[RISCV] Efficiently lower select %cond, %constant, %variable using zicond extension

Provides the following lowerings, which reduces instruction count by 1 for rv64:
(select c, c1, t) -> (add (czero_nez t - c1, c), c1)
(select c, t, c1) -> (add (czero_eqz t - c1, c), c1)

For the special case of c1 == -2048
(select c, -2048, t) -> (xor (czero_nez (xor t, -2048), c), -2048)
(select c, t, -2048) -> (xor (czero_eqz (xor t, -2048), c), -2048)

Author: bababuck

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -9088,6 +9088,32 @@ SDValue RISCVTargetLowering::lowerSELECT(SDValue Op, SelectionDAG &DAG) const {
       return DAG.getNode(ISD::ADD, DL, VT, CMOV, RHSVal);
     }
 
+    // (select c, c1, t) -> (add (czero_nez t - c1, c), c1)
+    // (select c, t, c1) -> (add (czero_eqz t - c1, c), c1)
+    if (isa<ConstantSDNode>(TrueV) != isa<ConstantSDNode>(FalseV)) {
+      bool IsCZERO_NEZ = isa<ConstantSDNode>(TrueV);
+      SDValue ConstVal = IsCZERO_NEZ ? TrueV : FalseV;
+      SDValue RegV = IsCZERO_NEZ ? FalseV : TrueV;
+      int64_t RawConstVal = cast<ConstantSDNode>(ConstVal)->getSExtValue();
+      // Fall back to XORI if Const == -0x800
+      if (RawConstVal == -0x800) {
+        SDValue XorOp = DAG.getNode(ISD::XOR, DL, VT, RegV, ConstVal);
+        SDValue CMOV =
+            DAG.getNode(IsCZERO_NEZ ? RISCVISD::CZERO_NEZ : RISCVISD::CZERO_EQZ,
+                        DL, VT, XorOp, CondV);
+        return DAG.getNode(ISD::XOR, DL, VT, CMOV, ConstVal);
+      }
+      // Efficient only if the constant and its negation fit into `ADDI`
+      // Prefer Add/Sub over Xor since can be compressed for small immediates
+      if (isInt<12>(RawConstVal)) {
+        SDValue SubOp = DAG.getNode(ISD::SUB, DL, VT, RegV, ConstVal);
+        SDValue CMOV =
+            DAG.getNode(IsCZERO_NEZ ? RISCVISD::CZERO_NEZ : RISCVISD::CZERO_EQZ,
+                        DL, VT, SubOp, CondV);
+        return DAG.getNode(ISD::ADD, DL, VT, CMOV, ConstVal);
+      }
+    }
+
     // (select c, t, f) -> (or (czero_eqz t, c), (czero_nez f, c))
     // Unless we have the short forward branch optimization.
     if (!Subtarget.hasConditionalMoveFusion())

llvm/test/CodeGen/RISCV/short-forward-branch-opt.ll

@@ -173,14 +173,21 @@ define signext i32 @test6(i32 signext %x, i32 signext %z) {
 ; NOSFB-NEXT:    or a0, a0, a1
 ; NOSFB-NEXT:    ret
 ;
-; SFB-LABEL: test6:
-; SFB:       # %bb.0:
-; SFB-NEXT:    li a2, -1
-; SFB-NEXT:    beqz a1, .LBB5_2
-; SFB-NEXT:  # %bb.1:
-; SFB-NEXT:    mv a0, a2
-; SFB-NEXT:  .LBB5_2:
-; SFB-NEXT:    ret
+; NOZICOND-LABEL: test6:
+; NOZICOND:       # %bb.0:
+; NOZICOND-NEXT:    li a2, -1
+; NOZICOND-NEXT:    beqz a1, .LBB5_2
+; NOZICOND-NEXT:  # %bb.1:
+; NOZICOND-NEXT:    mv a0, a2
+; NOZICOND-NEXT:  .LBB5_2:
+; NOZICOND-NEXT:    ret
+;
+; ZICOND-LABEL: test6:
+; ZICOND:       # %bb.0:
+; ZICOND-NEXT:    addi a0, a0, 1
+; ZICOND-NEXT:    czero.nez a0, a0, a1
+; ZICOND-NEXT:    addi a0, a0, -1
+; ZICOND-NEXT:    ret
   %c = icmp eq i32 %z, 0
   %b = select i1 %c, i32 %x, i32 -1
   ret i32 %b
@@ -195,14 +202,21 @@ define signext i32 @test7(i32 signext %x, i32 signext %z) {
 ; NOSFB-NEXT:    or a0, a0, a1
 ; NOSFB-NEXT:    ret
 ;
-; SFB-LABEL: test7:
-; SFB:       # %bb.0:
-; SFB-NEXT:    li a2, -1
-; SFB-NEXT:    bnez a1, .LBB6_2
-; SFB-NEXT:  # %bb.1:
-; SFB-NEXT:    mv a0, a2
-; SFB-NEXT:  .LBB6_2:
-; SFB-NEXT:    ret
+; NOZICOND-LABEL: test7:
+; NOZICOND:       # %bb.0:
+; NOZICOND-NEXT:    li a2, -1
+; NOZICOND-NEXT:    bnez a1, .LBB6_2
+; NOZICOND-NEXT:  # %bb.1:
+; NOZICOND-NEXT:    mv a0, a2
+; NOZICOND-NEXT:  .LBB6_2:
+; NOZICOND-NEXT:    ret
+;
+; ZICOND-LABEL: test7:
+; ZICOND:       # %bb.0:
+; ZICOND-NEXT:    addi a0, a0, 1
+; ZICOND-NEXT:    czero.eqz a0, a0, a1
+; ZICOND-NEXT:    addi a0, a0, -1
+; ZICOND-NEXT:    ret
   %c = icmp eq i32 %z, 0
   %b = select i1 %c, i32 -1, i32 %x
   ret i32 %b

llvm/test/CodeGen/RISCV/zicond-opts.ll

@@ -104,21 +104,19 @@ define i64 @select_imm_reg(i64 %0, i64 %1) {
 ; RV32ZICOND-LABEL: select_imm_reg:
 ; RV32ZICOND:       # %bb.0:
 ; RV32ZICOND-NEXT:    xori a0, a0, 2
-; RV32ZICOND-NEXT:    or a1, a0, a1
-; RV32ZICOND-NEXT:    li a0, 3
-; RV32ZICOND-NEXT:    czero.eqz a2, a2, a1
-; RV32ZICOND-NEXT:    czero.nez a0, a0, a1
-; RV32ZICOND-NEXT:    or a0, a0, a2
-; RV32ZICOND-NEXT:    czero.eqz a1, a3, a1
+; RV32ZICOND-NEXT:    addi a2, a2, -3
+; RV32ZICOND-NEXT:    or a0, a0, a1
+; RV32ZICOND-NEXT:    czero.eqz a1, a3, a0
+; RV32ZICOND-NEXT:    czero.eqz a0, a2, a0
+; RV32ZICOND-NEXT:    addi a0, a0, 3
 ; RV32ZICOND-NEXT:    ret
 ;
 ; RV64ZICOND-LABEL: select_imm_reg:
 ; RV64ZICOND:       # %bb.0:
 ; RV64ZICOND-NEXT:    addi a0, a0, -2
-; RV64ZICOND-NEXT:    li a2, 3
-; RV64ZICOND-NEXT:    czero.eqz a1, a1, a0
-; RV64ZICOND-NEXT:    czero.nez a0, a2, a0
-; RV64ZICOND-NEXT:    or a0, a0, a1
+; RV64ZICOND-NEXT:    addi a1, a1, -3
+; RV64ZICOND-NEXT:    czero.eqz a0, a1, a0
+; RV64ZICOND-NEXT:    addi a0, a0, 3
 ; RV64ZICOND-NEXT:    ret
   %3 = icmp eq i64 %0, 2
   %4 = select i1 %3, i64 3, i64 %1
@@ -130,23 +128,21 @@ define i64 @select_imm_reg_2048(i64 %0, i64 %1) {
 ; RV32ZICOND-LABEL: select_imm_reg_2048:
 ; RV32ZICOND:       # %bb.0:
 ; RV32ZICOND-NEXT:    xori a0, a0, 2
-; RV32ZICOND-NEXT:    li a4, -2048
+; RV32ZICOND-NEXT:    xori a2, a2, -2048
 ; RV32ZICOND-NEXT:    or a0, a0, a1
 ; RV32ZICOND-NEXT:    seqz a1, a0
-; RV32ZICOND-NEXT:    czero.eqz a2, a2, a0
-; RV32ZICOND-NEXT:    czero.nez a0, a4, a0
-; RV32ZICOND-NEXT:    or a0, a0, a2
+; RV32ZICOND-NEXT:    czero.eqz a0, a2, a0
 ; RV32ZICOND-NEXT:    neg a1, a1
 ; RV32ZICOND-NEXT:    or a1, a1, a3
+; RV32ZICOND-NEXT:    xori a0, a0, -2048
 ; RV32ZICOND-NEXT:    ret
 ;
 ; RV64ZICOND-LABEL: select_imm_reg_2048:
 ; RV64ZICOND:       # %bb.0:
 ; RV64ZICOND-NEXT:    addi a0, a0, -2
-; RV64ZICOND-NEXT:    li a2, -2048
-; RV64ZICOND-NEXT:    czero.eqz a1, a1, a0
-; RV64ZICOND-NEXT:    czero.nez a0, a2, a0
-; RV64ZICOND-NEXT:    or a0, a0, a1
+; RV64ZICOND-NEXT:    xori a1, a1, -2048
+; RV64ZICOND-NEXT:    czero.eqz a0, a1, a0
+; RV64ZICOND-NEXT:    xori a0, a0, -2048
 ; RV64ZICOND-NEXT:    ret
   %3 = icmp eq i64 %0, 2
   %4 = select i1 %3, i64 -2048, i64 %1