1 //===-- AArch64CleanupLocalDynamicTLSPass.cpp ---------------------*- C++ -*-=//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // Local-dynamic access to thread-local variables proceeds in three stages.
11 //
12 // 1. The offset of this Module's thread-local area from TPIDR_EL0 is calculated
13 //    in much the same way as a general-dynamic TLS-descriptor access against
14 //    the special symbol _TLS_MODULE_BASE.
15 // 2. The variable's offset from _TLS_MODULE_BASE_ is calculated using
16 //    instructions with "dtprel" modifiers.
17 // 3. These two are added, together with TPIDR_EL0, to obtain the variable's
18 //    true address.
19 //
20 // This is only better than general-dynamic access to the variable if two or
21 // more of the first stage TLS-descriptor calculations can be combined. This
22 // pass looks through a function and performs such combinations.
23 //
24 //===----------------------------------------------------------------------===//
25 #include "AArch64.h"
26 #include "AArch64InstrInfo.h"
27 #include "AArch64MachineFunctionInfo.h"
28 #include "llvm/CodeGen/MachineDominators.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineFunctionPass.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 using namespace llvm;
34 
35 #define TLSCLEANUP_PASS_NAME "AArch64 Local Dynamic TLS Access Clean-up"
36 
37 namespace {
38 struct LDTLSCleanup : public MachineFunctionPass {
39   static char ID;
40   LDTLSCleanup() : MachineFunctionPass(ID) {
41     initializeLDTLSCleanupPass(*PassRegistry::getPassRegistry());
42   }
43 
44   bool runOnMachineFunction(MachineFunction &MF) override {
45     if (skipFunction(MF.getFunction()))
46       return false;
47 
48     AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
49     if (AFI->getNumLocalDynamicTLSAccesses() < 2) {
50       // No point folding accesses if there isn't at least two.
51       return false;
52     }
53 
54     MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
55     return VisitNode(DT->getRootNode(), 0);
56   }
57 
58   // Visit the dominator subtree rooted at Node in pre-order.
59   // If TLSBaseAddrReg is non-null, then use that to replace any
60   // TLS_base_addr instructions. Otherwise, create the register
61   // when the first such instruction is seen, and then use it
62   // as we encounter more instructions.
63   bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
64     MachineBasicBlock *BB = Node->getBlock();
65     bool Changed = false;
66 
67     // Traverse the current block.
68     for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
69          ++I) {
70       switch (I->getOpcode()) {
71       case AArch64::TLSDESC_CALLSEQ:
72         // Make sure it's a local dynamic access.
73         if (!I->getOperand(0).isSymbol() ||
74             strcmp(I->getOperand(0).getSymbolName(), "_TLS_MODULE_BASE_"))
75           break;
76 
77         if (TLSBaseAddrReg)
78           I = replaceTLSBaseAddrCall(*I, TLSBaseAddrReg);
79         else
80           I = setRegister(*I, &TLSBaseAddrReg);
81         Changed = true;
82         break;
83       default:
84         break;
85       }
86     }
87 
88     // Visit the children of this block in the dominator tree.
89     for (MachineDomTreeNode *N : *Node) {
90       Changed |= VisitNode(N, TLSBaseAddrReg);
91     }
92 
93     return Changed;
94   }
95 
96   // Replace the TLS_base_addr instruction I with a copy from
97   // TLSBaseAddrReg, returning the new instruction.
98   MachineInstr *replaceTLSBaseAddrCall(MachineInstr &I,
99                                        unsigned TLSBaseAddrReg) {
100     MachineFunction *MF = I.getParent()->getParent();
101     const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
102 
103     // Insert a Copy from TLSBaseAddrReg to x0, which is where the rest of the
104     // code sequence assumes the address will be.
105     MachineInstr *Copy = BuildMI(*I.getParent(), I, I.getDebugLoc(),
106                                  TII->get(TargetOpcode::COPY), AArch64::X0)
107                              .addReg(TLSBaseAddrReg);
108 
109     // Erase the TLS_base_addr instruction.
110     I.eraseFromParent();
111 
112     return Copy;
113   }
114 
115   // Create a virtual register in *TLSBaseAddrReg, and populate it by
116   // inserting a copy instruction after I. Returns the new instruction.
117   MachineInstr *setRegister(MachineInstr &I, unsigned *TLSBaseAddrReg) {
118     MachineFunction *MF = I.getParent()->getParent();
119     const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
120 
121     // Create a virtual register for the TLS base address.
122     MachineRegisterInfo &RegInfo = MF->getRegInfo();
123     *TLSBaseAddrReg = RegInfo.createVirtualRegister(&AArch64::GPR64RegClass);
124 
125     // Insert a copy from X0 to TLSBaseAddrReg for later.
126     MachineInstr *Copy =
127         BuildMI(*I.getParent(), ++I.getIterator(), I.getDebugLoc(),
128                 TII->get(TargetOpcode::COPY), *TLSBaseAddrReg)
129             .addReg(AArch64::X0);
130 
131     return Copy;
132   }
133 
134   StringRef getPassName() const override { return TLSCLEANUP_PASS_NAME; }
135 
136   void getAnalysisUsage(AnalysisUsage &AU) const override {
137     AU.setPreservesCFG();
138     AU.addRequired<MachineDominatorTree>();
139     MachineFunctionPass::getAnalysisUsage(AU);
140   }
141 };
142 }
143 
144 INITIALIZE_PASS(LDTLSCleanup, "aarch64-local-dynamic-tls-cleanup",
145                 TLSCLEANUP_PASS_NAME, false, false)
146 
147 char LDTLSCleanup::ID = 0;
148 FunctionPass *llvm::createAArch64CleanupLocalDynamicTLSPass() {
149   return new LDTLSCleanup();
150 }
151