diff options
Diffstat (limited to 'src/common/dwarf/dwarf2reader.h')
-rw-r--r-- | src/common/dwarf/dwarf2reader.h | 468 |
1 files changed, 468 insertions, 0 deletions
diff --git a/src/common/dwarf/dwarf2reader.h b/src/common/dwarf/dwarf2reader.h index d99de334..136b8932 100644 --- a/src/common/dwarf/dwarf2reader.h +++ b/src/common/dwarf/dwarf2reader.h @@ -1,3 +1,5 @@ +// -*- mode: C++ -*- + // Copyright (c) 2010 Google Inc. All Rights Reserved. // // Redistribution and use in source and binary forms, with or without @@ -26,6 +28,8 @@ // (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> + // This file contains definitions related to the DWARF2/3 reader and // it's handler interfaces. // The DWARF2/3 specification can be found at @@ -398,6 +402,470 @@ class Dwarf2Handler { }; +// This class is a reader for DWARF's Call Frame Information. CFI +// describes how to unwind stack frames --- even for functions that do +// not follow fixed conventions for saving registers, whose frame size +// varies as they execute, etc. +// +// CFI describes, at each machine instruction, how to compute the +// stack frame's base address, how to find the return address, and +// where to find the saved values of the caller's registers (if the +// callee has stashed them somewhere to free up the registers for its +// own use). +// +// For example, suppose we have a function whose machine code looks +// like this (imagine an assembly language that looks like C, for a +// machine with 32-bit registers, and a stack that grows towards lower +// addresses): +// +// func: ; entry point; return address at sp +// func+0: sp = sp - 16 ; allocate space for stack frame +// func+1: sp[12] = r0 ; save r0 at sp+12 +// ... ; other code, not frame-related +// func+10: sp -= 4; *sp = x ; push some x on the stack +// ... ; other code, not frame-related +// func+20: r0 = sp[16] ; restore saved r0 +// func+21: sp += 20 ; pop whole stack frame +// func+22: pc = *sp; sp += 4 ; pop return address and jump to it +// +// DWARF CFI is (a very compressed representation of) a table with a +// row for each machine instruction address and a column for each +// register showing how to restore it, if possible. +// +// A special column named "CFA", for "Canonical Frame Address", tells how +// to compute the base address of the frame; registers' entries may +// refer to the CFA in describing where the registers are saved. +// +// Another special column, named "RA", represents the return address. +// +// For example, here is a complete (uncompressed) table describing the +// function above: +// +// insn cfa r0 r1 ... ra +// ======================================= +// func+0: sp cfa[0] +// func+1: sp+16 cfa[0] +// func+2: sp+16 cfa[-4] cfa[0] +// func+11: sp+20 cfa[-4] cfa[0] +// func+21: sp+20 cfa[0] +// func+22: sp cfa[0] +// +// Some things to note here: +// +// - Each row describes the state of affairs *before* executing the +// instruction at the given address. Thus, the row for func+0 +// describes the state before we allocate the stack frame. In the +// next row, the formula for computing the CFA has changed, +// reflecting that allocation. +// +// - The other entries are written in terms of the CFA; this allows +// them to remain unchanged as the stack pointer gets bumped around. +// For example, the rule for recovering the return address (the "ra" +// column) remains unchanged throughout the function, even as the +// stack pointer takes on three different offsets from the return +// address. +// +// - Although we haven't shown it, most calling conventions designate +// "callee-saves" and "caller-saves" registers. The callee must +// preserve the values of callee-saves registers; if it uses them, +// it must save their original values somewhere, and restore them +// before it returns. In contrast, the callee is free to trash +// caller-saves registers; if the callee uses these, it will +// probably not bother to save them anywhere, and the CFI will +// probably mark their values as "unrecoverable". +// +// (However, since the caller cannot assume the callee was going to +// save them, caller-saves registers are probably dead in the caller +// anyway, so compilers usually don't generate CFA for caller-saves +// registers.) +// +// - Exactly where the CFA points is a matter of convention that +// depends on the architecture and ABI in use. In the example, the +// CFA is the value the stack pointer had upon entry to the +// function, pointing at the saved return address. But on the x86, +// the call frame information generated by GCC follows the +// convention that the CFA is the address *after* the saved return +// address. +// +// But by definition, the CFA remains constant throughout the +// lifetime of the frame. This makes it a useful value for other +// columns to refer to. It is also gives debuggers a useful handle +// for identifying a frame. +// +// If you look at the table above, you'll notice that a given entry is +// often the same as the one immediately above it: most instructions +// change only one or two aspects of the stack frame, if they affect +// it at all. The DWARF format takes advantage of this fact, and +// reduces the size of the data by mentioning only the addresses and +// columns at which changes take place. So for the above, DWARF CFI +// data would only actually mention the following: +// +// insn cfa r0 r1 ... ra +// ======================================= +// func+0: sp cfa[0] +// func+1: sp+16 +// func+2: cfa[-4] +// func+11: sp+20 +// func+21: r0 +// func+22: sp +// +// In fact, this is the way the parser reports CFI to the consumer: as +// a series of statements of the form, "At address X, column Y changed +// to Z," and related conventions for describing the initial state. +// +// Naturally, it would be impractical to have to scan the entire +// program's CFI, noting changes as we go, just to recover the +// unwinding rules in effect at one particular instruction. To avoid +// this, CFI data is grouped into "entries", each of which covers a +// specified range of addresses and begins with a complete statement +// of the rules for all recoverable registers at that starting +// address. Each entry typically covers a single function. +// +// Thus, to compute the contents of a given row of the table --- that +// is, rules for recovering the CFA, RA, and registers at a given +// instruction --- the consumer should find the entry that covers that +// instruction's address, start with the initial state supplied at the +// beginning of the entry, and work forward until it has processed all +// the changes up to and including those for the present instruction. +// +// There are seven kinds of rules that can appear in an entry of the +// table: +// +// - "undefined": The given register is not preserved by the callee; +// its value cannot be recovered. +// +// - "same value": This register has the same value it did in the callee. +// +// - offset(N): The register is saved at offset N from the CFA. +// +// - val_offset(N): The value the register had in the caller is the +// CFA plus offset N. (This is usually only useful for describing +// the stack pointer.) +// +// - register(R): The register's value was saved in another register R. +// +// - expression(E): Evaluating the DWARF expression E using the +// current frame's registers' values yields the address at which the +// register was saved. +// +// - val_expression(E): Evaluating the DWARF expression E using the +// current frame's registers' values yields the value the register +// had in the caller. + +class CallFrameInfo { + public: + // The different kinds of entries one finds in CFI. Used internally, + // and for error reporting. + enum EntryKind { kUnknown, kCIE, kFDE }; + + // The handler class to which the parser hands the parsed call frame + // information. Defined below. + class Handler; + + // A reporter class, which CallFrameInfo uses to report errors + // encountered while parsing call frame information. Defined below. + class Reporter; + + // Create a DWARF CFI parser. BUFFER points to the contents of the + // .debug_frame section to parse; BUFFER_LENGTH is its length in + // bytes. REPORTER is an error reporter the parser should use to + // report problems. READER is a ByteReader instance that has the + // endianness and address size set properly. Report the data we find + // to HANDLER. + CallFrameInfo(const char *buffer, size_t buffer_length, + ByteReader *reader, Handler *handler, + Reporter *reporter) + : buffer_(buffer), + buffer_length_(buffer_length), + reader_(reader), + handler_(handler), + reporter_(reporter) { } + ~CallFrameInfo() { } + + // Parse the entries in BUFFER, reporting what we find to HANDLER. + // Return true if we reach the end of the section successfully, or + // false if we encounter an error. + bool Start(); + + // Return the textual name of KIND. For error reporting. + static const char *KindName(EntryKind kind); + + private: + + struct CIE; + + // A CFI entry, either an FDE or a CIE. + struct Entry { + // The starting offset of the entry in the section, for error + // reporting. + size_t offset; + + // The start of this entry in the buffer. + const char *start; + + // The end of this entry's common prologue (initial length and id), and + // the start of this entry's kind-specific fields. + const char *fields; + + // The start of this entry's instructions. + const char *instructions; + + // The address past the entry's last byte in the buffer. (Note that + // since offset points to the entry's initial length field, and the + // length field is the number of bytes after that field, this is not + // simply buffer_ + offset + length.) + const char *end; + + // The CIE pointer or CIE id field. + uint64 id; + + // The kind of entry we're parsing. + // + // This may be kUnknown at times, since we want to be able to + // count on it for error reporting even before we've finished + // parsing enough to tell what kind of entry we're looking at. + EntryKind kind; + + // The CIE that applies to this entry, if we've parsed it. If this is a + // CIE, then this field points to this structure. + CIE *cie; + }; + + // A common information entry (CIE). + struct CIE: public Entry { + uint8 version; // CFI data version number + string augmentation; // vendor format extension markers + uint64 code_alignment_factor; // scale for code address adjustments + int data_alignment_factor; // scale for stack pointer adjustments + unsigned return_address_register; // which register holds the return addr + }; + + // A frame description entry (FDE). + struct FDE: public Entry { + uint64 address; // start address of described code + uint64 size; // size of described code, in bytes + }; + + // Internal use. + class Rule; + class UndefinedRule; + class SameValueRule; + class OffsetRule; + class ValOffsetRule; + class RegisterRule; + class ExpressionRule; + class ValExpressionRule; + class RuleMap; + class State; + + // Parse the initial length and id of a CFI entry, either a CIE or an + // FDE. CURSOR points to the beginning of the data to parse. + // On success, populate ENTRY as appropriate, and return true. + // On failure, report the problem, and return false. + bool ReadEntryPrologue(const char *cursor, Entry *entry); + + // Parse the fields of a CIE after the entry prologue. Assume that the + // 'Entry' fields of CIE are populated; use CIE->fields and CIE->end as + // the start and limit for parsing. On success, populate the rest of + // *CIE, and return true; on failure, report the problem and return + // false. + bool ReadCIEFields(CIE *cie); + + // Parse the fields of an FDE after the entry prologue. Assume that the + // 'Entry' fields of *FDE are initialized; use FDE->fields and FDE->end + // as the start and limit for parsing. Assume that FDE->cie is fully + // initialized. On success, populate the rest of *FDE, and return true; + // on failure, report the problem and return false. + bool ReadFDEFields(FDE *fde); + + // Report that ENTRY is incomplete, and return false. This is just a + // trivial wrapper for invoking reporter_->Incomplete; it provides a + // little brevity. + bool ReportIncomplete(Entry *entry); + + // The contents of the DWARF .debug_info section we're parsing. + const char *buffer_; + size_t buffer_length_; + + // 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 handler class for CallFrameInfo. The a CFI parser calls the +// member functions of a handler object to report the data it finds. +class CallFrameInfo::Handler { + public: + // The pseudo-register number for the canonical frame address. + enum { kCFARegister = -1 }; + + Handler() { } + virtual ~Handler() { } + + // The parser has found CFI for the machine code at ADDRESS, + // extending for LENGTH bytes. OFFSET is the offset of the frame + // description entry in the section, for use in error messages. + // VERSION is the version number of the CFI format. AUGMENTATION is + // a string describing any producer-specific extensions present in + // the data. RETURN_ADDRESS is the number of the register that holds + // the address to which the function should return. + // + // Entry should return true to process this CFI, or false to skip to + // the next entry. + // + // The parser invokes Entry for each Frame Description Entry (FDE) + // it finds. The parser doesn't report Common Information Entries + // to the handler explicitly; instead, if the handler elects to + // process a given FDE, the parser reiterates the appropriate CIE's + // contents at the beginning of the FDE's rules. + virtual bool Entry(size_t offset, uint64 address, uint64 length, + uint8 version, const string &augmentation, + unsigned return_address) = 0; + + // When the Entry function returns true, the parser calls these + // handler functions repeatedly to describe the rules for recovering + // registers at each instruction in the given range of machine code. + // Immediately after a call to Entry, the handler should assume that + // the rule for each callee-saves register is "unchanged" --- that + // is, that the register still has the value it had in the caller. + // + // If a *Rule function returns true, we continue processing this entry's + // instructions. If a *Rule function returns false, we stop evaluating + // instructions, and skip to the next entry. Either way, we call End + // before going on to the next entry. + // + // In all of these functions, if the REG parameter is kCFARegister, then + // the rule describes how to find the canonical frame address. + // kCFARegister may be passed as a BASE_REGISTER argument, meaning that + // the canonical frame address should be used as the base address for the + // computation. All other REG values will be positive. + + // At ADDRESS, register REG's value is not recoverable. + virtual bool UndefinedRule(uint64 address, int reg) = 0; + + // At ADDRESS, register REG's value is the same as that it had in + // the caller. + virtual bool SameValueRule(uint64 address, int reg) = 0; + + // At ADDRESS, register REG has been saved at offset OFFSET from + // BASE_REGISTER. + virtual bool OffsetRule(uint64 address, int reg, + int base_register, long offset) = 0; + + // At ADDRESS, the caller's value of register REG is the current + // value of BASE_REGISTER plus OFFSET. (This rule doesn't provide an + // address at which the register's value is saved.) + virtual bool ValOffsetRule(uint64 address, int reg, + int base_register, long offset) = 0; + + // At ADDRESS, register REG has been saved in BASE_REGISTER. This differs + // from ValOffsetRule(ADDRESS, REG, BASE_REGISTER, 0), in that + // BASE_REGISTER is the "home" for REG's saved value: if you want to + // assign to a variable whose home is REG in the calling frame, you + // should put the value in BASE_REGISTER. + virtual bool RegisterRule(uint64 address, int reg, int base_register) = 0; + + // At ADDRESS, the DWARF expression EXPRESSION yields the address at + // which REG was saved. + virtual bool ExpressionRule(uint64 address, int reg, + const string &expression) = 0; + + // At ADDRESS, the DWARF expression EXPRESSION yields the caller's + // value for REG. (This rule doesn't provide an address at which the + // register's value is saved.) + virtual bool ValExpressionRule(uint64 address, int reg, + const string &expression) = 0; + + // Indicate that the rules for the address range reported by the + // last call to Entry are complete. End should return true if + // everything is okay, or false if an error has occurred and parsing + // should stop. + virtual bool End() = 0; +}; + +// The CallFrameInfo class makes calls on an instance of this class to +// report errors or warn about problems in the data it is parsing. The +// default definitions of these methods print a message to stderr, but +// you can make a derived class that overrides them. +class CallFrameInfo::Reporter { + public: + // Create an error reporter which attributes troubles to the section + // named SECTION in FILENAME. + // + // Normally SECTION would be .debug_frame, but the Mac puts CFI data + // in a Mach-O section named __debug_frame. If we support + // Linux-style exception handling data, we could be reading an + // .eh_frame section. + Reporter(const string &filename, + const string §ion = ".debug_frame") + : filename_(filename), section_(section) { } + virtual ~Reporter() { } + + // The CFI entry at OFFSET ends too early to be well-formed. KIND + // indicates what kind of entry it is; KIND can be kUnknown if we + // haven't parsed enough of the entry to tell yet. + virtual void Incomplete(uint64 offset, CallFrameInfo::EntryKind kind); + + // The FDE at OFFSET refers to the CIE at CIE_OFFSET, but the + // section is not that large. + virtual void CIEPointerOutOfRange(uint64 offset, uint64 cie_offset); + + // The FDE at OFFSET refers to the CIE at CIE_OFFSET, but the entry + // there is not a CIE. + virtual void BadCIEId(uint64 offset, uint64 cie_offset); + + // The FDE at OFFSET refers to a CIE with version number VERSION, + // which we don't recognize. We cannot parse DWARF CFI if it uses + // a version number we don't recognize. + virtual void UnrecognizedVersion(uint64 offset, int version); + + // The FDE at OFFSET refers to a CIE with augmentation AUGMENTATION, + // which we don't recognize. We cannot parse DWARF CFI if it uses + // augmentations we don't recognize. + virtual void UnrecognizedAugmentation(uint64 offset, + const string &augmentation); + + // The CIE at OFFSET contains a DW_CFA_restore instruction at + // INSN_OFFSET, which may not appear in a CIE. + virtual void RestoreInCIE(uint64 offset, uint64 insn_offset); + + // The entry at OFFSET, of kind KIND, has an unrecognized + // instruction at INSN_OFFSET. + virtual void BadInstruction(uint64 offset, CallFrameInfo::EntryKind kind, + uint64 insn_offset); + + // The instruction at INSN_OFFSET in the entry at OFFSET, of kind + // KIND, establishes a rule that cites the CFA, but we have not + // established a CFA rule yet. + virtual void NoCFARule(uint64 offset, CallFrameInfo::EntryKind kind, + uint64 insn_offset); + + // The instruction at INSN_OFFSET in the entry at OFFSET, of kind + // KIND, is a DW_CFA_restore_state instruction, but the stack of + // saved states is empty. + virtual void EmptyStateStack(uint64 offset, CallFrameInfo::EntryKind kind, + uint64 insn_offset); + + // The DW_CFA_remember_state instruction at INSN_OFFSET in the entry + // at OFFSET, of kind KIND, would restore a state that has no CFA + // rule, whereas the current state does have a CFA rule. This is + // bogus input, which the CallFrameInfo::Handler interface doesn't + // (and shouldn't) have any way to report. + virtual void ClearingCFARule(uint64 offset, CallFrameInfo::EntryKind kind, + uint64 insn_offset); + + protected: + // The name of the file whose CFI we're reading. + string filename_; + + // The name of the CFI section in that file. + string section_; +}; } // namespace dwarf2reader |