xref: /llvm-project/llvm/docs/SourceLevelDebugging.rst (revision eab6e94f912d014e6f19e1737ef81e36e4601faf)
1================================
2Source Level Debugging with LLVM
3================================
4
5.. contents::
6   :local:
7
8Introduction
9============
10
11This document is the central repository for all information pertaining to debug
12information in LLVM.  It describes the :ref:`actual format that the LLVM debug
13information takes <format>`, which is useful for those interested in creating
14front-ends or dealing directly with the information.  Further, this document
15provides specific examples of what debug information for C/C++ looks like.
16
17Philosophy behind LLVM debugging information
18--------------------------------------------
19
20The idea of the LLVM debugging information is to capture how the important
21pieces of the source-language's Abstract Syntax Tree map onto LLVM code.
22Several design aspects have shaped the solution that appears here.  The
23important ones are:
24
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.
28
29* LLVM optimizations should interact in :ref:`well-defined and easily described
30  ways <intro_debugopt>` with the debugging information.
31
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.
35
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.
39
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.
44
45The approach used by the LLVM implementation is to use a small set of
46:ref:`intrinsic functions <format_common_intrinsics>` to define a mapping
47between LLVM program objects and the source-level objects.  The description of
48the source-level program is maintained in LLVM metadata in an
49:ref:`implementation-defined format <ccxx_frontend>` (the C/C++ front-end
50currently uses working draft 7 of the `DWARF 3 standard
51<http://www.eagercon.com/dwarf/dwarf3std.htm>`_).
52
53When a program is being debugged, a debugger interacts with the user and turns
54the stored debug information into source-language specific information.  As
55such, a debugger must be aware of the source-language, and is thus tied to a
56specific language or family of languages.
57
58Debug information consumers
59---------------------------
60
61The role of debug information is to provide meta information normally stripped
62away during the compilation process.  This meta information provides an LLVM
63user a relationship between generated code and the original program source
64code.
65
66Currently, there are two backend consumers of debug info: DwarfDebug and
67CodeViewDebug. DwarfDebug produces DWARF suitable for use with GDB, LLDB, and
68other DWARF-based debuggers. :ref:`CodeViewDebug <codeview>` produces CodeView,
69the Microsoft debug info format, which is usable with Microsoft debuggers such
70as Visual Studio and WinDBG. LLVM's debug information format is mostly derived
71from and inspired by DWARF, but it is feasible to translate into other target
72debug info formats such as STABS.
73
74It would also be reasonable to use debug information to feed profiling tools
75for analysis of generated code, or, tools for reconstructing the original
76source from generated code.
77
78.. _intro_debugopt:
79
80Debug information and optimizations
81-----------------------------------
82
83An extremely high priority of LLVM debugging information is to make it interact
84well with optimizations and analysis.  In particular, the LLVM debug
85information provides the following guarantees:
86
87* LLVM debug information **always provides information to accurately read
88  the source-level state of the program**, regardless of which LLVM
89  optimizations have been run. :doc:`HowToUpdateDebugInfo` specifies how debug
90  info should be updated in various kinds of code transformations to avoid
91  breaking this guarantee, and how to preserve as much useful debug info as
92  possible.  Note that some optimizations may impact the ability to modify the
93  current state of the program with a debugger, such as setting program
94  variables, or calling functions that have been deleted.
95
96* As desired, LLVM optimizations can be upgraded to be aware of debugging
97  information, allowing them to update the debugging information as they
98  perform aggressive optimizations.  This means that, with effort, the LLVM
99  optimizers could optimize debug code just as well as non-debug code.
100
101* LLVM debug information does not prevent optimizations from
102  happening (for example inlining, basic block reordering/merging/cleanup,
103  tail duplication, etc).
104
105* LLVM debug information is automatically optimized along with the rest of
106  the program, using existing facilities.  For example, duplicate
107  information is automatically merged by the linker, and unused information
108  is automatically removed.
109
110Basically, the debug information allows you to compile a program with
111"``-O0 -g``" and get full debug information, allowing you to arbitrarily modify
112the program as it executes from a debugger.  Compiling a program with
113"``-O3 -g``" gives you full debug information that is always available and
114accurate for reading (e.g., you get accurate stack traces despite tail call
115elimination and inlining), but you might lose the ability to modify the program
116and call functions which were optimized out of the program, or inlined away
117completely.
118
119The :doc:`LLVM test-suite <TestSuiteMakefileGuide>` provides a framework to
120test the optimizer's handling of debugging information.  It can be run like
121this:
122
123.. code-block:: bash
124
125  % cd llvm/projects/test-suite/MultiSource/Benchmarks  # or some other level
126  % make TEST=dbgopt
127
128This will test impact of debugging information on optimization passes.  If
129debugging information influences optimization passes then it will be reported
130as a failure.  See :doc:`TestingGuide` for more information on LLVM test
131infrastructure and how to run various tests.
132
133.. _format:
134
135Debugging information format
136============================
137
138LLVM debugging information has been carefully designed to make it possible for
139the optimizer to optimize the program and debugging information without
140necessarily having to know anything about debugging information.  In
141particular, the use of metadata avoids duplicated debugging information from
142the beginning, and the global dead code elimination pass automatically deletes
143debugging information for a function if it decides to delete the function.
144
145To do this, most of the debugging information (descriptors for types,
146variables, functions, source files, etc) is inserted by the language front-end
147in the form of LLVM metadata.
148
149Debug information is designed to be agnostic about the target debugger and
150debugging information representation (e.g. DWARF/Stabs/etc).  It uses a generic
151pass to decode the information that represents variables, types, functions,
152namespaces, etc: this allows for arbitrary source-language semantics and
153type-systems to be used, as long as there is a module written for the target
154debugger to interpret the information.
155
156To provide basic functionality, the LLVM debugger does have to make some
157assumptions about the source-level language being debugged, though it keeps
158these to a minimum.  The only common features that the LLVM debugger assumes
159exist are `source files <LangRef.html#difile>`_, and `program objects
160<LangRef.html#diglobalvariable>`_.  These abstract objects are used by a
161debugger to form stack traces, show information about local variables, etc.
162
163This section of the documentation first describes the representation aspects
164common to any source-language.  :ref:`ccxx_frontend` describes the data layout
165conventions used by the C and C++ front-ends.
166
167Debug information descriptors are `specialized metadata nodes
168<LangRef.html#specialized-metadata>`_, first-class subclasses of ``Metadata``.
169
170.. _format_common_intrinsics:
171
172Debugger intrinsic functions
173----------------------------
174
175LLVM uses several intrinsic functions (name prefixed with "``llvm.dbg``") to
176track source local variables through optimization and code generation.
177
178``llvm.dbg.addr``
179^^^^^^^^^^^^^^^^^^^^
180
181.. code-block:: llvm
182
183  void @llvm.dbg.addr(metadata, metadata, metadata)
184
185This intrinsic provides information about a local element (e.g., variable).
186The first argument is metadata holding the address of variable, typically a
187static alloca in the function entry block.  The second argument is a
188`local variable <LangRef.html#dilocalvariable>`_ containing a description of
189the variable.  The third argument is a `complex expression
190<LangRef.html#diexpression>`_.  An `llvm.dbg.addr` intrinsic describes the
191*address* of a source variable.
192
193.. code-block:: text
194
195    %i.addr = alloca i32, align 4
196    call void @llvm.dbg.addr(metadata i32* %i.addr, metadata !1,
197                             metadata !DIExpression()), !dbg !2
198    !1 = !DILocalVariable(name: "i", ...) ; int i
199    !2 = !DILocation(...)
200    ...
201    %buffer = alloca [256 x i8], align 8
202    ; The address of i is buffer+64.
203    call void @llvm.dbg.addr(metadata [256 x i8]* %buffer, metadata !3,
204                             metadata !DIExpression(DW_OP_plus, 64)), !dbg !4
205    !3 = !DILocalVariable(name: "i", ...) ; int i
206    !4 = !DILocation(...)
207
208A frontend should generate exactly one call to ``llvm.dbg.addr`` at the point
209of declaration of a source variable. Optimization passes that fully promote the
210variable from memory to SSA values will replace this call with possibly
211multiple calls to `llvm.dbg.value`. Passes that delete stores are effectively
212partial promotion, and they will insert a mix of calls to ``llvm.dbg.value``
213and ``llvm.dbg.addr`` to track the source variable value when it is available.
214After optimization, there may be multiple calls to ``llvm.dbg.addr`` describing
215the program points where the variables lives in memory. All calls for the same
216concrete source variable must agree on the memory location.
217
218
219``llvm.dbg.declare``
220^^^^^^^^^^^^^^^^^^^^
221
222.. code-block:: llvm
223
224  void @llvm.dbg.declare(metadata, metadata, metadata)
225
226This intrinsic is identical to `llvm.dbg.addr`, except that there can only be
227one call to `llvm.dbg.declare` for a given concrete `local variable
228<LangRef.html#dilocalvariable>`_. It is not control-dependent, meaning that if
229a call to `llvm.dbg.declare` exists and has a valid location argument, that
230address is considered to be the true home of the variable across its entire
231lifetime. This makes it hard for optimizations to preserve accurate debug info
232in the presence of ``llvm.dbg.declare``, so we are transitioning away from it,
233and we plan to deprecate it in future LLVM releases.
234
235
236``llvm.dbg.value``
237^^^^^^^^^^^^^^^^^^
238
239.. code-block:: llvm
240
241  void @llvm.dbg.value(metadata, metadata, metadata)
242
243This intrinsic provides information when a user source variable is set to a new
244value.  The first argument is the new value (wrapped as metadata).  The second
245argument is a `local variable <LangRef.html#dilocalvariable>`_ containing a
246description of the variable.  The third argument is a `complex expression
247<LangRef.html#diexpression>`_.
248
249An `llvm.dbg.value` intrinsic describes the *value* of a source variable
250directly, not its address.  Note that the value operand of this intrinsic may
251be indirect (i.e, a pointer to the source variable), provided that interpreting
252the complex expression derives the direct value.
253
254Object lifetimes and scoping
255============================
256
257In many languages, the local variables in functions can have their lifetimes or
258scopes limited to a subset of a function.  In the C family of languages, for
259example, variables are only live (readable and writable) within the source
260block that they are defined in.  In functional languages, values are only
261readable after they have been defined.  Though this is a very obvious concept,
262it is non-trivial to model in LLVM, because it has no notion of scoping in this
263sense, and does not want to be tied to a language's scoping rules.
264
265In order to handle this, the LLVM debug format uses the metadata attached to
266llvm instructions to encode line number and scoping information.  Consider the
267following C fragment, for example:
268
269.. code-block:: c
270
271  1.  void foo() {
272  2.    int X = 21;
273  3.    int Y = 22;
274  4.    {
275  5.      int Z = 23;
276  6.      Z = X;
277  7.    }
278  8.    X = Y;
279  9.  }
280
281.. FIXME: Update the following example to use llvm.dbg.addr once that is the
282   default in clang.
283
284Compiled to LLVM, this function would be represented like this:
285
286.. code-block:: text
287
288  ; Function Attrs: nounwind ssp uwtable
289  define void @foo() #0 !dbg !4 {
290  entry:
291    %X = alloca i32, align 4
292    %Y = alloca i32, align 4
293    %Z = alloca i32, align 4
294    call void @llvm.dbg.declare(metadata i32* %X, metadata !11, metadata !13), !dbg !14
295    store i32 21, i32* %X, align 4, !dbg !14
296    call void @llvm.dbg.declare(metadata i32* %Y, metadata !15, metadata !13), !dbg !16
297    store i32 22, i32* %Y, align 4, !dbg !16
298    call void @llvm.dbg.declare(metadata i32* %Z, metadata !17, metadata !13), !dbg !19
299    store i32 23, i32* %Z, align 4, !dbg !19
300    %0 = load i32, i32* %X, align 4, !dbg !20
301    store i32 %0, i32* %Z, align 4, !dbg !21
302    %1 = load i32, i32* %Y, align 4, !dbg !22
303    store i32 %1, i32* %X, align 4, !dbg !23
304    ret void, !dbg !24
305  }
306
307  ; Function Attrs: nounwind readnone
308  declare void @llvm.dbg.declare(metadata, metadata, metadata) #1
309
310  attributes #0 = { nounwind ssp uwtable "less-precise-fpmad"="false" "frame-pointer"="all" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }
311  attributes #1 = { nounwind readnone }
312
313  !llvm.dbg.cu = !{!0}
314  !llvm.module.flags = !{!7, !8, !9}
315  !llvm.ident = !{!10}
316
317  !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)
318  !1 = !DIFile(filename: "/dev/stdin", directory: "/Users/dexonsmith/data/llvm/debug-info")
319  !2 = !{}
320  !3 = !{!4}
321  !4 = distinct !DISubprogram(name: "foo", scope: !1, file: !1, line: 1, type: !5, isLocal: false, isDefinition: true, scopeLine: 1, isOptimized: false, retainedNodes: !2)
322  !5 = !DISubroutineType(types: !6)
323  !6 = !{null}
324  !7 = !{i32 2, !"Dwarf Version", i32 2}
325  !8 = !{i32 2, !"Debug Info Version", i32 3}
326  !9 = !{i32 1, !"PIC Level", i32 2}
327  !10 = !{!"clang version 3.7.0 (trunk 231150) (llvm/trunk 231154)"}
328  !11 = !DILocalVariable(name: "X", scope: !4, file: !1, line: 2, type: !12)
329  !12 = !DIBasicType(name: "int", size: 32, align: 32, encoding: DW_ATE_signed)
330  !13 = !DIExpression()
331  !14 = !DILocation(line: 2, column: 9, scope: !4)
332  !15 = !DILocalVariable(name: "Y", scope: !4, file: !1, line: 3, type: !12)
333  !16 = !DILocation(line: 3, column: 9, scope: !4)
334  !17 = !DILocalVariable(name: "Z", scope: !18, file: !1, line: 5, type: !12)
335  !18 = distinct !DILexicalBlock(scope: !4, file: !1, line: 4, column: 5)
336  !19 = !DILocation(line: 5, column: 11, scope: !18)
337  !20 = !DILocation(line: 6, column: 11, scope: !18)
338  !21 = !DILocation(line: 6, column: 9, scope: !18)
339  !22 = !DILocation(line: 8, column: 9, scope: !4)
340  !23 = !DILocation(line: 8, column: 7, scope: !4)
341  !24 = !DILocation(line: 9, column: 3, scope: !4)
342
343
344This example illustrates a few important details about LLVM debugging
345information.  In particular, it shows how the ``llvm.dbg.declare`` intrinsic and
346location information, which are attached to an instruction, are applied
347together to allow a debugger to analyze the relationship between statements,
348variable definitions, and the code used to implement the function.
349
350.. code-block:: llvm
351
352  call void @llvm.dbg.declare(metadata i32* %X, metadata !11, metadata !13), !dbg !14
353    ; [debug line = 2:7] [debug variable = X]
354
355The first intrinsic ``%llvm.dbg.declare`` encodes debugging information for the
356variable ``X``.  The metadata ``!dbg !14`` attached to the intrinsic provides
357scope information for the variable ``X``.
358
359.. code-block:: text
360
361  !14 = !DILocation(line: 2, column: 9, scope: !4)
362  !4 = distinct !DISubprogram(name: "foo", scope: !1, file: !1, line: 1, type: !5,
363                              isLocal: false, isDefinition: true, scopeLine: 1,
364                              isOptimized: false, retainedNodes: !2)
365
366Here ``!14`` is metadata providing `location information
367<LangRef.html#dilocation>`_.  In this example, scope is encoded by ``!4``, a
368`subprogram descriptor <LangRef.html#disubprogram>`_.  This way the location
369information attached to the intrinsics indicates that the variable ``X`` is
370declared at line number 2 at a function level scope in function ``foo``.
371
372Now lets take another example.
373
374.. code-block:: llvm
375
376  call void @llvm.dbg.declare(metadata i32* %Z, metadata !17, metadata !13), !dbg !19
377    ; [debug line = 5:9] [debug variable = Z]
378
379The third intrinsic ``%llvm.dbg.declare`` encodes debugging information for
380variable ``Z``.  The metadata ``!dbg !19`` attached to the intrinsic provides
381scope information for the variable ``Z``.
382
383.. code-block:: text
384
385  !18 = distinct !DILexicalBlock(scope: !4, file: !1, line: 4, column: 5)
386  !19 = !DILocation(line: 5, column: 11, scope: !18)
387
388Here ``!19`` indicates that ``Z`` is declared at line number 5 and column
389number 11 inside of lexical scope ``!18``.  The lexical scope itself resides
390inside of subprogram ``!4`` described above.
391
392The scope information attached with each instruction provides a straightforward
393way to find instructions covered by a scope.
394
395Object lifetime in optimized code
396=================================
397
398In the example above, every variable assignment uniquely corresponds to a
399memory store to the variable's position on the stack. However in heavily
400optimized code LLVM promotes most variables into SSA values, which can
401eventually be placed in physical registers or memory locations. To track SSA
402values through compilation, when objects are promoted to SSA values an
403``llvm.dbg.value`` intrinsic is created for each assignment, recording the
404variable's new location. Compared with the ``llvm.dbg.declare`` intrinsic:
405
406* A dbg.value terminates the effect of any preceding dbg.values for (any
407  overlapping fragments of) the specified variable.
408* The dbg.value's position in the IR defines where in the instruction stream
409  the variable's value changes.
410* Operands can be constants, indicating the variable is assigned a
411  constant value.
412
413Care must be taken to update ``llvm.dbg.value`` intrinsics when optimization
414passes alter or move instructions and blocks -- the developer could observe such
415changes reflected in the value of variables when debugging the program. For any
416execution of the optimized program, the set of variable values presented to the
417developer by the debugger should not show a state that would never have existed
418in the execution of the unoptimized program, given the same input. Doing so
419risks misleading the developer by reporting a state that does not exist,
420damaging their understanding of the optimized program and undermining their
421trust in the debugger.
422
423Sometimes perfectly preserving variable locations is not possible, often when a
424redundant calculation is optimized out. In such cases, a ``llvm.dbg.value``
425with operand ``undef`` should be used, to terminate earlier variable locations
426and let the debugger present ``optimized out`` to the developer. Withholding
427these potentially stale variable values from the developer diminishes the
428amount of available debug information, but increases the reliability of the
429remaining information.
430
431To illustrate some potential issues, consider the following example:
432
433.. code-block:: llvm
434
435  define i32 @foo(i32 %bar, i1 %cond) {
436  entry:
437    call @llvm.dbg.value(metadata i32 0, metadata !1, metadata !2)
438    br i1 %cond, label %truebr, label %falsebr
439  truebr:
440    %tval = add i32 %bar, 1
441    call @llvm.dbg.value(metadata i32 %tval, metadata !1, metadata !2)
442    %g1 = call i32 @gazonk()
443    br label %exit
444  falsebr:
445    %fval = add i32 %bar, 2
446    call @llvm.dbg.value(metadata i32 %fval, metadata !1, metadata !2)
447    %g2 = call i32 @gazonk()
448    br label %exit
449  exit:
450    %merge = phi [ %tval, %truebr ], [ %fval, %falsebr ]
451    %g = phi [ %g1, %truebr ], [ %g2, %falsebr ]
452    call @llvm.dbg.value(metadata i32 %merge, metadata !1, metadata !2)
453    call @llvm.dbg.value(metadata i32 %g, metadata !3, metadata !2)
454    %plusten = add i32 %merge, 10
455    %toret = add i32 %plusten, %g
456    call @llvm.dbg.value(metadata i32 %toret, metadata !1, metadata !2)
457    ret i32 %toret
458  }
459
460Containing two source-level variables in ``!1`` and ``!3``. The function could,
461perhaps, be optimized into the following code:
462
463.. code-block:: llvm
464
465  define i32 @foo(i32 %bar, i1 %cond) {
466  entry:
467    %g = call i32 @gazonk()
468    %addoper = select i1 %cond, i32 11, i32 12
469    %plusten = add i32 %bar, %addoper
470    %toret = add i32 %plusten, %g
471    ret i32 %toret
472  }
473
474What ``llvm.dbg.value`` intrinsics should be placed to represent the original variable
475locations in this code? Unfortunately the second, third and fourth
476dbg.values for ``!1`` in the source function have had their operands
477(%tval, %fval, %merge) optimized out. Assuming we cannot recover them, we
478might consider this placement of dbg.values:
479
480.. code-block:: llvm
481
482  define i32 @foo(i32 %bar, i1 %cond) {
483  entry:
484    call @llvm.dbg.value(metadata i32 0, metadata !1, metadata !2)
485    %g = call i32 @gazonk()
486    call @llvm.dbg.value(metadata i32 %g, metadata !3, metadata !2)
487    %addoper = select i1 %cond, i32 11, i32 12
488    %plusten = add i32 %bar, %addoper
489    %toret = add i32 %plusten, %g
490    call @llvm.dbg.value(metadata i32 %toret, metadata !1, metadata !2)
491    ret i32 %toret
492  }
493
494However, this will cause ``!3`` to have the return value of ``@gazonk()`` at
495the same time as ``!1`` has the constant value zero -- a pair of assignments
496that never occurred in the unoptimized program. To avoid this, we must terminate
497the range that ``!1`` has the constant value assignment by inserting an undef
498dbg.value before the dbg.value for ``!3``:
499
500.. code-block:: llvm
501
502  define i32 @foo(i32 %bar, i1 %cond) {
503  entry:
504    call @llvm.dbg.value(metadata i32 0, metadata !1, metadata !2)
505    %g = call i32 @gazonk()
506    call @llvm.dbg.value(metadata i32 undef, metadata !1, metadata !2)
507    call @llvm.dbg.value(metadata i32 %g, metadata !3, metadata !2)
508    %addoper = select i1 %cond, i32 11, i32 12
509    %plusten = add i32 %bar, %addoper
510    %toret = add i32 %plusten, %g
511    call @llvm.dbg.value(metadata i32 %toret, metadata !1, metadata !2)
512    ret i32 %toret
513  }
514
515In general, if any dbg.value has its operand optimized out and cannot be
516recovered, then an undef dbg.value is necessary to terminate earlier variable
517locations. Additional undef dbg.values may be necessary when the debugger can
518observe re-ordering of assignments.
519
520How variable location metadata is transformed during CodeGen
521============================================================
522
523LLVM preserves debug information throughout mid-level and backend passes,
524ultimately producing a mapping between source-level information and
525instruction ranges. This
526is relatively straightforwards for line number information, as mapping
527instructions to line numbers is a simple association. For variable locations
528however the story is more complex. As each ``llvm.dbg.value`` intrinsic
529represents a source-level assignment of a value to a source variable, the
530variable location intrinsics effectively embed a small imperative program
531within the LLVM IR. By the end of CodeGen, this becomes a mapping from each
532variable to their machine locations over ranges of instructions.
533From IR to object emission, the major transformations which affect variable
534location fidelity are:
535
5361. Instruction Selection
5372. Register allocation
5383. Block layout
539
540each of which are discussed below. In addition, instruction scheduling can
541significantly change the ordering of the program, and occurs in a number of
542different passes.
543
544Some variable locations are not transformed during CodeGen. Stack locations
545specified by ``llvm.dbg.declare`` are valid and unchanging for the entire
546duration of the function, and are recorded in a simple MachineFunction table.
547Location changes in the prologue and epilogue of a function are also ignored:
548frame setup and destruction may take several instructions, require a
549disproportionate amount of debugging information in the output binary to
550describe, and should be stepped over by debuggers anyway.
551
552Variable locations in Instruction Selection and MIR
553---------------------------------------------------
554
555Instruction selection creates a MIR function from an IR function, and just as
556it transforms ``intermediate`` instructions into machine instructions, so must
557``intermediate`` variable locations become machine variable locations.
558Within IR, variable locations are always identified by a Value, but in MIR
559there can be different types of variable locations. In addition, some IR
560locations become unavailable, for example if the operation of multiple IR
561instructions are combined into one machine instruction (such as
562multiply-and-accumulate) then intermediate Values are lost. To track variable
563locations through instruction selection, they are first separated into
564locations that do not depend on code generation (constants, stack locations,
565allocated virtual registers) and those that do. For those that do, debug
566metadata is attached to SDNodes in SelectionDAGs. After instruction selection
567has occurred and a MIR function is created, if the SDNode associated with debug
568metadata is allocated a virtual register, that virtual register is used as the
569variable location. If the SDNode is folded into a machine instruction or
570otherwise transformed into a non-register, the variable location becomes
571unavailable.
572
573Locations that are unavailable are treated as if they have been optimized out:
574in IR the location would be assigned ``undef`` by a debug intrinsic, and in MIR
575the equivalent location is used.
576
577After MIR locations are assigned to each variable, machine pseudo-instructions
578corresponding to each ``llvm.dbg.value`` and ``llvm.dbg.addr`` intrinsic are
579inserted. There are two forms of this type of instruction.
580
581The first form, ``DBG_VALUE``, appears thus:
582
583.. code-block:: text
584
585  DBG_VALUE %1, $noreg, !123, !DIExpression()
586
587And has the following operands:
588 * The first operand can record the variable location as a register,
589   a frame index, an immediate, or the base address register if the original
590   debug intrinsic referred to memory. ``$noreg`` indicates the variable
591   location is undefined, equivalent to an ``undef`` dbg.value operand.
592 * The type of the second operand indicates whether the variable location is
593   directly referred to by the DBG_VALUE, or whether it is indirect. The
594   ``$noreg`` register signifies the former, an immediate operand (0) the
595   latter.
596 * Operand 3 is the Variable field of the original debug intrinsic.
597 * Operand 4 is the Expression field of the original debug intrinsic.
598
599The second form, ``DBG_VALUE_LIST``, appears thus:
600
601.. code-block:: text
602
603  DBG_VALUE_LIST !123, !DIExpression(DW_OP_LLVM_arg, 0, DW_OP_LLVM_arg, 1, DW_OP_plus), %1, %2
604
605And has the following operands:
606 * The first operand is the Variable field of the original debug intrinsic.
607 * The second operand is the Expression field of the original debug intrinsic.
608 * Any number of operands, from the 3rd onwards, record a sequence of variable
609   location operands, which may take any of the same values as the first
610   operand of the ``DBG_VALUE`` instruction above. These variable location
611   operands are inserted into the final DWARF Expression in positions indicated
612   by the DW_OP_LLVM_arg operator in the `DIExpression
613   <LangRef.html#diexpression>`.
614
615The position at which the DBG_VALUEs are inserted should correspond to the
616positions of their matching ``llvm.dbg.value`` intrinsics in the IR block.  As
617with optimization, LLVM aims to preserve the order in which variable
618assignments occurred in the source program. However SelectionDAG performs some
619instruction scheduling, which can reorder assignments (discussed below).
620Function parameter locations are moved to the beginning of the function if
621they're not already, to ensure they're immediately available on function entry.
622
623To demonstrate variable locations during instruction selection, consider
624the following example:
625
626.. code-block:: llvm
627
628  define i32 @foo(i32* %addr) {
629  entry:
630    call void @llvm.dbg.value(metadata i32 0, metadata !3, metadata !DIExpression()), !dbg !5
631    br label %bb1, !dbg !5
632
633  bb1:                                              ; preds = %bb1, %entry
634    %bar.0 = phi i32 [ 0, %entry ], [ %add, %bb1 ]
635    call void @llvm.dbg.value(metadata i32 %bar.0, metadata !3, metadata !DIExpression()), !dbg !5
636    %addr1 = getelementptr i32, i32 *%addr, i32 1, !dbg !5
637    call void @llvm.dbg.value(metadata i32 *%addr1, metadata !3, metadata !DIExpression()), !dbg !5
638    %loaded1 = load i32, i32* %addr1, !dbg !5
639    %addr2 = getelementptr i32, i32 *%addr, i32 %bar.0, !dbg !5
640    call void @llvm.dbg.value(metadata i32 *%addr2, metadata !3, metadata !DIExpression()), !dbg !5
641    %loaded2 = load i32, i32* %addr2, !dbg !5
642    %add = add i32 %bar.0, 1, !dbg !5
643    call void @llvm.dbg.value(metadata i32 %add, metadata !3, metadata !DIExpression()), !dbg !5
644    %added = add i32 %loaded1, %loaded2
645    %cond = icmp ult i32 %added, %bar.0, !dbg !5
646    br i1 %cond, label %bb1, label %bb2, !dbg !5
647
648  bb2:                                              ; preds = %bb1
649    ret i32 0, !dbg !5
650  }
651
652If one compiles this IR with ``llc -o - -start-after=codegen-prepare -stop-after=expand-isel-pseudos -mtriple=x86_64--``, the following MIR is produced:
653
654.. code-block:: text
655
656  bb.0.entry:
657    successors: %bb.1(0x80000000)
658    liveins: $rdi
659
660    %2:gr64 = COPY $rdi
661    %3:gr32 = MOV32r0 implicit-def dead $eflags
662    DBG_VALUE 0, $noreg, !3, !DIExpression(), debug-location !5
663
664  bb.1.bb1:
665    successors: %bb.1(0x7c000000), %bb.2(0x04000000)
666
667    %0:gr32 = PHI %3, %bb.0, %1, %bb.1
668    DBG_VALUE %0, $noreg, !3, !DIExpression(), debug-location !5
669    DBG_VALUE %2, $noreg, !3, !DIExpression(DW_OP_plus_uconst, 4, DW_OP_stack_value), debug-location !5
670    %4:gr32 = MOV32rm %2, 1, $noreg, 4, $noreg, debug-location !5 :: (load 4 from %ir.addr1)
671    %5:gr64_nosp = MOVSX64rr32 %0, debug-location !5
672    DBG_VALUE $noreg, $noreg, !3, !DIExpression(), debug-location !5
673    %1:gr32 = INC32r %0, implicit-def dead $eflags, debug-location !5
674    DBG_VALUE %1, $noreg, !3, !DIExpression(), debug-location !5
675    %6:gr32 = ADD32rm %4, %2, 4, killed %5, 0, $noreg, implicit-def dead $eflags :: (load 4 from %ir.addr2)
676    %7:gr32 = SUB32rr %6, %0, implicit-def $eflags, debug-location !5
677    JB_1 %bb.1, implicit $eflags, debug-location !5
678    JMP_1 %bb.2, debug-location !5
679
680  bb.2.bb2:
681    %8:gr32 = MOV32r0 implicit-def dead $eflags
682    $eax = COPY %8, debug-location !5
683    RET 0, $eax, debug-location !5
684
685Observe first that there is a DBG_VALUE instruction for every ``llvm.dbg.value``
686intrinsic in the source IR, ensuring no source level assignments go missing.
687Then consider the different ways in which variable locations have been recorded:
688
689* For the first dbg.value an immediate operand is used to record a zero value.
690* The dbg.value of the PHI instruction leads to a DBG_VALUE of virtual register
691  ``%0``.
692* The first GEP has its effect folded into the first load instruction
693  (as a 4-byte offset), but the variable location is salvaged by folding
694  the GEPs effect into the DIExpression.
695* The second GEP is also folded into the corresponding load. However, it is
696  insufficiently simple to be salvaged, and is emitted as a ``$noreg``
697  DBG_VALUE, indicating that the variable takes on an undefined location.
698* The final dbg.value has its Value placed in virtual register ``%1``.
699
700Instruction Scheduling
701----------------------
702
703A number of passes can reschedule instructions, notably instruction selection
704and the pre-and-post RA machine schedulers. Instruction scheduling can
705significantly change the nature of the program -- in the (very unlikely) worst
706case the instruction sequence could be completely reversed. In such
707circumstances LLVM follows the principle applied to optimizations, that it is
708better for the debugger not to display any state than a misleading state.
709Thus, whenever instructions are advanced in order of execution, any
710corresponding DBG_VALUE is kept in its original position, and if an instruction
711is delayed then the variable is given an undefined location for the duration
712of the delay. To illustrate, consider this pseudo-MIR:
713
714.. code-block:: text
715
716  %1:gr32 = MOV32rm %0, 1, $noreg, 4, $noreg, debug-location !5 :: (load 4 from %ir.addr1)
717  DBG_VALUE %1, $noreg, !1, !2
718  %4:gr32 = ADD32rr %3, %2, implicit-def dead $eflags
719  DBG_VALUE %4, $noreg, !3, !4
720  %7:gr32 = SUB32rr %6, %5, implicit-def dead $eflags
721  DBG_VALUE %7, $noreg, !5, !6
722
723Imagine that the SUB32rr were moved forward to give us the following MIR:
724
725.. code-block:: text
726
727  %7:gr32 = SUB32rr %6, %5, implicit-def dead $eflags
728  %1:gr32 = MOV32rm %0, 1, $noreg, 4, $noreg, debug-location !5 :: (load 4 from %ir.addr1)
729  DBG_VALUE %1, $noreg, !1, !2
730  %4:gr32 = ADD32rr %3, %2, implicit-def dead $eflags
731  DBG_VALUE %4, $noreg, !3, !4
732  DBG_VALUE %7, $noreg, !5, !6
733
734In this circumstance LLVM would leave the MIR as shown above. Were we to move
735the DBG_VALUE of virtual register %7 upwards with the SUB32rr, we would re-order
736assignments and introduce a new state of the program. Whereas with the solution
737above, the debugger will see one fewer combination of variable values, because
738``!3`` and ``!5`` will change value at the same time. This is preferred over
739misrepresenting the original program.
740
741In comparison, if one sunk the MOV32rm, LLVM would produce the following:
742
743.. code-block:: text
744
745  DBG_VALUE $noreg, $noreg, !1, !2
746  %4:gr32 = ADD32rr %3, %2, implicit-def dead $eflags
747  DBG_VALUE %4, $noreg, !3, !4
748  %7:gr32 = SUB32rr %6, %5, implicit-def dead $eflags
749  DBG_VALUE %7, $noreg, !5, !6
750  %1:gr32 = MOV32rm %0, 1, $noreg, 4, $noreg, debug-location !5 :: (load 4 from %ir.addr1)
751  DBG_VALUE %1, $noreg, !1, !2
752
753Here, to avoid presenting a state in which the first assignment to ``!1``
754disappears, the DBG_VALUE at the top of the block assigns the variable the
755undefined location, until its value is available at the end of the block where
756an additional DBG_VALUE is added. Were any other DBG_VALUE for ``!1`` to occur
757in the instructions that the MOV32rm was sunk past, the DBG_VALUE for ``%1``
758would be dropped and the debugger would never observe it in the variable. This
759accurately reflects that the value is not available during the corresponding
760portion of the original program.
761
762Variable locations during Register Allocation
763---------------------------------------------
764
765To avoid debug instructions interfering with the register allocator, the
766LiveDebugVariables pass extracts variable locations from a MIR function and
767deletes the corresponding DBG_VALUE instructions. Some localized copy
768propagation is performed within blocks. After register allocation, the
769VirtRegRewriter pass re-inserts DBG_VALUE instructions in their original
770positions, translating virtual register references into their physical
771machine locations. To avoid encoding incorrect variable locations, in this
772pass any DBG_VALUE of a virtual register that is not live, is replaced by
773the undefined location. The LiveDebugVariables may insert redundant DBG_VALUEs
774because of virtual register rewriting. These will be subsequently removed by
775the RemoveRedundantDebugValues pass.
776
777LiveDebugValues expansion of variable locations
778-----------------------------------------------
779
780After all optimizations have run and shortly before emission, the
781LiveDebugValues pass runs to achieve two aims:
782
783* To propagate the location of variables through copies and register spills,
784* For every block, to record every valid variable location in that block.
785
786After this pass the DBG_VALUE instruction changes meaning: rather than
787corresponding to a source-level assignment where the variable may change value,
788it asserts the location of a variable in a block, and loses effect outside the
789block. Propagating variable locations through copies and spills is
790straightforwards: determining the variable location in every basic block
791requires the consideration of control flow. Consider the following IR, which
792presents several difficulties:
793
794.. code-block:: text
795
796  define dso_local i32 @foo(i1 %cond, i32 %input) !dbg !12 {
797  entry:
798    br i1 %cond, label %truebr, label %falsebr
799
800  bb1:
801    %value = phi i32 [ %value1, %truebr ], [ %value2, %falsebr ]
802    br label %exit, !dbg !26
803
804  truebr:
805    call void @llvm.dbg.value(metadata i32 %input, metadata !30, metadata !DIExpression()), !dbg !24
806    call void @llvm.dbg.value(metadata i32 1, metadata !23, metadata !DIExpression()), !dbg !24
807    %value1 = add i32 %input, 1
808    br label %bb1
809
810  falsebr:
811    call void @llvm.dbg.value(metadata i32 %input, metadata !30, metadata !DIExpression()), !dbg !24
812    call void @llvm.dbg.value(metadata i32 2, metadata !23, metadata !DIExpression()), !dbg !24
813    %value = add i32 %input, 2
814    br label %bb1
815
816  exit:
817    ret i32 %value, !dbg !30
818  }
819
820Here the difficulties are:
821
822* The control flow is roughly the opposite of basic block order
823* The value of the ``!23`` variable merges into ``%bb1``, but there is no PHI
824  node
825
826As mentioned above, the ``llvm.dbg.value`` intrinsics essentially form an
827imperative program embedded in the IR, with each intrinsic defining a variable
828location. This *could* be converted to an SSA form by mem2reg, in the same way
829that it uses use-def chains to identify control flow merges and insert phi
830nodes for IR Values. However, because debug variable locations are defined for
831every machine instruction, in effect every IR instruction uses every variable
832location, which would lead to a large number of debugging intrinsics being
833generated.
834
835Examining the example above, variable ``!30`` is assigned ``%input`` on both
836conditional paths through the function, while ``!23`` is assigned differing
837constant values on either path. Where control flow merges in ``%bb1`` we would
838want ``!30`` to keep its location (``%input``), but ``!23`` to become undefined
839as we cannot determine at runtime what value it should have in %bb1 without
840inserting a PHI node. mem2reg does not insert the PHI node to avoid changing
841codegen when debugging is enabled, and does not insert the other dbg.values
842to avoid adding very large numbers of intrinsics.
843
844Instead, LiveDebugValues determines variable locations when control
845flow merges. A dataflow analysis is used to propagate locations between blocks:
846when control flow merges, if a variable has the same location in all
847predecessors then that location is propagated into the successor. If the
848predecessor locations disagree, the location becomes undefined.
849
850Once LiveDebugValues has run, every block should have all valid variable
851locations described by DBG_VALUE instructions within the block. Very little
852effort is then required by supporting classes (such as
853DbgEntityHistoryCalculator) to build a map of each instruction to every
854valid variable location, without the need to consider control flow. From
855the example above, it is otherwise difficult to determine that the location
856of variable ``!30`` should flow "up" into block ``%bb1``, but that the location
857of variable ``!23`` should not flow "down" into the ``%exit`` block.
858
859.. _ccxx_frontend:
860
861C/C++ front-end specific debug information
862==========================================
863
864The C and C++ front-ends represent information about the program in a
865format that is effectively identical to `DWARF <http://www.dwarfstd.org/>`_
866in terms of information content.  This allows code generators to
867trivially support native debuggers by generating standard dwarf
868information, and contains enough information for non-dwarf targets to
869translate it as needed.
870
871This section describes the forms used to represent C and C++ programs.  Other
872languages could pattern themselves after this (which itself is tuned to
873representing programs in the same way that DWARF does), or they could choose
874to provide completely different forms if they don't fit into the DWARF model.
875As support for debugging information gets added to the various LLVM
876source-language front-ends, the information used should be documented here.
877
878The following sections provide examples of a few C/C++ constructs and
879the debug information that would best describe those constructs.  The
880canonical references are the ``DINode`` classes defined in
881``include/llvm/IR/DebugInfoMetadata.h`` and the implementations of the
882helper functions in ``lib/IR/DIBuilder.cpp``.
883
884C/C++ source file information
885-----------------------------
886
887``llvm::Instruction`` provides easy access to metadata attached with an
888instruction.  One can extract line number information encoded in LLVM IR using
889``Instruction::getDebugLoc()`` and ``DILocation::getLine()``.
890
891.. code-block:: c++
892
893  if (DILocation *Loc = I->getDebugLoc()) { // Here I is an LLVM instruction
894    unsigned Line = Loc->getLine();
895    StringRef File = Loc->getFilename();
896    StringRef Dir = Loc->getDirectory();
897    bool ImplicitCode = Loc->isImplicitCode();
898  }
899
900When the flag ImplicitCode is true then it means that the Instruction has been
901added by the front-end but doesn't correspond to source code written by the user. For example
902
903.. code-block:: c++
904
905  if (MyBoolean) {
906    MyObject MO;
907    ...
908  }
909
910At the end of the scope the MyObject's destructor is called but it isn't written
911explicitly. This information is useful to avoid to have counters on brackets when
912making code coverage.
913
914C/C++ global variable information
915---------------------------------
916
917Given an integer global variable declared as follows:
918
919.. code-block:: c
920
921  _Alignas(8) int MyGlobal = 100;
922
923a C/C++ front-end would generate the following descriptors:
924
925.. code-block:: text
926
927  ;;
928  ;; Define the global itself.
929  ;;
930  @MyGlobal = global i32 100, align 8, !dbg !0
931
932  ;;
933  ;; List of debug info of globals
934  ;;
935  !llvm.dbg.cu = !{!1}
936
937  ;; Some unrelated metadata.
938  !llvm.module.flags = !{!6, !7}
939  !llvm.ident = !{!8}
940
941  ;; Define the global variable itself
942  !0 = distinct !DIGlobalVariable(name: "MyGlobal", scope: !1, file: !2, line: 1, type: !5, isLocal: false, isDefinition: true, align: 64)
943
944  ;; Define the compile unit.
945  !1 = distinct !DICompileUnit(language: DW_LANG_C99, file: !2,
946                               producer: "clang version 4.0.0",
947                               isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug,
948                               enums: !3, globals: !4)
949
950  ;;
951  ;; Define the file
952  ;;
953  !2 = !DIFile(filename: "/dev/stdin",
954               directory: "/Users/dexonsmith/data/llvm/debug-info")
955
956  ;; An empty array.
957  !3 = !{}
958
959  ;; The Array of Global Variables
960  !4 = !{!0}
961
962  ;;
963  ;; Define the type
964  ;;
965  !5 = !DIBasicType(name: "int", size: 32, encoding: DW_ATE_signed)
966
967  ;; Dwarf version to output.
968  !6 = !{i32 2, !"Dwarf Version", i32 4}
969
970  ;; Debug info schema version.
971  !7 = !{i32 2, !"Debug Info Version", i32 3}
972
973  ;; Compiler identification
974  !8 = !{!"clang version 4.0.0"}
975
976
977The align value in DIGlobalVariable description specifies variable alignment in
978case it was forced by C11 _Alignas(), C++11 alignas() keywords or compiler
979attribute __attribute__((aligned ())). In other case (when this field is missing)
980alignment is considered default. This is used when producing DWARF output
981for DW_AT_alignment value.
982
983C/C++ function information
984--------------------------
985
986Given a function declared as follows:
987
988.. code-block:: c
989
990  int main(int argc, char *argv[]) {
991    return 0;
992  }
993
994a C/C++ front-end would generate the following descriptors:
995
996.. code-block:: text
997
998  ;;
999  ;; Define the anchor for subprograms.
1000  ;;
1001  !4 = !DISubprogram(name: "main", scope: !1, file: !1, line: 1, type: !5,
1002                     isLocal: false, isDefinition: true, scopeLine: 1,
1003                     flags: DIFlagPrototyped, isOptimized: false,
1004                     retainedNodes: !2)
1005
1006  ;;
1007  ;; Define the subprogram itself.
1008  ;;
1009  define i32 @main(i32 %argc, i8** %argv) !dbg !4 {
1010  ...
1011  }
1012
1013C++ specific debug information
1014==============================
1015
1016C++ special member functions information
1017----------------------------------------
1018
1019DWARF v5 introduces attributes defined to enhance debugging information of C++ programs. LLVM can generate (or omit) these appropriate DWARF attributes. In C++ a special member function Ctors, Dtors, Copy/Move Ctors, assignment operators can be declared with C++11 keyword deleted. This is represented in LLVM using spFlags value DISPFlagDeleted.
1020
1021Given a class declaration with copy constructor declared as deleted:
1022
1023.. code-block:: c
1024
1025  class foo {
1026   public:
1027     foo(const foo&) = deleted;
1028  };
1029
1030A C++ frontend would generate following:
1031
1032.. code-block:: text
1033
1034  !17 = !DISubprogram(name: "foo", scope: !11, file: !1, line: 5, type: !18, scopeLine: 5, flags: DIFlagPublic | DIFlagPrototyped, spFlags: DISPFlagDeleted)
1035
1036and this will produce an additional DWARF attribute as:
1037
1038.. code-block:: text
1039
1040  DW_TAG_subprogram [7] *
1041    DW_AT_name [DW_FORM_strx1]    (indexed (00000006) string = "foo")
1042    DW_AT_decl_line [DW_FORM_data1]       (5)
1043    ...
1044    DW_AT_deleted [DW_FORM_flag_present]  (true)
1045
1046Fortran specific debug information
1047==================================
1048
1049Fortran function information
1050----------------------------
1051
1052There are a few DWARF attributes defined to support client debugging of Fortran programs.  LLVM can generate (or omit) the appropriate DWARF attributes for the prefix-specs of ELEMENTAL, PURE, IMPURE, RECURSIVE, and NON_RECURSIVE.  This is done by using the spFlags values: DISPFlagElemental, DISPFlagPure, and DISPFlagRecursive.
1053
1054.. code-block:: fortran
1055
1056  elemental function elem_func(a)
1057
1058a Fortran front-end would generate the following descriptors:
1059
1060.. code-block:: text
1061
1062  !11 = distinct !DISubprogram(name: "subroutine2", scope: !1, file: !1,
1063          line: 5, type: !8, scopeLine: 6,
1064          spFlags: DISPFlagDefinition | DISPFlagElemental, unit: !0,
1065          retainedNodes: !2)
1066
1067and this will materialize an additional DWARF attribute as:
1068
1069.. code-block:: text
1070
1071  DW_TAG_subprogram [3]
1072     DW_AT_low_pc [DW_FORM_addr]     (0x0000000000000010 ".text")
1073     DW_AT_high_pc [DW_FORM_data4]   (0x00000001)
1074     ...
1075     DW_AT_elemental [DW_FORM_flag_present]  (true)
1076
1077There are a few DWARF tags defined to represent Fortran specific constructs i.e DW_TAG_string_type for representing Fortran character(n). In LLVM this is represented as DIStringType.
1078
1079.. code-block:: fortran
1080
1081  character(len=*), intent(in) :: string
1082
1083a Fortran front-end would generate the following descriptors:
1084
1085.. code-block:: text
1086
1087  !DILocalVariable(name: "string", arg: 1, scope: !10, file: !3, line: 4, type: !15)
1088  !DIStringType(name: "character(*)!2", stringLength: !16, stringLengthExpression: !DIExpression(), size: 32)
1089
1090A fortran deferred-length character can also contain the information of raw storage of the characters in addition to the length of the string. This information is encoded in the  stringLocationExpression field. Based on this information, DW_AT_data_location attribute is emitted in a DW_TAG_string_type debug info.
1091
1092  !DIStringType(name: "character(*)!2", stringLengthExpression: !DIExpression(), stringLocationExpression: !DIExpression(DW_OP_push_object_address, DW_OP_deref), size: 32)
1093
1094and this will materialize in DWARF tags as:
1095
1096.. code-block:: text
1097
1098   DW_TAG_string_type
1099                DW_AT_name      ("character(*)!2")
1100                DW_AT_string_length     (0x00000064)
1101   0x00000064:    DW_TAG_variable
1102                  DW_AT_location      (DW_OP_fbreg +16)
1103                  DW_AT_type  (0x00000083 "integer*8")
1104                  DW_AT_data_location (DW_OP_push_object_address, DW_OP_deref)
1105                  ...
1106                  DW_AT_artificial    (true)
1107
1108A Fortran front-end may need to generate a *trampoline* function to call a
1109function defined in a different compilation unit. In this case, the front-end
1110can emit the following descriptor for the trampoline function:
1111
1112.. code-block:: text
1113
1114  !DISubprogram(name: "sub1_.t0p", linkageName: "sub1_.t0p", scope: !4, file: !4, type: !5, spFlags: DISPFlagLocalToUnit | DISPFlagDefinition, unit: !7, retainedNodes: !24, targetFuncName: "sub1_")
1115
1116The targetFuncName field is the name of the function that the trampoline
1117calls. This descriptor results in the following DWARF tag:
1118
1119.. code-block:: text
1120
1121  DW_TAG_subprogram
1122    ...
1123    DW_AT_linkage_name	("sub1_.t0p")
1124    DW_AT_name	("sub1_.t0p")
1125    DW_AT_trampoline	("sub1_")
1126
1127Debugging information format
1128============================
1129
1130Debugging Information Extension for Objective C Properties
1131----------------------------------------------------------
1132
1133Introduction
1134^^^^^^^^^^^^
1135
1136Objective C provides a simpler way to declare and define accessor methods using
1137declared properties.  The language provides features to declare a property and
1138to let compiler synthesize accessor methods.
1139
1140The debugger lets developer inspect Objective C interfaces and their instance
1141variables and class variables.  However, the debugger does not know anything
1142about the properties defined in Objective C interfaces.  The debugger consumes
1143information generated by compiler in DWARF format.  The format does not support
1144encoding of Objective C properties.  This proposal describes DWARF extensions to
1145encode Objective C properties, which the debugger can use to let developers
1146inspect Objective C properties.
1147
1148Proposal
1149^^^^^^^^
1150
1151Objective C properties exist separately from class members.  A property can be
1152defined only by "setter" and "getter" selectors, and be calculated anew on each
1153access.  Or a property can just be a direct access to some declared ivar.
1154Finally it can have an ivar "automatically synthesized" for it by the compiler,
1155in which case the property can be referred to in user code directly using the
1156standard C dereference syntax as well as through the property "dot" syntax, but
1157there is no entry in the ``@interface`` declaration corresponding to this ivar.
1158
1159To facilitate debugging, these properties we will add a new DWARF TAG into the
1160``DW_TAG_structure_type`` definition for the class to hold the description of a
1161given property, and a set of DWARF attributes that provide said description.
1162The property tag will also contain the name and declared type of the property.
1163
1164If there is a related ivar, there will also be a DWARF property attribute placed
1165in the ``DW_TAG_member`` DIE for that ivar referring back to the property TAG
1166for that property.  And in the case where the compiler synthesizes the ivar
1167directly, the compiler is expected to generate a ``DW_TAG_member`` for that
1168ivar (with the ``DW_AT_artificial`` set to 1), whose name will be the name used
1169to access this ivar directly in code, and with the property attribute pointing
1170back to the property it is backing.
1171
1172The following examples will serve as illustration for our discussion:
1173
1174.. code-block:: objc
1175
1176  @interface I1 {
1177    int n2;
1178  }
1179
1180  @property int p1;
1181  @property int p2;
1182  @end
1183
1184  @implementation I1
1185  @synthesize p1;
1186  @synthesize p2 = n2;
1187  @end
1188
1189This produces the following DWARF (this is a "pseudo dwarfdump" output):
1190
1191.. code-block:: none
1192
1193  0x00000100:  TAG_structure_type [7] *
1194                 AT_APPLE_runtime_class( 0x10 )
1195                 AT_name( "I1" )
1196                 AT_decl_file( "Objc_Property.m" )
1197                 AT_decl_line( 3 )
1198
1199  0x00000110    TAG_APPLE_property
1200                  AT_name ( "p1" )
1201                  AT_type ( {0x00000150} ( int ) )
1202
1203  0x00000120:   TAG_APPLE_property
1204                  AT_name ( "p2" )
1205                  AT_type ( {0x00000150} ( int ) )
1206
1207  0x00000130:   TAG_member [8]
1208                  AT_name( "_p1" )
1209                  AT_APPLE_property ( {0x00000110} "p1" )
1210                  AT_type( {0x00000150} ( int ) )
1211                  AT_artificial ( 0x1 )
1212
1213  0x00000140:    TAG_member [8]
1214                   AT_name( "n2" )
1215                   AT_APPLE_property ( {0x00000120} "p2" )
1216                   AT_type( {0x00000150} ( int ) )
1217
1218  0x00000150:  AT_type( ( int ) )
1219
1220Note, the current convention is that the name of the ivar for an
1221auto-synthesized property is the name of the property from which it derives
1222with an underscore prepended, as is shown in the example.  But we actually
1223don't need to know this convention, since we are given the name of the ivar
1224directly.
1225
1226Also, it is common practice in ObjC to have different property declarations in
1227the @interface and @implementation - e.g. to provide a read-only property in
1228the interface, and a read-write interface in the implementation.  In that case,
1229the compiler should emit whichever property declaration will be in force in the
1230current translation unit.
1231
1232Developers can decorate a property with attributes which are encoded using
1233``DW_AT_APPLE_property_attribute``.
1234
1235.. code-block:: objc
1236
1237  @property (readonly, nonatomic) int pr;
1238
1239.. code-block:: none
1240
1241  TAG_APPLE_property [8]
1242    AT_name( "pr" )
1243    AT_type ( {0x00000147} (int) )
1244    AT_APPLE_property_attribute (DW_APPLE_PROPERTY_readonly, DW_APPLE_PROPERTY_nonatomic)
1245
1246The setter and getter method names are attached to the property using
1247``DW_AT_APPLE_property_setter`` and ``DW_AT_APPLE_property_getter`` attributes.
1248
1249.. code-block:: objc
1250
1251  @interface I1
1252  @property (setter=myOwnP3Setter:) int p3;
1253  -(void)myOwnP3Setter:(int)a;
1254  @end
1255
1256  @implementation I1
1257  @synthesize p3;
1258  -(void)myOwnP3Setter:(int)a{ }
1259  @end
1260
1261The DWARF for this would be:
1262
1263.. code-block:: none
1264
1265  0x000003bd: TAG_structure_type [7] *
1266                AT_APPLE_runtime_class( 0x10 )
1267                AT_name( "I1" )
1268                AT_decl_file( "Objc_Property.m" )
1269                AT_decl_line( 3 )
1270
1271  0x000003cd      TAG_APPLE_property
1272                    AT_name ( "p3" )
1273                    AT_APPLE_property_setter ( "myOwnP3Setter:" )
1274                    AT_type( {0x00000147} ( int ) )
1275
1276  0x000003f3:     TAG_member [8]
1277                    AT_name( "_p3" )
1278                    AT_type ( {0x00000147} ( int ) )
1279                    AT_APPLE_property ( {0x000003cd} )
1280                    AT_artificial ( 0x1 )
1281
1282New DWARF Tags
1283^^^^^^^^^^^^^^
1284
1285+-----------------------+--------+
1286| TAG                   | Value  |
1287+=======================+========+
1288| DW_TAG_APPLE_property | 0x4200 |
1289+-----------------------+--------+
1290
1291New DWARF Attributes
1292^^^^^^^^^^^^^^^^^^^^
1293
1294+--------------------------------+--------+-----------+
1295| Attribute                      | Value  | Classes   |
1296+================================+========+===========+
1297| DW_AT_APPLE_property           | 0x3fed | Reference |
1298+--------------------------------+--------+-----------+
1299| DW_AT_APPLE_property_getter    | 0x3fe9 | String    |
1300+--------------------------------+--------+-----------+
1301| DW_AT_APPLE_property_setter    | 0x3fea | String    |
1302+--------------------------------+--------+-----------+
1303| DW_AT_APPLE_property_attribute | 0x3feb | Constant  |
1304+--------------------------------+--------+-----------+
1305
1306New DWARF Constants
1307^^^^^^^^^^^^^^^^^^^
1308
1309+--------------------------------------+-------+
1310| Name                                 | Value |
1311+======================================+=======+
1312| DW_APPLE_PROPERTY_readonly           | 0x01  |
1313+--------------------------------------+-------+
1314| DW_APPLE_PROPERTY_getter             | 0x02  |
1315+--------------------------------------+-------+
1316| DW_APPLE_PROPERTY_assign             | 0x04  |
1317+--------------------------------------+-------+
1318| DW_APPLE_PROPERTY_readwrite          | 0x08  |
1319+--------------------------------------+-------+
1320| DW_APPLE_PROPERTY_retain             | 0x10  |
1321+--------------------------------------+-------+
1322| DW_APPLE_PROPERTY_copy               | 0x20  |
1323+--------------------------------------+-------+
1324| DW_APPLE_PROPERTY_nonatomic          | 0x40  |
1325+--------------------------------------+-------+
1326| DW_APPLE_PROPERTY_setter             | 0x80  |
1327+--------------------------------------+-------+
1328| DW_APPLE_PROPERTY_atomic             | 0x100 |
1329+--------------------------------------+-------+
1330| DW_APPLE_PROPERTY_weak               | 0x200 |
1331+--------------------------------------+-------+
1332| DW_APPLE_PROPERTY_strong             | 0x400 |
1333+--------------------------------------+-------+
1334| DW_APPLE_PROPERTY_unsafe_unretained  | 0x800 |
1335+--------------------------------------+-------+
1336| DW_APPLE_PROPERTY_nullability        | 0x1000|
1337+--------------------------------------+-------+
1338| DW_APPLE_PROPERTY_null_resettable    | 0x2000|
1339+--------------------------------------+-------+
1340| DW_APPLE_PROPERTY_class              | 0x4000|
1341+--------------------------------------+-------+
1342
1343Name Accelerator Tables
1344-----------------------
1345
1346Introduction
1347^^^^^^^^^^^^
1348
1349The "``.debug_pubnames``" and "``.debug_pubtypes``" formats are not what a
1350debugger needs.  The "``pub``" in the section name indicates that the entries
1351in the table are publicly visible names only.  This means no static or hidden
1352functions show up in the "``.debug_pubnames``".  No static variables or private
1353class variables are in the "``.debug_pubtypes``".  Many compilers add different
1354things to these tables, so we can't rely upon the contents between gcc, icc, or
1355clang.
1356
1357The typical query given by users tends not to match up with the contents of
1358these tables.  For example, the DWARF spec states that "In the case of the name
1359of a function member or static data member of a C++ structure, class or union,
1360the name presented in the "``.debug_pubnames``" section is not the simple name
1361given by the ``DW_AT_name attribute`` of the referenced debugging information
1362entry, but rather the fully qualified name of the data or function member."
1363So the only names in these tables for complex C++ entries is a fully
1364qualified name.  Debugger users tend not to enter their search strings as
1365"``a::b::c(int,const Foo&) const``", but rather as "``c``", "``b::c``" , or
1366"``a::b::c``".  So the name entered in the name table must be demangled in
1367order to chop it up appropriately and additional names must be manually entered
1368into the table to make it effective as a name lookup table for debuggers to
1369use.
1370
1371All debuggers currently ignore the "``.debug_pubnames``" table as a result of
1372its inconsistent and useless public-only name content making it a waste of
1373space in the object file.  These tables, when they are written to disk, are not
1374sorted in any way, leaving every debugger to do its own parsing and sorting.
1375These tables also include an inlined copy of the string values in the table
1376itself making the tables much larger than they need to be on disk, especially
1377for large C++ programs.
1378
1379Can't we just fix the sections by adding all of the names we need to this
1380table? No, because that is not what the tables are defined to contain and we
1381won't know the difference between the old bad tables and the new good tables.
1382At best we could make our own renamed sections that contain all of the data we
1383need.
1384
1385These tables are also insufficient for what a debugger like LLDB needs.  LLDB
1386uses clang for its expression parsing where LLDB acts as a PCH.  LLDB is then
1387often asked to look for type "``foo``" or namespace "``bar``", or list items in
1388namespace "``baz``".  Namespaces are not included in the pubnames or pubtypes
1389tables.  Since clang asks a lot of questions when it is parsing an expression,
1390we need to be very fast when looking up names, as it happens a lot.  Having new
1391accelerator tables that are optimized for very quick lookups will benefit this
1392type of debugging experience greatly.
1393
1394We would like to generate name lookup tables that can be mapped into memory
1395from disk, and used as is, with little or no up-front parsing.  We would also
1396be able to control the exact content of these different tables so they contain
1397exactly what we need.  The Name Accelerator Tables were designed to fix these
1398issues.  In order to solve these issues we need to:
1399
1400* Have a format that can be mapped into memory from disk and used as is
1401* Lookups should be very fast
1402* Extensible table format so these tables can be made by many producers
1403* Contain all of the names needed for typical lookups out of the box
1404* Strict rules for the contents of tables
1405
1406Table size is important and the accelerator table format should allow the reuse
1407of strings from common string tables so the strings for the names are not
1408duplicated.  We also want to make sure the table is ready to be used as-is by
1409simply mapping the table into memory with minimal header parsing.
1410
1411The name lookups need to be fast and optimized for the kinds of lookups that
1412debuggers tend to do.  Optimally we would like to touch as few parts of the
1413mapped table as possible when doing a name lookup and be able to quickly find
1414the name entry we are looking for, or discover there are no matches.  In the
1415case of debuggers we optimized for lookups that fail most of the time.
1416
1417Each table that is defined should have strict rules on exactly what is in the
1418accelerator tables and documented so clients can rely on the content.
1419
1420Hash Tables
1421^^^^^^^^^^^
1422
1423Standard Hash Tables
1424""""""""""""""""""""
1425
1426Typical hash tables have a header, buckets, and each bucket points to the
1427bucket contents:
1428
1429.. code-block:: none
1430
1431  .------------.
1432  |  HEADER    |
1433  |------------|
1434  |  BUCKETS   |
1435  |------------|
1436  |  DATA      |
1437  `------------'
1438
1439The BUCKETS are an array of offsets to DATA for each hash:
1440
1441.. code-block:: none
1442
1443  .------------.
1444  | 0x00001000 | BUCKETS[0]
1445  | 0x00002000 | BUCKETS[1]
1446  | 0x00002200 | BUCKETS[2]
1447  | 0x000034f0 | BUCKETS[3]
1448  |            | ...
1449  | 0xXXXXXXXX | BUCKETS[n_buckets]
1450  '------------'
1451
1452So for ``bucket[3]`` in the example above, we have an offset into the table
14530x000034f0 which points to a chain of entries for the bucket.  Each bucket must
1454contain a next pointer, full 32 bit hash value, the string itself, and the data
1455for the current string value.
1456
1457.. code-block:: none
1458
1459              .------------.
1460  0x000034f0: | 0x00003500 | next pointer
1461              | 0x12345678 | 32 bit hash
1462              | "erase"    | string value
1463              | data[n]    | HashData for this bucket
1464              |------------|
1465  0x00003500: | 0x00003550 | next pointer
1466              | 0x29273623 | 32 bit hash
1467              | "dump"     | string value
1468              | data[n]    | HashData for this bucket
1469              |------------|
1470  0x00003550: | 0x00000000 | next pointer
1471              | 0x82638293 | 32 bit hash
1472              | "main"     | string value
1473              | data[n]    | HashData for this bucket
1474              `------------'
1475
1476The problem with this layout for debuggers is that we need to optimize for the
1477negative lookup case where the symbol we're searching for is not present.  So
1478if we were to lookup "``printf``" in the table above, we would make a 32-bit
1479hash for "``printf``", it might match ``bucket[3]``.  We would need to go to
1480the offset 0x000034f0 and start looking to see if our 32 bit hash matches.  To
1481do so, we need to read the next pointer, then read the hash, compare it, and
1482skip to the next bucket.  Each time we are skipping many bytes in memory and
1483touching new pages just to do the compare on the full 32 bit hash.  All of
1484these accesses then tell us that we didn't have a match.
1485
1486Name Hash Tables
1487""""""""""""""""
1488
1489To solve the issues mentioned above we have structured the hash tables a bit
1490differently: a header, buckets, an array of all unique 32 bit hash values,
1491followed by an array of hash value data offsets, one for each hash value, then
1492the data for all hash values:
1493
1494.. code-block:: none
1495
1496  .-------------.
1497  |  HEADER     |
1498  |-------------|
1499  |  BUCKETS    |
1500  |-------------|
1501  |  HASHES     |
1502  |-------------|
1503  |  OFFSETS    |
1504  |-------------|
1505  |  DATA       |
1506  `-------------'
1507
1508The ``BUCKETS`` in the name tables are an index into the ``HASHES`` array.  By
1509making all of the full 32 bit hash values contiguous in memory, we allow
1510ourselves to efficiently check for a match while touching as little memory as
1511possible.  Most often checking the 32 bit hash values is as far as the lookup
1512goes.  If it does match, it usually is a match with no collisions.  So for a
1513table with "``n_buckets``" buckets, and "``n_hashes``" unique 32 bit hash
1514values, we can clarify the contents of the ``BUCKETS``, ``HASHES`` and
1515``OFFSETS`` as:
1516
1517.. code-block:: none
1518
1519  .-------------------------.
1520  |  HEADER.magic           | uint32_t
1521  |  HEADER.version         | uint16_t
1522  |  HEADER.hash_function   | uint16_t
1523  |  HEADER.bucket_count    | uint32_t
1524  |  HEADER.hashes_count    | uint32_t
1525  |  HEADER.header_data_len | uint32_t
1526  |  HEADER_DATA            | HeaderData
1527  |-------------------------|
1528  |  BUCKETS                | uint32_t[n_buckets] // 32 bit hash indexes
1529  |-------------------------|
1530  |  HASHES                 | uint32_t[n_hashes] // 32 bit hash values
1531  |-------------------------|
1532  |  OFFSETS                | uint32_t[n_hashes] // 32 bit offsets to hash value data
1533  |-------------------------|
1534  |  ALL HASH DATA          |
1535  `-------------------------'
1536
1537So taking the exact same data from the standard hash example above we end up
1538with:
1539
1540.. code-block:: none
1541
1542              .------------.
1543              | HEADER     |
1544              |------------|
1545              |          0 | BUCKETS[0]
1546              |          2 | BUCKETS[1]
1547              |          5 | BUCKETS[2]
1548              |          6 | BUCKETS[3]
1549              |            | ...
1550              |        ... | BUCKETS[n_buckets]
1551              |------------|
1552              | 0x........ | HASHES[0]
1553              | 0x........ | HASHES[1]
1554              | 0x........ | HASHES[2]
1555              | 0x........ | HASHES[3]
1556              | 0x........ | HASHES[4]
1557              | 0x........ | HASHES[5]
1558              | 0x12345678 | HASHES[6]    hash for BUCKETS[3]
1559              | 0x29273623 | HASHES[7]    hash for BUCKETS[3]
1560              | 0x82638293 | HASHES[8]    hash for BUCKETS[3]
1561              | 0x........ | HASHES[9]
1562              | 0x........ | HASHES[10]
1563              | 0x........ | HASHES[11]
1564              | 0x........ | HASHES[12]
1565              | 0x........ | HASHES[13]
1566              | 0x........ | HASHES[n_hashes]
1567              |------------|
1568              | 0x........ | OFFSETS[0]
1569              | 0x........ | OFFSETS[1]
1570              | 0x........ | OFFSETS[2]
1571              | 0x........ | OFFSETS[3]
1572              | 0x........ | OFFSETS[4]
1573              | 0x........ | OFFSETS[5]
1574              | 0x000034f0 | OFFSETS[6]   offset for BUCKETS[3]
1575              | 0x00003500 | OFFSETS[7]   offset for BUCKETS[3]
1576              | 0x00003550 | OFFSETS[8]   offset for BUCKETS[3]
1577              | 0x........ | OFFSETS[9]
1578              | 0x........ | OFFSETS[10]
1579              | 0x........ | OFFSETS[11]
1580              | 0x........ | OFFSETS[12]
1581              | 0x........ | OFFSETS[13]
1582              | 0x........ | OFFSETS[n_hashes]
1583              |------------|
1584              |            |
1585              |            |
1586              |            |
1587              |            |
1588              |            |
1589              |------------|
1590  0x000034f0: | 0x00001203 | .debug_str ("erase")
1591              | 0x00000004 | A 32 bit array count - number of HashData with name "erase"
1592              | 0x........ | HashData[0]
1593              | 0x........ | HashData[1]
1594              | 0x........ | HashData[2]
1595              | 0x........ | HashData[3]
1596              | 0x00000000 | String offset into .debug_str (terminate data for hash)
1597              |------------|
1598  0x00003500: | 0x00001203 | String offset into .debug_str ("collision")
1599              | 0x00000002 | A 32 bit array count - number of HashData with name "collision"
1600              | 0x........ | HashData[0]
1601              | 0x........ | HashData[1]
1602              | 0x00001203 | String offset into .debug_str ("dump")
1603              | 0x00000003 | A 32 bit array count - number of HashData with name "dump"
1604              | 0x........ | HashData[0]
1605              | 0x........ | HashData[1]
1606              | 0x........ | HashData[2]
1607              | 0x00000000 | String offset into .debug_str (terminate data for hash)
1608              |------------|
1609  0x00003550: | 0x00001203 | String offset into .debug_str ("main")
1610              | 0x00000009 | A 32 bit array count - number of HashData with name "main"
1611              | 0x........ | HashData[0]
1612              | 0x........ | HashData[1]
1613              | 0x........ | HashData[2]
1614              | 0x........ | HashData[3]
1615              | 0x........ | HashData[4]
1616              | 0x........ | HashData[5]
1617              | 0x........ | HashData[6]
1618              | 0x........ | HashData[7]
1619              | 0x........ | HashData[8]
1620              | 0x00000000 | String offset into .debug_str (terminate data for hash)
1621              `------------'
1622
1623So we still have all of the same data, we just organize it more efficiently for
1624debugger lookup.  If we repeat the same "``printf``" lookup from above, we
1625would hash "``printf``" and find it matches ``BUCKETS[3]`` by taking the 32 bit
1626hash value and modulo it by ``n_buckets``.  ``BUCKETS[3]`` contains "6" which
1627is the index into the ``HASHES`` table.  We would then compare any consecutive
162832 bit hashes values in the ``HASHES`` array as long as the hashes would be in
1629``BUCKETS[3]``.  We do this by verifying that each subsequent hash value modulo
1630``n_buckets`` is still 3.  In the case of a failed lookup we would access the
1631memory for ``BUCKETS[3]``, and then compare a few consecutive 32 bit hashes
1632before we know that we have no match.  We don't end up marching through
1633multiple words of memory and we really keep the number of processor data cache
1634lines being accessed as small as possible.
1635
1636The string hash that is used for these lookup tables is the Daniel J.
1637Bernstein hash which is also used in the ELF ``GNU_HASH`` sections.  It is a
1638very good hash for all kinds of names in programs with very few hash
1639collisions.
1640
1641Empty buckets are designated by using an invalid hash index of ``UINT32_MAX``.
1642
1643Details
1644^^^^^^^
1645
1646These name hash tables are designed to be generic where specializations of the
1647table get to define additional data that goes into the header ("``HeaderData``"),
1648how the string value is stored ("``KeyType``") and the content of the data for each
1649hash value.
1650
1651Header Layout
1652"""""""""""""
1653
1654The header has a fixed part, and the specialized part.  The exact format of the
1655header is:
1656
1657.. code-block:: c
1658
1659  struct Header
1660  {
1661    uint32_t   magic;           // 'HASH' magic value to allow endian detection
1662    uint16_t   version;         // Version number
1663    uint16_t   hash_function;   // The hash function enumeration that was used
1664    uint32_t   bucket_count;    // The number of buckets in this hash table
1665    uint32_t   hashes_count;    // The total number of unique hash values and hash data offsets in this table
1666    uint32_t   header_data_len; // The bytes to skip to get to the hash indexes (buckets) for correct alignment
1667                                // Specifically the length of the following HeaderData field - this does not
1668                                // include the size of the preceding fields
1669    HeaderData header_data;     // Implementation specific header data
1670  };
1671
1672The header starts with a 32 bit "``magic``" value which must be ``'HASH'``
1673encoded as an ASCII integer.  This allows the detection of the start of the
1674hash table and also allows the table's byte order to be determined so the table
1675can be correctly extracted.  The "``magic``" value is followed by a 16 bit
1676``version`` number which allows the table to be revised and modified in the
1677future.  The current version number is 1. ``hash_function`` is a ``uint16_t``
1678enumeration that specifies which hash function was used to produce this table.
1679The current values for the hash function enumerations include:
1680
1681.. code-block:: c
1682
1683  enum HashFunctionType
1684  {
1685    eHashFunctionDJB = 0u, // Daniel J Bernstein hash function
1686  };
1687
1688``bucket_count`` is a 32 bit unsigned integer that represents how many buckets
1689are in the ``BUCKETS`` array.  ``hashes_count`` is the number of unique 32 bit
1690hash values that are in the ``HASHES`` array, and is the same number of offsets
1691are contained in the ``OFFSETS`` array.  ``header_data_len`` specifies the size
1692in bytes of the ``HeaderData`` that is filled in by specialized versions of
1693this table.
1694
1695Fixed Lookup
1696""""""""""""
1697
1698The header is followed by the buckets, hashes, offsets, and hash value data.
1699
1700.. code-block:: c
1701
1702  struct FixedTable
1703  {
1704    uint32_t buckets[Header.bucket_count];  // An array of hash indexes into the "hashes[]" array below
1705    uint32_t hashes [Header.hashes_count];  // Every unique 32 bit hash for the entire table is in this table
1706    uint32_t offsets[Header.hashes_count];  // An offset that corresponds to each item in the "hashes[]" array above
1707  };
1708
1709``buckets`` is an array of 32 bit indexes into the ``hashes`` array.  The
1710``hashes`` array contains all of the 32 bit hash values for all names in the
1711hash table.  Each hash in the ``hashes`` table has an offset in the ``offsets``
1712array that points to the data for the hash value.
1713
1714This table setup makes it very easy to repurpose these tables to contain
1715different data, while keeping the lookup mechanism the same for all tables.
1716This layout also makes it possible to save the table to disk and map it in
1717later and do very efficient name lookups with little or no parsing.
1718
1719DWARF lookup tables can be implemented in a variety of ways and can store a lot
1720of information for each name.  We want to make the DWARF tables extensible and
1721able to store the data efficiently so we have used some of the DWARF features
1722that enable efficient data storage to define exactly what kind of data we store
1723for each name.
1724
1725The ``HeaderData`` contains a definition of the contents of each HashData chunk.
1726We might want to store an offset to all of the debug information entries (DIEs)
1727for each name.  To keep things extensible, we create a list of items, or
1728Atoms, that are contained in the data for each name.  First comes the type of
1729the data in each atom:
1730
1731.. code-block:: c
1732
1733  enum AtomType
1734  {
1735    eAtomTypeNULL       = 0u,
1736    eAtomTypeDIEOffset  = 1u,   // DIE offset, check form for encoding
1737    eAtomTypeCUOffset   = 2u,   // DIE offset of the compiler unit header that contains the item in question
1738    eAtomTypeTag        = 3u,   // DW_TAG_xxx value, should be encoded as DW_FORM_data1 (if no tags exceed 255) or DW_FORM_data2
1739    eAtomTypeNameFlags  = 4u,   // Flags from enum NameFlags
1740    eAtomTypeTypeFlags  = 5u,   // Flags from enum TypeFlags
1741  };
1742
1743The enumeration values and their meanings are:
1744
1745.. code-block:: none
1746
1747  eAtomTypeNULL       - a termination atom that specifies the end of the atom list
1748  eAtomTypeDIEOffset  - an offset into the .debug_info section for the DWARF DIE for this name
1749  eAtomTypeCUOffset   - an offset into the .debug_info section for the CU that contains the DIE
1750  eAtomTypeDIETag     - The DW_TAG_XXX enumeration value so you don't have to parse the DWARF to see what it is
1751  eAtomTypeNameFlags  - Flags for functions and global variables (isFunction, isInlined, isExternal...)
1752  eAtomTypeTypeFlags  - Flags for types (isCXXClass, isObjCClass, ...)
1753
1754Then we allow each atom type to define the atom type and how the data for each
1755atom type data is encoded:
1756
1757.. code-block:: c
1758
1759  struct Atom
1760  {
1761    uint16_t type;  // AtomType enum value
1762    uint16_t form;  // DWARF DW_FORM_XXX defines
1763  };
1764
1765The ``form`` type above is from the DWARF specification and defines the exact
1766encoding of the data for the Atom type.  See the DWARF specification for the
1767``DW_FORM_`` definitions.
1768
1769.. code-block:: c
1770
1771  struct HeaderData
1772  {
1773    uint32_t die_offset_base;
1774    uint32_t atom_count;
1775    Atoms    atoms[atom_count0];
1776  };
1777
1778``HeaderData`` defines the base DIE offset that should be added to any atoms
1779that are encoded using the ``DW_FORM_ref1``, ``DW_FORM_ref2``,
1780``DW_FORM_ref4``, ``DW_FORM_ref8`` or ``DW_FORM_ref_udata``.  It also defines
1781what is contained in each ``HashData`` object -- ``Atom.form`` tells us how large
1782each field will be in the ``HashData`` and the ``Atom.type`` tells us how this data
1783should be interpreted.
1784
1785For the current implementations of the "``.apple_names``" (all functions +
1786globals), the "``.apple_types``" (names of all types that are defined), and
1787the "``.apple_namespaces``" (all namespaces), we currently set the ``Atom``
1788array to be:
1789
1790.. code-block:: c
1791
1792  HeaderData.atom_count = 1;
1793  HeaderData.atoms[0].type = eAtomTypeDIEOffset;
1794  HeaderData.atoms[0].form = DW_FORM_data4;
1795
1796This defines the contents to be the DIE offset (eAtomTypeDIEOffset) that is
1797encoded as a 32 bit value (DW_FORM_data4).  This allows a single name to have
1798multiple matching DIEs in a single file, which could come up with an inlined
1799function for instance.  Future tables could include more information about the
1800DIE such as flags indicating if the DIE is a function, method, block,
1801or inlined.
1802
1803The KeyType for the DWARF table is a 32 bit string table offset into the
1804".debug_str" table.  The ".debug_str" is the string table for the DWARF which
1805may already contain copies of all of the strings.  This helps make sure, with
1806help from the compiler, that we reuse the strings between all of the DWARF
1807sections and keeps the hash table size down.  Another benefit to having the
1808compiler generate all strings as DW_FORM_strp in the debug info, is that
1809DWARF parsing can be made much faster.
1810
1811After a lookup is made, we get an offset into the hash data.  The hash data
1812needs to be able to deal with 32 bit hash collisions, so the chunk of data
1813at the offset in the hash data consists of a triple:
1814
1815.. code-block:: c
1816
1817  uint32_t str_offset
1818  uint32_t hash_data_count
1819  HashData[hash_data_count]
1820
1821If "str_offset" is zero, then the bucket contents are done. 99.9% of the
1822hash data chunks contain a single item (no 32 bit hash collision):
1823
1824.. code-block:: none
1825
1826  .------------.
1827  | 0x00001023 | uint32_t KeyType (.debug_str[0x0001023] => "main")
1828  | 0x00000004 | uint32_t HashData count
1829  | 0x........ | uint32_t HashData[0] DIE offset
1830  | 0x........ | uint32_t HashData[1] DIE offset
1831  | 0x........ | uint32_t HashData[2] DIE offset
1832  | 0x........ | uint32_t HashData[3] DIE offset
1833  | 0x00000000 | uint32_t KeyType (end of hash chain)
1834  `------------'
1835
1836If there are collisions, you will have multiple valid string offsets:
1837
1838.. code-block:: none
1839
1840  .------------.
1841  | 0x00001023 | uint32_t KeyType (.debug_str[0x0001023] => "main")
1842  | 0x00000004 | uint32_t HashData count
1843  | 0x........ | uint32_t HashData[0] DIE offset
1844  | 0x........ | uint32_t HashData[1] DIE offset
1845  | 0x........ | uint32_t HashData[2] DIE offset
1846  | 0x........ | uint32_t HashData[3] DIE offset
1847  | 0x00002023 | uint32_t KeyType (.debug_str[0x0002023] => "print")
1848  | 0x00000002 | uint32_t HashData count
1849  | 0x........ | uint32_t HashData[0] DIE offset
1850  | 0x........ | uint32_t HashData[1] DIE offset
1851  | 0x00000000 | uint32_t KeyType (end of hash chain)
1852  `------------'
1853
1854Current testing with real world C++ binaries has shown that there is around 1
185532 bit hash collision per 100,000 name entries.
1856
1857Contents
1858^^^^^^^^
1859
1860As we said, we want to strictly define exactly what is included in the
1861different tables.  For DWARF, we have 3 tables: "``.apple_names``",
1862"``.apple_types``", and "``.apple_namespaces``".
1863
1864"``.apple_names``" sections should contain an entry for each DWARF DIE whose
1865``DW_TAG`` is a ``DW_TAG_label``, ``DW_TAG_inlined_subroutine``, or
1866``DW_TAG_subprogram`` that has address attributes: ``DW_AT_low_pc``,
1867``DW_AT_high_pc``, ``DW_AT_ranges`` or ``DW_AT_entry_pc``.  It also contains
1868``DW_TAG_variable`` DIEs that have a ``DW_OP_addr`` in the location (global and
1869static variables).  All global and static variables should be included,
1870including those scoped within functions and classes.  For example using the
1871following code:
1872
1873.. code-block:: c
1874
1875  static int var = 0;
1876
1877  void f ()
1878  {
1879    static int var = 0;
1880  }
1881
1882Both of the static ``var`` variables would be included in the table.  All
1883functions should emit both their full names and their basenames.  For C or C++,
1884the full name is the mangled name (if available) which is usually in the
1885``DW_AT_MIPS_linkage_name`` attribute, and the ``DW_AT_name`` contains the
1886function basename.  If global or static variables have a mangled name in a
1887``DW_AT_MIPS_linkage_name`` attribute, this should be emitted along with the
1888simple name found in the ``DW_AT_name`` attribute.
1889
1890"``.apple_types``" sections should contain an entry for each DWARF DIE whose
1891tag is one of:
1892
1893* DW_TAG_array_type
1894* DW_TAG_class_type
1895* DW_TAG_enumeration_type
1896* DW_TAG_pointer_type
1897* DW_TAG_reference_type
1898* DW_TAG_string_type
1899* DW_TAG_structure_type
1900* DW_TAG_subroutine_type
1901* DW_TAG_typedef
1902* DW_TAG_union_type
1903* DW_TAG_ptr_to_member_type
1904* DW_TAG_set_type
1905* DW_TAG_subrange_type
1906* DW_TAG_base_type
1907* DW_TAG_const_type
1908* DW_TAG_immutable_type
1909* DW_TAG_file_type
1910* DW_TAG_namelist
1911* DW_TAG_packed_type
1912* DW_TAG_volatile_type
1913* DW_TAG_restrict_type
1914* DW_TAG_atomic_type
1915* DW_TAG_interface_type
1916* DW_TAG_unspecified_type
1917* DW_TAG_shared_type
1918
1919Only entries with a ``DW_AT_name`` attribute are included, and the entry must
1920not be a forward declaration (``DW_AT_declaration`` attribute with a non-zero
1921value).  For example, using the following code:
1922
1923.. code-block:: c
1924
1925  int main ()
1926  {
1927    int *b = 0;
1928    return *b;
1929  }
1930
1931We get a few type DIEs:
1932
1933.. code-block:: none
1934
1935  0x00000067:     TAG_base_type [5]
1936                  AT_encoding( DW_ATE_signed )
1937                  AT_name( "int" )
1938                  AT_byte_size( 0x04 )
1939
1940  0x0000006e:     TAG_pointer_type [6]
1941                  AT_type( {0x00000067} ( int ) )
1942                  AT_byte_size( 0x08 )
1943
1944The DW_TAG_pointer_type is not included because it does not have a ``DW_AT_name``.
1945
1946"``.apple_namespaces``" section should contain all ``DW_TAG_namespace`` DIEs.
1947If we run into a namespace that has no name this is an anonymous namespace, and
1948the name should be output as "``(anonymous namespace)``" (without the quotes).
1949Why?  This matches the output of the ``abi::cxa_demangle()`` that is in the
1950standard C++ library that demangles mangled names.
1951
1952
1953Language Extensions and File Format Changes
1954^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1955
1956Objective-C Extensions
1957""""""""""""""""""""""
1958
1959"``.apple_objc``" section should contain all ``DW_TAG_subprogram`` DIEs for an
1960Objective-C class.  The name used in the hash table is the name of the
1961Objective-C class itself.  If the Objective-C class has a category, then an
1962entry is made for both the class name without the category, and for the class
1963name with the category.  So if we have a DIE at offset 0x1234 with a name of
1964method "``-[NSString(my_additions) stringWithSpecialString:]``", we would add
1965an entry for "``NSString``" that points to DIE 0x1234, and an entry for
1966"``NSString(my_additions)``" that points to 0x1234.  This allows us to quickly
1967track down all Objective-C methods for an Objective-C class when doing
1968expressions.  It is needed because of the dynamic nature of Objective-C where
1969anyone can add methods to a class.  The DWARF for Objective-C methods is also
1970emitted differently from C++ classes where the methods are not usually
1971contained in the class definition, they are scattered about across one or more
1972compile units.  Categories can also be defined in different shared libraries.
1973So we need to be able to quickly find all of the methods and class functions
1974given the Objective-C class name, or quickly find all methods and class
1975functions for a class + category name.  This table does not contain any
1976selector names, it just maps Objective-C class names (or class names +
1977category) to all of the methods and class functions.  The selectors are added
1978as function basenames in the "``.debug_names``" section.
1979
1980In the "``.apple_names``" section for Objective-C functions, the full name is
1981the entire function name with the brackets ("``-[NSString
1982stringWithCString:]``") and the basename is the selector only
1983("``stringWithCString:``").
1984
1985Mach-O Changes
1986""""""""""""""
1987
1988The sections names for the apple hash tables are for non-mach-o files.  For
1989mach-o files, the sections should be contained in the ``__DWARF`` segment with
1990names as follows:
1991
1992* "``.apple_names``" -> "``__apple_names``"
1993* "``.apple_types``" -> "``__apple_types``"
1994* "``.apple_namespaces``" -> "``__apple_namespac``" (16 character limit)
1995* "``.apple_objc``" -> "``__apple_objc``"
1996
1997.. _codeview:
1998
1999CodeView Debug Info Format
2000==========================
2001
2002LLVM supports emitting CodeView, the Microsoft debug info format, and this
2003section describes the design and implementation of that support.
2004
2005Format Background
2006-----------------
2007
2008CodeView as a format is clearly oriented around C++ debugging, and in C++, the
2009majority of debug information tends to be type information. Therefore, the
2010overriding design constraint of CodeView is the separation of type information
2011from other "symbol" information so that type information can be efficiently
2012merged across translation units. Both type information and symbol information is
2013generally stored as a sequence of records, where each record begins with a
201416-bit record size and a 16-bit record kind.
2015
2016Type information is usually stored in the ``.debug$T`` section of the object
2017file.  All other debug info, such as line info, string table, symbol info, and
2018inlinee info, is stored in one or more ``.debug$S`` sections. There may only be
2019one ``.debug$T`` section per object file, since all other debug info refers to
2020it. If a PDB (enabled by the ``/Zi`` MSVC option) was used during compilation,
2021the ``.debug$T`` section will contain only an ``LF_TYPESERVER2`` record pointing
2022to the PDB. When using PDBs, symbol information appears to remain in the object
2023file ``.debug$S`` sections.
2024
2025Type records are referred to by their index, which is the number of records in
2026the stream before a given record plus ``0x1000``. Many common basic types, such
2027as the basic integral types and unqualified pointers to them, are represented
2028using type indices less than ``0x1000``. Such basic types are built in to
2029CodeView consumers and do not require type records.
2030
2031Each type record may only contain type indices that are less than its own type
2032index. This ensures that the graph of type stream references is acyclic. While
2033the source-level type graph may contain cycles through pointer types (consider a
2034linked list struct), these cycles are removed from the type stream by always
2035referring to the forward declaration record of user-defined record types. Only
2036"symbol" records in the ``.debug$S`` streams may refer to complete,
2037non-forward-declaration type records.
2038
2039Working with CodeView
2040---------------------
2041
2042These are instructions for some common tasks for developers working to improve
2043LLVM's CodeView support. Most of them revolve around using the CodeView dumper
2044embedded in ``llvm-readobj``.
2045
2046* Testing MSVC's output::
2047
2048    $ cl -c -Z7 foo.cpp # Use /Z7 to keep types in the object file
2049    $ llvm-readobj --codeview foo.obj
2050
2051* Getting LLVM IR debug info out of Clang::
2052
2053    $ clang -g -gcodeview --target=x86_64-windows-msvc foo.cpp -S -emit-llvm
2054
2055  Use this to generate LLVM IR for LLVM test cases.
2056
2057* Generate and dump CodeView from LLVM IR metadata::
2058
2059    $ llc foo.ll -filetype=obj -o foo.obj
2060    $ llvm-readobj --codeview foo.obj > foo.txt
2061
2062  Use this pattern in lit test cases and FileCheck the output of llvm-readobj
2063
2064Improving LLVM's CodeView support is a process of finding interesting type
2065records, constructing a C++ test case that makes MSVC emit those records,
2066dumping the records, understanding them, and then generating equivalent records
2067in LLVM's backend.
2068