| 1 | //===-- Thumb2ITBlockPass.cpp - Insert Thumb-2 IT blocks ------------------===// |
| 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 | |
| 9 | #include "ARM.h" |
| 10 | #include "ARMMachineFunctionInfo.h" |
| 11 | #include "ARMSubtarget.h" |
| 12 | #include "Thumb2InstrInfo.h" |
| 13 | #include "llvm/ADT/SmallSet.h" |
| 14 | #include "llvm/ADT/SmallVector.h" |
| 15 | #include "llvm/ADT/Statistic.h" |
| 16 | #include "llvm/ADT/StringRef.h" |
| 17 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 18 | #include "llvm/CodeGen/MachineFunction.h" |
| 19 | #include "llvm/CodeGen/MachineFunctionPass.h" |
| 20 | #include "llvm/CodeGen/MachineInstr.h" |
| 21 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 22 | #include "llvm/CodeGen/MachineInstrBundle.h" |
| 23 | #include "llvm/CodeGen/MachineOperand.h" |
| 24 | #include "llvm/IR/DebugLoc.h" |
| 25 | #include "llvm/MC/MCInstrDesc.h" |
| 26 | #include <cassert> |
| 27 | #include <new> |
| 28 | |
| 29 | using namespace llvm; |
| 30 | |
| 31 | #define DEBUG_TYPE "thumb2-it" |
| 32 | #define PASS_NAME "Thumb IT blocks insertion pass" |
| 33 | |
| 34 | STATISTIC(NumITs, "Number of IT blocks inserted" ); |
| 35 | STATISTIC(NumMovedInsts, "Number of predicated instructions moved" ); |
| 36 | |
| 37 | using RegisterSet = SmallSet<unsigned, 4>; |
| 38 | |
| 39 | namespace { |
| 40 | |
| 41 | class Thumb2ITBlock : public MachineFunctionPass { |
| 42 | public: |
| 43 | static char ID; |
| 44 | |
| 45 | bool restrictIT; |
| 46 | const Thumb2InstrInfo *TII; |
| 47 | const TargetRegisterInfo *TRI; |
| 48 | ARMFunctionInfo *AFI; |
| 49 | |
| 50 | Thumb2ITBlock() : MachineFunctionPass(ID) {} |
| 51 | |
| 52 | bool runOnMachineFunction(MachineFunction &Fn) override; |
| 53 | |
| 54 | MachineFunctionProperties getRequiredProperties() const override { |
| 55 | return MachineFunctionProperties().setNoVRegs(); |
| 56 | } |
| 57 | |
| 58 | StringRef getPassName() const override { |
| 59 | return PASS_NAME; |
| 60 | } |
| 61 | |
| 62 | private: |
| 63 | bool MoveCopyOutOfITBlock(MachineInstr *MI, |
| 64 | ARMCC::CondCodes CC, ARMCC::CondCodes OCC, |
| 65 | RegisterSet &Defs, RegisterSet &Uses); |
| 66 | bool InsertITInstructions(MachineBasicBlock &Block); |
| 67 | }; |
| 68 | |
| 69 | char Thumb2ITBlock::ID = 0; |
| 70 | |
| 71 | } // end anonymous namespace |
| 72 | |
| 73 | INITIALIZE_PASS(Thumb2ITBlock, DEBUG_TYPE, PASS_NAME, false, false) |
| 74 | |
| 75 | /// TrackDefUses - Tracking what registers are being defined and used by |
| 76 | /// instructions in the IT block. This also tracks "dependencies", i.e. uses |
| 77 | /// in the IT block that are defined before the IT instruction. |
| 78 | static void TrackDefUses(MachineInstr *MI, RegisterSet &Defs, RegisterSet &Uses, |
| 79 | const TargetRegisterInfo *TRI) { |
| 80 | using RegList = SmallVector<unsigned, 4>; |
| 81 | RegList LocalDefs; |
| 82 | RegList LocalUses; |
| 83 | |
| 84 | for (auto &MO : MI->operands()) { |
| 85 | if (!MO.isReg()) |
| 86 | continue; |
| 87 | Register Reg = MO.getReg(); |
| 88 | if (!Reg || Reg == ARM::ITSTATE || Reg == ARM::SP) |
| 89 | continue; |
| 90 | if (MO.isUse()) |
| 91 | LocalUses.push_back(Elt: Reg); |
| 92 | else |
| 93 | LocalDefs.push_back(Elt: Reg); |
| 94 | } |
| 95 | |
| 96 | auto InsertUsesDefs = [&](RegList &Regs, RegisterSet &UsesDefs) { |
| 97 | for (unsigned Reg : Regs) |
| 98 | UsesDefs.insert_range(R: TRI->subregs_inclusive(Reg)); |
| 99 | }; |
| 100 | |
| 101 | InsertUsesDefs(LocalDefs, Defs); |
| 102 | InsertUsesDefs(LocalUses, Uses); |
| 103 | } |
| 104 | |
| 105 | /// Clear kill flags for any uses in the given set. This will likely |
| 106 | /// conservatively remove more kill flags than are necessary, but removing them |
| 107 | /// is safer than incorrect kill flags remaining on instructions. |
| 108 | static void ClearKillFlags(MachineInstr *MI, RegisterSet &Uses) { |
| 109 | for (MachineOperand &MO : MI->operands()) { |
| 110 | if (!MO.isReg() || MO.isDef() || !MO.isKill()) |
| 111 | continue; |
| 112 | if (!Uses.count(V: MO.getReg())) |
| 113 | continue; |
| 114 | MO.setIsKill(false); |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | static bool isCopy(MachineInstr *MI) { |
| 119 | switch (MI->getOpcode()) { |
| 120 | default: |
| 121 | return false; |
| 122 | case ARM::MOVr: |
| 123 | case ARM::MOVr_TC: |
| 124 | case ARM::tMOVr: |
| 125 | case ARM::t2MOVr: |
| 126 | return true; |
| 127 | } |
| 128 | } |
| 129 | |
| 130 | bool |
| 131 | Thumb2ITBlock::MoveCopyOutOfITBlock(MachineInstr *MI, |
| 132 | ARMCC::CondCodes CC, ARMCC::CondCodes OCC, |
| 133 | RegisterSet &Defs, RegisterSet &Uses) { |
| 134 | if (!isCopy(MI)) |
| 135 | return false; |
| 136 | // llvm models select's as two-address instructions. That means a copy |
| 137 | // is inserted before a t2MOVccr, etc. If the copy is scheduled in |
| 138 | // between selects we would end up creating multiple IT blocks. |
| 139 | assert(MI->getOperand(0).getSubReg() == 0 && |
| 140 | MI->getOperand(1).getSubReg() == 0 && |
| 141 | "Sub-register indices still around?" ); |
| 142 | |
| 143 | Register DstReg = MI->getOperand(i: 0).getReg(); |
| 144 | Register SrcReg = MI->getOperand(i: 1).getReg(); |
| 145 | |
| 146 | // First check if it's safe to move it. |
| 147 | if (Uses.count(V: DstReg) || Defs.count(V: SrcReg)) |
| 148 | return false; |
| 149 | |
| 150 | // If the CPSR is defined by this copy, then we don't want to move it. E.g., |
| 151 | // if we have: |
| 152 | // |
| 153 | // movs r1, r1 |
| 154 | // rsb r1, 0 |
| 155 | // movs r2, r2 |
| 156 | // rsb r2, 0 |
| 157 | // |
| 158 | // we don't want this to be converted to: |
| 159 | // |
| 160 | // movs r1, r1 |
| 161 | // movs r2, r2 |
| 162 | // itt mi |
| 163 | // rsb r1, 0 |
| 164 | // rsb r2, 0 |
| 165 | // |
| 166 | const MCInstrDesc &MCID = MI->getDesc(); |
| 167 | if (MI->hasOptionalDef() && |
| 168 | MI->getOperand(i: MCID.getNumOperands() - 1).getReg() == ARM::CPSR) |
| 169 | return false; |
| 170 | |
| 171 | // Then peek at the next instruction to see if it's predicated on CC or OCC. |
| 172 | // If not, then there is nothing to be gained by moving the copy. |
| 173 | MachineBasicBlock::iterator I = MI; |
| 174 | ++I; |
| 175 | MachineBasicBlock::iterator E = MI->getParent()->end(); |
| 176 | |
| 177 | while (I != E && I->isDebugInstr()) |
| 178 | ++I; |
| 179 | |
| 180 | if (I != E) { |
| 181 | Register NPredReg; |
| 182 | ARMCC::CondCodes NCC = getITInstrPredicate(MI: *I, PredReg&: NPredReg); |
| 183 | if (NCC == CC || NCC == OCC) |
| 184 | return true; |
| 185 | } |
| 186 | return false; |
| 187 | } |
| 188 | |
| 189 | bool Thumb2ITBlock::InsertITInstructions(MachineBasicBlock &MBB) { |
| 190 | bool Modified = false; |
| 191 | RegisterSet Defs, Uses; |
| 192 | MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end(); |
| 193 | |
| 194 | while (MBBI != E) { |
| 195 | MachineInstr *MI = &*MBBI; |
| 196 | DebugLoc dl = MI->getDebugLoc(); |
| 197 | Register PredReg; |
| 198 | ARMCC::CondCodes CC = getITInstrPredicate(MI: *MI, PredReg); |
| 199 | if (CC == ARMCC::AL) { |
| 200 | ++MBBI; |
| 201 | continue; |
| 202 | } |
| 203 | |
| 204 | Defs.clear(); |
| 205 | Uses.clear(); |
| 206 | TrackDefUses(MI, Defs, Uses, TRI); |
| 207 | |
| 208 | // Insert an IT instruction. |
| 209 | MachineInstrBuilder MIB = BuildMI(BB&: MBB, I: MBBI, MIMD: dl, MCID: TII->get(Opcode: ARM::t2IT)) |
| 210 | .addImm(Val: CC); |
| 211 | |
| 212 | // Add implicit use of ITSTATE to IT block instructions. |
| 213 | MI->addOperand(Op: MachineOperand::CreateReg(Reg: ARM::ITSTATE, isDef: false/*ifDef*/, |
| 214 | isImp: true/*isImp*/, isKill: false/*isKill*/)); |
| 215 | |
| 216 | MachineInstr *LastITMI = MI; |
| 217 | MachineBasicBlock::iterator InsertPos = MIB.getInstr(); |
| 218 | ++MBBI; |
| 219 | |
| 220 | // Form IT block. |
| 221 | ARMCC::CondCodes OCC = ARMCC::getOppositeCondition(CC); |
| 222 | unsigned Mask = 0, Pos = 3; |
| 223 | |
| 224 | // IT blocks are limited to one conditional op if -arm-restrict-it |
| 225 | // is set: skip the loop |
| 226 | if (!restrictIT) { |
| 227 | LLVM_DEBUG(dbgs() << "Allowing complex IT block\n" ); |
| 228 | // Branches, including tricky ones like LDM_RET, need to end an IT |
| 229 | // block so check the instruction we just put in the block. |
| 230 | for (; MBBI != E && Pos && |
| 231 | (!MI->isBranch() && !MI->isReturn()) ; ++MBBI) { |
| 232 | if (MBBI->isDebugInstr()) |
| 233 | continue; |
| 234 | |
| 235 | MachineInstr *NMI = &*MBBI; |
| 236 | MI = NMI; |
| 237 | |
| 238 | Register NPredReg; |
| 239 | ARMCC::CondCodes NCC = getITInstrPredicate(MI: *NMI, PredReg&: NPredReg); |
| 240 | if (NCC == CC || NCC == OCC) { |
| 241 | Mask |= ((NCC ^ CC) & 1) << Pos; |
| 242 | // Add implicit use of ITSTATE. |
| 243 | NMI->addOperand(Op: MachineOperand::CreateReg(Reg: ARM::ITSTATE, isDef: false/*ifDef*/, |
| 244 | isImp: true/*isImp*/, isKill: false/*isKill*/)); |
| 245 | LastITMI = NMI; |
| 246 | } else { |
| 247 | if (NCC == ARMCC::AL && |
| 248 | MoveCopyOutOfITBlock(MI: NMI, CC, OCC, Defs, Uses)) { |
| 249 | --MBBI; |
| 250 | MBB.remove(I: NMI); |
| 251 | MBB.insert(I: InsertPos, MI: NMI); |
| 252 | ClearKillFlags(MI, Uses); |
| 253 | ++NumMovedInsts; |
| 254 | continue; |
| 255 | } |
| 256 | break; |
| 257 | } |
| 258 | TrackDefUses(MI: NMI, Defs, Uses, TRI); |
| 259 | --Pos; |
| 260 | } |
| 261 | } |
| 262 | |
| 263 | // Finalize IT mask. |
| 264 | Mask |= (1 << Pos); |
| 265 | MIB.addImm(Val: Mask); |
| 266 | |
| 267 | // Last instruction in IT block kills ITSTATE. |
| 268 | LastITMI->findRegisterUseOperand(Reg: ARM::ITSTATE, /*TRI=*/nullptr) |
| 269 | ->setIsKill(); |
| 270 | |
| 271 | // Finalize the bundle. |
| 272 | finalizeBundle(MBB, FirstMI: InsertPos.getInstrIterator(), |
| 273 | LastMI: ++LastITMI->getIterator()); |
| 274 | |
| 275 | Modified = true; |
| 276 | ++NumITs; |
| 277 | } |
| 278 | |
| 279 | return Modified; |
| 280 | } |
| 281 | |
| 282 | bool Thumb2ITBlock::runOnMachineFunction(MachineFunction &Fn) { |
| 283 | const ARMSubtarget &STI = Fn.getSubtarget<ARMSubtarget>(); |
| 284 | if (!STI.isThumb2()) |
| 285 | return false; |
| 286 | AFI = Fn.getInfo<ARMFunctionInfo>(); |
| 287 | TII = static_cast<const Thumb2InstrInfo *>(STI.getInstrInfo()); |
| 288 | TRI = STI.getRegisterInfo(); |
| 289 | restrictIT = STI.restrictIT(); |
| 290 | |
| 291 | if (!AFI->isThumbFunction()) |
| 292 | return false; |
| 293 | |
| 294 | bool Modified = false; |
| 295 | for (auto &MBB : Fn ) |
| 296 | Modified |= InsertITInstructions(MBB); |
| 297 | |
| 298 | if (Modified) |
| 299 | AFI->setHasITBlocks(true); |
| 300 | |
| 301 | return Modified; |
| 302 | } |
| 303 | |
| 304 | /// createThumb2ITBlockPass - Returns an instance of the Thumb2 IT blocks |
| 305 | /// insertion pass. |
| 306 | FunctionPass *llvm::createThumb2ITBlockPass() { return new Thumb2ITBlock(); } |
| 307 | |