[RISCV] Select (add/or C, x) -> (add.uw C|0xffffffff00000000, x) (#143375)

pfusik · web-flow · commit 39a7664fc10d · 2025-06-10T13:28:49.000+02:00
Emits fewer instructions for certain constants.
diff --git a/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp b/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
@@ -3307,6 +3307,50 @@ bool RISCVDAGToDAGISel::selectSHXADD_UWOp(SDValue N, unsigned ShAmt,
   return false;
 }
 
+bool RISCVDAGToDAGISel::orIsAdd(const SDNode *N) const {
+  if (N->getFlags().hasDisjoint())
+    return true;
+  KnownBits Known0 = CurDAG->computeKnownBits(N->getOperand(0), 0);
+  KnownBits Known1 = CurDAG->computeKnownBits(N->getOperand(1), 0);
+  return KnownBits::haveNoCommonBitsSet(Known0, Known1);
+}
+
+bool RISCVDAGToDAGISel::selectImm64IfCheaper(int64_t Imm, int64_t OrigImm,
+                                             SDValue N, SDValue &Val) {
+  int OrigCost = RISCVMatInt::getIntMatCost(APInt(64, OrigImm), 64, *Subtarget,
+                                            /*CompressionCost=*/true);
+  int Cost = RISCVMatInt::getIntMatCost(APInt(64, Imm), 64, *Subtarget,
+                                        /*CompressionCost=*/true);
+  if (OrigCost <= Cost)
+    return false;
+
+  Val = selectImm(CurDAG, SDLoc(N), N->getSimpleValueType(0), Imm, *Subtarget);
+  return true;
+}
+
+bool RISCVDAGToDAGISel::selectZExtImm32(SDValue N, SDValue &Val) {
+  if (!isa<ConstantSDNode>(N))
+    return false;
+  int64_t Imm = cast<ConstantSDNode>(N)->getSExtValue();
+  if ((Imm >> 31) != 1)
+    return false;
+
+  for (const SDNode *U : N->users()) {
+    switch (U->getOpcode()) {
+    case ISD::ADD:
+      break;
+    case ISD::OR:
+      if (orIsAdd(U))
+        break;
+      return false;
+    default:
+      return false;
+    }
+  }
+
+  return selectImm64IfCheaper(0xffffffff00000000 | Imm, Imm, N, Val);
+}
+
 bool RISCVDAGToDAGISel::selectNegImm(SDValue N, SDValue &Val) {
   if (!isa<ConstantSDNode>(N))
     return false;
@@ -3330,15 +3374,7 @@ bool RISCVDAGToDAGISel::selectNegImm(SDValue N, SDValue &Val) {
     }
   }
 
-  int OrigImmCost = RISCVMatInt::getIntMatCost(APInt(64, Imm), 64, *Subtarget,
-                                               /*CompressionCost=*/true);
-  int NegImmCost = RISCVMatInt::getIntMatCost(APInt(64, -Imm), 64, *Subtarget,
-                                              /*CompressionCost=*/true);
-  if (OrigImmCost <= NegImmCost)
-    return false;
-
-  Val = selectImm(CurDAG, SDLoc(N), N->getSimpleValueType(0), -Imm, *Subtarget);
-  return true;
+  return selectImm64IfCheaper(-Imm, Imm, N, Val);
 }
 
 bool RISCVDAGToDAGISel::selectInvLogicImm(SDValue N, SDValue &Val) {
@@ -3373,19 +3409,15 @@ bool RISCVDAGToDAGISel::selectInvLogicImm(SDValue N, SDValue &Val) {
     }
   }
 
-  // For 64-bit constants, the instruction sequences get complex,
-  // so we select inverted only if it's cheaper.
-  if (!isInt<32>(Imm)) {
-    int OrigImmCost = RISCVMatInt::getIntMatCost(APInt(64, Imm), 64, *Subtarget,
-                                                 /*CompressionCost=*/true);
-    int NegImmCost = RISCVMatInt::getIntMatCost(APInt(64, ~Imm), 64, *Subtarget,
-                                                /*CompressionCost=*/true);
-    if (OrigImmCost <= NegImmCost)
-      return false;
+  if (isInt<32>(Imm)) {
+    Val =
+        selectImm(CurDAG, SDLoc(N), N->getSimpleValueType(0), ~Imm, *Subtarget);
+    return true;
   }
 
-  Val = selectImm(CurDAG, SDLoc(N), N->getSimpleValueType(0), ~Imm, *Subtarget);
-  return true;
+  // For 64-bit constants, the instruction sequences get complex,
+  // so we select inverted only if it's cheaper.
+  return selectImm64IfCheaper(~Imm, Imm, N, Val);
 }
 
 static bool vectorPseudoHasAllNBitUsers(SDNode *User, unsigned UserOpNo,
diff --git a/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.h b/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.h
@@ -118,9 +118,11 @@ class RISCVDAGToDAGISel : public SelectionDAGISel {
     return selectSHXADD_UWOp(N, ShAmt, Val);
   }
 
+  bool selectZExtImm32(SDValue N, SDValue &Val);
   bool selectNegImm(SDValue N, SDValue &Val);
   bool selectInvLogicImm(SDValue N, SDValue &Val);
 
+  bool orIsAdd(const SDNode *Node) const;
   bool hasAllNBitUsers(SDNode *Node, unsigned Bits,
                        const unsigned Depth = 0) const;
   bool hasAllBUsers(SDNode *Node) const { return hasAllNBitUsers(Node, 8); }
@@ -199,6 +201,8 @@ class RISCVDAGToDAGISel : public SelectionDAGISel {
   bool doPeepholeMergeVVMFold();
   bool doPeepholeNoRegPassThru();
   bool performCombineVMergeAndVOps(SDNode *N);
+  bool selectImm64IfCheaper(int64_t Imm, int64_t OrigImm, SDValue N,
+                            SDValue &Val);
 };
 
 class RISCVDAGToDAGISelLegacy : public SelectionDAGISelLegacy {
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfo.td b/llvm/lib/Target/RISCV/RISCVInstrInfo.td
@@ -1445,11 +1445,7 @@ def : PatGprUimmLog2XLen<sra, SRAI>;
 // Select 'or' as ADDI if the immediate bits are known to be 0 in $rs1. This
 // can improve compressibility.
 def or_is_add : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),[{
-  if (N->getFlags().hasDisjoint())
-    return true;
-  KnownBits Known0 = CurDAG->computeKnownBits(N->getOperand(0), 0);
-  KnownBits Known1 = CurDAG->computeKnownBits(N->getOperand(1), 0);
-  return KnownBits::haveNoCommonBitsSet(Known0, Known1);
+  return orIsAdd(N);
 }]>;
 def : PatGprSimm12<or_is_add, ADDI>;
 
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoVVLPatterns.td b/llvm/lib/Target/RISCV/RISCVInstrInfoVVLPatterns.td
@@ -653,11 +653,7 @@ let HasOneUse = 1 in {
                                                node:$E),
                                           (riscv_or_vl node:$A, node:$B, node:$C,
                                                        node:$D, node:$E), [{
-    if (N->getFlags().hasDisjoint())
-      return true;
-    KnownBits Known0 = CurDAG->computeKnownBits(N->getOperand(0), 0);
-    KnownBits Known1 = CurDAG->computeKnownBits(N->getOperand(1), 0);
-    return KnownBits::haveNoCommonBitsSet(Known0, Known1);
+    return orIsAdd(N);
   }]>;
   def riscv_sub_vl_oneuse : PatFrag<(ops node:$A, node:$B, node:$C, node:$D,
                                          node:$E),
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoZb.td b/llvm/lib/Target/RISCV/RISCVInstrInfoZb.td
@@ -704,9 +704,13 @@ def : CSImm12MulBy4Pat<SH2ADD>;
 def : CSImm12MulBy8Pat<SH3ADD>;
 } // Predicates = [HasStdExtZba]
 
+def zExtImm32 : ComplexPattern<i64, 1, "selectZExtImm32", [], [], 0>;
+
 multiclass ADD_UWPat<Instruction add_uw> {
   def : Pat<(i64 (add_like_non_imm12 (and GPR:$rs1, 0xFFFFFFFF), GPR:$rs2)),
             (add_uw GPR:$rs1, GPR:$rs2)>;
+  def : Pat<(i64 (add_like zExtImm32:$rs1, GPR:$rs2)),
+            (add_uw zExtImm32:$rs1, GPR:$rs2)>;
   def : Pat<(i64 (and GPR:$rs, 0xFFFFFFFF)), (add_uw GPR:$rs, (XLenVT X0))>;
 }
 
diff --git a/llvm/test/CodeGen/RISCV/rv64zba.ll b/llvm/test/CodeGen/RISCV/rv64zba.ll
@@ -4260,10 +4260,11 @@ define i64 @add_u32simm32_zextw(i64 %x) nounwind {
 ;
 ; RV64XANDESPERF-LABEL: add_u32simm32_zextw:
 ; RV64XANDESPERF:       # %bb.0: # %entry
+; RV64XANDESPERF-NEXT:    li a1, -2
+; RV64XANDESPERF-NEXT:    nds.lea.b.ze a0, a0, a1
 ; RV64XANDESPERF-NEXT:    li a1, 1
 ; RV64XANDESPERF-NEXT:    slli a1, a1, 32
 ; RV64XANDESPERF-NEXT:    addi a1, a1, -2
-; RV64XANDESPERF-NEXT:    add a0, a0, a1
 ; RV64XANDESPERF-NEXT:    addi a1, a1, 1
 ; RV64XANDESPERF-NEXT:    and a0, a0, a1
 ; RV64XANDESPERF-NEXT:    ret
@@ -4413,12 +4414,24 @@ define ptr @udiv1280_gep(ptr %p, i16 zeroext %i) {
 }
 
 define i64 @adduw_m1(i64 %x) {
-; CHECK-LABEL: adduw_m1:
-; CHECK:       # %bb.0:
-; CHECK-NEXT:    li a1, -1
-; CHECK-NEXT:    srli a1, a1, 32
-; CHECK-NEXT:    add a0, a0, a1
-; CHECK-NEXT:    ret
+; RV64I-LABEL: adduw_m1:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    li a1, -1
+; RV64I-NEXT:    srli a1, a1, 32
+; RV64I-NEXT:    add a0, a0, a1
+; RV64I-NEXT:    ret
+;
+; RV64ZBA-LABEL: adduw_m1:
+; RV64ZBA:       # %bb.0:
+; RV64ZBA-NEXT:    li a1, -1
+; RV64ZBA-NEXT:    add.uw a0, a1, a0
+; RV64ZBA-NEXT:    ret
+;
+; RV64XANDESPERF-LABEL: adduw_m1:
+; RV64XANDESPERF:       # %bb.0:
+; RV64XANDESPERF-NEXT:    li a1, -1
+; RV64XANDESPERF-NEXT:    nds.lea.b.ze a0, a0, a1
+; RV64XANDESPERF-NEXT:    ret
   %a = add i64 %x, 4294967295
   ret i64 %a
 }
@@ -4435,28 +4448,37 @@ define i64 @adduw_m3(i64 %x) {
 ; RV64ZBA-LABEL: adduw_m3:
 ; RV64ZBA:       # %bb.0:
 ; RV64ZBA-NEXT:    li a1, -3
-; RV64ZBA-NEXT:    zext.w a1, a1
-; RV64ZBA-NEXT:    add a0, a0, a1
+; RV64ZBA-NEXT:    add.uw a0, a1, a0
 ; RV64ZBA-NEXT:    ret
 ;
 ; RV64XANDESPERF-LABEL: adduw_m3:
 ; RV64XANDESPERF:       # %bb.0:
-; RV64XANDESPERF-NEXT:    li a1, 1
-; RV64XANDESPERF-NEXT:    slli a1, a1, 32
-; RV64XANDESPERF-NEXT:    addi a1, a1, -3
-; RV64XANDESPERF-NEXT:    add a0, a0, a1
+; RV64XANDESPERF-NEXT:    li a1, -3
+; RV64XANDESPERF-NEXT:    nds.lea.b.ze a0, a0, a1
 ; RV64XANDESPERF-NEXT:    ret
   %a = add i64 %x, 4294967293
   ret i64 %a
 }
 
 define i64 @adduw_3shl30(i64 %x) {
-; CHECK-LABEL: adduw_3shl30:
-; CHECK:       # %bb.0:
-; CHECK-NEXT:    li a1, 3
-; CHECK-NEXT:    slli a1, a1, 30
-; CHECK-NEXT:    add a0, a0, a1
-; CHECK-NEXT:    ret
+; RV64I-LABEL: adduw_3shl30:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    li a1, 3
+; RV64I-NEXT:    slli a1, a1, 30
+; RV64I-NEXT:    add a0, a0, a1
+; RV64I-NEXT:    ret
+;
+; RV64ZBA-LABEL: adduw_3shl30:
+; RV64ZBA:       # %bb.0:
+; RV64ZBA-NEXT:    lui a1, 786432
+; RV64ZBA-NEXT:    add.uw a0, a1, a0
+; RV64ZBA-NEXT:    ret
+;
+; RV64XANDESPERF-LABEL: adduw_3shl30:
+; RV64XANDESPERF:       # %bb.0:
+; RV64XANDESPERF-NEXT:    lui a1, 786432
+; RV64XANDESPERF-NEXT:    nds.lea.b.ze a0, a0, a1
+; RV64XANDESPERF-NEXT:    ret
   %a = add i64 %x, 3221225472
   ret i64 %a
 }
@@ -4475,19 +4497,16 @@ define i64 @adduw_m3_multiuse(i64 %x, i64 %y) {
 ; RV64ZBA-LABEL: adduw_m3_multiuse:
 ; RV64ZBA:       # %bb.0:
 ; RV64ZBA-NEXT:    li a2, -3
-; RV64ZBA-NEXT:    zext.w a2, a2
-; RV64ZBA-NEXT:    add a0, a0, a2
-; RV64ZBA-NEXT:    add a1, a1, a2
+; RV64ZBA-NEXT:    add.uw a0, a2, a0
+; RV64ZBA-NEXT:    add.uw a1, a2, a1
 ; RV64ZBA-NEXT:    or a0, a0, a1
 ; RV64ZBA-NEXT:    ret
 ;
 ; RV64XANDESPERF-LABEL: adduw_m3_multiuse:
 ; RV64XANDESPERF:       # %bb.0:
-; RV64XANDESPERF-NEXT:    li a2, 1
-; RV64XANDESPERF-NEXT:    slli a2, a2, 32
-; RV64XANDESPERF-NEXT:    addi a2, a2, -3
-; RV64XANDESPERF-NEXT:    add a0, a0, a2
-; RV64XANDESPERF-NEXT:    add a1, a1, a2
+; RV64XANDESPERF-NEXT:    li a2, -3
+; RV64XANDESPERF-NEXT:    nds.lea.b.ze a0, a0, a2
+; RV64XANDESPERF-NEXT:    nds.lea.b.ze a1, a1, a2
 ; RV64XANDESPERF-NEXT:    or a0, a0, a1
 ; RV64XANDESPERF-NEXT:    ret
   %a = add i64 %x, 4294967293
@@ -4532,13 +4551,27 @@ define i64 @add_or_m3(i64 %x) {
 }
 
 define i64 @append_32ones(i64 %x) {
-; CHECK-LABEL: append_32ones:
-; CHECK:       # %bb.0:
-; CHECK-NEXT:    slli a0, a0, 32
-; CHECK-NEXT:    li a1, -1
-; CHECK-NEXT:    srli a1, a1, 32
-; CHECK-NEXT:    or a0, a0, a1
-; CHECK-NEXT:    ret
+; RV64I-LABEL: append_32ones:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    slli a0, a0, 32
+; RV64I-NEXT:    li a1, -1
+; RV64I-NEXT:    srli a1, a1, 32
+; RV64I-NEXT:    or a0, a0, a1
+; RV64I-NEXT:    ret
+;
+; RV64ZBA-LABEL: append_32ones:
+; RV64ZBA:       # %bb.0:
+; RV64ZBA-NEXT:    slli a0, a0, 32
+; RV64ZBA-NEXT:    li a1, -1
+; RV64ZBA-NEXT:    add.uw a0, a1, a0
+; RV64ZBA-NEXT:    ret
+;
+; RV64XANDESPERF-LABEL: append_32ones:
+; RV64XANDESPERF:       # %bb.0:
+; RV64XANDESPERF-NEXT:    slli a0, a0, 32
+; RV64XANDESPERF-NEXT:    li a1, -1
+; RV64XANDESPERF-NEXT:    nds.lea.b.ze a0, a0, a1
+; RV64XANDESPERF-NEXT:    ret
   %s = shl i64 %x, 32
   %o = or i64 %s, 4294967295
   ret i64 %o
