| 1 | //===---------------------- ExecuteStage.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 defines the execution stage of an instruction pipeline. |
| 11 | /// |
| 12 | /// The ExecuteStage is responsible for managing the hardware scheduler |
| 13 | /// and issuing notifications that an instruction has been executed. |
| 14 | /// |
| 15 | //===----------------------------------------------------------------------===// |
| 16 | |
| 17 | #include "llvm/MCA/Stages/ExecuteStage.h" |
| 18 | #include "llvm/ADT/SmallVector.h" |
| 19 | #include "llvm/Support/Debug.h" |
| 20 | |
| 21 | #define DEBUG_TYPE "llvm-mca" |
| 22 | |
| 23 | namespace llvm { |
| 24 | namespace mca { |
| 25 | |
| 26 | HWStallEvent::GenericEventType toHWStallEventType(Scheduler::Status Status) { |
| 27 | switch (Status) { |
| 28 | case Scheduler::SC_LOAD_QUEUE_FULL: |
| 29 | return HWStallEvent::LoadQueueFull; |
| 30 | case Scheduler::SC_STORE_QUEUE_FULL: |
| 31 | return HWStallEvent::StoreQueueFull; |
| 32 | case Scheduler::SC_BUFFERS_FULL: |
| 33 | return HWStallEvent::SchedulerQueueFull; |
| 34 | case Scheduler::SC_DISPATCH_GROUP_STALL: |
| 35 | return HWStallEvent::DispatchGroupStall; |
| 36 | case Scheduler::SC_AVAILABLE: |
| 37 | return HWStallEvent::Invalid; |
| 38 | } |
| 39 | |
| 40 | llvm_unreachable("Don't know how to process this StallKind!" ); |
| 41 | } |
| 42 | |
| 43 | bool ExecuteStage::isAvailable(const InstRef &IR) const { |
| 44 | if (Scheduler::Status S = HWS.isAvailable(IR)) { |
| 45 | HWStallEvent::GenericEventType ET = toHWStallEventType(Status: S); |
| 46 | notifyEvent<HWStallEvent>(Event: HWStallEvent(ET, IR)); |
| 47 | return false; |
| 48 | } |
| 49 | |
| 50 | return true; |
| 51 | } |
| 52 | |
| 53 | Error ExecuteStage::issueInstruction(InstRef &IR) { |
| 54 | SmallVector<ResourceUse, 4> Used; |
| 55 | SmallVector<InstRef, 4> Pending; |
| 56 | SmallVector<InstRef, 4> Ready; |
| 57 | |
| 58 | HWS.issueInstruction(IR, Used, Pending, Ready); |
| 59 | Instruction &IS = *IR.getInstruction(); |
| 60 | NumIssuedOpcodes += IS.getNumMicroOps(); |
| 61 | |
| 62 | notifyReservedOrReleasedBuffers(IR, /* Reserved */ false); |
| 63 | |
| 64 | notifyInstructionIssued(IR, Used); |
| 65 | if (IS.isExecuted()) { |
| 66 | notifyInstructionExecuted(IR); |
| 67 | // FIXME: add a buffer of executed instructions. |
| 68 | if (Error S = moveToTheNextStage(IR)) |
| 69 | return S; |
| 70 | } |
| 71 | |
| 72 | for (const InstRef &I : Pending) |
| 73 | notifyInstructionPending(IR: I); |
| 74 | |
| 75 | for (const InstRef &I : Ready) |
| 76 | notifyInstructionReady(IR: I); |
| 77 | return ErrorSuccess(); |
| 78 | } |
| 79 | |
| 80 | Error ExecuteStage::issueReadyInstructions() { |
| 81 | InstRef IR = HWS.select(); |
| 82 | while (IR) { |
| 83 | if (Error Err = issueInstruction(IR)) |
| 84 | return Err; |
| 85 | |
| 86 | // Select the next instruction to issue. |
| 87 | IR = HWS.select(); |
| 88 | } |
| 89 | |
| 90 | return ErrorSuccess(); |
| 91 | } |
| 92 | |
| 93 | Error ExecuteStage::cycleStart() { |
| 94 | SmallVector<ResourceRef, 8> Freed; |
| 95 | SmallVector<InstRef, 4> Executed; |
| 96 | SmallVector<InstRef, 4> Pending; |
| 97 | SmallVector<InstRef, 4> Ready; |
| 98 | |
| 99 | HWS.cycleEvent(Freed, Executed, Pending, Ready); |
| 100 | NumDispatchedOpcodes = 0; |
| 101 | NumIssuedOpcodes = 0; |
| 102 | |
| 103 | for (const ResourceRef &RR : Freed) |
| 104 | notifyResourceAvailable(RR); |
| 105 | |
| 106 | for (InstRef &IR : Executed) { |
| 107 | notifyInstructionExecuted(IR); |
| 108 | // FIXME: add a buffer of executed instructions. |
| 109 | if (Error S = moveToTheNextStage(IR)) |
| 110 | return S; |
| 111 | } |
| 112 | |
| 113 | for (const InstRef &IR : Pending) |
| 114 | notifyInstructionPending(IR); |
| 115 | |
| 116 | for (const InstRef &IR : Ready) |
| 117 | notifyInstructionReady(IR); |
| 118 | |
| 119 | return issueReadyInstructions(); |
| 120 | } |
| 121 | |
| 122 | Error ExecuteStage::cycleEnd() { |
| 123 | if (!EnablePressureEvents) |
| 124 | return ErrorSuccess(); |
| 125 | |
| 126 | // Always conservatively report any backpressure events if the dispatch logic |
| 127 | // was stalled due to unavailable scheduler resources. |
| 128 | if (!HWS.hadTokenStall() && NumDispatchedOpcodes <= NumIssuedOpcodes) |
| 129 | return ErrorSuccess(); |
| 130 | |
| 131 | SmallVector<InstRef, 8> Insts; |
| 132 | uint64_t Mask = HWS.analyzeResourcePressure(Insts); |
| 133 | if (Mask) { |
| 134 | LLVM_DEBUG(dbgs() << "[E] Backpressure increased because of unavailable " |
| 135 | "pipeline resources: " |
| 136 | << format_hex(Mask, 16) << '\n'); |
| 137 | HWPressureEvent Ev(HWPressureEvent::RESOURCES, Insts, Mask); |
| 138 | notifyEvent(Event: Ev); |
| 139 | } |
| 140 | |
| 141 | SmallVector<InstRef, 8> RegDeps; |
| 142 | SmallVector<InstRef, 8> MemDeps; |
| 143 | HWS.analyzeDataDependencies(RegDeps, MemDeps); |
| 144 | if (RegDeps.size()) { |
| 145 | LLVM_DEBUG( |
| 146 | dbgs() << "[E] Backpressure increased by register dependencies\n" ); |
| 147 | HWPressureEvent Ev(HWPressureEvent::REGISTER_DEPS, RegDeps); |
| 148 | notifyEvent(Event: Ev); |
| 149 | } |
| 150 | |
| 151 | if (MemDeps.size()) { |
| 152 | LLVM_DEBUG(dbgs() << "[E] Backpressure increased by memory dependencies\n" ); |
| 153 | HWPressureEvent Ev(HWPressureEvent::MEMORY_DEPS, MemDeps); |
| 154 | notifyEvent(Event: Ev); |
| 155 | } |
| 156 | |
| 157 | return ErrorSuccess(); |
| 158 | } |
| 159 | |
| 160 | #ifndef NDEBUG |
| 161 | static void verifyInstructionEliminated(const InstRef &IR) { |
| 162 | const Instruction &Inst = *IR.getInstruction(); |
| 163 | assert(Inst.isEliminated() && "Instruction was not eliminated!" ); |
| 164 | assert(Inst.isReady() && "Instruction in an inconsistent state!" ); |
| 165 | |
| 166 | // Ensure that instructions eliminated at register renaming stage are in a |
| 167 | // consistent state. |
| 168 | assert(!Inst.getMayLoad() && !Inst.getMayStore() && |
| 169 | "Cannot eliminate a memory op!" ); |
| 170 | } |
| 171 | #endif |
| 172 | |
| 173 | Error ExecuteStage::handleInstructionEliminated(InstRef &IR) { |
| 174 | #ifndef NDEBUG |
| 175 | verifyInstructionEliminated(IR); |
| 176 | #endif |
| 177 | notifyInstructionPending(IR); |
| 178 | notifyInstructionReady(IR); |
| 179 | notifyInstructionIssued(IR, Used: {}); |
| 180 | IR.getInstruction()->forceExecuted(); |
| 181 | notifyInstructionExecuted(IR); |
| 182 | return moveToTheNextStage(IR); |
| 183 | } |
| 184 | |
| 185 | // Schedule the instruction for execution on the hardware. |
| 186 | Error ExecuteStage::execute(InstRef &IR) { |
| 187 | assert(isAvailable(IR) && "Scheduler is not available!" ); |
| 188 | |
| 189 | #ifndef NDEBUG |
| 190 | // Ensure that the HWS has not stored this instruction in its queues. |
| 191 | HWS.instructionCheck(IR); |
| 192 | #endif |
| 193 | |
| 194 | if (IR.getInstruction()->isEliminated()) |
| 195 | return handleInstructionEliminated(IR); |
| 196 | |
| 197 | // Reserve a slot in each buffered resource. Also, mark units with |
| 198 | // BufferSize=0 as reserved. Resources with a buffer size of zero will only |
| 199 | // be released after MCIS is issued, and all the ReleaseAtCycles for those |
| 200 | // units have been consumed. |
| 201 | bool IsReadyInstruction = HWS.dispatch(IR); |
| 202 | const Instruction &Inst = *IR.getInstruction(); |
| 203 | unsigned NumMicroOps = Inst.getNumMicroOps(); |
| 204 | NumDispatchedOpcodes += NumMicroOps; |
| 205 | notifyReservedOrReleasedBuffers(IR, /* Reserved */ true); |
| 206 | |
| 207 | if (!IsReadyInstruction) { |
| 208 | if (Inst.isPending()) |
| 209 | notifyInstructionPending(IR); |
| 210 | return ErrorSuccess(); |
| 211 | } |
| 212 | |
| 213 | notifyInstructionPending(IR); |
| 214 | |
| 215 | // If we did not return early, then the scheduler is ready for execution. |
| 216 | notifyInstructionReady(IR); |
| 217 | |
| 218 | // If we cannot issue immediately, the HWS will add IR to its ready queue for |
| 219 | // execution later, so we must return early here. |
| 220 | if (!HWS.mustIssueImmediately(IR)) |
| 221 | return ErrorSuccess(); |
| 222 | |
| 223 | // Issue IR to the underlying pipelines. |
| 224 | return issueInstruction(IR); |
| 225 | } |
| 226 | |
| 227 | void ExecuteStage::notifyInstructionExecuted(const InstRef &IR) const { |
| 228 | LLVM_DEBUG(dbgs() << "[E] Instruction Executed: #" << IR << '\n'); |
| 229 | notifyEvent<HWInstructionEvent>( |
| 230 | Event: HWInstructionEvent(HWInstructionEvent::Executed, IR)); |
| 231 | } |
| 232 | |
| 233 | void ExecuteStage::notifyInstructionPending(const InstRef &IR) const { |
| 234 | LLVM_DEBUG(dbgs() << "[E] Instruction Pending: #" << IR << '\n'); |
| 235 | notifyEvent<HWInstructionEvent>( |
| 236 | Event: HWInstructionEvent(HWInstructionEvent::Pending, IR)); |
| 237 | } |
| 238 | |
| 239 | void ExecuteStage::notifyInstructionReady(const InstRef &IR) const { |
| 240 | LLVM_DEBUG(dbgs() << "[E] Instruction Ready: #" << IR << '\n'); |
| 241 | notifyEvent<HWInstructionEvent>( |
| 242 | Event: HWInstructionEvent(HWInstructionEvent::Ready, IR)); |
| 243 | } |
| 244 | |
| 245 | void ExecuteStage::notifyResourceAvailable(const ResourceRef &RR) const { |
| 246 | LLVM_DEBUG(dbgs() << "[E] Resource Available: [" << RR.first << '.' |
| 247 | << RR.second << "]\n" ); |
| 248 | for (HWEventListener *Listener : getListeners()) |
| 249 | Listener->onResourceAvailable(RRef: RR); |
| 250 | } |
| 251 | |
| 252 | void ExecuteStage::notifyInstructionIssued( |
| 253 | const InstRef &IR, MutableArrayRef<ResourceUse> Used) const { |
| 254 | LLVM_DEBUG({ |
| 255 | dbgs() << "[E] Instruction Issued: #" << IR << '\n'; |
| 256 | for (const ResourceUse &Use : Used) { |
| 257 | assert(Use.second.getDenominator() == 1 && "Invalid cycles!" ); |
| 258 | dbgs() << "[E] Resource Used: [" << Use.first.first << '.' |
| 259 | << Use.first.second << "], " ; |
| 260 | dbgs() << "cycles: " << Use.second.getNumerator() << '\n'; |
| 261 | } |
| 262 | }); |
| 263 | |
| 264 | // Replace resource masks with valid resource processor IDs. |
| 265 | for (ResourceUse &Use : Used) |
| 266 | Use.first.first = HWS.getResourceID(Mask: Use.first.first); |
| 267 | |
| 268 | notifyEvent<HWInstructionEvent>(Event: HWInstructionIssuedEvent(IR, Used)); |
| 269 | } |
| 270 | |
| 271 | void ExecuteStage::notifyReservedOrReleasedBuffers(const InstRef &IR, |
| 272 | bool Reserved) const { |
| 273 | uint64_t UsedBuffers = IR.getInstruction()->getDesc().UsedBuffers; |
| 274 | if (!UsedBuffers) |
| 275 | return; |
| 276 | |
| 277 | SmallVector<unsigned, 4> BufferIDs(llvm::popcount(Value: UsedBuffers), 0); |
| 278 | for (unsigned &ID : BufferIDs) { |
| 279 | uint64_t CurrentBufferMask = UsedBuffers & (-UsedBuffers); |
| 280 | ID = HWS.getResourceID(Mask: CurrentBufferMask); |
| 281 | UsedBuffers ^= CurrentBufferMask; |
| 282 | } |
| 283 | |
| 284 | if (Reserved) { |
| 285 | for (HWEventListener *Listener : getListeners()) |
| 286 | Listener->onReservedBuffers(Inst: IR, Buffers: BufferIDs); |
| 287 | return; |
| 288 | } |
| 289 | |
| 290 | for (HWEventListener *Listener : getListeners()) |
| 291 | Listener->onReleasedBuffers(Inst: IR, Buffers: BufferIDs); |
| 292 | } |
| 293 | |
| 294 | } // namespace mca |
| 295 | } // namespace llvm |
| 296 | |