[llvm] [NVPTX] Use PRMT instruction to lower i16 bswap (PR #168968)

[Alex MacLean via llvm-commits]

================
@@ -2471,33 +2471,9 @@ include "NVPTXIntrinsics.td"
 //-----------------------------------
 // Notes
 //-----------------------------------
-// BSWAP is currently expanded. The following is a more efficient
-// - for < sm_20, use vector scalar mov, as tesla support native 16-bit register
-// - for sm_20, use pmpt (use vector scalar mov to get the pack and
-//   unpack). sm_20 supports native 32-bit register, but not native 16-bit
-// register.
-
-def : Pat <
-  (i32 (bswap i32:$a)),
-  (PRMT_B32rii $a, (i32 0), (i32 0x0123), PrmtNONE)>;
-
-def : Pat <
-  (v2i16 (bswap v2i16:$a)),
-  (PRMT_B32rii $a, (i32 0), (i32 0x2301), PrmtNONE)>;
-
-def : Pat <
-  (i64 (bswap i64:$a)),
-  (V2I32toI64
-    (PRMT_B32rii (I64toI32H_Sink $a), (i32 0), (i32 0x0123), PrmtNONE),
-    (PRMT_B32rii (I64toI32L_Sink $a), (i32 0), (i32 0x0123), PrmtNONE))>,
-  Requires<[hasPTX<71>]>;
-
-// Fall back to the old way if we don't have PTX 7.1.
-def : Pat <
-  (i64 (bswap i64:$a)),
-  (V2I32toI64
-    (PRMT_B32rii (I64toI32H $a), (i32 0), (i32 0x0123), PrmtNONE),
-    (PRMT_B32rii (I64toI32L $a), (i32 0), (i32 0x0123), PrmtNONE))>;
+// BSWAP is currently custom-lowered during operation legalization in
+// NVPTXISelLowering.cpp.
+// See the lowerBSWAP function in NVPTXISelLowering.cpp for details.
----------------
AlexMaclean wrote:

Nit: I don't think this is necessary.

https://github.com/llvm/llvm-project/pull/168968