[RISCV] Select signed bitfield insert for XAndesPerf (llvm#143356)

This patch is similar to llvm#142737

The XAndesPerf extension includes signed bitfield extraction
instruction `NDS.BFOS, which can extract the bits from 0 to Len - 1,
place them starting at bit Lsb, zero-filled the bits from 0 to Lsb -1, 
and sign-extend the result.

When Lsb == Msb, it is a special case where the Lsb will be set to 0
instead of being equal to the Msb.

Author: tclin914

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

@@ -683,6 +683,59 @@ bool RISCVDAGToDAGISel::trySignedBitfieldExtract(SDNode *Node) {
   return false;
 }
 
+bool RISCVDAGToDAGISel::trySignedBitfieldInsertInSign(SDNode *Node) {
+  // Only supported with XAndesPerf at the moment.
+  if (!Subtarget->hasVendorXAndesPerf())
+    return false;
+
+  auto *N1C = dyn_cast<ConstantSDNode>(Node->getOperand(1));
+  if (!N1C)
+    return false;
+
+  SDValue N0 = Node->getOperand(0);
+  if (!N0.hasOneUse())
+    return false;
+
+  auto BitfieldInsert = [&](SDValue N0, unsigned Msb, unsigned Lsb,
+                            const SDLoc &DL, MVT VT) {
+    unsigned Opc = RISCV::NDS_BFOS;
+    // If the Lsb is equal to the Msb, then the Lsb should be 0.
+    if (Lsb == Msb)
+      Lsb = 0;
+    return CurDAG->getMachineNode(Opc, DL, VT, N0.getOperand(0),
+                                  CurDAG->getTargetConstant(Lsb, DL, VT),
+                                  CurDAG->getTargetConstant(Msb, DL, VT));
+  };
+
+  SDLoc DL(Node);
+  MVT VT = Node->getSimpleValueType(0);
+  const unsigned RightShAmt = N1C->getZExtValue();
+
+  // Transform (sra (shl X, C1) C2) with C1 > C2
+  //        -> (NDS.BFOS X, lsb, msb)
+  if (N0.getOpcode() == ISD::SHL) {
+    auto *N01C = dyn_cast<ConstantSDNode>(N0->getOperand(1));
+    if (!N01C)
+      return false;
+
+    const unsigned LeftShAmt = N01C->getZExtValue();
+    // Make sure that this is a bitfield insertion (i.e., the shift-right
+    // amount should be less than the left-shift).
+    if (LeftShAmt <= RightShAmt)
+      return false;
+
+    const unsigned MsbPlusOne = VT.getSizeInBits() - RightShAmt;
+    const unsigned Msb = MsbPlusOne - 1;
+    const unsigned Lsb = LeftShAmt - RightShAmt;
+
+    SDNode *Sbi = BitfieldInsert(N0, Msb, Lsb, DL, VT);
+    ReplaceNode(Node, Sbi);
+    return true;
+  }
+
+  return false;
+}
+
 bool RISCVDAGToDAGISel::tryUnsignedBitfieldExtract(SDNode *Node,
                                                    const SDLoc &DL, MVT VT,
                                                    SDValue X, unsigned Msb,
@@ -1214,6 +1267,9 @@ void RISCVDAGToDAGISel::Select(SDNode *Node) {
     if (trySignedBitfieldExtract(Node))
       return;
 
+    if (trySignedBitfieldInsertInSign(Node))
+      return;
+
     // Optimize (sra (sext_inreg X, i16), C) ->
     //          (srai (slli X, (XLen-16), (XLen-16) + C)
     // And      (sra (sext_inreg X, i8), C) ->

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.h

@@ -77,6 +77,7 @@ class RISCVDAGToDAGISel : public SelectionDAGISel {
 
   bool tryShrinkShlLogicImm(SDNode *Node);
   bool trySignedBitfieldExtract(SDNode *Node);
+  bool trySignedBitfieldInsertInSign(SDNode *Node);
   bool tryUnsignedBitfieldExtract(SDNode *Node, const SDLoc &DL, MVT VT,
                                   SDValue X, unsigned Msb, unsigned Lsb);
   bool tryUnsignedBitfieldInsertInZero(SDNode *Node, const SDLoc &DL, MVT VT,

llvm/test/CodeGen/RISCV/rv32xandesperf.ll

@@ -154,6 +154,32 @@ define i32 @bfos_from_ashr_sexti16_i32(i16 %x) {
   ret i32 %ashr
 }
 
+; MSB = 0
+
+define i32 @bfos_from_ashr_shl_with_msb_zero_insert_i32(i32 %x) {
+; CHECK-LABEL: bfos_from_ashr_shl_with_msb_zero_insert_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    nds.bfos a0, a0, 0, 14
+; CHECK-NEXT:    ret
+  %shl = shl i32 %x, 31
+  %lshr = ashr i32 %shl, 17
+  ret i32 %lshr
+}
+
+; MSB < LSB
+
+define i32 @bfos_from_ashr_shl_insert_i32(i32 %x) {
+; CHECK-LABEL: bfos_from_ashr_shl_insert_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    nds.bfos a0, a0, 18, 20
+; CHECK-NEXT:    ret
+  %shl = shl i32 %x, 29
+  %lshr = ashr i32 %shl, 11
+  ret i32 %lshr
+}
+
+; sext
+
 define i32 @sexti1_i32(i32 %a) {
 ; CHECK-LABEL: sexti1_i32:
 ; CHECK:       # %bb.0:

llvm/test/CodeGen/RISCV/rv64xandesperf.ll

@@ -212,6 +212,52 @@ define i64 @bfos_from_ashr_sexti16_i64(i16 %x) {
   ret i64 %ashr
 }
 
+; MSB = 0
+
+define i32 @bfos_from_ashr_shl_with_msb_zero_insert_i32(i32 %x) {
+; CHECK-LABEL: bfos_from_ashr_shl_with_msb_zero_insert_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    nds.bfos a0, a0, 0, 14
+; CHECK-NEXT:    ret
+  %shl = shl i32 %x, 31
+  %lshr = ashr i32 %shl, 17
+  ret i32 %lshr
+}
+
+define i64 @bfos_from_ashr_shl_with_msb_zero_insert_i64(i64 %x) {
+; CHECK-LABEL: bfos_from_ashr_shl_with_msb_zero_insert_i64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    nds.bfos a0, a0, 0, 46
+; CHECK-NEXT:    ret
+  %shl = shl i64 %x, 63
+  %lshr = ashr i64 %shl, 17
+  ret i64 %lshr
+}
+
+; MSB < LSB
+
+define i32 @bfos_from_ashr_shl_insert_i32(i32 %x) {
+; CHECK-LABEL: bfos_from_ashr_shl_insert_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    nds.bfos a0, a0, 18, 20
+; CHECK-NEXT:    ret
+  %shl = shl i32 %x, 29
+  %lshr = ashr i32 %shl, 11
+  ret i32 %lshr
+}
+
+define i64 @bfos_from_ashr_shl_insert_i64(i64 %x) {
+; CHECK-LABEL: bfos_from_ashr_shl_insert_i64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    nds.bfos a0, a0, 18, 52
+; CHECK-NEXT:    ret
+  %shl = shl i64 %x, 29
+  %lshr = ashr i64 %shl, 11
+  ret i64 %lshr
+}
+
+; sext
+
 define signext i32 @sexti1_i32(i32 signext %a) {
 ; CHECK-LABEL: sexti1_i32:
 ; CHECK:       # %bb.0: