[llvm] [NVPTX] Improve lowering of v4i8 (PR #67866)

[Artem Belevich via llvm-commits]

================
@@ -3307,6 +3312,10 @@ let hasSideEffects = false in {
                                    Int16Regs:$d3, Int16Regs:$d4),
                              (ins Int64Regs:$s),
                              "mov.b64 \t{{$d1, $d2, $d3, $d4}}, $s;", []>;
+  def I32toV4I8  : NVPTXInst<(outs Int16Regs:$d1, Int16Regs:$d2,
+                                   Int16Regs:$d3, Int16Regs:$d4),
+                             (ins Int32Regs:$s),
+                             "mov.b32 \t{{$d1, $d2, $d3, $d4}}, $s;", []>;
----------------
Artem-B wrote:

Now I wonder if this works. Will check.

https://github.com/llvm/llvm-project/pull/67866