1 ================================
2 Source Level Debugging with LLVM
3 ================================
11 This document is the central repository for all information pertaining to debug
12 information in LLVM. It describes the :ref:`actual format that the LLVM debug
13 information takes <format>`, which is useful for those interested in creating
14 front-ends or dealing directly with the information. Further, this document
15 provides specific examples of what debug information for C/C++ looks like.
17 Philosophy behind LLVM debugging information
18 --------------------------------------------
20 The idea of the LLVM debugging information is to capture how the important
21 pieces of the source-language's Abstract Syntax Tree map onto LLVM code.
22 Several design aspects have shaped the solution that appears here. The
25 * Debugging information should have very little impact on the rest of the
26 compiler. No transformations, analyses, or code generators should need to
27 be modified because of debugging information.
29 * LLVM optimizations should interact in :ref:`well-defined and easily described
30 ways <intro_debugopt>` with the debugging information.
32 * Because LLVM is designed to support arbitrary programming languages,
33 LLVM-to-LLVM tools should not need to know anything about the semantics of
34 the source-level-language.
36 * Source-level languages are often **widely** different from one another.
37 LLVM should not put any restrictions of the flavor of the source-language,
38 and the debugging information should work with any language.
40 * With code generator support, it should be possible to use an LLVM compiler
41 to compile a program to native machine code and standard debugging
42 formats. This allows compatibility with traditional machine-code level
43 debuggers, like GDB or DBX.
45 The approach used by the LLVM implementation is to use a small set of
46 :ref:`intrinsic functions <format_common_intrinsics>` to define a mapping
47 between LLVM program objects and the source-level objects. The description of
48 the source-level program is maintained in LLVM metadata in an
49 :ref:`implementation-defined format <ccxx_frontend>` (the C/C++ front-end
50 currently uses working draft 7 of the `DWARF 3 standard
51 <http://www.eagercon.com/dwarf/dwarf3std.htm>`_).
53 When a program is being debugged, a debugger interacts with the user and turns
54 the stored debug information into source-language specific information. As
55 such, a debugger must be aware of the source-language, and is thus tied to a
56 specific language or family of languages.
58 Debug information consumers
59 ---------------------------
61 The role of debug information is to provide meta information normally stripped
62 away during the compilation process. This meta information provides an LLVM
63 user a relationship between generated code and the original program source
66 Currently, debug information is consumed by DwarfDebug to produce dwarf
67 information used by the gdb debugger. Other targets could use the same
68 information to produce stabs or other debug forms.
70 It would also be reasonable to use debug information to feed profiling tools
71 for analysis of generated code, or, tools for reconstructing the original
72 source from generated code.
74 TODO - expound a bit more.
78 Debugging optimized code
79 ------------------------
81 An extremely high priority of LLVM debugging information is to make it interact
82 well with optimizations and analysis. In particular, the LLVM debug
83 information provides the following guarantees:
85 * LLVM debug information **always provides information to accurately read
86 the source-level state of the program**, regardless of which LLVM
87 optimizations have been run, and without any modification to the
88 optimizations themselves. However, some optimizations may impact the
89 ability to modify the current state of the program with a debugger, such
90 as setting program variables, or calling functions that have been
93 * As desired, LLVM optimizations can be upgraded to be aware of the LLVM
94 debugging information, allowing them to update the debugging information
95 as they perform aggressive optimizations. This means that, with effort,
96 the LLVM optimizers could optimize debug code just as well as non-debug
99 * LLVM debug information does not prevent optimizations from
100 happening (for example inlining, basic block reordering/merging/cleanup,
101 tail duplication, etc).
103 * LLVM debug information is automatically optimized along with the rest of
104 the program, using existing facilities. For example, duplicate
105 information is automatically merged by the linker, and unused information
106 is automatically removed.
108 Basically, the debug information allows you to compile a program with
109 "``-O0 -g``" and get full debug information, allowing you to arbitrarily modify
110 the program as it executes from a debugger. Compiling a program with
111 "``-O3 -g``" gives you full debug information that is always available and
112 accurate for reading (e.g., you get accurate stack traces despite tail call
113 elimination and inlining), but you might lose the ability to modify the program
114 and call functions where were optimized out of the program, or inlined away
117 :ref:`LLVM test suite <test-suite-quickstart>` provides a framework to test
118 optimizer's handling of debugging information. It can be run like this:
122 % cd llvm/projects/test-suite/MultiSource/Benchmarks # or some other level
125 This will test impact of debugging information on optimization passes. If
126 debugging information influences optimization passes then it will be reported
127 as a failure. See :doc:`TestingGuide` for more information on LLVM test
128 infrastructure and how to run various tests.
132 Debugging information format
133 ============================
135 LLVM debugging information has been carefully designed to make it possible for
136 the optimizer to optimize the program and debugging information without
137 necessarily having to know anything about debugging information. In
138 particular, the use of metadata avoids duplicated debugging information from
139 the beginning, and the global dead code elimination pass automatically deletes
140 debugging information for a function if it decides to delete the function.
142 To do this, most of the debugging information (descriptors for types,
143 variables, functions, source files, etc) is inserted by the language front-end
144 in the form of LLVM metadata.
146 Debug information is designed to be agnostic about the target debugger and
147 debugging information representation (e.g. DWARF/Stabs/etc). It uses a generic
148 pass to decode the information that represents variables, types, functions,
149 namespaces, etc: this allows for arbitrary source-language semantics and
150 type-systems to be used, as long as there is a module written for the target
151 debugger to interpret the information.
153 To provide basic functionality, the LLVM debugger does have to make some
154 assumptions about the source-level language being debugged, though it keeps
155 these to a minimum. The only common features that the LLVM debugger assumes
156 exist are `source files <LangRef.html#difile>`_, and `program objects
157 <LangRef.html#diglobalvariable>`_. These abstract objects are used by a
158 debugger to form stack traces, show information about local variables, etc.
160 This section of the documentation first describes the representation aspects
161 common to any source-language. :ref:`ccxx_frontend` describes the data layout
162 conventions used by the C and C++ front-ends.
164 Debug information descriptors are `specialized metadata nodes
165 <LangRef.html#specialized-metadata>`_, first-class subclasses of ``Metadata``.
167 .. _format_common_intrinsics:
169 Debugger intrinsic functions
170 ----------------------------
172 LLVM uses several intrinsic functions (name prefixed with "``llvm.dbg``") to
173 provide debug information at various points in generated code.
180 void @llvm.dbg.declare(metadata, metadata, metadata)
182 This intrinsic provides information about a local element (e.g., variable).
183 The first argument is metadata holding the alloca for the variable. The second
184 argument is a `local variable <LangRef.html#dilocalvariable>`_ containing a
185 description of the variable. The third argument is a `complex expression
186 <LangRef.html#diexpression>`_.
193 void @llvm.dbg.value(metadata, i64, metadata, metadata)
195 This intrinsic provides information when a user source variable is set to a new
196 value. The first argument is the new value (wrapped as metadata). The second
197 argument is the offset in the user source variable where the new value is
198 written. The third argument is a `local variable
199 <LangRef.html#dilocalvariable>`_ containing a description of the variable. The
200 third argument is a `complex expression <LangRef.html#diexpression>`_.
202 Object lifetimes and scoping
203 ============================
205 In many languages, the local variables in functions can have their lifetimes or
206 scopes limited to a subset of a function. In the C family of languages, for
207 example, variables are only live (readable and writable) within the source
208 block that they are defined in. In functional languages, values are only
209 readable after they have been defined. Though this is a very obvious concept,
210 it is non-trivial to model in LLVM, because it has no notion of scoping in this
211 sense, and does not want to be tied to a language's scoping rules.
213 In order to handle this, the LLVM debug format uses the metadata attached to
214 llvm instructions to encode line number and scoping information. Consider the
215 following C fragment, for example:
229 Compiled to LLVM, this function would be represented like this:
233 ; Function Attrs: nounwind ssp uwtable
234 define void @foo() #0 !dbg !4 {
236 %X = alloca i32, align 4
237 %Y = alloca i32, align 4
238 %Z = alloca i32, align 4
239 call void @llvm.dbg.declare(metadata i32* %X, metadata !11, metadata !13), !dbg !14
240 store i32 21, i32* %X, align 4, !dbg !14
241 call void @llvm.dbg.declare(metadata i32* %Y, metadata !15, metadata !13), !dbg !16
242 store i32 22, i32* %Y, align 4, !dbg !16
243 call void @llvm.dbg.declare(metadata i32* %Z, metadata !17, metadata !13), !dbg !19
244 store i32 23, i32* %Z, align 4, !dbg !19
245 %0 = load i32, i32* %X, align 4, !dbg !20
246 store i32 %0, i32* %Z, align 4, !dbg !21
247 %1 = load i32, i32* %Y, align 4, !dbg !22
248 store i32 %1, i32* %X, align 4, !dbg !23
252 ; Function Attrs: nounwind readnone
253 declare void @llvm.dbg.declare(metadata, metadata, metadata) #1
255 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" }
256 attributes #1 = { nounwind readnone }
259 !llvm.module.flags = !{!7, !8, !9}
262 !0 = !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang version 3.7.0 (trunk 231150) (llvm/trunk 231154)", isOptimized: false, runtimeVersion: 0, emissionKind: 1, enums: !2, retainedTypes: !2, subprograms: !3, globals: !2, imports: !2)
263 !1 = !DIFile(filename: "/dev/stdin", directory: "/Users/dexonsmith/data/llvm/debug-info")
266 !4 = distinct !DISubprogram(name: "foo", scope: !1, file: !1, line: 1, type: !5, isLocal: false, isDefinition: true, scopeLine: 1, isOptimized: false, variables: !2)
267 !5 = !DISubroutineType(types: !6)
269 !7 = !{i32 2, !"Dwarf Version", i32 2}
270 !8 = !{i32 2, !"Debug Info Version", i32 3}
271 !9 = !{i32 1, !"PIC Level", i32 2}
272 !10 = !{!"clang version 3.7.0 (trunk 231150) (llvm/trunk 231154)"}
273 !11 = !DILocalVariable(name: "X", scope: !4, file: !1, line: 2, type: !12)
274 !12 = !DIBasicType(name: "int", size: 32, align: 32, encoding: DW_ATE_signed)
275 !13 = !DIExpression()
276 !14 = !DILocation(line: 2, column: 9, scope: !4)
277 !15 = !DILocalVariable(name: "Y", scope: !4, file: !1, line: 3, type: !12)
278 !16 = !DILocation(line: 3, column: 9, scope: !4)
279 !17 = !DILocalVariable(name: "Z", scope: !18, file: !1, line: 5, type: !12)
280 !18 = distinct !DILexicalBlock(scope: !4, file: !1, line: 4, column: 5)
281 !19 = !DILocation(line: 5, column: 11, scope: !18)
282 !20 = !DILocation(line: 6, column: 11, scope: !18)
283 !21 = !DILocation(line: 6, column: 9, scope: !18)
284 !22 = !DILocation(line: 8, column: 9, scope: !4)
285 !23 = !DILocation(line: 8, column: 7, scope: !4)
286 !24 = !DILocation(line: 9, column: 3, scope: !4)
289 This example illustrates a few important details about LLVM debugging
290 information. In particular, it shows how the ``llvm.dbg.declare`` intrinsic and
291 location information, which are attached to an instruction, are applied
292 together to allow a debugger to analyze the relationship between statements,
293 variable definitions, and the code used to implement the function.
297 call void @llvm.dbg.declare(metadata i32* %X, metadata !11, metadata !13), !dbg !14
298 ; [debug line = 2:7] [debug variable = X]
300 The first intrinsic ``%llvm.dbg.declare`` encodes debugging information for the
301 variable ``X``. The metadata ``!dbg !14`` attached to the intrinsic provides
302 scope information for the variable ``X``.
306 !14 = !DILocation(line: 2, column: 9, scope: !4)
307 !4 = distinct !DISubprogram(name: "foo", scope: !1, file: !1, line: 1, type: !5,
308 isLocal: false, isDefinition: true, scopeLine: 1,
309 isOptimized: false, variables: !2)
311 Here ``!14`` is metadata providing `location information
312 <LangRef.html#dilocation>`_. In this example, scope is encoded by ``!4``, a
313 `subprogram descriptor <LangRef.html#disubprogram>`_. This way the location
314 information attached to the intrinsics indicates that the variable ``X`` is
315 declared at line number 2 at a function level scope in function ``foo``.
317 Now lets take another example.
321 call void @llvm.dbg.declare(metadata i32* %Z, metadata !17, metadata !13), !dbg !19
322 ; [debug line = 5:9] [debug variable = Z]
324 The third intrinsic ``%llvm.dbg.declare`` encodes debugging information for
325 variable ``Z``. The metadata ``!dbg !19`` attached to the intrinsic provides
326 scope information for the variable ``Z``.
330 !18 = distinct !DILexicalBlock(scope: !4, file: !1, line: 4, column: 5)
331 !19 = !DILocation(line: 5, column: 11, scope: !18)
333 Here ``!19`` indicates that ``Z`` is declared at line number 5 and column
334 number 0 inside of lexical scope ``!18``. The lexical scope itself resides
335 inside of subprogram ``!4`` described above.
337 The scope information attached with each instruction provides a straightforward
338 way to find instructions covered by a scope.
342 C/C++ front-end specific debug information
343 ==========================================
345 The C and C++ front-ends represent information about the program in a format
346 that is effectively identical to `DWARF 3.0
347 <http://www.eagercon.com/dwarf/dwarf3std.htm>`_ in terms of information
348 content. This allows code generators to trivially support native debuggers by
349 generating standard dwarf information, and contains enough information for
350 non-dwarf targets to translate it as needed.
352 This section describes the forms used to represent C and C++ programs. Other
353 languages could pattern themselves after this (which itself is tuned to
354 representing programs in the same way that DWARF 3 does), or they could choose
355 to provide completely different forms if they don't fit into the DWARF model.
356 As support for debugging information gets added to the various LLVM
357 source-language front-ends, the information used should be documented here.
359 The following sections provide examples of a few C/C++ constructs and the debug
360 information that would best describe those constructs. The canonical
361 references are the ``DIDescriptor`` classes defined in
362 ``include/llvm/IR/DebugInfo.h`` and the implementations of the helper functions
363 in ``lib/IR/DIBuilder.cpp``.
365 C/C++ source file information
366 -----------------------------
368 ``llvm::Instruction`` provides easy access to metadata attached with an
369 instruction. One can extract line number information encoded in LLVM IR using
370 ``Instruction::getDebugLoc()`` and ``DILocation::getLine()``.
374 if (DILocation *Loc = I->getDebugLoc()) { // Here I is an LLVM instruction
375 unsigned Line = Loc->getLine();
376 StringRef File = Loc->getFilename();
377 StringRef Dir = Loc->getDirectory();
380 C/C++ global variable information
381 ---------------------------------
383 Given an integer global variable declared as follows:
389 a C/C++ front-end would generate the following descriptors:
394 ;; Define the global itself.
396 @MyGlobal = global i32 100, align 4
399 ;; List of debug info of globals
403 ;; Some unrelated metadata.
404 !llvm.module.flags = !{!6, !7}
406 ;; Define the compile unit.
407 !0 = !DICompileUnit(language: DW_LANG_C99, file: !1,
409 "clang version 3.7.0 (trunk 231150) (llvm/trunk 231154)",
410 isOptimized: false, runtimeVersion: 0, emissionKind: 1,
411 enums: !2, retainedTypes: !2, subprograms: !2, globals:
417 !1 = !DIFile(filename: "/dev/stdin",
418 directory: "/Users/dexonsmith/data/llvm/debug-info")
423 ;; The Array of Global Variables
427 ;; Define the global variable itself.
429 !4 = !DIGlobalVariable(name: "MyGlobal", scope: !0, file: !1, line: 1,
430 type: !5, isLocal: false, isDefinition: true,
431 variable: i32* @MyGlobal)
436 !5 = !DIBasicType(name: "int", size: 32, align: 32, encoding: DW_ATE_signed)
438 ;; Dwarf version to output.
439 !6 = !{i32 2, !"Dwarf Version", i32 2}
441 ;; Debug info schema version.
442 !7 = !{i32 2, !"Debug Info Version", i32 3}
444 C/C++ function information
445 --------------------------
447 Given a function declared as follows:
451 int main(int argc, char *argv[]) {
455 a C/C++ front-end would generate the following descriptors:
460 ;; Define the anchor for subprograms.
462 !4 = !DISubprogram(name: "main", scope: !1, file: !1, line: 1, type: !5,
463 isLocal: false, isDefinition: true, scopeLine: 1,
464 flags: DIFlagPrototyped, isOptimized: false,
468 ;; Define the subprogram itself.
470 define i32 @main(i32 %argc, i8** %argv) !dbg !4 {
474 Debugging information format
475 ============================
477 Debugging Information Extension for Objective C Properties
478 ----------------------------------------------------------
483 Objective C provides a simpler way to declare and define accessor methods using
484 declared properties. The language provides features to declare a property and
485 to let compiler synthesize accessor methods.
487 The debugger lets developer inspect Objective C interfaces and their instance
488 variables and class variables. However, the debugger does not know anything
489 about the properties defined in Objective C interfaces. The debugger consumes
490 information generated by compiler in DWARF format. The format does not support
491 encoding of Objective C properties. This proposal describes DWARF extensions to
492 encode Objective C properties, which the debugger can use to let developers
493 inspect Objective C properties.
498 Objective C properties exist separately from class members. A property can be
499 defined only by "setter" and "getter" selectors, and be calculated anew on each
500 access. Or a property can just be a direct access to some declared ivar.
501 Finally it can have an ivar "automatically synthesized" for it by the compiler,
502 in which case the property can be referred to in user code directly using the
503 standard C dereference syntax as well as through the property "dot" syntax, but
504 there is no entry in the ``@interface`` declaration corresponding to this ivar.
506 To facilitate debugging, these properties we will add a new DWARF TAG into the
507 ``DW_TAG_structure_type`` definition for the class to hold the description of a
508 given property, and a set of DWARF attributes that provide said description.
509 The property tag will also contain the name and declared type of the property.
511 If there is a related ivar, there will also be a DWARF property attribute placed
512 in the ``DW_TAG_member`` DIE for that ivar referring back to the property TAG
513 for that property. And in the case where the compiler synthesizes the ivar
514 directly, the compiler is expected to generate a ``DW_TAG_member`` for that
515 ivar (with the ``DW_AT_artificial`` set to 1), whose name will be the name used
516 to access this ivar directly in code, and with the property attribute pointing
517 back to the property it is backing.
519 The following examples will serve as illustration for our discussion:
536 This produces the following DWARF (this is a "pseudo dwarfdump" output):
540 0x00000100: TAG_structure_type [7] *
541 AT_APPLE_runtime_class( 0x10 )
543 AT_decl_file( "Objc_Property.m" )
546 0x00000110 TAG_APPLE_property
548 AT_type ( {0x00000150} ( int ) )
550 0x00000120: TAG_APPLE_property
552 AT_type ( {0x00000150} ( int ) )
554 0x00000130: TAG_member [8]
556 AT_APPLE_property ( {0x00000110} "p1" )
557 AT_type( {0x00000150} ( int ) )
558 AT_artificial ( 0x1 )
560 0x00000140: TAG_member [8]
562 AT_APPLE_property ( {0x00000120} "p2" )
563 AT_type( {0x00000150} ( int ) )
565 0x00000150: AT_type( ( int ) )
567 Note, the current convention is that the name of the ivar for an
568 auto-synthesized property is the name of the property from which it derives
569 with an underscore prepended, as is shown in the example. But we actually
570 don't need to know this convention, since we are given the name of the ivar
573 Also, it is common practice in ObjC to have different property declarations in
574 the @interface and @implementation - e.g. to provide a read-only property in
575 the interface,and a read-write interface in the implementation. In that case,
576 the compiler should emit whichever property declaration will be in force in the
577 current translation unit.
579 Developers can decorate a property with attributes which are encoded using
580 ``DW_AT_APPLE_property_attribute``.
584 @property (readonly, nonatomic) int pr;
588 TAG_APPLE_property [8]
590 AT_type ( {0x00000147} (int) )
591 AT_APPLE_property_attribute (DW_APPLE_PROPERTY_readonly, DW_APPLE_PROPERTY_nonatomic)
593 The setter and getter method names are attached to the property using
594 ``DW_AT_APPLE_property_setter`` and ``DW_AT_APPLE_property_getter`` attributes.
599 @property (setter=myOwnP3Setter:) int p3;
600 -(void)myOwnP3Setter:(int)a;
605 -(void)myOwnP3Setter:(int)a{ }
608 The DWARF for this would be:
612 0x000003bd: TAG_structure_type [7] *
613 AT_APPLE_runtime_class( 0x10 )
615 AT_decl_file( "Objc_Property.m" )
618 0x000003cd TAG_APPLE_property
620 AT_APPLE_property_setter ( "myOwnP3Setter:" )
621 AT_type( {0x00000147} ( int ) )
623 0x000003f3: TAG_member [8]
625 AT_type ( {0x00000147} ( int ) )
626 AT_APPLE_property ( {0x000003cd} )
627 AT_artificial ( 0x1 )
632 +-----------------------+--------+
634 +=======================+========+
635 | DW_TAG_APPLE_property | 0x4200 |
636 +-----------------------+--------+
641 +--------------------------------+--------+-----------+
642 | Attribute | Value | Classes |
643 +================================+========+===========+
644 | DW_AT_APPLE_property | 0x3fed | Reference |
645 +--------------------------------+--------+-----------+
646 | DW_AT_APPLE_property_getter | 0x3fe9 | String |
647 +--------------------------------+--------+-----------+
648 | DW_AT_APPLE_property_setter | 0x3fea | String |
649 +--------------------------------+--------+-----------+
650 | DW_AT_APPLE_property_attribute | 0x3feb | Constant |
651 +--------------------------------+--------+-----------+
656 +--------------------------------------+-------+
658 +======================================+=======+
659 | DW_APPLE_PROPERTY_readonly | 0x01 |
660 +--------------------------------------+-------+
661 | DW_APPLE_PROPERTY_getter | 0x02 |
662 +--------------------------------------+-------+
663 | DW_APPLE_PROPERTY_assign | 0x04 |
664 +--------------------------------------+-------+
665 | DW_APPLE_PROPERTY_readwrite | 0x08 |
666 +--------------------------------------+-------+
667 | DW_APPLE_PROPERTY_retain | 0x10 |
668 +--------------------------------------+-------+
669 | DW_APPLE_PROPERTY_copy | 0x20 |
670 +--------------------------------------+-------+
671 | DW_APPLE_PROPERTY_nonatomic | 0x40 |
672 +--------------------------------------+-------+
673 | DW_APPLE_PROPERTY_setter | 0x80 |
674 +--------------------------------------+-------+
675 | DW_APPLE_PROPERTY_atomic | 0x100 |
676 +--------------------------------------+-------+
677 | DW_APPLE_PROPERTY_weak | 0x200 |
678 +--------------------------------------+-------+
679 | DW_APPLE_PROPERTY_strong | 0x400 |
680 +--------------------------------------+-------+
681 | DW_APPLE_PROPERTY_unsafe_unretained | 0x800 |
682 +--------------------------------+-----+-------+
684 Name Accelerator Tables
685 -----------------------
690 The "``.debug_pubnames``" and "``.debug_pubtypes``" formats are not what a
691 debugger needs. The "``pub``" in the section name indicates that the entries
692 in the table are publicly visible names only. This means no static or hidden
693 functions show up in the "``.debug_pubnames``". No static variables or private
694 class variables are in the "``.debug_pubtypes``". Many compilers add different
695 things to these tables, so we can't rely upon the contents between gcc, icc, or
698 The typical query given by users tends not to match up with the contents of
699 these tables. For example, the DWARF spec states that "In the case of the name
700 of a function member or static data member of a C++ structure, class or union,
701 the name presented in the "``.debug_pubnames``" section is not the simple name
702 given by the ``DW_AT_name attribute`` of the referenced debugging information
703 entry, but rather the fully qualified name of the data or function member."
704 So the only names in these tables for complex C++ entries is a fully
705 qualified name. Debugger users tend not to enter their search strings as
706 "``a::b::c(int,const Foo&) const``", but rather as "``c``", "``b::c``" , or
707 "``a::b::c``". So the name entered in the name table must be demangled in
708 order to chop it up appropriately and additional names must be manually entered
709 into the table to make it effective as a name lookup table for debuggers to
712 All debuggers currently ignore the "``.debug_pubnames``" table as a result of
713 its inconsistent and useless public-only name content making it a waste of
714 space in the object file. These tables, when they are written to disk, are not
715 sorted in any way, leaving every debugger to do its own parsing and sorting.
716 These tables also include an inlined copy of the string values in the table
717 itself making the tables much larger than they need to be on disk, especially
718 for large C++ programs.
720 Can't we just fix the sections by adding all of the names we need to this
721 table? No, because that is not what the tables are defined to contain and we
722 won't know the difference between the old bad tables and the new good tables.
723 At best we could make our own renamed sections that contain all of the data we
726 These tables are also insufficient for what a debugger like LLDB needs. LLDB
727 uses clang for its expression parsing where LLDB acts as a PCH. LLDB is then
728 often asked to look for type "``foo``" or namespace "``bar``", or list items in
729 namespace "``baz``". Namespaces are not included in the pubnames or pubtypes
730 tables. Since clang asks a lot of questions when it is parsing an expression,
731 we need to be very fast when looking up names, as it happens a lot. Having new
732 accelerator tables that are optimized for very quick lookups will benefit this
733 type of debugging experience greatly.
735 We would like to generate name lookup tables that can be mapped into memory
736 from disk, and used as is, with little or no up-front parsing. We would also
737 be able to control the exact content of these different tables so they contain
738 exactly what we need. The Name Accelerator Tables were designed to fix these
739 issues. In order to solve these issues we need to:
741 * Have a format that can be mapped into memory from disk and used as is
742 * Lookups should be very fast
743 * Extensible table format so these tables can be made by many producers
744 * Contain all of the names needed for typical lookups out of the box
745 * Strict rules for the contents of tables
747 Table size is important and the accelerator table format should allow the reuse
748 of strings from common string tables so the strings for the names are not
749 duplicated. We also want to make sure the table is ready to be used as-is by
750 simply mapping the table into memory with minimal header parsing.
752 The name lookups need to be fast and optimized for the kinds of lookups that
753 debuggers tend to do. Optimally we would like to touch as few parts of the
754 mapped table as possible when doing a name lookup and be able to quickly find
755 the name entry we are looking for, or discover there are no matches. In the
756 case of debuggers we optimized for lookups that fail most of the time.
758 Each table that is defined should have strict rules on exactly what is in the
759 accelerator tables and documented so clients can rely on the content.
767 Typical hash tables have a header, buckets, and each bucket points to the
780 The BUCKETS are an array of offsets to DATA for each hash:
785 | 0x00001000 | BUCKETS[0]
786 | 0x00002000 | BUCKETS[1]
787 | 0x00002200 | BUCKETS[2]
788 | 0x000034f0 | BUCKETS[3]
790 | 0xXXXXXXXX | BUCKETS[n_buckets]
793 So for ``bucket[3]`` in the example above, we have an offset into the table
794 0x000034f0 which points to a chain of entries for the bucket. Each bucket must
795 contain a next pointer, full 32 bit hash value, the string itself, and the data
796 for the current string value.
801 0x000034f0: | 0x00003500 | next pointer
802 | 0x12345678 | 32 bit hash
803 | "erase" | string value
804 | data[n] | HashData for this bucket
806 0x00003500: | 0x00003550 | next pointer
807 | 0x29273623 | 32 bit hash
808 | "dump" | string value
809 | data[n] | HashData for this bucket
811 0x00003550: | 0x00000000 | next pointer
812 | 0x82638293 | 32 bit hash
813 | "main" | string value
814 | data[n] | HashData for this bucket
817 The problem with this layout for debuggers is that we need to optimize for the
818 negative lookup case where the symbol we're searching for is not present. So
819 if we were to lookup "``printf``" in the table above, we would make a 32 hash
820 for "``printf``", it might match ``bucket[3]``. We would need to go to the
821 offset 0x000034f0 and start looking to see if our 32 bit hash matches. To do
822 so, we need to read the next pointer, then read the hash, compare it, and skip
823 to the next bucket. Each time we are skipping many bytes in memory and
824 touching new cache pages just to do the compare on the full 32 bit hash. All
825 of these accesses then tell us that we didn't have a match.
830 To solve the issues mentioned above we have structured the hash tables a bit
831 differently: a header, buckets, an array of all unique 32 bit hash values,
832 followed by an array of hash value data offsets, one for each hash value, then
833 the data for all hash values:
849 The ``BUCKETS`` in the name tables are an index into the ``HASHES`` array. By
850 making all of the full 32 bit hash values contiguous in memory, we allow
851 ourselves to efficiently check for a match while touching as little memory as
852 possible. Most often checking the 32 bit hash values is as far as the lookup
853 goes. If it does match, it usually is a match with no collisions. So for a
854 table with "``n_buckets``" buckets, and "``n_hashes``" unique 32 bit hash
855 values, we can clarify the contents of the ``BUCKETS``, ``HASHES`` and
860 .-------------------------.
861 | HEADER.magic | uint32_t
862 | HEADER.version | uint16_t
863 | HEADER.hash_function | uint16_t
864 | HEADER.bucket_count | uint32_t
865 | HEADER.hashes_count | uint32_t
866 | HEADER.header_data_len | uint32_t
867 | HEADER_DATA | HeaderData
868 |-------------------------|
869 | BUCKETS | uint32_t[n_buckets] // 32 bit hash indexes
870 |-------------------------|
871 | HASHES | uint32_t[n_hashes] // 32 bit hash values
872 |-------------------------|
873 | OFFSETS | uint32_t[n_hashes] // 32 bit offsets to hash value data
874 |-------------------------|
876 `-------------------------'
878 So taking the exact same data from the standard hash example above we end up
891 | ... | BUCKETS[n_buckets]
893 | 0x........ | HASHES[0]
894 | 0x........ | HASHES[1]
895 | 0x........ | HASHES[2]
896 | 0x........ | HASHES[3]
897 | 0x........ | HASHES[4]
898 | 0x........ | HASHES[5]
899 | 0x12345678 | HASHES[6] hash for BUCKETS[3]
900 | 0x29273623 | HASHES[7] hash for BUCKETS[3]
901 | 0x82638293 | HASHES[8] hash for BUCKETS[3]
902 | 0x........ | HASHES[9]
903 | 0x........ | HASHES[10]
904 | 0x........ | HASHES[11]
905 | 0x........ | HASHES[12]
906 | 0x........ | HASHES[13]
907 | 0x........ | HASHES[n_hashes]
909 | 0x........ | OFFSETS[0]
910 | 0x........ | OFFSETS[1]
911 | 0x........ | OFFSETS[2]
912 | 0x........ | OFFSETS[3]
913 | 0x........ | OFFSETS[4]
914 | 0x........ | OFFSETS[5]
915 | 0x000034f0 | OFFSETS[6] offset for BUCKETS[3]
916 | 0x00003500 | OFFSETS[7] offset for BUCKETS[3]
917 | 0x00003550 | OFFSETS[8] offset for BUCKETS[3]
918 | 0x........ | OFFSETS[9]
919 | 0x........ | OFFSETS[10]
920 | 0x........ | OFFSETS[11]
921 | 0x........ | OFFSETS[12]
922 | 0x........ | OFFSETS[13]
923 | 0x........ | OFFSETS[n_hashes]
931 0x000034f0: | 0x00001203 | .debug_str ("erase")
932 | 0x00000004 | A 32 bit array count - number of HashData with name "erase"
933 | 0x........ | HashData[0]
934 | 0x........ | HashData[1]
935 | 0x........ | HashData[2]
936 | 0x........ | HashData[3]
937 | 0x00000000 | String offset into .debug_str (terminate data for hash)
939 0x00003500: | 0x00001203 | String offset into .debug_str ("collision")
940 | 0x00000002 | A 32 bit array count - number of HashData with name "collision"
941 | 0x........ | HashData[0]
942 | 0x........ | HashData[1]
943 | 0x00001203 | String offset into .debug_str ("dump")
944 | 0x00000003 | A 32 bit array count - number of HashData with name "dump"
945 | 0x........ | HashData[0]
946 | 0x........ | HashData[1]
947 | 0x........ | HashData[2]
948 | 0x00000000 | String offset into .debug_str (terminate data for hash)
950 0x00003550: | 0x00001203 | String offset into .debug_str ("main")
951 | 0x00000009 | A 32 bit array count - number of HashData with name "main"
952 | 0x........ | HashData[0]
953 | 0x........ | HashData[1]
954 | 0x........ | HashData[2]
955 | 0x........ | HashData[3]
956 | 0x........ | HashData[4]
957 | 0x........ | HashData[5]
958 | 0x........ | HashData[6]
959 | 0x........ | HashData[7]
960 | 0x........ | HashData[8]
961 | 0x00000000 | String offset into .debug_str (terminate data for hash)
964 So we still have all of the same data, we just organize it more efficiently for
965 debugger lookup. If we repeat the same "``printf``" lookup from above, we
966 would hash "``printf``" and find it matches ``BUCKETS[3]`` by taking the 32 bit
967 hash value and modulo it by ``n_buckets``. ``BUCKETS[3]`` contains "6" which
968 is the index into the ``HASHES`` table. We would then compare any consecutive
969 32 bit hashes values in the ``HASHES`` array as long as the hashes would be in
970 ``BUCKETS[3]``. We do this by verifying that each subsequent hash value modulo
971 ``n_buckets`` is still 3. In the case of a failed lookup we would access the
972 memory for ``BUCKETS[3]``, and then compare a few consecutive 32 bit hashes
973 before we know that we have no match. We don't end up marching through
974 multiple words of memory and we really keep the number of processor data cache
975 lines being accessed as small as possible.
977 The string hash that is used for these lookup tables is the Daniel J.
978 Bernstein hash which is also used in the ELF ``GNU_HASH`` sections. It is a
979 very good hash for all kinds of names in programs with very few hash
982 Empty buckets are designated by using an invalid hash index of ``UINT32_MAX``.
987 These name hash tables are designed to be generic where specializations of the
988 table get to define additional data that goes into the header ("``HeaderData``"),
989 how the string value is stored ("``KeyType``") and the content of the data for each
995 The header has a fixed part, and the specialized part. The exact format of the
1002 uint32_t magic; // 'HASH' magic value to allow endian detection
1003 uint16_t version; // Version number
1004 uint16_t hash_function; // The hash function enumeration that was used
1005 uint32_t bucket_count; // The number of buckets in this hash table
1006 uint32_t hashes_count; // The total number of unique hash values and hash data offsets in this table
1007 uint32_t header_data_len; // The bytes to skip to get to the hash indexes (buckets) for correct alignment
1008 // Specifically the length of the following HeaderData field - this does not
1009 // include the size of the preceding fields
1010 HeaderData header_data; // Implementation specific header data
1013 The header starts with a 32 bit "``magic``" value which must be ``'HASH'``
1014 encoded as an ASCII integer. This allows the detection of the start of the
1015 hash table and also allows the table's byte order to be determined so the table
1016 can be correctly extracted. The "``magic``" value is followed by a 16 bit
1017 ``version`` number which allows the table to be revised and modified in the
1018 future. The current version number is 1. ``hash_function`` is a ``uint16_t``
1019 enumeration that specifies which hash function was used to produce this table.
1020 The current values for the hash function enumerations include:
1024 enum HashFunctionType
1026 eHashFunctionDJB = 0u, // Daniel J Bernstein hash function
1029 ``bucket_count`` is a 32 bit unsigned integer that represents how many buckets
1030 are in the ``BUCKETS`` array. ``hashes_count`` is the number of unique 32 bit
1031 hash values that are in the ``HASHES`` array, and is the same number of offsets
1032 are contained in the ``OFFSETS`` array. ``header_data_len`` specifies the size
1033 in bytes of the ``HeaderData`` that is filled in by specialized versions of
1039 The header is followed by the buckets, hashes, offsets, and hash value data.
1045 uint32_t buckets[Header.bucket_count]; // An array of hash indexes into the "hashes[]" array below
1046 uint32_t hashes [Header.hashes_count]; // Every unique 32 bit hash for the entire table is in this table
1047 uint32_t offsets[Header.hashes_count]; // An offset that corresponds to each item in the "hashes[]" array above
1050 ``buckets`` is an array of 32 bit indexes into the ``hashes`` array. The
1051 ``hashes`` array contains all of the 32 bit hash values for all names in the
1052 hash table. Each hash in the ``hashes`` table has an offset in the ``offsets``
1053 array that points to the data for the hash value.
1055 This table setup makes it very easy to repurpose these tables to contain
1056 different data, while keeping the lookup mechanism the same for all tables.
1057 This layout also makes it possible to save the table to disk and map it in
1058 later and do very efficient name lookups with little or no parsing.
1060 DWARF lookup tables can be implemented in a variety of ways and can store a lot
1061 of information for each name. We want to make the DWARF tables extensible and
1062 able to store the data efficiently so we have used some of the DWARF features
1063 that enable efficient data storage to define exactly what kind of data we store
1066 The ``HeaderData`` contains a definition of the contents of each HashData chunk.
1067 We might want to store an offset to all of the debug information entries (DIEs)
1068 for each name. To keep things extensible, we create a list of items, or
1069 Atoms, that are contained in the data for each name. First comes the type of
1070 the data in each atom:
1077 eAtomTypeDIEOffset = 1u, // DIE offset, check form for encoding
1078 eAtomTypeCUOffset = 2u, // DIE offset of the compiler unit header that contains the item in question
1079 eAtomTypeTag = 3u, // DW_TAG_xxx value, should be encoded as DW_FORM_data1 (if no tags exceed 255) or DW_FORM_data2
1080 eAtomTypeNameFlags = 4u, // Flags from enum NameFlags
1081 eAtomTypeTypeFlags = 5u, // Flags from enum TypeFlags
1084 The enumeration values and their meanings are:
1086 .. code-block:: none
1088 eAtomTypeNULL - a termination atom that specifies the end of the atom list
1089 eAtomTypeDIEOffset - an offset into the .debug_info section for the DWARF DIE for this name
1090 eAtomTypeCUOffset - an offset into the .debug_info section for the CU that contains the DIE
1091 eAtomTypeDIETag - The DW_TAG_XXX enumeration value so you don't have to parse the DWARF to see what it is
1092 eAtomTypeNameFlags - Flags for functions and global variables (isFunction, isInlined, isExternal...)
1093 eAtomTypeTypeFlags - Flags for types (isCXXClass, isObjCClass, ...)
1095 Then we allow each atom type to define the atom type and how the data for each
1096 atom type data is encoded:
1102 uint16_t type; // AtomType enum value
1103 uint16_t form; // DWARF DW_FORM_XXX defines
1106 The ``form`` type above is from the DWARF specification and defines the exact
1107 encoding of the data for the Atom type. See the DWARF specification for the
1108 ``DW_FORM_`` definitions.
1114 uint32_t die_offset_base;
1115 uint32_t atom_count;
1116 Atoms atoms[atom_count0];
1119 ``HeaderData`` defines the base DIE offset that should be added to any atoms
1120 that are encoded using the ``DW_FORM_ref1``, ``DW_FORM_ref2``,
1121 ``DW_FORM_ref4``, ``DW_FORM_ref8`` or ``DW_FORM_ref_udata``. It also defines
1122 what is contained in each ``HashData`` object -- ``Atom.form`` tells us how large
1123 each field will be in the ``HashData`` and the ``Atom.type`` tells us how this data
1124 should be interpreted.
1126 For the current implementations of the "``.apple_names``" (all functions +
1127 globals), the "``.apple_types``" (names of all types that are defined), and
1128 the "``.apple_namespaces``" (all namespaces), we currently set the ``Atom``
1133 HeaderData.atom_count = 1;
1134 HeaderData.atoms[0].type = eAtomTypeDIEOffset;
1135 HeaderData.atoms[0].form = DW_FORM_data4;
1137 This defines the contents to be the DIE offset (eAtomTypeDIEOffset) that is
1138 encoded as a 32 bit value (DW_FORM_data4). This allows a single name to have
1139 multiple matching DIEs in a single file, which could come up with an inlined
1140 function for instance. Future tables could include more information about the
1141 DIE such as flags indicating if the DIE is a function, method, block,
1144 The KeyType for the DWARF table is a 32 bit string table offset into the
1145 ".debug_str" table. The ".debug_str" is the string table for the DWARF which
1146 may already contain copies of all of the strings. This helps make sure, with
1147 help from the compiler, that we reuse the strings between all of the DWARF
1148 sections and keeps the hash table size down. Another benefit to having the
1149 compiler generate all strings as DW_FORM_strp in the debug info, is that
1150 DWARF parsing can be made much faster.
1152 After a lookup is made, we get an offset into the hash data. The hash data
1153 needs to be able to deal with 32 bit hash collisions, so the chunk of data
1154 at the offset in the hash data consists of a triple:
1159 uint32_t hash_data_count
1160 HashData[hash_data_count]
1162 If "str_offset" is zero, then the bucket contents are done. 99.9% of the
1163 hash data chunks contain a single item (no 32 bit hash collision):
1165 .. code-block:: none
1168 | 0x00001023 | uint32_t KeyType (.debug_str[0x0001023] => "main")
1169 | 0x00000004 | uint32_t HashData count
1170 | 0x........ | uint32_t HashData[0] DIE offset
1171 | 0x........ | uint32_t HashData[1] DIE offset
1172 | 0x........ | uint32_t HashData[2] DIE offset
1173 | 0x........ | uint32_t HashData[3] DIE offset
1174 | 0x00000000 | uint32_t KeyType (end of hash chain)
1177 If there are collisions, you will have multiple valid string offsets:
1179 .. code-block:: none
1182 | 0x00001023 | uint32_t KeyType (.debug_str[0x0001023] => "main")
1183 | 0x00000004 | uint32_t HashData count
1184 | 0x........ | uint32_t HashData[0] DIE offset
1185 | 0x........ | uint32_t HashData[1] DIE offset
1186 | 0x........ | uint32_t HashData[2] DIE offset
1187 | 0x........ | uint32_t HashData[3] DIE offset
1188 | 0x00002023 | uint32_t KeyType (.debug_str[0x0002023] => "print")
1189 | 0x00000002 | uint32_t HashData count
1190 | 0x........ | uint32_t HashData[0] DIE offset
1191 | 0x........ | uint32_t HashData[1] DIE offset
1192 | 0x00000000 | uint32_t KeyType (end of hash chain)
1195 Current testing with real world C++ binaries has shown that there is around 1
1196 32 bit hash collision per 100,000 name entries.
1201 As we said, we want to strictly define exactly what is included in the
1202 different tables. For DWARF, we have 3 tables: "``.apple_names``",
1203 "``.apple_types``", and "``.apple_namespaces``".
1205 "``.apple_names``" sections should contain an entry for each DWARF DIE whose
1206 ``DW_TAG`` is a ``DW_TAG_label``, ``DW_TAG_inlined_subroutine``, or
1207 ``DW_TAG_subprogram`` that has address attributes: ``DW_AT_low_pc``,
1208 ``DW_AT_high_pc``, ``DW_AT_ranges`` or ``DW_AT_entry_pc``. It also contains
1209 ``DW_TAG_variable`` DIEs that have a ``DW_OP_addr`` in the location (global and
1210 static variables). All global and static variables should be included,
1211 including those scoped within functions and classes. For example using the
1223 Both of the static ``var`` variables would be included in the table. All
1224 functions should emit both their full names and their basenames. For C or C++,
1225 the full name is the mangled name (if available) which is usually in the
1226 ``DW_AT_MIPS_linkage_name`` attribute, and the ``DW_AT_name`` contains the
1227 function basename. If global or static variables have a mangled name in a
1228 ``DW_AT_MIPS_linkage_name`` attribute, this should be emitted along with the
1229 simple name found in the ``DW_AT_name`` attribute.
1231 "``.apple_types``" sections should contain an entry for each DWARF DIE whose
1236 * DW_TAG_enumeration_type
1237 * DW_TAG_pointer_type
1238 * DW_TAG_reference_type
1239 * DW_TAG_string_type
1240 * DW_TAG_structure_type
1241 * DW_TAG_subroutine_type
1244 * DW_TAG_ptr_to_member_type
1246 * DW_TAG_subrange_type
1251 * DW_TAG_packed_type
1252 * DW_TAG_volatile_type
1253 * DW_TAG_restrict_type
1254 * DW_TAG_interface_type
1255 * DW_TAG_unspecified_type
1256 * DW_TAG_shared_type
1258 Only entries with a ``DW_AT_name`` attribute are included, and the entry must
1259 not be a forward declaration (``DW_AT_declaration`` attribute with a non-zero
1260 value). For example, using the following code:
1270 We get a few type DIEs:
1272 .. code-block:: none
1274 0x00000067: TAG_base_type [5]
1275 AT_encoding( DW_ATE_signed )
1277 AT_byte_size( 0x04 )
1279 0x0000006e: TAG_pointer_type [6]
1280 AT_type( {0x00000067} ( int ) )
1281 AT_byte_size( 0x08 )
1283 The DW_TAG_pointer_type is not included because it does not have a ``DW_AT_name``.
1285 "``.apple_namespaces``" section should contain all ``DW_TAG_namespace`` DIEs.
1286 If we run into a namespace that has no name this is an anonymous namespace, and
1287 the name should be output as "``(anonymous namespace)``" (without the quotes).
1288 Why? This matches the output of the ``abi::cxa_demangle()`` that is in the
1289 standard C++ library that demangles mangled names.
1292 Language Extensions and File Format Changes
1293 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1295 Objective-C Extensions
1296 """"""""""""""""""""""
1298 "``.apple_objc``" section should contain all ``DW_TAG_subprogram`` DIEs for an
1299 Objective-C class. The name used in the hash table is the name of the
1300 Objective-C class itself. If the Objective-C class has a category, then an
1301 entry is made for both the class name without the category, and for the class
1302 name with the category. So if we have a DIE at offset 0x1234 with a name of
1303 method "``-[NSString(my_additions) stringWithSpecialString:]``", we would add
1304 an entry for "``NSString``" that points to DIE 0x1234, and an entry for
1305 "``NSString(my_additions)``" that points to 0x1234. This allows us to quickly
1306 track down all Objective-C methods for an Objective-C class when doing
1307 expressions. It is needed because of the dynamic nature of Objective-C where
1308 anyone can add methods to a class. The DWARF for Objective-C methods is also
1309 emitted differently from C++ classes where the methods are not usually
1310 contained in the class definition, they are scattered about across one or more
1311 compile units. Categories can also be defined in different shared libraries.
1312 So we need to be able to quickly find all of the methods and class functions
1313 given the Objective-C class name, or quickly find all methods and class
1314 functions for a class + category name. This table does not contain any
1315 selector names, it just maps Objective-C class names (or class names +
1316 category) to all of the methods and class functions. The selectors are added
1317 as function basenames in the "``.debug_names``" section.
1319 In the "``.apple_names``" section for Objective-C functions, the full name is
1320 the entire function name with the brackets ("``-[NSString
1321 stringWithCString:]``") and the basename is the selector only
1322 ("``stringWithCString:``").
1327 The sections names for the apple hash tables are for non-mach-o files. For
1328 mach-o files, the sections should be contained in the ``__DWARF`` segment with
1331 * "``.apple_names``" -> "``__apple_names``"
1332 * "``.apple_types``" -> "``__apple_types``"
1333 * "``.apple_namespaces``" -> "``__apple_namespac``" (16 character limit)
1334 * "``.apple_objc``" -> "``__apple_objc``"