[RISCV] Add a DAG combine to pre-promote (i32 (and (srl X, Y), 1)) with Zbs on RV64.

[craig.topper]

Summary

Type legalization will want to turn (srl X, Y) into RISCVISD::SRLW,
which will prevent us from using a BEXT instruction.

I don't think there is any precedent for type promotion checking
users to decide how to promote. Instead, I've added this DAG combine to
do it before type legalization.

[craig.topper]

Make sure shift amount is non-constant to prevent infinite loops.

[craig.topper]

Ping

[asb]

LGTM. Thanks Craig!

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

 static SDValue performSUBCombine(SDNode *N, SelectionDAG &DAG) {
   // fold (sub x, (select lhs, rhs, cc, 0, y)) ->
   //      (select lhs, rhs, cc, x, (sub x, y))
   SDValue N0 = N->getOperand(0);
   SDValue N1 = N->getOperand(1);
   return combineSelectAndUse(N, N1, N0, DAG, /*AllOnes*/ false);
 }
 
-static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG) {
+static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
+                                 const RISCVSubtarget &Subtarget) {
+  SDValue N0 = N->getOperand(0);
+  // Pre-promote (i32 (and (srl X, Y), 1)) on RV64 with Zbs without zero
+  // extending X. This is safe since we only need the LSB after the shift and
+  // shift amounts larger than 31 would produce poison. If we wait until
+  // type legalization, we'll create RISCVISD::SRLW and we can't recover it
+  // to use a BEXT instruction.
+  if (Subtarget.is64Bit() && Subtarget.hasStdExtZbs() &&
+      N->getValueType(0) == MVT::i32 && isOneConstant(N->getOperand(1)) &&
+      N0.getOpcode() == ISD::SRL && !isa<ConstantSDNode>(N0.getOperand(1)) &&
+      N0.hasOneUse()) {
+    SDLoc DL(N);
+    SDValue Op0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N0.getOperand(0));
+    SDValue Op1 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, N0.getOperand(1));
+    SDValue Srl = DAG.getNode(ISD::SRL, DL, MVT::i64, Op0, Op1);
+    SDValue And = DAG.getNode(ISD::AND, DL, MVT::i64, Srl,
+                              DAG.getConstant(1, DL, MVT::i64));
+    return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, And);
+  }
+
   if (SDValue V = combineBinOpToReduce(N, DAG))
     return V;
+
   // fold (and (select lhs, rhs, cc, -1, y), x) ->
   //      (select lhs, rhs, cc, x, (and x, y))
   return combineSelectAndUseCommutative(N, DAG, /*AllOnes*/ true);
 }
 
 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
                                 const RISCVSubtarget &Subtarget) {
   if (Subtarget.hasStdExtZbp()) {
case RISCVISD::FMV_X_ANYEXTW_RV64: {
     return DAG.getNode(ISD::AND, DL, VT, NewFMV,
                        DAG.getConstant(~SignBit, DL, VT));
   }
   case ISD::ADD:
     return performADDCombine(N, DAG, Subtarget);
   case ISD::SUB:
     return performSUBCombine(N, DAG);
   case ISD::AND:
-    return performANDCombine(N, DAG);
+    return performANDCombine(N, DAG, Subtarget);
   case ISD::OR:
     return performORCombine(N, DAG, Subtarget);
   case ISD::XOR:
     return performXORCombine(N, DAG);
   case ISD::FADD:
   case ISD::UMAX:
   case ISD::UMIN:
   case ISD::SMAX:

llvm/test/CodeGen/RISCV/rv64zbs.ll

 ; RV64I-LABEL: bext_i32:
 ; RV64I:       # %bb.0:
 ; RV64I-NEXT:    srlw a0, a0, a1
 ; RV64I-NEXT:    andi a0, a0, 1
 ; RV64I-NEXT:    ret
 ;
 ; RV64ZBS-LABEL: bext_i32:
 ; RV64ZBS:       # %bb.0:
-; RV64ZBS-NEXT:    srlw a0, a0, a1
-; RV64ZBS-NEXT:    andi a0, a0, 1
+; RV64ZBS-NEXT:    andi a1, a1, 31
+; RV64ZBS-NEXT:    bext a0, a0, a1
 ; RV64ZBS-NEXT:    ret
   %and = and i32 %b, 31
   %shr = lshr i32 %a, %and
   %and1 = and i32 %shr, 1
   ret i32 %and1
 }
 
 define signext i32 @bext_i32_no_mask(i32 signext %a, i32 signext %b) nounwind {
 ; RV64I-LABEL: bext_i32_no_mask:
 ; RV64I:       # %bb.0:
 ; RV64I-NEXT:    srlw a0, a0, a1
 ; RV64I-NEXT:    andi a0, a0, 1
 ; RV64I-NEXT:    ret
 ;
 ; RV64ZBS-LABEL: bext_i32_no_mask:
 ; RV64ZBS:       # %bb.0:
-; RV64ZBS-NEXT:    srlw a0, a0, a1
-; RV64ZBS-NEXT:    andi a0, a0, 1
+; RV64ZBS-NEXT:    bext a0, a0, a1
 ; RV64ZBS-NEXT:    ret
   %shr = lshr i32 %a, %b
   %and1 = and i32 %shr, 1
   ret i32 %and1
 }
 
 define i64 @bext_i64(i64 %a, i64 %b) nounwind {
 ; RV64I-LABEL: bext_i64: