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 Pointer16: |
30 | return "Pointer16" ; |
31 | case Pointer8: |
32 | return "Pointer8" ; |
33 | case Delta64: |
34 | return "Delta64" ; |
35 | case Delta32: |
36 | return "Delta32" ; |
37 | case Delta8: |
38 | return "Delta8" ; |
39 | case NegDelta64: |
40 | return "NegDelta64" ; |
41 | case NegDelta32: |
42 | return "NegDelta32" ; |
43 | case Delta64FromGOT: |
44 | return "Delta64FromGOT" ; |
45 | case PCRel32: |
46 | return "PCRel32" ; |
47 | case BranchPCRel32: |
48 | return "BranchPCRel32" ; |
49 | case BranchPCRel32ToPtrJumpStub: |
50 | return "BranchPCRel32ToPtrJumpStub" ; |
51 | case BranchPCRel32ToPtrJumpStubBypassable: |
52 | return "BranchPCRel32ToPtrJumpStubBypassable" ; |
53 | case RequestGOTAndTransformToDelta32: |
54 | return "RequestGOTAndTransformToDelta32" ; |
55 | case RequestGOTAndTransformToDelta64: |
56 | return "RequestGOTAndTransformToDelta64" ; |
57 | case RequestGOTAndTransformToDelta64FromGOT: |
58 | return "RequestGOTAndTransformToDelta64FromGOT" ; |
59 | case PCRel32GOTLoadREXRelaxable: |
60 | return "PCRel32GOTLoadREXRelaxable" ; |
61 | case RequestGOTAndTransformToPCRel32GOTLoadREXRelaxable: |
62 | return "RequestGOTAndTransformToPCRel32GOTLoadREXRelaxable" ; |
63 | case PCRel32GOTLoadRelaxable: |
64 | return "PCRel32GOTLoadRelaxable" ; |
65 | case RequestGOTAndTransformToPCRel32GOTLoadRelaxable: |
66 | return "RequestGOTAndTransformToPCRel32GOTLoadRelaxable" ; |
67 | case PCRel32TLVPLoadREXRelaxable: |
68 | return "PCRel32TLVPLoadREXRelaxable" ; |
69 | case RequestTLVPAndTransformToPCRel32TLVPLoadREXRelaxable: |
70 | return "RequestTLVPAndTransformToPCRel32TLVPLoadREXRelaxable" ; |
71 | default: |
72 | return getGenericEdgeKindName(K: static_cast<Edge::Kind>(K)); |
73 | } |
74 | } |
75 | |
76 | const char NullPointerContent[PointerSize] = {0x00, 0x00, 0x00, 0x00, |
77 | 0x00, 0x00, 0x00, 0x00}; |
78 | |
79 | const char PointerJumpStubContent[6] = { |
80 | static_cast<char>(0xFFu), 0x25, 0x00, 0x00, 0x00, 0x00}; |
81 | |
82 | Error optimizeGOTAndStubAccesses(LinkGraph &G) { |
83 | LLVM_DEBUG(dbgs() << "Optimizing GOT entries and stubs:\n" ); |
84 | |
85 | for (auto *B : G.blocks()) |
86 | for (auto &E : B->edges()) { |
87 | if (E.getKind() == x86_64::PCRel32GOTLoadRelaxable || |
88 | E.getKind() == x86_64::PCRel32GOTLoadREXRelaxable) { |
89 | #ifndef NDEBUG |
90 | bool REXPrefix = E.getKind() == x86_64::PCRel32GOTLoadREXRelaxable; |
91 | assert(E.getOffset() >= (REXPrefix ? 3u : 2u) && |
92 | "GOT edge occurs too early in block" ); |
93 | #endif |
94 | auto *FixupData = reinterpret_cast<uint8_t *>( |
95 | const_cast<char *>(B->getContent().data())) + |
96 | E.getOffset(); |
97 | const uint8_t Op = FixupData[-2]; |
98 | const uint8_t ModRM = FixupData[-1]; |
99 | |
100 | auto &GOTEntryBlock = E.getTarget().getBlock(); |
101 | assert(GOTEntryBlock.getSize() == G.getPointerSize() && |
102 | "GOT entry block should be pointer sized" ); |
103 | assert(GOTEntryBlock.edges_size() == 1 && |
104 | "GOT entry should only have one outgoing edge" ); |
105 | auto &GOTTarget = GOTEntryBlock.edges().begin()->getTarget(); |
106 | orc::ExecutorAddr TargetAddr = GOTTarget.getAddress(); |
107 | orc::ExecutorAddr EdgeAddr = B->getFixupAddress(E); |
108 | int64_t Displacement = TargetAddr - EdgeAddr + 4; |
109 | bool TargetInRangeForImmU32 = isUInt<32>(x: TargetAddr.getValue()); |
110 | bool DisplacementInRangeForImmS32 = isInt<32>(x: Displacement); |
111 | |
112 | // If both of the Target and displacement is out of range, then |
113 | // there isn't optimization chance. |
114 | if (!(TargetInRangeForImmU32 || DisplacementInRangeForImmS32)) |
115 | continue; |
116 | |
117 | // Transform "mov foo@GOTPCREL(%rip),%reg" to "lea foo(%rip),%reg". |
118 | if (Op == 0x8b && DisplacementInRangeForImmS32) { |
119 | FixupData[-2] = 0x8d; |
120 | E.setKind(x86_64::Delta32); |
121 | E.setTarget(GOTTarget); |
122 | E.setAddend(E.getAddend() - 4); |
123 | LLVM_DEBUG({ |
124 | dbgs() << " Replaced GOT load wih LEA:\n " ; |
125 | printEdge(dbgs(), *B, E, getEdgeKindName(E.getKind())); |
126 | dbgs() << "\n" ; |
127 | }); |
128 | continue; |
129 | } |
130 | |
131 | // Transform call/jmp instructions |
132 | if (Op == 0xff && TargetInRangeForImmU32) { |
133 | if (ModRM == 0x15) { |
134 | // ABI says we can convert "call *foo@GOTPCREL(%rip)" to "nop; call |
135 | // foo" But lld convert it to "addr32 call foo, because that makes |
136 | // result expression to be a single instruction. |
137 | FixupData[-2] = 0x67; |
138 | FixupData[-1] = 0xe8; |
139 | LLVM_DEBUG({ |
140 | dbgs() << " replaced call instruction's memory operand wih imm " |
141 | "operand:\n " ; |
142 | printEdge(dbgs(), *B, E, getEdgeKindName(E.getKind())); |
143 | dbgs() << "\n" ; |
144 | }); |
145 | } else { |
146 | // Transform "jmp *foo@GOTPCREL(%rip)" to "jmp foo; nop" |
147 | assert(ModRM == 0x25 && "Invalid ModRm for call/jmp instructions" ); |
148 | FixupData[-2] = 0xe9; |
149 | FixupData[3] = 0x90; |
150 | E.setOffset(E.getOffset() - 1); |
151 | LLVM_DEBUG({ |
152 | dbgs() << " replaced jmp instruction's memory operand wih imm " |
153 | "operand:\n " ; |
154 | printEdge(dbgs(), *B, E, getEdgeKindName(E.getKind())); |
155 | dbgs() << "\n" ; |
156 | }); |
157 | } |
158 | E.setKind(x86_64::Pointer32); |
159 | E.setTarget(GOTTarget); |
160 | continue; |
161 | } |
162 | } else if (E.getKind() == x86_64::BranchPCRel32ToPtrJumpStubBypassable) { |
163 | auto &StubBlock = E.getTarget().getBlock(); |
164 | assert(StubBlock.getSize() == sizeof(PointerJumpStubContent) && |
165 | "Stub block should be stub sized" ); |
166 | assert(StubBlock.edges_size() == 1 && |
167 | "Stub block should only have one outgoing edge" ); |
168 | |
169 | auto &GOTBlock = StubBlock.edges().begin()->getTarget().getBlock(); |
170 | assert(GOTBlock.getSize() == G.getPointerSize() && |
171 | "GOT block should be pointer sized" ); |
172 | assert(GOTBlock.edges_size() == 1 && |
173 | "GOT block should only have one outgoing edge" ); |
174 | |
175 | auto &GOTTarget = GOTBlock.edges().begin()->getTarget(); |
176 | orc::ExecutorAddr EdgeAddr = B->getAddress() + E.getOffset(); |
177 | orc::ExecutorAddr TargetAddr = GOTTarget.getAddress(); |
178 | |
179 | int64_t Displacement = TargetAddr - EdgeAddr + 4; |
180 | if (isInt<32>(x: Displacement)) { |
181 | E.setKind(x86_64::BranchPCRel32); |
182 | E.setTarget(GOTTarget); |
183 | LLVM_DEBUG({ |
184 | dbgs() << " Replaced stub branch with direct branch:\n " ; |
185 | printEdge(dbgs(), *B, E, getEdgeKindName(E.getKind())); |
186 | dbgs() << "\n" ; |
187 | }); |
188 | } |
189 | } |
190 | } |
191 | |
192 | return Error::success(); |
193 | } |
194 | |
195 | } // end namespace x86_64 |
196 | } // end namespace jitlink |
197 | } // end namespace llvm |
198 | |