1 //===----- x86_64.cpp - Generic JITLink x86-64 edge kinds, utilities ------===// 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 // Generic utilities for graphs representing x86-64 objects. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/ExecutionEngine/JITLink/x86_64.h" 14 15 #define DEBUG_TYPE "jitlink" 16 17 namespace llvm { 18 namespace jitlink { 19 namespace x86_64 { 20 21 const char *getEdgeKindName(Edge::Kind K) { 22 switch (K) { 23 case Pointer64: 24 return "Pointer64"; 25 case Pointer32: 26 return "Pointer32"; 27 case Pointer32Signed: 28 return "Pointer32Signed"; 29 case Delta64: 30 return "Delta64"; 31 case Delta32: 32 return "Delta32"; 33 case NegDelta64: 34 return "NegDelta64"; 35 case NegDelta32: 36 return "NegDelta32"; 37 case Delta64FromGOT: 38 return "Delta64FromGOT"; 39 case BranchPCRel32: 40 return "BranchPCRel32"; 41 case BranchPCRel32ToPtrJumpStub: 42 return "BranchPCRel32ToPtrJumpStub"; 43 case BranchPCRel32ToPtrJumpStubBypassable: 44 return "BranchPCRel32ToPtrJumpStubBypassable"; 45 case RequestGOTAndTransformToDelta32: 46 return "RequestGOTAndTransformToDelta32"; 47 case RequestGOTAndTransformToDelta64: 48 return "RequestGOTAndTransformToDelta64"; 49 case RequestGOTAndTransformToDelta64FromGOT: 50 return "RequestGOTAndTransformToDelta64FromGOT"; 51 case PCRel32GOTLoadREXRelaxable: 52 return "PCRel32GOTLoadREXRelaxable"; 53 case RequestGOTAndTransformToPCRel32GOTLoadREXRelaxable: 54 return "RequestGOTAndTransformToPCRel32GOTLoadREXRelaxable"; 55 case PCRel32GOTLoadRelaxable: 56 return "PCRel32GOTLoadRelaxable"; 57 case RequestGOTAndTransformToPCRel32GOTLoadRelaxable: 58 return "RequestGOTAndTransformToPCRel32GOTLoadRelaxable"; 59 case PCRel32TLVPLoadREXRelaxable: 60 return "PCRel32TLVPLoadREXRelaxable"; 61 case RequestTLVPAndTransformToPCRel32TLVPLoadREXRelaxable: 62 return "RequestTLVPAndTransformToPCRel32TLVPLoadREXRelaxable"; 63 default: 64 return getGenericEdgeKindName(static_cast<Edge::Kind>(K)); 65 } 66 } 67 68 const char NullPointerContent[PointerSize] = {0x00, 0x00, 0x00, 0x00, 69 0x00, 0x00, 0x00, 0x00}; 70 71 const char PointerJumpStubContent[6] = { 72 static_cast<char>(0xFFu), 0x25, 0x00, 0x00, 0x00, 0x00}; 73 74 Error optimizeGOTAndStubAccesses(LinkGraph &G) { 75 LLVM_DEBUG(dbgs() << "Optimizing GOT entries and stubs:\n"); 76 77 for (auto *B : G.blocks()) 78 for (auto &E : B->edges()) { 79 if (E.getKind() == x86_64::PCRel32GOTLoadRelaxable || 80 E.getKind() == x86_64::PCRel32GOTLoadREXRelaxable) { 81 #ifndef NDEBUG 82 bool REXPrefix = E.getKind() == x86_64::PCRel32GOTLoadREXRelaxable; 83 assert(E.getOffset() >= (REXPrefix ? 3u : 2u) && 84 "GOT edge occurs too early in block"); 85 #endif 86 auto *FixupData = reinterpret_cast<uint8_t *>( 87 const_cast<char *>(B->getContent().data())) + 88 E.getOffset(); 89 const uint8_t Op = FixupData[-2]; 90 const uint8_t ModRM = FixupData[-1]; 91 92 auto &GOTEntryBlock = E.getTarget().getBlock(); 93 assert(GOTEntryBlock.getSize() == G.getPointerSize() && 94 "GOT entry block should be pointer sized"); 95 assert(GOTEntryBlock.edges_size() == 1 && 96 "GOT entry should only have one outgoing edge"); 97 auto &GOTTarget = GOTEntryBlock.edges().begin()->getTarget(); 98 orc::ExecutorAddr TargetAddr = GOTTarget.getAddress(); 99 orc::ExecutorAddr EdgeAddr = B->getFixupAddress(E); 100 int64_t Displacement = TargetAddr - EdgeAddr + 4; 101 bool TargetInRangeForImmU32 = isInRangeForImmU32(TargetAddr.getValue()); 102 bool DisplacementInRangeForImmS32 = isInRangeForImmS32(Displacement); 103 104 // If both of the Target and displacement is out of range, then 105 // there isn't optimization chance. 106 if (!(TargetInRangeForImmU32 || DisplacementInRangeForImmS32)) 107 continue; 108 109 // Transform "mov foo@GOTPCREL(%rip),%reg" to "lea foo(%rip),%reg". 110 if (Op == 0x8b && DisplacementInRangeForImmS32) { 111 FixupData[-2] = 0x8d; 112 E.setKind(x86_64::Delta32); 113 E.setTarget(GOTTarget); 114 E.setAddend(E.getAddend() - 4); 115 LLVM_DEBUG({ 116 dbgs() << " Replaced GOT load wih LEA:\n "; 117 printEdge(dbgs(), *B, E, getEdgeKindName(E.getKind())); 118 dbgs() << "\n"; 119 }); 120 continue; 121 } 122 123 // Transform call/jmp instructions 124 if (Op == 0xff && TargetInRangeForImmU32) { 125 if (ModRM == 0x15) { 126 // ABI says we can convert "call *foo@GOTPCREL(%rip)" to "nop; call 127 // foo" But lld convert it to "addr32 call foo, because that makes 128 // result expression to be a single instruction. 129 FixupData[-2] = 0x67; 130 FixupData[-1] = 0xe8; 131 LLVM_DEBUG({ 132 dbgs() << " replaced call instruction's memory operand wih imm " 133 "operand:\n "; 134 printEdge(dbgs(), *B, E, getEdgeKindName(E.getKind())); 135 dbgs() << "\n"; 136 }); 137 } else { 138 // Transform "jmp *foo@GOTPCREL(%rip)" to "jmp foo; nop" 139 assert(ModRM == 0x25 && "Invalid ModRm for call/jmp instructions"); 140 FixupData[-2] = 0xe9; 141 FixupData[3] = 0x90; 142 E.setOffset(E.getOffset() - 1); 143 LLVM_DEBUG({ 144 dbgs() << " replaced jmp instruction's memory operand wih imm " 145 "operand:\n "; 146 printEdge(dbgs(), *B, E, getEdgeKindName(E.getKind())); 147 dbgs() << "\n"; 148 }); 149 } 150 E.setKind(x86_64::Pointer32); 151 E.setTarget(GOTTarget); 152 continue; 153 } 154 } else if (E.getKind() == x86_64::BranchPCRel32ToPtrJumpStubBypassable) { 155 auto &StubBlock = E.getTarget().getBlock(); 156 assert(StubBlock.getSize() == sizeof(PointerJumpStubContent) && 157 "Stub block should be stub sized"); 158 assert(StubBlock.edges_size() == 1 && 159 "Stub block should only have one outgoing edge"); 160 161 auto &GOTBlock = StubBlock.edges().begin()->getTarget().getBlock(); 162 assert(GOTBlock.getSize() == G.getPointerSize() && 163 "GOT block should be pointer sized"); 164 assert(GOTBlock.edges_size() == 1 && 165 "GOT block should only have one outgoing edge"); 166 167 auto &GOTTarget = GOTBlock.edges().begin()->getTarget(); 168 orc::ExecutorAddr EdgeAddr = B->getAddress() + E.getOffset(); 169 orc::ExecutorAddr TargetAddr = GOTTarget.getAddress(); 170 171 int64_t Displacement = TargetAddr - EdgeAddr + 4; 172 if (isInRangeForImmS32(Displacement)) { 173 E.setKind(x86_64::BranchPCRel32); 174 E.setTarget(GOTTarget); 175 LLVM_DEBUG({ 176 dbgs() << " Replaced stub branch with direct branch:\n "; 177 printEdge(dbgs(), *B, E, getEdgeKindName(E.getKind())); 178 dbgs() << "\n"; 179 }); 180 } 181 } 182 } 183 184 return Error::success(); 185 } 186 187 } // end namespace x86_64 188 } // end namespace jitlink 189 } // end namespace llvm 190