1//===- Offloading.cpp - Utilities for handling offloading code -*- C++ -*-===//
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#include "llvm/Object/OffloadBinary.h"
10
11#include "llvm/ADT/StringSwitch.h"
12#include "llvm/BinaryFormat/Magic.h"
13#include "llvm/IR/Constants.h"
14#include "llvm/IR/Module.h"
15#include "llvm/IRReader/IRReader.h"
16#include "llvm/MC/StringTableBuilder.h"
17#include "llvm/Object/Archive.h"
18#include "llvm/Object/ArchiveWriter.h"
19#include "llvm/Object/Binary.h"
20#include "llvm/Object/COFF.h"
21#include "llvm/Object/ELFObjectFile.h"
22#include "llvm/Object/Error.h"
23#include "llvm/Object/IRObjectFile.h"
24#include "llvm/Object/ObjectFile.h"
25#include "llvm/Support/Alignment.h"
26#include "llvm/Support/FileOutputBuffer.h"
27#include "llvm/Support/SourceMgr.h"
28
29using namespace llvm;
30using namespace llvm::object;
31
32namespace {
33
34/// Attempts to extract all the embedded device images contained inside the
35/// buffer \p Contents. The buffer is expected to contain a valid offloading
36/// binary format.
37Error extractOffloadFiles(MemoryBufferRef Contents,
38 SmallVectorImpl<OffloadFile> &Binaries) {
39 uint64_t Offset = 0;
40 // There could be multiple offloading binaries stored at this section.
41 while (Offset < Contents.getBuffer().size()) {
42 std::unique_ptr<MemoryBuffer> Buffer =
43 MemoryBuffer::getMemBuffer(InputData: Contents.getBuffer().drop_front(N: Offset), BufferName: "",
44 /*RequiresNullTerminator*/ false);
45 if (!isAddrAligned(Lhs: Align(OffloadBinary::getAlignment()),
46 Addr: Buffer->getBufferStart()))
47 Buffer = MemoryBuffer::getMemBufferCopy(InputData: Buffer->getBuffer(),
48 BufferName: Buffer->getBufferIdentifier());
49 auto BinaryOrErr = OffloadBinary::create(*Buffer);
50 if (!BinaryOrErr)
51 return BinaryOrErr.takeError();
52 OffloadBinary &Binary = **BinaryOrErr;
53
54 // Create a new owned binary with a copy of the original memory.
55 std::unique_ptr<MemoryBuffer> BufferCopy = MemoryBuffer::getMemBufferCopy(
56 InputData: Binary.getData().take_front(N: Binary.getSize()),
57 BufferName: Contents.getBufferIdentifier());
58 auto NewBinaryOrErr = OffloadBinary::create(*BufferCopy);
59 if (!NewBinaryOrErr)
60 return NewBinaryOrErr.takeError();
61 Binaries.emplace_back(Args: std::move(*NewBinaryOrErr), Args: std::move(BufferCopy));
62
63 Offset += Binary.getSize();
64 }
65
66 return Error::success();
67}
68
69// Extract offloading binaries from an Object file \p Obj.
70Error extractFromObject(const ObjectFile &Obj,
71 SmallVectorImpl<OffloadFile> &Binaries) {
72 assert((Obj.isELF() || Obj.isCOFF()) && "Invalid file type");
73
74 for (SectionRef Sec : Obj.sections()) {
75 // ELF files contain a section with the LLVM_OFFLOADING type.
76 if (Obj.isELF() &&
77 static_cast<ELFSectionRef>(Sec).getType() != ELF::SHT_LLVM_OFFLOADING)
78 continue;
79
80 // COFF has no section types so we rely on the name of the section.
81 if (Obj.isCOFF()) {
82 Expected<StringRef> NameOrErr = Sec.getName();
83 if (!NameOrErr)
84 return NameOrErr.takeError();
85
86 if (!NameOrErr->starts_with(Prefix: ".llvm.offloading"))
87 continue;
88 }
89
90 Expected<StringRef> Buffer = Sec.getContents();
91 if (!Buffer)
92 return Buffer.takeError();
93
94 MemoryBufferRef Contents(*Buffer, Obj.getFileName());
95 if (Error Err = extractOffloadFiles(Contents, Binaries))
96 return Err;
97 }
98
99 return Error::success();
100}
101
102Error extractFromBitcode(MemoryBufferRef Buffer,
103 SmallVectorImpl<OffloadFile> &Binaries) {
104 LLVMContext Context;
105 SMDiagnostic Err;
106 std::unique_ptr<Module> M = getLazyIRModule(
107 Buffer: MemoryBuffer::getMemBuffer(Ref: Buffer, /*RequiresNullTerminator=*/false), Err,
108 Context);
109 if (!M)
110 return createStringError(EC: inconvertibleErrorCode(),
111 S: "Failed to create module");
112
113 // Extract offloading data from globals referenced by the
114 // `llvm.embedded.object` metadata with the `.llvm.offloading` section.
115 auto *MD = M->getNamedMetadata(Name: "llvm.embedded.objects");
116 if (!MD)
117 return Error::success();
118
119 for (const MDNode *Op : MD->operands()) {
120 if (Op->getNumOperands() < 2)
121 continue;
122
123 MDString *SectionID = dyn_cast<MDString>(Val: Op->getOperand(I: 1));
124 if (!SectionID || SectionID->getString() != ".llvm.offloading")
125 continue;
126
127 GlobalVariable *GV =
128 mdconst::dyn_extract_or_null<GlobalVariable>(MD: Op->getOperand(I: 0));
129 if (!GV)
130 continue;
131
132 auto *CDS = dyn_cast<ConstantDataSequential>(Val: GV->getInitializer());
133 if (!CDS)
134 continue;
135
136 MemoryBufferRef Contents(CDS->getAsString(), M->getName());
137 if (Error Err = extractOffloadFiles(Contents, Binaries))
138 return Err;
139 }
140
141 return Error::success();
142}
143
144Error extractFromArchive(const Archive &Library,
145 SmallVectorImpl<OffloadFile> &Binaries) {
146 // Try to extract device code from each file stored in the static archive.
147 Error Err = Error::success();
148 for (auto Child : Library.children(Err)) {
149 auto ChildBufferOrErr = Child.getMemoryBufferRef();
150 if (!ChildBufferOrErr)
151 return ChildBufferOrErr.takeError();
152 std::unique_ptr<MemoryBuffer> ChildBuffer =
153 MemoryBuffer::getMemBuffer(Ref: *ChildBufferOrErr, RequiresNullTerminator: false);
154
155 // Check if the buffer has the required alignment.
156 if (!isAddrAligned(Lhs: Align(OffloadBinary::getAlignment()),
157 Addr: ChildBuffer->getBufferStart()))
158 ChildBuffer = MemoryBuffer::getMemBufferCopy(
159 InputData: ChildBufferOrErr->getBuffer(),
160 BufferName: ChildBufferOrErr->getBufferIdentifier());
161
162 if (Error Err = extractOffloadBinaries(Buffer: *ChildBuffer, Binaries))
163 return Err;
164 }
165
166 if (Err)
167 return Err;
168 return Error::success();
169}
170
171} // namespace
172
173Expected<std::unique_ptr<OffloadBinary>>
174OffloadBinary::create(MemoryBufferRef Buf) {
175 if (Buf.getBufferSize() < sizeof(Header) + sizeof(Entry))
176 return errorCodeToError(EC: object_error::parse_failed);
177
178 // Check for 0x10FF1OAD magic bytes.
179 if (identify_magic(magic: Buf.getBuffer()) != file_magic::offload_binary)
180 return errorCodeToError(EC: object_error::parse_failed);
181
182 // Make sure that the data has sufficient alignment.
183 if (!isAddrAligned(Lhs: Align(getAlignment()), Addr: Buf.getBufferStart()))
184 return errorCodeToError(EC: object_error::parse_failed);
185
186 const char *Start = Buf.getBufferStart();
187 const Header *TheHeader = reinterpret_cast<const Header *>(Start);
188 if (TheHeader->Version != OffloadBinary::Version)
189 return errorCodeToError(EC: object_error::parse_failed);
190
191 if (TheHeader->Size > Buf.getBufferSize() ||
192 TheHeader->Size < sizeof(Entry) || TheHeader->Size < sizeof(Header))
193 return errorCodeToError(EC: object_error::unexpected_eof);
194
195 if (TheHeader->EntryOffset > TheHeader->Size - sizeof(Entry) ||
196 TheHeader->EntrySize > TheHeader->Size - sizeof(Header))
197 return errorCodeToError(EC: object_error::unexpected_eof);
198
199 const Entry *TheEntry =
200 reinterpret_cast<const Entry *>(&Start[TheHeader->EntryOffset]);
201
202 if (TheEntry->ImageOffset > Buf.getBufferSize() ||
203 TheEntry->StringOffset > Buf.getBufferSize())
204 return errorCodeToError(EC: object_error::unexpected_eof);
205
206 return std::unique_ptr<OffloadBinary>(
207 new OffloadBinary(Buf, TheHeader, TheEntry));
208}
209
210SmallString<0> OffloadBinary::write(const OffloadingImage &OffloadingData) {
211 // Create a null-terminated string table with all the used strings.
212 StringTableBuilder StrTab(StringTableBuilder::ELF);
213 for (auto &KeyAndValue : OffloadingData.StringData) {
214 StrTab.add(S: KeyAndValue.first);
215 StrTab.add(S: KeyAndValue.second);
216 }
217 StrTab.finalize();
218
219 uint64_t StringEntrySize =
220 sizeof(StringEntry) * OffloadingData.StringData.size();
221
222 // Make sure the image we're wrapping around is aligned as well.
223 uint64_t BinaryDataSize = alignTo(Value: sizeof(Header) + sizeof(Entry) +
224 StringEntrySize + StrTab.getSize(),
225 Align: getAlignment());
226
227 // Create the header and fill in the offsets. The entry will be directly
228 // placed after the header in memory. Align the size to the alignment of the
229 // header so this can be placed contiguously in a single section.
230 Header TheHeader;
231 TheHeader.Size = alignTo(
232 Value: BinaryDataSize + OffloadingData.Image->getBufferSize(), Align: getAlignment());
233 TheHeader.EntryOffset = sizeof(Header);
234 TheHeader.EntrySize = sizeof(Entry);
235
236 // Create the entry using the string table offsets. The string table will be
237 // placed directly after the entry in memory, and the image after that.
238 Entry TheEntry;
239 TheEntry.TheImageKind = OffloadingData.TheImageKind;
240 TheEntry.TheOffloadKind = OffloadingData.TheOffloadKind;
241 TheEntry.Flags = OffloadingData.Flags;
242 TheEntry.StringOffset = sizeof(Header) + sizeof(Entry);
243 TheEntry.NumStrings = OffloadingData.StringData.size();
244
245 TheEntry.ImageOffset = BinaryDataSize;
246 TheEntry.ImageSize = OffloadingData.Image->getBufferSize();
247
248 SmallString<0> Data;
249 Data.reserve(N: TheHeader.Size);
250 raw_svector_ostream OS(Data);
251 OS << StringRef(reinterpret_cast<char *>(&TheHeader), sizeof(Header));
252 OS << StringRef(reinterpret_cast<char *>(&TheEntry), sizeof(Entry));
253 for (auto &KeyAndValue : OffloadingData.StringData) {
254 uint64_t Offset = sizeof(Header) + sizeof(Entry) + StringEntrySize;
255 StringEntry Map{.KeyOffset: Offset + StrTab.getOffset(S: KeyAndValue.first),
256 .ValueOffset: Offset + StrTab.getOffset(S: KeyAndValue.second)};
257 OS << StringRef(reinterpret_cast<char *>(&Map), sizeof(StringEntry));
258 }
259 StrTab.write(OS);
260 // Add padding to required image alignment.
261 OS.write_zeros(NumZeros: TheEntry.ImageOffset - OS.tell());
262 OS << OffloadingData.Image->getBuffer();
263
264 // Add final padding to required alignment.
265 assert(TheHeader.Size >= OS.tell() && "Too much data written?");
266 OS.write_zeros(NumZeros: TheHeader.Size - OS.tell());
267 assert(TheHeader.Size == OS.tell() && "Size mismatch");
268
269 return Data;
270}
271
272Error object::extractOffloadBinaries(MemoryBufferRef Buffer,
273 SmallVectorImpl<OffloadFile> &Binaries) {
274 file_magic Type = identify_magic(magic: Buffer.getBuffer());
275 switch (Type) {
276 case file_magic::bitcode:
277 return extractFromBitcode(Buffer, Binaries);
278 case file_magic::elf_relocatable:
279 case file_magic::elf_executable:
280 case file_magic::elf_shared_object:
281 case file_magic::coff_object: {
282 Expected<std::unique_ptr<ObjectFile>> ObjFile =
283 ObjectFile::createObjectFile(Object: Buffer, Type);
284 if (!ObjFile)
285 return ObjFile.takeError();
286 return extractFromObject(Obj: *ObjFile->get(), Binaries);
287 }
288 case file_magic::archive: {
289 Expected<std::unique_ptr<llvm::object::Archive>> LibFile =
290 object::Archive::create(Source: Buffer);
291 if (!LibFile)
292 return LibFile.takeError();
293 return extractFromArchive(Library: *LibFile->get(), Binaries);
294 }
295 case file_magic::offload_binary:
296 return extractOffloadFiles(Contents: Buffer, Binaries);
297 default:
298 return Error::success();
299 }
300}
301
302OffloadKind object::getOffloadKind(StringRef Name) {
303 return llvm::StringSwitch<OffloadKind>(Name)
304 .Case(S: "openmp", Value: OFK_OpenMP)
305 .Case(S: "cuda", Value: OFK_Cuda)
306 .Case(S: "hip", Value: OFK_HIP)
307 .Default(Value: OFK_None);
308}
309
310StringRef object::getOffloadKindName(OffloadKind Kind) {
311 switch (Kind) {
312 case OFK_OpenMP:
313 return "openmp";
314 case OFK_Cuda:
315 return "cuda";
316 case OFK_HIP:
317 return "hip";
318 default:
319 return "none";
320 }
321}
322
323ImageKind object::getImageKind(StringRef Name) {
324 return llvm::StringSwitch<ImageKind>(Name)
325 .Case(S: "o", Value: IMG_Object)
326 .Case(S: "bc", Value: IMG_Bitcode)
327 .Case(S: "cubin", Value: IMG_Cubin)
328 .Case(S: "fatbin", Value: IMG_Fatbinary)
329 .Case(S: "s", Value: IMG_PTX)
330 .Default(Value: IMG_None);
331}
332
333StringRef object::getImageKindName(ImageKind Kind) {
334 switch (Kind) {
335 case IMG_Object:
336 return "o";
337 case IMG_Bitcode:
338 return "bc";
339 case IMG_Cubin:
340 return "cubin";
341 case IMG_Fatbinary:
342 return "fatbin";
343 case IMG_PTX:
344 return "s";
345 default:
346 return "";
347 }
348}
349
350bool object::areTargetsCompatible(const OffloadFile::TargetID &LHS,
351 const OffloadFile::TargetID &RHS) {
352 // Exact matches are not considered compatible because they are the same
353 // target. We are interested in different targets that are compatible.
354 if (LHS == RHS)
355 return false;
356
357 // The triples must match at all times.
358 if (LHS.first != RHS.first)
359 return false;
360
361 // If the architecture is "all" we assume it is always compatible.
362 if (LHS.second == "generic" || RHS.second == "generic")
363 return true;
364
365 // Only The AMDGPU target requires additional checks.
366 llvm::Triple T(LHS.first);
367 if (!T.isAMDGPU())
368 return false;
369
370 // The base processor must always match.
371 if (LHS.second.split(Separator: ":").first != RHS.second.split(Separator: ":").first)
372 return false;
373
374 // Check combintions of on / off features that must match.
375 if (LHS.second.contains(Other: "xnack+") && RHS.second.contains(Other: "xnack-"))
376 return false;
377 if (LHS.second.contains(Other: "xnack-") && RHS.second.contains(Other: "xnack+"))
378 return false;
379 if (LHS.second.contains(Other: "sramecc-") && RHS.second.contains(Other: "sramecc+"))
380 return false;
381 if (LHS.second.contains(Other: "sramecc+") && RHS.second.contains(Other: "sramecc-"))
382 return false;
383 return true;
384}
385