//===---------- PPCTLSDynamicCall.cpp - TLS Dynamic Call Fixup ------------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// // // This pass expands ADDItls{ld,gd}LADDR[32] machine instructions into // separate ADDItls[gd]L[32] and GETtlsADDR[32] instructions, both of // which define GPR3. A copy is added from GPR3 to the target virtual // register of the original instruction. The GETtlsADDR[32] is really // a call instruction, so its target register is constrained to be GPR3. // This is not true of ADDItls[gd]L[32], but there is a legacy linker // optimization bug that requires the target register of the addi of // a local- or general-dynamic TLS access sequence to be GPR3. // // This is done in a late pass so that TLS variable accesses can be // fully commoned by MachineCSE. // //===----------------------------------------------------------------------===// #include "PPC.h" #include "PPCInstrBuilder.h" #include "PPCInstrInfo.h" #include "PPCTargetMachine.h" #include "llvm/CodeGen/LiveIntervals.h" #include "llvm/CodeGen/MachineFrameInfo.h" #include "llvm/CodeGen/MachineFunctionPass.h" #include "llvm/CodeGen/MachineInstrBuilder.h" #include "llvm/InitializePasses.h" #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" using namespace llvm; #define DEBUG_TYPE "ppc-tls-dynamic-call" namespace { struct PPCTLSDynamicCall : public MachineFunctionPass { static char ID; PPCTLSDynamicCall() : MachineFunctionPass(ID) { initializePPCTLSDynamicCallPass(*PassRegistry::getPassRegistry()); } const PPCInstrInfo *TII; protected: bool processBlock(MachineBasicBlock &MBB) { bool Changed = false; bool NeedFence = true; const PPCSubtarget &Subtarget = MBB.getParent()->getSubtarget(); bool Is64Bit = Subtarget.isPPC64(); bool IsAIX = Subtarget.isAIXABI(); bool IsLargeModel = Subtarget.getTargetMachine().getCodeModel() == CodeModel::Large; bool IsPCREL = false; MachineFunction *MF = MBB.getParent(); MachineRegisterInfo &RegInfo = MF->getRegInfo(); for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end(); I != IE;) { MachineInstr &MI = *I; IsPCREL = isPCREL(MI); // There are a number of slight differences in code generation // when we call .__get_tpointer (32-bit AIX TLS). bool IsTLSTPRelMI = MI.getOpcode() == PPC::GETtlsTpointer32AIX; bool IsTLSLDAIXMI = (MI.getOpcode() == PPC::TLSLDAIX8 || MI.getOpcode() == PPC::TLSLDAIX); if (MI.getOpcode() != PPC::ADDItlsgdLADDR && MI.getOpcode() != PPC::ADDItlsldLADDR && MI.getOpcode() != PPC::ADDItlsgdLADDR32 && MI.getOpcode() != PPC::ADDItlsldLADDR32 && MI.getOpcode() != PPC::TLSGDAIX && MI.getOpcode() != PPC::TLSGDAIX8 && !IsTLSTPRelMI && !IsPCREL && !IsTLSLDAIXMI) { // Although we create ADJCALLSTACKDOWN and ADJCALLSTACKUP // as scheduling fences, we skip creating fences if we already // have existing ADJCALLSTACKDOWN/UP to avoid nesting, // which causes verification error with -verify-machineinstrs. if (MI.getOpcode() == PPC::ADJCALLSTACKDOWN) NeedFence = false; else if (MI.getOpcode() == PPC::ADJCALLSTACKUP) NeedFence = true; ++I; continue; } LLVM_DEBUG(dbgs() << "TLS Dynamic Call Fixup:\n " << MI); Register OutReg = MI.getOperand(0).getReg(); Register InReg = PPC::NoRegister; Register GPR3 = Is64Bit ? PPC::X3 : PPC::R3; Register GPR4 = Is64Bit ? PPC::X4 : PPC::R4; if (!IsPCREL && !IsTLSTPRelMI) InReg = MI.getOperand(1).getReg(); DebugLoc DL = MI.getDebugLoc(); unsigned Opc1, Opc2; switch (MI.getOpcode()) { default: llvm_unreachable("Opcode inconsistency error"); case PPC::ADDItlsgdLADDR: Opc1 = PPC::ADDItlsgdL; Opc2 = PPC::GETtlsADDR; break; case PPC::ADDItlsldLADDR: Opc1 = PPC::ADDItlsldL; Opc2 = PPC::GETtlsldADDR; break; case PPC::ADDItlsgdLADDR32: Opc1 = PPC::ADDItlsgdL32; Opc2 = PPC::GETtlsADDR32; break; case PPC::ADDItlsldLADDR32: Opc1 = PPC::ADDItlsldL32; Opc2 = PPC::GETtlsldADDR32; break; case PPC::TLSLDAIX: // TLSLDAIX is expanded to one copy and GET_TLS_MOD, so we only set // Opc2 here. Opc2 = PPC::GETtlsMOD32AIX; break; case PPC::TLSLDAIX8: // TLSLDAIX8 is expanded to one copy and GET_TLS_MOD, so we only set // Opc2 here. Opc2 = PPC::GETtlsMOD64AIX; break; case PPC::TLSGDAIX8: // TLSGDAIX8 is expanded to two copies and GET_TLS_ADDR, so we only // set Opc2 here. Opc2 = PPC::GETtlsADDR64AIX; break; case PPC::TLSGDAIX: // TLSGDAIX is expanded to two copies and GET_TLS_ADDR, so we only // set Opc2 here. Opc2 = PPC::GETtlsADDR32AIX; break; case PPC::GETtlsTpointer32AIX: // GETtlsTpointer32AIX is expanded to a call to GET_TPOINTER on AIX // 32-bit mode within PPCAsmPrinter. This instruction does not need // to change, so Opc2 is set to the same instruction opcode. Opc2 = PPC::GETtlsTpointer32AIX; break; case PPC::PADDI8pc: assert(IsPCREL && "Expecting General/Local Dynamic PCRel"); Opc1 = PPC::PADDI8pc; Opc2 = MI.getOperand(2).getTargetFlags() == PPCII::MO_GOT_TLSGD_PCREL_FLAG ? PPC::GETtlsADDRPCREL : PPC::GETtlsldADDRPCREL; } // We create ADJCALLSTACKUP and ADJCALLSTACKDOWN around _tls_get_addr // as scheduling fence to avoid it is scheduled before // mflr in the prologue and the address in LR is clobbered (PR25839). // We don't really need to save data to the stack - the clobbered // registers are already saved when the SDNode (e.g. PPCaddiTlsgdLAddr) // gets translated to the pseudo instruction (e.g. ADDItlsgdLADDR). if (NeedFence) { MBB.getParent()->getFrameInfo().setAdjustsStack(true); BuildMI(MBB, I, DL, TII->get(PPC::ADJCALLSTACKDOWN)).addImm(0) .addImm(0); } if (IsAIX) { if (IsTLSLDAIXMI) { // The relative order between the node that loads the variable // offset from the TOC, and the .__tls_get_mod node is being tuned // here. It is better to put the variable offset TOC load after the // call, since this node can use clobbers r4/r5. // Search for the pattern of the two nodes that load from the TOC // (either for the variable offset or for the module handle), and // then move the variable offset TOC load right before the node that // uses the OutReg of the .__tls_get_mod node. unsigned LDTocOp = Is64Bit ? (IsLargeModel ? PPC::LDtocL : PPC::LDtoc) : (IsLargeModel ? PPC::LWZtocL : PPC::LWZtoc); if (!RegInfo.use_empty(OutReg)) { std::set Uses; // Collect all instructions that use the OutReg. for (MachineOperand &MO : RegInfo.use_operands(OutReg)) Uses.insert(MO.getParent()); // Find the first user (e.g.: lwax/stfdx) of the OutReg within the // current BB. MachineBasicBlock::iterator UseIter = MBB.begin(); for (MachineBasicBlock::iterator IE = MBB.end(); UseIter != IE; ++UseIter) if (Uses.count(&*UseIter)) break; // Additional handling is required when UserIter (the first user // of OutReg) is pointing to a valid node that loads from the TOC. // Check the pattern and do the movement if the pattern matches. if (UseIter != MBB.end()) { // Collect all associated nodes that load from the TOC. Use // hasOneDef() to guard against unexpected scenarios. std::set LoadFromTocs; for (MachineOperand &MO : UseIter->operands()) if (MO.isReg() && MO.isUse()) { Register MOReg = MO.getReg(); if (RegInfo.hasOneDef(MOReg)) { MachineInstr *Temp = RegInfo.getOneDef(MOReg)->getParent(); // For the current TLSLDAIX node, get the corresponding // node that loads from the TOC for the InReg. Otherwise, // Temp probably pointed to the variable offset TOC load // we would like to move. if (Temp == &MI && RegInfo.hasOneDef(InReg)) Temp = RegInfo.getOneDef(InReg)->getParent(); if (Temp->getOpcode() == LDTocOp) LoadFromTocs.insert(Temp); } else { // FIXME: analyze this scenario if there is one. LoadFromTocs.clear(); break; } } // Check the two nodes that loaded from the TOC: one should be // "_$TLSML", and the other will be moved before the node that // uses the OutReg of the .__tls_get_mod node. if (LoadFromTocs.size() == 2) { MachineBasicBlock::iterator TLSMLIter = MBB.end(); MachineBasicBlock::iterator OffsetIter = MBB.end(); // Make sure the two nodes that loaded from the TOC are within // the current BB, and that one of them is from the "_$TLSML" // pseudo symbol, while the other is from the variable. for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end(); I != IE; ++I) if (LoadFromTocs.count(&*I)) { MachineOperand MO = I->getOperand(1); if (MO.isGlobal() && MO.getGlobal()->hasName() && MO.getGlobal()->getName() == "_$TLSML") TLSMLIter = I; else OffsetIter = I; } // Perform the movement when the desired scenario has been // identified, which should be when both of the iterators are // valid. if (TLSMLIter != MBB.end() && OffsetIter != MBB.end()) OffsetIter->moveBefore(&*UseIter); } } } // The module-handle is copied into r3. The copy is followed by // GETtlsMOD32AIX/GETtlsMOD64AIX. BuildMI(MBB, I, DL, TII->get(TargetOpcode::COPY), GPR3) .addReg(InReg); // The call to .__tls_get_mod. BuildMI(MBB, I, DL, TII->get(Opc2), GPR3).addReg(GPR3); } else if (!IsTLSTPRelMI) { // The variable offset and region handle (for TLSGD) are copied in // r4 and r3. The copies are followed by // GETtlsADDR32AIX/GETtlsADDR64AIX. BuildMI(MBB, I, DL, TII->get(TargetOpcode::COPY), GPR4) .addReg(MI.getOperand(1).getReg()); BuildMI(MBB, I, DL, TII->get(TargetOpcode::COPY), GPR3) .addReg(MI.getOperand(2).getReg()); BuildMI(MBB, I, DL, TII->get(Opc2), GPR3).addReg(GPR3).addReg(GPR4); } else // The opcode of GETtlsTpointer32AIX does not change, because later // this instruction will be expanded into a call to .__get_tpointer, // which will return the thread pointer into r3. BuildMI(MBB, I, DL, TII->get(Opc2), GPR3); } else { MachineInstr *Addi; if (IsPCREL) { Addi = BuildMI(MBB, I, DL, TII->get(Opc1), GPR3).addImm(0); } else { // Expand into two ops built prior to the existing instruction. assert(InReg != PPC::NoRegister && "Operand must be a register"); Addi = BuildMI(MBB, I, DL, TII->get(Opc1), GPR3).addReg(InReg); } Addi->addOperand(MI.getOperand(2)); MachineInstr *Call = (BuildMI(MBB, I, DL, TII->get(Opc2), GPR3).addReg(GPR3)); if (IsPCREL) Call->addOperand(MI.getOperand(2)); else Call->addOperand(MI.getOperand(3)); } if (NeedFence) BuildMI(MBB, I, DL, TII->get(PPC::ADJCALLSTACKUP)).addImm(0).addImm(0); BuildMI(MBB, I, DL, TII->get(TargetOpcode::COPY), OutReg) .addReg(GPR3); // Move past the original instruction and remove it. ++I; MI.removeFromParent(); Changed = true; } return Changed; } public: bool isPCREL(const MachineInstr &MI) { return (MI.getOpcode() == PPC::PADDI8pc) && (MI.getOperand(2).getTargetFlags() == PPCII::MO_GOT_TLSGD_PCREL_FLAG || MI.getOperand(2).getTargetFlags() == PPCII::MO_GOT_TLSLD_PCREL_FLAG); } bool runOnMachineFunction(MachineFunction &MF) override { TII = MF.getSubtarget().getInstrInfo(); bool Changed = false; for (MachineBasicBlock &B : llvm::make_early_inc_range(MF)) if (processBlock(B)) Changed = true; return Changed; } void getAnalysisUsage(AnalysisUsage &AU) const override { AU.addRequired(); AU.addRequired(); MachineFunctionPass::getAnalysisUsage(AU); } }; } INITIALIZE_PASS_BEGIN(PPCTLSDynamicCall, DEBUG_TYPE, "PowerPC TLS Dynamic Call Fixup", false, false) INITIALIZE_PASS_DEPENDENCY(LiveIntervalsWrapperPass) INITIALIZE_PASS_DEPENDENCY(SlotIndexesWrapperPass) INITIALIZE_PASS_END(PPCTLSDynamicCall, DEBUG_TYPE, "PowerPC TLS Dynamic Call Fixup", false, false) char PPCTLSDynamicCall::ID = 0; FunctionPass* llvm::createPPCTLSDynamicCallPass() { return new PPCTLSDynamicCall(); }