1 .. _gep: 2 3 ======================================= 4 The Often Misunderstood GEP Instruction 5 ======================================= 6 7 .. contents:: 8 :local: 9 10 Introduction 11 ============ 12 13 This document seeks to dispel the mystery and confusion surrounding LLVM's 14 `GetElementPtr <LangRef.html#i_getelementptr>`_ (GEP) instruction. Questions 15 about the wily GEP instruction are probably the most frequently occurring 16 questions once a developer gets down to coding with LLVM. Here we lay out the 17 sources of confusion and show that the GEP instruction is really quite simple. 18 19 Address Computation 20 =================== 21 22 When people are first confronted with the GEP instruction, they tend to relate 23 it to known concepts from other programming paradigms, most notably C array 24 indexing and field selection. GEP closely resembles C array indexing and field 25 selection, however it's is a little different and this leads to the following 26 questions. 27 28 What is the first index of the GEP instruction? 29 ----------------------------------------------- 30 31 Quick answer: The index stepping through the first operand. 32 33 The confusion with the first index usually arises from thinking about the 34 GetElementPtr instruction as if it was a C index operator. They aren't the 35 same. For example, when we write, in "C": 36 37 .. code-block:: c++ 38 39 AType *Foo; 40 ... 41 X = &Foo->F; 42 43 it is natural to think that there is only one index, the selection of the field 44 ``F``. However, in this example, ``Foo`` is a pointer. That pointer 45 must be indexed explicitly in LLVM. C, on the other hand, indices through it 46 transparently. To arrive at the same address location as the C code, you would 47 provide the GEP instruction with two index operands. The first operand indexes 48 through the pointer; the second operand indexes the field ``F`` of the 49 structure, just as if you wrote: 50 51 .. code-block:: c++ 52 53 X = &Foo[0].F; 54 55 Sometimes this question gets rephrased as: 56 57 .. _GEP index through first pointer: 58 59 *Why is it okay to index through the first pointer, but subsequent pointers 60 won't be dereferenced?* 61 62 The answer is simply because memory does not have to be accessed to perform the 63 computation. The first operand to the GEP instruction must be a value of a 64 pointer type. The value of the pointer is provided directly to the GEP 65 instruction as an operand without any need for accessing memory. It must, 66 therefore be indexed and requires an index operand. Consider this example: 67 68 .. code-block:: c++ 69 70 struct munger_struct { 71 int f1; 72 int f2; 73 }; 74 void munge(struct munger_struct *P) { 75 P[0].f1 = P[1].f1 + P[2].f2; 76 } 77 ... 78 munger_struct Array[3]; 79 ... 80 munge(Array); 81 82 In this "C" example, the front end compiler (llvm-gcc) will generate three GEP 83 instructions for the three indices through "P" in the assignment statement. The 84 function argument ``P`` will be the first operand of each of these GEP 85 instructions. The second operand indexes through that pointer. The third 86 operand will be the field offset into the ``struct munger_struct`` type, for 87 either the ``f1`` or ``f2`` field. So, in LLVM assembly the ``munge`` function 88 looks like: 89 90 .. code-block:: llvm 91 92 void %munge(%struct.munger_struct* %P) { 93 entry: 94 %tmp = getelementptr %struct.munger_struct* %P, i32 1, i32 0 95 %tmp = load i32* %tmp 96 %tmp6 = getelementptr %struct.munger_struct* %P, i32 2, i32 1 97 %tmp7 = load i32* %tmp6 98 %tmp8 = add i32 %tmp7, %tmp 99 %tmp9 = getelementptr %struct.munger_struct* %P, i32 0, i32 0 100 store i32 %tmp8, i32* %tmp9 101 ret void 102 } 103 104 In each case the first operand is the pointer through which the GEP instruction 105 starts. The same is true whether the first operand is an argument, allocated 106 memory, or a global variable. 107 108 To make this clear, let's consider a more obtuse example: 109 110 .. code-block:: llvm 111 112 %MyVar = uninitialized global i32 113 ... 114 %idx1 = getelementptr i32* %MyVar, i64 0 115 %idx2 = getelementptr i32* %MyVar, i64 1 116 %idx3 = getelementptr i32* %MyVar, i64 2 117 118 These GEP instructions are simply making address computations from the base 119 address of ``MyVar``. They compute, as follows (using C syntax): 120 121 .. code-block:: c++ 122 123 idx1 = (char*) &MyVar + 0 124 idx2 = (char*) &MyVar + 4 125 idx3 = (char*) &MyVar + 8 126 127 Since the type ``i32`` is known to be four bytes long, the indices 0, 1 and 2 128 translate into memory offsets of 0, 4, and 8, respectively. No memory is 129 accessed to make these computations because the address of ``%MyVar`` is passed 130 directly to the GEP instructions. 131 132 The obtuse part of this example is in the cases of ``%idx2`` and ``%idx3``. They 133 result in the computation of addresses that point to memory past the end of the 134 ``%MyVar`` global, which is only one ``i32`` long, not three ``i32``\s long. 135 While this is legal in LLVM, it is inadvisable because any load or store with 136 the pointer that results from these GEP instructions would produce undefined 137 results. 138 139 Why is the extra 0 index required? 140 ---------------------------------- 141 142 Quick answer: there are no superfluous indices. 143 144 This question arises most often when the GEP instruction is applied to a global 145 variable which is always a pointer type. For example, consider this: 146 147 .. code-block:: llvm 148 149 %MyStruct = uninitialized global { float*, i32 } 150 ... 151 %idx = getelementptr { float*, i32 }* %MyStruct, i64 0, i32 1 152 153 The GEP above yields an ``i32*`` by indexing the ``i32`` typed field of the 154 structure ``%MyStruct``. When people first look at it, they wonder why the ``i64 155 0`` index is needed. However, a closer inspection of how globals and GEPs work 156 reveals the need. Becoming aware of the following facts will dispel the 157 confusion: 158 159 #. The type of ``%MyStruct`` is *not* ``{ float*, i32 }`` but rather ``{ float*, 160 i32 }*``. That is, ``%MyStruct`` is a pointer to a structure containing a 161 pointer to a ``float`` and an ``i32``. 162 163 #. Point #1 is evidenced by noticing the type of the first operand of the GEP 164 instruction (``%MyStruct``) which is ``{ float*, i32 }*``. 165 166 #. The first index, ``i64 0`` is required to step over the global variable 167 ``%MyStruct``. Since the first argument to the GEP instruction must always 168 be a value of pointer type, the first index steps through that pointer. A 169 value of 0 means 0 elements offset from that pointer. 170 171 #. The second index, ``i32 1`` selects the second field of the structure (the 172 ``i32``). 173 174 What is dereferenced by GEP? 175 ---------------------------- 176 177 Quick answer: nothing. 178 179 The GetElementPtr instruction dereferences nothing. That is, it doesn't access 180 memory in any way. That's what the Load and Store instructions are for. GEP is 181 only involved in the computation of addresses. For example, consider this: 182 183 .. code-block:: llvm 184 185 %MyVar = uninitialized global { [40 x i32 ]* } 186 ... 187 %idx = getelementptr { [40 x i32]* }* %MyVar, i64 0, i32 0, i64 0, i64 17 188 189 In this example, we have a global variable, ``%MyVar`` that is a pointer to a 190 structure containing a pointer to an array of 40 ints. The GEP instruction seems 191 to be accessing the 18th integer of the structure's array of ints. However, this 192 is actually an illegal GEP instruction. It won't compile. The reason is that the 193 pointer in the structure <i>must</i> be dereferenced in order to index into the 194 array of 40 ints. Since the GEP instruction never accesses memory, it is 195 illegal. 196 197 In order to access the 18th integer in the array, you would need to do the 198 following: 199 200 .. code-block:: llvm 201 202 %idx = getelementptr { [40 x i32]* }* %, i64 0, i32 0 203 %arr = load [40 x i32]** %idx 204 %idx = getelementptr [40 x i32]* %arr, i64 0, i64 17 205 206 In this case, we have to load the pointer in the structure with a load 207 instruction before we can index into the array. If the example was changed to: 208 209 .. code-block:: llvm 210 211 %MyVar = uninitialized global { [40 x i32 ] } 212 ... 213 %idx = getelementptr { [40 x i32] }*, i64 0, i32 0, i64 17 214 215 then everything works fine. In this case, the structure does not contain a 216 pointer and the GEP instruction can index through the global variable, into the 217 first field of the structure and access the 18th ``i32`` in the array there. 218 219 Why don't GEP x,0,0,1 and GEP x,1 alias? 220 ---------------------------------------- 221 222 Quick Answer: They compute different address locations. 223 224 If you look at the first indices in these GEP instructions you find that they 225 are different (0 and 1), therefore the address computation diverges with that 226 index. Consider this example: 227 228 .. code-block:: llvm 229 230 %MyVar = global { [10 x i32 ] } 231 %idx1 = getelementptr { [10 x i32 ] }* %MyVar, i64 0, i32 0, i64 1 232 %idx2 = getelementptr { [10 x i32 ] }* %MyVar, i64 1 233 234 In this example, ``idx1`` computes the address of the second integer in the 235 array that is in the structure in ``%MyVar``, that is ``MyVar+4``. The type of 236 ``idx1`` is ``i32*``. However, ``idx2`` computes the address of *the next* 237 structure after ``%MyVar``. The type of ``idx2`` is ``{ [10 x i32] }*`` and its 238 value is equivalent to ``MyVar + 40`` because it indexes past the ten 4-byte 239 integers in ``MyVar``. Obviously, in such a situation, the pointers don't 240 alias. 241 242 Why do GEP x,1,0,0 and GEP x,1 alias? 243 ------------------------------------- 244 245 Quick Answer: They compute the same address location. 246 247 These two GEP instructions will compute the same address because indexing 248 through the 0th element does not change the address. However, it does change the 249 type. Consider this example: 250 251 .. code-block:: llvm 252 253 %MyVar = global { [10 x i32 ] } 254 %idx1 = getelementptr { [10 x i32 ] }* %MyVar, i64 1, i32 0, i64 0 255 %idx2 = getelementptr { [10 x i32 ] }* %MyVar, i64 1 256 257 In this example, the value of ``%idx1`` is ``%MyVar+40`` and its type is 258 ``i32*``. The value of ``%idx2`` is also ``MyVar+40`` but its type is ``{ [10 x 259 i32] }*``. 260 261 Can GEP index into vector elements? 262 ----------------------------------- 263 264 This hasn't always been forcefully disallowed, though it's not recommended. It 265 leads to awkward special cases in the optimizers, and fundamental inconsistency 266 in the IR. In the future, it will probably be outright disallowed. 267 268 What effect do address spaces have on GEPs? 269 ------------------------------------------- 270 271 None, except that the address space qualifier on the first operand pointer type 272 always matches the address space qualifier on the result type. 273 274 How is GEP different from ``ptrtoint``, arithmetic, and ``inttoptr``? 275 --------------------------------------------------------------------- 276 277 It's very similar; there are only subtle differences. 278 279 With ptrtoint, you have to pick an integer type. One approach is to pick i64; 280 this is safe on everything LLVM supports (LLVM internally assumes pointers are 281 never wider than 64 bits in many places), and the optimizer will actually narrow 282 the i64 arithmetic down to the actual pointer size on targets which don't 283 support 64-bit arithmetic in most cases. However, there are some cases where it 284 doesn't do this. With GEP you can avoid this problem. 285 286 Also, GEP carries additional pointer aliasing rules. It's invalid to take a GEP 287 from one object, address into a different separately allocated object, and 288 dereference it. IR producers (front-ends) must follow this rule, and consumers 289 (optimizers, specifically alias analysis) benefit from being able to rely on 290 it. See the `Rules`_ section for more information. 291 292 And, GEP is more concise in common cases. 293 294 However, for the underlying integer computation implied, there is no 295 difference. 296 297 298 I'm writing a backend for a target which needs custom lowering for GEP. How do I do this? 299 ----------------------------------------------------------------------------------------- 300 301 You don't. The integer computation implied by a GEP is target-independent. 302 Typically what you'll need to do is make your backend pattern-match expressions 303 trees involving ADD, MUL, etc., which are what GEP is lowered into. This has the 304 advantage of letting your code work correctly in more cases. 305 306 GEP does use target-dependent parameters for the size and layout of data types, 307 which targets can customize. 308 309 If you require support for addressing units which are not 8 bits, you'll need to 310 fix a lot of code in the backend, with GEP lowering being only a small piece of 311 the overall picture. 312 313 How does VLA addressing work with GEPs? 314 --------------------------------------- 315 316 GEPs don't natively support VLAs. LLVM's type system is entirely static, and GEP 317 address computations are guided by an LLVM type. 318 319 VLA indices can be implemented as linearized indices. For example, an expression 320 like ``X[a][b][c]``, must be effectively lowered into a form like 321 ``X[a*m+b*n+c]``, so that it appears to the GEP as a single-dimensional array 322 reference. 323 324 This means if you want to write an analysis which understands array indices and 325 you want to support VLAs, your code will have to be prepared to reverse-engineer 326 the linearization. One way to solve this problem is to use the ScalarEvolution 327 library, which always presents VLA and non-VLA indexing in the same manner. 328 329 .. _Rules: 330 331 Rules 332 ===== 333 334 What happens if an array index is out of bounds? 335 ------------------------------------------------ 336 337 There are two senses in which an array index can be out of bounds. 338 339 First, there's the array type which comes from the (static) type of the first 340 operand to the GEP. Indices greater than the number of elements in the 341 corresponding static array type are valid. There is no problem with out of 342 bounds indices in this sense. Indexing into an array only depends on the size of 343 the array element, not the number of elements. 344 345 A common example of how this is used is arrays where the size is not known. 346 It's common to use array types with zero length to represent these. The fact 347 that the static type says there are zero elements is irrelevant; it's perfectly 348 valid to compute arbitrary element indices, as the computation only depends on 349 the size of the array element, not the number of elements. Note that zero-sized 350 arrays are not a special case here. 351 352 This sense is unconnected with ``inbounds`` keyword. The ``inbounds`` keyword is 353 designed to describe low-level pointer arithmetic overflow conditions, rather 354 than high-level array indexing rules. 355 356 Analysis passes which wish to understand array indexing should not assume that 357 the static array type bounds are respected. 358 359 The second sense of being out of bounds is computing an address that's beyond 360 the actual underlying allocated object. 361 362 With the ``inbounds`` keyword, the result value of the GEP is undefined if the 363 address is outside the actual underlying allocated object and not the address 364 one-past-the-end. 365 366 Without the ``inbounds`` keyword, there are no restrictions on computing 367 out-of-bounds addresses. Obviously, performing a load or a store requires an 368 address of allocated and sufficiently aligned memory. But the GEP itself is only 369 concerned with computing addresses. 370 371 Can array indices be negative? 372 ------------------------------ 373 374 Yes. This is basically a special case of array indices being out of bounds. 375 376 Can I compare two values computed with GEPs? 377 -------------------------------------------- 378 379 Yes. If both addresses are within the same allocated object, or 380 one-past-the-end, you'll get the comparison result you expect. If either is 381 outside of it, integer arithmetic wrapping may occur, so the comparison may not 382 be meaningful. 383 384 Can I do GEP with a different pointer type than the type of the underlying object? 385 ---------------------------------------------------------------------------------- 386 387 Yes. There are no restrictions on bitcasting a pointer value to an arbitrary 388 pointer type. The types in a GEP serve only to define the parameters for the 389 underlying integer computation. They need not correspond with the actual type of 390 the underlying object. 391 392 Furthermore, loads and stores don't have to use the same types as the type of 393 the underlying object. Types in this context serve only to specify memory size 394 and alignment. Beyond that there are merely a hint to the optimizer indicating 395 how the value will likely be used. 396 397 Can I cast an object's address to integer and add it to null? 398 ------------------------------------------------------------- 399 400 You can compute an address that way, but if you use GEP to do the add, you can't 401 use that pointer to actually access the object, unless the object is managed 402 outside of LLVM. 403 404 The underlying integer computation is sufficiently defined; null has a defined 405 value --- zero --- and you can add whatever value you want to it. 406 407 However, it's invalid to access (load from or store to) an LLVM-aware object 408 with such a pointer. This includes ``GlobalVariables``, ``Allocas``, and objects 409 pointed to by noalias pointers. 410 411 If you really need this functionality, you can do the arithmetic with explicit 412 integer instructions, and use inttoptr to convert the result to an address. Most 413 of GEP's special aliasing rules do not apply to pointers computed from ptrtoint, 414 arithmetic, and inttoptr sequences. 415 416 Can I compute the distance between two objects, and add that value to one address to compute the other address? 417 --------------------------------------------------------------------------------------------------------------- 418 419 As with arithmetic on null, You can use GEP to compute an address that way, but 420 you can't use that pointer to actually access the object if you do, unless the 421 object is managed outside of LLVM. 422 423 Also as above, ptrtoint and inttoptr provide an alternative way to do this which 424 do not have this restriction. 425 426 Can I do type-based alias analysis on LLVM IR? 427 ---------------------------------------------- 428 429 You can't do type-based alias analysis using LLVM's built-in type system, 430 because LLVM has no restrictions on mixing types in addressing, loads or stores. 431 432 LLVM's type-based alias analysis pass uses metadata to describe a different type 433 system (such as the C type system), and performs type-based aliasing on top of 434 that. Further details are in the `language reference <LangRef.html#tbaa>`_. 435 436 What happens if a GEP computation overflows? 437 -------------------------------------------- 438 439 If the GEP lacks the ``inbounds`` keyword, the value is the result from 440 evaluating the implied two's complement integer computation. However, since 441 there's no guarantee of where an object will be allocated in the address space, 442 such values have limited meaning. 443 444 If the GEP has the ``inbounds`` keyword, the result value is undefined (a "trap 445 value") if the GEP overflows (i.e. wraps around the end of the address space). 446 447 As such, there are some ramifications of this for inbounds GEPs: scales implied 448 by array/vector/pointer indices are always known to be "nsw" since they are 449 signed values that are scaled by the element size. These values are also 450 allowed to be negative (e.g. "``gep i32 *%P, i32 -1``") but the pointer itself 451 is logically treated as an unsigned value. This means that GEPs have an 452 asymmetric relation between the pointer base (which is treated as unsigned) and 453 the offset applied to it (which is treated as signed). The result of the 454 additions within the offset calculation cannot have signed overflow, but when 455 applied to the base pointer, there can be signed overflow. 456 457 How can I tell if my front-end is following the rules? 458 ------------------------------------------------------ 459 460 There is currently no checker for the getelementptr rules. Currently, the only 461 way to do this is to manually check each place in your front-end where 462 GetElementPtr operators are created. 463 464 It's not possible to write a checker which could find all rule violations 465 statically. It would be possible to write a checker which works by instrumenting 466 the code with dynamic checks though. Alternatively, it would be possible to 467 write a static checker which catches a subset of possible problems. However, no 468 such checker exists today. 469 470 Rationale 471 ========= 472 473 Why is GEP designed this way? 474 ----------------------------- 475 476 The design of GEP has the following goals, in rough unofficial order of 477 priority: 478 479 * Support C, C-like languages, and languages which can be conceptually lowered 480 into C (this covers a lot). 481 482 * Support optimizations such as those that are common in C compilers. In 483 particular, GEP is a cornerstone of LLVM's `pointer aliasing 484 model <LangRef.html#pointeraliasing>`_. 485 486 * Provide a consistent method for computing addresses so that address 487 computations don't need to be a part of load and store instructions in the IR. 488 489 * Support non-C-like languages, to the extent that it doesn't interfere with 490 other goals. 491 492 * Minimize target-specific information in the IR. 493 494 Why do struct member indices always use ``i32``? 495 ------------------------------------------------ 496 497 The specific type i32 is probably just a historical artifact, however it's wide 498 enough for all practical purposes, so there's been no need to change it. It 499 doesn't necessarily imply i32 address arithmetic; it's just an identifier which 500 identifies a field in a struct. Requiring that all struct indices be the same 501 reduces the range of possibilities for cases where two GEPs are effectively the 502 same but have distinct operand types. 503 504 What's an uglygep? 505 ------------------ 506 507 Some LLVM optimizers operate on GEPs by internally lowering them into more 508 primitive integer expressions, which allows them to be combined with other 509 integer expressions and/or split into multiple separate integer expressions. If 510 they've made non-trivial changes, translating back into LLVM IR can involve 511 reverse-engineering the structure of the addressing in order to fit it into the 512 static type of the original first operand. It isn't always possibly to fully 513 reconstruct this structure; sometimes the underlying addressing doesn't 514 correspond with the static type at all. In such cases the optimizer instead will 515 emit a GEP with the base pointer casted to a simple address-unit pointer, using 516 the name "uglygep". This isn't pretty, but it's just as valid, and it's 517 sufficient to preserve the pointer aliasing guarantees that GEP provides. 518 519 Summary 520 ======= 521 522 In summary, here's some things to always remember about the GetElementPtr 523 instruction: 524 525 526 #. The GEP instruction never accesses memory, it only provides pointer 527 computations. 528 529 #. The first operand to the GEP instruction is always a pointer and it must be 530 indexed. 531 532 #. There are no superfluous indices for the GEP instruction. 533 534 #. Trailing zero indices are superfluous for pointer aliasing, but not for the 535 types of the pointers. 536 537 #. Leading zero indices are not superfluous for pointer aliasing nor the types 538 of the pointers. 539
