1 //======- X86RetpolineThunks.cpp - Construct retpoline thunks for x86 --=====//
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
7 //===----------------------------------------------------------------------===//
10 /// Pass that injects an MI thunk implementing a "retpoline". This is
11 /// a RET-implemented trampoline that is used to lower indirect calls in a way
12 /// that prevents speculation on some x86 processors and can be used to mitigate
13 /// security vulnerabilities due to targeted speculative execution and side
14 /// channels such as CVE-2017-5715.
16 /// TODO(chandlerc): All of this code could use better comments and
19 //===----------------------------------------------------------------------===//
22 #include "X86InstrBuilder.h"
23 #include "X86Subtarget.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineModuleInfo.h"
27 #include "llvm/CodeGen/Passes.h"
28 #include "llvm/CodeGen/TargetPassConfig.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/Module.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/raw_ostream.h"
38 #define DEBUG_TYPE "x86-retpoline-thunks"
40 static const char ThunkNamePrefix
[] = "__llvm_retpoline_";
41 static const char R11ThunkName
[] = "__llvm_retpoline_r11";
42 static const char EAXThunkName
[] = "__llvm_retpoline_eax";
43 static const char ECXThunkName
[] = "__llvm_retpoline_ecx";
44 static const char EDXThunkName
[] = "__llvm_retpoline_edx";
45 static const char EDIThunkName
[] = "__llvm_retpoline_edi";
48 class X86RetpolineThunks
: public MachineFunctionPass
{
52 X86RetpolineThunks() : MachineFunctionPass(ID
) {}
54 StringRef
getPassName() const override
{ return "X86 Retpoline Thunks"; }
56 bool doInitialization(Module
&M
) override
;
57 bool runOnMachineFunction(MachineFunction
&F
) override
;
59 void getAnalysisUsage(AnalysisUsage
&AU
) const override
{
60 MachineFunctionPass::getAnalysisUsage(AU
);
61 AU
.addRequired
<MachineModuleInfo
>();
62 AU
.addPreserved
<MachineModuleInfo
>();
66 MachineModuleInfo
*MMI
;
67 const TargetMachine
*TM
;
69 const X86Subtarget
*STI
;
70 const X86InstrInfo
*TII
;
74 void createThunkFunction(Module
&M
, StringRef Name
);
75 void insertRegReturnAddrClobber(MachineBasicBlock
&MBB
, unsigned Reg
);
76 void populateThunk(MachineFunction
&MF
, unsigned Reg
);
79 } // end anonymous namespace
81 FunctionPass
*llvm::createX86RetpolineThunksPass() {
82 return new X86RetpolineThunks();
85 char X86RetpolineThunks::ID
= 0;
87 bool X86RetpolineThunks::doInitialization(Module
&M
) {
88 InsertedThunks
= false;
92 bool X86RetpolineThunks::runOnMachineFunction(MachineFunction
&MF
) {
93 LLVM_DEBUG(dbgs() << getPassName() << '\n');
95 TM
= &MF
.getTarget();;
96 STI
= &MF
.getSubtarget
<X86Subtarget
>();
97 TII
= STI
->getInstrInfo();
98 Is64Bit
= TM
->getTargetTriple().getArch() == Triple::x86_64
;
100 MMI
= &getAnalysis
<MachineModuleInfo
>();
101 Module
&M
= const_cast<Module
&>(*MMI
->getModule());
103 // If this function is not a thunk, check to see if we need to insert
105 if (!MF
.getName().startswith(ThunkNamePrefix
)) {
106 // If we've already inserted a thunk, nothing else to do.
110 // Only add a thunk if one of the functions has the retpoline feature
111 // enabled in its subtarget, and doesn't enable external thunks.
112 // FIXME: Conditionalize on indirect calls so we don't emit a thunk when
113 // nothing will end up calling it.
114 // FIXME: It's a little silly to look at every function just to enumerate
115 // the subtargets, but eventually we'll want to look at them for indirect
116 // calls, so maybe this is OK.
117 if ((!STI
->useRetpolineIndirectCalls() &&
118 !STI
->useRetpolineIndirectBranches()) ||
119 STI
->useRetpolineExternalThunk())
122 // Otherwise, we need to insert the thunk.
123 // WARNING: This is not really a well behaving thing to do in a function
124 // pass. We extract the module and insert a new function (and machine
125 // function) directly into the module.
127 createThunkFunction(M
, R11ThunkName
);
129 for (StringRef Name
:
130 {EAXThunkName
, ECXThunkName
, EDXThunkName
, EDIThunkName
})
131 createThunkFunction(M
, Name
);
132 InsertedThunks
= true;
136 // If this *is* a thunk function, we need to populate it with the correct MI.
138 assert(MF
.getName() == "__llvm_retpoline_r11" &&
139 "Should only have an r11 thunk on 64-bit targets");
141 // __llvm_retpoline_r11:
142 // callq .Lr11_call_target
143 // .Lr11_capture_spec:
146 // jmp .Lr11_capture_spec
148 // .Lr11_call_target:
151 populateThunk(MF
, X86::R11
);
153 // For 32-bit targets we need to emit a collection of thunks for various
154 // possible scratch registers as well as a fallback that uses EDI, which is
155 // normally callee saved.
156 // __llvm_retpoline_eax:
157 // calll .Leax_call_target
158 // .Leax_capture_spec:
160 // jmp .Leax_capture_spec
162 // .Leax_call_target:
163 // movl %eax, (%esp) # Clobber return addr
166 // __llvm_retpoline_ecx:
171 // __llvm_retpoline_edx:
176 // __llvm_retpoline_edi:
180 if (MF
.getName() == EAXThunkName
)
181 populateThunk(MF
, X86::EAX
);
182 else if (MF
.getName() == ECXThunkName
)
183 populateThunk(MF
, X86::ECX
);
184 else if (MF
.getName() == EDXThunkName
)
185 populateThunk(MF
, X86::EDX
);
186 else if (MF
.getName() == EDIThunkName
)
187 populateThunk(MF
, X86::EDI
);
189 llvm_unreachable("Invalid thunk name on x86-32!");
195 void X86RetpolineThunks::createThunkFunction(Module
&M
, StringRef Name
) {
196 assert(Name
.startswith(ThunkNamePrefix
) &&
197 "Created a thunk with an unexpected prefix!");
199 LLVMContext
&Ctx
= M
.getContext();
200 auto Type
= FunctionType::get(Type::getVoidTy(Ctx
), false);
202 Function::Create(Type
, GlobalValue::LinkOnceODRLinkage
, Name
, &M
);
203 F
->setVisibility(GlobalValue::HiddenVisibility
);
204 F
->setComdat(M
.getOrInsertComdat(Name
));
206 // Add Attributes so that we don't create a frame, unwind information, or
209 B
.addAttribute(llvm::Attribute::NoUnwind
);
210 B
.addAttribute(llvm::Attribute::Naked
);
211 F
->addAttributes(llvm::AttributeList::FunctionIndex
, B
);
213 // Populate our function a bit so that we can verify.
214 BasicBlock
*Entry
= BasicBlock::Create(Ctx
, "entry", F
);
215 IRBuilder
<> Builder(Entry
);
217 Builder
.CreateRetVoid();
219 // MachineFunctions/MachineBasicBlocks aren't created automatically for the
220 // IR-level constructs we already made. Create them and insert them into the
222 MachineFunction
&MF
= MMI
->getOrCreateMachineFunction(*F
);
223 MachineBasicBlock
*EntryMBB
= MF
.CreateMachineBasicBlock(Entry
);
225 // Insert EntryMBB into MF. It's not in the module until we do this.
226 MF
.insert(MF
.end(), EntryMBB
);
229 void X86RetpolineThunks::insertRegReturnAddrClobber(MachineBasicBlock
&MBB
,
231 const unsigned MovOpc
= Is64Bit
? X86::MOV64mr
: X86::MOV32mr
;
232 const unsigned SPReg
= Is64Bit
? X86::RSP
: X86::ESP
;
233 addRegOffset(BuildMI(&MBB
, DebugLoc(), TII
->get(MovOpc
)), SPReg
, false, 0)
237 void X86RetpolineThunks::populateThunk(MachineFunction
&MF
,
239 // Set MF properties. We never use vregs...
240 MF
.getProperties().set(MachineFunctionProperties::Property::NoVRegs
);
242 // Grab the entry MBB and erase any other blocks. O0 codegen appears to
243 // generate two bbs for the entry block.
244 MachineBasicBlock
*Entry
= &MF
.front();
246 while (MF
.size() > 1)
247 MF
.erase(std::next(MF
.begin()));
249 MachineBasicBlock
*CaptureSpec
= MF
.CreateMachineBasicBlock(Entry
->getBasicBlock());
250 MachineBasicBlock
*CallTarget
= MF
.CreateMachineBasicBlock(Entry
->getBasicBlock());
251 MCSymbol
*TargetSym
= MF
.getContext().createTempSymbol();
252 MF
.push_back(CaptureSpec
);
253 MF
.push_back(CallTarget
);
255 const unsigned CallOpc
= Is64Bit
? X86::CALL64pcrel32
: X86::CALLpcrel32
;
256 const unsigned RetOpc
= Is64Bit
? X86::RETQ
: X86::RETL
;
258 Entry
->addLiveIn(Reg
);
259 BuildMI(Entry
, DebugLoc(), TII
->get(CallOpc
)).addSym(TargetSym
);
261 // The MIR verifier thinks that the CALL in the entry block will fall through
262 // to CaptureSpec, so mark it as the successor. Technically, CaptureTarget is
263 // the successor, but the MIR verifier doesn't know how to cope with that.
264 Entry
->addSuccessor(CaptureSpec
);
266 // In the capture loop for speculation, we want to stop the processor from
267 // speculating as fast as possible. On Intel processors, the PAUSE instruction
268 // will block speculation without consuming any execution resources. On AMD
269 // processors, the PAUSE instruction is (essentially) a nop, so we also use an
270 // LFENCE instruction which they have advised will stop speculation as well
271 // with minimal resource utilization. We still end the capture with a jump to
272 // form an infinite loop to fully guarantee that no matter what implementation
273 // of the x86 ISA, speculating this code path never escapes.
274 BuildMI(CaptureSpec
, DebugLoc(), TII
->get(X86::PAUSE
));
275 BuildMI(CaptureSpec
, DebugLoc(), TII
->get(X86::LFENCE
));
276 BuildMI(CaptureSpec
, DebugLoc(), TII
->get(X86::JMP_1
)).addMBB(CaptureSpec
);
277 CaptureSpec
->setHasAddressTaken();
278 CaptureSpec
->addSuccessor(CaptureSpec
);
280 CallTarget
->addLiveIn(Reg
);
281 CallTarget
->setHasAddressTaken();
282 CallTarget
->setAlignment(4);
283 insertRegReturnAddrClobber(*CallTarget
, Reg
);
284 CallTarget
->back().setPreInstrSymbol(MF
, TargetSym
);
285 BuildMI(CallTarget
, DebugLoc(), TII
->get(RetOpc
));