[llvm] [RISCV] Introduce VLOptimizer pass (PR #108640)
Philip Reames via llvm-commits
llvm-commits at lists.llvm.org
Tue Sep 17 12:29:27 PDT 2024
================
@@ -0,0 +1,1589 @@
+//===-------------- RISCVVLOptimizer.cpp - VL Optimizer -------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===---------------------------------------------------------------------===//
+//
+// This pass reduces the VL where possible at the MI level, before VSETVLI
+// instructions are inserted.
+//
+// The purpose of this optimization is to make the VL argument, for instructions
+// that have a VL argument, as small as possible. This is implemented by
+// visiting each instruction in reverse order and checking that if it has a VL
+// argument, whether the VL can be reduced.
+//
+//===---------------------------------------------------------------------===//
+
+#include "RISCV.h"
+#include "RISCVMachineFunctionInfo.h"
+#include "RISCVSubtarget.h"
+#include "llvm/ADT/SetVector.h"
+#include "llvm/CodeGen/MachineDominators.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/InitializePasses.h"
+
+#include <algorithm>
+
+using namespace llvm;
+
+#define DEBUG_TYPE "riscv-vl-optimizer"
+#define PASS_NAME "RISC-V VL Optimizer"
+
+namespace {
+
+class RISCVVLOptimizer : public MachineFunctionPass {
+ const MachineRegisterInfo *MRI;
+ const MachineDominatorTree *MDT;
+
+public:
+ static char ID;
+
+ RISCVVLOptimizer() : MachineFunctionPass(ID) {}
+
+ bool runOnMachineFunction(MachineFunction &MF) override;
+
+ void getAnalysisUsage(AnalysisUsage &AU) const override {
+ AU.setPreservesCFG();
+ AU.addRequired<MachineDominatorTreeWrapperPass>();
+ MachineFunctionPass::getAnalysisUsage(AU);
+ }
+
+ StringRef getPassName() const override { return PASS_NAME; }
+
+private:
+ bool tryReduceVL(MachineInstr &MI);
+ bool isCandidate(const MachineInstr &MI) const;
+};
+
+} // end anonymous namespace
+
+char RISCVVLOptimizer::ID = 0;
+INITIALIZE_PASS_BEGIN(RISCVVLOptimizer, DEBUG_TYPE, PASS_NAME, false, false)
+INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass)
+INITIALIZE_PASS_END(RISCVVLOptimizer, DEBUG_TYPE, PASS_NAME, false, false)
+
+FunctionPass *llvm::createRISCVVLOptimizerPass() {
+ return new RISCVVLOptimizer();
+}
+
+/// Return true if R is a physical or virtual vector register, false otherwise.
+static bool isVectorRegClass(Register R, const MachineRegisterInfo *MRI) {
+ if (R.isPhysical())
+ return RISCV::VRRegClass.contains(R);
+ const TargetRegisterClass *RC = MRI->getRegClass(R);
+ return RISCVRI::isVRegClass(RC->TSFlags);
+}
+
+/// Represents the EMUL and EEW of a MachineOperand.
+struct OperandInfo {
+ enum class State {
+ Unknown,
+ Known,
+ } S;
+
+ // Represent as 1,2,4,8, ... and fractional indicator. This is because
+ // EMUL can take on values that don't map to RISCVII::VLMUL values exactly.
+ // For example, a mask operand can have an EMUL less than MF8.
+ std::pair<unsigned, bool> EMUL;
+
+ unsigned Log2EEW;
+
+ OperandInfo(RISCVII::VLMUL EMUL, unsigned Log2EEW)
+ : S(State::Known), EMUL(RISCVVType::decodeVLMUL(EMUL)), Log2EEW(Log2EEW) {
+ }
+
+ OperandInfo(std::pair<unsigned, bool> EMUL, unsigned Log2EEW)
+ : S(State::Known), EMUL(EMUL), Log2EEW(Log2EEW) {}
+
+ OperandInfo(State S) : S(S) {
+ assert(S != State::Known &&
+ "This constructor may only be used to construct "
+ "an Unknown OperandInfo");
+ }
+
+ bool isUnknown() const { return S == State::Unknown; }
+ bool isKnown() const { return S == State::Known; }
+
+ static bool EMULAndEEWAreEqual(const OperandInfo &A, const OperandInfo &B) {
+ assert(A.isKnown() && B.isKnown() && "Both operands must be known");
+ return A.Log2EEW == B.Log2EEW && A.EMUL.first == B.EMUL.first &&
+ A.EMUL.second == B.EMUL.second;
+ }
+
+ void print(raw_ostream &OS) const {
+ if (isUnknown()) {
+ OS << "Unknown";
+ return;
+ }
+ OS << "EMUL: ";
+ if (EMUL.second)
+ OS << "m";
+ OS << "f" << EMUL.first;
+ OS << ", EEW: " << (1 << Log2EEW);
+ }
+};
+
+static raw_ostream &operator<<(raw_ostream &OS, const OperandInfo &OI) {
+ OI.print(OS);
+ return OS;
+}
+
+namespace llvm {
+namespace RISCVVType {
+/// Return the RISCVII::VLMUL that is two times VLMul.
+/// Precondition: VLMul is not LMUL_RESERVED or LMUL_8.
+static RISCVII::VLMUL twoTimesVLMUL(RISCVII::VLMUL VLMul) {
+ switch (VLMul) {
+ case RISCVII::VLMUL::LMUL_F8:
+ return RISCVII::VLMUL::LMUL_F4;
+ case RISCVII::VLMUL::LMUL_F4:
+ return RISCVII::VLMUL::LMUL_F2;
+ case RISCVII::VLMUL::LMUL_F2:
+ return RISCVII::VLMUL::LMUL_1;
+ case RISCVII::VLMUL::LMUL_1:
+ return RISCVII::VLMUL::LMUL_2;
+ case RISCVII::VLMUL::LMUL_2:
+ return RISCVII::VLMUL::LMUL_4;
+ case RISCVII::VLMUL::LMUL_4:
+ return RISCVII::VLMUL::LMUL_8;
+ case RISCVII::VLMUL::LMUL_8:
+ default:
+ llvm_unreachable("Could not multiply VLMul by 2");
+ }
+}
+
+/// Return the RISCVII::VLMUL that is VLMul / 2.
+/// Precondition: VLMul is not LMUL_RESERVED or LMUL_MF8.
+static RISCVII::VLMUL halfVLMUL(RISCVII::VLMUL VLMul) {
+ switch (VLMul) {
+ case RISCVII::VLMUL::LMUL_F4:
+ return RISCVII::VLMUL::LMUL_F8;
+ case RISCVII::VLMUL::LMUL_F2:
+ return RISCVII::VLMUL::LMUL_F4;
+ case RISCVII::VLMUL::LMUL_1:
+ return RISCVII::VLMUL::LMUL_F2;
+ case RISCVII::VLMUL::LMUL_2:
+ return RISCVII::VLMUL::LMUL_1;
+ case RISCVII::VLMUL::LMUL_4:
+ return RISCVII::VLMUL::LMUL_2;
+ case RISCVII::VLMUL::LMUL_8:
+ return RISCVII::VLMUL::LMUL_4;
+ case RISCVII::VLMUL::LMUL_F8:
+ default:
+ llvm_unreachable("Could not divide VLMul by 2");
+ }
+}
+
+/// Return EMUL = (EEW / SEW) * LMUL where EEW comes from Log2EEW and LMUL and
+/// SEW are from the TSFlags of MI.
+static std::pair<unsigned, bool>
+getEMULEqualsEEWDivSEWTimesLMUL(unsigned Log2EEW, const MachineInstr &MI) {
+ RISCVII::VLMUL MIVLMUL = RISCVII::getLMul(MI.getDesc().TSFlags);
+ auto [MILMUL, MILMULIsFractional] = RISCVVType::decodeVLMUL(MIVLMUL);
+ unsigned MILog2SEW =
+ MI.getOperand(RISCVII::getSEWOpNum(MI.getDesc())).getImm();
+ unsigned MISEW = 1 << MILog2SEW;
+
+ unsigned EEW = 1 << Log2EEW;
+ // Calculate (EEW/SEW)*LMUL preserving fractions less than 1. Use GCD
+ // to put fraction in simplest form.
+ unsigned Num = EEW, Denom = MISEW;
+ int GCD = MILMULIsFractional ? std::gcd(Num, Denom * MILMUL)
+ : std::gcd(Num * MILMUL, Denom);
+ Num = MILMULIsFractional ? Num / GCD : Num * MILMUL / GCD;
+ Denom = MILMULIsFractional ? Denom * MILMUL / GCD : Denom / GCD;
+ return std::make_pair(Num > Denom ? Num : Denom, Denom > Num);
+}
+} // end namespace RISCVVType
+} // end namespace llvm
+
+static bool isOpN(const MachineOperand &MO, unsigned OpN) {
+ const MachineInstr &MI = *MO.getParent();
+ bool HasPassthru = RISCVII::isFirstDefTiedToFirstUse(MI.getDesc());
+
+ if (HasPassthru)
+ return MO.getOperandNo() == OpN + 1;
+
+ return MO.getOperandNo() == OpN;
+}
+
+/// An index segment load or store operand has the form v.*seg<nf>ei<eeew>.v.
+/// Data has EEW=SEW, EMUL=LMUL. Index has EEW=<eew>, EMUL=(EEW/SEW)*LMUL. LMUL
+/// and SEW comes from TSFlags of MI.
+static OperandInfo getIndexSegmentLoadStoreOperandInfo(unsigned Log2EEW,
+ const MachineInstr &MI,
+ const MachineOperand &MO,
+ bool IsLoad) {
+ // Operand 0 is data register
+ // Data vector register group has EEW=SEW, EMUL=LMUL.
+ if (MO.getOperandNo() == 0) {
+ RISCVII::VLMUL MIVLMul = RISCVII::getLMul(MI.getDesc().TSFlags);
+ unsigned MILog2SEW =
+ MI.getOperand(RISCVII::getSEWOpNum(MI.getDesc())).getImm();
+ return OperandInfo(MIVLMul, MILog2SEW);
+ }
+
+ // Operand 2 is index vector register
+ // v.*seg<nf>ei<eeew>.v
+ // Index vector register group has EEW=<eew>, EMUL=(EEW/SEW)*LMUL.
+ if (isOpN(MO, 2))
+ return OperandInfo(RISCVVType::getEMULEqualsEEWDivSEWTimesLMUL(Log2EEW, MI),
+ Log2EEW);
+
+ llvm_unreachable("Could not get OperandInfo for non-vector register of an "
+ "indexed segment load or store instruction");
+}
+
+/// Dest has EEW=SEW and EMUL=LMUL. Source EEW=SEW/Factor (i.e. F2 => EEW/2).
+/// Source has EMUL=(EEW/SEW)*LMUL. LMUL and SEW comes from TSFlags of MI.
+static OperandInfo getIntegerExtensionOperandInfo(unsigned Factor,
+ const MachineInstr &MI,
+ const MachineOperand &MO) {
+ RISCVII::VLMUL MIVLMul = RISCVII::getLMul(MI.getDesc().TSFlags);
+ unsigned MILog2SEW =
+ MI.getOperand(RISCVII::getSEWOpNum(MI.getDesc())).getImm();
+
+ if (MO.getOperandNo() == 0)
+ return OperandInfo(MIVLMul, MILog2SEW);
+
+ unsigned MISEW = 1 << MILog2SEW;
+ unsigned EEW = MISEW / Factor;
+ unsigned Log2EEW = Log2_32(EEW);
+
+ return OperandInfo(RISCVVType::getEMULEqualsEEWDivSEWTimesLMUL(Log2EEW, MI),
+ Log2EEW);
+}
+
+/// Check whether MO is a mask operand of MI.
+static bool isMaskOperand(const MachineInstr &MI, const MachineOperand &MO,
+ const MachineRegisterInfo *MRI) {
+
+ if (!MO.isReg() || !isVectorRegClass(MO.getReg(), MRI))
+ return false;
+
+ const MCInstrDesc &Desc = MI.getDesc();
+ return Desc.operands()[MO.getOperandNo()].RegClass == RISCV::VMV0RegClassID;
+}
+
+/// Return the OperandInfo for MO, which is an operand of MI.
+static OperandInfo getOperandInfo(const MachineInstr &MI,
+ const MachineOperand &MO,
+ const MachineRegisterInfo *MRI) {
+ const RISCVVPseudosTable::PseudoInfo *RVV =
+ RISCVVPseudosTable::getPseudoInfo(MI.getOpcode());
+ assert(RVV && "Could not find MI in PseudoTable");
+
+ // MI has a VLMUL and SEW associated with it. The RVV specification defines
+ // the LMUL and SEW of each operand and definition in relation to MI.VLMUL and
+ // MI.SEW.
+ RISCVII::VLMUL MIVLMul = RISCVII::getLMul(MI.getDesc().TSFlags);
+ unsigned MILog2SEW =
+ MI.getOperand(RISCVII::getSEWOpNum(MI.getDesc())).getImm();
+
+ const bool HasPassthru = RISCVII::isFirstDefTiedToFirstUse(MI.getDesc());
+
+ // We bail out early for instructions that have passthru with non NoRegister,
+ // which means they are using TU policy. We are not interested in these
+ // since they must preserve the entire register content.
+ if (HasPassthru && MO.getOperandNo() == MI.getNumExplicitDefs() &&
+ (MO.getReg() != RISCV::NoRegister))
+ return OperandInfo(OperandInfo::State::Unknown);
+
+ bool IsMODef = MO.getOperandNo() == 0;
+ bool IsOp1 = isOpN(MO, 1);
+ bool IsOp2 = isOpN(MO, 2);
+ bool IsOp3 = isOpN(MO, 3);
+
+ // All mask operands have EEW=1, EMUL=(EEW/SEW)*LMUL
+ if (isMaskOperand(MI, MO, MRI))
+ return OperandInfo(RISCVVType::getEMULEqualsEEWDivSEWTimesLMUL(0, MI), 0);
+
+ // switch against BaseInstr to reduce number of cases that need to be
+ // considered.
+ switch (RVV->BaseInstr) {
+
+ // 6. Configuration-Setting Instructions
+ // Configuration setting instructions do not read or write vector registers
+ case RISCV::VSETIVLI:
+ case RISCV::VSETVL:
+ case RISCV::VSETVLI:
+ llvm_unreachable("Configuration setting instructions do not read or write "
+ "vector registers");
+
+ // 7. Vector Loads and Stores
+ // 7.4. Vector Unit-Stride Instructions
+ // 7.5. Vector Strided Instructions
+ // 7.7. Unit-stride Fault-Only-First Loads
+ /// Dest EEW encoded in the instruction and EMUL=(EEW/SEW)*LMUL
+ case RISCV::VLE8_V:
+ case RISCV::VSE8_V:
+ case RISCV::VLM_V:
+ case RISCV::VSM_V:
+ case RISCV::VLSE8_V:
+ case RISCV::VSSE8_V:
+ case RISCV::VLE8FF_V:
+ return OperandInfo(RISCVVType::getEMULEqualsEEWDivSEWTimesLMUL(3, MI), 3);
+ case RISCV::VLE16_V:
+ case RISCV::VSE16_V:
+ case RISCV::VLSE16_V:
+ case RISCV::VSSE16_V:
+ case RISCV::VLE16FF_V:
+ return OperandInfo(RISCVVType::getEMULEqualsEEWDivSEWTimesLMUL(4, MI), 4);
+ case RISCV::VLE32_V:
+ case RISCV::VSE32_V:
+ case RISCV::VLSE32_V:
+ case RISCV::VSSE32_V:
+ case RISCV::VLE32FF_V:
+ return OperandInfo(RISCVVType::getEMULEqualsEEWDivSEWTimesLMUL(5, MI), 5);
+ case RISCV::VLE64_V:
+ case RISCV::VSE64_V:
+ case RISCV::VLSE64_V:
+ case RISCV::VSSE64_V:
+ case RISCV::VLE64FF_V:
+ return OperandInfo(RISCVVType::getEMULEqualsEEWDivSEWTimesLMUL(6, MI), 6);
+
+ // 7.6. Vector Indexed Instructions
+ // Data EEW=SEW, EMUL=LMUL. Index EEW=<eew> and EMUL=(EEW/SEW)*LMUL
+ case RISCV::VLUXEI8_V:
+ case RISCV::VLOXEI8_V:
+ case RISCV::VSUXEI8_V:
+ case RISCV::VSOXEI8_V:
+ if (MO.getOperandNo() == 0)
+ return OperandInfo(MIVLMul, MILog2SEW);
+ return OperandInfo(RISCVVType::getEMULEqualsEEWDivSEWTimesLMUL(3, MI), 3);
+ case RISCV::VLUXEI16_V:
+ case RISCV::VLOXEI16_V:
+ case RISCV::VSUXEI16_V:
+ case RISCV::VSOXEI16_V:
+ if (MO.getOperandNo() == 0)
+ return OperandInfo(MIVLMul, MILog2SEW);
+ return OperandInfo(RISCVVType::getEMULEqualsEEWDivSEWTimesLMUL(4, MI), 4);
+ case RISCV::VLUXEI32_V:
+ case RISCV::VLOXEI32_V:
+ case RISCV::VSUXEI32_V:
+ case RISCV::VSOXEI32_V:
+ if (MO.getOperandNo() == 0)
+ return OperandInfo(MIVLMul, MILog2SEW);
+ return OperandInfo(RISCVVType::getEMULEqualsEEWDivSEWTimesLMUL(5, MI), 5);
+ case RISCV::VLUXEI64_V:
+ case RISCV::VLOXEI64_V:
+ case RISCV::VSUXEI64_V:
+ case RISCV::VSOXEI64_V:
+ if (MO.getOperandNo() == 0)
+ return OperandInfo(MIVLMul, MILog2SEW);
+ return OperandInfo(RISCVVType::getEMULEqualsEEWDivSEWTimesLMUL(6, MI), 6);
+
+ // 7.8. Vector Load/Store Segment Instructions
+ // 7.8.1. Vector Unit-Stride Segment Loads and Stores
+ // v.*seg<nf>e<eew>.*
+ // EEW=eew, EMUL=LMUL
+ case RISCV::VLSEG2E8_V:
+ case RISCV::VLSEG2E8FF_V:
+ case RISCV::VLSEG3E8_V:
+ case RISCV::VLSEG3E8FF_V:
+ case RISCV::VLSEG4E8_V:
+ case RISCV::VLSEG4E8FF_V:
+ case RISCV::VLSEG5E8_V:
+ case RISCV::VLSEG5E8FF_V:
+ case RISCV::VLSEG6E8_V:
+ case RISCV::VLSEG6E8FF_V:
+ case RISCV::VLSEG7E8_V:
+ case RISCV::VLSEG7E8FF_V:
+ case RISCV::VLSEG8E8_V:
+ case RISCV::VLSEG8E8FF_V:
+ case RISCV::VSSEG2E8_V:
+ case RISCV::VSSEG3E8_V:
+ case RISCV::VSSEG4E8_V:
+ case RISCV::VSSEG5E8_V:
+ case RISCV::VSSEG6E8_V:
+ case RISCV::VSSEG7E8_V:
+ case RISCV::VSSEG8E8_V:
+ return OperandInfo(MIVLMul, 3);
+ case RISCV::VLSEG2E16_V:
+ case RISCV::VLSEG2E16FF_V:
+ case RISCV::VLSEG3E16_V:
+ case RISCV::VLSEG3E16FF_V:
+ case RISCV::VLSEG4E16_V:
+ case RISCV::VLSEG4E16FF_V:
+ case RISCV::VLSEG5E16_V:
+ case RISCV::VLSEG5E16FF_V:
+ case RISCV::VLSEG6E16_V:
+ case RISCV::VLSEG6E16FF_V:
+ case RISCV::VLSEG7E16_V:
+ case RISCV::VLSEG7E16FF_V:
+ case RISCV::VLSEG8E16_V:
+ case RISCV::VLSEG8E16FF_V:
+ case RISCV::VSSEG2E16_V:
+ case RISCV::VSSEG3E16_V:
+ case RISCV::VSSEG4E16_V:
+ case RISCV::VSSEG5E16_V:
+ case RISCV::VSSEG6E16_V:
+ case RISCV::VSSEG7E16_V:
+ case RISCV::VSSEG8E16_V:
+ return OperandInfo(MIVLMul, 4);
+ case RISCV::VLSEG2E32_V:
+ case RISCV::VLSEG2E32FF_V:
+ case RISCV::VLSEG3E32_V:
+ case RISCV::VLSEG3E32FF_V:
+ case RISCV::VLSEG4E32_V:
+ case RISCV::VLSEG4E32FF_V:
+ case RISCV::VLSEG5E32_V:
+ case RISCV::VLSEG5E32FF_V:
+ case RISCV::VLSEG6E32_V:
+ case RISCV::VLSEG6E32FF_V:
+ case RISCV::VLSEG7E32_V:
+ case RISCV::VLSEG7E32FF_V:
+ case RISCV::VLSEG8E32_V:
+ case RISCV::VLSEG8E32FF_V:
+ case RISCV::VSSEG2E32_V:
+ case RISCV::VSSEG3E32_V:
+ case RISCV::VSSEG4E32_V:
+ case RISCV::VSSEG5E32_V:
+ case RISCV::VSSEG6E32_V:
+ case RISCV::VSSEG7E32_V:
+ case RISCV::VSSEG8E32_V:
+ return OperandInfo(MIVLMul, 5);
+ case RISCV::VLSEG2E64_V:
+ case RISCV::VLSEG2E64FF_V:
+ case RISCV::VLSEG3E64_V:
+ case RISCV::VLSEG3E64FF_V:
+ case RISCV::VLSEG4E64_V:
+ case RISCV::VLSEG4E64FF_V:
+ case RISCV::VLSEG5E64_V:
+ case RISCV::VLSEG5E64FF_V:
+ case RISCV::VLSEG6E64_V:
+ case RISCV::VLSEG6E64FF_V:
+ case RISCV::VLSEG7E64_V:
+ case RISCV::VLSEG7E64FF_V:
+ case RISCV::VLSEG8E64_V:
+ case RISCV::VLSEG8E64FF_V:
+ case RISCV::VSSEG2E64_V:
+ case RISCV::VSSEG3E64_V:
+ case RISCV::VSSEG4E64_V:
+ case RISCV::VSSEG5E64_V:
+ case RISCV::VSSEG6E64_V:
+ case RISCV::VSSEG7E64_V:
+ case RISCV::VSSEG8E64_V:
+ return OperandInfo(MIVLMul, 6);
+
+ // 7.8.2. Vector Strided Segment Loads and Stores
+ case RISCV::VLSSEG2E8_V:
+ case RISCV::VLSSEG3E8_V:
+ case RISCV::VLSSEG4E8_V:
+ case RISCV::VLSSEG5E8_V:
+ case RISCV::VLSSEG6E8_V:
+ case RISCV::VLSSEG7E8_V:
+ case RISCV::VLSSEG8E8_V:
+ case RISCV::VSSSEG2E8_V:
+ case RISCV::VSSSEG3E8_V:
+ case RISCV::VSSSEG4E8_V:
+ case RISCV::VSSSEG5E8_V:
+ case RISCV::VSSSEG6E8_V:
+ case RISCV::VSSSEG7E8_V:
+ case RISCV::VSSSEG8E8_V:
+ return OperandInfo(MIVLMul, 3);
+ case RISCV::VLSSEG2E16_V:
+ case RISCV::VLSSEG3E16_V:
+ case RISCV::VLSSEG4E16_V:
+ case RISCV::VLSSEG5E16_V:
+ case RISCV::VLSSEG6E16_V:
+ case RISCV::VLSSEG7E16_V:
+ case RISCV::VLSSEG8E16_V:
+ case RISCV::VSSSEG2E16_V:
+ case RISCV::VSSSEG3E16_V:
+ case RISCV::VSSSEG4E16_V:
+ case RISCV::VSSSEG5E16_V:
+ case RISCV::VSSSEG6E16_V:
+ case RISCV::VSSSEG7E16_V:
+ case RISCV::VSSSEG8E16_V:
+ return OperandInfo(MIVLMul, 4);
+ case RISCV::VLSSEG2E32_V:
+ case RISCV::VLSSEG3E32_V:
+ case RISCV::VLSSEG4E32_V:
+ case RISCV::VLSSEG5E32_V:
+ case RISCV::VLSSEG6E32_V:
+ case RISCV::VLSSEG7E32_V:
+ case RISCV::VLSSEG8E32_V:
+ case RISCV::VSSSEG2E32_V:
+ case RISCV::VSSSEG3E32_V:
+ case RISCV::VSSSEG4E32_V:
+ case RISCV::VSSSEG5E32_V:
+ case RISCV::VSSSEG6E32_V:
+ case RISCV::VSSSEG7E32_V:
+ case RISCV::VSSSEG8E32_V:
+ return OperandInfo(MIVLMul, 5);
+ case RISCV::VLSSEG2E64_V:
+ case RISCV::VLSSEG3E64_V:
+ case RISCV::VLSSEG4E64_V:
+ case RISCV::VLSSEG5E64_V:
+ case RISCV::VLSSEG6E64_V:
+ case RISCV::VLSSEG7E64_V:
+ case RISCV::VLSSEG8E64_V:
+ case RISCV::VSSSEG2E64_V:
+ case RISCV::VSSSEG3E64_V:
+ case RISCV::VSSSEG4E64_V:
+ case RISCV::VSSSEG5E64_V:
+ case RISCV::VSSSEG6E64_V:
+ case RISCV::VSSSEG7E64_V:
+ case RISCV::VSSSEG8E64_V:
+ return OperandInfo(MIVLMul, 6);
+
+ // 7.8.3. Vector Indexed Segment Loads and Stores
+ case RISCV::VLUXSEG2EI8_V:
+ case RISCV::VLUXSEG3EI8_V:
+ case RISCV::VLUXSEG4EI8_V:
+ case RISCV::VLUXSEG5EI8_V:
+ case RISCV::VLUXSEG6EI8_V:
+ case RISCV::VLUXSEG7EI8_V:
+ case RISCV::VLUXSEG8EI8_V:
+ case RISCV::VLOXSEG2EI8_V:
+ case RISCV::VLOXSEG3EI8_V:
+ case RISCV::VLOXSEG4EI8_V:
+ case RISCV::VLOXSEG5EI8_V:
+ case RISCV::VLOXSEG6EI8_V:
+ case RISCV::VLOXSEG7EI8_V:
+ case RISCV::VLOXSEG8EI8_V:
+ return getIndexSegmentLoadStoreOperandInfo(3, MI, MO, /* IsLoad */ true);
+ case RISCV::VSUXSEG2EI8_V:
+ case RISCV::VSUXSEG3EI8_V:
+ case RISCV::VSUXSEG4EI8_V:
+ case RISCV::VSUXSEG5EI8_V:
+ case RISCV::VSUXSEG6EI8_V:
+ case RISCV::VSUXSEG7EI8_V:
+ case RISCV::VSUXSEG8EI8_V:
+ case RISCV::VSOXSEG2EI8_V:
+ case RISCV::VSOXSEG3EI8_V:
+ case RISCV::VSOXSEG4EI8_V:
+ case RISCV::VSOXSEG5EI8_V:
+ case RISCV::VSOXSEG6EI8_V:
+ case RISCV::VSOXSEG7EI8_V:
+ case RISCV::VSOXSEG8EI8_V:
+ return getIndexSegmentLoadStoreOperandInfo(3, MI, MO, /* IsLoad */ false);
+ case RISCV::VLUXSEG2EI16_V:
+ case RISCV::VLUXSEG3EI16_V:
+ case RISCV::VLUXSEG4EI16_V:
+ case RISCV::VLUXSEG5EI16_V:
+ case RISCV::VLUXSEG6EI16_V:
+ case RISCV::VLUXSEG7EI16_V:
+ case RISCV::VLUXSEG8EI16_V:
+ case RISCV::VLOXSEG2EI16_V:
+ case RISCV::VLOXSEG3EI16_V:
+ case RISCV::VLOXSEG4EI16_V:
+ case RISCV::VLOXSEG5EI16_V:
+ case RISCV::VLOXSEG6EI16_V:
+ case RISCV::VLOXSEG7EI16_V:
+ case RISCV::VLOXSEG8EI16_V:
+ return getIndexSegmentLoadStoreOperandInfo(4, MI, MO, /* IsLoad */ true);
+ case RISCV::VSUXSEG2EI16_V:
+ case RISCV::VSUXSEG3EI16_V:
+ case RISCV::VSUXSEG4EI16_V:
+ case RISCV::VSUXSEG5EI16_V:
+ case RISCV::VSUXSEG6EI16_V:
+ case RISCV::VSUXSEG7EI16_V:
+ case RISCV::VSUXSEG8EI16_V:
+ case RISCV::VSOXSEG2EI16_V:
+ case RISCV::VSOXSEG3EI16_V:
+ case RISCV::VSOXSEG4EI16_V:
+ case RISCV::VSOXSEG5EI16_V:
+ case RISCV::VSOXSEG6EI16_V:
+ case RISCV::VSOXSEG7EI16_V:
+ case RISCV::VSOXSEG8EI16_V:
+ return getIndexSegmentLoadStoreOperandInfo(4, MI, MO, /* IsLoad */ false);
+ case RISCV::VLUXSEG2EI32_V:
+ case RISCV::VLUXSEG3EI32_V:
+ case RISCV::VLUXSEG4EI32_V:
+ case RISCV::VLUXSEG5EI32_V:
+ case RISCV::VLUXSEG6EI32_V:
+ case RISCV::VLUXSEG7EI32_V:
+ case RISCV::VLUXSEG8EI32_V:
+ case RISCV::VLOXSEG2EI32_V:
+ case RISCV::VLOXSEG3EI32_V:
+ case RISCV::VLOXSEG4EI32_V:
+ case RISCV::VLOXSEG5EI32_V:
+ case RISCV::VLOXSEG6EI32_V:
+ case RISCV::VLOXSEG7EI32_V:
+ case RISCV::VLOXSEG8EI32_V:
+ return getIndexSegmentLoadStoreOperandInfo(5, MI, MO, /* IsLoad */ true);
+ case RISCV::VSUXSEG2EI32_V:
+ case RISCV::VSUXSEG3EI32_V:
+ case RISCV::VSUXSEG4EI32_V:
+ case RISCV::VSUXSEG5EI32_V:
+ case RISCV::VSUXSEG6EI32_V:
+ case RISCV::VSUXSEG7EI32_V:
+ case RISCV::VSUXSEG8EI32_V:
+ case RISCV::VSOXSEG2EI32_V:
+ case RISCV::VSOXSEG3EI32_V:
+ case RISCV::VSOXSEG4EI32_V:
+ case RISCV::VSOXSEG5EI32_V:
+ case RISCV::VSOXSEG6EI32_V:
+ case RISCV::VSOXSEG7EI32_V:
+ case RISCV::VSOXSEG8EI32_V:
+ return getIndexSegmentLoadStoreOperandInfo(5, MI, MO, /* IsLoad */ false);
+ case RISCV::VLUXSEG2EI64_V:
+ case RISCV::VLUXSEG3EI64_V:
+ case RISCV::VLUXSEG4EI64_V:
+ case RISCV::VLUXSEG5EI64_V:
+ case RISCV::VLUXSEG6EI64_V:
+ case RISCV::VLUXSEG7EI64_V:
+ case RISCV::VLUXSEG8EI64_V:
+ case RISCV::VLOXSEG2EI64_V:
+ case RISCV::VLOXSEG3EI64_V:
+ case RISCV::VLOXSEG4EI64_V:
+ case RISCV::VLOXSEG5EI64_V:
+ case RISCV::VLOXSEG6EI64_V:
+ case RISCV::VLOXSEG7EI64_V:
+ case RISCV::VLOXSEG8EI64_V:
+ return getIndexSegmentLoadStoreOperandInfo(6, MI, MO, /* IsLoad */ true);
+ case RISCV::VSUXSEG2EI64_V:
+ case RISCV::VSUXSEG3EI64_V:
+ case RISCV::VSUXSEG4EI64_V:
+ case RISCV::VSUXSEG5EI64_V:
+ case RISCV::VSUXSEG6EI64_V:
+ case RISCV::VSUXSEG7EI64_V:
+ case RISCV::VSUXSEG8EI64_V:
+ case RISCV::VSOXSEG2EI64_V:
+ case RISCV::VSOXSEG3EI64_V:
+ case RISCV::VSOXSEG4EI64_V:
+ case RISCV::VSOXSEG5EI64_V:
+ case RISCV::VSOXSEG6EI64_V:
+ case RISCV::VSOXSEG7EI64_V:
+ case RISCV::VSOXSEG8EI64_V:
+ return getIndexSegmentLoadStoreOperandInfo(6, MI, MO, /* IsLoad */ false);
+
+ // 7.9. Vector Load/Store Whole Register Instructions
+ // EMUL=nr. EEW=eew. Since in-register byte layouts are idential to in-memory
+ // byte layouts, the same data is writen to destination register regardless
+ // of EEW. eew is just a hint to the hardware and has not functional impact.
+ // Therefore, it is be okay if we ignore eew and always use the same EEW to
+ // create more optimization opportunities.
+ // FIXME: Instead of using any SEW, we really ought to return the SEW in the
+ // instruction and add a field to OperandInfo that says the SEW is just a hint
+ // so that this optimization can use any sew to construct a ratio.
+ case RISCV::VL1RE8_V:
+ case RISCV::VL1RE16_V:
+ case RISCV::VL1RE32_V:
+ case RISCV::VL1RE64_V:
+ case RISCV::VS1R_V:
+ return OperandInfo(RISCVII::VLMUL::LMUL_1, 0);
+ case RISCV::VL2RE8_V:
+ case RISCV::VL2RE16_V:
+ case RISCV::VL2RE32_V:
+ case RISCV::VL2RE64_V:
+ case RISCV::VS2R_V:
+ return OperandInfo(RISCVII::VLMUL::LMUL_2, 0);
+ case RISCV::VL4RE8_V:
+ case RISCV::VL4RE16_V:
+ case RISCV::VL4RE32_V:
+ case RISCV::VL4RE64_V:
+ case RISCV::VS4R_V:
+ return OperandInfo(RISCVII::VLMUL::LMUL_4, 0);
+ case RISCV::VL8RE8_V:
+ case RISCV::VL8RE16_V:
+ case RISCV::VL8RE32_V:
+ case RISCV::VL8RE64_V:
+ case RISCV::VS8R_V:
+ return OperandInfo(RISCVII::VLMUL::LMUL_8, 0);
----------------
preames wrote:
Ok, another idea...
Their LMUL should match the register class right? We don't have fractional register classes, but we also don't have fraction whole register moves? So maybe combine this with isRVVWholeLoadStore from RISCVInstrInfo.cpp for the selection, then use the register class?
(I'm just looking for ways to remove redundancy. If we can share code with other users, that increases the odds we'd catch a bug.)
Again, this is minor, and I'd be happy to iterate on this type of topic in tree.
https://github.com/llvm/llvm-project/pull/108640
More information about the llvm-commits
mailing list