5 files changed, 108 insertions, 8 deletions
diff --git a/llvm/lib/Target/RISCV/RISCVExpandPseudoInsts.cpp b/llvm/lib/Target/RISCV/RISCVExpandPseudoInsts.cpp
index cb57c43..d4d9e54 100644
--- a/llvm/lib/Target/RISCV/RISCVExpandPseudoInsts.cpp
+++ b/llvm/lib/Target/RISCV/RISCVExpandPseudoInsts.cpp
@@ -193,7 +193,7 @@ bool RISCVExpandPseudo::expandCCOp(MachineBasicBlock &MBB,
   // we need to invert the branch condition to jump over TrueBB when the
   // condition is false.
   auto CC = static_cast<RISCVCC::CondCode>(MI.getOperand(3).getImm());
-  CC = RISCVCC::getOppositeBranchCondition(CC);
+  CC = RISCVCC::getInverseBranchCondition(CC);
 
   // Insert branch instruction.
   BuildMI(MBB, MBBI, DL, TII->get(RISCVCC::getBrCond(CC)))
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp b/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
index 56db09a..70b6c7e 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
@@ -1023,6 +1023,37 @@ static void parseCondBranch(MachineInstr &LastInst, MachineBasicBlock *&Target,
   Cond.push_back(LastInst.getOperand(1));
 }
 
+static unsigned getInverseXqcicmOpcode(unsigned Opcode) {
+  switch (Opcode) {
+  default:
+    llvm_unreachable("Unexpected Opcode");
+  case RISCV::QC_MVEQ:
+    return RISCV::QC_MVNE;
+  case RISCV::QC_MVNE:
+    return RISCV::QC_MVEQ;
+  case RISCV::QC_MVLT:
+    return RISCV::QC_MVGE;
+  case RISCV::QC_MVGE:
+    return RISCV::QC_MVLT;
+  case RISCV::QC_MVLTU:
+    return RISCV::QC_MVGEU;
+  case RISCV::QC_MVGEU:
+    return RISCV::QC_MVLTU;
+  case RISCV::QC_MVEQI:
+    return RISCV::QC_MVNEI;
+  case RISCV::QC_MVNEI:
+    return RISCV::QC_MVEQI;
+  case RISCV::QC_MVLTI:
+    return RISCV::QC_MVGEI;
+  case RISCV::QC_MVGEI:
+    return RISCV::QC_MVLTI;
+  case RISCV::QC_MVLTUI:
+    return RISCV::QC_MVGEUI;
+  case RISCV::QC_MVGEUI:
+    return RISCV::QC_MVLTUI;
+  }
+}
+
 unsigned RISCVCC::getBrCond(RISCVCC::CondCode CC, unsigned SelectOpc) {
   switch (SelectOpc) {
   default:
@@ -1134,7 +1165,7 @@ unsigned RISCVCC::getBrCond(RISCVCC::CondCode CC, unsigned SelectOpc) {
   }
 }
 
-RISCVCC::CondCode RISCVCC::getOppositeBranchCondition(RISCVCC::CondCode CC) {
+RISCVCC::CondCode RISCVCC::getInverseBranchCondition(RISCVCC::CondCode CC) {
   switch (CC) {
   default:
     llvm_unreachable("Unrecognized conditional branch");
@@ -1554,7 +1585,7 @@ bool RISCVInstrInfo::optimizeCondBranch(MachineInstr &MI) const {
     return Register();
   };
 
-  unsigned NewOpc = RISCVCC::getBrCond(getOppositeBranchCondition(CC));
+  unsigned NewOpc = RISCVCC::getBrCond(getInverseBranchCondition(CC));
 
   // Might be case 1.
   // Don't change 0 to 1 since we can use x0.
@@ -1801,7 +1832,7 @@ RISCVInstrInfo::optimizeSelect(MachineInstr &MI,
   // Add condition code, inverting if necessary.
   auto CC = static_cast<RISCVCC::CondCode>(MI.getOperand(3).getImm());
   if (Invert)
-    CC = RISCVCC::getOppositeBranchCondition(CC);
+    CC = RISCVCC::getInverseBranchCondition(CC);
   NewMI.addImm(CC);
 
   // Copy the false register.
@@ -3762,6 +3793,19 @@ bool RISCVInstrInfo::findCommutedOpIndices(const MachineInstr &MI,
       return false;
     // Operands 1 and 2 are commutable, if we switch the opcode.
     return fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 1, 2);
+  case RISCV::QC_MVEQ:
+  case RISCV::QC_MVNE:
+  case RISCV::QC_MVLT:
+  case RISCV::QC_MVGE:
+  case RISCV::QC_MVLTU:
+  case RISCV::QC_MVGEU:
+  case RISCV::QC_MVEQI:
+  case RISCV::QC_MVNEI:
+  case RISCV::QC_MVLTI:
+  case RISCV::QC_MVGEI:
+  case RISCV::QC_MVLTUI:
+  case RISCV::QC_MVGEUI:
+    return fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 1, 4);
   case RISCV::TH_MULA:
   case RISCV::TH_MULAW:
   case RISCV::TH_MULAH:
@@ -3974,11 +4018,28 @@ MachineInstr *RISCVInstrInfo::commuteInstructionImpl(MachineInstr &MI,
     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, false, OpIdx1,
                                                    OpIdx2);
   }
+  case RISCV::QC_MVEQ:
+  case RISCV::QC_MVNE:
+  case RISCV::QC_MVLT:
+  case RISCV::QC_MVGE:
+  case RISCV::QC_MVLTU:
+  case RISCV::QC_MVGEU:
+  case RISCV::QC_MVEQI:
+  case RISCV::QC_MVNEI:
+  case RISCV::QC_MVLTI:
+  case RISCV::QC_MVGEI:
+  case RISCV::QC_MVLTUI:
+  case RISCV::QC_MVGEUI: {
+    auto &WorkingMI = cloneIfNew(MI);
+    WorkingMI.setDesc(get(getInverseXqcicmOpcode(MI.getOpcode())));
+    return TargetInstrInfo::commuteInstructionImpl(WorkingMI, false, OpIdx1,
+                                                   OpIdx2);
+  }
   case RISCV::PseudoCCMOVGPRNoX0:
   case RISCV::PseudoCCMOVGPR: {
     // CCMOV can be commuted by inverting the condition.
     auto CC = static_cast<RISCVCC::CondCode>(MI.getOperand(3).getImm());
-    CC = RISCVCC::getOppositeBranchCondition(CC);
+    CC = RISCVCC::getInverseBranchCondition(CC);
     auto &WorkingMI = cloneIfNew(MI);
     WorkingMI.getOperand(3).setImm(CC);
     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI*/ false,
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfo.h b/llvm/lib/Target/RISCV/RISCVInstrInfo.h
index 2bc499b..42a0c4c 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfo.h
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfo.h
@@ -44,7 +44,7 @@ enum CondCode {
   COND_INVALID
 };
 
-CondCode getOppositeBranchCondition(CondCode);
+CondCode getInverseBranchCondition(CondCode);
 unsigned getBrCond(CondCode CC, unsigned SelectOpc = 0);
 
 } // end of namespace RISCVCC
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoXqci.td b/llvm/lib/Target/RISCV/RISCVInstrInfoXqci.td
index 13b02d1..ff4a040 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoXqci.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoXqci.td
@@ -604,7 +604,7 @@ class QCILICC<bits<3> funct3, bits<2> funct2, DAGOperand InTyRs2, string opcodes
   let Inst{31-25} = {simm, funct2};
 }
 
-let hasSideEffects = 0, mayLoad = 0, mayStore = 0 in
+let hasSideEffects = 0, mayLoad = 0, mayStore = 0, isCommutable = 1 in
 class QCIMVCC<bits<3> funct3, string opcodestr>
     : RVInstR4<0b00, funct3, OPC_CUSTOM_2, (outs GPRNoX0:$rd_wb),
                (ins GPRNoX0:$rd, GPRNoX0:$rs1, GPRNoX0:$rs2, GPRNoX0:$rs3),
@@ -612,7 +612,7 @@ class QCIMVCC<bits<3> funct3, string opcodestr>
   let Constraints = "$rd = $rd_wb";
 }
 
-let hasSideEffects = 0, mayLoad = 0, mayStore = 0 in
+let hasSideEffects = 0, mayLoad = 0, mayStore = 0, isCommutable = 1 in
 class QCIMVCCI<bits<3> funct3, string opcodestr, DAGOperand immType>
     : RVInstR4<0b10, funct3, OPC_CUSTOM_2, (outs GPRNoX0:$rd_wb),
                (ins GPRNoX0:$rd, GPRNoX0:$rs1, immType:$imm, GPRNoX0:$rs3),
diff --git a/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp b/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
index 563f3bb..d4124ae 100644
--- a/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
+++ b/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
@@ -167,6 +167,42 @@ static bool canUseShiftPair(Instruction *Inst, const APInt &Imm) {
   return false;
 }
 
+// If this is i64 AND is part of (X & -(1 << C1) & 0xffffffff) == C2 << C1),
+// DAGCombiner can convert this to (sraiw X, C1) == sext(C2) for RV64. On RV32,
+// the type will be split so only the lower 32 bits need to be compared using
+// (srai/srli X, C) == C2.
+static bool canUseShiftCmp(Instruction *Inst, const APInt &Imm) {
+  if (!Inst->hasOneUse())
+    return false;
+
+  // Look for equality comparison.
+  auto *Cmp = dyn_cast<ICmpInst>(*Inst->user_begin());
+  if (!Cmp || !Cmp->isEquality())
+    return false;
+
+  // Right hand side of comparison should be a constant.
+  auto *C = dyn_cast<ConstantInt>(Cmp->getOperand(1));
+  if (!C)
+    return false;
+
+  uint64_t Mask = Imm.getZExtValue();
+
+  // Mask should be of the form -(1 << C) in the lower 32 bits.
+  if (!isUInt<32>(Mask) || !isPowerOf2_32(-uint32_t(Mask)))
+    return false;
+
+  // Comparison constant should be a subset of Mask.
+  uint64_t CmpC = C->getZExtValue();
+  if ((CmpC & Mask) != CmpC)
+    return false;
+
+  // We'll need to sign extend the comparison constant and shift it right. Make
+  // sure the new constant can use addi/xori+seqz/snez.
+  unsigned ShiftBits = llvm::countr_zero(Mask);
+  int64_t NewCmpC = SignExtend64<32>(CmpC) >> ShiftBits;
+  return NewCmpC >= -2048 && NewCmpC <= 2048;
+}
+
 InstructionCost RISCVTTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx,
                                                 const APInt &Imm, Type *Ty,
                                                 TTI::TargetCostKind CostKind,
@@ -224,6 +260,9 @@ InstructionCost RISCVTTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx,
     if (Inst && Idx == 1 && Imm.getBitWidth() <= ST->getXLen() &&
         canUseShiftPair(Inst, Imm))
       return TTI::TCC_Free;
+    if (Inst && Idx == 1 && Imm.getBitWidth() == 64 &&
+        canUseShiftCmp(Inst, Imm))
+      return TTI::TCC_Free;
     Takes12BitImm = true;
     break;
   case Instruction::Add: