Source Level Debugging with LLVM


This document is the central repository for all information pertaining to debug information in LLVM. It describes the actual format that the LLVM debug information takes, which is useful for those interested in creating front-ends or dealing directly with the information. Further, this document provides specific examples of what debug information for C/C++ looks like.

Philosophy behind LLVM debugging information

The idea of the LLVM debugging information is to capture how the important pieces of the source-language’s Abstract Syntax Tree map onto LLVM code. Several design aspects have shaped the solution that appears here. The important ones are:

  • Debugging information should have very little impact on the rest of the compiler. No transformations, analyses, or code generators should need to be modified because of debugging information.
  • LLVM optimizations should interact in well-defined and easily described ways with the debugging information.
  • Because LLVM is designed to support arbitrary programming languages, LLVM-to-LLVM tools should not need to know anything about the semantics of the source-level-language.
  • Source-level languages are often widely different from one another. LLVM should not put any restrictions of the flavor of the source-language, and the debugging information should work with any language.
  • With code generator support, it should be possible to use an LLVM compiler to compile a program to native machine code and standard debugging formats. This allows compatibility with traditional machine-code level debuggers, like GDB or DBX.

The approach used by the LLVM implementation is to use a small set of intrinsic functions to define a mapping between LLVM program objects and the source-level objects. The description of the source-level program is maintained in LLVM metadata in an implementation-defined format (the C/C++ front-end currently uses working draft 7 of the DWARF 3 standard).

When a program is being debugged, a debugger interacts with the user and turns the stored debug information into source-language specific information. As such, a debugger must be aware of the source-language, and is thus tied to a specific language or family of languages.

Debug information consumers

The role of debug information is to provide meta information normally stripped away during the compilation process. This meta information provides an LLVM user a relationship between generated code and the original program source code.

Currently, there are two backend consumers of debug info: DwarfDebug and CodeViewDebug. DwarfDebug produces DWARF sutable for use with GDB, LLDB, and other DWARF-based debuggers. CodeViewDebug produces CodeView, the Microsoft debug info format, which is usable with Microsoft debuggers such as Visual Studio and WinDBG. LLVM’s debug information format is mostly derived from and inspired by DWARF, but it is feasible to translate into other target debug info formats such as STABS.

It would also be reasonable to use debug information to feed profiling tools for analysis of generated code, or, tools for reconstructing the original source from generated code.

Debugging optimized code

An extremely high priority of LLVM debugging information is to make it interact well with optimizations and analysis. In particular, the LLVM debug information provides the following guarantees:

  • LLVM debug information always provides information to accurately read the source-level state of the program, regardless of which LLVM optimizations have been run, and without any modification to the optimizations themselves. However, some optimizations may impact the ability to modify the current state of the program with a debugger, such as setting program variables, or calling functions that have been deleted.
  • As desired, LLVM optimizations can be upgraded to be aware of the LLVM debugging information, allowing them to update the debugging information as they perform aggressive optimizations. This means that, with effort, the LLVM optimizers could optimize debug code just as well as non-debug code.
  • LLVM debug information does not prevent optimizations from happening (for example inlining, basic block reordering/merging/cleanup, tail duplication, etc).
  • LLVM debug information is automatically optimized along with the rest of the program, using existing facilities. For example, duplicate information is automatically merged by the linker, and unused information is automatically removed.

Basically, the debug information allows you to compile a program with “-O0 -g” and get full debug information, allowing you to arbitrarily modify the program as it executes from a debugger. Compiling a program with “-O3 -g” gives you full debug information that is always available and accurate for reading (e.g., you get accurate stack traces despite tail call elimination and inlining), but you might lose the ability to modify the program and call functions where were optimized out of the program, or inlined away completely.

LLVM test suite provides a framework to test optimizer’s handling of debugging information. It can be run like this:

% cd llvm/projects/test-suite/MultiSource/Benchmarks  # or some other level
% make TEST=dbgopt

This will test impact of debugging information on optimization passes. If debugging information influences optimization passes then it will be reported as a failure. See LLVM Testing Infrastructure Guide for more information on LLVM test infrastructure and how to run various tests.

Debugging information format

LLVM debugging information has been carefully designed to make it possible for the optimizer to optimize the program and debugging information without necessarily having to know anything about debugging information. In particular, the use of metadata avoids duplicated debugging information from the beginning, and the global dead code elimination pass automatically deletes debugging information for a function if it decides to delete the function.

To do this, most of the debugging information (descriptors for types, variables, functions, source files, etc) is inserted by the language front-end in the form of LLVM metadata.

Debug information is designed to be agnostic about the target debugger and debugging information representation (e.g. DWARF/Stabs/etc). It uses a generic pass to decode the information that represents variables, types, functions, namespaces, etc: this allows for arbitrary source-language semantics and type-systems to be used, as long as there is a module written for the target debugger to interpret the information.

To provide basic functionality, the LLVM debugger does have to make some assumptions about the source-level language being debugged, though it keeps these to a minimum. The only common features that the LLVM debugger assumes exist are source files, and program objects. These abstract objects are used by a debugger to form stack traces, show information about local variables, etc.

This section of the documentation first describes the representation aspects common to any source-language. C/C++ front-end specific debug information describes the data layout conventions used by the C and C++ front-ends.

Debug information descriptors are specialized metadata nodes, first-class subclasses of Metadata.

Debugger intrinsic functions

LLVM uses several intrinsic functions (name prefixed with “llvm.dbg”) to provide debug information at various points in generated code.


void @llvm.dbg.declare(metadata, metadata, metadata)

This intrinsic provides information about a local element (e.g., variable). The first argument is metadata holding the alloca for the variable. The second argument is a local variable containing a description of the variable. The third argument is a complex expression.


void @llvm.dbg.value(metadata, i64, metadata, metadata)

This intrinsic provides information when a user source variable is set to a new value. The first argument is the new value (wrapped as metadata). The second argument is the offset in the user source variable where the new value is written. The third argument is a local variable containing a description of the variable. The fourth argument is a complex expression.

Object lifetimes and scoping

In many languages, the local variables in functions can have their lifetimes or scopes limited to a subset of a function. In the C family of languages, for example, variables are only live (readable and writable) within the source block that they are defined in. In functional languages, values are only readable after they have been defined. Though this is a very obvious concept, it is non-trivial to model in LLVM, because it has no notion of scoping in this sense, and does not want to be tied to a language’s scoping rules.

In order to handle this, the LLVM debug format uses the metadata attached to llvm instructions to encode line number and scoping information. Consider the following C fragment, for example:

1.  void foo() {
2.    int X = 21;
3.    int Y = 22;
4.    {
5.      int Z = 23;
6.      Z = X;
7.    }
8.    X = Y;
9.  }

Compiled to LLVM, this function would be represented like this:

; Function Attrs: nounwind ssp uwtable
define void @foo() #0 !dbg !4 {
  %X = alloca i32, align 4
  %Y = alloca i32, align 4
  %Z = alloca i32, align 4
  call void @llvm.dbg.declare(metadata i32* %X, metadata !11, metadata !13), !dbg !14
  store i32 21, i32* %X, align 4, !dbg !14
  call void @llvm.dbg.declare(metadata i32* %Y, metadata !15, metadata !13), !dbg !16
  store i32 22, i32* %Y, align 4, !dbg !16
  call void @llvm.dbg.declare(metadata i32* %Z, metadata !17, metadata !13), !dbg !19
  store i32 23, i32* %Z, align 4, !dbg !19
  %0 = load i32, i32* %X, align 4, !dbg !20
  store i32 %0, i32* %Z, align 4, !dbg !21
  %1 = load i32, i32* %Y, align 4, !dbg !22
  store i32 %1, i32* %X, align 4, !dbg !23
  ret void, !dbg !24

; Function Attrs: nounwind readnone
declare void @llvm.dbg.declare(metadata, metadata, metadata) #1

attributes #0 = { nounwind ssp uwtable "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }
attributes #1 = { nounwind readnone }

! = !{!0}
!llvm.module.flags = !{!7, !8, !9}
!llvm.ident = !{!10}

!0 = !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang version 3.7.0 (trunk 231150) (llvm/trunk 231154)", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, enums: !2, retainedTypes: !2, subprograms: !3, globals: !2, imports: !2)
!1 = !DIFile(filename: "/dev/stdin", directory: "/Users/dexonsmith/data/llvm/debug-info")
!2 = !{}
!3 = !{!4}
!4 = distinct !DISubprogram(name: "foo", scope: !1, file: !1, line: 1, type: !5, isLocal: false, isDefinition: true, scopeLine: 1, isOptimized: false, variables: !2)
!5 = !DISubroutineType(types: !6)
!6 = !{null}
!7 = !{i32 2, !"Dwarf Version", i32 2}
!8 = !{i32 2, !"Debug Info Version", i32 3}
!9 = !{i32 1, !"PIC Level", i32 2}
!10 = !{!"clang version 3.7.0 (trunk 231150) (llvm/trunk 231154)"}
!11 = !DILocalVariable(name: "X", scope: !4, file: !1, line: 2, type: !12)
!12 = !DIBasicType(name: "int", size: 32, align: 32, encoding: DW_ATE_signed)
!13 = !DIExpression()
!14 = !DILocation(line: 2, column: 9, scope: !4)
!15 = !DILocalVariable(name: "Y", scope: !4, file: !1, line: 3, type: !12)
!16 = !DILocation(line: 3, column: 9, scope: !4)
!17 = !DILocalVariable(name: "Z", scope: !18, file: !1, line: 5, type: !12)
!18 = distinct !DILexicalBlock(scope: !4, file: !1, line: 4, column: 5)
!19 = !DILocation(line: 5, column: 11, scope: !18)
!20 = !DILocation(line: 6, column: 11, scope: !18)
!21 = !DILocation(line: 6, column: 9, scope: !18)
!22 = !DILocation(line: 8, column: 9, scope: !4)
!23 = !DILocation(line: 8, column: 7, scope: !4)
!24 = !DILocation(line: 9, column: 3, scope: !4)

This example illustrates a few important details about LLVM debugging information. In particular, it shows how the llvm.dbg.declare intrinsic and location information, which are attached to an instruction, are applied together to allow a debugger to analyze the relationship between statements, variable definitions, and the code used to implement the function.

call void @llvm.dbg.declare(metadata i32* %X, metadata !11, metadata !13), !dbg !14
  ; [debug line = 2:7] [debug variable = X]

The first intrinsic %llvm.dbg.declare encodes debugging information for the variable X. The metadata !dbg !14 attached to the intrinsic provides scope information for the variable X.

!14 = !DILocation(line: 2, column: 9, scope: !4)
!4 = distinct !DISubprogram(name: "foo", scope: !1, file: !1, line: 1, type: !5,
                            isLocal: false, isDefinition: true, scopeLine: 1,
                            isOptimized: false, variables: !2)

Here !14 is metadata providing location information. In this example, scope is encoded by !4, a subprogram descriptor. This way the location information attached to the intrinsics indicates that the variable X is declared at line number 2 at a function level scope in function foo.

Now lets take another example.

call void @llvm.dbg.declare(metadata i32* %Z, metadata !17, metadata !13), !dbg !19
  ; [debug line = 5:9] [debug variable = Z]

The third intrinsic %llvm.dbg.declare encodes debugging information for variable Z. The metadata !dbg !19 attached to the intrinsic provides scope information for the variable Z.

!18 = distinct !DILexicalBlock(scope: !4, file: !1, line: 4, column: 5)
!19 = !DILocation(line: 5, column: 11, scope: !18)

Here !19 indicates that Z is declared at line number 5 and column number 0 inside of lexical scope !18. The lexical scope itself resides inside of subprogram !4 described above.

The scope information attached with each instruction provides a straightforward way to find instructions covered by a scope.

C/C++ front-end specific debug information

The C and C++ front-ends represent information about the program in a format that is effectively identical to DWARF 3.0 in terms of information content. This allows code generators to trivially support native debuggers by generating standard dwarf information, and contains enough information for non-dwarf targets to translate it as needed.

This section describes the forms used to represent C and C++ programs. Other languages could pattern themselves after this (which itself is tuned to representing programs in the same way that DWARF 3 does), or they could choose to provide completely different forms if they don’t fit into the DWARF model. As support for debugging information gets added to the various LLVM source-language front-ends, the information used should be documented here.

The following sections provide examples of a few C/C++ constructs and the debug information that would best describe those constructs. The canonical references are the DIDescriptor classes defined in include/llvm/IR/DebugInfo.h and the implementations of the helper functions in lib/IR/DIBuilder.cpp.

C/C++ source file information

llvm::Instruction provides easy access to metadata attached with an instruction. One can extract line number information encoded in LLVM IR using Instruction::getDebugLoc() and DILocation::getLine().

if (DILocation *Loc = I->getDebugLoc()) { // Here I is an LLVM instruction
  unsigned Line = Loc->getLine();
  StringRef File = Loc->getFilename();
  StringRef Dir = Loc->getDirectory();

C/C++ global variable information

Given an integer global variable declared as follows:

int MyGlobal = 100;

a C/C++ front-end would generate the following descriptors:

;; Define the global itself.
@MyGlobal = global i32 100, align 4

;; List of debug info of globals
! = !{!0}

;; Some unrelated metadata.
!llvm.module.flags = !{!6, !7}

;; Define the compile unit.
!0 = !DICompileUnit(language: DW_LANG_C99, file: !1,
                    "clang version 3.7.0 (trunk 231150) (llvm/trunk 231154)",
                    isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug,
                    enums: !2, retainedTypes: !2, subprograms: !2, globals:
                    !3, imports: !2)

;; Define the file
!1 = !DIFile(filename: "/dev/stdin",
             directory: "/Users/dexonsmith/data/llvm/debug-info")

;; An empty array.
!2 = !{}

;; The Array of Global Variables
!3 = !{!4}

;; Define the global variable itself.
!4 = !DIGlobalVariable(name: "MyGlobal", scope: !0, file: !1, line: 1,
                       type: !5, isLocal: false, isDefinition: true,
                       variable: i32* @MyGlobal)

;; Define the type
!5 = !DIBasicType(name: "int", size: 32, align: 32, encoding: DW_ATE_signed)

;; Dwarf version to output.
!6 = !{i32 2, !"Dwarf Version", i32 2}

;; Debug info schema version.
!7 = !{i32 2, !"Debug Info Version", i32 3}

C/C++ function information

Given a function declared as follows:

int main(int argc, char *argv[]) {
  return 0;

a C/C++ front-end would generate the following descriptors:

;; Define the anchor for subprograms.
!4 = !DISubprogram(name: "main", scope: !1, file: !1, line: 1, type: !5,
                   isLocal: false, isDefinition: true, scopeLine: 1,
                   flags: DIFlagPrototyped, isOptimized: false,
                   variables: !2)

;; Define the subprogram itself.
define i32 @main(i32 %argc, i8** %argv) !dbg !4 {

Debugging information format

Debugging Information Extension for Objective C Properties


Objective C provides a simpler way to declare and define accessor methods using declared properties. The language provides features to declare a property and to let compiler synthesize accessor methods.

The debugger lets developer inspect Objective C interfaces and their instance variables and class variables. However, the debugger does not know anything about the properties defined in Objective C interfaces. The debugger consumes information generated by compiler in DWARF format. The format does not support encoding of Objective C properties. This proposal describes DWARF extensions to encode Objective C properties, which the debugger can use to let developers inspect Objective C properties.


Objective C properties exist separately from class members. A property can be defined only by “setter” and “getter” selectors, and be calculated anew on each access. Or a property can just be a direct access to some declared ivar. Finally it can have an ivar “automatically synthesized” for it by the compiler, in which case the property can be referred to in user code directly using the standard C dereference syntax as well as through the property “dot” syntax, but there is no entry in the @interface declaration corresponding to this ivar.

To facilitate debugging, these properties we will add a new DWARF TAG into the DW_TAG_structure_type definition for the class to hold the description of a given property, and a set of DWARF attributes that provide said description. The property tag will also contain the name and declared type of the property.

If there is a related ivar, there will also be a DWARF property attribute placed in the DW_TAG_member DIE for that ivar referring back to the property TAG for that property. And in the case where the compiler synthesizes the ivar directly, the compiler is expected to generate a DW_TAG_member for that ivar (with the DW_AT_artificial set to 1), whose name will be the name used to access this ivar directly in code, and with the property attribute pointing back to the property it is backing.

The following examples will serve as illustration for our discussion:

@interface I1 {
  int n2;

@property int p1;
@property int p2;

@implementation I1
@synthesize p1;
@synthesize p2 = n2;

This produces the following DWARF (this is a “pseudo dwarfdump” output):

0x00000100:  TAG_structure_type [7] *
               AT_APPLE_runtime_class( 0x10 )
               AT_name( "I1" )
               AT_decl_file( "Objc_Property.m" )
               AT_decl_line( 3 )

0x00000110    TAG_APPLE_property
                AT_name ( "p1" )
                AT_type ( {0x00000150} ( int ) )

0x00000120:   TAG_APPLE_property
                AT_name ( "p2" )
                AT_type ( {0x00000150} ( int ) )

0x00000130:   TAG_member [8]
                AT_name( "_p1" )
                AT_APPLE_property ( {0x00000110} "p1" )
                AT_type( {0x00000150} ( int ) )
                AT_artificial ( 0x1 )

0x00000140:    TAG_member [8]
                 AT_name( "n2" )
                 AT_APPLE_property ( {0x00000120} "p2" )
                 AT_type( {0x00000150} ( int ) )

0x00000150:  AT_type( ( int ) )

Note, the current convention is that the name of the ivar for an auto-synthesized property is the name of the property from which it derives with an underscore prepended, as is shown in the example. But we actually don’t need to know this convention, since we are given the name of the ivar directly.

Also, it is common practice in ObjC to have different property declarations in the @interface and @implementation - e.g. to provide a read-only property in the interface,and a read-write interface in the implementation. In that case, the compiler should emit whichever property declaration will be in force in the current translation unit.

Developers can decorate a property with attributes which are encoded using DW_AT_APPLE_property_attribute.

@property (readonly, nonatomic) int pr;
TAG_APPLE_property [8]
  AT_name( "pr" )
  AT_type ( {0x00000147} (int) )
  AT_APPLE_property_attribute (DW_APPLE_PROPERTY_readonly, DW_APPLE_PROPERTY_nonatomic)

The setter and getter method names are attached to the property using DW_AT_APPLE_property_setter and DW_AT_APPLE_property_getter attributes.

@interface I1
@property (setter=myOwnP3Setter:) int p3;

@implementation I1
@synthesize p3;
-(void)myOwnP3Setter:(int)a{ }

The DWARF for this would be:

0x000003bd: TAG_structure_type [7] *
              AT_APPLE_runtime_class( 0x10 )
              AT_name( "I1" )
              AT_decl_file( "Objc_Property.m" )
              AT_decl_line( 3 )

0x000003cd      TAG_APPLE_property
                  AT_name ( "p3" )
                  AT_APPLE_property_setter ( "myOwnP3Setter:" )
                  AT_type( {0x00000147} ( int ) )

0x000003f3:     TAG_member [8]
                  AT_name( "_p3" )
                  AT_type ( {0x00000147} ( int ) )
                  AT_APPLE_property ( {0x000003cd} )
                  AT_artificial ( 0x1 )

New DWARF Tags

TAG Value
DW_TAG_APPLE_property 0x4200

New DWARF Attributes

Attribute Value Classes
DW_AT_APPLE_property 0x3fed Reference
DW_AT_APPLE_property_getter 0x3fe9 String
DW_AT_APPLE_property_setter 0x3fea String
DW_AT_APPLE_property_attribute 0x3feb Constant

New DWARF Constants

Name Value
DW_APPLE_PROPERTY_readonly 0x01
DW_APPLE_PROPERTY_readwrite 0x08
DW_APPLE_PROPERTY_nonatomic 0x40
DW_APPLE_PROPERTY_atomic 0x100
DW_APPLE_PROPERTY_strong 0x400
DW_APPLE_PROPERTY_unsafe_unretained 0x800
DW_APPLE_PROPERTY_nullability 0x1000
DW_APPLE_PROPERTY_null_resettable 0x2000
DW_APPLE_PROPERTY_class 0x4000

Name Accelerator Tables


The “.debug_pubnames” and “.debug_pubtypes” formats are not what a debugger needs. The “pub” in the section name indicates that the entries in the table are publicly visible names only. This means no static or hidden functions show up in the “.debug_pubnames”. No static variables or private class variables are in the “.debug_pubtypes”. Many compilers add different things to these tables, so we can’t rely upon the contents between gcc, icc, or clang.

The typical query given by users tends not to match up with the contents of these tables. For example, the DWARF spec states that “In the case of the name of a function member or static data member of a C++ structure, class or union, the name presented in the “.debug_pubnames” section is not the simple name given by the DW_AT_name attribute of the referenced debugging information entry, but rather the fully qualified name of the data or function member.” So the only names in these tables for complex C++ entries is a fully qualified name. Debugger users tend not to enter their search strings as “a::b::c(int,const Foo&) const”, but rather as “c”, “b::c” , or “a::b::c”. So the name entered in the name table must be demangled in order to chop it up appropriately and additional names must be manually entered into the table to make it effective as a name lookup table for debuggers to use.

All debuggers currently ignore the “.debug_pubnames” table as a result of its inconsistent and useless public-only name content making it a waste of space in the object file. These tables, when they are written to disk, are not sorted in any way, leaving every debugger to do its own parsing and sorting. These tables also include an inlined copy of the string values in the table itself making the tables much larger than they need to be on disk, especially for large C++ programs.

Can’t we just fix the sections by adding all of the names we need to this table? No, because that is not what the tables are defined to contain and we won’t know the difference between the old bad tables and the new good tables. At best we could make our own renamed sections that contain all of the data we need.

These tables are also insufficient for what a debugger like LLDB needs. LLDB uses clang for its expression parsing where LLDB acts as a PCH. LLDB is then often asked to look for type “foo” or namespace “bar”, or list items in namespace “baz”. Namespaces are not included in the pubnames or pubtypes tables. Since clang asks a lot of questions when it is parsing an expression, we need to be very fast when looking up names, as it happens a lot. Having new accelerator tables that are optimized for very quick lookups will benefit this type of debugging experience greatly.

We would like to generate name lookup tables that can be mapped into memory from disk, and used as is, with little or no up-front parsing. We would also be able to control the exact content of these different tables so they contain exactly what we need. The Name Accelerator Tables were designed to fix these issues. In order to solve these issues we need to:

  • Have a format that can be mapped into memory from disk and used as is
  • Lookups should be very fast
  • Extensible table format so these tables can be made by many producers
  • Contain all of the names needed for typical lookups out of the box
  • Strict rules for the contents of tables

Table size is important and the accelerator table format should allow the reuse of strings from common string tables so the strings for the names are not duplicated. We also want to make sure the table is ready to be used as-is by simply mapping the table into memory with minimal header parsing.

The name lookups need to be fast and optimized for the kinds of lookups that debuggers tend to do. Optimally we would like to touch as few parts of the mapped table as possible when doing a name lookup and be able to quickly find the name entry we are looking for, or discover there are no matches. In the case of debuggers we optimized for lookups that fail most of the time.

Each table that is defined should have strict rules on exactly what is in the accelerator tables and documented so clients can rely on the content.

Hash Tables

Standard Hash Tables

Typical hash tables have a header, buckets, and each bucket points to the bucket contents:

|  HEADER    |
|  BUCKETS   |
|  DATA      |

The BUCKETS are an array of offsets to DATA for each hash:

| 0x00001000 | BUCKETS[0]
| 0x00002000 | BUCKETS[1]
| 0x00002200 | BUCKETS[2]
| 0x000034f0 | BUCKETS[3]
|            | ...
| 0xXXXXXXXX | BUCKETS[n_buckets]

So for bucket[3] in the example above, we have an offset into the table 0x000034f0 which points to a chain of entries for the bucket. Each bucket must contain a next pointer, full 32 bit hash value, the string itself, and the data for the current string value.

0x000034f0: | 0x00003500 | next pointer
            | 0x12345678 | 32 bit hash
            | "erase"    | string value
            | data[n]    | HashData for this bucket
0x00003500: | 0x00003550 | next pointer
            | 0x29273623 | 32 bit hash
            | "dump"     | string value
            | data[n]    | HashData for this bucket
0x00003550: | 0x00000000 | next pointer
            | 0x82638293 | 32 bit hash
            | "main"     | string value
            | data[n]    | HashData for this bucket

The problem with this layout for debuggers is that we need to optimize for the negative lookup case where the symbol we’re searching for is not present. So if we were to lookup “printf” in the table above, we would make a 32 hash for “printf”, it might match bucket[3]. We would need to go to the offset 0x000034f0 and start looking to see if our 32 bit hash matches. To do so, we need to read the next pointer, then read the hash, compare it, and skip to the next bucket. Each time we are skipping many bytes in memory and touching new cache pages just to do the compare on the full 32 bit hash. All of these accesses then tell us that we didn’t have a match.

Name Hash Tables

To solve the issues mentioned above we have structured the hash tables a bit differently: a header, buckets, an array of all unique 32 bit hash values, followed by an array of hash value data offsets, one for each hash value, then the data for all hash values:

|  HEADER     |
|  BUCKETS    |
|  HASHES     |
|  OFFSETS    |
|  DATA       |

The BUCKETS in the name tables are an index into the HASHES array. By making all of the full 32 bit hash values contiguous in memory, we allow ourselves to efficiently check for a match while touching as little memory as possible. Most often checking the 32 bit hash values is as far as the lookup goes. If it does match, it usually is a match with no collisions. So for a table with “n_buckets” buckets, and “n_hashes” unique 32 bit hash values, we can clarify the contents of the BUCKETS, HASHES and OFFSETS as:

|  HEADER.magic           | uint32_t
|  HEADER.version         | uint16_t
|  HEADER.hash_function   | uint16_t
|  HEADER.bucket_count    | uint32_t
|  HEADER.hashes_count    | uint32_t
|  HEADER.header_data_len | uint32_t
|  HEADER_DATA            | HeaderData
|  BUCKETS                | uint32_t[n_buckets] // 32 bit hash indexes
|  HASHES                 | uint32_t[n_hashes] // 32 bit hash values
|  OFFSETS                | uint32_t[n_hashes] // 32 bit offsets to hash value data
|  ALL HASH DATA          |

So taking the exact same data from the standard hash example above we end up with:

            | HEADER     |
            |          0 | BUCKETS[0]
            |          2 | BUCKETS[1]
            |          5 | BUCKETS[2]
            |          6 | BUCKETS[3]
            |            | ...
            |        ... | BUCKETS[n_buckets]
            | 0x........ | HASHES[0]
            | 0x........ | HASHES[1]
            | 0x........ | HASHES[2]
            | 0x........ | HASHES[3]
            | 0x........ | HASHES[4]
            | 0x........ | HASHES[5]
            | 0x12345678 | HASHES[6]    hash for BUCKETS[3]
            | 0x29273623 | HASHES[7]    hash for BUCKETS[3]
            | 0x82638293 | HASHES[8]    hash for BUCKETS[3]
            | 0x........ | HASHES[9]
            | 0x........ | HASHES[10]
            | 0x........ | HASHES[11]
            | 0x........ | HASHES[12]
            | 0x........ | HASHES[13]
            | 0x........ | HASHES[n_hashes]
            | 0x........ | OFFSETS[0]
            | 0x........ | OFFSETS[1]
            | 0x........ | OFFSETS[2]
            | 0x........ | OFFSETS[3]
            | 0x........ | OFFSETS[4]
            | 0x........ | OFFSETS[5]
            | 0x000034f0 | OFFSETS[6]   offset for BUCKETS[3]
            | 0x00003500 | OFFSETS[7]   offset for BUCKETS[3]
            | 0x00003550 | OFFSETS[8]   offset for BUCKETS[3]
            | 0x........ | OFFSETS[9]
            | 0x........ | OFFSETS[10]
            | 0x........ | OFFSETS[11]
            | 0x........ | OFFSETS[12]
            | 0x........ | OFFSETS[13]
            | 0x........ | OFFSETS[n_hashes]
            |            |
            |            |
            |            |
            |            |
            |            |
0x000034f0: | 0x00001203 | .debug_str ("erase")
            | 0x00000004 | A 32 bit array count - number of HashData with name "erase"
            | 0x........ | HashData[0]
            | 0x........ | HashData[1]
            | 0x........ | HashData[2]
            | 0x........ | HashData[3]
            | 0x00000000 | String offset into .debug_str (terminate data for hash)
0x00003500: | 0x00001203 | String offset into .debug_str ("collision")
            | 0x00000002 | A 32 bit array count - number of HashData with name "collision"
            | 0x........ | HashData[0]
            | 0x........ | HashData[1]
            | 0x00001203 | String offset into .debug_str ("dump")
            | 0x00000003 | A 32 bit array count - number of HashData with name "dump"
            | 0x........ | HashData[0]
            | 0x........ | HashData[1]
            | 0x........ | HashData[2]
            | 0x00000000 | String offset into .debug_str (terminate data for hash)
0x00003550: | 0x00001203 | String offset into .debug_str ("main")
            | 0x00000009 | A 32 bit array count - number of HashData with name "main"
            | 0x........ | HashData[0]
            | 0x........ | HashData[1]
            | 0x........ | HashData[2]
            | 0x........ | HashData[3]
            | 0x........ | HashData[4]
            | 0x........ | HashData[5]
            | 0x........ | HashData[6]
            | 0x........ | HashData[7]
            | 0x........ | HashData[8]
            | 0x00000000 | String offset into .debug_str (terminate data for hash)

So we still have all of the same data, we just organize it more efficiently for debugger lookup. If we repeat the same “printf” lookup from above, we would hash “printf” and find it matches BUCKETS[3] by taking the 32 bit hash value and modulo it by n_buckets. BUCKETS[3] contains “6” which is the index into the HASHES table. We would then compare any consecutive 32 bit hashes values in the HASHES array as long as the hashes would be in BUCKETS[3]. We do this by verifying that each subsequent hash value modulo n_buckets is still 3. In the case of a failed lookup we would access the memory for BUCKETS[3], and then compare a few consecutive 32 bit hashes before we know that we have no match. We don’t end up marching through multiple words of memory and we really keep the number of processor data cache lines being accessed as small as possible.

The string hash that is used for these lookup tables is the Daniel J. Bernstein hash which is also used in the ELF GNU_HASH sections. It is a very good hash for all kinds of names in programs with very few hash collisions.

Empty buckets are designated by using an invalid hash index of UINT32_MAX.


These name hash tables are designed to be generic where specializations of the table get to define additional data that goes into the header (“HeaderData”), how the string value is stored (“KeyType”) and the content of the data for each hash value.

Header Layout

The header has a fixed part, and the specialized part. The exact format of the header is:

struct Header
  uint32_t   magic;           // 'HASH' magic value to allow endian detection
  uint16_t   version;         // Version number
  uint16_t   hash_function;   // The hash function enumeration that was used
  uint32_t   bucket_count;    // The number of buckets in this hash table
  uint32_t   hashes_count;    // The total number of unique hash values and hash data offsets in this table
  uint32_t   header_data_len; // The bytes to skip to get to the hash indexes (buckets) for correct alignment
                              // Specifically the length of the following HeaderData field - this does not
                              // include the size of the preceding fields
  HeaderData header_data;     // Implementation specific header data

The header starts with a 32 bit “magic” value which must be 'HASH' encoded as an ASCII integer. This allows the detection of the start of the hash table and also allows the table’s byte order to be determined so the table can be correctly extracted. The “magic” value is followed by a 16 bit version number which allows the table to be revised and modified in the future. The current version number is 1. hash_function is a uint16_t enumeration that specifies which hash function was used to produce this table. The current values for the hash function enumerations include:

enum HashFunctionType
  eHashFunctionDJB = 0u, // Daniel J Bernstein hash function

bucket_count is a 32 bit unsigned integer that represents how many buckets are in the BUCKETS array. hashes_count is the number of unique 32 bit hash values that are in the HASHES array, and is the same number of offsets are contained in the OFFSETS array. header_data_len specifies the size in bytes of the HeaderData that is filled in by specialized versions of this table.

Fixed Lookup

The header is followed by the buckets, hashes, offsets, and hash value data.

struct FixedTable
  uint32_t buckets[Header.bucket_count];  // An array of hash indexes into the "hashes[]" array below
  uint32_t hashes [Header.hashes_count];  // Every unique 32 bit hash for the entire table is in this table
  uint32_t offsets[Header.hashes_count];  // An offset that corresponds to each item in the "hashes[]" array above

buckets is an array of 32 bit indexes into the hashes array. The hashes array contains all of the 32 bit hash values for all names in the hash table. Each hash in the hashes table has an offset in the offsets array that points to the data for the hash value.

This table setup makes it very easy to repurpose these tables to contain different data, while keeping the lookup mechanism the same for all tables. This layout also makes it possible to save the table to disk and map it in later and do very efficient name lookups with little or no parsing.

DWARF lookup tables can be implemented in a variety of ways and can store a lot of information for each name. We want to make the DWARF tables extensible and able to store the data efficiently so we have used some of the DWARF features that enable efficient data storage to define exactly what kind of data we store for each name.

The HeaderData contains a definition of the contents of each HashData chunk. We might want to store an offset to all of the debug information entries (DIEs) for each name. To keep things extensible, we create a list of items, or Atoms, that are contained in the data for each name. First comes the type of the data in each atom:

enum AtomType
  eAtomTypeNULL       = 0u,
  eAtomTypeDIEOffset  = 1u,   // DIE offset, check form for encoding
  eAtomTypeCUOffset   = 2u,   // DIE offset of the compiler unit header that contains the item in question
  eAtomTypeTag        = 3u,   // DW_TAG_xxx value, should be encoded as DW_FORM_data1 (if no tags exceed 255) or DW_FORM_data2
  eAtomTypeNameFlags  = 4u,   // Flags from enum NameFlags
  eAtomTypeTypeFlags  = 5u,   // Flags from enum TypeFlags

The enumeration values and their meanings are:

eAtomTypeNULL       - a termination atom that specifies the end of the atom list
eAtomTypeDIEOffset  - an offset into the .debug_info section for the DWARF DIE for this name
eAtomTypeCUOffset   - an offset into the .debug_info section for the CU that contains the DIE
eAtomTypeDIETag     - The DW_TAG_XXX enumeration value so you don't have to parse the DWARF to see what it is
eAtomTypeNameFlags  - Flags for functions and global variables (isFunction, isInlined, isExternal...)
eAtomTypeTypeFlags  - Flags for types (isCXXClass, isObjCClass, ...)

Then we allow each atom type to define the atom type and how the data for each atom type data is encoded:

struct Atom
  uint16_t type;  // AtomType enum value
  uint16_t form;  // DWARF DW_FORM_XXX defines

The form type above is from the DWARF specification and defines the exact encoding of the data for the Atom type. See the DWARF specification for the DW_FORM_ definitions.

struct HeaderData
  uint32_t die_offset_base;
  uint32_t atom_count;
  Atoms    atoms[atom_count0];

HeaderData defines the base DIE offset that should be added to any atoms that are encoded using the DW_FORM_ref1, DW_FORM_ref2, DW_FORM_ref4, DW_FORM_ref8 or DW_FORM_ref_udata. It also defines what is contained in each HashData object – Atom.form tells us how large each field will be in the HashData and the Atom.type tells us how this data should be interpreted.

For the current implementations of the “.apple_names” (all functions + globals), the “.apple_types” (names of all types that are defined), and the “.apple_namespaces” (all namespaces), we currently set the Atom array to be:

HeaderData.atom_count = 1;
HeaderData.atoms[0].type = eAtomTypeDIEOffset;
HeaderData.atoms[0].form = DW_FORM_data4;

This defines the contents to be the DIE offset (eAtomTypeDIEOffset) that is encoded as a 32 bit value (DW_FORM_data4). This allows a single name to have multiple matching DIEs in a single file, which could come up with an inlined function for instance. Future tables could include more information about the DIE such as flags indicating if the DIE is a function, method, block, or inlined.

The KeyType for the DWARF table is a 32 bit string table offset into the ”.debug_str” table. The ”.debug_str” is the string table for the DWARF which may already contain copies of all of the strings. This helps make sure, with help from the compiler, that we reuse the strings between all of the DWARF sections and keeps the hash table size down. Another benefit to having the compiler generate all strings as DW_FORM_strp in the debug info, is that DWARF parsing can be made much faster.

After a lookup is made, we get an offset into the hash data. The hash data needs to be able to deal with 32 bit hash collisions, so the chunk of data at the offset in the hash data consists of a triple:

uint32_t str_offset
uint32_t hash_data_count

If “str_offset” is zero, then the bucket contents are done. 99.9% of the hash data chunks contain a single item (no 32 bit hash collision):

| 0x00001023 | uint32_t KeyType (.debug_str[0x0001023] => "main")
| 0x00000004 | uint32_t HashData count
| 0x........ | uint32_t HashData[0] DIE offset
| 0x........ | uint32_t HashData[1] DIE offset
| 0x........ | uint32_t HashData[2] DIE offset
| 0x........ | uint32_t HashData[3] DIE offset
| 0x00000000 | uint32_t KeyType (end of hash chain)

If there are collisions, you will have multiple valid string offsets:

| 0x00001023 | uint32_t KeyType (.debug_str[0x0001023] => "main")
| 0x00000004 | uint32_t HashData count
| 0x........ | uint32_t HashData[0] DIE offset
| 0x........ | uint32_t HashData[1] DIE offset
| 0x........ | uint32_t HashData[2] DIE offset
| 0x........ | uint32_t HashData[3] DIE offset
| 0x00002023 | uint32_t KeyType (.debug_str[0x0002023] => "print")
| 0x00000002 | uint32_t HashData count
| 0x........ | uint32_t HashData[0] DIE offset
| 0x........ | uint32_t HashData[1] DIE offset
| 0x00000000 | uint32_t KeyType (end of hash chain)

Current testing with real world C++ binaries has shown that there is around 1 32 bit hash collision per 100,000 name entries.


As we said, we want to strictly define exactly what is included in the different tables. For DWARF, we have 3 tables: “.apple_names”, “.apple_types”, and “.apple_namespaces”.

.apple_names” sections should contain an entry for each DWARF DIE whose DW_TAG is a DW_TAG_label, DW_TAG_inlined_subroutine, or DW_TAG_subprogram that has address attributes: DW_AT_low_pc, DW_AT_high_pc, DW_AT_ranges or DW_AT_entry_pc. It also contains DW_TAG_variable DIEs that have a DW_OP_addr in the location (global and static variables). All global and static variables should be included, including those scoped within functions and classes. For example using the following code:

static int var = 0;

void f ()
  static int var = 0;

Both of the static var variables would be included in the table. All functions should emit both their full names and their basenames. For C or C++, the full name is the mangled name (if available) which is usually in the DW_AT_MIPS_linkage_name attribute, and the DW_AT_name contains the function basename. If global or static variables have a mangled name in a DW_AT_MIPS_linkage_name attribute, this should be emitted along with the simple name found in the DW_AT_name attribute.

.apple_types” sections should contain an entry for each DWARF DIE whose tag is one of:

  • DW_TAG_array_type
  • DW_TAG_class_type
  • DW_TAG_enumeration_type
  • DW_TAG_pointer_type
  • DW_TAG_reference_type
  • DW_TAG_string_type
  • DW_TAG_structure_type
  • DW_TAG_subroutine_type
  • DW_TAG_typedef
  • DW_TAG_union_type
  • DW_TAG_ptr_to_member_type
  • DW_TAG_set_type
  • DW_TAG_subrange_type
  • DW_TAG_base_type
  • DW_TAG_const_type
  • DW_TAG_file_type
  • DW_TAG_namelist
  • DW_TAG_packed_type
  • DW_TAG_volatile_type
  • DW_TAG_restrict_type
  • DW_TAG_interface_type
  • DW_TAG_unspecified_type
  • DW_TAG_shared_type

Only entries with a DW_AT_name attribute are included, and the entry must not be a forward declaration (DW_AT_declaration attribute with a non-zero value). For example, using the following code:

int main ()
  int *b = 0;
  return *b;

We get a few type DIEs:

0x00000067:     TAG_base_type [5]
                AT_encoding( DW_ATE_signed )
                AT_name( "int" )
                AT_byte_size( 0x04 )

0x0000006e:     TAG_pointer_type [6]
                AT_type( {0x00000067} ( int ) )
                AT_byte_size( 0x08 )

The DW_TAG_pointer_type is not included because it does not have a DW_AT_name.

.apple_namespaces” section should contain all DW_TAG_namespace DIEs. If we run into a namespace that has no name this is an anonymous namespace, and the name should be output as “(anonymous namespace)” (without the quotes). Why? This matches the output of the abi::cxa_demangle() that is in the standard C++ library that demangles mangled names.

Language Extensions and File Format Changes

Objective-C Extensions

.apple_objc” section should contain all DW_TAG_subprogram DIEs for an Objective-C class. The name used in the hash table is the name of the Objective-C class itself. If the Objective-C class has a category, then an entry is made for both the class name without the category, and for the class name with the category. So if we have a DIE at offset 0x1234 with a name of method “-[NSString(my_additions) stringWithSpecialString:]”, we would add an entry for “NSString” that points to DIE 0x1234, and an entry for “NSString(my_additions)” that points to 0x1234. This allows us to quickly track down all Objective-C methods for an Objective-C class when doing expressions. It is needed because of the dynamic nature of Objective-C where anyone can add methods to a class. The DWARF for Objective-C methods is also emitted differently from C++ classes where the methods are not usually contained in the class definition, they are scattered about across one or more compile units. Categories can also be defined in different shared libraries. So we need to be able to quickly find all of the methods and class functions given the Objective-C class name, or quickly find all methods and class functions for a class + category name. This table does not contain any selector names, it just maps Objective-C class names (or class names + category) to all of the methods and class functions. The selectors are added as function basenames in the “.debug_names” section.

In the “.apple_names” section for Objective-C functions, the full name is the entire function name with the brackets (“-[NSString stringWithCString:]”) and the basename is the selector only (“stringWithCString:”).

Mach-O Changes

The sections names for the apple hash tables are for non-mach-o files. For mach-o files, the sections should be contained in the __DWARF segment with names as follows:

  • .apple_names” -> “__apple_names
  • .apple_types” -> “__apple_types
  • .apple_namespaces” -> “__apple_namespac” (16 character limit)
  • .apple_objc” -> “__apple_objc

CodeView Debug Info Format

LLVM supports emitting CodeView, the Microsoft debug info format, and this section describes the design and implementation of that support.

Format Background

CodeView as a format is clearly oriented around C++ debugging, and in C++, the majority of debug information tends to be type information. Therefore, the overriding design constraint of CodeView is the separation of type information from other “symbol” information so that type information can be efficiently merged across translation units. Both type information and symbol information is generally stored as a sequence of records, where each record begins with a 16-bit record size and a 16-bit record kind.

Type information is usually stored in the .debug$T section of the object file. All other debug info, such as line info, string table, symbol info, and inlinee info, is stored in one or more .debug$S sections. There may only be one .debug$T section per object file, since all other debug info refers to it. If a PDB (enabled by the /Zi MSVC option) was used during compilation, the .debug$T section will contain only an LF_TYPESERVER2 record pointing to the PDB. When using PDBs, symbol information appears to remain in the object file .debug$S sections.

Type records are referred to by their index, which is the number of records in the stream before a given record plus 0x1000. Many common basic types, such as the basic integral types and unqualified pointers to them, are represented using type indices less than 0x1000. Such basic types are built in to CodeView consumers and do not require type records.

Each type record may only contain type indices that are less than its own type index. This ensures that the graph of type stream references is acyclic. While the source-level type graph may contain cycles through pointer types (consider a linked list struct), these cycles are removed from the type stream by always referring to the forward declaration record of user-defined record types. Only “symbol” records in the .debug$S streams may refer to complete, non-forward-declaration type records.

Working with CodeView

These are instructions for some common tasks for developers working to improve LLVM’s CodeView support. Most of them revolve around using the CodeView dumper embedded in llvm-readobj.

  • Testing MSVC’s output:

    $ cl -c -Z7 foo.cpp # Use /Z7 to keep types in the object file
    $ llvm-readobj -codeview foo.obj
  • Getting LLVM IR debug info out of Clang:

    $ clang -g -gcodeview --target=x86_64-windows-msvc foo.cpp -S -emit-llvm

    Use this to generate LLVM IR for LLVM test cases.

  • Generate and dump CodeView from LLVM IR metadata:

    $ llc foo.ll -filetype=obj -o foo.obj
    $ llvm-readobj -codeview foo.obj > foo.txt

    Use this pattern in lit test cases and FileCheck the output of llvm-readobj

Improving LLVM’s CodeView support is a process of finding interesting type records, constructing a C++ test case that makes MSVC emit those records, dumping the records, understanding them, and then generating equivalent records in LLVM’s backend.