DAG:  cvt.u32.u16     [[R0:%r[0-9]+]], [[RE0]];
+; CHECK-DAG:  prmt.b32        [[P1:%r[0-9]+]], [[R0]], [[R1]], 0x3340U;
+; CHECK-DAG:  prmt.b32        [[P2:%r[0-9]+]], [[P1]], [[P0]], 0x5410U;
+; CHECK-DAG:  st.param.b32  [func_retval0], [[P2]];
----------------
AlexMaclean wrote:

Agreed. In general our handling of i8 vectors is pretty poor. I think this is essentially just a new manifestation of what is already a fairly widespread issue. I'm working on replacing how we load and store params for calls with conventional load/store instructions and I think that will recover this case. 

https://github.com/llvm/llvm-project/pull/145585