1 //===---------- PPCTLSDynamicCall.cpp - TLS Dynamic Call Fixup ------------===// 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 // This pass expands ADDItls{ld,gd}LADDR[32] machine instructions into 10 // separate ADDItls[gd]L[32] and GETtlsADDR[32] instructions, both of 11 // which define GPR3. A copy is added from GPR3 to the target virtual 12 // register of the original instruction. The GETtlsADDR[32] is really 13 // a call instruction, so its target register is constrained to be GPR3. 14 // This is not true of ADDItls[gd]L[32], but there is a legacy linker 15 // optimization bug that requires the target register of the addi of 16 // a local- or general-dynamic TLS access sequence to be GPR3. 17 // 18 // This is done in a late pass so that TLS variable accesses can be 19 // fully commoned by MachineCSE. 20 // 21 //===----------------------------------------------------------------------===// 22 23 #include "PPC.h" 24 #include "PPCInstrBuilder.h" 25 #include "PPCInstrInfo.h" 26 #include "PPCTargetMachine.h" 27 #include "llvm/CodeGen/LiveIntervals.h" 28 #include "llvm/CodeGen/MachineFrameInfo.h" 29 #include "llvm/CodeGen/MachineFunctionPass.h" 30 #include "llvm/CodeGen/MachineInstrBuilder.h" 31 #include "llvm/InitializePasses.h" 32 #include "llvm/Support/Debug.h" 33 #include "llvm/Support/raw_ostream.h" 34 35 using namespace llvm; 36 37 #define DEBUG_TYPE "ppc-tls-dynamic-call" 38 39 namespace { 40 struct PPCTLSDynamicCall : public MachineFunctionPass { 41 static char ID; 42 PPCTLSDynamicCall() : MachineFunctionPass(ID) { 43 initializePPCTLSDynamicCallPass(*PassRegistry::getPassRegistry()); 44 } 45 46 const PPCInstrInfo *TII; 47 48 protected: 49 bool processBlock(MachineBasicBlock &MBB) { 50 bool Changed = false; 51 bool NeedFence = true; 52 const PPCSubtarget &Subtarget = 53 MBB.getParent()->getSubtarget<PPCSubtarget>(); 54 bool Is64Bit = Subtarget.isPPC64(); 55 bool IsAIX = Subtarget.isAIXABI(); 56 bool IsLargeModel = 57 Subtarget.getTargetMachine().getCodeModel() == CodeModel::Large; 58 bool IsPCREL = false; 59 MachineFunction *MF = MBB.getParent(); 60 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 61 62 for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end(); 63 I != IE;) { 64 MachineInstr &MI = *I; 65 IsPCREL = isPCREL(MI); 66 // There are a number of slight differences in code generation 67 // when we call .__get_tpointer (32-bit AIX TLS). 68 bool IsTLSTPRelMI = MI.getOpcode() == PPC::GETtlsTpointer32AIX; 69 bool IsTLSLDAIXMI = (MI.getOpcode() == PPC::TLSLDAIX8 || 70 MI.getOpcode() == PPC::TLSLDAIX); 71 72 if (MI.getOpcode() != PPC::ADDItlsgdLADDR && 73 MI.getOpcode() != PPC::ADDItlsldLADDR && 74 MI.getOpcode() != PPC::ADDItlsgdLADDR32 && 75 MI.getOpcode() != PPC::ADDItlsldLADDR32 && 76 MI.getOpcode() != PPC::TLSGDAIX && 77 MI.getOpcode() != PPC::TLSGDAIX8 && !IsTLSTPRelMI && !IsPCREL && 78 !IsTLSLDAIXMI) { 79 // Although we create ADJCALLSTACKDOWN and ADJCALLSTACKUP 80 // as scheduling fences, we skip creating fences if we already 81 // have existing ADJCALLSTACKDOWN/UP to avoid nesting, 82 // which causes verification error with -verify-machineinstrs. 83 if (MI.getOpcode() == PPC::ADJCALLSTACKDOWN) 84 NeedFence = false; 85 else if (MI.getOpcode() == PPC::ADJCALLSTACKUP) 86 NeedFence = true; 87 88 ++I; 89 continue; 90 } 91 92 LLVM_DEBUG(dbgs() << "TLS Dynamic Call Fixup:\n " << MI); 93 94 Register OutReg = MI.getOperand(0).getReg(); 95 Register InReg = PPC::NoRegister; 96 Register GPR3 = Is64Bit ? PPC::X3 : PPC::R3; 97 Register GPR4 = Is64Bit ? PPC::X4 : PPC::R4; 98 if (!IsPCREL && !IsTLSTPRelMI) 99 InReg = MI.getOperand(1).getReg(); 100 DebugLoc DL = MI.getDebugLoc(); 101 102 unsigned Opc1, Opc2; 103 switch (MI.getOpcode()) { 104 default: 105 llvm_unreachable("Opcode inconsistency error"); 106 case PPC::ADDItlsgdLADDR: 107 Opc1 = PPC::ADDItlsgdL; 108 Opc2 = PPC::GETtlsADDR; 109 break; 110 case PPC::ADDItlsldLADDR: 111 Opc1 = PPC::ADDItlsldL; 112 Opc2 = PPC::GETtlsldADDR; 113 break; 114 case PPC::ADDItlsgdLADDR32: 115 Opc1 = PPC::ADDItlsgdL32; 116 Opc2 = PPC::GETtlsADDR32; 117 break; 118 case PPC::ADDItlsldLADDR32: 119 Opc1 = PPC::ADDItlsldL32; 120 Opc2 = PPC::GETtlsldADDR32; 121 break; 122 case PPC::TLSLDAIX: 123 // TLSLDAIX is expanded to one copy and GET_TLS_MOD, so we only set 124 // Opc2 here. 125 Opc2 = PPC::GETtlsMOD32AIX; 126 break; 127 case PPC::TLSLDAIX8: 128 // TLSLDAIX8 is expanded to one copy and GET_TLS_MOD, so we only set 129 // Opc2 here. 130 Opc2 = PPC::GETtlsMOD64AIX; 131 break; 132 case PPC::TLSGDAIX8: 133 // TLSGDAIX8 is expanded to two copies and GET_TLS_ADDR, so we only 134 // set Opc2 here. 135 Opc2 = PPC::GETtlsADDR64AIX; 136 break; 137 case PPC::TLSGDAIX: 138 // TLSGDAIX is expanded to two copies and GET_TLS_ADDR, so we only 139 // set Opc2 here. 140 Opc2 = PPC::GETtlsADDR32AIX; 141 break; 142 case PPC::GETtlsTpointer32AIX: 143 // GETtlsTpointer32AIX is expanded to a call to GET_TPOINTER on AIX 144 // 32-bit mode within PPCAsmPrinter. This instruction does not need 145 // to change, so Opc2 is set to the same instruction opcode. 146 Opc2 = PPC::GETtlsTpointer32AIX; 147 break; 148 case PPC::PADDI8pc: 149 assert(IsPCREL && "Expecting General/Local Dynamic PCRel"); 150 Opc1 = PPC::PADDI8pc; 151 Opc2 = MI.getOperand(2).getTargetFlags() == 152 PPCII::MO_GOT_TLSGD_PCREL_FLAG 153 ? PPC::GETtlsADDRPCREL 154 : PPC::GETtlsldADDRPCREL; 155 } 156 157 // We create ADJCALLSTACKUP and ADJCALLSTACKDOWN around _tls_get_addr 158 // as scheduling fence to avoid it is scheduled before 159 // mflr in the prologue and the address in LR is clobbered (PR25839). 160 // We don't really need to save data to the stack - the clobbered 161 // registers are already saved when the SDNode (e.g. PPCaddiTlsgdLAddr) 162 // gets translated to the pseudo instruction (e.g. ADDItlsgdLADDR). 163 if (NeedFence) { 164 MBB.getParent()->getFrameInfo().setAdjustsStack(true); 165 BuildMI(MBB, I, DL, TII->get(PPC::ADJCALLSTACKDOWN)).addImm(0) 166 .addImm(0); 167 } 168 169 if (IsAIX) { 170 if (IsTLSLDAIXMI) { 171 // The relative order between the node that loads the variable 172 // offset from the TOC, and the .__tls_get_mod node is being tuned 173 // here. It is better to put the variable offset TOC load after the 174 // call, since this node can use clobbers r4/r5. 175 // Search for the pattern of the two nodes that load from the TOC 176 // (either for the variable offset or for the module handle), and 177 // then move the variable offset TOC load right before the node that 178 // uses the OutReg of the .__tls_get_mod node. 179 unsigned LDTocOp = 180 Is64Bit ? (IsLargeModel ? PPC::LDtocL : PPC::LDtoc) 181 : (IsLargeModel ? PPC::LWZtocL : PPC::LWZtoc); 182 if (!RegInfo.use_empty(OutReg)) { 183 std::set<MachineInstr *> Uses; 184 // Collect all instructions that use the OutReg. 185 for (MachineOperand &MO : RegInfo.use_operands(OutReg)) 186 Uses.insert(MO.getParent()); 187 // Find the first user (e.g.: lwax/stfdx) of the OutReg within the 188 // current BB. 189 MachineBasicBlock::iterator UseIter = MBB.begin(); 190 for (MachineBasicBlock::iterator IE = MBB.end(); UseIter != IE; 191 ++UseIter) 192 if (Uses.count(&*UseIter)) 193 break; 194 195 // Additional handling is required when UserIter (the first user 196 // of OutReg) is pointing to a valid node that loads from the TOC. 197 // Check the pattern and do the movement if the pattern matches. 198 if (UseIter != MBB.end()) { 199 // Collect all associated nodes that load from the TOC. Use 200 // hasOneDef() to guard against unexpected scenarios. 201 std::set<MachineInstr *> LoadFromTocs; 202 for (MachineOperand &MO : UseIter->operands()) 203 if (MO.isReg() && MO.isUse()) { 204 Register MOReg = MO.getReg(); 205 if (RegInfo.hasOneDef(MOReg)) { 206 MachineInstr *Temp = 207 RegInfo.getOneDef(MOReg)->getParent(); 208 // For the current TLSLDAIX node, get the corresponding 209 // node that loads from the TOC for the InReg. Otherwise, 210 // Temp probably pointed to the variable offset TOC load 211 // we would like to move. 212 if (Temp == &MI && RegInfo.hasOneDef(InReg)) 213 Temp = RegInfo.getOneDef(InReg)->getParent(); 214 if (Temp->getOpcode() == LDTocOp) 215 LoadFromTocs.insert(Temp); 216 } else { 217 // FIXME: analyze this scenario if there is one. 218 LoadFromTocs.clear(); 219 break; 220 } 221 } 222 223 // Check the two nodes that loaded from the TOC: one should be 224 // "_$TLSML", and the other will be moved before the node that 225 // uses the OutReg of the .__tls_get_mod node. 226 if (LoadFromTocs.size() == 2) { 227 MachineBasicBlock::iterator TLSMLIter = MBB.end(); 228 MachineBasicBlock::iterator OffsetIter = MBB.end(); 229 // Make sure the two nodes that loaded from the TOC are within 230 // the current BB, and that one of them is from the "_$TLSML" 231 // pseudo symbol, while the other is from the variable. 232 for (MachineBasicBlock::iterator I = MBB.begin(), 233 IE = MBB.end(); 234 I != IE; ++I) 235 if (LoadFromTocs.count(&*I)) { 236 MachineOperand MO = I->getOperand(1); 237 if (MO.isGlobal() && MO.getGlobal()->hasName() && 238 MO.getGlobal()->getName() == "_$TLSML") 239 TLSMLIter = I; 240 else 241 OffsetIter = I; 242 } 243 // Perform the movement when the desired scenario has been 244 // identified, which should be when both of the iterators are 245 // valid. 246 if (TLSMLIter != MBB.end() && OffsetIter != MBB.end()) 247 OffsetIter->moveBefore(&*UseIter); 248 } 249 } 250 } 251 // The module-handle is copied into r3. The copy is followed by 252 // GETtlsMOD32AIX/GETtlsMOD64AIX. 253 BuildMI(MBB, I, DL, TII->get(TargetOpcode::COPY), GPR3) 254 .addReg(InReg); 255 // The call to .__tls_get_mod. 256 BuildMI(MBB, I, DL, TII->get(Opc2), GPR3).addReg(GPR3); 257 } else if (!IsTLSTPRelMI) { 258 // The variable offset and region handle (for TLSGD) are copied in 259 // r4 and r3. The copies are followed by 260 // GETtlsADDR32AIX/GETtlsADDR64AIX. 261 BuildMI(MBB, I, DL, TII->get(TargetOpcode::COPY), GPR4) 262 .addReg(MI.getOperand(1).getReg()); 263 BuildMI(MBB, I, DL, TII->get(TargetOpcode::COPY), GPR3) 264 .addReg(MI.getOperand(2).getReg()); 265 BuildMI(MBB, I, DL, TII->get(Opc2), GPR3).addReg(GPR3).addReg(GPR4); 266 } else 267 // The opcode of GETtlsTpointer32AIX does not change, because later 268 // this instruction will be expanded into a call to .__get_tpointer, 269 // which will return the thread pointer into r3. 270 BuildMI(MBB, I, DL, TII->get(Opc2), GPR3); 271 } else { 272 MachineInstr *Addi; 273 if (IsPCREL) { 274 Addi = BuildMI(MBB, I, DL, TII->get(Opc1), GPR3).addImm(0); 275 } else { 276 // Expand into two ops built prior to the existing instruction. 277 assert(InReg != PPC::NoRegister && "Operand must be a register"); 278 Addi = BuildMI(MBB, I, DL, TII->get(Opc1), GPR3).addReg(InReg); 279 } 280 281 Addi->addOperand(MI.getOperand(2)); 282 283 MachineInstr *Call = 284 (BuildMI(MBB, I, DL, TII->get(Opc2), GPR3).addReg(GPR3)); 285 if (IsPCREL) 286 Call->addOperand(MI.getOperand(2)); 287 else 288 Call->addOperand(MI.getOperand(3)); 289 } 290 if (NeedFence) 291 BuildMI(MBB, I, DL, TII->get(PPC::ADJCALLSTACKUP)).addImm(0).addImm(0); 292 293 BuildMI(MBB, I, DL, TII->get(TargetOpcode::COPY), OutReg) 294 .addReg(GPR3); 295 296 // Move past the original instruction and remove it. 297 ++I; 298 MI.removeFromParent(); 299 300 Changed = true; 301 } 302 303 return Changed; 304 } 305 306 public: 307 bool isPCREL(const MachineInstr &MI) { 308 return (MI.getOpcode() == PPC::PADDI8pc) && 309 (MI.getOperand(2).getTargetFlags() == 310 PPCII::MO_GOT_TLSGD_PCREL_FLAG || 311 MI.getOperand(2).getTargetFlags() == 312 PPCII::MO_GOT_TLSLD_PCREL_FLAG); 313 } 314 315 bool runOnMachineFunction(MachineFunction &MF) override { 316 TII = MF.getSubtarget<PPCSubtarget>().getInstrInfo(); 317 318 bool Changed = false; 319 320 for (MachineBasicBlock &B : llvm::make_early_inc_range(MF)) 321 if (processBlock(B)) 322 Changed = true; 323 324 return Changed; 325 } 326 327 void getAnalysisUsage(AnalysisUsage &AU) const override { 328 AU.addRequired<LiveIntervalsWrapperPass>(); 329 AU.addRequired<SlotIndexesWrapperPass>(); 330 MachineFunctionPass::getAnalysisUsage(AU); 331 } 332 }; 333 } 334 335 INITIALIZE_PASS_BEGIN(PPCTLSDynamicCall, DEBUG_TYPE, 336 "PowerPC TLS Dynamic Call Fixup", false, false) 337 INITIALIZE_PASS_DEPENDENCY(LiveIntervalsWrapperPass) 338 INITIALIZE_PASS_DEPENDENCY(SlotIndexesWrapperPass) 339 INITIALIZE_PASS_END(PPCTLSDynamicCall, DEBUG_TYPE, 340 "PowerPC TLS Dynamic Call Fixup", false, false) 341 342 char PPCTLSDynamicCall::ID = 0; 343 FunctionPass* 344 llvm::createPPCTLSDynamicCallPass() { return new PPCTLSDynamicCall(); } 345