1 | //===----- X86AvoidTrailingCall.cpp - Insert int3 after trailing calls ----===// |
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 | // The Windows x64 unwinder decodes the instruction stream during unwinding. |
10 | // The unwinder decodes forward from the current PC to detect epilogue code |
11 | // patterns. |
12 | // |
13 | // First, this means that there must be an instruction after every |
14 | // call instruction for the unwinder to decode. LLVM must maintain the invariant |
15 | // that the last instruction of a function or funclet is not a call, or the |
16 | // unwinder may decode into the next function. Similarly, a call may not |
17 | // immediately precede an epilogue code pattern. As of this writing, the |
18 | // SEH_Epilogue pseudo instruction takes care of that. |
19 | // |
20 | // Second, all non-tail call jump targets must be within the *half-open* |
21 | // interval of the bounds of the function. The unwinder distinguishes between |
22 | // internal jump instructions and tail calls in an epilogue sequence by checking |
23 | // the jump target against the function bounds from the .pdata section. This |
24 | // means that the last regular MBB of an LLVM function must not be empty if |
25 | // there are regular jumps targeting it. |
26 | // |
27 | // This pass upholds these invariants by ensuring that blocks at the end of a |
28 | // function or funclet are a) not empty and b) do not end in a CALL instruction. |
29 | // |
30 | // Unwinder implementation for reference: |
31 | // https://github.com/dotnet/coreclr/blob/a9f3fc16483eecfc47fb79c362811d870be02249/src/unwinder/amd64/unwinder_amd64.cpp#L1015 |
32 | // |
33 | //===----------------------------------------------------------------------===// |
34 | |
35 | #include "X86.h" |
36 | #include "X86InstrInfo.h" |
37 | #include "X86Subtarget.h" |
38 | #include "llvm/CodeGen/MachineFunctionPass.h" |
39 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
40 | |
41 | #define AVOIDCALL_DESC "X86 avoid trailing call pass" |
42 | #define AVOIDCALL_NAME "x86-avoid-trailing-call" |
43 | |
44 | #define DEBUG_TYPE AVOIDCALL_NAME |
45 | |
46 | using namespace llvm; |
47 | |
48 | namespace { |
49 | class X86AvoidTrailingCallPass : public MachineFunctionPass { |
50 | public: |
51 | X86AvoidTrailingCallPass() : MachineFunctionPass(ID) {} |
52 | |
53 | bool runOnMachineFunction(MachineFunction &MF) override; |
54 | |
55 | static char ID; |
56 | |
57 | private: |
58 | StringRef getPassName() const override { return AVOIDCALL_DESC; } |
59 | }; |
60 | } // end anonymous namespace |
61 | |
62 | char X86AvoidTrailingCallPass::ID = 0; |
63 | |
64 | FunctionPass *llvm::createX86AvoidTrailingCallPass() { |
65 | return new X86AvoidTrailingCallPass(); |
66 | } |
67 | |
68 | INITIALIZE_PASS(X86AvoidTrailingCallPass, AVOIDCALL_NAME, AVOIDCALL_DESC, false, false) |
69 | |
70 | // A real instruction is a non-meta, non-pseudo instruction. Some pseudos |
71 | // expand to nothing, and some expand to code. This logic conservatively assumes |
72 | // they might expand to nothing. |
73 | static bool isCallOrRealInstruction(MachineInstr &MI) { |
74 | return MI.isCall() || (!MI.isPseudo() && !MI.isMetaInstruction()); |
75 | } |
76 | |
77 | // Return true if this is a call instruction, but not a tail call. |
78 | static bool isCallInstruction(const MachineInstr &MI) { |
79 | return MI.isCall() && !MI.isReturn(); |
80 | } |
81 | |
82 | bool X86AvoidTrailingCallPass::runOnMachineFunction(MachineFunction &MF) { |
83 | const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>(); |
84 | const X86InstrInfo &TII = *STI.getInstrInfo(); |
85 | assert(STI.isTargetWin64() && "pass only runs on Win64" ); |
86 | |
87 | // We don't need to worry about any of the invariants described above if there |
88 | // is no unwind info (CFI). |
89 | if (!MF.hasWinCFI()) |
90 | return false; |
91 | |
92 | // FIXME: Perhaps this pass should also replace SEH_Epilogue by inserting nops |
93 | // before epilogues. |
94 | |
95 | bool Changed = false; |
96 | for (MachineBasicBlock &MBB : MF) { |
97 | // Look for basic blocks that precede funclet entries or are at the end of |
98 | // the function. |
99 | MachineBasicBlock *NextMBB = MBB.getNextNode(); |
100 | if (NextMBB && !NextMBB->isEHFuncletEntry()) |
101 | continue; |
102 | |
103 | // Find the last real instruction in this block. |
104 | auto LastRealInstr = llvm::find_if(Range: reverse(C&: MBB), P: isCallOrRealInstruction); |
105 | |
106 | // If the block is empty or the last real instruction is a call instruction, |
107 | // insert an int3. If there is a call instruction, insert the int3 between |
108 | // the call and any labels or other meta instructions. If the block is |
109 | // empty, insert at block end. |
110 | bool IsEmpty = LastRealInstr == MBB.rend(); |
111 | bool IsCall = !IsEmpty && isCallInstruction(MI: *LastRealInstr); |
112 | if (IsEmpty || IsCall) { |
113 | LLVM_DEBUG({ |
114 | if (IsCall) { |
115 | dbgs() << "inserting int3 after trailing call instruction:\n" ; |
116 | LastRealInstr->dump(); |
117 | dbgs() << '\n'; |
118 | } else { |
119 | dbgs() << "inserting int3 in trailing empty MBB:\n" ; |
120 | MBB.dump(); |
121 | } |
122 | }); |
123 | |
124 | MachineBasicBlock::iterator MBBI = MBB.end(); |
125 | DebugLoc DL; |
126 | if (IsCall) { |
127 | MBBI = std::next(x: LastRealInstr.getReverse()); |
128 | DL = LastRealInstr->getDebugLoc(); |
129 | } |
130 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII.get(Opcode: X86::INT3)); |
131 | Changed = true; |
132 | } |
133 | } |
134 | |
135 | return Changed; |
136 | } |
137 | |