1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/tools/profiler/LulDwarfExt.h Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,1272 @@
1.4 +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
1.5 +/* vim: set ts=8 sts=2 et sw=2 tw=80: */
1.6 +
1.7 +// Copyright 2006, 2010 Google Inc. All Rights Reserved.
1.8 +//
1.9 +// Redistribution and use in source and binary forms, with or without
1.10 +// modification, are permitted provided that the following conditions are
1.11 +// met:
1.12 +//
1.13 +// * Redistributions of source code must retain the above copyright
1.14 +// notice, this list of conditions and the following disclaimer.
1.15 +// * Redistributions in binary form must reproduce the above
1.16 +// copyright notice, this list of conditions and the following disclaimer
1.17 +// in the documentation and/or other materials provided with the
1.18 +// distribution.
1.19 +// * Neither the name of Google Inc. nor the names of its
1.20 +// contributors may be used to endorse or promote products derived from
1.21 +// this software without specific prior written permission.
1.22 +//
1.23 +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
1.24 +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
1.25 +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
1.26 +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
1.27 +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
1.28 +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
1.29 +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
1.30 +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
1.31 +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
1.32 +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
1.33 +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
1.34 +
1.35 +// Original author: Jim Blandy <jimb@mozilla.com> <jimb@red-bean.com>
1.36 +
1.37 +// This file is derived from the following files in
1.38 +// toolkit/crashreporter/google-breakpad:
1.39 +// src/common/dwarf/types.h
1.40 +// src/common/dwarf/dwarf2enums.h
1.41 +// src/common/dwarf/bytereader.h
1.42 +// src/common/dwarf_cfi_to_module.h
1.43 +// src/common/dwarf/dwarf2reader.h
1.44 +
1.45 +#ifndef LulDwarfExt_h
1.46 +#define LulDwarfExt_h
1.47 +
1.48 +#include <stdint.h>
1.49 +
1.50 +#include "mozilla/Assertions.h"
1.51 +
1.52 +#include "LulDwarfSummariser.h"
1.53 +
1.54 +typedef signed char int8;
1.55 +typedef short int16;
1.56 +typedef int int32;
1.57 +typedef long long int64;
1.58 +
1.59 +typedef unsigned char uint8;
1.60 +typedef unsigned short uint16;
1.61 +typedef unsigned int uint32;
1.62 +typedef unsigned long long uint64;
1.63 +
1.64 +#ifdef __PTRDIFF_TYPE__
1.65 +typedef __PTRDIFF_TYPE__ intptr;
1.66 +typedef unsigned __PTRDIFF_TYPE__ uintptr;
1.67 +#else
1.68 +#error "Can't find pointer-sized integral types."
1.69 +#endif
1.70 +
1.71 +
1.72 +namespace lul {
1.73 +
1.74 +// Exception handling frame description pointer formats, as described
1.75 +// by the Linux Standard Base Core Specification 4.0, section 11.5,
1.76 +// DWARF Extensions.
1.77 +enum DwarfPointerEncoding
1.78 + {
1.79 + DW_EH_PE_absptr = 0x00,
1.80 + DW_EH_PE_omit = 0xff,
1.81 + DW_EH_PE_uleb128 = 0x01,
1.82 + DW_EH_PE_udata2 = 0x02,
1.83 + DW_EH_PE_udata4 = 0x03,
1.84 + DW_EH_PE_udata8 = 0x04,
1.85 + DW_EH_PE_sleb128 = 0x09,
1.86 + DW_EH_PE_sdata2 = 0x0A,
1.87 + DW_EH_PE_sdata4 = 0x0B,
1.88 + DW_EH_PE_sdata8 = 0x0C,
1.89 + DW_EH_PE_pcrel = 0x10,
1.90 + DW_EH_PE_textrel = 0x20,
1.91 + DW_EH_PE_datarel = 0x30,
1.92 + DW_EH_PE_funcrel = 0x40,
1.93 + DW_EH_PE_aligned = 0x50,
1.94 +
1.95 + // The GNU toolchain sources define this enum value as well,
1.96 + // simply to help classify the lower nybble values into signed and
1.97 + // unsigned groups.
1.98 + DW_EH_PE_signed = 0x08,
1.99 +
1.100 + // This is not documented in LSB 4.0, but it is used in both the
1.101 + // Linux and OS X toolchains. It can be added to any other
1.102 + // encoding (except DW_EH_PE_aligned), and indicates that the
1.103 + // encoded value represents the address at which the true address
1.104 + // is stored, not the true address itself.
1.105 + DW_EH_PE_indirect = 0x80
1.106 + };
1.107 +
1.108 +
1.109 +// We can't use the obvious name of LITTLE_ENDIAN and BIG_ENDIAN
1.110 +// because it conflicts with a macro
1.111 +enum Endianness {
1.112 + ENDIANNESS_BIG,
1.113 + ENDIANNESS_LITTLE
1.114 +};
1.115 +
1.116 +// A ByteReader knows how to read single- and multi-byte values of
1.117 +// various endiannesses, sizes, and encodings, as used in DWARF
1.118 +// debugging information and Linux C++ exception handling data.
1.119 +class ByteReader {
1.120 + public:
1.121 + // Construct a ByteReader capable of reading one-, two-, four-, and
1.122 + // eight-byte values according to ENDIANNESS, absolute machine-sized
1.123 + // addresses, DWARF-style "initial length" values, signed and
1.124 + // unsigned LEB128 numbers, and Linux C++ exception handling data's
1.125 + // encoded pointers.
1.126 + explicit ByteReader(enum Endianness endianness);
1.127 + virtual ~ByteReader();
1.128 +
1.129 + // Read a single byte from BUFFER and return it as an unsigned 8 bit
1.130 + // number.
1.131 + uint8 ReadOneByte(const char* buffer) const;
1.132 +
1.133 + // Read two bytes from BUFFER and return them as an unsigned 16 bit
1.134 + // number, using this ByteReader's endianness.
1.135 + uint16 ReadTwoBytes(const char* buffer) const;
1.136 +
1.137 + // Read four bytes from BUFFER and return them as an unsigned 32 bit
1.138 + // number, using this ByteReader's endianness. This function returns
1.139 + // a uint64 so that it is compatible with ReadAddress and
1.140 + // ReadOffset. The number it returns will never be outside the range
1.141 + // of an unsigned 32 bit integer.
1.142 + uint64 ReadFourBytes(const char* buffer) const;
1.143 +
1.144 + // Read eight bytes from BUFFER and return them as an unsigned 64
1.145 + // bit number, using this ByteReader's endianness.
1.146 + uint64 ReadEightBytes(const char* buffer) const;
1.147 +
1.148 + // Read an unsigned LEB128 (Little Endian Base 128) number from
1.149 + // BUFFER and return it as an unsigned 64 bit integer. Set LEN to
1.150 + // the number of bytes read.
1.151 + //
1.152 + // The unsigned LEB128 representation of an integer N is a variable
1.153 + // number of bytes:
1.154 + //
1.155 + // - If N is between 0 and 0x7f, then its unsigned LEB128
1.156 + // representation is a single byte whose value is N.
1.157 + //
1.158 + // - Otherwise, its unsigned LEB128 representation is (N & 0x7f) |
1.159 + // 0x80, followed by the unsigned LEB128 representation of N /
1.160 + // 128, rounded towards negative infinity.
1.161 + //
1.162 + // In other words, we break VALUE into groups of seven bits, put
1.163 + // them in little-endian order, and then write them as eight-bit
1.164 + // bytes with the high bit on all but the last.
1.165 + uint64 ReadUnsignedLEB128(const char* buffer, size_t* len) const;
1.166 +
1.167 + // Read a signed LEB128 number from BUFFER and return it as an
1.168 + // signed 64 bit integer. Set LEN to the number of bytes read.
1.169 + //
1.170 + // The signed LEB128 representation of an integer N is a variable
1.171 + // number of bytes:
1.172 + //
1.173 + // - If N is between -0x40 and 0x3f, then its signed LEB128
1.174 + // representation is a single byte whose value is N in two's
1.175 + // complement.
1.176 + //
1.177 + // - Otherwise, its signed LEB128 representation is (N & 0x7f) |
1.178 + // 0x80, followed by the signed LEB128 representation of N / 128,
1.179 + // rounded towards negative infinity.
1.180 + //
1.181 + // In other words, we break VALUE into groups of seven bits, put
1.182 + // them in little-endian order, and then write them as eight-bit
1.183 + // bytes with the high bit on all but the last.
1.184 + int64 ReadSignedLEB128(const char* buffer, size_t* len) const;
1.185 +
1.186 + // Indicate that addresses on this architecture are SIZE bytes long. SIZE
1.187 + // must be either 4 or 8. (DWARF allows addresses to be any number of
1.188 + // bytes in length from 1 to 255, but we only support 32- and 64-bit
1.189 + // addresses at the moment.) You must call this before using the
1.190 + // ReadAddress member function.
1.191 + //
1.192 + // For data in a .debug_info section, or something that .debug_info
1.193 + // refers to like line number or macro data, the compilation unit
1.194 + // header's address_size field indicates the address size to use. Call
1.195 + // frame information doesn't indicate its address size (a shortcoming of
1.196 + // the spec); you must supply the appropriate size based on the
1.197 + // architecture of the target machine.
1.198 + void SetAddressSize(uint8 size);
1.199 +
1.200 + // Return the current address size, in bytes. This is either 4,
1.201 + // indicating 32-bit addresses, or 8, indicating 64-bit addresses.
1.202 + uint8 AddressSize() const { return address_size_; }
1.203 +
1.204 + // Read an address from BUFFER and return it as an unsigned 64 bit
1.205 + // integer, respecting this ByteReader's endianness and address size. You
1.206 + // must call SetAddressSize before calling this function.
1.207 + uint64 ReadAddress(const char* buffer) const;
1.208 +
1.209 + // DWARF actually defines two slightly different formats: 32-bit DWARF
1.210 + // and 64-bit DWARF. This is *not* related to the size of registers or
1.211 + // addresses on the target machine; it refers only to the size of section
1.212 + // offsets and data lengths appearing in the DWARF data. One only needs
1.213 + // 64-bit DWARF when the debugging data itself is larger than 4GiB.
1.214 + // 32-bit DWARF can handle x86_64 or PPC64 code just fine, unless the
1.215 + // debugging data itself is very large.
1.216 + //
1.217 + // DWARF information identifies itself as 32-bit or 64-bit DWARF: each
1.218 + // compilation unit and call frame information entry begins with an
1.219 + // "initial length" field, which, in addition to giving the length of the
1.220 + // data, also indicates the size of section offsets and lengths appearing
1.221 + // in that data. The ReadInitialLength member function, below, reads an
1.222 + // initial length and sets the ByteReader's offset size as a side effect.
1.223 + // Thus, in the normal process of reading DWARF data, the appropriate
1.224 + // offset size is set automatically. So, you should only need to call
1.225 + // SetOffsetSize if you are using the same ByteReader to jump from the
1.226 + // midst of one block of DWARF data into another.
1.227 +
1.228 + // Read a DWARF "initial length" field from START, and return it as
1.229 + // an unsigned 64 bit integer, respecting this ByteReader's
1.230 + // endianness. Set *LEN to the length of the initial length in
1.231 + // bytes, either four or twelve. As a side effect, set this
1.232 + // ByteReader's offset size to either 4 (if we see a 32-bit DWARF
1.233 + // initial length) or 8 (if we see a 64-bit DWARF initial length).
1.234 + //
1.235 + // A DWARF initial length is either:
1.236 + //
1.237 + // - a byte count stored as an unsigned 32-bit value less than
1.238 + // 0xffffff00, indicating that the data whose length is being
1.239 + // measured uses the 32-bit DWARF format, or
1.240 + //
1.241 + // - The 32-bit value 0xffffffff, followed by a 64-bit byte count,
1.242 + // indicating that the data whose length is being measured uses
1.243 + // the 64-bit DWARF format.
1.244 + uint64 ReadInitialLength(const char* start, size_t* len);
1.245 +
1.246 + // Read an offset from BUFFER and return it as an unsigned 64 bit
1.247 + // integer, respecting the ByteReader's endianness. In 32-bit DWARF, the
1.248 + // offset is 4 bytes long; in 64-bit DWARF, the offset is eight bytes
1.249 + // long. You must call ReadInitialLength or SetOffsetSize before calling
1.250 + // this function; see the comments above for details.
1.251 + uint64 ReadOffset(const char* buffer) const;
1.252 +
1.253 + // Return the current offset size, in bytes.
1.254 + // A return value of 4 indicates that we are reading 32-bit DWARF.
1.255 + // A return value of 8 indicates that we are reading 64-bit DWARF.
1.256 + uint8 OffsetSize() const { return offset_size_; }
1.257 +
1.258 + // Indicate that section offsets and lengths are SIZE bytes long. SIZE
1.259 + // must be either 4 (meaning 32-bit DWARF) or 8 (meaning 64-bit DWARF).
1.260 + // Usually, you should not call this function yourself; instead, let a
1.261 + // call to ReadInitialLength establish the data's offset size
1.262 + // automatically.
1.263 + void SetOffsetSize(uint8 size);
1.264 +
1.265 + // The Linux C++ ABI uses a variant of DWARF call frame information
1.266 + // for exception handling. This data is included in the program's
1.267 + // address space as the ".eh_frame" section, and intepreted at
1.268 + // runtime to walk the stack, find exception handlers, and run
1.269 + // cleanup code. The format is mostly the same as DWARF CFI, with
1.270 + // some adjustments made to provide the additional
1.271 + // exception-handling data, and to make the data easier to work with
1.272 + // in memory --- for example, to allow it to be placed in read-only
1.273 + // memory even when describing position-independent code.
1.274 + //
1.275 + // In particular, exception handling data can select a number of
1.276 + // different encodings for pointers that appear in the data, as
1.277 + // described by the DwarfPointerEncoding enum. There are actually
1.278 + // four axes(!) to the encoding:
1.279 + //
1.280 + // - The pointer size: pointers can be 2, 4, or 8 bytes long, or use
1.281 + // the DWARF LEB128 encoding.
1.282 + //
1.283 + // - The pointer's signedness: pointers can be signed or unsigned.
1.284 + //
1.285 + // - The pointer's base address: the data stored in the exception
1.286 + // handling data can be the actual address (that is, an absolute
1.287 + // pointer), or relative to one of a number of different base
1.288 + // addreses --- including that of the encoded pointer itself, for
1.289 + // a form of "pc-relative" addressing.
1.290 + //
1.291 + // - The pointer may be indirect: it may be the address where the
1.292 + // true pointer is stored. (This is used to refer to things via
1.293 + // global offset table entries, program linkage table entries, or
1.294 + // other tricks used in position-independent code.)
1.295 + //
1.296 + // There are also two options that fall outside that matrix
1.297 + // altogether: the pointer may be omitted, or it may have padding to
1.298 + // align it on an appropriate address boundary. (That last option
1.299 + // may seem like it should be just another axis, but it is not.)
1.300 +
1.301 + // Indicate that the exception handling data is loaded starting at
1.302 + // SECTION_BASE, and that the start of its buffer in our own memory
1.303 + // is BUFFER_BASE. This allows us to find the address that a given
1.304 + // byte in our buffer would have when loaded into the program the
1.305 + // data describes. We need this to resolve DW_EH_PE_pcrel pointers.
1.306 + void SetCFIDataBase(uint64 section_base, const char *buffer_base);
1.307 +
1.308 + // Indicate that the base address of the program's ".text" section
1.309 + // is TEXT_BASE. We need this to resolve DW_EH_PE_textrel pointers.
1.310 + void SetTextBase(uint64 text_base);
1.311 +
1.312 + // Indicate that the base address for DW_EH_PE_datarel pointers is
1.313 + // DATA_BASE. The proper value depends on the ABI; it is usually the
1.314 + // address of the global offset table, held in a designated register in
1.315 + // position-independent code. You will need to look at the startup code
1.316 + // for the target system to be sure. I tried; my eyes bled.
1.317 + void SetDataBase(uint64 data_base);
1.318 +
1.319 + // Indicate that the base address for the FDE we are processing is
1.320 + // FUNCTION_BASE. This is the start address of DW_EH_PE_funcrel
1.321 + // pointers. (This encoding does not seem to be used by the GNU
1.322 + // toolchain.)
1.323 + void SetFunctionBase(uint64 function_base);
1.324 +
1.325 + // Indicate that we are no longer processing any FDE, so any use of
1.326 + // a DW_EH_PE_funcrel encoding is an error.
1.327 + void ClearFunctionBase();
1.328 +
1.329 + // Return true if ENCODING is a valid pointer encoding.
1.330 + bool ValidEncoding(DwarfPointerEncoding encoding) const;
1.331 +
1.332 + // Return true if we have all the information we need to read a
1.333 + // pointer that uses ENCODING. This checks that the appropriate
1.334 + // SetFooBase function for ENCODING has been called.
1.335 + bool UsableEncoding(DwarfPointerEncoding encoding) const;
1.336 +
1.337 + // Read an encoded pointer from BUFFER using ENCODING; return the
1.338 + // absolute address it represents, and set *LEN to the pointer's
1.339 + // length in bytes, including any padding for aligned pointers.
1.340 + //
1.341 + // This function calls 'abort' if ENCODING is invalid or refers to a
1.342 + // base address this reader hasn't been given, so you should check
1.343 + // with ValidEncoding and UsableEncoding first if you would rather
1.344 + // die in a more helpful way.
1.345 + uint64 ReadEncodedPointer(const char *buffer, DwarfPointerEncoding encoding,
1.346 + size_t *len) const;
1.347 +
1.348 + private:
1.349 +
1.350 + // Function pointer type for our address and offset readers.
1.351 + typedef uint64 (ByteReader::*AddressReader)(const char*) const;
1.352 +
1.353 + // Read an offset from BUFFER and return it as an unsigned 64 bit
1.354 + // integer. DWARF2/3 define offsets as either 4 or 8 bytes,
1.355 + // generally depending on the amount of DWARF2/3 info present.
1.356 + // This function pointer gets set by SetOffsetSize.
1.357 + AddressReader offset_reader_;
1.358 +
1.359 + // Read an address from BUFFER and return it as an unsigned 64 bit
1.360 + // integer. DWARF2/3 allow addresses to be any size from 0-255
1.361 + // bytes currently. Internally we support 4 and 8 byte addresses,
1.362 + // and will CHECK on anything else.
1.363 + // This function pointer gets set by SetAddressSize.
1.364 + AddressReader address_reader_;
1.365 +
1.366 + Endianness endian_;
1.367 + uint8 address_size_;
1.368 + uint8 offset_size_;
1.369 +
1.370 + // Base addresses for Linux C++ exception handling data's encoded pointers.
1.371 + bool have_section_base_, have_text_base_, have_data_base_;
1.372 + bool have_function_base_;
1.373 + uint64 section_base_;
1.374 + uint64 text_base_, data_base_, function_base_;
1.375 + const char *buffer_base_;
1.376 +};
1.377 +
1.378 +
1.379 +inline uint8 ByteReader::ReadOneByte(const char* buffer) const {
1.380 + return buffer[0];
1.381 +}
1.382 +
1.383 +inline uint16 ByteReader::ReadTwoBytes(const char* signed_buffer) const {
1.384 + const unsigned char *buffer
1.385 + = reinterpret_cast<const unsigned char *>(signed_buffer);
1.386 + const uint16 buffer0 = buffer[0];
1.387 + const uint16 buffer1 = buffer[1];
1.388 + if (endian_ == ENDIANNESS_LITTLE) {
1.389 + return buffer0 | buffer1 << 8;
1.390 + } else {
1.391 + return buffer1 | buffer0 << 8;
1.392 + }
1.393 +}
1.394 +
1.395 +inline uint64 ByteReader::ReadFourBytes(const char* signed_buffer) const {
1.396 + const unsigned char *buffer
1.397 + = reinterpret_cast<const unsigned char *>(signed_buffer);
1.398 + const uint32 buffer0 = buffer[0];
1.399 + const uint32 buffer1 = buffer[1];
1.400 + const uint32 buffer2 = buffer[2];
1.401 + const uint32 buffer3 = buffer[3];
1.402 + if (endian_ == ENDIANNESS_LITTLE) {
1.403 + return buffer0 | buffer1 << 8 | buffer2 << 16 | buffer3 << 24;
1.404 + } else {
1.405 + return buffer3 | buffer2 << 8 | buffer1 << 16 | buffer0 << 24;
1.406 + }
1.407 +}
1.408 +
1.409 +inline uint64 ByteReader::ReadEightBytes(const char* signed_buffer) const {
1.410 + const unsigned char *buffer
1.411 + = reinterpret_cast<const unsigned char *>(signed_buffer);
1.412 + const uint64 buffer0 = buffer[0];
1.413 + const uint64 buffer1 = buffer[1];
1.414 + const uint64 buffer2 = buffer[2];
1.415 + const uint64 buffer3 = buffer[3];
1.416 + const uint64 buffer4 = buffer[4];
1.417 + const uint64 buffer5 = buffer[5];
1.418 + const uint64 buffer6 = buffer[6];
1.419 + const uint64 buffer7 = buffer[7];
1.420 + if (endian_ == ENDIANNESS_LITTLE) {
1.421 + return buffer0 | buffer1 << 8 | buffer2 << 16 | buffer3 << 24 |
1.422 + buffer4 << 32 | buffer5 << 40 | buffer6 << 48 | buffer7 << 56;
1.423 + } else {
1.424 + return buffer7 | buffer6 << 8 | buffer5 << 16 | buffer4 << 24 |
1.425 + buffer3 << 32 | buffer2 << 40 | buffer1 << 48 | buffer0 << 56;
1.426 + }
1.427 +}
1.428 +
1.429 +// Read an unsigned LEB128 number. Each byte contains 7 bits of
1.430 +// information, plus one bit saying whether the number continues or
1.431 +// not.
1.432 +
1.433 +inline uint64 ByteReader::ReadUnsignedLEB128(const char* buffer,
1.434 + size_t* len) const {
1.435 + uint64 result = 0;
1.436 + size_t num_read = 0;
1.437 + unsigned int shift = 0;
1.438 + unsigned char byte;
1.439 +
1.440 + do {
1.441 + byte = *buffer++;
1.442 + num_read++;
1.443 +
1.444 + result |= (static_cast<uint64>(byte & 0x7f)) << shift;
1.445 +
1.446 + shift += 7;
1.447 +
1.448 + } while (byte & 0x80);
1.449 +
1.450 + *len = num_read;
1.451 +
1.452 + return result;
1.453 +}
1.454 +
1.455 +// Read a signed LEB128 number. These are like regular LEB128
1.456 +// numbers, except the last byte may have a sign bit set.
1.457 +
1.458 +inline int64 ByteReader::ReadSignedLEB128(const char* buffer,
1.459 + size_t* len) const {
1.460 + int64 result = 0;
1.461 + unsigned int shift = 0;
1.462 + size_t num_read = 0;
1.463 + unsigned char byte;
1.464 +
1.465 + do {
1.466 + byte = *buffer++;
1.467 + num_read++;
1.468 + result |= (static_cast<uint64>(byte & 0x7f) << shift);
1.469 + shift += 7;
1.470 + } while (byte & 0x80);
1.471 +
1.472 + if ((shift < 8 * sizeof (result)) && (byte & 0x40))
1.473 + result |= -((static_cast<int64>(1)) << shift);
1.474 + *len = num_read;
1.475 + return result;
1.476 +}
1.477 +
1.478 +inline uint64 ByteReader::ReadOffset(const char* buffer) const {
1.479 + MOZ_ASSERT(this->offset_reader_);
1.480 + return (this->*offset_reader_)(buffer);
1.481 +}
1.482 +
1.483 +inline uint64 ByteReader::ReadAddress(const char* buffer) const {
1.484 + MOZ_ASSERT(this->address_reader_);
1.485 + return (this->*address_reader_)(buffer);
1.486 +}
1.487 +
1.488 +inline void ByteReader::SetCFIDataBase(uint64 section_base,
1.489 + const char *buffer_base) {
1.490 + section_base_ = section_base;
1.491 + buffer_base_ = buffer_base;
1.492 + have_section_base_ = true;
1.493 +}
1.494 +
1.495 +inline void ByteReader::SetTextBase(uint64 text_base) {
1.496 + text_base_ = text_base;
1.497 + have_text_base_ = true;
1.498 +}
1.499 +
1.500 +inline void ByteReader::SetDataBase(uint64 data_base) {
1.501 + data_base_ = data_base;
1.502 + have_data_base_ = true;
1.503 +}
1.504 +
1.505 +inline void ByteReader::SetFunctionBase(uint64 function_base) {
1.506 + function_base_ = function_base;
1.507 + have_function_base_ = true;
1.508 +}
1.509 +
1.510 +inline void ByteReader::ClearFunctionBase() {
1.511 + have_function_base_ = false;
1.512 +}
1.513 +
1.514 +
1.515 +// (derived from)
1.516 +// dwarf_cfi_to_module.h: Define the DwarfCFIToModule class, which
1.517 +// accepts parsed DWARF call frame info and adds it to a Summariser object.
1.518 +
1.519 +// This class is a reader for DWARF's Call Frame Information. CFI
1.520 +// describes how to unwind stack frames --- even for functions that do
1.521 +// not follow fixed conventions for saving registers, whose frame size
1.522 +// varies as they execute, etc.
1.523 +//
1.524 +// CFI describes, at each machine instruction, how to compute the
1.525 +// stack frame's base address, how to find the return address, and
1.526 +// where to find the saved values of the caller's registers (if the
1.527 +// callee has stashed them somewhere to free up the registers for its
1.528 +// own use).
1.529 +//
1.530 +// For example, suppose we have a function whose machine code looks
1.531 +// like this (imagine an assembly language that looks like C, for a
1.532 +// machine with 32-bit registers, and a stack that grows towards lower
1.533 +// addresses):
1.534 +//
1.535 +// func: ; entry point; return address at sp
1.536 +// func+0: sp = sp - 16 ; allocate space for stack frame
1.537 +// func+1: sp[12] = r0 ; save r0 at sp+12
1.538 +// ... ; other code, not frame-related
1.539 +// func+10: sp -= 4; *sp = x ; push some x on the stack
1.540 +// ... ; other code, not frame-related
1.541 +// func+20: r0 = sp[16] ; restore saved r0
1.542 +// func+21: sp += 20 ; pop whole stack frame
1.543 +// func+22: pc = *sp; sp += 4 ; pop return address and jump to it
1.544 +//
1.545 +// DWARF CFI is (a very compressed representation of) a table with a
1.546 +// row for each machine instruction address and a column for each
1.547 +// register showing how to restore it, if possible.
1.548 +//
1.549 +// A special column named "CFA", for "Canonical Frame Address", tells how
1.550 +// to compute the base address of the frame; registers' entries may
1.551 +// refer to the CFA in describing where the registers are saved.
1.552 +//
1.553 +// Another special column, named "RA", represents the return address.
1.554 +//
1.555 +// For example, here is a complete (uncompressed) table describing the
1.556 +// function above:
1.557 +//
1.558 +// insn cfa r0 r1 ... ra
1.559 +// =======================================
1.560 +// func+0: sp cfa[0]
1.561 +// func+1: sp+16 cfa[0]
1.562 +// func+2: sp+16 cfa[-4] cfa[0]
1.563 +// func+11: sp+20 cfa[-4] cfa[0]
1.564 +// func+21: sp+20 cfa[0]
1.565 +// func+22: sp cfa[0]
1.566 +//
1.567 +// Some things to note here:
1.568 +//
1.569 +// - Each row describes the state of affairs *before* executing the
1.570 +// instruction at the given address. Thus, the row for func+0
1.571 +// describes the state before we allocate the stack frame. In the
1.572 +// next row, the formula for computing the CFA has changed,
1.573 +// reflecting that allocation.
1.574 +//
1.575 +// - The other entries are written in terms of the CFA; this allows
1.576 +// them to remain unchanged as the stack pointer gets bumped around.
1.577 +// For example, the rule for recovering the return address (the "ra"
1.578 +// column) remains unchanged throughout the function, even as the
1.579 +// stack pointer takes on three different offsets from the return
1.580 +// address.
1.581 +//
1.582 +// - Although we haven't shown it, most calling conventions designate
1.583 +// "callee-saves" and "caller-saves" registers. The callee must
1.584 +// preserve the values of callee-saves registers; if it uses them,
1.585 +// it must save their original values somewhere, and restore them
1.586 +// before it returns. In contrast, the callee is free to trash
1.587 +// caller-saves registers; if the callee uses these, it will
1.588 +// probably not bother to save them anywhere, and the CFI will
1.589 +// probably mark their values as "unrecoverable".
1.590 +//
1.591 +// (However, since the caller cannot assume the callee was going to
1.592 +// save them, caller-saves registers are probably dead in the caller
1.593 +// anyway, so compilers usually don't generate CFA for caller-saves
1.594 +// registers.)
1.595 +//
1.596 +// - Exactly where the CFA points is a matter of convention that
1.597 +// depends on the architecture and ABI in use. In the example, the
1.598 +// CFA is the value the stack pointer had upon entry to the
1.599 +// function, pointing at the saved return address. But on the x86,
1.600 +// the call frame information generated by GCC follows the
1.601 +// convention that the CFA is the address *after* the saved return
1.602 +// address.
1.603 +//
1.604 +// But by definition, the CFA remains constant throughout the
1.605 +// lifetime of the frame. This makes it a useful value for other
1.606 +// columns to refer to. It is also gives debuggers a useful handle
1.607 +// for identifying a frame.
1.608 +//
1.609 +// If you look at the table above, you'll notice that a given entry is
1.610 +// often the same as the one immediately above it: most instructions
1.611 +// change only one or two aspects of the stack frame, if they affect
1.612 +// it at all. The DWARF format takes advantage of this fact, and
1.613 +// reduces the size of the data by mentioning only the addresses and
1.614 +// columns at which changes take place. So for the above, DWARF CFI
1.615 +// data would only actually mention the following:
1.616 +//
1.617 +// insn cfa r0 r1 ... ra
1.618 +// =======================================
1.619 +// func+0: sp cfa[0]
1.620 +// func+1: sp+16
1.621 +// func+2: cfa[-4]
1.622 +// func+11: sp+20
1.623 +// func+21: r0
1.624 +// func+22: sp
1.625 +//
1.626 +// In fact, this is the way the parser reports CFI to the consumer: as
1.627 +// a series of statements of the form, "At address X, column Y changed
1.628 +// to Z," and related conventions for describing the initial state.
1.629 +//
1.630 +// Naturally, it would be impractical to have to scan the entire
1.631 +// program's CFI, noting changes as we go, just to recover the
1.632 +// unwinding rules in effect at one particular instruction. To avoid
1.633 +// this, CFI data is grouped into "entries", each of which covers a
1.634 +// specified range of addresses and begins with a complete statement
1.635 +// of the rules for all recoverable registers at that starting
1.636 +// address. Each entry typically covers a single function.
1.637 +//
1.638 +// Thus, to compute the contents of a given row of the table --- that
1.639 +// is, rules for recovering the CFA, RA, and registers at a given
1.640 +// instruction --- the consumer should find the entry that covers that
1.641 +// instruction's address, start with the initial state supplied at the
1.642 +// beginning of the entry, and work forward until it has processed all
1.643 +// the changes up to and including those for the present instruction.
1.644 +//
1.645 +// There are seven kinds of rules that can appear in an entry of the
1.646 +// table:
1.647 +//
1.648 +// - "undefined": The given register is not preserved by the callee;
1.649 +// its value cannot be recovered.
1.650 +//
1.651 +// - "same value": This register has the same value it did in the callee.
1.652 +//
1.653 +// - offset(N): The register is saved at offset N from the CFA.
1.654 +//
1.655 +// - val_offset(N): The value the register had in the caller is the
1.656 +// CFA plus offset N. (This is usually only useful for describing
1.657 +// the stack pointer.)
1.658 +//
1.659 +// - register(R): The register's value was saved in another register R.
1.660 +//
1.661 +// - expression(E): Evaluating the DWARF expression E using the
1.662 +// current frame's registers' values yields the address at which the
1.663 +// register was saved.
1.664 +//
1.665 +// - val_expression(E): Evaluating the DWARF expression E using the
1.666 +// current frame's registers' values yields the value the register
1.667 +// had in the caller.
1.668 +
1.669 +class CallFrameInfo {
1.670 + public:
1.671 + // The different kinds of entries one finds in CFI. Used internally,
1.672 + // and for error reporting.
1.673 + enum EntryKind { kUnknown, kCIE, kFDE, kTerminator };
1.674 +
1.675 + // The handler class to which the parser hands the parsed call frame
1.676 + // information. Defined below.
1.677 + class Handler;
1.678 +
1.679 + // A reporter class, which CallFrameInfo uses to report errors
1.680 + // encountered while parsing call frame information. Defined below.
1.681 + class Reporter;
1.682 +
1.683 + // Create a DWARF CFI parser. BUFFER points to the contents of the
1.684 + // .debug_frame section to parse; BUFFER_LENGTH is its length in bytes.
1.685 + // REPORTER is an error reporter the parser should use to report
1.686 + // problems. READER is a ByteReader instance that has the endianness and
1.687 + // address size set properly. Report the data we find to HANDLER.
1.688 + //
1.689 + // This class can also parse Linux C++ exception handling data, as found
1.690 + // in '.eh_frame' sections. This data is a variant of DWARF CFI that is
1.691 + // placed in loadable segments so that it is present in the program's
1.692 + // address space, and is interpreted by the C++ runtime to search the
1.693 + // call stack for a handler interested in the exception being thrown,
1.694 + // actually pop the frames, and find cleanup code to run.
1.695 + //
1.696 + // There are two differences between the call frame information described
1.697 + // in the DWARF standard and the exception handling data Linux places in
1.698 + // the .eh_frame section:
1.699 + //
1.700 + // - Exception handling data uses uses a different format for call frame
1.701 + // information entry headers. The distinguished CIE id, the way FDEs
1.702 + // refer to their CIEs, and the way the end of the series of entries is
1.703 + // determined are all slightly different.
1.704 + //
1.705 + // If the constructor's EH_FRAME argument is true, then the
1.706 + // CallFrameInfo parses the entry headers as Linux C++ exception
1.707 + // handling data. If EH_FRAME is false or omitted, the CallFrameInfo
1.708 + // parses standard DWARF call frame information.
1.709 + //
1.710 + // - Linux C++ exception handling data uses CIE augmentation strings
1.711 + // beginning with 'z' to specify the presence of additional data after
1.712 + // the CIE and FDE headers and special encodings used for addresses in
1.713 + // frame description entries.
1.714 + //
1.715 + // CallFrameInfo can handle 'z' augmentations in either DWARF CFI or
1.716 + // exception handling data if you have supplied READER with the base
1.717 + // addresses needed to interpret the pointer encodings that 'z'
1.718 + // augmentations can specify. See the ByteReader interface for details
1.719 + // about the base addresses. See the CallFrameInfo::Handler interface
1.720 + // for details about the additional information one might find in
1.721 + // 'z'-augmented data.
1.722 + //
1.723 + // Thus:
1.724 + //
1.725 + // - If you are parsing standard DWARF CFI, as found in a .debug_frame
1.726 + // section, you should pass false for the EH_FRAME argument, or omit
1.727 + // it, and you need not worry about providing READER with the
1.728 + // additional base addresses.
1.729 + //
1.730 + // - If you want to parse Linux C++ exception handling data from a
1.731 + // .eh_frame section, you should pass EH_FRAME as true, and call
1.732 + // READER's Set*Base member functions before calling our Start method.
1.733 + //
1.734 + // - If you want to parse DWARF CFI that uses the 'z' augmentations
1.735 + // (although I don't think any toolchain ever emits such data), you
1.736 + // could pass false for EH_FRAME, but call READER's Set*Base members.
1.737 + //
1.738 + // The extensions the Linux C++ ABI makes to DWARF for exception
1.739 + // handling are described here, rather poorly:
1.740 + // http://refspecs.linux-foundation.org/LSB_4.0.0/LSB-Core-generic/LSB-Core-generic/dwarfext.html
1.741 + // http://refspecs.linux-foundation.org/LSB_4.0.0/LSB-Core-generic/LSB-Core-generic/ehframechpt.html
1.742 + //
1.743 + // The mechanics of C++ exception handling, personality routines,
1.744 + // and language-specific data areas are described here, rather nicely:
1.745 + // http://www.codesourcery.com/public/cxx-abi/abi-eh.html
1.746 +
1.747 + CallFrameInfo(const char *buffer, size_t buffer_length,
1.748 + ByteReader *reader, Handler *handler, Reporter *reporter,
1.749 + bool eh_frame = false)
1.750 + : buffer_(buffer), buffer_length_(buffer_length),
1.751 + reader_(reader), handler_(handler), reporter_(reporter),
1.752 + eh_frame_(eh_frame) { }
1.753 +
1.754 + ~CallFrameInfo() { }
1.755 +
1.756 + // Parse the entries in BUFFER, reporting what we find to HANDLER.
1.757 + // Return true if we reach the end of the section successfully, or
1.758 + // false if we encounter an error.
1.759 + bool Start();
1.760 +
1.761 + // Return the textual name of KIND. For error reporting.
1.762 + static const char *KindName(EntryKind kind);
1.763 +
1.764 + private:
1.765 +
1.766 + struct CIE;
1.767 +
1.768 + // A CFI entry, either an FDE or a CIE.
1.769 + struct Entry {
1.770 + // The starting offset of the entry in the section, for error
1.771 + // reporting.
1.772 + size_t offset;
1.773 +
1.774 + // The start of this entry in the buffer.
1.775 + const char *start;
1.776 +
1.777 + // Which kind of entry this is.
1.778 + //
1.779 + // We want to be able to use this for error reporting even while we're
1.780 + // in the midst of parsing. Error reporting code may assume that kind,
1.781 + // offset, and start fields are valid, although kind may be kUnknown.
1.782 + EntryKind kind;
1.783 +
1.784 + // The end of this entry's common prologue (initial length and id), and
1.785 + // the start of this entry's kind-specific fields.
1.786 + const char *fields;
1.787 +
1.788 + // The start of this entry's instructions.
1.789 + const char *instructions;
1.790 +
1.791 + // The address past the entry's last byte in the buffer. (Note that
1.792 + // since offset points to the entry's initial length field, and the
1.793 + // length field is the number of bytes after that field, this is not
1.794 + // simply buffer_ + offset + length.)
1.795 + const char *end;
1.796 +
1.797 + // For both DWARF CFI and .eh_frame sections, this is the CIE id in a
1.798 + // CIE, and the offset of the associated CIE in an FDE.
1.799 + uint64 id;
1.800 +
1.801 + // The CIE that applies to this entry, if we've parsed it. If this is a
1.802 + // CIE, then this field points to this structure.
1.803 + CIE *cie;
1.804 + };
1.805 +
1.806 + // A common information entry (CIE).
1.807 + struct CIE: public Entry {
1.808 + uint8 version; // CFI data version number
1.809 + std::string augmentation; // vendor format extension markers
1.810 + uint64 code_alignment_factor; // scale for code address adjustments
1.811 + int data_alignment_factor; // scale for stack pointer adjustments
1.812 + unsigned return_address_register; // which register holds the return addr
1.813 +
1.814 + // True if this CIE includes Linux C++ ABI 'z' augmentation data.
1.815 + bool has_z_augmentation;
1.816 +
1.817 + // Parsed 'z' augmentation data. These are meaningful only if
1.818 + // has_z_augmentation is true.
1.819 + bool has_z_lsda; // The 'z' augmentation included 'L'.
1.820 + bool has_z_personality; // The 'z' augmentation included 'P'.
1.821 + bool has_z_signal_frame; // The 'z' augmentation included 'S'.
1.822 +
1.823 + // If has_z_lsda is true, this is the encoding to be used for language-
1.824 + // specific data area pointers in FDEs.
1.825 + DwarfPointerEncoding lsda_encoding;
1.826 +
1.827 + // If has_z_personality is true, this is the encoding used for the
1.828 + // personality routine pointer in the augmentation data.
1.829 + DwarfPointerEncoding personality_encoding;
1.830 +
1.831 + // If has_z_personality is true, this is the address of the personality
1.832 + // routine --- or, if personality_encoding & DW_EH_PE_indirect, the
1.833 + // address where the personality routine's address is stored.
1.834 + uint64 personality_address;
1.835 +
1.836 + // This is the encoding used for addresses in the FDE header and
1.837 + // in DW_CFA_set_loc instructions. This is always valid, whether
1.838 + // or not we saw a 'z' augmentation string; its default value is
1.839 + // DW_EH_PE_absptr, which is what normal DWARF CFI uses.
1.840 + DwarfPointerEncoding pointer_encoding;
1.841 + };
1.842 +
1.843 + // A frame description entry (FDE).
1.844 + struct FDE: public Entry {
1.845 + uint64 address; // start address of described code
1.846 + uint64 size; // size of described code, in bytes
1.847 +
1.848 + // If cie->has_z_lsda is true, then this is the language-specific data
1.849 + // area's address --- or its address's address, if cie->lsda_encoding
1.850 + // has the DW_EH_PE_indirect bit set.
1.851 + uint64 lsda_address;
1.852 + };
1.853 +
1.854 + // Internal use.
1.855 + class Rule;
1.856 + class UndefinedRule;
1.857 + class SameValueRule;
1.858 + class OffsetRule;
1.859 + class ValOffsetRule;
1.860 + class RegisterRule;
1.861 + class ExpressionRule;
1.862 + class ValExpressionRule;
1.863 + class RuleMap;
1.864 + class State;
1.865 +
1.866 + // Parse the initial length and id of a CFI entry, either a CIE, an FDE,
1.867 + // or a .eh_frame end-of-data mark. CURSOR points to the beginning of the
1.868 + // data to parse. On success, populate ENTRY as appropriate, and return
1.869 + // true. On failure, report the problem, and return false. Even if we
1.870 + // return false, set ENTRY->end to the first byte after the entry if we
1.871 + // were able to figure that out, or NULL if we weren't.
1.872 + bool ReadEntryPrologue(const char *cursor, Entry *entry);
1.873 +
1.874 + // Parse the fields of a CIE after the entry prologue, including any 'z'
1.875 + // augmentation data. Assume that the 'Entry' fields of CIE are
1.876 + // populated; use CIE->fields and CIE->end as the start and limit for
1.877 + // parsing. On success, populate the rest of *CIE, and return true; on
1.878 + // failure, report the problem and return false.
1.879 + bool ReadCIEFields(CIE *cie);
1.880 +
1.881 + // Parse the fields of an FDE after the entry prologue, including any 'z'
1.882 + // augmentation data. Assume that the 'Entry' fields of *FDE are
1.883 + // initialized; use FDE->fields and FDE->end as the start and limit for
1.884 + // parsing. Assume that FDE->cie is fully initialized. On success,
1.885 + // populate the rest of *FDE, and return true; on failure, report the
1.886 + // problem and return false.
1.887 + bool ReadFDEFields(FDE *fde);
1.888 +
1.889 + // Report that ENTRY is incomplete, and return false. This is just a
1.890 + // trivial wrapper for invoking reporter_->Incomplete; it provides a
1.891 + // little brevity.
1.892 + bool ReportIncomplete(Entry *entry);
1.893 +
1.894 + // Return true if ENCODING has the DW_EH_PE_indirect bit set.
1.895 + static bool IsIndirectEncoding(DwarfPointerEncoding encoding) {
1.896 + return encoding & DW_EH_PE_indirect;
1.897 + }
1.898 +
1.899 + // The contents of the DWARF .debug_info section we're parsing.
1.900 + const char *buffer_;
1.901 + size_t buffer_length_;
1.902 +
1.903 + // For reading multi-byte values with the appropriate endianness.
1.904 + ByteReader *reader_;
1.905 +
1.906 + // The handler to which we should report the data we find.
1.907 + Handler *handler_;
1.908 +
1.909 + // For reporting problems in the info we're parsing.
1.910 + Reporter *reporter_;
1.911 +
1.912 + // True if we are processing .eh_frame-format data.
1.913 + bool eh_frame_;
1.914 +};
1.915 +
1.916 +
1.917 +// The handler class for CallFrameInfo. The a CFI parser calls the
1.918 +// member functions of a handler object to report the data it finds.
1.919 +class CallFrameInfo::Handler {
1.920 + public:
1.921 + // The pseudo-register number for the canonical frame address.
1.922 + enum { kCFARegister = DW_REG_CFA };
1.923 +
1.924 + Handler() { }
1.925 + virtual ~Handler() { }
1.926 +
1.927 + // The parser has found CFI for the machine code at ADDRESS,
1.928 + // extending for LENGTH bytes. OFFSET is the offset of the frame
1.929 + // description entry in the section, for use in error messages.
1.930 + // VERSION is the version number of the CFI format. AUGMENTATION is
1.931 + // a string describing any producer-specific extensions present in
1.932 + // the data. RETURN_ADDRESS is the number of the register that holds
1.933 + // the address to which the function should return.
1.934 + //
1.935 + // Entry should return true to process this CFI, or false to skip to
1.936 + // the next entry.
1.937 + //
1.938 + // The parser invokes Entry for each Frame Description Entry (FDE)
1.939 + // it finds. The parser doesn't report Common Information Entries
1.940 + // to the handler explicitly; instead, if the handler elects to
1.941 + // process a given FDE, the parser reiterates the appropriate CIE's
1.942 + // contents at the beginning of the FDE's rules.
1.943 + virtual bool Entry(size_t offset, uint64 address, uint64 length,
1.944 + uint8 version, const std::string &augmentation,
1.945 + unsigned return_address) = 0;
1.946 +
1.947 + // When the Entry function returns true, the parser calls these
1.948 + // handler functions repeatedly to describe the rules for recovering
1.949 + // registers at each instruction in the given range of machine code.
1.950 + // Immediately after a call to Entry, the handler should assume that
1.951 + // the rule for each callee-saves register is "unchanged" --- that
1.952 + // is, that the register still has the value it had in the caller.
1.953 + //
1.954 + // If a *Rule function returns true, we continue processing this entry's
1.955 + // instructions. If a *Rule function returns false, we stop evaluating
1.956 + // instructions, and skip to the next entry. Either way, we call End
1.957 + // before going on to the next entry.
1.958 + //
1.959 + // In all of these functions, if the REG parameter is kCFARegister, then
1.960 + // the rule describes how to find the canonical frame address.
1.961 + // kCFARegister may be passed as a BASE_REGISTER argument, meaning that
1.962 + // the canonical frame address should be used as the base address for the
1.963 + // computation. All other REG values will be positive.
1.964 +
1.965 + // At ADDRESS, register REG's value is not recoverable.
1.966 + virtual bool UndefinedRule(uint64 address, int reg) = 0;
1.967 +
1.968 + // At ADDRESS, register REG's value is the same as that it had in
1.969 + // the caller.
1.970 + virtual bool SameValueRule(uint64 address, int reg) = 0;
1.971 +
1.972 + // At ADDRESS, register REG has been saved at offset OFFSET from
1.973 + // BASE_REGISTER.
1.974 + virtual bool OffsetRule(uint64 address, int reg,
1.975 + int base_register, long offset) = 0;
1.976 +
1.977 + // At ADDRESS, the caller's value of register REG is the current
1.978 + // value of BASE_REGISTER plus OFFSET. (This rule doesn't provide an
1.979 + // address at which the register's value is saved.)
1.980 + virtual bool ValOffsetRule(uint64 address, int reg,
1.981 + int base_register, long offset) = 0;
1.982 +
1.983 + // At ADDRESS, register REG has been saved in BASE_REGISTER. This differs
1.984 + // from ValOffsetRule(ADDRESS, REG, BASE_REGISTER, 0), in that
1.985 + // BASE_REGISTER is the "home" for REG's saved value: if you want to
1.986 + // assign to a variable whose home is REG in the calling frame, you
1.987 + // should put the value in BASE_REGISTER.
1.988 + virtual bool RegisterRule(uint64 address, int reg, int base_register) = 0;
1.989 +
1.990 + // At ADDRESS, the DWARF expression EXPRESSION yields the address at
1.991 + // which REG was saved.
1.992 + virtual bool ExpressionRule(uint64 address, int reg,
1.993 + const std::string &expression) = 0;
1.994 +
1.995 + // At ADDRESS, the DWARF expression EXPRESSION yields the caller's
1.996 + // value for REG. (This rule doesn't provide an address at which the
1.997 + // register's value is saved.)
1.998 + virtual bool ValExpressionRule(uint64 address, int reg,
1.999 + const std::string &expression) = 0;
1.1000 +
1.1001 + // Indicate that the rules for the address range reported by the
1.1002 + // last call to Entry are complete. End should return true if
1.1003 + // everything is okay, or false if an error has occurred and parsing
1.1004 + // should stop.
1.1005 + virtual bool End() = 0;
1.1006 +
1.1007 + // Handler functions for Linux C++ exception handling data. These are
1.1008 + // only called if the data includes 'z' augmentation strings.
1.1009 +
1.1010 + // The Linux C++ ABI uses an extension of the DWARF CFI format to
1.1011 + // walk the stack to propagate exceptions from the throw to the
1.1012 + // appropriate catch, and do the appropriate cleanups along the way.
1.1013 + // CFI entries used for exception handling have two additional data
1.1014 + // associated with them:
1.1015 + //
1.1016 + // - The "language-specific data area" describes which exception
1.1017 + // types the function has 'catch' clauses for, and indicates how
1.1018 + // to go about re-entering the function at the appropriate catch
1.1019 + // clause. If the exception is not caught, it describes the
1.1020 + // destructors that must run before the frame is popped.
1.1021 + //
1.1022 + // - The "personality routine" is responsible for interpreting the
1.1023 + // language-specific data area's contents, and deciding whether
1.1024 + // the exception should continue to propagate down the stack,
1.1025 + // perhaps after doing some cleanup for this frame, or whether the
1.1026 + // exception will be caught here.
1.1027 + //
1.1028 + // In principle, the language-specific data area is opaque to
1.1029 + // everybody but the personality routine. In practice, these values
1.1030 + // may be useful or interesting to readers with extra context, and
1.1031 + // we have to at least skip them anyway, so we might as well report
1.1032 + // them to the handler.
1.1033 +
1.1034 + // This entry's exception handling personality routine's address is
1.1035 + // ADDRESS. If INDIRECT is true, then ADDRESS is the address at
1.1036 + // which the routine's address is stored. The default definition for
1.1037 + // this handler function simply returns true, allowing parsing of
1.1038 + // the entry to continue.
1.1039 + virtual bool PersonalityRoutine(uint64 address, bool indirect) {
1.1040 + return true;
1.1041 + }
1.1042 +
1.1043 + // This entry's language-specific data area (LSDA) is located at
1.1044 + // ADDRESS. If INDIRECT is true, then ADDRESS is the address at
1.1045 + // which the area's address is stored. The default definition for
1.1046 + // this handler function simply returns true, allowing parsing of
1.1047 + // the entry to continue.
1.1048 + virtual bool LanguageSpecificDataArea(uint64 address, bool indirect) {
1.1049 + return true;
1.1050 + }
1.1051 +
1.1052 + // This entry describes a signal trampoline --- this frame is the
1.1053 + // caller of a signal handler. The default definition for this
1.1054 + // handler function simply returns true, allowing parsing of the
1.1055 + // entry to continue.
1.1056 + //
1.1057 + // The best description of the rationale for and meaning of signal
1.1058 + // trampoline CFI entries seems to be in the GCC bug database:
1.1059 + // http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26208
1.1060 + virtual bool SignalHandler() { return true; }
1.1061 +};
1.1062 +
1.1063 +
1.1064 +// The CallFrameInfo class makes calls on an instance of this class to
1.1065 +// report errors or warn about problems in the data it is parsing.
1.1066 +// These messages are sent to the message sink |aLog| provided to the
1.1067 +// constructor.
1.1068 +class CallFrameInfo::Reporter {
1.1069 + public:
1.1070 + // Create an error reporter which attributes troubles to the section
1.1071 + // named SECTION in FILENAME.
1.1072 + //
1.1073 + // Normally SECTION would be .debug_frame, but the Mac puts CFI data
1.1074 + // in a Mach-O section named __debug_frame. If we support
1.1075 + // Linux-style exception handling data, we could be reading an
1.1076 + // .eh_frame section.
1.1077 + Reporter(void (*aLog)(const char*),
1.1078 + const std::string &filename,
1.1079 + const std::string §ion = ".debug_frame")
1.1080 + : log_(aLog), filename_(filename), section_(section) { }
1.1081 + virtual ~Reporter() { }
1.1082 +
1.1083 + // The CFI entry at OFFSET ends too early to be well-formed. KIND
1.1084 + // indicates what kind of entry it is; KIND can be kUnknown if we
1.1085 + // haven't parsed enough of the entry to tell yet.
1.1086 + virtual void Incomplete(uint64 offset, CallFrameInfo::EntryKind kind);
1.1087 +
1.1088 + // The .eh_frame data has a four-byte zero at OFFSET where the next
1.1089 + // entry's length would be; this is a terminator. However, the buffer
1.1090 + // length as given to the CallFrameInfo constructor says there should be
1.1091 + // more data.
1.1092 + virtual void EarlyEHTerminator(uint64 offset);
1.1093 +
1.1094 + // The FDE at OFFSET refers to the CIE at CIE_OFFSET, but the
1.1095 + // section is not that large.
1.1096 + virtual void CIEPointerOutOfRange(uint64 offset, uint64 cie_offset);
1.1097 +
1.1098 + // The FDE at OFFSET refers to the CIE at CIE_OFFSET, but the entry
1.1099 + // there is not a CIE.
1.1100 + virtual void BadCIEId(uint64 offset, uint64 cie_offset);
1.1101 +
1.1102 + // The FDE at OFFSET refers to a CIE with version number VERSION,
1.1103 + // which we don't recognize. We cannot parse DWARF CFI if it uses
1.1104 + // a version number we don't recognize.
1.1105 + virtual void UnrecognizedVersion(uint64 offset, int version);
1.1106 +
1.1107 + // The FDE at OFFSET refers to a CIE with augmentation AUGMENTATION,
1.1108 + // which we don't recognize. We cannot parse DWARF CFI if it uses
1.1109 + // augmentations we don't recognize.
1.1110 + virtual void UnrecognizedAugmentation(uint64 offset,
1.1111 + const std::string &augmentation);
1.1112 +
1.1113 + // The pointer encoding ENCODING, specified by the CIE at OFFSET, is not
1.1114 + // a valid encoding.
1.1115 + virtual void InvalidPointerEncoding(uint64 offset, uint8 encoding);
1.1116 +
1.1117 + // The pointer encoding ENCODING, specified by the CIE at OFFSET, depends
1.1118 + // on a base address which has not been supplied.
1.1119 + virtual void UnusablePointerEncoding(uint64 offset, uint8 encoding);
1.1120 +
1.1121 + // The CIE at OFFSET contains a DW_CFA_restore instruction at
1.1122 + // INSN_OFFSET, which may not appear in a CIE.
1.1123 + virtual void RestoreInCIE(uint64 offset, uint64 insn_offset);
1.1124 +
1.1125 + // The entry at OFFSET, of kind KIND, has an unrecognized
1.1126 + // instruction at INSN_OFFSET.
1.1127 + virtual void BadInstruction(uint64 offset, CallFrameInfo::EntryKind kind,
1.1128 + uint64 insn_offset);
1.1129 +
1.1130 + // The instruction at INSN_OFFSET in the entry at OFFSET, of kind
1.1131 + // KIND, establishes a rule that cites the CFA, but we have not
1.1132 + // established a CFA rule yet.
1.1133 + virtual void NoCFARule(uint64 offset, CallFrameInfo::EntryKind kind,
1.1134 + uint64 insn_offset);
1.1135 +
1.1136 + // The instruction at INSN_OFFSET in the entry at OFFSET, of kind
1.1137 + // KIND, is a DW_CFA_restore_state instruction, but the stack of
1.1138 + // saved states is empty.
1.1139 + virtual void EmptyStateStack(uint64 offset, CallFrameInfo::EntryKind kind,
1.1140 + uint64 insn_offset);
1.1141 +
1.1142 + // The DW_CFA_remember_state instruction at INSN_OFFSET in the entry
1.1143 + // at OFFSET, of kind KIND, would restore a state that has no CFA
1.1144 + // rule, whereas the current state does have a CFA rule. This is
1.1145 + // bogus input, which the CallFrameInfo::Handler interface doesn't
1.1146 + // (and shouldn't) have any way to report.
1.1147 + virtual void ClearingCFARule(uint64 offset, CallFrameInfo::EntryKind kind,
1.1148 + uint64 insn_offset);
1.1149 +
1.1150 + private:
1.1151 + // A logging sink function, as supplied by LUL's user.
1.1152 + void (*log_)(const char*);
1.1153 +
1.1154 + protected:
1.1155 + // The name of the file whose CFI we're reading.
1.1156 + std::string filename_;
1.1157 +
1.1158 + // The name of the CFI section in that file.
1.1159 + std::string section_;
1.1160 +};
1.1161 +
1.1162 +
1.1163 +using lul::CallFrameInfo;
1.1164 +using lul::Summariser;
1.1165 +
1.1166 +// A class that accepts parsed call frame information from the DWARF
1.1167 +// CFI parser and populates a google_breakpad::Module object with the
1.1168 +// contents.
1.1169 +class DwarfCFIToModule: public CallFrameInfo::Handler {
1.1170 + public:
1.1171 +
1.1172 + // DwarfCFIToModule uses an instance of this class to report errors
1.1173 + // detected while converting DWARF CFI to Breakpad STACK CFI records.
1.1174 + class Reporter {
1.1175 + public:
1.1176 + // Create a reporter that writes messages to the message sink
1.1177 + // |aLog|. FILE is the name of the file we're processing, and
1.1178 + // SECTION is the name of the section within that file that we're
1.1179 + // looking at (.debug_frame, .eh_frame, etc.).
1.1180 + Reporter(void (*aLog)(const char*),
1.1181 + const std::string &file, const std::string §ion)
1.1182 + : log_(aLog), file_(file), section_(section) { }
1.1183 + virtual ~Reporter() { }
1.1184 +
1.1185 + // The DWARF CFI entry at OFFSET says that REG is undefined, but the
1.1186 + // Breakpad symbol file format cannot express this.
1.1187 + virtual void UndefinedNotSupported(size_t offset,
1.1188 + const UniqueString* reg);
1.1189 +
1.1190 + // The DWARF CFI entry at OFFSET says that REG uses a DWARF
1.1191 + // expression to find its value, but DwarfCFIToModule is not
1.1192 + // capable of translating DWARF expressions to Breakpad postfix
1.1193 + // expressions.
1.1194 + virtual void ExpressionsNotSupported(size_t offset,
1.1195 + const UniqueString* reg);
1.1196 +
1.1197 + private:
1.1198 + // A logging sink function, as supplied by LUL's user.
1.1199 + void (*log_)(const char*);
1.1200 + protected:
1.1201 + std::string file_, section_;
1.1202 + };
1.1203 +
1.1204 + // Register name tables. If TABLE is a vector returned by one of these
1.1205 + // functions, then TABLE[R] is the name of the register numbered R in
1.1206 + // DWARF call frame information.
1.1207 + class RegisterNames {
1.1208 + public:
1.1209 + // Intel's "x86" or IA-32.
1.1210 + static const unsigned int I386();
1.1211 +
1.1212 + // AMD x86_64, AMD64, Intel EM64T, or Intel 64
1.1213 + static const unsigned int X86_64();
1.1214 +
1.1215 + // ARM.
1.1216 + static const unsigned int ARM();
1.1217 + };
1.1218 +
1.1219 + // Create a handler for the dwarf2reader::CallFrameInfo parser that
1.1220 + // records the stack unwinding information it receives in SUMM.
1.1221 + //
1.1222 + // Use REGISTER_NAMES[I] as the name of register number I; *this
1.1223 + // keeps a reference to the vector, so the vector should remain
1.1224 + // alive for as long as the DwarfCFIToModule does.
1.1225 + //
1.1226 + // Use REPORTER for reporting problems encountered in the conversion
1.1227 + // process.
1.1228 + DwarfCFIToModule(const unsigned int num_dw_regs,
1.1229 + Reporter *reporter,
1.1230 + /*OUT*/Summariser* summ)
1.1231 + : summ_(summ), num_dw_regs_(num_dw_regs), reporter_(reporter),
1.1232 + return_address_(-1) {
1.1233 + }
1.1234 + virtual ~DwarfCFIToModule() {}
1.1235 +
1.1236 + virtual bool Entry(size_t offset, uint64 address, uint64 length,
1.1237 + uint8 version, const std::string &augmentation,
1.1238 + unsigned return_address);
1.1239 + virtual bool UndefinedRule(uint64 address, int reg);
1.1240 + virtual bool SameValueRule(uint64 address, int reg);
1.1241 + virtual bool OffsetRule(uint64 address, int reg,
1.1242 + int base_register, long offset);
1.1243 + virtual bool ValOffsetRule(uint64 address, int reg,
1.1244 + int base_register, long offset);
1.1245 + virtual bool RegisterRule(uint64 address, int reg, int base_register);
1.1246 + virtual bool ExpressionRule(uint64 address, int reg,
1.1247 + const std::string &expression);
1.1248 + virtual bool ValExpressionRule(uint64 address, int reg,
1.1249 + const std::string &expression);
1.1250 + virtual bool End();
1.1251 +
1.1252 + private:
1.1253 + // Return the name to use for register REG.
1.1254 + const UniqueString* RegisterName(int i);
1.1255 +
1.1256 + // The Summariser to which we should give entries
1.1257 + Summariser* summ_;
1.1258 +
1.1259 + // The number of Dwarf-defined register names for this architecture.
1.1260 + const unsigned int num_dw_regs_;
1.1261 +
1.1262 + // The reporter to use to report problems.
1.1263 + Reporter *reporter_;
1.1264 +
1.1265 + // The section offset of the current frame description entry, for
1.1266 + // use in error messages.
1.1267 + size_t entry_offset_;
1.1268 +
1.1269 + // The return address column for that entry.
1.1270 + unsigned return_address_;
1.1271 +};
1.1272 +
1.1273 +} // namespace lul
1.1274 +
1.1275 +#endif // LulDwarfExt_h
