| 1 | //===- XRayInstrumentation.cpp - Adds XRay instrumentation to functions. --===// |
| 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 file implements a MachineFunctionPass that inserts the appropriate |
| 10 | // XRay instrumentation instructions. We look for XRay-specific attributes |
| 11 | // on the function to determine whether we should insert the replacement |
| 12 | // operations. |
| 13 | // |
| 14 | //===---------------------------------------------------------------------===// |
| 15 | |
| 16 | #include "llvm/CodeGen/XRayInstrumentation.h" |
| 17 | #include "llvm/ADT/STLExtras.h" |
| 18 | #include "llvm/ADT/SmallVector.h" |
| 19 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 20 | #include "llvm/CodeGen/MachineDominators.h" |
| 21 | #include "llvm/CodeGen/MachineFunction.h" |
| 22 | #include "llvm/CodeGen/MachineFunctionAnalysis.h" |
| 23 | #include "llvm/CodeGen/MachineFunctionPass.h" |
| 24 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 25 | #include "llvm/CodeGen/MachineLoopInfo.h" |
| 26 | #include "llvm/CodeGen/MachinePassManager.h" |
| 27 | #include "llvm/CodeGen/TargetInstrInfo.h" |
| 28 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
| 29 | #include "llvm/IR/Attributes.h" |
| 30 | #include "llvm/IR/DiagnosticInfo.h" |
| 31 | #include "llvm/IR/Function.h" |
| 32 | #include "llvm/InitializePasses.h" |
| 33 | #include "llvm/Pass.h" |
| 34 | #include "llvm/Target/TargetMachine.h" |
| 35 | #include "llvm/TargetParser/Triple.h" |
| 36 | |
| 37 | using namespace llvm; |
| 38 | |
| 39 | namespace { |
| 40 | |
| 41 | struct InstrumentationOptions { |
| 42 | // Whether to emit PATCHABLE_TAIL_CALL. |
| 43 | bool HandleTailcall; |
| 44 | |
| 45 | // Whether to emit PATCHABLE_RET/PATCHABLE_FUNCTION_EXIT for all forms of |
| 46 | // return, e.g. conditional return. |
| 47 | bool HandleAllReturns; |
| 48 | }; |
| 49 | |
| 50 | struct XRayInstrumentationLegacy : public MachineFunctionPass { |
| 51 | static char ID; |
| 52 | |
| 53 | XRayInstrumentationLegacy() : MachineFunctionPass(ID) { |
| 54 | initializeXRayInstrumentationLegacyPass(*PassRegistry::getPassRegistry()); |
| 55 | } |
| 56 | |
| 57 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 58 | AU.setPreservesCFG(); |
| 59 | AU.addPreserved<MachineLoopInfoWrapperPass>(); |
| 60 | AU.addPreserved<MachineDominatorTreeWrapperPass>(); |
| 61 | MachineFunctionPass::getAnalysisUsage(AU); |
| 62 | } |
| 63 | |
| 64 | bool runOnMachineFunction(MachineFunction &MF) override; |
| 65 | }; |
| 66 | |
| 67 | struct XRayInstrumentation { |
| 68 | XRayInstrumentation(MachineDominatorTree *MDT, MachineLoopInfo *MLI) |
| 69 | : MDT(MDT), MLI(MLI) {} |
| 70 | |
| 71 | bool run(MachineFunction &MF); |
| 72 | |
| 73 | // Methods for use in the NPM and legacy passes, can be removed once migration |
| 74 | // is complete. |
| 75 | static bool alwaysInstrument(Function &F) { |
| 76 | auto InstrAttr = F.getFnAttribute(Kind: "function-instrument" ); |
| 77 | return InstrAttr.isStringAttribute() && |
| 78 | InstrAttr.getValueAsString() == "xray-always" ; |
| 79 | } |
| 80 | |
| 81 | static bool needMDTAndMLIAnalyses(Function &F) { |
| 82 | auto IgnoreLoopsAttr = F.getFnAttribute(Kind: "xray-ignore-loops" ); |
| 83 | auto AlwaysInstrument = XRayInstrumentation::alwaysInstrument(F); |
| 84 | return !AlwaysInstrument && !IgnoreLoopsAttr.isValid(); |
| 85 | } |
| 86 | |
| 87 | private: |
| 88 | // Replace the original RET instruction with the exit sled code ("patchable |
| 89 | // ret" pseudo-instruction), so that at runtime XRay can replace the sled |
| 90 | // with a code jumping to XRay trampoline, which calls the tracing handler |
| 91 | // and, in the end, issues the RET instruction. |
| 92 | // This is the approach to go on CPUs which have a single RET instruction, |
| 93 | // like x86/x86_64. |
| 94 | void replaceRetWithPatchableRet(MachineFunction &MF, |
| 95 | const TargetInstrInfo *TII, |
| 96 | InstrumentationOptions); |
| 97 | |
| 98 | // Prepend the original return instruction with the exit sled code ("patchable |
| 99 | // function exit" pseudo-instruction), preserving the original return |
| 100 | // instruction just after the exit sled code. |
| 101 | // This is the approach to go on CPUs which have multiple options for the |
| 102 | // return instruction, like ARM. For such CPUs we can't just jump into the |
| 103 | // XRay trampoline and issue a single return instruction there. We rather |
| 104 | // have to call the trampoline and return from it to the original return |
| 105 | // instruction of the function being instrumented. |
| 106 | void prependRetWithPatchableExit(MachineFunction &MF, |
| 107 | const TargetInstrInfo *TII, |
| 108 | InstrumentationOptions); |
| 109 | |
| 110 | MachineDominatorTree *MDT; |
| 111 | MachineLoopInfo *MLI; |
| 112 | }; |
| 113 | |
| 114 | } // end anonymous namespace |
| 115 | |
| 116 | void XRayInstrumentation::replaceRetWithPatchableRet( |
| 117 | MachineFunction &MF, const TargetInstrInfo *TII, |
| 118 | InstrumentationOptions op) { |
| 119 | // We look for *all* terminators and returns, then replace those with |
| 120 | // PATCHABLE_RET instructions. |
| 121 | SmallVector<MachineInstr *, 4> Terminators; |
| 122 | for (auto &MBB : MF) { |
| 123 | for (auto &T : MBB.terminators()) { |
| 124 | unsigned Opc = 0; |
| 125 | if (T.isReturn() && |
| 126 | (op.HandleAllReturns || T.getOpcode() == TII->getReturnOpcode())) { |
| 127 | // Replace return instructions with: |
| 128 | // PATCHABLE_RET <Opcode>, <Operand>... |
| 129 | Opc = TargetOpcode::PATCHABLE_RET; |
| 130 | } |
| 131 | if (TII->isTailCall(Inst: T) && op.HandleTailcall) { |
| 132 | // Treat the tail call as a return instruction, which has a |
| 133 | // different-looking sled than the normal return case. |
| 134 | Opc = TargetOpcode::PATCHABLE_TAIL_CALL; |
| 135 | } |
| 136 | if (Opc != 0) { |
| 137 | auto MIB = BuildMI(BB&: MBB, I&: T, MIMD: T.getDebugLoc(), MCID: TII->get(Opcode: Opc)) |
| 138 | .addImm(Val: T.getOpcode()); |
| 139 | for (auto &MO : T.operands()) |
| 140 | MIB.add(MO); |
| 141 | Terminators.push_back(Elt: &T); |
| 142 | if (T.shouldUpdateAdditionalCallInfo()) |
| 143 | MF.eraseAdditionalCallInfo(MI: &T); |
| 144 | } |
| 145 | } |
| 146 | } |
| 147 | |
| 148 | for (auto &I : Terminators) |
| 149 | I->eraseFromParent(); |
| 150 | } |
| 151 | |
| 152 | void XRayInstrumentation::prependRetWithPatchableExit( |
| 153 | MachineFunction &MF, const TargetInstrInfo *TII, |
| 154 | InstrumentationOptions op) { |
| 155 | for (auto &MBB : MF) |
| 156 | for (auto &T : MBB.terminators()) { |
| 157 | unsigned Opc = 0; |
| 158 | if (T.isReturn() && |
| 159 | (op.HandleAllReturns || T.getOpcode() == TII->getReturnOpcode())) { |
| 160 | Opc = TargetOpcode::PATCHABLE_FUNCTION_EXIT; |
| 161 | } |
| 162 | if (TII->isTailCall(Inst: T) && op.HandleTailcall) { |
| 163 | Opc = TargetOpcode::PATCHABLE_TAIL_CALL; |
| 164 | } |
| 165 | if (Opc != 0) { |
| 166 | // Prepend the return instruction with PATCHABLE_FUNCTION_EXIT or |
| 167 | // PATCHABLE_TAIL_CALL . |
| 168 | BuildMI(BB&: MBB, I&: T, MIMD: T.getDebugLoc(), MCID: TII->get(Opcode: Opc)); |
| 169 | } |
| 170 | } |
| 171 | } |
| 172 | |
| 173 | PreservedAnalyses |
| 174 | XRayInstrumentationPass::run(MachineFunction &MF, |
| 175 | MachineFunctionAnalysisManager &MFAM) { |
| 176 | MachineDominatorTree *MDT = nullptr; |
| 177 | MachineLoopInfo *MLI = nullptr; |
| 178 | |
| 179 | if (XRayInstrumentation::needMDTAndMLIAnalyses(F&: MF.getFunction())) { |
| 180 | MDT = MFAM.getCachedResult<MachineDominatorTreeAnalysis>(IR&: MF); |
| 181 | MLI = MFAM.getCachedResult<MachineLoopAnalysis>(IR&: MF); |
| 182 | } |
| 183 | |
| 184 | if (!XRayInstrumentation(MDT, MLI).run(MF)) |
| 185 | return PreservedAnalyses::all(); |
| 186 | |
| 187 | auto PA = getMachineFunctionPassPreservedAnalyses(); |
| 188 | PA.preserveSet<CFGAnalyses>(); |
| 189 | return PA; |
| 190 | } |
| 191 | |
| 192 | bool XRayInstrumentationLegacy::runOnMachineFunction(MachineFunction &MF) { |
| 193 | MachineDominatorTree *MDT = nullptr; |
| 194 | MachineLoopInfo *MLI = nullptr; |
| 195 | if (XRayInstrumentation::needMDTAndMLIAnalyses(F&: MF.getFunction())) { |
| 196 | auto *MDTWrapper = |
| 197 | getAnalysisIfAvailable<MachineDominatorTreeWrapperPass>(); |
| 198 | MDT = MDTWrapper ? &MDTWrapper->getDomTree() : nullptr; |
| 199 | auto *MLIWrapper = getAnalysisIfAvailable<MachineLoopInfoWrapperPass>(); |
| 200 | MLI = MLIWrapper ? &MLIWrapper->getLI() : nullptr; |
| 201 | } |
| 202 | return XRayInstrumentation(MDT, MLI).run(MF); |
| 203 | } |
| 204 | |
| 205 | bool XRayInstrumentation::run(MachineFunction &MF) { |
| 206 | auto &F = MF.getFunction(); |
| 207 | auto InstrAttr = F.getFnAttribute(Kind: "function-instrument" ); |
| 208 | bool AlwaysInstrument = alwaysInstrument(F); |
| 209 | bool NeverInstrument = InstrAttr.isStringAttribute() && |
| 210 | InstrAttr.getValueAsString() == "xray-never" ; |
| 211 | if (NeverInstrument && !AlwaysInstrument) |
| 212 | return false; |
| 213 | auto IgnoreLoopsAttr = F.getFnAttribute(Kind: "xray-ignore-loops" ); |
| 214 | |
| 215 | uint64_t XRayThreshold = 0; |
| 216 | if (!AlwaysInstrument) { |
| 217 | bool IgnoreLoops = IgnoreLoopsAttr.isValid(); |
| 218 | XRayThreshold = F.getFnAttributeAsParsedInteger( |
| 219 | Kind: "xray-instruction-threshold" , Default: std::numeric_limits<uint64_t>::max()); |
| 220 | if (XRayThreshold == std::numeric_limits<uint64_t>::max()) |
| 221 | return false; |
| 222 | |
| 223 | // Count the number of MachineInstr`s in MachineFunction |
| 224 | uint64_t MICount = 0; |
| 225 | for (const auto &MBB : MF) |
| 226 | MICount += MBB.size(); |
| 227 | |
| 228 | bool TooFewInstrs = MICount < XRayThreshold; |
| 229 | |
| 230 | if (!IgnoreLoops) { |
| 231 | // Get MachineDominatorTree or compute it on the fly if it's unavailable |
| 232 | MachineDominatorTree ComputedMDT; |
| 233 | if (!MDT) { |
| 234 | ComputedMDT.recalculate(Func&: MF); |
| 235 | MDT = &ComputedMDT; |
| 236 | } |
| 237 | |
| 238 | // Get MachineLoopInfo or compute it on the fly if it's unavailable |
| 239 | MachineLoopInfo ComputedMLI; |
| 240 | if (!MLI) { |
| 241 | ComputedMLI.analyze(DomTree: *MDT); |
| 242 | MLI = &ComputedMLI; |
| 243 | } |
| 244 | |
| 245 | // Check if we have a loop. |
| 246 | // FIXME: Maybe make this smarter, and see whether the loops are dependent |
| 247 | // on inputs or side-effects? |
| 248 | if (MLI->empty() && TooFewInstrs) |
| 249 | return false; // Function is too small and has no loops. |
| 250 | } else if (TooFewInstrs) { |
| 251 | // Function is too small |
| 252 | return false; |
| 253 | } |
| 254 | } |
| 255 | |
| 256 | // We look for the first non-empty MachineBasicBlock, so that we can insert |
| 257 | // the function instrumentation in the appropriate place. |
| 258 | auto MBI = llvm::find_if( |
| 259 | Range&: MF, P: [&](const MachineBasicBlock &MBB) { return !MBB.empty(); }); |
| 260 | if (MBI == MF.end()) |
| 261 | return false; // The function is empty. |
| 262 | |
| 263 | auto *TII = MF.getSubtarget().getInstrInfo(); |
| 264 | auto &FirstMBB = *MBI; |
| 265 | auto &FirstMI = *FirstMBB.begin(); |
| 266 | |
| 267 | if (!MF.getSubtarget().isXRaySupported()) { |
| 268 | |
| 269 | const Function &Fn = FirstMBB.getParent()->getFunction(); |
| 270 | Fn.getContext().diagnose(DI: DiagnosticInfoUnsupported( |
| 271 | Fn, "An attempt to perform XRay instrumentation for an" |
| 272 | " unsupported target." )); |
| 273 | |
| 274 | return false; |
| 275 | } |
| 276 | |
| 277 | if (!F.hasFnAttribute(Kind: "xray-skip-entry" )) { |
| 278 | // First, insert an PATCHABLE_FUNCTION_ENTER as the first instruction of the |
| 279 | // MachineFunction. |
| 280 | BuildMI(BB&: FirstMBB, I&: FirstMI, MIMD: FirstMI.getDebugLoc(), |
| 281 | MCID: TII->get(Opcode: TargetOpcode::PATCHABLE_FUNCTION_ENTER)); |
| 282 | } |
| 283 | |
| 284 | if (!F.hasFnAttribute(Kind: "xray-skip-exit" )) { |
| 285 | switch (MF.getTarget().getTargetTriple().getArch()) { |
| 286 | case Triple::ArchType::arm: |
| 287 | case Triple::ArchType::thumb: |
| 288 | case Triple::ArchType::aarch64: |
| 289 | case Triple::ArchType::hexagon: |
| 290 | case Triple::ArchType::loongarch64: |
| 291 | case Triple::ArchType::mips: |
| 292 | case Triple::ArchType::mipsel: |
| 293 | case Triple::ArchType::mips64: |
| 294 | case Triple::ArchType::mips64el: |
| 295 | case Triple::ArchType::riscv32: |
| 296 | case Triple::ArchType::riscv64: { |
| 297 | // For the architectures which don't have a single return instruction |
| 298 | InstrumentationOptions op; |
| 299 | // AArch64 and RISC-V support patching tail calls. |
| 300 | op.HandleTailcall = MF.getTarget().getTargetTriple().isAArch64() || |
| 301 | MF.getTarget().getTargetTriple().isRISCV(); |
| 302 | op.HandleAllReturns = true; |
| 303 | prependRetWithPatchableExit(MF, TII, op); |
| 304 | break; |
| 305 | } |
| 306 | case Triple::ArchType::ppc64le: |
| 307 | case Triple::ArchType::systemz: { |
| 308 | // PPC has conditional returns. Turn them into branch and plain returns. |
| 309 | InstrumentationOptions op; |
| 310 | op.HandleTailcall = false; |
| 311 | op.HandleAllReturns = true; |
| 312 | replaceRetWithPatchableRet(MF, TII, op); |
| 313 | break; |
| 314 | } |
| 315 | default: { |
| 316 | // For the architectures that have a single return instruction (such as |
| 317 | // RETQ on x86_64). |
| 318 | InstrumentationOptions op; |
| 319 | op.HandleTailcall = true; |
| 320 | op.HandleAllReturns = false; |
| 321 | replaceRetWithPatchableRet(MF, TII, op); |
| 322 | break; |
| 323 | } |
| 324 | } |
| 325 | } |
| 326 | return true; |
| 327 | } |
| 328 | |
| 329 | char XRayInstrumentationLegacy::ID = 0; |
| 330 | char &llvm::XRayInstrumentationID = XRayInstrumentationLegacy::ID; |
| 331 | INITIALIZE_PASS_BEGIN(XRayInstrumentationLegacy, "xray-instrumentation" , |
| 332 | "Insert XRay ops" , false, false) |
| 333 | INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass) |
| 334 | INITIALIZE_PASS_END(XRayInstrumentationLegacy, "xray-instrumentation" , |
| 335 | "Insert XRay ops" , false, false) |
| 336 | |