1 // Copyright (c) 2011 The Chromium Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file.
5 #include "courgette/encoded_program.h"
12 #include "base/environment.h"
13 #include "base/logging.h"
14 #include "base/memory/scoped_ptr.h"
15 #include "base/strings/string_number_conversions.h"
16 #include "base/strings/string_util.h"
17 #include "courgette/courgette.h"
18 #include "courgette/disassembler_elf_32_arm.h"
19 #include "courgette/streams.h"
20 #include "courgette/types_elf.h"
25 const int kStreamMisc
= 0;
26 const int kStreamOps
= 1;
27 const int kStreamBytes
= 2;
28 const int kStreamAbs32Indexes
= 3;
29 const int kStreamRel32Indexes
= 4;
30 const int kStreamAbs32Addresses
= 5;
31 const int kStreamRel32Addresses
= 6;
32 const int kStreamCopyCounts
= 7;
33 const int kStreamOriginAddresses
= kStreamMisc
;
35 const int kStreamLimit
= 9;
37 // Constructor is here rather than in the header. Although the constructor
38 // appears to do nothing it is fact quite large because of the implicit calls to
39 // field constructors. Ditto for the destructor.
40 EncodedProgram::EncodedProgram() : image_base_(0) {}
41 EncodedProgram::~EncodedProgram() {}
43 // Serializes a vector of integral values using Varint32 coding.
45 CheckBool
WriteVector(const V
& items
, SinkStream
* buffer
) {
46 size_t count
= items
.size();
47 bool ok
= buffer
->WriteSizeVarint32(count
);
48 for (size_t i
= 0; ok
&& i
< count
; ++i
) {
49 ok
= buffer
->WriteSizeVarint32(items
[i
]);
55 bool ReadVector(V
* items
, SourceStream
* buffer
) {
57 if (!buffer
->ReadVarint32(&count
))
62 bool ok
= items
->reserve(count
);
63 for (size_t i
= 0; ok
&& i
< count
; ++i
) {
65 ok
= buffer
->ReadVarint32(&item
);
67 ok
= items
->push_back(static_cast<typename
V::value_type
>(item
));
73 // Serializes a vector, using delta coding followed by Varint32 coding.
75 CheckBool
WriteU32Delta(const V
& set
, SinkStream
* buffer
) {
76 size_t count
= set
.size();
77 bool ok
= buffer
->WriteSizeVarint32(count
);
79 for (size_t i
= 0; ok
&& i
< count
; ++i
) {
80 uint32 current
= set
[i
];
81 uint32 delta
= current
- prev
;
82 ok
= buffer
->WriteVarint32(delta
);
89 static CheckBool
ReadU32Delta(V
* set
, SourceStream
* buffer
) {
92 if (!buffer
->ReadVarint32(&count
))
96 bool ok
= set
->reserve(count
);
99 for (size_t i
= 0; ok
&& i
< count
; ++i
) {
101 ok
= buffer
->ReadVarint32(&delta
);
103 uint32 current
= prev
+ delta
;
104 ok
= set
->push_back(current
);
112 // Write a vector as the byte representation of the contents.
114 // (This only really makes sense for a type T that has sizeof(T)==1, otherwise
115 // serialized representation is not endian-agnostic. But it is useful to keep
116 // the possibility of a greater size for experiments comparing Varint32 encoding
117 // of a vector of larger integrals vs a plain form.)
120 CheckBool
WriteVectorU8(const V
& items
, SinkStream
* buffer
) {
121 size_t count
= items
.size();
122 bool ok
= buffer
->WriteSizeVarint32(count
);
123 if (count
!= 0 && ok
) {
124 size_t byte_count
= count
* sizeof(typename
V::value_type
);
125 ok
= buffer
->Write(static_cast<const void*>(&items
[0]), byte_count
);
131 bool ReadVectorU8(V
* items
, SourceStream
* buffer
) {
133 if (!buffer
->ReadVarint32(&count
))
137 bool ok
= items
->resize(count
, 0);
138 if (ok
&& count
!= 0) {
139 size_t byte_count
= count
* sizeof(typename
V::value_type
);
140 return buffer
->Read(static_cast<void*>(&((*items
)[0])), byte_count
);
145 ////////////////////////////////////////////////////////////////////////////////
147 CheckBool
EncodedProgram::DefineRel32Label(int index
, RVA value
) {
148 return DefineLabelCommon(&rel32_rva_
, index
, value
);
151 CheckBool
EncodedProgram::DefineAbs32Label(int index
, RVA value
) {
152 return DefineLabelCommon(&abs32_rva_
, index
, value
);
155 static const RVA kUnassignedRVA
= static_cast<RVA
>(-1);
157 CheckBool
EncodedProgram::DefineLabelCommon(RvaVector
* rvas
,
161 if (static_cast<int>(rvas
->size()) <= index
)
162 ok
= rvas
->resize(index
+ 1, kUnassignedRVA
);
165 DCHECK_EQ((*rvas
)[index
], kUnassignedRVA
)
166 << "DefineLabel double assigned " << index
;
167 (*rvas
)[index
] = rva
;
173 void EncodedProgram::EndLabels() {
174 FinishLabelsCommon(&abs32_rva_
);
175 FinishLabelsCommon(&rel32_rva_
);
178 void EncodedProgram::FinishLabelsCommon(RvaVector
* rvas
) {
179 // Replace all unassigned slots with the value at the previous index so they
180 // delta-encode to zero. (There might be better values than zero. The way to
181 // get that is have the higher level assembly program assign the unassigned
184 size_t size
= rvas
->size();
185 for (size_t i
= 0; i
< size
; ++i
) {
186 if ((*rvas
)[i
] == kUnassignedRVA
)
187 (*rvas
)[i
] = previous
;
189 previous
= (*rvas
)[i
];
193 CheckBool
EncodedProgram::AddOrigin(RVA origin
) {
194 return ops_
.push_back(ORIGIN
) && origins_
.push_back(origin
);
197 CheckBool
EncodedProgram::AddCopy(size_t count
, const void* bytes
) {
198 const uint8
* source
= static_cast<const uint8
*>(bytes
);
202 // Fold adjacent COPY instructions into one. This nearly halves the size of
203 // an EncodedProgram with only COPY1 instructions since there are approx plain
204 // 16 bytes per reloc. This has a working-set benefit during decompression.
205 // For compression of files with large differences this makes a small (4%)
206 // improvement in size. For files with small differences this degrades the
207 // compressed size by 1.3%
209 if (ops_
.back() == COPY1
) {
211 ok
= copy_counts_
.push_back(1);
213 if (ok
&& ops_
.back() == COPY
) {
214 copy_counts_
.back() += count
;
215 for (size_t i
= 0; ok
&& i
< count
; ++i
) {
216 ok
= copy_bytes_
.push_back(source
[i
]);
224 ok
= ops_
.push_back(COPY1
) && copy_bytes_
.push_back(source
[0]);
226 ok
= ops_
.push_back(COPY
) && copy_counts_
.push_back(count
);
227 for (size_t i
= 0; ok
&& i
< count
; ++i
) {
228 ok
= copy_bytes_
.push_back(source
[i
]);
236 CheckBool
EncodedProgram::AddAbs32(int label_index
) {
237 return ops_
.push_back(ABS32
) && abs32_ix_
.push_back(label_index
);
240 CheckBool
EncodedProgram::AddRel32(int label_index
) {
241 return ops_
.push_back(REL32
) && rel32_ix_
.push_back(label_index
);
244 CheckBool
EncodedProgram::AddRel32ARM(uint16 op
, int label_index
) {
245 return ops_
.push_back(static_cast<OP
>(op
)) &&
246 rel32_ix_
.push_back(label_index
);
249 CheckBool
EncodedProgram::AddPeMakeRelocs(ExecutableType kind
) {
250 if (kind
== EXE_WIN_32_X86
)
251 return ops_
.push_back(MAKE_PE_RELOCATION_TABLE
);
252 return ops_
.push_back(MAKE_PE64_RELOCATION_TABLE
);
255 CheckBool
EncodedProgram::AddElfMakeRelocs() {
256 return ops_
.push_back(MAKE_ELF_RELOCATION_TABLE
);
259 CheckBool
EncodedProgram::AddElfARMMakeRelocs() {
260 return ops_
.push_back(MAKE_ELF_ARM_RELOCATION_TABLE
);
263 void EncodedProgram::DebuggingSummary() {
264 VLOG(1) << "EncodedProgram Summary"
265 << "\n image base " << image_base_
266 << "\n abs32 rvas " << abs32_rva_
.size()
267 << "\n rel32 rvas " << rel32_rva_
.size()
268 << "\n ops " << ops_
.size()
269 << "\n origins " << origins_
.size()
270 << "\n copy_counts " << copy_counts_
.size()
271 << "\n copy_bytes " << copy_bytes_
.size()
272 << "\n abs32_ix " << abs32_ix_
.size()
273 << "\n rel32_ix " << rel32_ix_
.size();
276 ////////////////////////////////////////////////////////////////////////////////
278 // For algorithm refinement purposes it is useful to write subsets of the file
279 // format. This gives us the ability to estimate the entropy of the
280 // differential compression of the individual streams, which can provide
281 // invaluable insights. The default, of course, is to include all the streams.
284 INCLUDE_ABS32_ADDRESSES
= 0x0001,
285 INCLUDE_REL32_ADDRESSES
= 0x0002,
286 INCLUDE_ABS32_INDEXES
= 0x0010,
287 INCLUDE_REL32_INDEXES
= 0x0020,
288 INCLUDE_OPS
= 0x0100,
289 INCLUDE_BYTES
= 0x0200,
290 INCLUDE_COPY_COUNTS
= 0x0400,
291 INCLUDE_MISC
= 0x1000
294 static FieldSelect
GetFieldSelect() {
295 // TODO(sra): Use better configuration.
296 scoped_ptr
<base::Environment
> env(base::Environment::Create());
298 env
->GetVar("A_FIELDS", &s
);
300 if (!base::StringToUint64(s
, &fields
))
301 return static_cast<FieldSelect
>(~0);
302 return static_cast<FieldSelect
>(fields
);
305 CheckBool
EncodedProgram::WriteTo(SinkStreamSet
* streams
) {
306 FieldSelect select
= GetFieldSelect();
308 // The order of fields must be consistent in WriteTo and ReadFrom, regardless
309 // of the streams used. The code can be configured with all kStreamXXX
310 // constants the same.
312 // If we change the code to pipeline reading with assembly (to avoid temporary
313 // storage vectors by consuming operands directly from the stream) then we
314 // need to read the base address and the random access address tables first,
315 // the rest can be interleaved.
317 if (select
& INCLUDE_MISC
) {
318 // TODO(sra): write 64 bits.
319 if (!streams
->stream(kStreamMisc
)->WriteVarint32(
320 static_cast<uint32
>(image_base_
))) {
327 if (select
& INCLUDE_ABS32_ADDRESSES
) {
328 success
&= WriteU32Delta(abs32_rva_
,
329 streams
->stream(kStreamAbs32Addresses
));
332 if (select
& INCLUDE_REL32_ADDRESSES
) {
333 success
&= WriteU32Delta(rel32_rva_
,
334 streams
->stream(kStreamRel32Addresses
));
337 if (select
& INCLUDE_MISC
)
338 success
&= WriteVector(origins_
, streams
->stream(kStreamOriginAddresses
));
340 if (select
& INCLUDE_OPS
) {
342 success
&= streams
->stream(kStreamOps
)->Reserve(ops_
.size() + 5);
343 success
&= WriteVector(ops_
, streams
->stream(kStreamOps
));
346 if (select
& INCLUDE_COPY_COUNTS
)
347 success
&= WriteVector(copy_counts_
, streams
->stream(kStreamCopyCounts
));
349 if (select
& INCLUDE_BYTES
)
350 success
&= WriteVectorU8(copy_bytes_
, streams
->stream(kStreamBytes
));
352 if (select
& INCLUDE_ABS32_INDEXES
)
353 success
&= WriteVector(abs32_ix_
, streams
->stream(kStreamAbs32Indexes
));
355 if (select
& INCLUDE_REL32_INDEXES
)
356 success
&= WriteVector(rel32_ix_
, streams
->stream(kStreamRel32Indexes
));
361 bool EncodedProgram::ReadFrom(SourceStreamSet
* streams
) {
362 // TODO(sra): read 64 bits.
364 if (!streams
->stream(kStreamMisc
)->ReadVarint32(&temp
))
368 if (!ReadU32Delta(&abs32_rva_
, streams
->stream(kStreamAbs32Addresses
)))
370 if (!ReadU32Delta(&rel32_rva_
, streams
->stream(kStreamRel32Addresses
)))
372 if (!ReadVector(&origins_
, streams
->stream(kStreamOriginAddresses
)))
374 if (!ReadVector(&ops_
, streams
->stream(kStreamOps
)))
376 if (!ReadVector(©_counts_
, streams
->stream(kStreamCopyCounts
)))
378 if (!ReadVectorU8(©_bytes_
, streams
->stream(kStreamBytes
)))
380 if (!ReadVector(&abs32_ix_
, streams
->stream(kStreamAbs32Indexes
)))
382 if (!ReadVector(&rel32_ix_
, streams
->stream(kStreamRel32Indexes
)))
385 // Check that streams have been completely consumed.
386 for (int i
= 0; i
< kStreamLimit
; ++i
) {
387 if (streams
->stream(i
)->Remaining() > 0)
394 // Safe, non-throwing version of std::vector::at(). Returns 'true' for success,
395 // 'false' for out-of-bounds index error.
396 template<typename V
, typename T
>
397 bool VectorAt(const V
& v
, size_t index
, T
* output
) {
398 if (index
>= v
.size())
404 CheckBool
EncodedProgram::EvaluateRel32ARM(OP op
,
407 SinkStream
* output
) {
408 switch (op
& 0x0000F000) {
411 if (!VectorAt(rel32_ix_
, ix_rel32_ix
, &index
))
415 if (!VectorAt(rel32_rva_
, index
, &rva
))
417 uint32 decompressed_op
;
418 if (!DisassemblerElf32ARM::Decompress(ARM_OFF8
,
419 static_cast<uint16
>(op
),
420 static_cast<uint32
>(rva
-
425 uint16 op16
= static_cast<uint16
>(decompressed_op
);
426 if (!output
->Write(&op16
, 2))
433 if (!VectorAt(rel32_ix_
, ix_rel32_ix
, &index
))
437 if (!VectorAt(rel32_rva_
, index
, &rva
))
439 uint32 decompressed_op
;
440 if (!DisassemblerElf32ARM::Decompress(ARM_OFF11
, (uint16
) op
,
441 (uint32
) (rva
- current_rva
),
445 uint16 op16
= static_cast<uint16
>(decompressed_op
);
446 if (!output
->Write(&op16
, 2))
453 if (!VectorAt(rel32_ix_
, ix_rel32_ix
, &index
))
457 if (!VectorAt(rel32_rva_
, index
, &rva
))
459 uint32 decompressed_op
;
460 if (!DisassemblerElf32ARM::Decompress(ARM_OFF24
, (uint16
) op
,
461 (uint32
) (rva
- current_rva
),
465 if (!output
->Write(&decompressed_op
, 4))
472 if (!VectorAt(rel32_ix_
, ix_rel32_ix
, &index
))
476 if (!VectorAt(rel32_rva_
, index
, &rva
))
478 uint32 decompressed_op
;
479 if (!DisassemblerElf32ARM::Decompress(ARM_OFF25
, (uint16
) op
,
480 (uint32
) (rva
- current_rva
),
484 uint32 words
= (decompressed_op
<< 16) | (decompressed_op
>> 16);
485 if (!output
->Write(&words
, 4))
492 if (!VectorAt(rel32_ix_
, ix_rel32_ix
, &index
))
496 if (!VectorAt(rel32_rva_
, index
, &rva
))
498 uint32 decompressed_op
;
499 if (!DisassemblerElf32ARM::Decompress(ARM_OFF21
, (uint16
) op
,
500 (uint32
) (rva
- current_rva
),
504 uint32 words
= (decompressed_op
<< 16) | (decompressed_op
>> 16);
505 if (!output
->Write(&words
, 4))
517 CheckBool
EncodedProgram::AssembleTo(SinkStream
* final_buffer
) {
518 // For the most part, the assembly process walks the various tables.
519 // ix_mumble is the index into the mumble table.
520 size_t ix_origins
= 0;
521 size_t ix_copy_counts
= 0;
522 size_t ix_copy_bytes
= 0;
523 size_t ix_abs32_ix
= 0;
524 size_t ix_rel32_ix
= 0;
528 bool pending_pe_relocation_table
= false;
529 uint8 pending_pe_relocation_table_type
= 0x03; // IMAGE_REL_BASED_HIGHLOW
530 Elf32_Word pending_elf_relocation_table_type
= 0;
531 SinkStream bytes_following_relocation_table
;
533 SinkStream
* output
= final_buffer
;
535 for (size_t ix_ops
= 0; ix_ops
< ops_
.size(); ++ix_ops
) {
536 OP op
= ops_
[ix_ops
];
540 if (!EvaluateRel32ARM(op
, ix_rel32_ix
, current_rva
, output
))
546 if (!VectorAt(origins_
, ix_origins
, §ion_rva
))
549 current_rva
= section_rva
;
555 if (!VectorAt(copy_counts_
, ix_copy_counts
, &count
))
558 for (size_t i
= 0; i
< count
; ++i
) {
560 if (!VectorAt(copy_bytes_
, ix_copy_bytes
, &b
))
563 if (!output
->Write(&b
, 1))
566 current_rva
+= static_cast<RVA
>(count
);
572 if (!VectorAt(copy_bytes_
, ix_copy_bytes
, &b
))
575 if (!output
->Write(&b
, 1))
583 if (!VectorAt(rel32_ix_
, ix_rel32_ix
, &index
))
587 if (!VectorAt(rel32_rva_
, index
, &rva
))
589 uint32 offset
= (rva
- (current_rva
+ 4));
590 if (!output
->Write(&offset
, 4))
598 if (!VectorAt(abs32_ix_
, ix_abs32_ix
, &index
))
602 if (!VectorAt(abs32_rva_
, index
, &rva
))
604 uint32 abs32
= static_cast<uint32
>(rva
+ image_base_
);
605 if (!abs32_relocs_
.push_back(current_rva
) || !output
->Write(&abs32
, 4))
611 case MAKE_PE_RELOCATION_TABLE
: {
612 // We can see the base relocation anywhere, but we only have the
613 // information to generate it at the very end. So we divert the bytes
614 // we are generating to a temporary stream.
615 if (pending_pe_relocation_table
)
616 return false; // Can't have two base relocation tables.
618 pending_pe_relocation_table
= true;
619 output
= &bytes_following_relocation_table
;
621 // There is a potential problem *if* the instruction stream contains
622 // some REL32 relocations following the base relocation and in the same
623 // section. We don't know the size of the table, so 'current_rva' will
624 // be wrong, causing REL32 offsets to be miscalculated. This never
625 // happens; the base relocation table is usually in a section of its
626 // own, a data-only section, and following everything else in the
627 // executable except some padding zero bytes. We could fix this by
628 // emitting an ORIGIN after the MAKE_BASE_RELOCATION_TABLE.
631 case MAKE_PE64_RELOCATION_TABLE
: {
632 if (pending_pe_relocation_table
)
633 return false; // Can't have two base relocation tables.
635 pending_pe_relocation_table
= true;
636 pending_pe_relocation_table_type
= 0x0A; // IMAGE_REL_BASED_DIR64
637 output
= &bytes_following_relocation_table
;
641 case MAKE_ELF_ARM_RELOCATION_TABLE
: {
642 // We can see the base relocation anywhere, but we only have the
643 // information to generate it at the very end. So we divert the bytes
644 // we are generating to a temporary stream.
645 if (pending_elf_relocation_table_type
)
646 return false; // Can't have two base relocation tables.
648 pending_elf_relocation_table_type
= R_ARM_RELATIVE
;
649 output
= &bytes_following_relocation_table
;
653 case MAKE_ELF_RELOCATION_TABLE
: {
654 // We can see the base relocation anywhere, but we only have the
655 // information to generate it at the very end. So we divert the bytes
656 // we are generating to a temporary stream.
657 if (pending_elf_relocation_table_type
)
658 return false; // Can't have two base relocation tables.
660 pending_elf_relocation_table_type
= R_386_RELATIVE
;
661 output
= &bytes_following_relocation_table
;
667 if (pending_pe_relocation_table
) {
668 if (!GeneratePeRelocations(final_buffer
,
669 pending_pe_relocation_table_type
) ||
670 !final_buffer
->Append(&bytes_following_relocation_table
))
674 if (pending_elf_relocation_table_type
) {
675 if (!GenerateElfRelocations(pending_elf_relocation_table_type
,
677 !final_buffer
->Append(&bytes_following_relocation_table
))
681 // Final verification check: did we consume all lists?
682 if (ix_copy_counts
!= copy_counts_
.size())
684 if (ix_copy_bytes
!= copy_bytes_
.size())
686 if (ix_abs32_ix
!= abs32_ix_
.size())
688 if (ix_rel32_ix
!= rel32_ix_
.size())
694 // RelocBlock has the layout of a block of relocations in the base relocation
695 // table file format.
697 struct RelocBlockPOD
{
700 uint16 relocs
[4096]; // Allow up to one relocation per byte of a 4k page.
703 static_assert(offsetof(RelocBlockPOD
, relocs
) == 8, "reloc block header size");
708 pod
.page_rva
= 0xFFFFFFFF;
712 void Add(uint16 item
) {
713 pod
.relocs
[(pod
.block_size
-8)/2] = item
;
717 CheckBool
Flush(SinkStream
* buffer
) WARN_UNUSED_RESULT
{
719 if (pod
.block_size
!= 8) {
720 if (pod
.block_size
% 4 != 0) { // Pad to make size multiple of 4 bytes.
723 ok
= buffer
->Write(&pod
, pod
.block_size
);
731 CheckBool
EncodedProgram::GeneratePeRelocations(SinkStream
* buffer
,
733 std::sort(abs32_relocs_
.begin(), abs32_relocs_
.end());
738 for (size_t i
= 0; ok
&& i
< abs32_relocs_
.size(); ++i
) {
739 uint32 rva
= abs32_relocs_
[i
];
740 uint32 page_rva
= rva
& ~0xFFF;
741 if (page_rva
!= block
.pod
.page_rva
) {
742 ok
&= block
.Flush(buffer
);
743 block
.pod
.page_rva
= page_rva
;
746 block
.Add(((static_cast<uint16
>(type
)) << 12) | (rva
& 0xFFF));
748 ok
&= block
.Flush(buffer
);
752 CheckBool
EncodedProgram::GenerateElfRelocations(Elf32_Word r_info
,
753 SinkStream
* buffer
) {
754 std::sort(abs32_relocs_
.begin(), abs32_relocs_
.end());
756 Elf32_Rel relocation_block
;
758 relocation_block
.r_info
= r_info
;
761 for (size_t i
= 0; ok
&& i
< abs32_relocs_
.size(); ++i
) {
762 relocation_block
.r_offset
= abs32_relocs_
[i
];
763 ok
= buffer
->Write(&relocation_block
, sizeof(Elf32_Rel
));
768 ////////////////////////////////////////////////////////////////////////////////
770 Status
WriteEncodedProgram(EncodedProgram
* encoded
, SinkStreamSet
* sink
) {
771 if (!encoded
->WriteTo(sink
))
772 return C_STREAM_ERROR
;
776 Status
ReadEncodedProgram(SourceStreamSet
* streams
, EncodedProgram
** output
) {
777 EncodedProgram
* encoded
= new EncodedProgram();
778 if (encoded
->ReadFrom(streams
)) {
783 return C_DESERIALIZATION_FAILED
;
786 Status
Assemble(EncodedProgram
* encoded
, SinkStream
* buffer
) {
787 bool assembled
= encoded
->AssembleTo(buffer
);
790 return C_ASSEMBLY_FAILED
;
793 void DeleteEncodedProgram(EncodedProgram
* encoded
) {
797 } // namespace courgette