[AArch64] Lower __builtin_bswap16 to rev16 if return value is 16-bit

Fixes #77222.

Author: adprasad-nvidia

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -22137,6 +22137,22 @@ static SDValue performExtendCombine(SDNode *N,
       N->getOperand(0)->getOpcode() == ISD::SETCC)
     return performSignExtendSetCCCombine(N, DCI, DAG);
 
+  // If we see (any_extend (bswap ...)) with bswap returning an i16, we know
+  // that the top half of the result register must be unused, due to the
+  // any_extend. This means that we can replace this pattern with (rev16
+  // (any_extend ...)). This saves a machine instruction compared to (lsr (rev
+  // ...)), which is what this pattern would otherwise be lowered to.
+  if (N->getOpcode() == ISD::ANY_EXTEND &&
+      N->getOperand(0).getOpcode() == ISD::BSWAP &&
+      N->getOperand(0).getValueType().isScalarInteger() &&
+      N->getOperand(0).getValueType().getFixedSizeInBits() == 16) {
+    SDNode *BswapNode = N->getOperand(0).getNode();
+    SDValue NewAnyExtend = DAG.getNode(ISD::ANY_EXTEND, SDLoc(BswapNode),
+                                       EVT(MVT::i32), BswapNode->getOperand(0));
+    return DAG.getNode(AArch64ISD::REV16, SDLoc(N), N->getValueType(0),
+                       NewAnyExtend);
+  }
+
   return SDValue();
 }
 

llvm/lib/Target/AArch64/AArch64InstrInfo.td

@@ -758,6 +758,8 @@ def AArch64mvni_msl : SDNode<"AArch64ISD::MVNImsl", SDT_AArch64MOVIshift>;
 def AArch64movi : SDNode<"AArch64ISD::MOVI", SDT_AArch64MOVIedit>;
 def AArch64fmov : SDNode<"AArch64ISD::FMOV", SDT_AArch64MOVIedit>;
 
+def AArch64rev16_scalar : SDNode<"AArch64ISD::REV16", SDTIntUnaryOp>;
+
 def AArch64rev16 : SDNode<"AArch64ISD::REV16", SDT_AArch64UnaryVec>;
 def AArch64rev32 : SDNode<"AArch64ISD::REV32", SDT_AArch64UnaryVec>;
 def AArch64rev64 : SDNode<"AArch64ISD::REV64", SDT_AArch64UnaryVec>;
@@ -2840,6 +2842,8 @@ def : Pat<(bswap (rotr GPR64:$Rn, (i64 32))), (REV32Xr GPR64:$Rn)>;
 def : Pat<(srl (bswap top16Zero:$Rn), (i64 16)), (REV16Wr GPR32:$Rn)>;
 def : Pat<(srl (bswap top32Zero:$Rn), (i64 32)), (REV32Xr GPR64:$Rn)>;
 
+def : Pat<(AArch64rev16_scalar GPR32:$Rn), (REV16Wr GPR32:$Rn)>;
+
 def : Pat<(or (and (srl GPR64:$Rn, (i64 8)), (i64 0x00ff00ff00ff00ff)),
               (and (shl GPR64:$Rn, (i64 8)), (i64 0xff00ff00ff00ff00))),
           (REV16Xr GPR64:$Rn)>;

llvm/test/CodeGen/AArch64/bswap.ll

@@ -3,16 +3,32 @@
 ; RUN: llc -mtriple=aarch64 -global-isel %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-GI
 
 ; ====== Scalar Tests =====
-define i16 @bswap_i16(i16 %a){
-; CHECK-LABEL: bswap_i16:
+define i16 @bswap_i16_to_i16(i16 %a){
+; CHECK-SD-LABEL: bswap_i16_to_i16:
+; CHECK-SD:       // %bb.0:
+; CHECK-SD-NEXT:    rev16 w0, w0
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: bswap_i16_to_i16:
+; CHECK-GI:       // %bb.0:
+; CHECK-GI-NEXT:    rev w8, w0
+; CHECK-GI-NEXT:    lsr w0, w8, #16
+; CHECK-GI-NEXT:    ret
+    %3 = call i16 @llvm.bswap.i16(i16 %a)
+    ret i16 %3
+}
+declare i16 @llvm.bswap.i16(i16)
+
+define i32 @bswap_i16_to_i32(i16 %a){
+; CHECK-LABEL: bswap_i16_to_i32:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    rev w8, w0
 ; CHECK-NEXT:    lsr w0, w8, #16
 ; CHECK-NEXT:    ret
-    %3 = call i16 @llvm.bswap.i16(i16 %a)
-    ret i16 %3
+  %3 = call i16 @llvm.bswap.i16(i16 %a)
+  %4 = zext i16 %3 to i32
+  ret i32 %4
 }
-declare i16 @llvm.bswap.i16(i16)
 
 define i32 @bswap_i32(i32 %a){
 ; CHECK-LABEL: bswap_i32: