526 lines · plain
1=======================================2The Often Misunderstood GEP Instruction3=======================================4 5.. contents::6 :local:7 8Introduction9============10 11This document seeks to dispel the mystery and confusion surrounding LLVM's12`GetElementPtr <LangRef.html#getelementptr-instruction>`_ (GEP) instruction.13Questions about the wily GEP instruction are probably the most frequent14questions once a developer gets down to coding with LLVM. Here we lay15out the sources of confusion and show that the GEP instruction is really quite16simple.17 18Address Computation19===================20 21When people are first confronted with the GEP instruction, they tend to relate22it to known concepts from other programming paradigms, most notably C array23indexing and field selection. GEP closely resembles C array indexing and field24selection, however it is a little different and this leads to the following25questions.26 27What is the first index of the GEP instruction?28-----------------------------------------------29 30Quick answer: The index stepping through the second operand.31 32The confusion with the first index usually arises from thinking about the33GetElementPtr instruction as if it were a C index operator. They aren't the34same. For example, when we write, in C:35 36.. code-block:: c++37 38 AType *Foo;39 ...40 X = &Foo->F;41 42it is natural to think that there is only one index, the selection of the field43``F``. However, in this example, ``Foo`` is a pointer. That pointer44must be indexed explicitly in LLVM. C, on the other hand, indexes through it45transparently. To arrive at the same address location as the C code, you would46provide the GEP instruction with two index operands. The first operand indexes47through the pointer; the second operand indexes the field ``F`` of the48structure, just as if you wrote:49 50.. code-block:: c++51 52 X = &Foo[0].F;53 54Sometimes this question gets rephrased as:55 56.. _GEP index through first pointer:57 58 *Why is it okay to index through the first pointer, but subsequent pointers59 won't be dereferenced?*60 61The answer is simply because memory does not have to be accessed to perform the62computation. The second operand to the GEP instruction must be a value of a63pointer type. The value of the pointer is provided directly to the GEP64instruction as an operand without any need for accessing memory. It must,65therefore, be indexed and requires an index operand. Consider this example:66 67.. code-block:: c++68 69 struct munger_struct {70 int f1;71 int f2;72 };73 void munge(struct munger_struct *P) {74 P[0].f1 = P[1].f1 + P[2].f2;75 }76 ...77 struct munger_struct Array[3];78 ...79 munge(Array);80 81In this "C" example, the front end compiler (Clang) will generate three GEP82instructions for the three indices through "P" in the assignment statement. The83function argument ``P`` will be the second operand of each of these GEP84instructions. The third operand indexes through that pointer. The fourth85operand will be the field offset into the ``struct munger_struct`` type, for86either the ``f1`` or ``f2`` field. So, in LLVM assembly the ``munge`` function87looks like:88 89.. code-block:: llvm90 91 define void @munge(ptr %P) {92 entry:93 %tmp = getelementptr %struct.munger_struct, ptr %P, i32 1, i32 094 %tmp1 = load i32, ptr %tmp95 %tmp2 = getelementptr %struct.munger_struct, ptr %P, i32 2, i32 196 %tmp3 = load i32, ptr %tmp297 %tmp4 = add i32 %tmp3, %tmp198 %tmp5 = getelementptr %struct.munger_struct, ptr %P, i32 0, i32 099 store i32 %tmp4, ptr %tmp5100 ret void101 }102 103In each case the second operand is the pointer through which the GEP instruction104starts. The same is true whether the second operand is an argument, allocated105memory, or a global variable.106 107To make this clear, let's consider a more obtuse example:108 109.. code-block:: text110 111 @MyVar = external global i32112 ...113 %idx1 = getelementptr i32, ptr @MyVar, i64 0114 %idx2 = getelementptr i32, ptr @MyVar, i64 1115 %idx3 = getelementptr i32, ptr @MyVar, i64 2116 117These GEP instructions are simply making address computations from the base118address of ``MyVar``. They compute, as follows (using C syntax):119 120.. code-block:: c++121 122 idx1 = (char*) &MyVar + 0123 idx2 = (char*) &MyVar + 4124 idx3 = (char*) &MyVar + 8125 126Since the type ``i32`` is known to be four bytes long, the indices 0, 1 and 2127translate into memory offsets of 0, 4, and 8, respectively. No memory is128accessed to make these computations because the address of ``@MyVar`` is passed129directly to the GEP instructions.130 131The obtuse part of this example is in the cases of ``%idx2`` and ``%idx3``. They132result in the computation of addresses that point to memory past the end of the133``@MyVar`` global, which is only one ``i32`` long, not three ``i32``\s long.134While this is legal in LLVM, it is inadvisable because any load or store with135the pointer that results from these GEP instructions would trigger undefined136behavior (UB).137 138Why is the extra 0 index required?139----------------------------------140 141Quick answer: there are no superfluous indices.142 143This question arises most often when the GEP instruction is applied to a global144variable which is always a pointer type. For example, consider this:145 146.. code-block:: text147 148 %MyStruct = external global { ptr, i32 }149 ...150 %idx = getelementptr { ptr, i32 }, ptr %MyStruct, i64 0, i32 1151 152The GEP above yields a ``ptr`` by indexing the ``i32`` typed field of the153structure ``%MyStruct``. When people first look at it, they wonder why the ``i641540`` index is needed. However, a closer inspection of how globals and GEPs work155reveals the need. Becoming aware of the following facts will dispel the156confusion:157 158#. The type of ``%MyStruct`` is *not* ``{ ptr, i32 }`` but rather ``ptr``.159 That is, ``%MyStruct`` is a pointer (to a structure), not a structure itself.160 161#. Point #1 is evidenced by noticing the type of the second operand of the GEP162 instruction (``%MyStruct``) which is ``ptr``.163 164#. The first index, ``i64 0`` is required to step over the global variable165 ``%MyStruct``. Since the second argument to the GEP instruction must always166 be a value of pointer type, the first index steps through that pointer. A167 value of 0 means 0 elements offset from that pointer.168 169#. The second index, ``i32 1`` selects the second field of the structure (the170 ``i32``).171 172What is dereferenced by GEP?173----------------------------174 175Quick answer: nothing.176 177The GetElementPtr instruction dereferences nothing. That is, it doesn't access178memory in any way. That's what the Load and Store instructions are for. GEP is179only involved in the computation of addresses. For example, consider this:180 181.. code-block:: text182 183 @MyVar = external global { i32, ptr }184 ...185 %idx = getelementptr { i32, ptr }, ptr @MyVar, i64 0, i32 1186 %arr = load ptr, ptr %idx187 %idx = getelementptr [40 x i32], ptr %arr, i64 0, i64 17188 189In this example, we have a global variable, ``@MyVar``, which is a pointer to190a structure containing a pointer. Let's assume that this inner pointer points191to an array of type ``[40 x i32]``. The above IR will first compute the address192of the inner pointer, then load the pointer, and then compute the address of193the 18th array element.194 195This cannot be expressed in a single GEP instruction, because it requires196a memory dereference in between. However, the following example would work197fine:198 199.. code-block:: text200 201 @MyVar = external global { i32, [40 x i32 ] }202 ...203 %idx = getelementptr { i32, [40 x i32] }, ptr @MyVar, i64 0, i32 1, i64 17204 205In this case, the structure does not contain a pointer and the GEP instruction206can index through the global variable, into the second field of the structure207and access the 18th ``i32`` in the array there.208 209Why don't GEP x,0,0,1 and GEP x,1 alias?210----------------------------------------211 212Quick Answer: They compute different address locations.213 214If you look at the first indices in these GEP instructions you find that they215are different (0 and 1), therefore the address computation diverges with that216index. Consider this example:217 218.. code-block:: llvm219 220 @MyVar = external global { [10 x i32] }221 %idx1 = getelementptr { [10 x i32] }, ptr @MyVar, i64 0, i32 0, i64 1222 %idx2 = getelementptr { [10 x i32] }, ptr @MyVar, i64 1223 224In this example, ``idx1`` computes the address of the second integer in the225array that is in the structure in ``@MyVar``, that is ``MyVar+4``. However,226``idx2`` computes the address of *the next* structure after ``@MyVar``, that is227``MyVar+40``, because it indexes past the ten 4-byte integers in ``MyVar``.228Obviously, in such a situation, the pointers don't alias.229 230Why do GEP x,1,0,0 and GEP x,1 alias?231-------------------------------------232 233Quick Answer: They compute the same address location.234 235These two GEP instructions will compute the same address because indexing236through the 0th element does not change the address. Consider this example:237 238.. code-block:: llvm239 240 @MyVar = global { [10 x i32] }241 %idx1 = getelementptr { [10 x i32] }, ptr @MyVar, i64 1, i32 0, i64 0242 %idx2 = getelementptr { [10 x i32] }, ptr @MyVar, i64 1243 244In this example, the value of ``%idx1`` is ``MyVar+40``, and the value of245``%idx2`` is also ``MyVar+40``.246 247Can GEP index into vector elements?248-----------------------------------249 250This hasn't always been forcefully disallowed, though it's not recommended. It251leads to awkward special cases in the optimizers, and fundamental inconsistency252in the IR. In the future, it will probably be outright disallowed.253 254What effect do address spaces have on GEPs?255-------------------------------------------256 257None, except that the address space qualifier on the second operand pointer type258always matches the address space qualifier on the result type.259 260How is GEP different from ``ptrtoint``, arithmetic, and ``inttoptr``?261---------------------------------------------------------------------262 263It's very similar; there are only subtle differences.264 265With ptrtoint, you have to pick an integer type. One approach is to pick i64;266this is safe on everything LLVM supports (LLVM internally assumes pointers are267never wider than 64 bits in many places), and the optimizer will actually narrow268the i64 arithmetic down to the actual pointer size on targets which don't269support 64-bit arithmetic in most cases. However, there are some cases where it270doesn't do this. With GEP you can avoid this problem.271 272Also, GEP carries additional pointer aliasing rules. It's invalid to take a GEP273from one object, address into a different separately allocated object, and274dereference it. IR producers (front-ends) must follow this rule, and consumers275(optimizers, specifically alias analysis) benefit from being able to rely on276it. See the `Rules`_ section for more information.277 278And, GEP is more concise in common cases.279 280However, for the underlying integer computation implied, there is no281difference.282 283 284I'm writing a backend for a target which needs custom lowering for GEP. How do I do this?285-----------------------------------------------------------------------------------------286 287You don't. The integer computation implied by a GEP is target-independent.288Typically what you'll need to do is make your backend pattern-match expression289trees involving ADD, MUL, etc., which are what GEP is lowered into. This has the290advantage of letting your code work correctly in more cases.291 292GEP does use target-dependent parameters for the size and layout of data types,293which targets can customize.294 295If you require support for addressing units which are not 8 bits, you'll need to296fix a lot of code in the backend, with GEP lowering being only a small piece of297the overall picture.298 299How does VLA addressing work with GEPs?300---------------------------------------301 302GEPs don't natively support VLAs. LLVM's type system is entirely static, and GEP303address computations are guided by an LLVM type.304 305VLA indices can be implemented as linearized indices. For example, an expression306like ``X[a][b][c]``, must be effectively lowered into a form like307``X[a*m+b*n+c]``, so that it appears to the GEP as a single-dimensional array308reference.309 310This means if you want to write an analysis which understands array indices and311you want to support VLAs, your code will have to be prepared to reverse-engineer312the linearization. One way to solve this problem is to use the ScalarEvolution313library, which always presents VLA and non-VLA indexing in the same manner.314 315.. _Rules:316 317Rules318=====319 320What happens if an array index is out of bounds?321------------------------------------------------322 323There are two senses in which an array index can be out of bounds.324 325First, there's the array type which comes from the (static) type of the first326operand to the GEP. Indices greater than the number of elements in the327corresponding static array type are valid. There is no problem with out of328bounds indices in this sense. Indexing into an array only depends on the size of329the array element, not the number of elements.330 331A common example of how this is used is arrays where the size is not known.332It's common to use array types with zero length to represent these. The fact333that the static type says there are zero elements is irrelevant; it's perfectly334valid to compute arbitrary element indices, as the computation only depends on335the size of the array element, not the number of elements. Note that zero-sized336arrays are not a special case here.337 338This sense is unconnected with ``inbounds`` keyword. The ``inbounds`` keyword is339designed to describe low-level pointer arithmetic overflow conditions, rather340than high-level array indexing rules.341 342Analysis passes which wish to understand array indexing should not assume that343the static array type bounds are respected.344 345The second sense of being out of bounds is computing an address that's beyond346the actual underlying allocated object.347 348With the ``inbounds`` keyword, the result value of the GEP is ``poison`` if the349address is outside the actual underlying allocated object and not the address350one-past-the-end.351 352Without the ``inbounds`` keyword, there are no restrictions on computing353out-of-bounds addresses. Obviously, performing a load or a store requires an354address of allocated and sufficiently aligned memory. But the GEP itself is only355concerned with computing addresses.356 357Can array indices be negative?358------------------------------359 360Yes. This is basically a special case of array indices being out of bounds.361 362Can I compare two values computed with GEPs?363--------------------------------------------364 365Yes. If both addresses are within the same allocated object, or366one-past-the-end, you'll get the comparison result you expect. If either is367outside of it, integer arithmetic wrapping may occur, so the comparison may not368be meaningful.369 370Can I do GEP with a different pointer type than the type of the underlying object?371----------------------------------------------------------------------------------372 373Yes. There are no restrictions on bitcasting a pointer value to an arbitrary374pointer type. The types in a GEP serve only to define the parameters for the375underlying integer computation. They need not correspond with the actual type of376the underlying object.377 378Furthermore, loads and stores don't have to use the same types as the type of379the underlying object. Types in this context serve only to specify memory size380and alignment. Beyond that they are merely a hint to the optimizer indicating381how the value will likely be used.382 383Can I cast an object's address to integer and add it to null?384-------------------------------------------------------------385 386You can compute an address that way, but if you use GEP to do the add, you can't387use that pointer to actually access the object, unless the object is managed388outside of LLVM.389 390The underlying integer computation is sufficiently defined; null has a defined391value --- zero --- and you can add whatever value you want to it.392 393However, it's invalid to access (load from or store to) an LLVM-aware object394with such a pointer. This includes ``GlobalVariables``, ``Allocas``, and objects395pointed to by noalias pointers.396 397If you really need this functionality, you can do the arithmetic with explicit398integer instructions, and use inttoptr to convert the result to an address. Most399of GEP's special aliasing rules do not apply to pointers computed from ptrtoint,400arithmetic, and inttoptr sequences.401 402Can I compute the distance between two objects, and add that value to one address to compute the other address?403---------------------------------------------------------------------------------------------------------------404 405As with arithmetic on null, you can use GEP to compute an address that way, but406you can't use that pointer to actually access the object if you do, unless the407object is managed outside of LLVM.408 409Also as above, ptrtoint and inttoptr provide an alternative way to do this which410do not have this restriction.411 412Can I do type-based alias analysis on LLVM IR?413----------------------------------------------414 415You can't do type-based alias analysis using LLVM's built-in type system,416because LLVM has no restrictions on mixing types in addressing, loads or stores.417 418LLVM's type-based alias analysis pass uses metadata to describe a different type419system (such as the C type system), and performs type-based aliasing on top of420that. Further details are in the421`language reference <LangRef.html#tbaa-metadata>`_.422 423What happens if a GEP computation overflows?424--------------------------------------------425 426If the GEP lacks the ``inbounds`` keyword, the value is the result from427evaluating the implied two's complement integer computation. However, since428there's no guarantee of where an object will be allocated in the address space,429such values have limited meaning.430 431If the GEP has the ``inbounds`` keyword, the result value is ``poison``432if the GEP overflows (i.e. wraps around the end of the address space).433 434As such, there are some ramifications of this for inbounds GEPs: scales implied435by array/vector/pointer indices are always known to be "nsw" since they are436signed values that are scaled by the element size. These values are also437allowed to be negative (e.g. "``gep i32, ptr %P, i32 -1``") but the pointer438itself is logically treated as an unsigned value. This means that GEPs have an439asymmetric relation between the pointer base (which is treated as unsigned) and440the offset applied to it (which is treated as signed). The result of the441additions within the offset calculation cannot have signed overflow, but when442applied to the base pointer, there can be signed overflow.443 444How can I tell if my front-end is following the rules?445------------------------------------------------------446 447There is currently no checker for the getelementptr rules. Currently, the only448way to do this is to manually check each place in your front-end where449GetElementPtr operators are created.450 451It's not possible to write a checker which could find all rule violations452statically. It would be possible to write a checker which works by instrumenting453the code with dynamic checks though. Alternatively, it would be possible to454write a static checker which catches a subset of possible problems. However, no455such checker exists today.456 457Rationale458=========459 460Why is GEP designed this way?461-----------------------------462 463The design of GEP has the following goals, in rough unofficial order of464priority:465 466* Support C, C-like languages, and languages which can be conceptually lowered467 into C (this covers a lot).468 469* Support optimizations such as those that are common in C compilers. In470 particular, GEP is a cornerstone of LLVM's `pointer aliasing471 model <LangRef.html#pointeraliasing>`_.472 473* Provide a consistent method for computing addresses so that address474 computations don't need to be a part of load and store instructions in the IR.475 476* Support non-C-like languages, to the extent that it doesn't interfere with477 other goals.478 479* Minimize target-specific information in the IR.480 481Why do struct member indices always use ``i32``?482------------------------------------------------483 484The specific type i32 is probably just a historical artifact, however it's wide485enough for all practical purposes, so there's been no need to change it. It486doesn't necessarily imply i32 address arithmetic; it's just an identifier which487identifies a field in a struct. Requiring that all struct indices be the same488reduces the range of possibilities for cases where two GEPs are effectively the489same but have distinct operand types.490 491What's an uglygep?492------------------493 494Some LLVM optimizers operate on GEPs by internally lowering them into more495primitive integer expressions, which allows them to be combined with other496integer expressions and/or split into multiple separate integer expressions. If497they've made non-trivial changes, translating back into LLVM IR can involve498reverse-engineering the structure of the addressing in order to fit it into the499static type of the original first operand. It isn't always possible to fully500reconstruct this structure; sometimes the underlying addressing doesn't501correspond with the static type at all. In such cases the optimizer instead will502emit a GEP with the base pointer cast to a simple address-unit pointer, using503the name "uglygep". This isn't pretty, but it's just as valid, and it's504sufficient to preserve the pointer aliasing guarantees that GEP provides.505 506Summary507=======508 509In summary, here are some things to always remember about the GetElementPtr510instruction:511 512 513#. The GEP instruction never accesses memory, it only provides pointer514 computations.515 516#. The second operand to the GEP instruction is always a pointer and it must be517 indexed.518 519#. There are no superfluous indices for the GEP instruction.520 521#. Trailing zero indices are superfluous for pointer aliasing, but not for the522 types of the pointers.523 524#. Leading zero indices are not superfluous for pointer aliasing nor the types525 of the pointers.526