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.declare`` 179^^^^^^^^^^^^^^^^^^^^ 180 181.. code-block:: llvm 182 183 void @llvm.dbg.declare(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.declare` 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.declare(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.declare(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.declare`` 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 multiple 211calls to `llvm.dbg.value`. Passes that delete stores are effectively partial 212promotion, and they will insert a mix of calls to ``llvm.dbg.value`` to track 213the source variable value when it is available. After optimization, there may be 214multiple calls to ``llvm.dbg.declare`` describing the program points where the 215variables lives in memory. All calls for the same concrete source variable must 216agree on the memory location. 217 218 219``llvm.dbg.value`` 220^^^^^^^^^^^^^^^^^^ 221 222.. code-block:: llvm 223 224 void @llvm.dbg.value(metadata, metadata, metadata) 225 226This intrinsic provides information when a user source variable is set to a new 227value. The first argument is the new value (wrapped as metadata). The second 228argument is a `local variable <LangRef.html#dilocalvariable>`_ containing a 229description of the variable. The third argument is a `complex expression 230<LangRef.html#diexpression>`_. 231 232An `llvm.dbg.value` intrinsic describes the *value* of a source variable 233directly, not its address. Note that the value operand of this intrinsic may 234be indirect (i.e, a pointer to the source variable), provided that interpreting 235the complex expression derives the direct value. 236 237``llvm.dbg.assign`` 238^^^^^^^^^^^^^^^^^^^ 239.. toctree:: 240 :hidden: 241 242 AssignmentTracking 243 244.. code-block:: llvm 245 246 void @llvm.dbg.assign(Value *Value, 247 DIExpression *ValueExpression, 248 DILocalVariable *Variable, 249 DIAssignID *ID, 250 Value *Address, 251 DIExpression *AddressExpression) 252 253This intrinsic marks the position in IR where a source assignment occured. It 254encodes the value of the variable. It references the store, if any, that 255performs the assignment, and the destination address. 256 257The first three arguments are the same as for an ``llvm.dbg.value``. The fourth 258argument is a ``DIAssignID`` used to reference a store. The fifth is the 259destination of the store (wrapped as metadata), and the sixth is a `complex 260expression <LangRef.html#diexpression>`_ that modfies it. 261 262The formal LLVM-IR signature is: 263 264.. code-block:: llvm 265 266 void @llvm.dbg.assign(metadata, metadata, metadata, metadata, metadata, metadata) 267 268 269See :doc:`AssignmentTracking` for more info. 270 271Object lifetimes and scoping 272============================ 273 274In many languages, the local variables in functions can have their lifetimes or 275scopes limited to a subset of a function. In the C family of languages, for 276example, variables are only live (readable and writable) within the source 277block that they are defined in. In functional languages, values are only 278readable after they have been defined. Though this is a very obvious concept, 279it is non-trivial to model in LLVM, because it has no notion of scoping in this 280sense, and does not want to be tied to a language's scoping rules. 281 282In order to handle this, the LLVM debug format uses the metadata attached to 283llvm instructions to encode line number and scoping information. Consider the 284following C fragment, for example: 285 286.. code-block:: c 287 288 1. void foo() { 289 2. int X = 21; 290 3. int Y = 22; 291 4. { 292 5. int Z = 23; 293 6. Z = X; 294 7. } 295 8. X = Y; 296 9. } 297 298Compiled to LLVM, this function would be represented like this: 299 300.. code-block:: text 301 302 ; Function Attrs: nounwind ssp uwtable 303 define void @foo() #0 !dbg !4 { 304 entry: 305 %X = alloca i32, align 4 306 %Y = alloca i32, align 4 307 %Z = alloca i32, align 4 308 call void @llvm.dbg.declare(metadata i32* %X, metadata !11, metadata !13), !dbg !14 309 store i32 21, i32* %X, align 4, !dbg !14 310 call void @llvm.dbg.declare(metadata i32* %Y, metadata !15, metadata !13), !dbg !16 311 store i32 22, i32* %Y, align 4, !dbg !16 312 call void @llvm.dbg.declare(metadata i32* %Z, metadata !17, metadata !13), !dbg !19 313 store i32 23, i32* %Z, align 4, !dbg !19 314 %0 = load i32, i32* %X, align 4, !dbg !20 315 store i32 %0, i32* %Z, align 4, !dbg !21 316 %1 = load i32, i32* %Y, align 4, !dbg !22 317 store i32 %1, i32* %X, align 4, !dbg !23 318 ret void, !dbg !24 319 } 320 321 ; Function Attrs: nounwind readnone 322 declare void @llvm.dbg.declare(metadata, metadata, metadata) #1 323 324 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" } 325 attributes #1 = { nounwind readnone } 326 327 !llvm.dbg.cu = !{!0} 328 !llvm.module.flags = !{!7, !8, !9} 329 !llvm.ident = !{!10} 330 331 !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) 332 !1 = !DIFile(filename: "/dev/stdin", directory: "/Users/dexonsmith/data/llvm/debug-info") 333 !2 = !{} 334 !3 = !{!4} 335 !4 = distinct !DISubprogram(name: "foo", scope: !1, file: !1, line: 1, type: !5, isLocal: false, isDefinition: true, scopeLine: 1, isOptimized: false, retainedNodes: !2) 336 !5 = !DISubroutineType(types: !6) 337 !6 = !{null} 338 !7 = !{i32 2, !"Dwarf Version", i32 2} 339 !8 = !{i32 2, !"Debug Info Version", i32 3} 340 !9 = !{i32 1, !"PIC Level", i32 2} 341 !10 = !{!"clang version 3.7.0 (trunk 231150) (llvm/trunk 231154)"} 342 !11 = !DILocalVariable(name: "X", scope: !4, file: !1, line: 2, type: !12) 343 !12 = !DIBasicType(name: "int", size: 32, align: 32, encoding: DW_ATE_signed) 344 !13 = !DIExpression() 345 !14 = !DILocation(line: 2, column: 9, scope: !4) 346 !15 = !DILocalVariable(name: "Y", scope: !4, file: !1, line: 3, type: !12) 347 !16 = !DILocation(line: 3, column: 9, scope: !4) 348 !17 = !DILocalVariable(name: "Z", scope: !18, file: !1, line: 5, type: !12) 349 !18 = distinct !DILexicalBlock(scope: !4, file: !1, line: 4, column: 5) 350 !19 = !DILocation(line: 5, column: 11, scope: !18) 351 !20 = !DILocation(line: 6, column: 11, scope: !18) 352 !21 = !DILocation(line: 6, column: 9, scope: !18) 353 !22 = !DILocation(line: 8, column: 9, scope: !4) 354 !23 = !DILocation(line: 8, column: 7, scope: !4) 355 !24 = !DILocation(line: 9, column: 3, scope: !4) 356 357 358This example illustrates a few important details about LLVM debugging 359information. In particular, it shows how the ``llvm.dbg.declare`` intrinsic and 360location information, which are attached to an instruction, are applied 361together to allow a debugger to analyze the relationship between statements, 362variable definitions, and the code used to implement the function. 363 364.. code-block:: llvm 365 366 call void @llvm.dbg.declare(metadata i32* %X, metadata !11, metadata !13), !dbg !14 367 ; [debug line = 2:7] [debug variable = X] 368 369The first intrinsic ``%llvm.dbg.declare`` encodes debugging information for the 370variable ``X``. The metadata ``!dbg !14`` attached to the intrinsic provides 371scope information for the variable ``X``. 372 373.. code-block:: text 374 375 !14 = !DILocation(line: 2, column: 9, scope: !4) 376 !4 = distinct !DISubprogram(name: "foo", scope: !1, file: !1, line: 1, type: !5, 377 isLocal: false, isDefinition: true, scopeLine: 1, 378 isOptimized: false, retainedNodes: !2) 379 380Here ``!14`` is metadata providing `location information 381<LangRef.html#dilocation>`_. In this example, scope is encoded by ``!4``, a 382`subprogram descriptor <LangRef.html#disubprogram>`_. This way the location 383information attached to the intrinsics indicates that the variable ``X`` is 384declared at line number 2 at a function level scope in function ``foo``. 385 386Now lets take another example. 387 388.. code-block:: llvm 389 390 call void @llvm.dbg.declare(metadata i32* %Z, metadata !17, metadata !13), !dbg !19 391 ; [debug line = 5:9] [debug variable = Z] 392 393The third intrinsic ``%llvm.dbg.declare`` encodes debugging information for 394variable ``Z``. The metadata ``!dbg !19`` attached to the intrinsic provides 395scope information for the variable ``Z``. 396 397.. code-block:: text 398 399 !18 = distinct !DILexicalBlock(scope: !4, file: !1, line: 4, column: 5) 400 !19 = !DILocation(line: 5, column: 11, scope: !18) 401 402Here ``!19`` indicates that ``Z`` is declared at line number 5 and column 403number 11 inside of lexical scope ``!18``. The lexical scope itself resides 404inside of subprogram ``!4`` described above. 405 406The scope information attached with each instruction provides a straightforward 407way to find instructions covered by a scope. 408 409Object lifetime in optimized code 410================================= 411 412In the example above, every variable assignment uniquely corresponds to a 413memory store to the variable's position on the stack. However in heavily 414optimized code LLVM promotes most variables into SSA values, which can 415eventually be placed in physical registers or memory locations. To track SSA 416values through compilation, when objects are promoted to SSA values an 417``llvm.dbg.value`` intrinsic is created for each assignment, recording the 418variable's new location. Compared with the ``llvm.dbg.declare`` intrinsic: 419 420* A dbg.value terminates the effect of any preceding dbg.values for (any 421 overlapping fragments of) the specified variable. 422* The dbg.value's position in the IR defines where in the instruction stream 423 the variable's value changes. 424* Operands can be constants, indicating the variable is assigned a 425 constant value. 426 427Care must be taken to update ``llvm.dbg.value`` intrinsics when optimization 428passes alter or move instructions and blocks -- the developer could observe such 429changes reflected in the value of variables when debugging the program. For any 430execution of the optimized program, the set of variable values presented to the 431developer by the debugger should not show a state that would never have existed 432in the execution of the unoptimized program, given the same input. Doing so 433risks misleading the developer by reporting a state that does not exist, 434damaging their understanding of the optimized program and undermining their 435trust in the debugger. 436 437Sometimes perfectly preserving variable locations is not possible, often when a 438redundant calculation is optimized out. In such cases, a ``llvm.dbg.value`` 439with operand ``undef`` should be used, to terminate earlier variable locations 440and let the debugger present ``optimized out`` to the developer. Withholding 441these potentially stale variable values from the developer diminishes the 442amount of available debug information, but increases the reliability of the 443remaining information. 444 445To illustrate some potential issues, consider the following example: 446 447.. code-block:: llvm 448 449 define i32 @foo(i32 %bar, i1 %cond) { 450 entry: 451 call @llvm.dbg.value(metadata i32 0, metadata !1, metadata !2) 452 br i1 %cond, label %truebr, label %falsebr 453 truebr: 454 %tval = add i32 %bar, 1 455 call @llvm.dbg.value(metadata i32 %tval, metadata !1, metadata !2) 456 %g1 = call i32 @gazonk() 457 br label %exit 458 falsebr: 459 %fval = add i32 %bar, 2 460 call @llvm.dbg.value(metadata i32 %fval, metadata !1, metadata !2) 461 %g2 = call i32 @gazonk() 462 br label %exit 463 exit: 464 %merge = phi [ %tval, %truebr ], [ %fval, %falsebr ] 465 %g = phi [ %g1, %truebr ], [ %g2, %falsebr ] 466 call @llvm.dbg.value(metadata i32 %merge, metadata !1, metadata !2) 467 call @llvm.dbg.value(metadata i32 %g, metadata !3, metadata !2) 468 %plusten = add i32 %merge, 10 469 %toret = add i32 %plusten, %g 470 call @llvm.dbg.value(metadata i32 %toret, metadata !1, metadata !2) 471 ret i32 %toret 472 } 473 474Containing two source-level variables in ``!1`` and ``!3``. The function could, 475perhaps, be optimized into the following code: 476 477.. code-block:: llvm 478 479 define i32 @foo(i32 %bar, i1 %cond) { 480 entry: 481 %g = call i32 @gazonk() 482 %addoper = select i1 %cond, i32 11, i32 12 483 %plusten = add i32 %bar, %addoper 484 %toret = add i32 %plusten, %g 485 ret i32 %toret 486 } 487 488What ``llvm.dbg.value`` intrinsics should be placed to represent the original variable 489locations in this code? Unfortunately the second, third and fourth 490dbg.values for ``!1`` in the source function have had their operands 491(%tval, %fval, %merge) optimized out. Assuming we cannot recover them, we 492might consider this placement of dbg.values: 493 494.. code-block:: llvm 495 496 define i32 @foo(i32 %bar, i1 %cond) { 497 entry: 498 call @llvm.dbg.value(metadata i32 0, metadata !1, metadata !2) 499 %g = call i32 @gazonk() 500 call @llvm.dbg.value(metadata i32 %g, metadata !3, metadata !2) 501 %addoper = select i1 %cond, i32 11, i32 12 502 %plusten = add i32 %bar, %addoper 503 %toret = add i32 %plusten, %g 504 call @llvm.dbg.value(metadata i32 %toret, metadata !1, metadata !2) 505 ret i32 %toret 506 } 507 508However, this will cause ``!3`` to have the return value of ``@gazonk()`` at 509the same time as ``!1`` has the constant value zero -- a pair of assignments 510that never occurred in the unoptimized program. To avoid this, we must terminate 511the range that ``!1`` has the constant value assignment by inserting an undef 512dbg.value before the dbg.value for ``!3``: 513 514.. code-block:: llvm 515 516 define i32 @foo(i32 %bar, i1 %cond) { 517 entry: 518 call @llvm.dbg.value(metadata i32 0, metadata !1, metadata !2) 519 %g = call i32 @gazonk() 520 call @llvm.dbg.value(metadata i32 undef, metadata !1, metadata !2) 521 call @llvm.dbg.value(metadata i32 %g, metadata !3, metadata !2) 522 %addoper = select i1 %cond, i32 11, i32 12 523 %plusten = add i32 %bar, %addoper 524 %toret = add i32 %plusten, %g 525 call @llvm.dbg.value(metadata i32 %toret, metadata !1, metadata !2) 526 ret i32 %toret 527 } 528 529In general, if any dbg.value has its operand optimized out and cannot be 530recovered, then an undef dbg.value is necessary to terminate earlier variable 531locations. Additional undef dbg.values may be necessary when the debugger can 532observe re-ordering of assignments. 533 534How variable location metadata is transformed during CodeGen 535============================================================ 536 537LLVM preserves debug information throughout mid-level and backend passes, 538ultimately producing a mapping between source-level information and 539instruction ranges. This 540is relatively straightforwards for line number information, as mapping 541instructions to line numbers is a simple association. For variable locations 542however the story is more complex. As each ``llvm.dbg.value`` intrinsic 543represents a source-level assignment of a value to a source variable, the 544variable location intrinsics effectively embed a small imperative program 545within the LLVM IR. By the end of CodeGen, this becomes a mapping from each 546variable to their machine locations over ranges of instructions. 547From IR to object emission, the major transformations which affect variable 548location fidelity are: 549 5501. Instruction Selection 5512. Register allocation 5523. Block layout 553 554each of which are discussed below. In addition, instruction scheduling can 555significantly change the ordering of the program, and occurs in a number of 556different passes. 557 558Some variable locations are not transformed during CodeGen. Stack locations 559specified by ``llvm.dbg.declare`` are valid and unchanging for the entire 560duration of the function, and are recorded in a simple MachineFunction table. 561Location changes in the prologue and epilogue of a function are also ignored: 562frame setup and destruction may take several instructions, require a 563disproportionate amount of debugging information in the output binary to 564describe, and should be stepped over by debuggers anyway. 565 566Variable locations in Instruction Selection and MIR 567--------------------------------------------------- 568 569Instruction selection creates a MIR function from an IR function, and just as 570it transforms ``intermediate`` instructions into machine instructions, so must 571``intermediate`` variable locations become machine variable locations. 572Within IR, variable locations are always identified by a Value, but in MIR 573there can be different types of variable locations. In addition, some IR 574locations become unavailable, for example if the operation of multiple IR 575instructions are combined into one machine instruction (such as 576multiply-and-accumulate) then intermediate Values are lost. To track variable 577locations through instruction selection, they are first separated into 578locations that do not depend on code generation (constants, stack locations, 579allocated virtual registers) and those that do. For those that do, debug 580metadata is attached to SDNodes in SelectionDAGs. After instruction selection 581has occurred and a MIR function is created, if the SDNode associated with debug 582metadata is allocated a virtual register, that virtual register is used as the 583variable location. If the SDNode is folded into a machine instruction or 584otherwise transformed into a non-register, the variable location becomes 585unavailable. 586 587Locations that are unavailable are treated as if they have been optimized out: 588in IR the location would be assigned ``undef`` by a debug intrinsic, and in MIR 589the equivalent location is used. 590 591After MIR locations are assigned to each variable, machine pseudo-instructions 592corresponding to each ``llvm.dbg.value`` intrinsic are inserted. There are two 593forms of this type of instruction. 594 595The first form, ``DBG_VALUE``, appears thus: 596 597.. code-block:: text 598 599 DBG_VALUE %1, $noreg, !123, !DIExpression() 600 601And has the following operands: 602 * The first operand can record the variable location as a register, 603 a frame index, an immediate, or the base address register if the original 604 debug intrinsic referred to memory. ``$noreg`` indicates the variable 605 location is undefined, equivalent to an ``undef`` dbg.value operand. 606 * The type of the second operand indicates whether the variable location is 607 directly referred to by the DBG_VALUE, or whether it is indirect. The 608 ``$noreg`` register signifies the former, an immediate operand (0) the 609 latter. 610 * Operand 3 is the Variable field of the original debug intrinsic. 611 * Operand 4 is the Expression field of the original debug intrinsic. 612 613The second form, ``DBG_VALUE_LIST``, appears thus: 614 615.. code-block:: text 616 617 DBG_VALUE_LIST !123, !DIExpression(DW_OP_LLVM_arg, 0, DW_OP_LLVM_arg, 1, DW_OP_plus), %1, %2 618 619And has the following operands: 620 * The first operand is the Variable field of the original debug intrinsic. 621 * The second operand is the Expression field of the original debug intrinsic. 622 * Any number of operands, from the 3rd onwards, record a sequence of variable 623 location operands, which may take any of the same values as the first 624 operand of the ``DBG_VALUE`` instruction above. These variable location 625 operands are inserted into the final DWARF Expression in positions indicated 626 by the DW_OP_LLVM_arg operator in the `DIExpression 627 <LangRef.html#diexpression>`. 628 629The position at which the DBG_VALUEs are inserted should correspond to the 630positions of their matching ``llvm.dbg.value`` intrinsics in the IR block. As 631with optimization, LLVM aims to preserve the order in which variable 632assignments occurred in the source program. However SelectionDAG performs some 633instruction scheduling, which can reorder assignments (discussed below). 634Function parameter locations are moved to the beginning of the function if 635they're not already, to ensure they're immediately available on function entry. 636 637To demonstrate variable locations during instruction selection, consider 638the following example: 639 640.. code-block:: llvm 641 642 define i32 @foo(i32* %addr) { 643 entry: 644 call void @llvm.dbg.value(metadata i32 0, metadata !3, metadata !DIExpression()), !dbg !5 645 br label %bb1, !dbg !5 646 647 bb1: ; preds = %bb1, %entry 648 %bar.0 = phi i32 [ 0, %entry ], [ %add, %bb1 ] 649 call void @llvm.dbg.value(metadata i32 %bar.0, metadata !3, metadata !DIExpression()), !dbg !5 650 %addr1 = getelementptr i32, i32 *%addr, i32 1, !dbg !5 651 call void @llvm.dbg.value(metadata i32 *%addr1, metadata !3, metadata !DIExpression()), !dbg !5 652 %loaded1 = load i32, i32* %addr1, !dbg !5 653 %addr2 = getelementptr i32, i32 *%addr, i32 %bar.0, !dbg !5 654 call void @llvm.dbg.value(metadata i32 *%addr2, metadata !3, metadata !DIExpression()), !dbg !5 655 %loaded2 = load i32, i32* %addr2, !dbg !5 656 %add = add i32 %bar.0, 1, !dbg !5 657 call void @llvm.dbg.value(metadata i32 %add, metadata !3, metadata !DIExpression()), !dbg !5 658 %added = add i32 %loaded1, %loaded2 659 %cond = icmp ult i32 %added, %bar.0, !dbg !5 660 br i1 %cond, label %bb1, label %bb2, !dbg !5 661 662 bb2: ; preds = %bb1 663 ret i32 0, !dbg !5 664 } 665 666If one compiles this IR with ``llc -o - -start-after=codegen-prepare -stop-after=expand-isel-pseudos -mtriple=x86_64--``, the following MIR is produced: 667 668.. code-block:: text 669 670 bb.0.entry: 671 successors: %bb.1(0x80000000) 672 liveins: $rdi 673 674 %2:gr64 = COPY $rdi 675 %3:gr32 = MOV32r0 implicit-def dead $eflags 676 DBG_VALUE 0, $noreg, !3, !DIExpression(), debug-location !5 677 678 bb.1.bb1: 679 successors: %bb.1(0x7c000000), %bb.2(0x04000000) 680 681 %0:gr32 = PHI %3, %bb.0, %1, %bb.1 682 DBG_VALUE %0, $noreg, !3, !DIExpression(), debug-location !5 683 DBG_VALUE %2, $noreg, !3, !DIExpression(DW_OP_plus_uconst, 4, DW_OP_stack_value), debug-location !5 684 %4:gr32 = MOV32rm %2, 1, $noreg, 4, $noreg, debug-location !5 :: (load 4 from %ir.addr1) 685 %5:gr64_nosp = MOVSX64rr32 %0, debug-location !5 686 DBG_VALUE $noreg, $noreg, !3, !DIExpression(), debug-location !5 687 %1:gr32 = INC32r %0, implicit-def dead $eflags, debug-location !5 688 DBG_VALUE %1, $noreg, !3, !DIExpression(), debug-location !5 689 %6:gr32 = ADD32rm %4, %2, 4, killed %5, 0, $noreg, implicit-def dead $eflags :: (load 4 from %ir.addr2) 690 %7:gr32 = SUB32rr %6, %0, implicit-def $eflags, debug-location !5 691 JB_1 %bb.1, implicit $eflags, debug-location !5 692 JMP_1 %bb.2, debug-location !5 693 694 bb.2.bb2: 695 %8:gr32 = MOV32r0 implicit-def dead $eflags 696 $eax = COPY %8, debug-location !5 697 RET 0, $eax, debug-location !5 698 699Observe first that there is a DBG_VALUE instruction for every ``llvm.dbg.value`` 700intrinsic in the source IR, ensuring no source level assignments go missing. 701Then consider the different ways in which variable locations have been recorded: 702 703* For the first dbg.value an immediate operand is used to record a zero value. 704* The dbg.value of the PHI instruction leads to a DBG_VALUE of virtual register 705 ``%0``. 706* The first GEP has its effect folded into the first load instruction 707 (as a 4-byte offset), but the variable location is salvaged by folding 708 the GEPs effect into the DIExpression. 709* The second GEP is also folded into the corresponding load. However, it is 710 insufficiently simple to be salvaged, and is emitted as a ``$noreg`` 711 DBG_VALUE, indicating that the variable takes on an undefined location. 712* The final dbg.value has its Value placed in virtual register ``%1``. 713 714Instruction Scheduling 715---------------------- 716 717A number of passes can reschedule instructions, notably instruction selection 718and the pre-and-post RA machine schedulers. Instruction scheduling can 719significantly change the nature of the program -- in the (very unlikely) worst 720case the instruction sequence could be completely reversed. In such 721circumstances LLVM follows the principle applied to optimizations, that it is 722better for the debugger not to display any state than a misleading state. 723Thus, whenever instructions are advanced in order of execution, any 724corresponding DBG_VALUE is kept in its original position, and if an instruction 725is delayed then the variable is given an undefined location for the duration 726of the delay. To illustrate, consider this pseudo-MIR: 727 728.. code-block:: text 729 730 %1:gr32 = MOV32rm %0, 1, $noreg, 4, $noreg, debug-location !5 :: (load 4 from %ir.addr1) 731 DBG_VALUE %1, $noreg, !1, !2 732 %4:gr32 = ADD32rr %3, %2, implicit-def dead $eflags 733 DBG_VALUE %4, $noreg, !3, !4 734 %7:gr32 = SUB32rr %6, %5, implicit-def dead $eflags 735 DBG_VALUE %7, $noreg, !5, !6 736 737Imagine that the SUB32rr were moved forward to give us the following MIR: 738 739.. code-block:: text 740 741 %7:gr32 = SUB32rr %6, %5, implicit-def dead $eflags 742 %1:gr32 = MOV32rm %0, 1, $noreg, 4, $noreg, debug-location !5 :: (load 4 from %ir.addr1) 743 DBG_VALUE %1, $noreg, !1, !2 744 %4:gr32 = ADD32rr %3, %2, implicit-def dead $eflags 745 DBG_VALUE %4, $noreg, !3, !4 746 DBG_VALUE %7, $noreg, !5, !6 747 748In this circumstance LLVM would leave the MIR as shown above. Were we to move 749the DBG_VALUE of virtual register %7 upwards with the SUB32rr, we would re-order 750assignments and introduce a new state of the program. Whereas with the solution 751above, the debugger will see one fewer combination of variable values, because 752``!3`` and ``!5`` will change value at the same time. This is preferred over 753misrepresenting the original program. 754 755In comparison, if one sunk the MOV32rm, LLVM would produce the following: 756 757.. code-block:: text 758 759 DBG_VALUE $noreg, $noreg, !1, !2 760 %4:gr32 = ADD32rr %3, %2, implicit-def dead $eflags 761 DBG_VALUE %4, $noreg, !3, !4 762 %7:gr32 = SUB32rr %6, %5, implicit-def dead $eflags 763 DBG_VALUE %7, $noreg, !5, !6 764 %1:gr32 = MOV32rm %0, 1, $noreg, 4, $noreg, debug-location !5 :: (load 4 from %ir.addr1) 765 DBG_VALUE %1, $noreg, !1, !2 766 767Here, to avoid presenting a state in which the first assignment to ``!1`` 768disappears, the DBG_VALUE at the top of the block assigns the variable the 769undefined location, until its value is available at the end of the block where 770an additional DBG_VALUE is added. Were any other DBG_VALUE for ``!1`` to occur 771in the instructions that the MOV32rm was sunk past, the DBG_VALUE for ``%1`` 772would be dropped and the debugger would never observe it in the variable. This 773accurately reflects that the value is not available during the corresponding 774portion of the original program. 775 776Variable locations during Register Allocation 777--------------------------------------------- 778 779To avoid debug instructions interfering with the register allocator, the 780LiveDebugVariables pass extracts variable locations from a MIR function and 781deletes the corresponding DBG_VALUE instructions. Some localized copy 782propagation is performed within blocks. After register allocation, the 783VirtRegRewriter pass re-inserts DBG_VALUE instructions in their original 784positions, translating virtual register references into their physical 785machine locations. To avoid encoding incorrect variable locations, in this 786pass any DBG_VALUE of a virtual register that is not live, is replaced by 787the undefined location. The LiveDebugVariables may insert redundant DBG_VALUEs 788because of virtual register rewriting. These will be subsequently removed by 789the RemoveRedundantDebugValues pass. 790 791LiveDebugValues expansion of variable locations 792----------------------------------------------- 793 794After all optimizations have run and shortly before emission, the 795LiveDebugValues pass runs to achieve two aims: 796 797* To propagate the location of variables through copies and register spills, 798* For every block, to record every valid variable location in that block. 799 800After this pass the DBG_VALUE instruction changes meaning: rather than 801corresponding to a source-level assignment where the variable may change value, 802it asserts the location of a variable in a block, and loses effect outside the 803block. Propagating variable locations through copies and spills is 804straightforwards: determining the variable location in every basic block 805requires the consideration of control flow. Consider the following IR, which 806presents several difficulties: 807 808.. code-block:: text 809 810 define dso_local i32 @foo(i1 %cond, i32 %input) !dbg !12 { 811 entry: 812 br i1 %cond, label %truebr, label %falsebr 813 814 bb1: 815 %value = phi i32 [ %value1, %truebr ], [ %value2, %falsebr ] 816 br label %exit, !dbg !26 817 818 truebr: 819 call void @llvm.dbg.value(metadata i32 %input, metadata !30, metadata !DIExpression()), !dbg !24 820 call void @llvm.dbg.value(metadata i32 1, metadata !23, metadata !DIExpression()), !dbg !24 821 %value1 = add i32 %input, 1 822 br label %bb1 823 824 falsebr: 825 call void @llvm.dbg.value(metadata i32 %input, metadata !30, metadata !DIExpression()), !dbg !24 826 call void @llvm.dbg.value(metadata i32 2, metadata !23, metadata !DIExpression()), !dbg !24 827 %value = add i32 %input, 2 828 br label %bb1 829 830 exit: 831 ret i32 %value, !dbg !30 832 } 833 834Here the difficulties are: 835 836* The control flow is roughly the opposite of basic block order 837* The value of the ``!23`` variable merges into ``%bb1``, but there is no PHI 838 node 839 840As mentioned above, the ``llvm.dbg.value`` intrinsics essentially form an 841imperative program embedded in the IR, with each intrinsic defining a variable 842location. This *could* be converted to an SSA form by mem2reg, in the same way 843that it uses use-def chains to identify control flow merges and insert phi 844nodes for IR Values. However, because debug variable locations are defined for 845every machine instruction, in effect every IR instruction uses every variable 846location, which would lead to a large number of debugging intrinsics being 847generated. 848 849Examining the example above, variable ``!30`` is assigned ``%input`` on both 850conditional paths through the function, while ``!23`` is assigned differing 851constant values on either path. Where control flow merges in ``%bb1`` we would 852want ``!30`` to keep its location (``%input``), but ``!23`` to become undefined 853as we cannot determine at runtime what value it should have in %bb1 without 854inserting a PHI node. mem2reg does not insert the PHI node to avoid changing 855codegen when debugging is enabled, and does not insert the other dbg.values 856to avoid adding very large numbers of intrinsics. 857 858Instead, LiveDebugValues determines variable locations when control 859flow merges. A dataflow analysis is used to propagate locations between blocks: 860when control flow merges, if a variable has the same location in all 861predecessors then that location is propagated into the successor. If the 862predecessor locations disagree, the location becomes undefined. 863 864Once LiveDebugValues has run, every block should have all valid variable 865locations described by DBG_VALUE instructions within the block. Very little 866effort is then required by supporting classes (such as 867DbgEntityHistoryCalculator) to build a map of each instruction to every 868valid variable location, without the need to consider control flow. From 869the example above, it is otherwise difficult to determine that the location 870of variable ``!30`` should flow "up" into block ``%bb1``, but that the location 871of variable ``!23`` should not flow "down" into the ``%exit`` block. 872 873.. _ccxx_frontend: 874 875C/C++ front-end specific debug information 876========================================== 877 878The C and C++ front-ends represent information about the program in a 879format that is effectively identical to `DWARF <http://www.dwarfstd.org/>`_ 880in terms of information content. This allows code generators to 881trivially support native debuggers by generating standard dwarf 882information, and contains enough information for non-dwarf targets to 883translate it as needed. 884 885This section describes the forms used to represent C and C++ programs. Other 886languages could pattern themselves after this (which itself is tuned to 887representing programs in the same way that DWARF does), or they could choose 888to provide completely different forms if they don't fit into the DWARF model. 889As support for debugging information gets added to the various LLVM 890source-language front-ends, the information used should be documented here. 891 892The following sections provide examples of a few C/C++ constructs and 893the debug information that would best describe those constructs. The 894canonical references are the ``DINode`` classes defined in 895``include/llvm/IR/DebugInfoMetadata.h`` and the implementations of the 896helper functions in ``lib/IR/DIBuilder.cpp``. 897 898C/C++ source file information 899----------------------------- 900 901``llvm::Instruction`` provides easy access to metadata attached with an 902instruction. One can extract line number information encoded in LLVM IR using 903``Instruction::getDebugLoc()`` and ``DILocation::getLine()``. 904 905.. code-block:: c++ 906 907 if (DILocation *Loc = I->getDebugLoc()) { // Here I is an LLVM instruction 908 unsigned Line = Loc->getLine(); 909 StringRef File = Loc->getFilename(); 910 StringRef Dir = Loc->getDirectory(); 911 bool ImplicitCode = Loc->isImplicitCode(); 912 } 913 914When the flag ImplicitCode is true then it means that the Instruction has been 915added by the front-end but doesn't correspond to source code written by the user. For example 916 917.. code-block:: c++ 918 919 if (MyBoolean) { 920 MyObject MO; 921 ... 922 } 923 924At the end of the scope the MyObject's destructor is called but it isn't written 925explicitly. This information is useful to avoid to have counters on brackets when 926making code coverage. 927 928C/C++ global variable information 929--------------------------------- 930 931Given an integer global variable declared as follows: 932 933.. code-block:: c 934 935 _Alignas(8) int MyGlobal = 100; 936 937a C/C++ front-end would generate the following descriptors: 938 939.. code-block:: text 940 941 ;; 942 ;; Define the global itself. 943 ;; 944 @MyGlobal = global i32 100, align 8, !dbg !0 945 946 ;; 947 ;; List of debug info of globals 948 ;; 949 !llvm.dbg.cu = !{!1} 950 951 ;; Some unrelated metadata. 952 !llvm.module.flags = !{!6, !7} 953 !llvm.ident = !{!8} 954 955 ;; Define the global variable itself 956 !0 = distinct !DIGlobalVariable(name: "MyGlobal", scope: !1, file: !2, line: 1, type: !5, isLocal: false, isDefinition: true, align: 64) 957 958 ;; Define the compile unit. 959 !1 = distinct !DICompileUnit(language: DW_LANG_C99, file: !2, 960 producer: "clang version 4.0.0", 961 isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, 962 enums: !3, globals: !4) 963 964 ;; 965 ;; Define the file 966 ;; 967 !2 = !DIFile(filename: "/dev/stdin", 968 directory: "/Users/dexonsmith/data/llvm/debug-info") 969 970 ;; An empty array. 971 !3 = !{} 972 973 ;; The Array of Global Variables 974 !4 = !{!0} 975 976 ;; 977 ;; Define the type 978 ;; 979 !5 = !DIBasicType(name: "int", size: 32, encoding: DW_ATE_signed) 980 981 ;; Dwarf version to output. 982 !6 = !{i32 2, !"Dwarf Version", i32 4} 983 984 ;; Debug info schema version. 985 !7 = !{i32 2, !"Debug Info Version", i32 3} 986 987 ;; Compiler identification 988 !8 = !{!"clang version 4.0.0"} 989 990 991The align value in DIGlobalVariable description specifies variable alignment in 992case it was forced by C11 _Alignas(), C++11 alignas() keywords or compiler 993attribute __attribute__((aligned ())). In other case (when this field is missing) 994alignment is considered default. This is used when producing DWARF output 995for DW_AT_alignment value. 996 997C/C++ function information 998-------------------------- 999 1000Given a function declared as follows: 1001 1002.. code-block:: c 1003 1004 int main(int argc, char *argv[]) { 1005 return 0; 1006 } 1007 1008a C/C++ front-end would generate the following descriptors: 1009 1010.. code-block:: text 1011 1012 ;; 1013 ;; Define the anchor for subprograms. 1014 ;; 1015 !4 = !DISubprogram(name: "main", scope: !1, file: !1, line: 1, type: !5, 1016 isLocal: false, isDefinition: true, scopeLine: 1, 1017 flags: DIFlagPrototyped, isOptimized: false, 1018 retainedNodes: !2) 1019 1020 ;; 1021 ;; Define the subprogram itself. 1022 ;; 1023 define i32 @main(i32 %argc, i8** %argv) !dbg !4 { 1024 ... 1025 } 1026 1027C++ specific debug information 1028============================== 1029 1030C++ special member functions information 1031---------------------------------------- 1032 1033DWARF 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. 1034 1035Given a class declaration with copy constructor declared as deleted: 1036 1037.. code-block:: c 1038 1039 class foo { 1040 public: 1041 foo(const foo&) = deleted; 1042 }; 1043 1044A C++ frontend would generate following: 1045 1046.. code-block:: text 1047 1048 !17 = !DISubprogram(name: "foo", scope: !11, file: !1, line: 5, type: !18, scopeLine: 5, flags: DIFlagPublic | DIFlagPrototyped, spFlags: DISPFlagDeleted) 1049 1050and this will produce an additional DWARF attribute as: 1051 1052.. code-block:: text 1053 1054 DW_TAG_subprogram [7] * 1055 DW_AT_name [DW_FORM_strx1] (indexed (00000006) string = "foo") 1056 DW_AT_decl_line [DW_FORM_data1] (5) 1057 ... 1058 DW_AT_deleted [DW_FORM_flag_present] (true) 1059 1060Fortran specific debug information 1061================================== 1062 1063Fortran function information 1064---------------------------- 1065 1066There 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. 1067 1068.. code-block:: fortran 1069 1070 elemental function elem_func(a) 1071 1072a Fortran front-end would generate the following descriptors: 1073 1074.. code-block:: text 1075 1076 !11 = distinct !DISubprogram(name: "subroutine2", scope: !1, file: !1, 1077 line: 5, type: !8, scopeLine: 6, 1078 spFlags: DISPFlagDefinition | DISPFlagElemental, unit: !0, 1079 retainedNodes: !2) 1080 1081and this will materialize an additional DWARF attribute as: 1082 1083.. code-block:: text 1084 1085 DW_TAG_subprogram [3] 1086 DW_AT_low_pc [DW_FORM_addr] (0x0000000000000010 ".text") 1087 DW_AT_high_pc [DW_FORM_data4] (0x00000001) 1088 ... 1089 DW_AT_elemental [DW_FORM_flag_present] (true) 1090 1091There 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. 1092 1093.. code-block:: fortran 1094 1095 character(len=*), intent(in) :: string 1096 1097a Fortran front-end would generate the following descriptors: 1098 1099.. code-block:: text 1100 1101 !DILocalVariable(name: "string", arg: 1, scope: !10, file: !3, line: 4, type: !15) 1102 !DIStringType(name: "character(*)!2", stringLength: !16, stringLengthExpression: !DIExpression(), size: 32) 1103 1104A 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. 1105 1106 !DIStringType(name: "character(*)!2", stringLengthExpression: !DIExpression(), stringLocationExpression: !DIExpression(DW_OP_push_object_address, DW_OP_deref), size: 32) 1107 1108and this will materialize in DWARF tags as: 1109 1110.. code-block:: text 1111 1112 DW_TAG_string_type 1113 DW_AT_name ("character(*)!2") 1114 DW_AT_string_length (0x00000064) 1115 0x00000064: DW_TAG_variable 1116 DW_AT_location (DW_OP_fbreg +16) 1117 DW_AT_type (0x00000083 "integer*8") 1118 DW_AT_data_location (DW_OP_push_object_address, DW_OP_deref) 1119 ... 1120 DW_AT_artificial (true) 1121 1122A Fortran front-end may need to generate a *trampoline* function to call a 1123function defined in a different compilation unit. In this case, the front-end 1124can emit the following descriptor for the trampoline function: 1125 1126.. code-block:: text 1127 1128 !DISubprogram(name: "sub1_.t0p", linkageName: "sub1_.t0p", scope: !4, file: !4, type: !5, spFlags: DISPFlagLocalToUnit | DISPFlagDefinition, unit: !7, retainedNodes: !24, targetFuncName: "sub1_") 1129 1130The targetFuncName field is the name of the function that the trampoline 1131calls. This descriptor results in the following DWARF tag: 1132 1133.. code-block:: text 1134 1135 DW_TAG_subprogram 1136 ... 1137 DW_AT_linkage_name ("sub1_.t0p") 1138 DW_AT_name ("sub1_.t0p") 1139 DW_AT_trampoline ("sub1_") 1140 1141Debugging information format 1142============================ 1143 1144Debugging Information Extension for Objective C Properties 1145---------------------------------------------------------- 1146 1147Introduction 1148^^^^^^^^^^^^ 1149 1150Objective C provides a simpler way to declare and define accessor methods using 1151declared properties. The language provides features to declare a property and 1152to let compiler synthesize accessor methods. 1153 1154The debugger lets developer inspect Objective C interfaces and their instance 1155variables and class variables. However, the debugger does not know anything 1156about the properties defined in Objective C interfaces. The debugger consumes 1157information generated by compiler in DWARF format. The format does not support 1158encoding of Objective C properties. This proposal describes DWARF extensions to 1159encode Objective C properties, which the debugger can use to let developers 1160inspect Objective C properties. 1161 1162Proposal 1163^^^^^^^^ 1164 1165Objective C properties exist separately from class members. A property can be 1166defined only by "setter" and "getter" selectors, and be calculated anew on each 1167access. Or a property can just be a direct access to some declared ivar. 1168Finally it can have an ivar "automatically synthesized" for it by the compiler, 1169in which case the property can be referred to in user code directly using the 1170standard C dereference syntax as well as through the property "dot" syntax, but 1171there is no entry in the ``@interface`` declaration corresponding to this ivar. 1172 1173To facilitate debugging, these properties we will add a new DWARF TAG into the 1174``DW_TAG_structure_type`` definition for the class to hold the description of a 1175given property, and a set of DWARF attributes that provide said description. 1176The property tag will also contain the name and declared type of the property. 1177 1178If there is a related ivar, there will also be a DWARF property attribute placed 1179in the ``DW_TAG_member`` DIE for that ivar referring back to the property TAG 1180for that property. And in the case where the compiler synthesizes the ivar 1181directly, the compiler is expected to generate a ``DW_TAG_member`` for that 1182ivar (with the ``DW_AT_artificial`` set to 1), whose name will be the name used 1183to access this ivar directly in code, and with the property attribute pointing 1184back to the property it is backing. 1185 1186The following examples will serve as illustration for our discussion: 1187 1188.. code-block:: objc 1189 1190 @interface I1 { 1191 int n2; 1192 } 1193 1194 @property int p1; 1195 @property int p2; 1196 @end 1197 1198 @implementation I1 1199 @synthesize p1; 1200 @synthesize p2 = n2; 1201 @end 1202 1203This produces the following DWARF (this is a "pseudo dwarfdump" output): 1204 1205.. code-block:: none 1206 1207 0x00000100: TAG_structure_type [7] * 1208 AT_APPLE_runtime_class( 0x10 ) 1209 AT_name( "I1" ) 1210 AT_decl_file( "Objc_Property.m" ) 1211 AT_decl_line( 3 ) 1212 1213 0x00000110 TAG_APPLE_property 1214 AT_name ( "p1" ) 1215 AT_type ( {0x00000150} ( int ) ) 1216 1217 0x00000120: TAG_APPLE_property 1218 AT_name ( "p2" ) 1219 AT_type ( {0x00000150} ( int ) ) 1220 1221 0x00000130: TAG_member [8] 1222 AT_name( "_p1" ) 1223 AT_APPLE_property ( {0x00000110} "p1" ) 1224 AT_type( {0x00000150} ( int ) ) 1225 AT_artificial ( 0x1 ) 1226 1227 0x00000140: TAG_member [8] 1228 AT_name( "n2" ) 1229 AT_APPLE_property ( {0x00000120} "p2" ) 1230 AT_type( {0x00000150} ( int ) ) 1231 1232 0x00000150: AT_type( ( int ) ) 1233 1234Note, the current convention is that the name of the ivar for an 1235auto-synthesized property is the name of the property from which it derives 1236with an underscore prepended, as is shown in the example. But we actually 1237don't need to know this convention, since we are given the name of the ivar 1238directly. 1239 1240Also, it is common practice in ObjC to have different property declarations in 1241the @interface and @implementation - e.g. to provide a read-only property in 1242the interface, and a read-write interface in the implementation. In that case, 1243the compiler should emit whichever property declaration will be in force in the 1244current translation unit. 1245 1246Developers can decorate a property with attributes which are encoded using 1247``DW_AT_APPLE_property_attribute``. 1248 1249.. code-block:: objc 1250 1251 @property (readonly, nonatomic) int pr; 1252 1253.. code-block:: none 1254 1255 TAG_APPLE_property [8] 1256 AT_name( "pr" ) 1257 AT_type ( {0x00000147} (int) ) 1258 AT_APPLE_property_attribute (DW_APPLE_PROPERTY_readonly, DW_APPLE_PROPERTY_nonatomic) 1259 1260The setter and getter method names are attached to the property using 1261``DW_AT_APPLE_property_setter`` and ``DW_AT_APPLE_property_getter`` attributes. 1262 1263.. code-block:: objc 1264 1265 @interface I1 1266 @property (setter=myOwnP3Setter:) int p3; 1267 -(void)myOwnP3Setter:(int)a; 1268 @end 1269 1270 @implementation I1 1271 @synthesize p3; 1272 -(void)myOwnP3Setter:(int)a{ } 1273 @end 1274 1275The DWARF for this would be: 1276 1277.. code-block:: none 1278 1279 0x000003bd: TAG_structure_type [7] * 1280 AT_APPLE_runtime_class( 0x10 ) 1281 AT_name( "I1" ) 1282 AT_decl_file( "Objc_Property.m" ) 1283 AT_decl_line( 3 ) 1284 1285 0x000003cd TAG_APPLE_property 1286 AT_name ( "p3" ) 1287 AT_APPLE_property_setter ( "myOwnP3Setter:" ) 1288 AT_type( {0x00000147} ( int ) ) 1289 1290 0x000003f3: TAG_member [8] 1291 AT_name( "_p3" ) 1292 AT_type ( {0x00000147} ( int ) ) 1293 AT_APPLE_property ( {0x000003cd} ) 1294 AT_artificial ( 0x1 ) 1295 1296New DWARF Tags 1297^^^^^^^^^^^^^^ 1298 1299+-----------------------+--------+ 1300| TAG | Value | 1301+=======================+========+ 1302| DW_TAG_APPLE_property | 0x4200 | 1303+-----------------------+--------+ 1304 1305New DWARF Attributes 1306^^^^^^^^^^^^^^^^^^^^ 1307 1308+--------------------------------+--------+-----------+ 1309| Attribute | Value | Classes | 1310+================================+========+===========+ 1311| DW_AT_APPLE_property | 0x3fed | Reference | 1312+--------------------------------+--------+-----------+ 1313| DW_AT_APPLE_property_getter | 0x3fe9 | String | 1314+--------------------------------+--------+-----------+ 1315| DW_AT_APPLE_property_setter | 0x3fea | String | 1316+--------------------------------+--------+-----------+ 1317| DW_AT_APPLE_property_attribute | 0x3feb | Constant | 1318+--------------------------------+--------+-----------+ 1319 1320New DWARF Constants 1321^^^^^^^^^^^^^^^^^^^ 1322 1323+--------------------------------------+-------+ 1324| Name | Value | 1325+======================================+=======+ 1326| DW_APPLE_PROPERTY_readonly | 0x01 | 1327+--------------------------------------+-------+ 1328| DW_APPLE_PROPERTY_getter | 0x02 | 1329+--------------------------------------+-------+ 1330| DW_APPLE_PROPERTY_assign | 0x04 | 1331+--------------------------------------+-------+ 1332| DW_APPLE_PROPERTY_readwrite | 0x08 | 1333+--------------------------------------+-------+ 1334| DW_APPLE_PROPERTY_retain | 0x10 | 1335+--------------------------------------+-------+ 1336| DW_APPLE_PROPERTY_copy | 0x20 | 1337+--------------------------------------+-------+ 1338| DW_APPLE_PROPERTY_nonatomic | 0x40 | 1339+--------------------------------------+-------+ 1340| DW_APPLE_PROPERTY_setter | 0x80 | 1341+--------------------------------------+-------+ 1342| DW_APPLE_PROPERTY_atomic | 0x100 | 1343+--------------------------------------+-------+ 1344| DW_APPLE_PROPERTY_weak | 0x200 | 1345+--------------------------------------+-------+ 1346| DW_APPLE_PROPERTY_strong | 0x400 | 1347+--------------------------------------+-------+ 1348| DW_APPLE_PROPERTY_unsafe_unretained | 0x800 | 1349+--------------------------------------+-------+ 1350| DW_APPLE_PROPERTY_nullability | 0x1000| 1351+--------------------------------------+-------+ 1352| DW_APPLE_PROPERTY_null_resettable | 0x2000| 1353+--------------------------------------+-------+ 1354| DW_APPLE_PROPERTY_class | 0x4000| 1355+--------------------------------------+-------+ 1356 1357Name Accelerator Tables 1358----------------------- 1359 1360Introduction 1361^^^^^^^^^^^^ 1362 1363The "``.debug_pubnames``" and "``.debug_pubtypes``" formats are not what a 1364debugger needs. The "``pub``" in the section name indicates that the entries 1365in the table are publicly visible names only. This means no static or hidden 1366functions show up in the "``.debug_pubnames``". No static variables or private 1367class variables are in the "``.debug_pubtypes``". Many compilers add different 1368things to these tables, so we can't rely upon the contents between gcc, icc, or 1369clang. 1370 1371The typical query given by users tends not to match up with the contents of 1372these tables. For example, the DWARF spec states that "In the case of the name 1373of a function member or static data member of a C++ structure, class or union, 1374the name presented in the "``.debug_pubnames``" section is not the simple name 1375given by the ``DW_AT_name attribute`` of the referenced debugging information 1376entry, but rather the fully qualified name of the data or function member." 1377So the only names in these tables for complex C++ entries is a fully 1378qualified name. Debugger users tend not to enter their search strings as 1379"``a::b::c(int,const Foo&) const``", but rather as "``c``", "``b::c``" , or 1380"``a::b::c``". So the name entered in the name table must be demangled in 1381order to chop it up appropriately and additional names must be manually entered 1382into the table to make it effective as a name lookup table for debuggers to 1383use. 1384 1385All debuggers currently ignore the "``.debug_pubnames``" table as a result of 1386its inconsistent and useless public-only name content making it a waste of 1387space in the object file. These tables, when they are written to disk, are not 1388sorted in any way, leaving every debugger to do its own parsing and sorting. 1389These tables also include an inlined copy of the string values in the table 1390itself making the tables much larger than they need to be on disk, especially 1391for large C++ programs. 1392 1393Can't we just fix the sections by adding all of the names we need to this 1394table? No, because that is not what the tables are defined to contain and we 1395won't know the difference between the old bad tables and the new good tables. 1396At best we could make our own renamed sections that contain all of the data we 1397need. 1398 1399These tables are also insufficient for what a debugger like LLDB needs. LLDB 1400uses clang for its expression parsing where LLDB acts as a PCH. LLDB is then 1401often asked to look for type "``foo``" or namespace "``bar``", or list items in 1402namespace "``baz``". Namespaces are not included in the pubnames or pubtypes 1403tables. Since clang asks a lot of questions when it is parsing an expression, 1404we need to be very fast when looking up names, as it happens a lot. Having new 1405accelerator tables that are optimized for very quick lookups will benefit this 1406type of debugging experience greatly. 1407 1408We would like to generate name lookup tables that can be mapped into memory 1409from disk, and used as is, with little or no up-front parsing. We would also 1410be able to control the exact content of these different tables so they contain 1411exactly what we need. The Name Accelerator Tables were designed to fix these 1412issues. In order to solve these issues we need to: 1413 1414* Have a format that can be mapped into memory from disk and used as is 1415* Lookups should be very fast 1416* Extensible table format so these tables can be made by many producers 1417* Contain all of the names needed for typical lookups out of the box 1418* Strict rules for the contents of tables 1419 1420Table size is important and the accelerator table format should allow the reuse 1421of strings from common string tables so the strings for the names are not 1422duplicated. We also want to make sure the table is ready to be used as-is by 1423simply mapping the table into memory with minimal header parsing. 1424 1425The name lookups need to be fast and optimized for the kinds of lookups that 1426debuggers tend to do. Optimally we would like to touch as few parts of the 1427mapped table as possible when doing a name lookup and be able to quickly find 1428the name entry we are looking for, or discover there are no matches. In the 1429case of debuggers we optimized for lookups that fail most of the time. 1430 1431Each table that is defined should have strict rules on exactly what is in the 1432accelerator tables and documented so clients can rely on the content. 1433 1434Hash Tables 1435^^^^^^^^^^^ 1436 1437Standard Hash Tables 1438"""""""""""""""""""" 1439 1440Typical hash tables have a header, buckets, and each bucket points to the 1441bucket contents: 1442 1443.. code-block:: none 1444 1445 .------------. 1446 | HEADER | 1447 |------------| 1448 | BUCKETS | 1449 |------------| 1450 | DATA | 1451 `------------' 1452 1453The BUCKETS are an array of offsets to DATA for each hash: 1454 1455.. code-block:: none 1456 1457 .------------. 1458 | 0x00001000 | BUCKETS[0] 1459 | 0x00002000 | BUCKETS[1] 1460 | 0x00002200 | BUCKETS[2] 1461 | 0x000034f0 | BUCKETS[3] 1462 | | ... 1463 | 0xXXXXXXXX | BUCKETS[n_buckets] 1464 '------------' 1465 1466So for ``bucket[3]`` in the example above, we have an offset into the table 14670x000034f0 which points to a chain of entries for the bucket. Each bucket must 1468contain a next pointer, full 32 bit hash value, the string itself, and the data 1469for the current string value. 1470 1471.. code-block:: none 1472 1473 .------------. 1474 0x000034f0: | 0x00003500 | next pointer 1475 | 0x12345678 | 32 bit hash 1476 | "erase" | string value 1477 | data[n] | HashData for this bucket 1478 |------------| 1479 0x00003500: | 0x00003550 | next pointer 1480 | 0x29273623 | 32 bit hash 1481 | "dump" | string value 1482 | data[n] | HashData for this bucket 1483 |------------| 1484 0x00003550: | 0x00000000 | next pointer 1485 | 0x82638293 | 32 bit hash 1486 | "main" | string value 1487 | data[n] | HashData for this bucket 1488 `------------' 1489 1490The problem with this layout for debuggers is that we need to optimize for the 1491negative lookup case where the symbol we're searching for is not present. So 1492if we were to lookup "``printf``" in the table above, we would make a 32-bit 1493hash for "``printf``", it might match ``bucket[3]``. We would need to go to 1494the offset 0x000034f0 and start looking to see if our 32 bit hash matches. To 1495do so, we need to read the next pointer, then read the hash, compare it, and 1496skip to the next bucket. Each time we are skipping many bytes in memory and 1497touching new pages just to do the compare on the full 32 bit hash. All of 1498these accesses then tell us that we didn't have a match. 1499 1500Name Hash Tables 1501"""""""""""""""" 1502 1503To solve the issues mentioned above we have structured the hash tables a bit 1504differently: a header, buckets, an array of all unique 32 bit hash values, 1505followed by an array of hash value data offsets, one for each hash value, then 1506the data for all hash values: 1507 1508.. code-block:: none 1509 1510 .-------------. 1511 | HEADER | 1512 |-------------| 1513 | BUCKETS | 1514 |-------------| 1515 | HASHES | 1516 |-------------| 1517 | OFFSETS | 1518 |-------------| 1519 | DATA | 1520 `-------------' 1521 1522The ``BUCKETS`` in the name tables are an index into the ``HASHES`` array. By 1523making all of the full 32 bit hash values contiguous in memory, we allow 1524ourselves to efficiently check for a match while touching as little memory as 1525possible. Most often checking the 32 bit hash values is as far as the lookup 1526goes. If it does match, it usually is a match with no collisions. So for a 1527table with "``n_buckets``" buckets, and "``n_hashes``" unique 32 bit hash 1528values, we can clarify the contents of the ``BUCKETS``, ``HASHES`` and 1529``OFFSETS`` as: 1530 1531.. code-block:: none 1532 1533 .-------------------------. 1534 | HEADER.magic | uint32_t 1535 | HEADER.version | uint16_t 1536 | HEADER.hash_function | uint16_t 1537 | HEADER.bucket_count | uint32_t 1538 | HEADER.hashes_count | uint32_t 1539 | HEADER.header_data_len | uint32_t 1540 | HEADER_DATA | HeaderData 1541 |-------------------------| 1542 | BUCKETS | uint32_t[n_buckets] // 32 bit hash indexes 1543 |-------------------------| 1544 | HASHES | uint32_t[n_hashes] // 32 bit hash values 1545 |-------------------------| 1546 | OFFSETS | uint32_t[n_hashes] // 32 bit offsets to hash value data 1547 |-------------------------| 1548 | ALL HASH DATA | 1549 `-------------------------' 1550 1551So taking the exact same data from the standard hash example above we end up 1552with: 1553 1554.. code-block:: none 1555 1556 .------------. 1557 | HEADER | 1558 |------------| 1559 | 0 | BUCKETS[0] 1560 | 2 | BUCKETS[1] 1561 | 5 | BUCKETS[2] 1562 | 6 | BUCKETS[3] 1563 | | ... 1564 | ... | BUCKETS[n_buckets] 1565 |------------| 1566 | 0x........ | HASHES[0] 1567 | 0x........ | HASHES[1] 1568 | 0x........ | HASHES[2] 1569 | 0x........ | HASHES[3] 1570 | 0x........ | HASHES[4] 1571 | 0x........ | HASHES[5] 1572 | 0x12345678 | HASHES[6] hash for BUCKETS[3] 1573 | 0x29273623 | HASHES[7] hash for BUCKETS[3] 1574 | 0x82638293 | HASHES[8] hash for BUCKETS[3] 1575 | 0x........ | HASHES[9] 1576 | 0x........ | HASHES[10] 1577 | 0x........ | HASHES[11] 1578 | 0x........ | HASHES[12] 1579 | 0x........ | HASHES[13] 1580 | 0x........ | HASHES[n_hashes] 1581 |------------| 1582 | 0x........ | OFFSETS[0] 1583 | 0x........ | OFFSETS[1] 1584 | 0x........ | OFFSETS[2] 1585 | 0x........ | OFFSETS[3] 1586 | 0x........ | OFFSETS[4] 1587 | 0x........ | OFFSETS[5] 1588 | 0x000034f0 | OFFSETS[6] offset for BUCKETS[3] 1589 | 0x00003500 | OFFSETS[7] offset for BUCKETS[3] 1590 | 0x00003550 | OFFSETS[8] offset for BUCKETS[3] 1591 | 0x........ | OFFSETS[9] 1592 | 0x........ | OFFSETS[10] 1593 | 0x........ | OFFSETS[11] 1594 | 0x........ | OFFSETS[12] 1595 | 0x........ | OFFSETS[13] 1596 | 0x........ | OFFSETS[n_hashes] 1597 |------------| 1598 | | 1599 | | 1600 | | 1601 | | 1602 | | 1603 |------------| 1604 0x000034f0: | 0x00001203 | .debug_str ("erase") 1605 | 0x00000004 | A 32 bit array count - number of HashData with name "erase" 1606 | 0x........ | HashData[0] 1607 | 0x........ | HashData[1] 1608 | 0x........ | HashData[2] 1609 | 0x........ | HashData[3] 1610 | 0x00000000 | String offset into .debug_str (terminate data for hash) 1611 |------------| 1612 0x00003500: | 0x00001203 | String offset into .debug_str ("collision") 1613 | 0x00000002 | A 32 bit array count - number of HashData with name "collision" 1614 | 0x........ | HashData[0] 1615 | 0x........ | HashData[1] 1616 | 0x00001203 | String offset into .debug_str ("dump") 1617 | 0x00000003 | A 32 bit array count - number of HashData with name "dump" 1618 | 0x........ | HashData[0] 1619 | 0x........ | HashData[1] 1620 | 0x........ | HashData[2] 1621 | 0x00000000 | String offset into .debug_str (terminate data for hash) 1622 |------------| 1623 0x00003550: | 0x00001203 | String offset into .debug_str ("main") 1624 | 0x00000009 | A 32 bit array count - number of HashData with name "main" 1625 | 0x........ | HashData[0] 1626 | 0x........ | HashData[1] 1627 | 0x........ | HashData[2] 1628 | 0x........ | HashData[3] 1629 | 0x........ | HashData[4] 1630 | 0x........ | HashData[5] 1631 | 0x........ | HashData[6] 1632 | 0x........ | HashData[7] 1633 | 0x........ | HashData[8] 1634 | 0x00000000 | String offset into .debug_str (terminate data for hash) 1635 `------------' 1636 1637So we still have all of the same data, we just organize it more efficiently for 1638debugger lookup. If we repeat the same "``printf``" lookup from above, we 1639would hash "``printf``" and find it matches ``BUCKETS[3]`` by taking the 32 bit 1640hash value and modulo it by ``n_buckets``. ``BUCKETS[3]`` contains "6" which 1641is the index into the ``HASHES`` table. We would then compare any consecutive 164232 bit hashes values in the ``HASHES`` array as long as the hashes would be in 1643``BUCKETS[3]``. We do this by verifying that each subsequent hash value modulo 1644``n_buckets`` is still 3. In the case of a failed lookup we would access the 1645memory for ``BUCKETS[3]``, and then compare a few consecutive 32 bit hashes 1646before we know that we have no match. We don't end up marching through 1647multiple words of memory and we really keep the number of processor data cache 1648lines being accessed as small as possible. 1649 1650The string hash that is used for these lookup tables is the Daniel J. 1651Bernstein hash which is also used in the ELF ``GNU_HASH`` sections. It is a 1652very good hash for all kinds of names in programs with very few hash 1653collisions. 1654 1655Empty buckets are designated by using an invalid hash index of ``UINT32_MAX``. 1656 1657Details 1658^^^^^^^ 1659 1660These name hash tables are designed to be generic where specializations of the 1661table get to define additional data that goes into the header ("``HeaderData``"), 1662how the string value is stored ("``KeyType``") and the content of the data for each 1663hash value. 1664 1665Header Layout 1666""""""""""""" 1667 1668The header has a fixed part, and the specialized part. The exact format of the 1669header is: 1670 1671.. code-block:: c 1672 1673 struct Header 1674 { 1675 uint32_t magic; // 'HASH' magic value to allow endian detection 1676 uint16_t version; // Version number 1677 uint16_t hash_function; // The hash function enumeration that was used 1678 uint32_t bucket_count; // The number of buckets in this hash table 1679 uint32_t hashes_count; // The total number of unique hash values and hash data offsets in this table 1680 uint32_t header_data_len; // The bytes to skip to get to the hash indexes (buckets) for correct alignment 1681 // Specifically the length of the following HeaderData field - this does not 1682 // include the size of the preceding fields 1683 HeaderData header_data; // Implementation specific header data 1684 }; 1685 1686The header starts with a 32 bit "``magic``" value which must be ``'HASH'`` 1687encoded as an ASCII integer. This allows the detection of the start of the 1688hash table and also allows the table's byte order to be determined so the table 1689can be correctly extracted. The "``magic``" value is followed by a 16 bit 1690``version`` number which allows the table to be revised and modified in the 1691future. The current version number is 1. ``hash_function`` is a ``uint16_t`` 1692enumeration that specifies which hash function was used to produce this table. 1693The current values for the hash function enumerations include: 1694 1695.. code-block:: c 1696 1697 enum HashFunctionType 1698 { 1699 eHashFunctionDJB = 0u, // Daniel J Bernstein hash function 1700 }; 1701 1702``bucket_count`` is a 32 bit unsigned integer that represents how many buckets 1703are in the ``BUCKETS`` array. ``hashes_count`` is the number of unique 32 bit 1704hash values that are in the ``HASHES`` array, and is the same number of offsets 1705are contained in the ``OFFSETS`` array. ``header_data_len`` specifies the size 1706in bytes of the ``HeaderData`` that is filled in by specialized versions of 1707this table. 1708 1709Fixed Lookup 1710"""""""""""" 1711 1712The header is followed by the buckets, hashes, offsets, and hash value data. 1713 1714.. code-block:: c 1715 1716 struct FixedTable 1717 { 1718 uint32_t buckets[Header.bucket_count]; // An array of hash indexes into the "hashes[]" array below 1719 uint32_t hashes [Header.hashes_count]; // Every unique 32 bit hash for the entire table is in this table 1720 uint32_t offsets[Header.hashes_count]; // An offset that corresponds to each item in the "hashes[]" array above 1721 }; 1722 1723``buckets`` is an array of 32 bit indexes into the ``hashes`` array. The 1724``hashes`` array contains all of the 32 bit hash values for all names in the 1725hash table. Each hash in the ``hashes`` table has an offset in the ``offsets`` 1726array that points to the data for the hash value. 1727 1728This table setup makes it very easy to repurpose these tables to contain 1729different data, while keeping the lookup mechanism the same for all tables. 1730This layout also makes it possible to save the table to disk and map it in 1731later and do very efficient name lookups with little or no parsing. 1732 1733DWARF lookup tables can be implemented in a variety of ways and can store a lot 1734of information for each name. We want to make the DWARF tables extensible and 1735able to store the data efficiently so we have used some of the DWARF features 1736that enable efficient data storage to define exactly what kind of data we store 1737for each name. 1738 1739The ``HeaderData`` contains a definition of the contents of each HashData chunk. 1740We might want to store an offset to all of the debug information entries (DIEs) 1741for each name. To keep things extensible, we create a list of items, or 1742Atoms, that are contained in the data for each name. First comes the type of 1743the data in each atom: 1744 1745.. code-block:: c 1746 1747 enum AtomType 1748 { 1749 eAtomTypeNULL = 0u, 1750 eAtomTypeDIEOffset = 1u, // DIE offset, check form for encoding 1751 eAtomTypeCUOffset = 2u, // DIE offset of the compiler unit header that contains the item in question 1752 eAtomTypeTag = 3u, // DW_TAG_xxx value, should be encoded as DW_FORM_data1 (if no tags exceed 255) or DW_FORM_data2 1753 eAtomTypeNameFlags = 4u, // Flags from enum NameFlags 1754 eAtomTypeTypeFlags = 5u, // Flags from enum TypeFlags 1755 }; 1756 1757The enumeration values and their meanings are: 1758 1759.. code-block:: none 1760 1761 eAtomTypeNULL - a termination atom that specifies the end of the atom list 1762 eAtomTypeDIEOffset - an offset into the .debug_info section for the DWARF DIE for this name 1763 eAtomTypeCUOffset - an offset into the .debug_info section for the CU that contains the DIE 1764 eAtomTypeDIETag - The DW_TAG_XXX enumeration value so you don't have to parse the DWARF to see what it is 1765 eAtomTypeNameFlags - Flags for functions and global variables (isFunction, isInlined, isExternal...) 1766 eAtomTypeTypeFlags - Flags for types (isCXXClass, isObjCClass, ...) 1767 1768Then we allow each atom type to define the atom type and how the data for each 1769atom type data is encoded: 1770 1771.. code-block:: c 1772 1773 struct Atom 1774 { 1775 uint16_t type; // AtomType enum value 1776 uint16_t form; // DWARF DW_FORM_XXX defines 1777 }; 1778 1779The ``form`` type above is from the DWARF specification and defines the exact 1780encoding of the data for the Atom type. See the DWARF specification for the 1781``DW_FORM_`` definitions. 1782 1783.. code-block:: c 1784 1785 struct HeaderData 1786 { 1787 uint32_t die_offset_base; 1788 uint32_t atom_count; 1789 Atoms atoms[atom_count0]; 1790 }; 1791 1792``HeaderData`` defines the base DIE offset that should be added to any atoms 1793that are encoded using the ``DW_FORM_ref1``, ``DW_FORM_ref2``, 1794``DW_FORM_ref4``, ``DW_FORM_ref8`` or ``DW_FORM_ref_udata``. It also defines 1795what is contained in each ``HashData`` object -- ``Atom.form`` tells us how large 1796each field will be in the ``HashData`` and the ``Atom.type`` tells us how this data 1797should be interpreted. 1798 1799For the current implementations of the "``.apple_names``" (all functions + 1800globals), the "``.apple_types``" (names of all types that are defined), and 1801the "``.apple_namespaces``" (all namespaces), we currently set the ``Atom`` 1802array to be: 1803 1804.. code-block:: c 1805 1806 HeaderData.atom_count = 1; 1807 HeaderData.atoms[0].type = eAtomTypeDIEOffset; 1808 HeaderData.atoms[0].form = DW_FORM_data4; 1809 1810This defines the contents to be the DIE offset (eAtomTypeDIEOffset) that is 1811encoded as a 32 bit value (DW_FORM_data4). This allows a single name to have 1812multiple matching DIEs in a single file, which could come up with an inlined 1813function for instance. Future tables could include more information about the 1814DIE such as flags indicating if the DIE is a function, method, block, 1815or inlined. 1816 1817The KeyType for the DWARF table is a 32 bit string table offset into the 1818".debug_str" table. The ".debug_str" is the string table for the DWARF which 1819may already contain copies of all of the strings. This helps make sure, with 1820help from the compiler, that we reuse the strings between all of the DWARF 1821sections and keeps the hash table size down. Another benefit to having the 1822compiler generate all strings as DW_FORM_strp in the debug info, is that 1823DWARF parsing can be made much faster. 1824 1825After a lookup is made, we get an offset into the hash data. The hash data 1826needs to be able to deal with 32 bit hash collisions, so the chunk of data 1827at the offset in the hash data consists of a triple: 1828 1829.. code-block:: c 1830 1831 uint32_t str_offset 1832 uint32_t hash_data_count 1833 HashData[hash_data_count] 1834 1835If "str_offset" is zero, then the bucket contents are done. 99.9% of the 1836hash data chunks contain a single item (no 32 bit hash collision): 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 | 0x00000000 | uint32_t KeyType (end of hash chain) 1848 `------------' 1849 1850If there are collisions, you will have multiple valid string offsets: 1851 1852.. code-block:: none 1853 1854 .------------. 1855 | 0x00001023 | uint32_t KeyType (.debug_str[0x0001023] => "main") 1856 | 0x00000004 | uint32_t HashData count 1857 | 0x........ | uint32_t HashData[0] DIE offset 1858 | 0x........ | uint32_t HashData[1] DIE offset 1859 | 0x........ | uint32_t HashData[2] DIE offset 1860 | 0x........ | uint32_t HashData[3] DIE offset 1861 | 0x00002023 | uint32_t KeyType (.debug_str[0x0002023] => "print") 1862 | 0x00000002 | uint32_t HashData count 1863 | 0x........ | uint32_t HashData[0] DIE offset 1864 | 0x........ | uint32_t HashData[1] DIE offset 1865 | 0x00000000 | uint32_t KeyType (end of hash chain) 1866 `------------' 1867 1868Current testing with real world C++ binaries has shown that there is around 1 186932 bit hash collision per 100,000 name entries. 1870 1871Contents 1872^^^^^^^^ 1873 1874As we said, we want to strictly define exactly what is included in the 1875different tables. For DWARF, we have 3 tables: "``.apple_names``", 1876"``.apple_types``", and "``.apple_namespaces``". 1877 1878"``.apple_names``" sections should contain an entry for each DWARF DIE whose 1879``DW_TAG`` is a ``DW_TAG_label``, ``DW_TAG_inlined_subroutine``, or 1880``DW_TAG_subprogram`` that has address attributes: ``DW_AT_low_pc``, 1881``DW_AT_high_pc``, ``DW_AT_ranges`` or ``DW_AT_entry_pc``. It also contains 1882``DW_TAG_variable`` DIEs that have a ``DW_OP_addr`` in the location (global and 1883static variables). All global and static variables should be included, 1884including those scoped within functions and classes. For example using the 1885following code: 1886 1887.. code-block:: c 1888 1889 static int var = 0; 1890 1891 void f () 1892 { 1893 static int var = 0; 1894 } 1895 1896Both of the static ``var`` variables would be included in the table. All 1897functions should emit both their full names and their basenames. For C or C++, 1898the full name is the mangled name (if available) which is usually in the 1899``DW_AT_MIPS_linkage_name`` attribute, and the ``DW_AT_name`` contains the 1900function basename. If global or static variables have a mangled name in a 1901``DW_AT_MIPS_linkage_name`` attribute, this should be emitted along with the 1902simple name found in the ``DW_AT_name`` attribute. 1903 1904"``.apple_types``" sections should contain an entry for each DWARF DIE whose 1905tag is one of: 1906 1907* DW_TAG_array_type 1908* DW_TAG_class_type 1909* DW_TAG_enumeration_type 1910* DW_TAG_pointer_type 1911* DW_TAG_reference_type 1912* DW_TAG_string_type 1913* DW_TAG_structure_type 1914* DW_TAG_subroutine_type 1915* DW_TAG_typedef 1916* DW_TAG_union_type 1917* DW_TAG_ptr_to_member_type 1918* DW_TAG_set_type 1919* DW_TAG_subrange_type 1920* DW_TAG_base_type 1921* DW_TAG_const_type 1922* DW_TAG_immutable_type 1923* DW_TAG_file_type 1924* DW_TAG_namelist 1925* DW_TAG_packed_type 1926* DW_TAG_volatile_type 1927* DW_TAG_restrict_type 1928* DW_TAG_atomic_type 1929* DW_TAG_interface_type 1930* DW_TAG_unspecified_type 1931* DW_TAG_shared_type 1932 1933Only entries with a ``DW_AT_name`` attribute are included, and the entry must 1934not be a forward declaration (``DW_AT_declaration`` attribute with a non-zero 1935value). For example, using the following code: 1936 1937.. code-block:: c 1938 1939 int main () 1940 { 1941 int *b = 0; 1942 return *b; 1943 } 1944 1945We get a few type DIEs: 1946 1947.. code-block:: none 1948 1949 0x00000067: TAG_base_type [5] 1950 AT_encoding( DW_ATE_signed ) 1951 AT_name( "int" ) 1952 AT_byte_size( 0x04 ) 1953 1954 0x0000006e: TAG_pointer_type [6] 1955 AT_type( {0x00000067} ( int ) ) 1956 AT_byte_size( 0x08 ) 1957 1958The DW_TAG_pointer_type is not included because it does not have a ``DW_AT_name``. 1959 1960"``.apple_namespaces``" section should contain all ``DW_TAG_namespace`` DIEs. 1961If we run into a namespace that has no name this is an anonymous namespace, and 1962the name should be output as "``(anonymous namespace)``" (without the quotes). 1963Why? This matches the output of the ``abi::cxa_demangle()`` that is in the 1964standard C++ library that demangles mangled names. 1965 1966 1967Language Extensions and File Format Changes 1968^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 1969 1970Objective-C Extensions 1971"""""""""""""""""""""" 1972 1973"``.apple_objc``" section should contain all ``DW_TAG_subprogram`` DIEs for an 1974Objective-C class. The name used in the hash table is the name of the 1975Objective-C class itself. If the Objective-C class has a category, then an 1976entry is made for both the class name without the category, and for the class 1977name with the category. So if we have a DIE at offset 0x1234 with a name of 1978method "``-[NSString(my_additions) stringWithSpecialString:]``", we would add 1979an entry for "``NSString``" that points to DIE 0x1234, and an entry for 1980"``NSString(my_additions)``" that points to 0x1234. This allows us to quickly 1981track down all Objective-C methods for an Objective-C class when doing 1982expressions. It is needed because of the dynamic nature of Objective-C where 1983anyone can add methods to a class. The DWARF for Objective-C methods is also 1984emitted differently from C++ classes where the methods are not usually 1985contained in the class definition, they are scattered about across one or more 1986compile units. Categories can also be defined in different shared libraries. 1987So we need to be able to quickly find all of the methods and class functions 1988given the Objective-C class name, or quickly find all methods and class 1989functions for a class + category name. This table does not contain any 1990selector names, it just maps Objective-C class names (or class names + 1991category) to all of the methods and class functions. The selectors are added 1992as function basenames in the "``.debug_names``" section. 1993 1994In the "``.apple_names``" section for Objective-C functions, the full name is 1995the entire function name with the brackets ("``-[NSString 1996stringWithCString:]``") and the basename is the selector only 1997("``stringWithCString:``"). 1998 1999Mach-O Changes 2000"""""""""""""" 2001 2002The sections names for the apple hash tables are for non-mach-o files. For 2003mach-o files, the sections should be contained in the ``__DWARF`` segment with 2004names as follows: 2005 2006* "``.apple_names``" -> "``__apple_names``" 2007* "``.apple_types``" -> "``__apple_types``" 2008* "``.apple_namespaces``" -> "``__apple_namespac``" (16 character limit) 2009* "``.apple_objc``" -> "``__apple_objc``" 2010 2011.. _codeview: 2012 2013CodeView Debug Info Format 2014========================== 2015 2016LLVM supports emitting CodeView, the Microsoft debug info format, and this 2017section describes the design and implementation of that support. 2018 2019Format Background 2020----------------- 2021 2022CodeView as a format is clearly oriented around C++ debugging, and in C++, the 2023majority of debug information tends to be type information. Therefore, the 2024overriding design constraint of CodeView is the separation of type information 2025from other "symbol" information so that type information can be efficiently 2026merged across translation units. Both type information and symbol information is 2027generally stored as a sequence of records, where each record begins with a 202816-bit record size and a 16-bit record kind. 2029 2030Type information is usually stored in the ``.debug$T`` section of the object 2031file. All other debug info, such as line info, string table, symbol info, and 2032inlinee info, is stored in one or more ``.debug$S`` sections. There may only be 2033one ``.debug$T`` section per object file, since all other debug info refers to 2034it. If a PDB (enabled by the ``/Zi`` MSVC option) was used during compilation, 2035the ``.debug$T`` section will contain only an ``LF_TYPESERVER2`` record pointing 2036to the PDB. When using PDBs, symbol information appears to remain in the object 2037file ``.debug$S`` sections. 2038 2039Type records are referred to by their index, which is the number of records in 2040the stream before a given record plus ``0x1000``. Many common basic types, such 2041as the basic integral types and unqualified pointers to them, are represented 2042using type indices less than ``0x1000``. Such basic types are built in to 2043CodeView consumers and do not require type records. 2044 2045Each type record may only contain type indices that are less than its own type 2046index. This ensures that the graph of type stream references is acyclic. While 2047the source-level type graph may contain cycles through pointer types (consider a 2048linked list struct), these cycles are removed from the type stream by always 2049referring to the forward declaration record of user-defined record types. Only 2050"symbol" records in the ``.debug$S`` streams may refer to complete, 2051non-forward-declaration type records. 2052 2053Working with CodeView 2054--------------------- 2055 2056These are instructions for some common tasks for developers working to improve 2057LLVM's CodeView support. Most of them revolve around using the CodeView dumper 2058embedded in ``llvm-readobj``. 2059 2060* Testing MSVC's output:: 2061 2062 $ cl -c -Z7 foo.cpp # Use /Z7 to keep types in the object file 2063 $ llvm-readobj --codeview foo.obj 2064 2065* Getting LLVM IR debug info out of Clang:: 2066 2067 $ clang -g -gcodeview --target=x86_64-windows-msvc foo.cpp -S -emit-llvm 2068 2069 Use this to generate LLVM IR for LLVM test cases. 2070 2071* Generate and dump CodeView from LLVM IR metadata:: 2072 2073 $ llc foo.ll -filetype=obj -o foo.obj 2074 $ llvm-readobj --codeview foo.obj > foo.txt 2075 2076 Use this pattern in lit test cases and FileCheck the output of llvm-readobj 2077 2078Improving LLVM's CodeView support is a process of finding interesting type 2079records, constructing a C++ test case that makes MSVC emit those records, 2080dumping the records, understanding them, and then generating equivalent records 2081in LLVM's backend. 2082