1.. _gep: 2 3======================================= 4The Often Misunderstood GEP Instruction 5======================================= 6 7.. contents:: 8 :local: 9 10Introduction 11============ 12 13This document seeks to dispel the mystery and confusion surrounding LLVM's 14`GetElementPtr <LangRef.html#i_getelementptr>`_ (GEP) instruction. Questions 15about the wily GEP instruction are probably the most frequently occurring 16questions once a developer gets down to coding with LLVM. Here we lay out the 17sources of confusion and show that the GEP instruction is really quite simple. 18 19Address Computation 20=================== 21 22When people are first confronted with the GEP instruction, they tend to relate 23it to known concepts from other programming paradigms, most notably C array 24indexing and field selection. GEP closely resembles C array indexing and field 25selection, however it's is a little different and this leads to the following 26questions. 27 28What is the first index of the GEP instruction? 29----------------------------------------------- 30 31Quick answer: The index stepping through the first operand. 32 33The confusion with the first index usually arises from thinking about the 34GetElementPtr instruction as if it was a C index operator. They aren't the 35same. For example, when we write, in "C": 36 37.. code-block:: c++ 38 39 AType *Foo; 40 ... 41 X = &Foo->F; 42 43it 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 45must be indexed explicitly in LLVM. C, on the other hand, indices through it 46transparently. To arrive at the same address location as the C code, you would 47provide the GEP instruction with two index operands. The first operand indexes 48through the pointer; the second operand indexes the field ``F`` of the 49structure, just as if you wrote: 50 51.. code-block:: c++ 52 53 X = &Foo[0].F; 54 55Sometimes 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 62The answer is simply because memory does not have to be accessed to perform the 63computation. The first operand to the GEP instruction must be a value of a 64pointer type. The value of the pointer is provided directly to the GEP 65instruction as an operand without any need for accessing memory. It must, 66therefore 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 82In this "C" example, the front end compiler (llvm-gcc) will generate three GEP 83instructions for the three indices through "P" in the assignment statement. The 84function argument ``P`` will be the first operand of each of these GEP 85instructions. The second operand indexes through that pointer. The third 86operand will be the field offset into the ``struct munger_struct`` type, for 87either the ``f1`` or ``f2`` field. So, in LLVM assembly the ``munge`` function 88looks 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 104In each case the first operand is the pointer through which the GEP instruction 105starts. The same is true whether the first operand is an argument, allocated 106memory, or a global variable. 107 108To 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 118These GEP instructions are simply making address computations from the base 119address 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 127Since the type ``i32`` is known to be four bytes long, the indices 0, 1 and 2 128translate into memory offsets of 0, 4, and 8, respectively. No memory is 129accessed to make these computations because the address of ``%MyVar`` is passed 130directly to the GEP instructions. 131 132The obtuse part of this example is in the cases of ``%idx2`` and ``%idx3``. They 133result 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. 135While this is legal in LLVM, it is inadvisable because any load or store with 136the pointer that results from these GEP instructions would produce undefined 137results. 138 139Why is the extra 0 index required? 140---------------------------------- 141 142Quick answer: there are no superfluous indices. 143 144This question arises most often when the GEP instruction is applied to a global 145variable 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 153The GEP above yields an ``i32*`` by indexing the ``i32`` typed field of the 154structure ``%MyStruct``. When people first look at it, they wonder why the ``i64 1550`` index is needed. However, a closer inspection of how globals and GEPs work 156reveals the need. Becoming aware of the following facts will dispel the 157confusion: 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 174What is dereferenced by GEP? 175---------------------------- 176 177Quick answer: nothing. 178 179The GetElementPtr instruction dereferences nothing. That is, it doesn't access 180memory in any way. That's what the Load and Store instructions are for. GEP is 181only 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 189In this example, we have a global variable, ``%MyVar`` that is a pointer to a 190structure containing a pointer to an array of 40 ints. The GEP instruction seems 191to be accessing the 18th integer of the structure's array of ints. However, this 192is actually an illegal GEP instruction. It won't compile. The reason is that the 193pointer in the structure <i>must</i> be dereferenced in order to index into the 194array of 40 ints. Since the GEP instruction never accesses memory, it is 195illegal. 196 197In order to access the 18th integer in the array, you would need to do the 198following: 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 206In this case, we have to load the pointer in the structure with a load 207instruction 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 215then everything works fine. In this case, the structure does not contain a 216pointer and the GEP instruction can index through the global variable, into the 217first field of the structure and access the 18th ``i32`` in the array there. 218 219Why don't GEP x,0,0,1 and GEP x,1 alias? 220---------------------------------------- 221 222Quick Answer: They compute different address locations. 223 224If you look at the first indices in these GEP instructions you find that they 225are different (0 and 1), therefore the address computation diverges with that 226index. 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 234In this example, ``idx1`` computes the address of the second integer in the 235array 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* 237structure after ``%MyVar``. The type of ``idx2`` is ``{ [10 x i32] }*`` and its 238value is equivalent to ``MyVar + 40`` because it indexes past the ten 4-byte 239integers in ``MyVar``. Obviously, in such a situation, the pointers don't 240alias. 241 242Why do GEP x,1,0,0 and GEP x,1 alias? 243------------------------------------- 244 245Quick Answer: They compute the same address location. 246 247These two GEP instructions will compute the same address because indexing 248through the 0th element does not change the address. However, it does change the 249type. 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 257In 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 259i32] }*``. 260 261Can GEP index into vector elements? 262----------------------------------- 263 264This hasn't always been forcefully disallowed, though it's not recommended. It 265leads to awkward special cases in the optimizers, and fundamental inconsistency 266in the IR. In the future, it will probably be outright disallowed. 267 268What effect do address spaces have on GEPs? 269------------------------------------------- 270 271None, except that the address space qualifier on the first operand pointer type 272always matches the address space qualifier on the result type. 273 274How is GEP different from ``ptrtoint``, arithmetic, and ``inttoptr``? 275--------------------------------------------------------------------- 276 277It's very similar; there are only subtle differences. 278 279With ptrtoint, you have to pick an integer type. One approach is to pick i64; 280this is safe on everything LLVM supports (LLVM internally assumes pointers are 281never wider than 64 bits in many places), and the optimizer will actually narrow 282the i64 arithmetic down to the actual pointer size on targets which don't 283support 64-bit arithmetic in most cases. However, there are some cases where it 284doesn't do this. With GEP you can avoid this problem. 285 286Also, GEP carries additional pointer aliasing rules. It's invalid to take a GEP 287from one object, address into a different separately allocated object, and 288dereference it. IR producers (front-ends) must follow this rule, and consumers 289(optimizers, specifically alias analysis) benefit from being able to rely on 290it. See the `Rules`_ section for more information. 291 292And, GEP is more concise in common cases. 293 294However, for the underlying integer computation implied, there is no 295difference. 296 297 298I'm writing a backend for a target which needs custom lowering for GEP. How do I do this? 299----------------------------------------------------------------------------------------- 300 301You don't. The integer computation implied by a GEP is target-independent. 302Typically what you'll need to do is make your backend pattern-match expressions 303trees involving ADD, MUL, etc., which are what GEP is lowered into. This has the 304advantage of letting your code work correctly in more cases. 305 306GEP does use target-dependent parameters for the size and layout of data types, 307which targets can customize. 308 309If you require support for addressing units which are not 8 bits, you'll need to 310fix a lot of code in the backend, with GEP lowering being only a small piece of 311the overall picture. 312 313How does VLA addressing work with GEPs? 314--------------------------------------- 315 316GEPs don't natively support VLAs. LLVM's type system is entirely static, and GEP 317address computations are guided by an LLVM type. 318 319VLA indices can be implemented as linearized indices. For example, an expression 320like ``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 322reference. 323 324This means if you want to write an analysis which understands array indices and 325you want to support VLAs, your code will have to be prepared to reverse-engineer 326the linearization. One way to solve this problem is to use the ScalarEvolution 327library, which always presents VLA and non-VLA indexing in the same manner. 328 329.. _Rules: 330 331Rules 332===== 333 334What happens if an array index is out of bounds? 335------------------------------------------------ 336 337There are two senses in which an array index can be out of bounds. 338 339First, there's the array type which comes from the (static) type of the first 340operand to the GEP. Indices greater than the number of elements in the 341corresponding static array type are valid. There is no problem with out of 342bounds indices in this sense. Indexing into an array only depends on the size of 343the array element, not the number of elements. 344 345A common example of how this is used is arrays where the size is not known. 346It's common to use array types with zero length to represent these. The fact 347that the static type says there are zero elements is irrelevant; it's perfectly 348valid to compute arbitrary element indices, as the computation only depends on 349the size of the array element, not the number of elements. Note that zero-sized 350arrays are not a special case here. 351 352This sense is unconnected with ``inbounds`` keyword. The ``inbounds`` keyword is 353designed to describe low-level pointer arithmetic overflow conditions, rather 354than high-level array indexing rules. 355 356Analysis passes which wish to understand array indexing should not assume that 357the static array type bounds are respected. 358 359The second sense of being out of bounds is computing an address that's beyond 360the actual underlying allocated object. 361 362With the ``inbounds`` keyword, the result value of the GEP is undefined if the 363address is outside the actual underlying allocated object and not the address 364one-past-the-end. 365 366Without the ``inbounds`` keyword, there are no restrictions on computing 367out-of-bounds addresses. Obviously, performing a load or a store requires an 368address of allocated and sufficiently aligned memory. But the GEP itself is only 369concerned with computing addresses. 370 371Can array indices be negative? 372------------------------------ 373 374Yes. This is basically a special case of array indices being out of bounds. 375 376Can I compare two values computed with GEPs? 377-------------------------------------------- 378 379Yes. If both addresses are within the same allocated object, or 380one-past-the-end, you'll get the comparison result you expect. If either is 381outside of it, integer arithmetic wrapping may occur, so the comparison may not 382be meaningful. 383 384Can I do GEP with a different pointer type than the type of the underlying object? 385---------------------------------------------------------------------------------- 386 387Yes. There are no restrictions on bitcasting a pointer value to an arbitrary 388pointer type. The types in a GEP serve only to define the parameters for the 389underlying integer computation. They need not correspond with the actual type of 390the underlying object. 391 392Furthermore, loads and stores don't have to use the same types as the type of 393the underlying object. Types in this context serve only to specify memory size 394and alignment. Beyond that there are merely a hint to the optimizer indicating 395how the value will likely be used. 396 397Can I cast an object's address to integer and add it to null? 398------------------------------------------------------------- 399 400You can compute an address that way, but if you use GEP to do the add, you can't 401use that pointer to actually access the object, unless the object is managed 402outside of LLVM. 403 404The underlying integer computation is sufficiently defined; null has a defined 405value --- zero --- and you can add whatever value you want to it. 406 407However, it's invalid to access (load from or store to) an LLVM-aware object 408with such a pointer. This includes ``GlobalVariables``, ``Allocas``, and objects 409pointed to by noalias pointers. 410 411If you really need this functionality, you can do the arithmetic with explicit 412integer instructions, and use inttoptr to convert the result to an address. Most 413of GEP's special aliasing rules do not apply to pointers computed from ptrtoint, 414arithmetic, and inttoptr sequences. 415 416Can I compute the distance between two objects, and add that value to one address to compute the other address? 417--------------------------------------------------------------------------------------------------------------- 418 419As with arithmetic on null, You can use GEP to compute an address that way, but 420you can't use that pointer to actually access the object if you do, unless the 421object is managed outside of LLVM. 422 423Also as above, ptrtoint and inttoptr provide an alternative way to do this which 424do not have this restriction. 425 426Can I do type-based alias analysis on LLVM IR? 427---------------------------------------------- 428 429You can't do type-based alias analysis using LLVM's built-in type system, 430because LLVM has no restrictions on mixing types in addressing, loads or stores. 431 432LLVM's type-based alias analysis pass uses metadata to describe a different type 433system (such as the C type system), and performs type-based aliasing on top of 434that. Further details are in the `language reference <LangRef.html#tbaa>`_. 435 436What happens if a GEP computation overflows? 437-------------------------------------------- 438 439If the GEP lacks the ``inbounds`` keyword, the value is the result from 440evaluating the implied two's complement integer computation. However, since 441there's no guarantee of where an object will be allocated in the address space, 442such values have limited meaning. 443 444If the GEP has the ``inbounds`` keyword, the result value is undefined (a "trap 445value") if the GEP overflows (i.e. wraps around the end of the address space). 446 447As such, there are some ramifications of this for inbounds GEPs: scales implied 448by array/vector/pointer indices are always known to be "nsw" since they are 449signed values that are scaled by the element size. These values are also 450allowed to be negative (e.g. "``gep i32 *%P, i32 -1``") but the pointer itself 451is logically treated as an unsigned value. This means that GEPs have an 452asymmetric relation between the pointer base (which is treated as unsigned) and 453the offset applied to it (which is treated as signed). The result of the 454additions within the offset calculation cannot have signed overflow, but when 455applied to the base pointer, there can be signed overflow. 456 457How can I tell if my front-end is following the rules? 458------------------------------------------------------ 459 460There is currently no checker for the getelementptr rules. Currently, the only 461way to do this is to manually check each place in your front-end where 462GetElementPtr operators are created. 463 464It's not possible to write a checker which could find all rule violations 465statically. It would be possible to write a checker which works by instrumenting 466the code with dynamic checks though. Alternatively, it would be possible to 467write a static checker which catches a subset of possible problems. However, no 468such checker exists today. 469 470Rationale 471========= 472 473Why is GEP designed this way? 474----------------------------- 475 476The design of GEP has the following goals, in rough unofficial order of 477priority: 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 494Why do struct member indices always use ``i32``? 495------------------------------------------------ 496 497The specific type i32 is probably just a historical artifact, however it's wide 498enough for all practical purposes, so there's been no need to change it. It 499doesn't necessarily imply i32 address arithmetic; it's just an identifier which 500identifies a field in a struct. Requiring that all struct indices be the same 501reduces the range of possibilities for cases where two GEPs are effectively the 502same but have distinct operand types. 503 504What's an uglygep? 505------------------ 506 507Some LLVM optimizers operate on GEPs by internally lowering them into more 508primitive integer expressions, which allows them to be combined with other 509integer expressions and/or split into multiple separate integer expressions. If 510they've made non-trivial changes, translating back into LLVM IR can involve 511reverse-engineering the structure of the addressing in order to fit it into the 512static type of the original first operand. It isn't always possibly to fully 513reconstruct this structure; sometimes the underlying addressing doesn't 514correspond with the static type at all. In such cases the optimizer instead will 515emit a GEP with the base pointer casted to a simple address-unit pointer, using 516the name "uglygep". This isn't pretty, but it's just as valid, and it's 517sufficient to preserve the pointer aliasing guarantees that GEP provides. 518 519Summary 520======= 521 522In summary, here's some things to always remember about the GetElementPtr 523instruction: 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