1 ===================================== 2 Accurate Garbage Collection with LLVM 3 ===================================== 4 5 .. contents:: 6 :local: 7 8 Introduction 9 ============ 10 11 Garbage collection is a widely used technique that frees the programmer from 12 having to know the lifetimes of heap objects, making software easier to produce 13 and maintain. Many programming languages rely on garbage collection for 14 automatic memory management. There are two primary forms of garbage collection: 15 conservative and accurate. 16 17 Conservative garbage collection often does not require any special support from 18 either the language or the compiler: it can handle non-type-safe programming 19 languages (such as C/C++) and does not require any special information from the 20 compiler. The `Boehm collector 21 <http://www.hpl.hp.com/personal/Hans_Boehm/gc/>`__ is an example of a 22 state-of-the-art conservative collector. 23 24 Accurate garbage collection requires the ability to identify all pointers in the 25 program at run-time (which requires that the source-language be type-safe in 26 most cases). Identifying pointers at run-time requires compiler support to 27 locate all places that hold live pointer variables at run-time, including the 28 :ref:`processor stack and registers <gcroot>`. 29 30 Conservative garbage collection is attractive because it does not require any 31 special compiler support, but it does have problems. In particular, because the 32 conservative garbage collector cannot *know* that a particular word in the 33 machine is a pointer, it cannot move live objects in the heap (preventing the 34 use of compacting and generational GC algorithms) and it can occasionally suffer 35 from memory leaks due to integer values that happen to point to objects in the 36 program. In addition, some aggressive compiler transformations can break 37 conservative garbage collectors (though these seem rare in practice). 38 39 Accurate garbage collectors do not suffer from any of these problems, but they 40 can suffer from degraded scalar optimization of the program. In particular, 41 because the runtime must be able to identify and update all pointers active in 42 the program, some optimizations are less effective. In practice, however, the 43 locality and performance benefits of using aggressive garbage collection 44 techniques dominates any low-level losses. 45 46 This document describes the mechanisms and interfaces provided by LLVM to 47 support accurate garbage collection. 48 49 Goals and non-goals 50 ------------------- 51 52 LLVM's intermediate representation provides :ref:`garbage collection intrinsics 53 <gc_intrinsics>` that offer support for a broad class of collector models. For 54 instance, the intrinsics permit: 55 56 * semi-space collectors 57 58 * mark-sweep collectors 59 60 * generational collectors 61 62 * reference counting 63 64 * incremental collectors 65 66 * concurrent collectors 67 68 * cooperative collectors 69 70 We hope that the primitive support built into the LLVM IR is sufficient to 71 support a broad class of garbage collected languages including Scheme, ML, Java, 72 C#, Perl, Python, Lua, Ruby, other scripting languages, and more. 73 74 However, LLVM does not itself provide a garbage collector --- this should be 75 part of your language's runtime library. LLVM provides a framework for compile 76 time :ref:`code generation plugins <plugin>`. The role of these plugins is to 77 generate code and data structures which conforms to the *binary interface* 78 specified by the *runtime library*. This is similar to the relationship between 79 LLVM and DWARF debugging info, for example. The difference primarily lies in 80 the lack of an established standard in the domain of garbage collection --- thus 81 the plugins. 82 83 The aspects of the binary interface with which LLVM's GC support is 84 concerned are: 85 86 * Creation of GC-safe points within code where collection is allowed to execute 87 safely. 88 89 * Computation of the stack map. For each safe point in the code, object 90 references within the stack frame must be identified so that the collector may 91 traverse and perhaps update them. 92 93 * Write barriers when storing object references to the heap. These are commonly 94 used to optimize incremental scans in generational collectors. 95 96 * Emission of read barriers when loading object references. These are useful 97 for interoperating with concurrent collectors. 98 99 There are additional areas that LLVM does not directly address: 100 101 * Registration of global roots with the runtime. 102 103 * Registration of stack map entries with the runtime. 104 105 * The functions used by the program to allocate memory, trigger a collection, 106 etc. 107 108 * Computation or compilation of type maps, or registration of them with the 109 runtime. These are used to crawl the heap for object references. 110 111 In general, LLVM's support for GC does not include features which can be 112 adequately addressed with other features of the IR and does not specify a 113 particular binary interface. On the plus side, this means that you should be 114 able to integrate LLVM with an existing runtime. On the other hand, it leaves a 115 lot of work for the developer of a novel language. However, it's easy to get 116 started quickly and scale up to a more sophisticated implementation as your 117 compiler matures. 118 119 Getting started 120 =============== 121 122 Using a GC with LLVM implies many things, for example: 123 124 * Write a runtime library or find an existing one which implements a GC heap. 125 126 #. Implement a memory allocator. 127 128 #. Design a binary interface for the stack map, used to identify references 129 within a stack frame on the machine stack.\* 130 131 #. Implement a stack crawler to discover functions on the call stack.\* 132 133 #. Implement a registry for global roots. 134 135 #. Design a binary interface for type maps, used to identify references 136 within heap objects. 137 138 #. Implement a collection routine bringing together all of the above. 139 140 * Emit compatible code from your compiler. 141 142 * Initialization in the main function. 143 144 * Use the ``gc "..."`` attribute to enable GC code generation (or 145 ``F.setGC("...")``). 146 147 * Use ``@llvm.gcroot`` to mark stack roots. 148 149 * Use ``@llvm.gcread`` and/or ``@llvm.gcwrite`` to manipulate GC references, 150 if necessary. 151 152 * Allocate memory using the GC allocation routine provided by the runtime 153 library. 154 155 * Generate type maps according to your runtime's binary interface. 156 157 * Write a compiler plugin to interface LLVM with the runtime library.\* 158 159 * Lower ``@llvm.gcread`` and ``@llvm.gcwrite`` to appropriate code 160 sequences.\* 161 162 * Compile LLVM's stack map to the binary form expected by the runtime. 163 164 * Load the plugin into the compiler. Use ``llc -load`` or link the plugin 165 statically with your language's compiler.\* 166 167 * Link program executables with the runtime. 168 169 To help with several of these tasks (those indicated with a \*), LLVM includes a 170 highly portable, built-in ShadowStack code generator. It is compiled into 171 ``llc`` and works even with the interpreter and C backends. 172 173 In your compiler 174 ---------------- 175 176 To turn the shadow stack on for your functions, first call: 177 178 .. code-block:: c++ 179 180 F.setGC("shadow-stack"); 181 182 for each function your compiler emits. Since the shadow stack is built into 183 LLVM, you do not need to load a plugin. 184 185 Your compiler must also use ``@llvm.gcroot`` as documented. Don't forget to 186 create a root for each intermediate value that is generated when evaluating an 187 expression. In ``h(f(), g())``, the result of ``f()`` could easily be collected 188 if evaluating ``g()`` triggers a collection. 189 190 There's no need to use ``@llvm.gcread`` and ``@llvm.gcwrite`` over plain 191 ``load`` and ``store`` for now. You will need them when switching to a more 192 advanced GC. 193 194 In your runtime 195 --------------- 196 197 The shadow stack doesn't imply a memory allocation algorithm. A semispace 198 collector or building atop ``malloc`` are great places to start, and can be 199 implemented with very little code. 200 201 When it comes time to collect, however, your runtime needs to traverse the stack 202 roots, and for this it needs to integrate with the shadow stack. Luckily, doing 203 so is very simple. (This code is heavily commented to help you understand the 204 data structure, but there are only 20 lines of meaningful code.) 205 206 .. code-block:: c++ 207 208 /// @brief The map for a single function's stack frame. One of these is 209 /// compiled as constant data into the executable for each function. 210 /// 211 /// Storage of metadata values is elided if the %metadata parameter to 212 /// @llvm.gcroot is null. 213 struct FrameMap { 214 int32_t NumRoots; //< Number of roots in stack frame. 215 int32_t NumMeta; //< Number of metadata entries. May be < NumRoots. 216 const void *Meta[0]; //< Metadata for each root. 217 }; 218 219 /// @brief A link in the dynamic shadow stack. One of these is embedded in 220 /// the stack frame of each function on the call stack. 221 struct StackEntry { 222 StackEntry *Next; //< Link to next stack entry (the caller's). 223 const FrameMap *Map; //< Pointer to constant FrameMap. 224 void *Roots[0]; //< Stack roots (in-place array). 225 }; 226 227 /// @brief The head of the singly-linked list of StackEntries. Functions push 228 /// and pop onto this in their prologue and epilogue. 229 /// 230 /// Since there is only a global list, this technique is not threadsafe. 231 StackEntry *llvm_gc_root_chain; 232 233 /// @brief Calls Visitor(root, meta) for each GC root on the stack. 234 /// root and meta are exactly the values passed to 235 /// @llvm.gcroot. 236 /// 237 /// Visitor could be a function to recursively mark live objects. Or it 238 /// might copy them to another heap or generation. 239 /// 240 /// @param Visitor A function to invoke for every GC root on the stack. 241 void visitGCRoots(void (*Visitor)(void **Root, const void *Meta)) { 242 for (StackEntry *R = llvm_gc_root_chain; R; R = R->Next) { 243 unsigned i = 0; 244 245 // For roots [0, NumMeta), the metadata pointer is in the FrameMap. 246 for (unsigned e = R->Map->NumMeta; i != e; ++i) 247 Visitor(&R->Roots[i], R->Map->Meta[i]); 248 249 // For roots [NumMeta, NumRoots), the metadata pointer is null. 250 for (unsigned e = R->Map->NumRoots; i != e; ++i) 251 Visitor(&R->Roots[i], NULL); 252 } 253 } 254 255 About the shadow stack 256 ---------------------- 257 258 Unlike many GC algorithms which rely on a cooperative code generator to compile 259 stack maps, this algorithm carefully maintains a linked list of stack roots 260 [:ref:`Henderson2002 <henderson02>`]. This so-called "shadow stack" mirrors the 261 machine stack. Maintaining this data structure is slower than using a stack map 262 compiled into the executable as constant data, but has a significant portability 263 advantage because it requires no special support from the target code generator, 264 and does not require tricky platform-specific code to crawl the machine stack. 265 266 The tradeoff for this simplicity and portability is: 267 268 * High overhead per function call. 269 270 * Not thread-safe. 271 272 Still, it's an easy way to get started. After your compiler and runtime are up 273 and running, writing a :ref:`plugin <plugin>` will allow you to take advantage 274 of :ref:`more advanced GC features <collector-algos>` of LLVM in order to 275 improve performance. 276 277 .. _gc_intrinsics: 278 279 IR features 280 =========== 281 282 This section describes the garbage collection facilities provided by the 283 :doc:`LLVM intermediate representation <LangRef>`. The exact behavior of these 284 IR features is specified by the binary interface implemented by a :ref:`code 285 generation plugin <plugin>`, not by this document. 286 287 These facilities are limited to those strictly necessary; they are not intended 288 to be a complete interface to any garbage collector. A program will need to 289 interface with the GC library using the facilities provided by that program. 290 291 Specifying GC code generation: ``gc "..."`` 292 ------------------------------------------- 293 294 .. code-block:: llvm 295 296 define ty @name(...) gc "name" { ... 297 298 The ``gc`` function attribute is used to specify the desired GC style to the 299 compiler. Its programmatic equivalent is the ``setGC`` method of ``Function``. 300 301 Setting ``gc "name"`` on a function triggers a search for a matching code 302 generation plugin "*name*"; it is that plugin which defines the exact nature of 303 the code generated to support GC. If none is found, the compiler will raise an 304 error. 305 306 Specifying the GC style on a per-function basis allows LLVM to link together 307 programs that use different garbage collection algorithms (or none at all). 308 309 .. _gcroot: 310 311 Identifying GC roots on the stack: ``llvm.gcroot`` 312 -------------------------------------------------- 313 314 .. code-block:: llvm 315 316 void @llvm.gcroot(i8** %ptrloc, i8* %metadata) 317 318 The ``llvm.gcroot`` intrinsic is used to inform LLVM that a stack variable 319 references an object on the heap and is to be tracked for garbage collection. 320 The exact impact on generated code is specified by a :ref:`compiler plugin 321 <plugin>`. All calls to ``llvm.gcroot`` **must** reside inside the first basic 322 block. 323 324 A compiler which uses mem2reg to raise imperative code using ``alloca`` into SSA 325 form need only add a call to ``@llvm.gcroot`` for those variables which a 326 pointers into the GC heap. 327 328 It is also important to mark intermediate values with ``llvm.gcroot``. For 329 example, consider ``h(f(), g())``. Beware leaking the result of ``f()`` in the 330 case that ``g()`` triggers a collection. Note, that stack variables must be 331 initialized and marked with ``llvm.gcroot`` in function's prologue. 332 333 The first argument **must** be a value referring to an alloca instruction or a 334 bitcast of an alloca. The second contains a pointer to metadata that should be 335 associated with the pointer, and **must** be a constant or global value 336 address. If your target collector uses tags, use a null pointer for metadata. 337 338 The ``%metadata`` argument can be used to avoid requiring heap objects to have 339 'isa' pointers or tag bits. [Appel89_, Goldberg91_, Tolmach94_] If specified, 340 its value will be tracked along with the location of the pointer in the stack 341 frame. 342 343 Consider the following fragment of Java code: 344 345 .. code-block:: java 346 347 { 348 Object X; // A null-initialized reference to an object 349 ... 350 } 351 352 This block (which may be located in the middle of a function or in a loop nest), 353 could be compiled to this LLVM code: 354 355 .. code-block:: llvm 356 357 Entry: 358 ;; In the entry block for the function, allocate the 359 ;; stack space for X, which is an LLVM pointer. 360 %X = alloca %Object* 361 362 ;; Tell LLVM that the stack space is a stack root. 363 ;; Java has type-tags on objects, so we pass null as metadata. 364 %tmp = bitcast %Object** %X to i8** 365 call void @llvm.gcroot(i8** %tmp, i8* null) 366 ... 367 368 ;; "CodeBlock" is the block corresponding to the start 369 ;; of the scope above. 370 CodeBlock: 371 ;; Java null-initializes pointers. 372 store %Object* null, %Object** %X 373 374 ... 375 376 ;; As the pointer goes out of scope, store a null value into 377 ;; it, to indicate that the value is no longer live. 378 store %Object* null, %Object** %X 379 ... 380 381 Reading and writing references in the heap 382 ------------------------------------------ 383 384 Some collectors need to be informed when the mutator (the program that needs 385 garbage collection) either reads a pointer from or writes a pointer to a field 386 of a heap object. The code fragments inserted at these points are called *read 387 barriers* and *write barriers*, respectively. The amount of code that needs to 388 be executed is usually quite small and not on the critical path of any 389 computation, so the overall performance impact of the barrier is tolerable. 390 391 Barriers often require access to the *object pointer* rather than the *derived 392 pointer* (which is a pointer to the field within the object). Accordingly, 393 these intrinsics take both pointers as separate arguments for completeness. In 394 this snippet, ``%object`` is the object pointer, and ``%derived`` is the derived 395 pointer: 396 397 .. code-block:: llvm 398 399 ;; An array type. 400 %class.Array = type { %class.Object, i32, [0 x %class.Object*] } 401 ... 402 403 ;; Load the object pointer from a gcroot. 404 %object = load %class.Array** %object_addr 405 406 ;; Compute the derived pointer. 407 %derived = getelementptr %object, i32 0, i32 2, i32 %n 408 409 LLVM does not enforce this relationship between the object and derived pointer 410 (although a :ref:`plugin <plugin>` might). However, it would be an unusual 411 collector that violated it. 412 413 The use of these intrinsics is naturally optional if the target GC does require 414 the corresponding barrier. Such a GC plugin will replace the intrinsic calls 415 with the corresponding ``load`` or ``store`` instruction if they are used. 416 417 Write barrier: ``llvm.gcwrite`` 418 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 419 420 .. code-block:: llvm 421 422 void @llvm.gcwrite(i8* %value, i8* %object, i8** %derived) 423 424 For write barriers, LLVM provides the ``llvm.gcwrite`` intrinsic function. It 425 has exactly the same semantics as a non-volatile ``store`` to the derived 426 pointer (the third argument). The exact code generated is specified by a 427 compiler :ref:`plugin <plugin>`. 428 429 Many important algorithms require write barriers, including generational and 430 concurrent collectors. Additionally, write barriers could be used to implement 431 reference counting. 432 433 Read barrier: ``llvm.gcread`` 434 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 435 436 .. code-block:: llvm 437 438 i8* @llvm.gcread(i8* %object, i8** %derived) 439 440 For read barriers, LLVM provides the ``llvm.gcread`` intrinsic function. It has 441 exactly the same semantics as a non-volatile ``load`` from the derived pointer 442 (the second argument). The exact code generated is specified by a 443 :ref:`compiler plugin <plugin>`. 444 445 Read barriers are needed by fewer algorithms than write barriers, and may have a 446 greater performance impact since pointer reads are more frequent than writes. 447 448 .. _plugin: 449 450 Implementing a collector plugin 451 =============================== 452 453 User code specifies which GC code generation to use with the ``gc`` function 454 attribute or, equivalently, with the ``setGC`` method of ``Function``. 455 456 To implement a GC plugin, it is necessary to subclass ``llvm::GCStrategy``, 457 which can be accomplished in a few lines of boilerplate code. LLVM's 458 infrastructure provides access to several important algorithms. For an 459 uncontroversial collector, all that remains may be to compile LLVM's computed 460 stack map to assembly code (using the binary representation expected by the 461 runtime library). This can be accomplished in about 100 lines of code. 462 463 This is not the appropriate place to implement a garbage collected heap or a 464 garbage collector itself. That code should exist in the language's runtime 465 library. The compiler plugin is responsible for generating code which conforms 466 to the binary interface defined by library, most essentially the :ref:`stack map 467 <stack-map>`. 468 469 To subclass ``llvm::GCStrategy`` and register it with the compiler: 470 471 .. code-block:: c++ 472 473 // lib/MyGC/MyGC.cpp - Example LLVM GC plugin 474 475 #include "llvm/CodeGen/GCStrategy.h" 476 #include "llvm/CodeGen/GCMetadata.h" 477 #include "llvm/Support/Compiler.h" 478 479 using namespace llvm; 480 481 namespace { 482 class LLVM_LIBRARY_VISIBILITY MyGC : public GCStrategy { 483 public: 484 MyGC() {} 485 }; 486 487 GCRegistry::Add<MyGC> 488 X("mygc", "My bespoke garbage collector."); 489 } 490 491 This boilerplate collector does nothing. More specifically: 492 493 * ``llvm.gcread`` calls are replaced with the corresponding ``load`` 494 instruction. 495 496 * ``llvm.gcwrite`` calls are replaced with the corresponding ``store`` 497 instruction. 498 499 * No safe points are added to the code. 500 501 * The stack map is not compiled into the executable. 502 503 Using the LLVM makefiles, this code 504 can be compiled as a plugin using a simple makefile: 505 506 .. code-block:: make 507 508 # lib/MyGC/Makefile 509 510 LEVEL := ../.. 511 LIBRARYNAME = MyGC 512 LOADABLE_MODULE = 1 513 514 include $(LEVEL)/Makefile.common 515 516 Once the plugin is compiled, code using it may be compiled using ``llc 517 -load=MyGC.so`` (though MyGC.so may have some other platform-specific 518 extension): 519 520 :: 521 522 $ cat sample.ll 523 define void @f() gc "mygc" { 524 entry: 525 ret void 526 } 527 $ llvm-as < sample.ll | llc -load=MyGC.so 528 529 It is also possible to statically link the collector plugin into tools, such as 530 a language-specific compiler front-end. 531 532 .. _collector-algos: 533 534 Overview of available features 535 ------------------------------ 536 537 ``GCStrategy`` provides a range of features through which a plugin may do useful 538 work. Some of these are callbacks, some are algorithms that can be enabled, 539 disabled, or customized. This matrix summarizes the supported (and planned) 540 features and correlates them with the collection techniques which typically 541 require them. 542 543 .. |v| unicode:: 0x2714 544 :trim: 545 546 .. |x| unicode:: 0x2718 547 :trim: 548 549 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 550 | Algorithm | Done | Shadow | refcount | mark- | copying | incremental | threaded | concurrent | 551 | | | stack | | sweep | | | | | 552 +============+======+========+==========+=======+=========+=============+==========+============+ 553 | stack map | |v| | | | |x| | |x| | |x| | |x| | |x| | 554 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 555 | initialize | |v| | |x| | |x| | |x| | |x| | |x| | |x| | |x| | 556 | roots | | | | | | | | | 557 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 558 | derived | NO | | | | | | **N**\* | **N**\* | 559 | pointers | | | | | | | | | 560 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 561 | **custom | |v| | | | | | | | | 562 | lowering** | | | | | | | | | 563 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 564 | *gcroot* | |v| | |x| | |x| | | | | | | 565 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 566 | *gcwrite* | |v| | | |x| | | | |x| | | |x| | 567 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 568 | *gcread* | |v| | | | | | | | |x| | 569 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 570 | **safe | | | | | | | | | 571 | points** | | | | | | | | | 572 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 573 | *in | |v| | | | |x| | |x| | |x| | |x| | |x| | 574 | calls* | | | | | | | | | 575 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 576 | *before | |v| | | | | | | |x| | |x| | 577 | calls* | | | | | | | | | 578 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 579 | *for | NO | | | | | | **N** | **N** | 580 | loops* | | | | | | | | | 581 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 582 | *before | |v| | | | | | | |x| | |x| | 583 | escape* | | | | | | | | | 584 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 585 | emit code | NO | | | | | | **N** | **N** | 586 | at safe | | | | | | | | | 587 | points | | | | | | | | | 588 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 589 | **output** | | | | | | | | | 590 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 591 | *assembly* | |v| | | | |x| | |x| | |x| | |x| | |x| | 592 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 593 | *JIT* | NO | | | **?** | **?** | **?** | **?** | **?** | 594 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 595 | *obj* | NO | | | **?** | **?** | **?** | **?** | **?** | 596 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 597 | live | NO | | | **?** | **?** | **?** | **?** | **?** | 598 | analysis | | | | | | | | | 599 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 600 | register | NO | | | **?** | **?** | **?** | **?** | **?** | 601 | map | | | | | | | | | 602 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 603 | \* Derived pointers only pose a hasard to copying collections. | 604 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 605 | **?** denotes a feature which could be utilized if available. | 606 +------------+------+--------+----------+-------+---------+-------------+----------+------------+ 607 608 To be clear, the collection techniques above are defined as: 609 610 Shadow Stack 611 The mutator carefully maintains a linked list of stack roots. 612 613 Reference Counting 614 The mutator maintains a reference count for each object and frees an object 615 when its count falls to zero. 616 617 Mark-Sweep 618 When the heap is exhausted, the collector marks reachable objects starting 619 from the roots, then deallocates unreachable objects in a sweep phase. 620 621 Copying 622 As reachability analysis proceeds, the collector copies objects from one heap 623 area to another, compacting them in the process. Copying collectors enable 624 highly efficient "bump pointer" allocation and can improve locality of 625 reference. 626 627 Incremental 628 (Including generational collectors.) Incremental collectors generally have all 629 the properties of a copying collector (regardless of whether the mature heap 630 is compacting), but bring the added complexity of requiring write barriers. 631 632 Threaded 633 Denotes a multithreaded mutator; the collector must still stop the mutator 634 ("stop the world") before beginning reachability analysis. Stopping a 635 multithreaded mutator is a complicated problem. It generally requires highly 636 platform-specific code in the runtime, and the production of carefully 637 designed machine code at safe points. 638 639 Concurrent 640 In this technique, the mutator and the collector run concurrently, with the 641 goal of eliminating pause times. In a *cooperative* collector, the mutator 642 further aids with collection should a pause occur, allowing collection to take 643 advantage of multiprocessor hosts. The "stop the world" problem of threaded 644 collectors is generally still present to a limited extent. Sophisticated 645 marking algorithms are necessary. Read barriers may be necessary. 646 647 As the matrix indicates, LLVM's garbage collection infrastructure is already 648 suitable for a wide variety of collectors, but does not currently extend to 649 multithreaded programs. This will be added in the future as there is 650 interest. 651 652 .. _stack-map: 653 654 Computing stack maps 655 -------------------- 656 657 LLVM automatically computes a stack map. One of the most important features 658 of a ``GCStrategy`` is to compile this information into the executable in 659 the binary representation expected by the runtime library. 660 661 The stack map consists of the location and identity of each GC root in the 662 each function in the module. For each root: 663 664 * ``RootNum``: The index of the root. 665 666 * ``StackOffset``: The offset of the object relative to the frame pointer. 667 668 * ``RootMetadata``: The value passed as the ``%metadata`` parameter to the 669 ``@llvm.gcroot`` intrinsic. 670 671 Also, for the function as a whole: 672 673 * ``getFrameSize()``: The overall size of the function's initial stack frame, 674 not accounting for any dynamic allocation. 675 676 * ``roots_size()``: The count of roots in the function. 677 678 To access the stack map, use ``GCFunctionMetadata::roots_begin()`` and 679 -``end()`` from the :ref:`GCMetadataPrinter <assembly>`: 680 681 .. code-block:: c++ 682 683 for (iterator I = begin(), E = end(); I != E; ++I) { 684 GCFunctionInfo *FI = *I; 685 unsigned FrameSize = FI->getFrameSize(); 686 size_t RootCount = FI->roots_size(); 687 688 for (GCFunctionInfo::roots_iterator RI = FI->roots_begin(), 689 RE = FI->roots_end(); 690 RI != RE; ++RI) { 691 int RootNum = RI->Num; 692 int RootStackOffset = RI->StackOffset; 693 Constant *RootMetadata = RI->Metadata; 694 } 695 } 696 697 If the ``llvm.gcroot`` intrinsic is eliminated before code generation by a 698 custom lowering pass, LLVM will compute an empty stack map. This may be useful 699 for collector plugins which implement reference counting or a shadow stack. 700 701 .. _init-roots: 702 703 Initializing roots to null: ``InitRoots`` 704 ----------------------------------------- 705 706 .. code-block:: c++ 707 708 MyGC::MyGC() { 709 InitRoots = true; 710 } 711 712 When set, LLVM will automatically initialize each root to ``null`` upon entry to 713 the function. This prevents the GC's sweep phase from visiting uninitialized 714 pointers, which will almost certainly cause it to crash. This initialization 715 occurs before custom lowering, so the two may be used together. 716 717 Since LLVM does not yet compute liveness information, there is no means of 718 distinguishing an uninitialized stack root from an initialized one. Therefore, 719 this feature should be used by all GC plugins. It is enabled by default. 720 721 Custom lowering of intrinsics: ``CustomRoots``, ``CustomReadBarriers``, and ``CustomWriteBarriers`` 722 --------------------------------------------------------------------------------------------------- 723 724 For GCs which use barriers or unusual treatment of stack roots, these flags 725 allow the collector to perform arbitrary transformations of the LLVM IR: 726 727 .. code-block:: c++ 728 729 class MyGC : public GCStrategy { 730 public: 731 MyGC() { 732 CustomRoots = true; 733 CustomReadBarriers = true; 734 CustomWriteBarriers = true; 735 } 736 737 virtual bool initializeCustomLowering(Module &M); 738 virtual bool performCustomLowering(Function &F); 739 }; 740 741 If any of these flags are set, then LLVM suppresses its default lowering for the 742 corresponding intrinsics and instead calls ``performCustomLowering``. 743 744 LLVM's default action for each intrinsic is as follows: 745 746 * ``llvm.gcroot``: Leave it alone. The code generator must see it or the stack 747 map will not be computed. 748 749 * ``llvm.gcread``: Substitute a ``load`` instruction. 750 751 * ``llvm.gcwrite``: Substitute a ``store`` instruction. 752 753 If ``CustomReadBarriers`` or ``CustomWriteBarriers`` are specified, then 754 ``performCustomLowering`` **must** eliminate the corresponding barriers. 755 756 ``performCustomLowering`` must comply with the same restrictions as 757 :ref:`FunctionPass::runOnFunction <writing-an-llvm-pass-runOnFunction>` 758 Likewise, ``initializeCustomLowering`` has the same semantics as 759 :ref:`Pass::doInitialization(Module&) 760 <writing-an-llvm-pass-doInitialization-mod>` 761 762 The following can be used as a template: 763 764 .. code-block:: c++ 765 766 #include "llvm/IR/Module.h" 767 #include "llvm/IR/IntrinsicInst.h" 768 769 bool MyGC::initializeCustomLowering(Module &M) { 770 return false; 771 } 772 773 bool MyGC::performCustomLowering(Function &F) { 774 bool MadeChange = false; 775 776 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 777 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) 778 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(II++)) 779 if (Function *F = CI->getCalledFunction()) 780 switch (F->getIntrinsicID()) { 781 case Intrinsic::gcwrite: 782 // Handle llvm.gcwrite. 783 CI->eraseFromParent(); 784 MadeChange = true; 785 break; 786 case Intrinsic::gcread: 787 // Handle llvm.gcread. 788 CI->eraseFromParent(); 789 MadeChange = true; 790 break; 791 case Intrinsic::gcroot: 792 // Handle llvm.gcroot. 793 CI->eraseFromParent(); 794 MadeChange = true; 795 break; 796 } 797 798 return MadeChange; 799 } 800 801 .. _safe-points: 802 803 Generating safe points: ``NeededSafePoints`` 804 -------------------------------------------- 805 806 LLVM can compute four kinds of safe points: 807 808 .. code-block:: c++ 809 810 namespace GC { 811 /// PointKind - The type of a collector-safe point. 812 /// 813 enum PointKind { 814 Loop, //< Instr is a loop (backwards branch). 815 Return, //< Instr is a return instruction. 816 PreCall, //< Instr is a call instruction. 817 PostCall //< Instr is the return address of a call. 818 }; 819 } 820 821 A collector can request any combination of the four by setting the 822 ``NeededSafePoints`` mask: 823 824 .. code-block:: c++ 825 826 MyGC::MyGC() { 827 NeededSafePoints = 1 << GC::Loop 828 | 1 << GC::Return 829 | 1 << GC::PreCall 830 | 1 << GC::PostCall; 831 } 832 833 It can then use the following routines to access safe points. 834 835 .. code-block:: c++ 836 837 for (iterator I = begin(), E = end(); I != E; ++I) { 838 GCFunctionInfo *MD = *I; 839 size_t PointCount = MD->size(); 840 841 for (GCFunctionInfo::iterator PI = MD->begin(), 842 PE = MD->end(); PI != PE; ++PI) { 843 GC::PointKind PointKind = PI->Kind; 844 unsigned PointNum = PI->Num; 845 } 846 } 847 848 Almost every collector requires ``PostCall`` safe points, since these correspond 849 to the moments when the function is suspended during a call to a subroutine. 850 851 Threaded programs generally require ``Loop`` safe points to guarantee that the 852 application will reach a safe point within a bounded amount of time, even if it 853 is executing a long-running loop which contains no function calls. 854 855 Threaded collectors may also require ``Return`` and ``PreCall`` safe points to 856 implement "stop the world" techniques using self-modifying code, where it is 857 important that the program not exit the function without reaching a safe point 858 (because only the topmost function has been patched). 859 860 .. _assembly: 861 862 Emitting assembly code: ``GCMetadataPrinter`` 863 --------------------------------------------- 864 865 LLVM allows a plugin to print arbitrary assembly code before and after the rest 866 of a module's assembly code. At the end of the module, the GC can compile the 867 LLVM stack map into assembly code. (At the beginning, this information is not 868 yet computed.) 869 870 Since AsmWriter and CodeGen are separate components of LLVM, a separate abstract 871 base class and registry is provided for printing assembly code, the 872 ``GCMetadaPrinter`` and ``GCMetadataPrinterRegistry``. The AsmWriter will look 873 for such a subclass if the ``GCStrategy`` sets ``UsesMetadata``: 874 875 .. code-block:: c++ 876 877 MyGC::MyGC() { 878 UsesMetadata = true; 879 } 880 881 This separation allows JIT-only clients to be smaller. 882 883 Note that LLVM does not currently have analogous APIs to support code generation 884 in the JIT, nor using the object writers. 885 886 .. code-block:: c++ 887 888 // lib/MyGC/MyGCPrinter.cpp - Example LLVM GC printer 889 890 #include "llvm/CodeGen/GCMetadataPrinter.h" 891 #include "llvm/Support/Compiler.h" 892 893 using namespace llvm; 894 895 namespace { 896 class LLVM_LIBRARY_VISIBILITY MyGCPrinter : public GCMetadataPrinter { 897 public: 898 virtual void beginAssembly(AsmPrinter &AP); 899 900 virtual void finishAssembly(AsmPrinter &AP); 901 }; 902 903 GCMetadataPrinterRegistry::Add<MyGCPrinter> 904 X("mygc", "My bespoke garbage collector."); 905 } 906 907 The collector should use ``AsmPrinter`` to print portable assembly code. The 908 collector itself contains the stack map for the entire module, and may access 909 the ``GCFunctionInfo`` using its own ``begin()`` and ``end()`` methods. Here's 910 a realistic example: 911 912 .. code-block:: c++ 913 914 #include "llvm/CodeGen/AsmPrinter.h" 915 #include "llvm/IR/Function.h" 916 #include "llvm/IR/DataLayout.h" 917 #include "llvm/Target/TargetAsmInfo.h" 918 #include "llvm/Target/TargetMachine.h" 919 920 void MyGCPrinter::beginAssembly(AsmPrinter &AP) { 921 // Nothing to do. 922 } 923 924 void MyGCPrinter::finishAssembly(AsmPrinter &AP) { 925 MCStreamer &OS = AP.OutStreamer; 926 unsigned IntPtrSize = AP.TM.getDataLayout()->getPointerSize(); 927 928 // Put this in the data section. 929 OS.SwitchSection(AP.getObjFileLowering().getDataSection()); 930 931 // For each function... 932 for (iterator FI = begin(), FE = end(); FI != FE; ++FI) { 933 GCFunctionInfo &MD = **FI; 934 935 // A compact GC layout. Emit this data structure: 936 // 937 // struct { 938 // int32_t PointCount; 939 // void *SafePointAddress[PointCount]; 940 // int32_t StackFrameSize; // in words 941 // int32_t StackArity; 942 // int32_t LiveCount; 943 // int32_t LiveOffsets[LiveCount]; 944 // } __gcmap_<FUNCTIONNAME>; 945 946 // Align to address width. 947 AP.EmitAlignment(IntPtrSize == 4 ? 2 : 3); 948 949 // Emit PointCount. 950 OS.AddComment("safe point count"); 951 AP.EmitInt32(MD.size()); 952 953 // And each safe point... 954 for (GCFunctionInfo::iterator PI = MD.begin(), 955 PE = MD.end(); PI != PE; ++PI) { 956 // Emit the address of the safe point. 957 OS.AddComment("safe point address"); 958 MCSymbol *Label = PI->Label; 959 AP.EmitLabelPlusOffset(Label/*Hi*/, 0/*Offset*/, 4/*Size*/); 960 } 961 962 // Stack information never change in safe points! Only print info from the 963 // first call-site. 964 GCFunctionInfo::iterator PI = MD.begin(); 965 966 // Emit the stack frame size. 967 OS.AddComment("stack frame size (in words)"); 968 AP.EmitInt32(MD.getFrameSize() / IntPtrSize); 969 970 // Emit stack arity, i.e. the number of stacked arguments. 971 unsigned RegisteredArgs = IntPtrSize == 4 ? 5 : 6; 972 unsigned StackArity = MD.getFunction().arg_size() > RegisteredArgs ? 973 MD.getFunction().arg_size() - RegisteredArgs : 0; 974 OS.AddComment("stack arity"); 975 AP.EmitInt32(StackArity); 976 977 // Emit the number of live roots in the function. 978 OS.AddComment("live root count"); 979 AP.EmitInt32(MD.live_size(PI)); 980 981 // And for each live root... 982 for (GCFunctionInfo::live_iterator LI = MD.live_begin(PI), 983 LE = MD.live_end(PI); 984 LI != LE; ++LI) { 985 // Emit live root's offset within the stack frame. 986 OS.AddComment("stack index (offset / wordsize)"); 987 AP.EmitInt32(LI->StackOffset); 988 } 989 } 990 } 991 992 References 993 ========== 994 995 .. _appel89: 996 997 [Appel89] Runtime Tags Aren't Necessary. Andrew W. Appel. Lisp and Symbolic 998 Computation 19(7):703-705, July 1989. 999 1000 .. _goldberg91: 1001 1002 [Goldberg91] Tag-free garbage collection for strongly typed programming 1003 languages. Benjamin Goldberg. ACM SIGPLAN PLDI'91. 1004 1005 .. _tolmach94: 1006 1007 [Tolmach94] Tag-free garbage collection using explicit type parameters. Andrew 1008 Tolmach. Proceedings of the 1994 ACM conference on LISP and functional 1009 programming. 1010 1011 .. _henderson02: 1012 1013 [Henderson2002] `Accurate Garbage Collection in an Uncooperative Environment 1014 <http://citeseer.ist.psu.edu/henderson02accurate.html>`__ 1015
