| 1 | //===--------------------- DispatchStage.cpp --------------------*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | /// \file |
| 9 | /// |
| 10 | /// This file models the dispatch component of an instruction pipeline. |
| 11 | /// |
| 12 | /// The DispatchStage is responsible for updating instruction dependencies |
| 13 | /// and communicating to the simulated instruction scheduler that an instruction |
| 14 | /// is ready to be scheduled for execution. |
| 15 | /// |
| 16 | //===----------------------------------------------------------------------===// |
| 17 | |
| 18 | #include "llvm/MCA/Stages/DispatchStage.h" |
| 19 | #include "llvm/MCA/HWEventListener.h" |
| 20 | #include "llvm/Support/Debug.h" |
| 21 | |
| 22 | #define DEBUG_TYPE "llvm-mca" |
| 23 | |
| 24 | namespace llvm { |
| 25 | namespace mca { |
| 26 | |
| 27 | DispatchStage::DispatchStage(const MCSubtargetInfo &Subtarget, |
| 28 | const MCRegisterInfo &MRI, |
| 29 | unsigned MaxDispatchWidth, RetireControlUnit &R, |
| 30 | RegisterFile &F) |
| 31 | : DispatchWidth(MaxDispatchWidth), AvailableEntries(MaxDispatchWidth), |
| 32 | CarryOver(0U), STI(Subtarget), RCU(R), PRF(F) { |
| 33 | if (!DispatchWidth) |
| 34 | DispatchWidth = Subtarget.getSchedModel().IssueWidth; |
| 35 | } |
| 36 | |
| 37 | void DispatchStage::notifyInstructionDispatched(const InstRef &IR, |
| 38 | ArrayRef<unsigned> UsedRegs, |
| 39 | unsigned UOps) const { |
| 40 | LLVM_DEBUG(dbgs() << "[E] Instruction Dispatched: #" << IR << '\n'); |
| 41 | notifyEvent<HWInstructionEvent>( |
| 42 | Event: HWInstructionDispatchedEvent(IR, UsedRegs, UOps)); |
| 43 | } |
| 44 | |
| 45 | bool DispatchStage::checkPRF(const InstRef &IR) const { |
| 46 | SmallVector<MCPhysReg, 4> RegDefs; |
| 47 | for (const WriteState &RegDef : IR.getInstruction()->getDefs()) |
| 48 | RegDefs.emplace_back(Args: RegDef.getRegisterID()); |
| 49 | |
| 50 | const unsigned RegisterMask = PRF.isAvailable(Regs: RegDefs); |
| 51 | // A mask with all zeroes means: register files are available. |
| 52 | if (RegisterMask) { |
| 53 | notifyEvent<HWStallEvent>( |
| 54 | Event: HWStallEvent(HWStallEvent::RegisterFileStall, IR)); |
| 55 | return false; |
| 56 | } |
| 57 | |
| 58 | return true; |
| 59 | } |
| 60 | |
| 61 | bool DispatchStage::checkRCU(const InstRef &IR) const { |
| 62 | const unsigned NumMicroOps = IR.getInstruction()->getNumMicroOps(); |
| 63 | if (RCU.isAvailable(Quantity: NumMicroOps)) |
| 64 | return true; |
| 65 | notifyEvent<HWStallEvent>( |
| 66 | Event: HWStallEvent(HWStallEvent::RetireControlUnitStall, IR)); |
| 67 | return false; |
| 68 | } |
| 69 | |
| 70 | bool DispatchStage::canDispatch(const InstRef &IR) const { |
| 71 | bool CanDispatch = checkRCU(IR); |
| 72 | CanDispatch &= checkPRF(IR); |
| 73 | CanDispatch &= checkNextStage(IR); |
| 74 | return CanDispatch; |
| 75 | } |
| 76 | |
| 77 | Error DispatchStage::dispatch(InstRef IR) { |
| 78 | assert(!CarryOver && "Cannot dispatch another instruction!" ); |
| 79 | Instruction &IS = *IR.getInstruction(); |
| 80 | const unsigned NumMicroOps = IS.getNumMicroOps(); |
| 81 | if (NumMicroOps > DispatchWidth) { |
| 82 | assert(AvailableEntries == DispatchWidth); |
| 83 | AvailableEntries = 0; |
| 84 | CarryOver = NumMicroOps - DispatchWidth; |
| 85 | CarriedOver = IR; |
| 86 | } else { |
| 87 | assert(AvailableEntries >= NumMicroOps); |
| 88 | AvailableEntries -= NumMicroOps; |
| 89 | } |
| 90 | |
| 91 | // Check if this instructions ends the dispatch group. |
| 92 | if (IS.getEndGroup()) |
| 93 | AvailableEntries = 0; |
| 94 | |
| 95 | // Check if this is an optimizable reg-reg move or an XCHG-like instruction. |
| 96 | if (IS.isOptimizableMove()) |
| 97 | if (PRF.tryEliminateMoveOrSwap(Writes: IS.getDefs(), Reads: IS.getUses())) |
| 98 | IS.setEliminated(); |
| 99 | |
| 100 | // A dependency-breaking instruction doesn't have to wait on the register |
| 101 | // input operands, and it is often optimized at register renaming stage. |
| 102 | // Update RAW dependencies if this instruction is not a dependency-breaking |
| 103 | // instruction. A dependency-breaking instruction is a zero-latency |
| 104 | // instruction that doesn't consume hardware resources. |
| 105 | // An example of dependency-breaking instruction on X86 is a zero-idiom XOR. |
| 106 | // |
| 107 | // We also don't update data dependencies for instructions that have been |
| 108 | // eliminated at register renaming stage. |
| 109 | if (!IS.isEliminated()) { |
| 110 | for (ReadState &RS : IS.getUses()) |
| 111 | PRF.addRegisterRead(RS, STI); |
| 112 | } |
| 113 | |
| 114 | // By default, a dependency-breaking zero-idiom is expected to be optimized |
| 115 | // at register renaming stage. That means, no physical register is allocated |
| 116 | // to the instruction. |
| 117 | SmallVector<unsigned, 4> RegisterFiles(PRF.getNumRegisterFiles()); |
| 118 | for (WriteState &WS : IS.getDefs()) |
| 119 | PRF.addRegisterWrite(Write: WriteRef(IR.getSourceIndex(), &WS), UsedPhysRegs: RegisterFiles); |
| 120 | |
| 121 | // Reserve entries in the reorder buffer. |
| 122 | unsigned RCUTokenID = RCU.dispatch(IS: IR); |
| 123 | // Notify the instruction that it has been dispatched. |
| 124 | IS.dispatch(RCUTokenID); |
| 125 | |
| 126 | // Notify listeners of the "instruction dispatched" event, |
| 127 | // and move IR to the next stage. |
| 128 | notifyInstructionDispatched(IR, UsedRegs: RegisterFiles, |
| 129 | UOps: std::min(a: DispatchWidth, b: NumMicroOps)); |
| 130 | return moveToTheNextStage(IR); |
| 131 | } |
| 132 | |
| 133 | Error DispatchStage::cycleStart() { |
| 134 | // The retire stage is responsible for calling method `cycleStart` |
| 135 | // on the PRF. |
| 136 | if (!CarryOver) { |
| 137 | AvailableEntries = DispatchWidth; |
| 138 | return ErrorSuccess(); |
| 139 | } |
| 140 | |
| 141 | AvailableEntries = CarryOver >= DispatchWidth ? 0 : DispatchWidth - CarryOver; |
| 142 | unsigned DispatchedOpcodes = DispatchWidth - AvailableEntries; |
| 143 | CarryOver -= DispatchedOpcodes; |
| 144 | assert(CarriedOver && "Invalid dispatched instruction" ); |
| 145 | |
| 146 | SmallVector<unsigned, 8> RegisterFiles(PRF.getNumRegisterFiles(), 0U); |
| 147 | notifyInstructionDispatched(IR: CarriedOver, UsedRegs: RegisterFiles, UOps: DispatchedOpcodes); |
| 148 | if (!CarryOver) |
| 149 | CarriedOver = InstRef(); |
| 150 | return ErrorSuccess(); |
| 151 | } |
| 152 | |
| 153 | bool DispatchStage::isAvailable(const InstRef &IR) const { |
| 154 | // Conservatively bail out if there are no available dispatch entries. |
| 155 | if (!AvailableEntries) |
| 156 | return false; |
| 157 | |
| 158 | const Instruction &Inst = *IR.getInstruction(); |
| 159 | unsigned NumMicroOps = Inst.getNumMicroOps(); |
| 160 | unsigned Required = std::min(a: NumMicroOps, b: DispatchWidth); |
| 161 | if (Required > AvailableEntries) |
| 162 | return false; |
| 163 | |
| 164 | if (Inst.getBeginGroup() && AvailableEntries != DispatchWidth) |
| 165 | return false; |
| 166 | |
| 167 | // The dispatch logic doesn't internally buffer instructions. It only accepts |
| 168 | // instructions that can be successfully moved to the next stage during this |
| 169 | // same cycle. |
| 170 | return canDispatch(IR); |
| 171 | } |
| 172 | |
| 173 | Error DispatchStage::execute(InstRef &IR) { |
| 174 | assert(canDispatch(IR) && "Cannot dispatch another instruction!" ); |
| 175 | return dispatch(IR); |
| 176 | } |
| 177 | |
| 178 | #ifndef NDEBUG |
| 179 | void DispatchStage::dump() const { |
| 180 | PRF.dump(); |
| 181 | RCU.dump(); |
| 182 | } |
| 183 | #endif |
| 184 | } // namespace mca |
| 185 | } // namespace llvm |
| 186 | |