|
| 1 | +//===----- HexagonLoopAlign.cpp - Generate loop alignment directives -----===// |
| 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 | +// Inspect a basic block and if its single basic block loop with a small |
| 9 | +// number of instructions, set the prefLoopAlignment to 32 bytes (5). |
| 10 | +//===----------------------------------------------------------------------===// |
| 11 | + |
| 12 | +#define DEBUG_TYPE "hexagon-loop-align" |
| 13 | + |
| 14 | +#include "HexagonTargetMachine.h" |
| 15 | +#include "llvm/CodeGen/MachineBlockFrequencyInfo.h" |
| 16 | +#include "llvm/CodeGen/MachineBranchProbabilityInfo.h" |
| 17 | +#include "llvm/CodeGen/SchedulerRegistry.h" |
| 18 | +#include "llvm/Support/Debug.h" |
| 19 | + |
| 20 | +using namespace llvm; |
| 21 | + |
| 22 | +static cl::opt<bool> |
| 23 | + DisableLoopAlign("disable-hexagon-loop-align", cl::Hidden, |
| 24 | + cl::desc("Disable Hexagon loop alignment pass")); |
| 25 | + |
| 26 | +static cl::opt<uint32_t> HVXLoopAlignLimitUB( |
| 27 | + "hexagon-hvx-loop-align-limit-ub", cl::Hidden, cl::init(16), |
| 28 | + cl::desc("Set hexagon hvx loop upper bound align limit")); |
| 29 | + |
| 30 | +static cl::opt<uint32_t> TinyLoopAlignLimitUB( |
| 31 | + "hexagon-tiny-loop-align-limit-ub", cl::Hidden, cl::init(16), |
| 32 | + cl::desc("Set hexagon tiny-core loop upper bound align limit")); |
| 33 | + |
| 34 | +static cl::opt<uint32_t> |
| 35 | + LoopAlignLimitUB("hexagon-loop-align-limit-ub", cl::Hidden, cl::init(8), |
| 36 | + cl::desc("Set hexagon loop upper bound align limit")); |
| 37 | + |
| 38 | +static cl::opt<uint32_t> |
| 39 | + LoopAlignLimitLB("hexagon-loop-align-limit-lb", cl::Hidden, cl::init(4), |
| 40 | + cl::desc("Set hexagon loop lower bound align limit")); |
| 41 | + |
| 42 | +static cl::opt<uint32_t> |
| 43 | + LoopBndlAlignLimit("hexagon-loop-bundle-align-limit", cl::Hidden, |
| 44 | + cl::init(4), |
| 45 | + cl::desc("Set hexagon loop align bundle limit")); |
| 46 | + |
| 47 | +static cl::opt<uint32_t> TinyLoopBndlAlignLimit( |
| 48 | + "hexagon-tiny-loop-bundle-align-limit", cl::Hidden, cl::init(8), |
| 49 | + cl::desc("Set hexagon tiny-core loop align bundle limit")); |
| 50 | + |
| 51 | +static cl::opt<uint32_t> |
| 52 | + LoopEdgeThreshold("hexagon-loop-edge-threshold", cl::Hidden, cl::init(7500), |
| 53 | + cl::desc("Set hexagon loop align edge theshold")); |
| 54 | + |
| 55 | +namespace llvm { |
| 56 | +FunctionPass *createHexagonLoopAlign(); |
| 57 | +void initializeHexagonLoopAlignPass(PassRegistry &); |
| 58 | +} // namespace llvm |
| 59 | + |
| 60 | +namespace { |
| 61 | + |
| 62 | +class HexagonLoopAlign : public MachineFunctionPass { |
| 63 | + const HexagonSubtarget *HST = nullptr; |
| 64 | + const TargetMachine *HTM = nullptr; |
| 65 | + const HexagonInstrInfo *HII = nullptr; |
| 66 | + |
| 67 | +public: |
| 68 | + static char ID; |
| 69 | + HexagonLoopAlign() : MachineFunctionPass(ID) { |
| 70 | + initializeHexagonLoopAlignPass(*PassRegistry::getPassRegistry()); |
| 71 | + } |
| 72 | + bool shouldBalignLoop(MachineBasicBlock &BB, bool AboveThres); |
| 73 | + bool isSingleLoop(MachineBasicBlock &MBB); |
| 74 | + bool attemptToBalignSmallLoop(MachineFunction &MF, MachineBasicBlock &MBB); |
| 75 | + |
| 76 | + void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 77 | + AU.addRequired<MachineBranchProbabilityInfo>(); |
| 78 | + AU.addRequired<MachineBlockFrequencyInfo>(); |
| 79 | + MachineFunctionPass::getAnalysisUsage(AU); |
| 80 | + } |
| 81 | + |
| 82 | + StringRef getPassName() const override { return "Hexagon LoopAlign pass"; } |
| 83 | + bool runOnMachineFunction(MachineFunction &MF) override; |
| 84 | +}; |
| 85 | + |
| 86 | +char HexagonLoopAlign::ID = 0; |
| 87 | + |
| 88 | +bool HexagonLoopAlign::shouldBalignLoop(MachineBasicBlock &BB, |
| 89 | + bool AboveThres) { |
| 90 | + bool isVec = false; |
| 91 | + unsigned InstCnt = 0; |
| 92 | + unsigned BndlCnt = 0; |
| 93 | + |
| 94 | + for (MachineBasicBlock::instr_iterator II = BB.instr_begin(), |
| 95 | + IE = BB.instr_end(); |
| 96 | + II != IE; ++II) { |
| 97 | + |
| 98 | + // End if the instruction is endloop. |
| 99 | + if (HII->isEndLoopN(II->getOpcode())) |
| 100 | + break; |
| 101 | + // Count the number of bundles. |
| 102 | + if (II->isBundle()) { |
| 103 | + BndlCnt++; |
| 104 | + continue; |
| 105 | + } |
| 106 | + // Skip over debug instructions. |
| 107 | + if (II->isDebugInstr()) |
| 108 | + continue; |
| 109 | + // Check if there are any HVX instructions in loop. |
| 110 | + isVec |= HII->isHVXVec(*II); |
| 111 | + // Count the number of instructions. |
| 112 | + InstCnt++; |
| 113 | + } |
| 114 | + |
| 115 | + LLVM_DEBUG({ |
| 116 | + dbgs() << "Bundle Count : " << BndlCnt << "\n"; |
| 117 | + dbgs() << "Instruction Count : " << InstCnt << "\n"; |
| 118 | + }); |
| 119 | + |
| 120 | + unsigned LimitUB = 0; |
| 121 | + unsigned LimitBndl = LoopBndlAlignLimit; |
| 122 | + // The conditions in the order of priority. |
| 123 | + if (HST->isTinyCore()) { |
| 124 | + LimitUB = TinyLoopAlignLimitUB; |
| 125 | + LimitBndl = TinyLoopBndlAlignLimit; |
| 126 | + } else if (isVec) |
| 127 | + LimitUB = HVXLoopAlignLimitUB; |
| 128 | + else if (AboveThres) |
| 129 | + LimitUB = LoopAlignLimitUB; |
| 130 | + |
| 131 | + // if the upper bound is not set to a value, implies we didn't meet |
| 132 | + // the criteria. |
| 133 | + if (LimitUB == 0) |
| 134 | + return false; |
| 135 | + |
| 136 | + return InstCnt >= LoopAlignLimitLB && InstCnt <= LimitUB && |
| 137 | + BndlCnt <= LimitBndl; |
| 138 | +} |
| 139 | + |
| 140 | +bool HexagonLoopAlign::isSingleLoop(MachineBasicBlock &MBB) { |
| 141 | + int Succs = MBB.succ_size(); |
| 142 | + return (MBB.isSuccessor(&MBB) && (Succs == 2)); |
| 143 | +} |
| 144 | + |
| 145 | +bool HexagonLoopAlign::attemptToBalignSmallLoop(MachineFunction &MF, |
| 146 | + MachineBasicBlock &MBB) { |
| 147 | + if (!isSingleLoop(MBB)) |
| 148 | + return false; |
| 149 | + |
| 150 | + const MachineBranchProbabilityInfo *MBPI = |
| 151 | + &getAnalysis<MachineBranchProbabilityInfo>(); |
| 152 | + const MachineBlockFrequencyInfo *MBFI = |
| 153 | + &getAnalysis<MachineBlockFrequencyInfo>(); |
| 154 | + |
| 155 | + // Compute frequency of back edge, |
| 156 | + BlockFrequency BlockFreq = MBFI->getBlockFreq(&MBB); |
| 157 | + BranchProbability BrProb = MBPI->getEdgeProbability(&MBB, &MBB); |
| 158 | + BlockFrequency EdgeFreq = BlockFreq * BrProb; |
| 159 | + LLVM_DEBUG({ |
| 160 | + dbgs() << "Loop Align Pass:\n"; |
| 161 | + dbgs() << "\tedge with freq(" << EdgeFreq.getFrequency() << ")\n"; |
| 162 | + }); |
| 163 | + |
| 164 | + bool AboveThres = EdgeFreq.getFrequency() > LoopEdgeThreshold; |
| 165 | + if (shouldBalignLoop(MBB, AboveThres)) { |
| 166 | + // We found a loop, change its alignment to be 32 (5). |
| 167 | + MBB.setAlignment(llvm::Align(1 << 5)); |
| 168 | + return true; |
| 169 | + } |
| 170 | + return false; |
| 171 | +} |
| 172 | + |
| 173 | +// Inspect each basic block, and if its a single BB loop, see if it |
| 174 | +// meets the criteria for increasing alignment to 32. |
| 175 | + |
| 176 | +bool HexagonLoopAlign::runOnMachineFunction(MachineFunction &MF) { |
| 177 | + |
| 178 | + HST = &MF.getSubtarget<HexagonSubtarget>(); |
| 179 | + HII = HST->getInstrInfo(); |
| 180 | + HTM = &MF.getTarget(); |
| 181 | + |
| 182 | + if (skipFunction(MF.getFunction())) |
| 183 | + return false; |
| 184 | + if (DisableLoopAlign) |
| 185 | + return false; |
| 186 | + |
| 187 | + // This optimization is performed at |
| 188 | + // i) -O2 and above, and when the loop has a HVX instruction. |
| 189 | + // ii) -O3 |
| 190 | + if (HST->useHVXOps()) { |
| 191 | + if (HTM->getOptLevel() < CodeGenOptLevel::Default) |
| 192 | + return false; |
| 193 | + } else { |
| 194 | + if (HTM->getOptLevel() < CodeGenOptLevel::Aggressive) |
| 195 | + return false; |
| 196 | + } |
| 197 | + |
| 198 | + bool Changed = false; |
| 199 | + for (MachineFunction::iterator MBBi = MF.begin(), MBBe = MF.end(); |
| 200 | + MBBi != MBBe; ++MBBi) { |
| 201 | + MachineBasicBlock &MBB = *MBBi; |
| 202 | + Changed |= attemptToBalignSmallLoop(MF, MBB); |
| 203 | + } |
| 204 | + return Changed; |
| 205 | +} |
| 206 | + |
| 207 | +} // namespace |
| 208 | + |
| 209 | +INITIALIZE_PASS(HexagonLoopAlign, "hexagon-loop-align", |
| 210 | + "Hexagon LoopAlign pass", false, false) |
| 211 | + |
| 212 | +//===----------------------------------------------------------------------===// |
| 213 | +// Public Constructor Functions |
| 214 | +//===----------------------------------------------------------------------===// |
| 215 | + |
| 216 | +FunctionPass *llvm::createHexagonLoopAlign() { return new HexagonLoopAlign(); } |
0 commit comments