// Copyright (c) 2010 Google Inc. All Rights Reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are // met: // // * Redistributions of source code must retain the above copyright // notice, this list of conditions and the following disclaimer. // * Redistributions in binary form must reproduce the above // copyright notice, this list of conditions and the following disclaimer // in the documentation and/or other materials provided with the // distribution. // * Neither the name of Google Inc. nor the names of its // contributors may be used to endorse or promote products derived from // this software without specific prior written permission. // // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT // OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT // LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, // DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY // THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
// CFI reader author: Jim Blandy <jimb@mozilla.com> <jimb@red-bean.com> // Original author: Jim Blandy <jimb@mozilla.com> <jimb@red-bean.com>
// Implementation of dwarf2reader::LineInfo, dwarf2reader::CompilationUnit, // and dwarf2reader::CallFrameInfo. See dwarf2reader.h for details.
// This file is derived from the following files in // toolkit/crashreporter/google-breakpad: // src/common/dwarf/bytereader.cc // src/common/dwarf/dwarf2reader.cc // src/common/dwarf_cfi_to_module.cc
// In DWARF2/3, if the initial length is all 1 bits, then the offset // size is 8 and we need to read the next 8 bytes for the real length. if (initial_length == 0xffffffff) {
SetOffsetSize(8);
*len = 12; return ReadOffset(start);
} else {
SetOffsetSize(4);
*len = 4;
} return initial_length;
}
bool ByteReader::ValidEncoding(DwarfPointerEncoding encoding) const { if (encoding == DW_EH_PE_omit) returntrue; if (encoding == DW_EH_PE_aligned) returntrue; if ((encoding & 0x7) > DW_EH_PE_udata8) returnfalse; if ((encoding & 0x70) > DW_EH_PE_funcrel) returnfalse; returntrue;
}
bool ByteReader::UsableEncoding(DwarfPointerEncoding encoding) const { switch (encoding & 0x70) { case DW_EH_PE_absptr: returntrue; case DW_EH_PE_pcrel: return have_section_base_; case DW_EH_PE_textrel: return have_text_base_; case DW_EH_PE_datarel: return have_data_base_; case DW_EH_PE_funcrel: return have_function_base_; default: returnfalse;
}
}
uint64 ByteReader::ReadEncodedPointer(constchar* buffer,
DwarfPointerEncoding encoding,
size_t* len) const { // UsableEncoding doesn't approve of DW_EH_PE_omit, so we shouldn't // see it here.
MOZ_ASSERT(encoding != DW_EH_PE_omit);
// The Linux Standards Base 4.0 does not make this clear, but the // GNU tools (gcc/unwind-pe.h; readelf/dwarf.c; gdb/dwarf2-frame.c) // agree that aligned pointers are always absolute, machine-sized, // machine-signed pointers. if (encoding == DW_EH_PE_aligned) {
MOZ_ASSERT(have_section_base_);
// We don't need to align BUFFER in *our* address space. Rather, we // need to find the next position in our buffer that would be aligned // when the .eh_frame section the buffer contains is loaded into the // program's memory. So align assuming that buffer_base_ gets loaded at // address section_base_, where section_base_ itself may or may not be // aligned.
// First, find the offset to START from the closest prior aligned // address.
uint64 skew = section_base_ & (AddressSize() - 1); // Now find the offset from that aligned address to buffer.
uint64 offset = skew + (buffer - buffer_base_); // Round up to the next boundary.
uint64 aligned = (offset + AddressSize() - 1) & -AddressSize(); // Convert back to a pointer. constchar* aligned_buffer = buffer_base_ + (aligned - skew); // Finally, store the length and actually fetch the pointer.
*len = aligned_buffer - buffer + AddressSize(); return ReadAddress(aligned_buffer);
}
// Extract the value first, ignoring whether it's a pointer or an // offset relative to some base.
uint64 offset; switch (encoding & 0x0f) { case DW_EH_PE_absptr: // DW_EH_PE_absptr is weird, as it is used as a meaningful value for // both the high and low nybble of encoding bytes. When it appears in // the high nybble, it means that the pointer is absolute, not an // offset from some base address. When it appears in the low nybble, // as here, it means that the pointer is stored as a normal // machine-sized and machine-signed address. A low nybble of // DW_EH_PE_absptr does not imply that the pointer is absolute; it is // correct for us to treat the value as an offset from a base address // if the upper nybble is not DW_EH_PE_absptr.
offset = ReadAddress(buffer);
*len = AddressSize(); break;
case DW_EH_PE_uleb128:
offset = ReadUnsignedLEB128(buffer, len); break;
case DW_EH_PE_udata2:
offset = ReadTwoBytes(buffer);
*len = 2; break;
case DW_EH_PE_udata4:
offset = ReadFourBytes(buffer);
*len = 4; break;
case DW_EH_PE_udata8:
offset = ReadEightBytes(buffer);
*len = 8; break;
case DW_EH_PE_sleb128:
offset = ReadSignedLEB128(buffer, len); break;
// A DWARF rule for recovering the address or value of a register, or // computing the canonical frame address. This is an 8-way sum-of-products // type. Excluding the INVALID variant, there is one subclass of this for // each '*Rule' member function in CallFrameInfo::Handler. // // This could logically be nested within State, but then the qualified names // get horrendous.
class CallFrameInfo::Rule final { public: enum Tag {
INVALID,
Undefined,
SameValue,
Offset,
ValOffset, Register,
Expression,
ValExpression
};
private: // tag_ (below) indicates the form of the expression. There are 7 forms // plus INVALID. All non-INVALID expressions denote a machine-word-sized // value at unwind time. The description below assumes the presence of, at // unwind time: // // * a function R, which takes a Dwarf register number and returns its value // in the callee frame (the one we are unwinding out of). // // * a function EvalDwarfExpr, which evaluates a Dwarf expression. // // Register numbers are encoded using the target ABI's Dwarf // register-numbering conventions. Except where otherwise noted, a register // value may also be the special value CallFrameInfo::Handler::kCFARegister // ("the CFA"). // // The expression forms are represented using tag_, word1_ and word2_. The // forms and denoted values are as follows: // // * INVALID: not a valid expression. // valid fields: (none) // denotes: no value // // * Undefined: denotes no value. This is used for a register whose value // cannot be recovered. // valid fields: (none) // denotes: no value // // * SameValue: the register's value is the same as in the callee. // valid fields: (none) // denotes: R(the register that this Rule is associated with, // not stored here) // // * Offset: the register's value is in memory at word2_ bytes away from // Dwarf register number word1_. word2_ is interpreted as a *signed* // offset. // valid fields: word1_=DwarfReg, word2=Offset // denotes: *(R(word1_) + word2_) // // * ValOffset: same as Offset, without the dereference. // valid fields: word1_=DwarfReg, word2=Offset // denotes: R(word1_) + word2_ // // * Register: the register's value is in some other register, // which may not be the CFA. // valid fields: word1_=DwarfReg // denotes: R(word1_) // // * Expression: the register's value is in memory at a location that can be // computed from the Dwarf expression contained in the word2_ bytes // starting at word1_. Note these locations are into the area of the .so // temporarily mmaped info for debuginfo reading and have no validity once // debuginfo reading has finished. // valid fields: ExprStart=word1_, ExprLen=word2_ // denotes: *(EvalDwarfExpr(word1_, word2_)) // // * ValExpression: same as Expression, without the dereference. // valid fields: ExprStart=word1_, ExprLen=word2_ // denotes: EvalDwarfExpr(word1_, word2_) //
// 3 words (or less) for representation. Unused word1_/word2_ fields must // be set to zero.
Tag tag_;
uintptr_t word1_;
uintptr_t word2_;
// To ensure that word1_ can hold a pointer to an expression string.
static_assert(sizeof(constchar*) <= sizeof(word1_)); // To ensure that word2_ can hold any string length or memory offset.
static_assert(sizeof(size_t) <= sizeof(word2_));
// This class denotes an 8-way sum-of-product type, and accessing invalid // fields is meaningless. The accessors and constructors below enforce // that. bool isCanonical() const { switch (tag_) { case Tag::INVALID: case Tag::Undefined: case Tag::SameValue: return word1_ == 0 && word2_ == 0; case Tag::Offset: case Tag::ValOffset: returntrue; case Tag::Register: return word2_ == 0; case Tag::Expression: case Tag::ValExpression: returntrue; default:
MOZ_CRASH();
}
}
public:
Tag tag() const { return tag_; } int dwreg() const { switch (tag_) { case Tag::Offset: case Tag::ValOffset: case Tag::Register: return (int)word1_; default:
MOZ_CRASH();
}
}
intptr_t offset() const { switch (tag_) { case Tag::Offset: case Tag::ValOffset: return (intptr_t)word2_; default:
MOZ_CRASH();
}
}
ImageSlice expr() const { switch (tag_) { case Tag::Expression: case Tag::ValExpression: return ImageSlice((constchar*)word1_, (size_t)word2_); default:
MOZ_CRASH();
}
}
// Tell HANDLER that, at ADDRESS in the program, REG can be // recovered using this rule. If REG is kCFARegister, then this rule // describes how to compute the canonical frame address. Return what the // HANDLER member function returned. bool Handle(Handler* handler, uint64 address, int reg) const {
MOZ_ASSERT(isVALID());
MOZ_ASSERT(isCanonical()); switch (tag_) { case Tag::Undefined: return handler->UndefinedRule(address, reg); case Tag::SameValue: return handler->SameValueRule(address, reg); case Tag::Offset: return handler->OffsetRule(address, reg, word1_, word2_); case Tag::ValOffset: return handler->ValOffsetRule(address, reg, word1_, word2_); case Tag::Register: return handler->RegisterRule(address, reg, word1_); case Tag::Expression: return handler->ExpressionRule(
address, reg, ImageSlice((constchar*)word1_, (size_t)word2_)); case Tag::ValExpression: return handler->ValExpressionRule(
address, reg, ImageSlice((constchar*)word1_, (size_t)word2_)); default:
MOZ_CRASH();
}
}
void SetBaseRegister(unsigned reg) {
MOZ_ASSERT(isVALID());
MOZ_ASSERT(isCanonical()); switch (tag_) { case Tag::ValOffset:
word1_ = reg; break; case Tag::Offset: // We don't actually need SetBaseRegister or SetOffset here, since they // are only ever applied to CFA rules, for DW_CFA_def_cfa_offset, and it // doesn't make sense to use OffsetRule for computing the CFA: it // computes the address at which a register is saved, not a value. // (fallthrough) case Tag::Undefined: case Tag::SameValue: case Tag::Register: case Tag::Expression: case Tag::ValExpression: // Do nothing break; default:
MOZ_CRASH();
}
}
void SetOffset(longlong offset) {
MOZ_ASSERT(isVALID());
MOZ_ASSERT(isCanonical()); switch (tag_) { case Tag::ValOffset:
word2_ = offset; break; case Tag::Offset: // Same comment as in SetBaseRegister applies // (fallthrough) case Tag::Undefined: case Tag::SameValue: case Tag::Register: case Tag::Expression: case Tag::ValExpression: // Do nothing break; default:
MOZ_CRASH();
}
}
// For debugging only
string show() const { char buf[100];
string s = ""; switch (tag_) { case Tag::INVALID:
s = "INVALID"; break; case Tag::Undefined:
s = "Undefined"; break; case Tag::SameValue:
s = "SameValue"; break; case Tag::Offset:
s = "Offset{..}"; break; case Tag::ValOffset:
sprintf(buf, "ValOffset{reg=%d offs=%lld}", (int)word1_,
(longlongint)word2_);
s = string(buf); break; case Tag::Register:
s = "Register{..}"; break; case Tag::Expression:
s = "Expression{..}"; break; case Tag::ValExpression:
s = "ValExpression{..}"; break; default:
MOZ_CRASH();
} return s;
}
};
// `RuleMapLowLevel` is a simple class that maps from `int` (register numbers) // to `Rule`. This is implemented as a vector of `<int, Rule>` pairs, with a // 12-element inline capacity. From a big-O perspective this is obviously a // terrible way to implement an associative map. This workload is however // quite special in that the maximum number of elements is normally 7 (on // x86_64-linux), and so this implementation is much faster than one based on // std::map with its attendant R-B-tree node allocation and balancing // overheads. // // An iterator that enumerates the mapping in increasing order of the `int` // keys is provided. This ordered iteration facility is required by // CallFrameInfo::RuleMap::HandleTransitionTo, which needs to iterate through // two such maps simultaneously and in-order so as to compare them.
// All `Rule`s in the map must satisfy `isVALID()`. That conveniently means // that `Rule::mkINVALID()` can be used to indicate "not found` in `get()`.
class CallFrameInfo::RuleMapLowLevel { using Entry = pair<int, Rule>;
// The inline capacity of 12 is carefully chosen. It would be wise to make // careful measurements of time, instruction count, allocation count and // allocated bytes before changing it. For x86_64-linux, a value of 8 is // marginally better; using 12 increases the total heap bytes allocated by // around 20%. For arm64-linux, a value of 24 is better; using 12 increases // the total blocks allocated by around 20%. But it's a not bad tradeoff // for both targets, and in any case is vastly superior to the previous // scheme of using `std::map`.
mozilla::Vector<Entry, 12> entries_;
public: void clear() { entries_.clear(); }
RuleMapLowLevel() { clear(); }
RuleMapLowLevel& operator=(const RuleMapLowLevel& rhs) {
entries_.clear(); for (size_t i = 0; i < rhs.entries_.length(); i++) { bool ok = entries_.append(rhs.entries_[i]);
MOZ_RELEASE_ASSERT(ok);
} return *this;
}
void set(int reg, Rule rule) {
MOZ_ASSERT(rule.isVALID()); // Find the place where it should go, if any
size_t i = 0;
size_t nEnt = entries_.length(); while (i < nEnt && entries_[i].first < reg) {
i++;
} if (i == nEnt) { // No entry exists, and all the existing ones are for lower register // numbers. So just add it at the end. bool ok = entries_.append(Entry(reg, rule));
MOZ_RELEASE_ASSERT(ok);
} else { // It needs to live at location `i`, and ..
MOZ_ASSERT(i < nEnt); if (entries_[i].first == reg) { // .. there's already an old entry, so just update it.
entries_[i].second = rule;
} else { // .. there's no previous entry, so shift `i` and all those following // it one place to the right, and put the new entry at `i`. Doing it // manually is measurably cheaper than using `Vector::insert`.
MOZ_ASSERT(entries_[i].first > reg); bool ok = entries_.append(Entry(999999, Rule::mkINVALID()));
MOZ_RELEASE_ASSERT(ok); for (size_t j = nEnt; j >= i + 1; j--) {
entries_[j] = entries_[j - 1];
}
entries_[i] = Entry(reg, rule);
}
} // Check in-order-ness and validity. for (size_t i = 0; i < entries_.length(); i++) {
MOZ_ASSERT(entries_[i].second.isVALID());
MOZ_ASSERT_IF(i > 0, entries_[i - 1].first < entries_[i].first);
}
MOZ_ASSERT(get(reg).isVALID());
}
// Find the entry for `reg`, or return `Rule::mkINVALID()` if not found.
Rule get(int reg) const {
size_t nEnt = entries_.length(); // "early exit" in the case where `entries_[i].first > reg` was tested on // x86_64 and found to be slightly slower than just testing all entries, // presumably because the reduced amount of searching was not offset by // the cost of an extra test per iteration. for (size_t i = 0; i < nEnt; i++) { if (entries_[i].first == reg) {
CallFrameInfo::Rule ret = entries_[i].second;
MOZ_ASSERT(ret.isVALID()); return ret;
}
} return CallFrameInfo::Rule::mkINVALID();
}
// A very simple in-order iteration facility. class Iter { const RuleMapLowLevel* rmll_;
size_t nextIx_;
public: explicit Iter(const RuleMapLowLevel* rmll) : rmll_(rmll), nextIx_(0) {} bool avail() const { return nextIx_ < rmll_->entries_.length(); } bool finished() const { return !avail(); } // Move the iterator to the next entry. void step() {
MOZ_RELEASE_ASSERT(nextIx_ < rmll_->entries_.length());
nextIx_++;
} // Get the value at the current iteration point, but don't advance to the // next entry.
pair<int, Rule> peek() {
MOZ_RELEASE_ASSERT(nextIx_ < rmll_->entries_.length()); return rmll_->entries_[nextIx_];
}
};
};
// A map from register numbers to rules. This is a wrapper around // `RuleMapLowLevel`, with added logic for dealing with the "special" CFA // rule, and with `HandleTransitionTo`, which effectively computes the // difference between two `RuleMaps`.
// Set the rule for computing the CFA to RULE. void SetCFARule(Rule rule) { cfa_rule_ = rule; }
// Return the current CFA rule. Be careful not to modify it -- it's returned // by value. If you want to modify the CFA rule, use CFARuleRef() instead. // We use these two for DW_CFA_def_cfa_offset and DW_CFA_def_cfa_register, // and for detecting references to the CFA before a rule for it has been // established.
Rule CFARule() const { return cfa_rule_; }
Rule* CFARuleRef() { return &cfa_rule_; }
// Return the rule for REG, or the INVALID rule if there is none.
Rule RegisterRule(int reg) const;
// Set the rule for computing REG to RULE. void SetRegisterRule(int reg, Rule rule);
// Make all the appropriate calls to HANDLER as if we were changing from // this RuleMap to NEW_RULES at ADDRESS. We use this to implement // DW_CFA_restore_state, where lots of rules can change simultaneously. // Return true if all handlers returned true; otherwise, return false. bool HandleTransitionTo(Handler* handler, uint64 address, const RuleMap& new_rules) const;
private: // Remove all register rules and clear cfa_rule_. void Clear();
// The rule for computing the canonical frame address.
Rule cfa_rule_;
// A map from register numbers to postfix expressions to recover // their values.
RuleMapLowLevel registers_;
};
bool CallFrameInfo::RuleMap::HandleTransitionTo(
Handler* handler, uint64 address, const RuleMap& new_rules) const { // Transition from cfa_rule_ to new_rules.cfa_rule_. if (cfa_rule_.isVALID() && new_rules.cfa_rule_.isVALID()) { if (cfa_rule_ != new_rules.cfa_rule_ &&
!new_rules.cfa_rule_.Handle(handler, address, Handler::kCFARegister)) { returnfalse;
}
} elseif (cfa_rule_.isVALID()) { // this RuleMap has a CFA rule but new_rules doesn't. // CallFrameInfo::Handler has no way to handle this --- and shouldn't; // it's garbage input. The instruction interpreter should have // detected this and warned, so take no action here.
} elseif (new_rules.cfa_rule_.isVALID()) { // This shouldn't be possible: NEW_RULES is some prior state, and // there's no way to remove entries.
MOZ_ASSERT(0);
} else { // Both CFA rules are empty. No action needed.
}
// Traverse the two maps in order by register number, and report // whatever differences we find.
RuleMapLowLevel::Iter old_it(®isters_);
RuleMapLowLevel::Iter new_it(&new_rules.registers_); while (!old_it.finished() && !new_it.finished()) {
pair<int, Rule> old_pair = old_it.peek();
pair<int, Rule> new_pair = new_it.peek(); if (old_pair.first < new_pair.first) { // This RuleMap has an entry for old.first, but NEW_RULES doesn't. // // This isn't really the right thing to do, but since CFI generally // only mentions callee-saves registers, and GCC's convention for // callee-saves registers is that they are unchanged, it's a good // approximation. if (!handler->SameValueRule(address, old_pair.first)) { returnfalse;
}
old_it.step();
} elseif (old_pair.first > new_pair.first) { // NEW_RULES has an entry for new_pair.first, but this RuleMap // doesn't. This shouldn't be possible: NEW_RULES is some prior // state, and there's no way to remove entries.
MOZ_ASSERT(0);
} else { // Both maps have an entry for this register. Report the new // rule if it is different. if (old_pair.second != new_pair.second &&
!new_pair.second.Handle(handler, address, new_pair.first)) { returnfalse;
}
new_it.step();
old_it.step();
}
} // Finish off entries from this RuleMap with no counterparts in new_rules. while (!old_it.finished()) {
pair<int, Rule> old_pair = old_it.peek(); if (!handler->SameValueRule(address, old_pair.first)) returnfalse;
old_it.step();
} // Since we only make transitions from a rule set to some previously // saved rule set, and we can only add rules to the map, NEW_RULES // must have fewer rules than *this.
MOZ_ASSERT(new_it.finished());
returntrue;
}
// Remove all register rules and clear cfa_rule_. void CallFrameInfo::RuleMap::Clear() {
cfa_rule_ = Rule::mkINVALID();
registers_.clear();
}
// The state of the call frame information interpreter as it processes // instructions from a CIE and FDE. class CallFrameInfo::State { public: // Create a call frame information interpreter state with the given // reporter, reader, handler, and initial call frame info address.
State(ByteReader* reader, Handler* handler, Reporter* reporter,
uint64 address)
: reader_(reader),
handler_(handler),
reporter_(reporter),
address_(address),
entry_(NULL),
cursor_(NULL),
saved_rules_(NULL) {}
~State() { if (saved_rules_) delete saved_rules_;
}
// Interpret instructions from CIE, save the resulting rule set for // DW_CFA_restore instructions, and return true. On error, report // the problem to reporter_ and return false. bool InterpretCIE(const CIE& cie);
// Interpret instructions from FDE, and return true. On error, // report the problem to reporter_ and return false. bool InterpretFDE(const FDE& fde);
private: // The operands of a CFI instruction, for ParseOperands. struct Operands { unsigned register_number; // A register number.
uint64 offset; // An offset or address. long signed_offset; // A signed offset.
ImageSlice expression; // A DWARF expression.
};
// Parse CFI instruction operands from STATE's instruction stream as // described by FORMAT. On success, populate OPERANDS with the // results, and return true. On failure, report the problem and // return false. // // Each character of FORMAT should be one of the following: // // 'r' unsigned LEB128 register number (OPERANDS->register_number) // 'o' unsigned LEB128 offset (OPERANDS->offset) // 's' signed LEB128 offset (OPERANDS->signed_offset) // 'a' machine-size address (OPERANDS->offset) // (If the CIE has a 'z' augmentation string, 'a' uses the // encoding specified by the 'R' argument.) // '1' a one-byte offset (OPERANDS->offset) // '2' a two-byte offset (OPERANDS->offset) // '4' a four-byte offset (OPERANDS->offset) // '8' an eight-byte offset (OPERANDS->offset) // 'e' a DW_FORM_block holding a (OPERANDS->expression) // DWARF expression bool ParseOperands(constchar* format, Operands* operands);
// Interpret one CFI instruction from STATE's instruction stream, update // STATE, report any rule changes to handler_, and return true. On // failure, report the problem and return false.
MOZ_ALWAYS_INLINE bool DoInstruction();
// Repeatedly call `DoInstruction`, until either: // * it returns `false`, which indicates some kind of failure, // in which case return `false` from here too, or // * we've run out of instructions (that is, `cursor_ >= entry_->end`), // in which case return `true`. // This is marked as never-inline because it is the only place that // `DoInstruction` is called from, and we want to maximise the chances that // `DoInstruction` is inlined into this routine.
MOZ_NEVER_INLINE bool DoInstructions();
// The following Do* member functions are subroutines of DoInstruction, // factoring out the actual work of operations that have several // different encodings.
// Set the CFA rule to be the value of BASE_REGISTER plus OFFSET, and // return true. On failure, report and return false. (Used for // DW_CFA_def_cfa and DW_CFA_def_cfa_sf.) bool DoDefCFA(unsigned base_register, long offset);
// Change the offset of the CFA rule to OFFSET, and return true. On // failure, report and return false. (Subroutine for // DW_CFA_def_cfa_offset and DW_CFA_def_cfa_offset_sf.) bool DoDefCFAOffset(long offset);
// Specify that REG can be recovered using RULE, and return true. On // failure, report and return false. bool DoRule(unsigned reg, Rule rule);
// Specify that REG can be found at OFFSET from the CFA, and return true. // On failure, report and return false. (Subroutine for DW_CFA_offset, // DW_CFA_offset_extended, and DW_CFA_offset_extended_sf.) bool DoOffset(unsigned reg, long offset);
// Specify that the caller's value for REG is the CFA plus OFFSET, // and return true. On failure, report and return false. (Subroutine // for DW_CFA_val_offset and DW_CFA_val_offset_sf.) bool DoValOffset(unsigned reg, long offset);
// Restore REG to the rule established in the CIE, and return true. On // failure, report and return false. (Subroutine for DW_CFA_restore and // DW_CFA_restore_extended.) bool DoRestore(unsigned reg);
// Return the section offset of the instruction at cursor. For use // in error messages.
uint64 CursorOffset() { return entry_->offset + (cursor_ - entry_->start); }
// Report that entry_ is incomplete, and return false. For brevity. bool ReportIncomplete() {
reporter_->Incomplete(entry_->offset, entry_->kind); returnfalse;
}
// For reading multi-byte values with the appropriate endianness.
ByteReader* reader_;
// The handler to which we should report the data we find.
Handler* handler_;
// For reporting problems in the info we're parsing.
Reporter* reporter_;
// The code address to which the next instruction in the stream applies.
uint64 address_;
// The entry whose instructions we are currently processing. This is // first a CIE, and then an FDE. const Entry* entry_;
// The next instruction to process. constchar* cursor_;
// The current set of rules.
RuleMap rules_;
// The set of rules established by the CIE, used by DW_CFA_restore // and DW_CFA_restore_extended. We set this after interpreting the // CIE's instructions.
RuleMap cie_rules_;
// A stack of saved states, for DW_CFA_remember_state and // DW_CFA_restore_state.
std::stack<RuleMap>* saved_rules_;
};
bool CallFrameInfo::State::InterpretCIE(const CIE& cie) {
entry_ = &cie;
cursor_ = entry_->instructions; if (!DoInstructions()) { returnfalse;
} // Note the rules established by the CIE, for use by DW_CFA_restore // and DW_CFA_restore_extended.
cie_rules_ = rules_; returntrue;
}
// Find a register at an offset from the CFA. case DW_CFA_offset: if (!ParseOperands("o", &ops) ||
!DoOffset(opcode & 0x3f, ops.offset * cie->data_alignment_factor)) returnfalse; break;
// Restore the rule established for a register by the CIE. case DW_CFA_restore: if (!DoRestore(opcode & 0x3f)) returnfalse; break;
// The 'if' above should have excluded this possibility. default:
MOZ_ASSERT(0);
}
// Return here, so the big switch below won't be indented. returntrue;
}
switch (opcode) { // Set the address. case DW_CFA_set_loc: if (!ParseOperands("a", &ops)) returnfalse;
address_ = ops.offset; break;
// Advance the address. case DW_CFA_advance_loc1: if (!ParseOperands("1", &ops)) returnfalse;
address_ += ops.offset * cie->code_alignment_factor; break;
// Advance the address. case DW_CFA_advance_loc2: if (!ParseOperands("2", &ops)) returnfalse;
address_ += ops.offset * cie->code_alignment_factor; break;
// Advance the address. case DW_CFA_advance_loc4: if (!ParseOperands("4", &ops)) returnfalse;
address_ += ops.offset * cie->code_alignment_factor; break;
// Advance the address. case DW_CFA_MIPS_advance_loc8: if (!ParseOperands("8", &ops)) returnfalse;
address_ += ops.offset * cie->code_alignment_factor; break;
// Compute the CFA by adding an offset to a register. case DW_CFA_def_cfa: if (!ParseOperands("ro", &ops) ||
!DoDefCFA(ops.register_number, ops.offset)) returnfalse; break;
// Compute the CFA by adding an offset to a register. case DW_CFA_def_cfa_sf: if (!ParseOperands("rs", &ops) ||
!DoDefCFA(ops.register_number,
ops.signed_offset * cie->data_alignment_factor)) returnfalse; break;
// Change the base register used to compute the CFA. case DW_CFA_def_cfa_register: {
Rule* cfa_rule = rules_.CFARuleRef(); if (!cfa_rule->isVALID()) {
reporter_->NoCFARule(entry_->offset, entry_->kind, CursorOffset()); returnfalse;
} if (!ParseOperands("r", &ops)) returnfalse;
cfa_rule->SetBaseRegister(ops.register_number); if (!cfa_rule->Handle(handler_, address_, Handler::kCFARegister)) returnfalse; break;
}
// Change the offset used to compute the CFA. case DW_CFA_def_cfa_offset: if (!ParseOperands("o", &ops) || !DoDefCFAOffset(ops.offset)) returnfalse; break;
// Change the offset used to compute the CFA. case DW_CFA_def_cfa_offset_sf: if (!ParseOperands("s", &ops) ||
!DoDefCFAOffset(ops.signed_offset * cie->data_alignment_factor)) returnfalse; break;
// Specify an expression whose value is the CFA. case DW_CFA_def_cfa_expression: { if (!ParseOperands("e", &ops)) returnfalse;
Rule rule = Rule::mkValExpressionRule(ops.expression);
rules_.SetCFARule(rule); if (!rule.Handle(handler_, address_, Handler::kCFARegister)) returnfalse; break;
}
// The register's value cannot be recovered. case DW_CFA_undefined: { if (!ParseOperands("r", &ops) ||
!DoRule(ops.register_number, Rule::mkUndefinedRule())) returnfalse; break;
}
// The register's value is unchanged from its value in the caller. case DW_CFA_same_value: { if (!ParseOperands("r", &ops) ||
!DoRule(ops.register_number, Rule::mkSameValueRule())) returnfalse; break;
}
// Find a register at an offset from the CFA. case DW_CFA_offset_extended: if (!ParseOperands("ro", &ops) ||
!DoOffset(ops.register_number,
ops.offset * cie->data_alignment_factor)) returnfalse; break;
// The register is saved at an offset from the CFA. case DW_CFA_offset_extended_sf: if (!ParseOperands("rs", &ops) ||
!DoOffset(ops.register_number,
ops.signed_offset * cie->data_alignment_factor)) returnfalse; break;
// The register is saved at an offset from the CFA. case DW_CFA_GNU_negative_offset_extended: if (!ParseOperands("ro", &ops) ||
!DoOffset(ops.register_number,
-ops.offset * cie->data_alignment_factor)) returnfalse; break;
// The register's value is the sum of the CFA plus an offset. case DW_CFA_val_offset: if (!ParseOperands("ro", &ops) ||
!DoValOffset(ops.register_number,
ops.offset * cie->data_alignment_factor)) returnfalse; break;
// The register's value is the sum of the CFA plus an offset. case DW_CFA_val_offset_sf: if (!ParseOperands("rs", &ops) ||
!DoValOffset(ops.register_number,
ops.signed_offset * cie->data_alignment_factor)) returnfalse; break;
// The register has been saved in another register. case DW_CFA_register: { if (!ParseOperands("ro", &ops) ||
!DoRule(ops.register_number, Rule::mkRegisterRule(ops.offset))) returnfalse; break;
}
// An expression yields the address at which the register is saved. case DW_CFA_expression: { if (!ParseOperands("re", &ops) ||
!DoRule(ops.register_number, Rule::mkExpressionRule(ops.expression))) returnfalse; break;
}
// An expression yields the caller's value for the register. case DW_CFA_val_expression: { if (!ParseOperands("re", &ops) ||
!DoRule(ops.register_number,
Rule::mkValExpressionRule(ops.expression))) returnfalse; break;
}
// Restore the rule established for a register by the CIE. case DW_CFA_restore_extended: if (!ParseOperands("r", &ops) || !DoRestore(ops.register_number)) returnfalse; break;
// Save the current set of rules on a stack. case DW_CFA_remember_state: if (!saved_rules_) {
saved_rules_ = new std::stack<RuleMap>();
}
saved_rules_->push(rules_); break;
// Pop the current set of rules off the stack. case DW_CFA_restore_state: { if (!saved_rules_ || saved_rules_->empty()) {
reporter_->EmptyStateStack(entry_->offset, entry_->kind,
CursorOffset()); returnfalse;
} const RuleMap& new_rules = saved_rules_->top(); if (rules_.CFARule().isVALID() && !new_rules.CFARule().isVALID()) {
reporter_->ClearingCFARule(entry_->offset, entry_->kind,
CursorOffset()); returnfalse;
}
rules_.HandleTransitionTo(handler_, address_, new_rules);
rules_ = new_rules;
saved_rules_->pop(); break;
}
// No operation. (Padding instruction.) case DW_CFA_nop: break;
// A SPARC register window save: Registers 8 through 15 (%o0-%o7) // are saved in registers 24 through 31 (%i0-%i7), and registers // 16 through 31 (%l0-%l7 and %i0-%i7) are saved at CFA offsets // (0-15 * the register size). The register numbers must be // hard-coded. A GNU extension, and not a pretty one. case DW_CFA_GNU_window_save: { // Save %o0-%o7 in %i0-%i7. for (int i = 8; i < 16; i++) if (!DoRule(i, Rule::mkRegisterRule(i + 16))) returnfalse; // Save %l0-%l7 and %i0-%i7 at the CFA. for (int i = 16; i < 32; i++) // Assume that the byte reader's address size is the same as // the architecture's register size. !@#%*^ hilarious. if (!DoRule(i, Rule::mkOffsetRule(Handler::kCFARegister,
(i - 16) * reader_->AddressSize()))) returnfalse; break;
}
// I'm not sure what this is. GDB doesn't use it for unwinding. case DW_CFA_GNU_args_size: if (!ParseOperands("o", &ops)) returnfalse; break;
// An opcode we don't recognize. default: {
reporter_->BadInstruction(entry_->offset, entry_->kind, CursorOffset()); returnfalse;
}
}
returntrue;
}
// See declaration above for rationale re the no-inline directive.
MOZ_NEVER_INLINE bool CallFrameInfo::State::DoInstructions() { while (cursor_ < entry_->end) { if (!DoInstruction()) { returnfalse;
}
} returntrue;
}
bool CallFrameInfo::State::DoValOffset(unsigned reg, long offset) { if (!rules_.CFARule().isVALID()) {
reporter_->NoCFARule(entry_->offset, entry_->kind, CursorOffset()); returnfalse;
} return DoRule(reg, Rule::mkValOffsetRule(Handler::kCFARegister, offset));
}
bool CallFrameInfo::State::DoRestore(unsigned reg) { // DW_CFA_restore and DW_CFA_restore_extended don't make sense in a CIE. if (entry_->kind == kCIE) {
reporter_->RestoreInCIE(entry_->offset, CursorOffset()); returnfalse;
}
Rule rule = cie_rules_.RegisterRule(reg); if (!rule.isVALID()) { // This isn't really the right thing to do, but since CFI generally // only mentions callee-saves registers, and GCC's convention for // callee-saves registers is that they are unchanged, it's a good // approximation.
rule = Rule::mkSameValueRule();
} return DoRule(reg, rule);
}
// Initialize enough of ENTRY for use in error reporting.
entry->offset = cursor - buffer_;
entry->start = cursor;
entry->kind = kUnknown;
entry->end = NULL;
// Read the initial length. This sets reader_'s offset size.
size_t length_size;
uint64 length = reader_->ReadInitialLength(cursor, &length_size); if (length_size > size_t(buffer_end - cursor)) return ReportIncomplete(entry);
cursor += length_size;
// In a .eh_frame section, a length of zero marks the end of the series // of entries. if (length == 0 && eh_frame_) {
entry->kind = kTerminator;
entry->end = cursor; returntrue;
}
// Validate the length. if (length > size_t(buffer_end - cursor)) return ReportIncomplete(entry);
// The length is the number of bytes after the initial length field; // we have that position handy at this point, so compute the end // now. (If we're parsing 64-bit-offset DWARF on a 32-bit machine, // and the length didn't fit in a size_t, we would have rejected it // above.)
entry->end = cursor + length;
// Parse the next field: either the offset of a CIE or a CIE id.
size_t offset_size = reader_->OffsetSize(); if (offset_size > size_t(entry->end - cursor)) return ReportIncomplete(entry);
entry->id = reader_->ReadOffset(cursor);
// Don't advance cursor past id field yet; in .eh_frame data we need // the id's position to compute the section offset of an FDE's CIE.
// Now we can decide what kind of entry this is. if (eh_frame_) { // In .eh_frame data, an ID of zero marks the entry as a CIE, and // anything else is an offset from the id field of the FDE to the start // of the CIE. if (entry->id == 0) {
entry->kind = kCIE;
} else {
entry->kind = kFDE; // Turn the offset from the id into an offset from the buffer's start.
entry->id = (cursor - buffer_) - entry->id;
}
} else { // In DWARF CFI data, an ID of ~0 (of the appropriate width, given the // offset size for the entry) marks the entry as a CIE, and anything // else is the offset of the CIE from the beginning of the section. if (offset_size == 4)
entry->kind = (entry->id == 0xffffffff) ? kCIE : kFDE; else {
MOZ_ASSERT(offset_size == 8);
entry->kind = (entry->id == 0xffffffffffffffffULL) ? kCIE : kFDE;
}
}
// Now advance cursor past the id.
cursor += offset_size;
// The fields specific to this kind of entry start here.
entry->fields = cursor;
// Parse the version number. if (cie->end - cursor < 1) return ReportIncomplete(cie);
cie->version = reader_->ReadOneByte(cursor);
cursor++;
// If we don't recognize the version, we can't parse any more fields of the // CIE. For DWARF CFI, we handle versions 1 through 4 (there was never a // version 2 of CFI data). For .eh_frame, we handle versions 1 and 4 as well; // the difference between those versions seems to be the same as for // .debug_frame. if (cie->version < 1 || cie->version > 4) {
reporter_->UnrecognizedVersion(cie->offset, cie->version); returnfalse;
}
// Is this CFI augmented? if (!cie->augmentation.empty()) { // Is it an augmentation we recognize? if (cie->augmentation[0] == DW_Z_augmentation_start) { // Linux C++ ABI 'z' augmentation, used for exception handling data.
cie->has_z_augmentation = true;
} else { // Not an augmentation we recognize. Augmentations can have arbitrary // effects on the form of rest of the content, so we have to give up.
reporter_->UnrecognizedAugmentation(cie->offset, cie->augmentation); returnfalse;
}
}
if (cie->version >= 4) { // Check that the address_size and segment_size fields are plausible. if (cie->end - cursor < 2) { return ReportIncomplete(cie);
}
uint8_t address_size = reader_->ReadOneByte(cursor);
cursor++; if (address_size != sizeof(void*)) { // This is not per-se invalid CFI. But we can reasonably expect to // be running on a target of the same word size as the CFI is for, // so we reject this case.
reporter_->InvalidDwarf4Artefact(cie->offset, "Invalid address_size"); returnfalse;
}
uint8_t segment_size = reader_->ReadOneByte(cursor);
cursor++; if (segment_size != 0) { // This is also not per-se invalid CFI, but we don't currently handle // the case of non-zero |segment_size|.
reporter_->InvalidDwarf4Artefact(cie->offset, "Invalid segment_size"); returnfalse;
} // We only continue parsing if |segment_size| is zero. If this routine // is ever changed to allow non-zero |segment_size|, then // ReadFDEFields() below will have to be changed to match, per comments // there.
}
// Parse the data alignment factor.
cie->data_alignment_factor = reader_->ReadSignedLEB128(cursor, &len); if (size_t(cie->end - cursor) < len) return ReportIncomplete(cie);
cursor += len;
// Parse the return address register. This is a ubyte in version 1, and // a ULEB128 in version 3. if (cie->version == 1) { if (cursor >= cie->end) return ReportIncomplete(cie);
cie->return_address_register = uint8(*cursor++);
} else {
cie->return_address_register = reader_->ReadUnsignedLEB128(cursor, &len); if (size_t(cie->end - cursor) < len) return ReportIncomplete(cie);
cursor += len;
}
// If we have a 'z' augmentation string, find the augmentation data and // use the augmentation string to parse it. if (cie->has_z_augmentation) {
uint64_t data_size = reader_->ReadUnsignedLEB128(cursor, &len); if (size_t(cie->end - cursor) < len + data_size) return ReportIncomplete(cie);
cursor += len; constchar* data = cursor;
cursor += data_size; constchar* data_end = cursor;
// Walk the augmentation string, and extract values from the // augmentation data as the string directs. for (size_t i = 1; i < cie->augmentation.size(); i++) { switch (cie->augmentation[i]) { case DW_Z_has_LSDA: // The CIE's augmentation data holds the language-specific data // area pointer's encoding, and the FDE's augmentation data holds // the pointer itself.
cie->has_z_lsda = true; // Fetch the LSDA encoding from the augmentation data. if (data >= data_end) return ReportIncomplete(cie);
cie->lsda_encoding = DwarfPointerEncoding(*data++); if (!reader_->ValidEncoding(cie->lsda_encoding)) {
reporter_->InvalidPointerEncoding(cie->offset, cie->lsda_encoding); returnfalse;
} // Don't check if the encoding is usable here --- we haven't // read the FDE's fields yet, so we're not prepared for // DW_EH_PE_funcrel, although that's a fine encoding for the // LSDA to use, since it appears in the FDE. break;
case DW_Z_has_personality_routine: // The CIE's augmentation data holds the personality routine // pointer's encoding, followed by the pointer itself.
cie->has_z_personality = true; // Fetch the personality routine pointer's encoding from the // augmentation data. if (data >= data_end) return ReportIncomplete(cie);
cie->personality_encoding = DwarfPointerEncoding(*data++); if (!reader_->ValidEncoding(cie->personality_encoding)) {
reporter_->InvalidPointerEncoding(cie->offset,
cie->personality_encoding); returnfalse;
} if (!reader_->UsableEncoding(cie->personality_encoding)) {
reporter_->UnusablePointerEncoding(cie->offset,
cie->personality_encoding); returnfalse;
} // Fetch the personality routine's pointer itself from the data.
cie->personality_address = reader_->ReadEncodedPointer(
data, cie->personality_encoding, &len); if (len > size_t(data_end - data)) return ReportIncomplete(cie);
data += len; break;
case DW_Z_has_FDE_address_encoding: // The CIE's augmentation data holds the pointer encoding to use // for addresses in the FDE. if (data >= data_end) return ReportIncomplete(cie);
cie->pointer_encoding = DwarfPointerEncoding(*data++); if (!reader_->ValidEncoding(cie->pointer_encoding)) {
reporter_->InvalidPointerEncoding(cie->offset,
cie->pointer_encoding); returnfalse;
} if (!reader_->UsableEncoding(cie->pointer_encoding)) {
reporter_->UnusablePointerEncoding(cie->offset,
cie->pointer_encoding); returnfalse;
} break;
case DW_Z_is_signal_trampoline: // Frames using this CIE are signal delivery frames.
cie->has_z_signal_frame = true; break;
default: // An augmentation we don't recognize.
reporter_->UnrecognizedAugmentation(cie->offset, cie->augmentation); returnfalse;
}
}
}
// The CIE's instructions start here.
cie->instructions = cursor;
// At this point, for Dwarf 4 and above, we are assuming that the // associated CIE has its |segment_size| field equal to zero. This is // checked for in ReadCIEFields() above. If ReadCIEFields() is ever // changed to allow non-zero |segment_size| CIEs then we will have to read // the segment_selector value at this point.
// For the length, we strip off the upper nybble of the encoding used for // the starting address.
DwarfPointerEncoding length_encoding =
DwarfPointerEncoding(fde->cie->pointer_encoding & 0x0f);
fde->size = reader_->ReadEncodedPointer(cursor, length_encoding, &size); if (size > size_t(fde->end - cursor)) return ReportIncomplete(fde);
cursor += size;
// If the CIE has a 'z' augmentation string, then augmentation data // appears here. if (fde->cie->has_z_augmentation) {
uint64_t data_size = reader_->ReadUnsignedLEB128(cursor, &size); if (size_t(fde->end - cursor) < size + data_size) return ReportIncomplete(fde);
cursor += size;
// In the abstract, we should walk the augmentation string, and extract // items from the FDE's augmentation data as we encounter augmentation // string characters that specify their presence: the ordering of items // in the augmentation string determines the arrangement of values in // the augmentation data. // // In practice, there's only ever one value in FDE augmentation data // that we support --- the LSDA pointer --- and we have to bail if we // see any unrecognized augmentation string characters. So if there is // anything here at all, we know what it is, and where it starts. if (fde->cie->has_z_lsda) { // Check whether the LSDA's pointer encoding is usable now: only once // we've parsed the FDE's starting address do we call reader_-> // SetFunctionBase, so that the DW_EH_PE_funcrel encoding becomes // usable. if (!reader_->UsableEncoding(fde->cie->lsda_encoding)) {
reporter_->UnusablePointerEncoding(fde->cie->offset,
fde->cie->lsda_encoding); returnfalse;
}
fde->lsda_address =
reader_->ReadEncodedPointer(cursor, fde->cie->lsda_encoding, &size); if (size > data_size) return ReportIncomplete(fde); // Ideally, we would also complain here if there were unconsumed // augmentation data.
}
cursor += data_size;
}
// The FDE's instructions start after those.
fde->instructions = cursor;
// Traverse all the entries in buffer_, skipping CIEs and offering // FDEs to the handler. for (cursor = buffer_; cursor < buffer_end;
cursor = entry_end, all_ok = all_ok && ok) {
FDE fde;
// Make it easy to skip this entry with 'continue': assume that // things are not okay until we've checked all the data, and // prepare the address of the next entry.
ok = false;
// Read the entry's prologue. if (!ReadEntryPrologue(cursor, &fde)) { if (!fde.end) { // If we couldn't even figure out this entry's extent, then we // must stop processing entries altogether.
all_ok = false; break;
}
entry_end = fde.end; continue;
}
// The next iteration picks up after this entry.
entry_end = fde.end;
// Did we see an .eh_frame terminating mark? if (fde.kind == kTerminator) { // If there appears to be more data left in the section after the // terminating mark, warn the user. But this is just a warning; // we leave all_ok true. if (fde.end < buffer_end) reporter_->EarlyEHTerminator(fde.offset); break;
}
// In this loop, we skip CIEs. We only parse them fully when we // parse an FDE that refers to them. This limits our memory // consumption (beyond the buffer itself) to that needed to // process the largest single entry. if (fde.kind != kFDE) {
ok = true; continue;
}
// Validate the CIE pointer. if (fde.id > buffer_length_) {
reporter_->CIEPointerOutOfRange(fde.offset, fde.id); continue;
}
CIE cie;
// Parse this FDE's CIE header. if (!ReadEntryPrologue(buffer_ + fde.id, &cie)) continue; // This had better be an actual CIE. if (cie.kind != kCIE) {
reporter_->BadCIEId(fde.offset, fde.id); continue;
} if (!ReadCIEFields(&cie)) continue;
// We now have the values that govern both the CIE and the FDE.
cie.cie = &cie;
fde.cie = &cie;
// Parse the FDE's header. if (!ReadFDEFields(&fde)) continue;
// Call Entry to ask the consumer if they're interested. if (!handler_->Entry(fde.offset, fde.address, fde.size, cie.version,
cie.augmentation, cie.return_address_register)) { // The handler isn't interested in this entry. That's not an error.
ok = true; continue;
}
if (cie.has_z_augmentation) { // Report the personality routine address, if we have one. if (cie.has_z_personality) { if (!handler_->PersonalityRoutine(
cie.personality_address,
IsIndirectEncoding(cie.personality_encoding))) continue;
}
// Report the language-specific data area address, if we have one. if (cie.has_z_lsda) { if (!handler_->LanguageSpecificDataArea(
fde.lsda_address, IsIndirectEncoding(cie.lsda_encoding)))
--> --------------------
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.