[RISCV] Fix mgather -> riscv.masked.strided.load combine not extending indices (...
[llvm-project.git] / llvm / lib / ExecutionEngine / JITLink / x86_64.cpp
blob273ac7b372a7ff30cf2d08eb0e9eeb058585e473
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.
11 //===----------------------------------------------------------------------===//
13 #include "llvm/ExecutionEngine/JITLink/x86_64.h"
15 #define DEBUG_TYPE "jitlink"
17 namespace llvm {
18 namespace jitlink {
19 namespace x86_64 {
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 NegDelta64:
38 return "NegDelta64";
39 case NegDelta32:
40 return "NegDelta32";
41 case Delta64FromGOT:
42 return "Delta64FromGOT";
43 case PCRel32:
44 return "PCRel32";
45 case BranchPCRel32:
46 return "BranchPCRel32";
47 case BranchPCRel32ToPtrJumpStub:
48 return "BranchPCRel32ToPtrJumpStub";
49 case BranchPCRel32ToPtrJumpStubBypassable:
50 return "BranchPCRel32ToPtrJumpStubBypassable";
51 case RequestGOTAndTransformToDelta32:
52 return "RequestGOTAndTransformToDelta32";
53 case RequestGOTAndTransformToDelta64:
54 return "RequestGOTAndTransformToDelta64";
55 case RequestGOTAndTransformToDelta64FromGOT:
56 return "RequestGOTAndTransformToDelta64FromGOT";
57 case PCRel32GOTLoadREXRelaxable:
58 return "PCRel32GOTLoadREXRelaxable";
59 case RequestGOTAndTransformToPCRel32GOTLoadREXRelaxable:
60 return "RequestGOTAndTransformToPCRel32GOTLoadREXRelaxable";
61 case PCRel32GOTLoadRelaxable:
62 return "PCRel32GOTLoadRelaxable";
63 case RequestGOTAndTransformToPCRel32GOTLoadRelaxable:
64 return "RequestGOTAndTransformToPCRel32GOTLoadRelaxable";
65 case PCRel32TLVPLoadREXRelaxable:
66 return "PCRel32TLVPLoadREXRelaxable";
67 case RequestTLVPAndTransformToPCRel32TLVPLoadREXRelaxable:
68 return "RequestTLVPAndTransformToPCRel32TLVPLoadREXRelaxable";
69 default:
70 return getGenericEdgeKindName(static_cast<Edge::Kind>(K));
74 const char NullPointerContent[PointerSize] = {0x00, 0x00, 0x00, 0x00,
75 0x00, 0x00, 0x00, 0x00};
77 const char PointerJumpStubContent[6] = {
78 static_cast<char>(0xFFu), 0x25, 0x00, 0x00, 0x00, 0x00};
80 Error optimizeGOTAndStubAccesses(LinkGraph &G) {
81 LLVM_DEBUG(dbgs() << "Optimizing GOT entries and stubs:\n");
83 for (auto *B : G.blocks())
84 for (auto &E : B->edges()) {
85 if (E.getKind() == x86_64::PCRel32GOTLoadRelaxable ||
86 E.getKind() == x86_64::PCRel32GOTLoadREXRelaxable) {
87 #ifndef NDEBUG
88 bool REXPrefix = E.getKind() == x86_64::PCRel32GOTLoadREXRelaxable;
89 assert(E.getOffset() >= (REXPrefix ? 3u : 2u) &&
90 "GOT edge occurs too early in block");
91 #endif
92 auto *FixupData = reinterpret_cast<uint8_t *>(
93 const_cast<char *>(B->getContent().data())) +
94 E.getOffset();
95 const uint8_t Op = FixupData[-2];
96 const uint8_t ModRM = FixupData[-1];
98 auto &GOTEntryBlock = E.getTarget().getBlock();
99 assert(GOTEntryBlock.getSize() == G.getPointerSize() &&
100 "GOT entry block should be pointer sized");
101 assert(GOTEntryBlock.edges_size() == 1 &&
102 "GOT entry should only have one outgoing edge");
103 auto &GOTTarget = GOTEntryBlock.edges().begin()->getTarget();
104 orc::ExecutorAddr TargetAddr = GOTTarget.getAddress();
105 orc::ExecutorAddr EdgeAddr = B->getFixupAddress(E);
106 int64_t Displacement = TargetAddr - EdgeAddr + 4;
107 bool TargetInRangeForImmU32 = isUInt<32>(TargetAddr.getValue());
108 bool DisplacementInRangeForImmS32 = isInt<32>(Displacement);
110 // If both of the Target and displacement is out of range, then
111 // there isn't optimization chance.
112 if (!(TargetInRangeForImmU32 || DisplacementInRangeForImmS32))
113 continue;
115 // Transform "mov foo@GOTPCREL(%rip),%reg" to "lea foo(%rip),%reg".
116 if (Op == 0x8b && DisplacementInRangeForImmS32) {
117 FixupData[-2] = 0x8d;
118 E.setKind(x86_64::Delta32);
119 E.setTarget(GOTTarget);
120 E.setAddend(E.getAddend() - 4);
121 LLVM_DEBUG({
122 dbgs() << " Replaced GOT load wih LEA:\n ";
123 printEdge(dbgs(), *B, E, getEdgeKindName(E.getKind()));
124 dbgs() << "\n";
126 continue;
129 // Transform call/jmp instructions
130 if (Op == 0xff && TargetInRangeForImmU32) {
131 if (ModRM == 0x15) {
132 // ABI says we can convert "call *foo@GOTPCREL(%rip)" to "nop; call
133 // foo" But lld convert it to "addr32 call foo, because that makes
134 // result expression to be a single instruction.
135 FixupData[-2] = 0x67;
136 FixupData[-1] = 0xe8;
137 LLVM_DEBUG({
138 dbgs() << " replaced call instruction's memory operand wih imm "
139 "operand:\n ";
140 printEdge(dbgs(), *B, E, getEdgeKindName(E.getKind()));
141 dbgs() << "\n";
143 } else {
144 // Transform "jmp *foo@GOTPCREL(%rip)" to "jmp foo; nop"
145 assert(ModRM == 0x25 && "Invalid ModRm for call/jmp instructions");
146 FixupData[-2] = 0xe9;
147 FixupData[3] = 0x90;
148 E.setOffset(E.getOffset() - 1);
149 LLVM_DEBUG({
150 dbgs() << " replaced jmp instruction's memory operand wih imm "
151 "operand:\n ";
152 printEdge(dbgs(), *B, E, getEdgeKindName(E.getKind()));
153 dbgs() << "\n";
156 E.setKind(x86_64::Pointer32);
157 E.setTarget(GOTTarget);
158 continue;
160 } else if (E.getKind() == x86_64::BranchPCRel32ToPtrJumpStubBypassable) {
161 auto &StubBlock = E.getTarget().getBlock();
162 assert(StubBlock.getSize() == sizeof(PointerJumpStubContent) &&
163 "Stub block should be stub sized");
164 assert(StubBlock.edges_size() == 1 &&
165 "Stub block should only have one outgoing edge");
167 auto &GOTBlock = StubBlock.edges().begin()->getTarget().getBlock();
168 assert(GOTBlock.getSize() == G.getPointerSize() &&
169 "GOT block should be pointer sized");
170 assert(GOTBlock.edges_size() == 1 &&
171 "GOT block should only have one outgoing edge");
173 auto &GOTTarget = GOTBlock.edges().begin()->getTarget();
174 orc::ExecutorAddr EdgeAddr = B->getAddress() + E.getOffset();
175 orc::ExecutorAddr TargetAddr = GOTTarget.getAddress();
177 int64_t Displacement = TargetAddr - EdgeAddr + 4;
178 if (isInt<32>(Displacement)) {
179 E.setKind(x86_64::BranchPCRel32);
180 E.setTarget(GOTTarget);
181 LLVM_DEBUG({
182 dbgs() << " Replaced stub branch with direct branch:\n ";
183 printEdge(dbgs(), *B, E, getEdgeKindName(E.getKind()));
184 dbgs() << "\n";
190 return Error::success();
193 } // end namespace x86_64
194 } // end namespace jitlink
195 } // end namespace llvm