• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1=======================================
2The Often Misunderstood GEP Instruction
3=======================================
4
5.. contents::
6   :local:
7
8Introduction
9============
10
11This document seeks to dispel the mystery and confusion surrounding LLVM's
12`GetElementPtr <LangRef.html#i_getelementptr>`_ (GEP) instruction.  Questions
13about the wily GEP instruction are probably the most frequently occurring
14questions once a developer gets down to coding with LLVM. Here we lay out the
15sources of confusion and show that the GEP instruction is really quite simple.
16
17Address Computation
18===================
19
20When people are first confronted with the GEP instruction, they tend to relate
21it to known concepts from other programming paradigms, most notably C array
22indexing and field selection. GEP closely resembles C array indexing and field
23selection, however it is a little different and this leads to the following
24questions.
25
26What is the first index of the GEP instruction?
27-----------------------------------------------
28
29Quick answer: The index stepping through the first operand.
30
31The confusion with the first index usually arises from thinking about the
32GetElementPtr instruction as if it was a C index operator. They aren't the
33same. For example, when we write, in "C":
34
35.. code-block:: c++
36
37  AType *Foo;
38  ...
39  X = &Foo->F;
40
41it is natural to think that there is only one index, the selection of the field
42``F``.  However, in this example, ``Foo`` is a pointer. That pointer
43must be indexed explicitly in LLVM. C, on the other hand, indices through it
44transparently.  To arrive at the same address location as the C code, you would
45provide the GEP instruction with two index operands. The first operand indexes
46through the pointer; the second operand indexes the field ``F`` of the
47structure, just as if you wrote:
48
49.. code-block:: c++
50
51  X = &Foo[0].F;
52
53Sometimes this question gets rephrased as:
54
55.. _GEP index through first pointer:
56
57  *Why is it okay to index through the first pointer, but subsequent pointers
58  won't be dereferenced?*
59
60The answer is simply because memory does not have to be accessed to perform the
61computation. The first operand to the GEP instruction must be a value of a
62pointer type. The value of the pointer is provided directly to the GEP
63instruction as an operand without any need for accessing memory. It must,
64therefore be indexed and requires an index operand. Consider this example:
65
66.. code-block:: c++
67
68  struct munger_struct {
69    int f1;
70    int f2;
71  };
72  void munge(struct munger_struct *P) {
73    P[0].f1 = P[1].f1 + P[2].f2;
74  }
75  ...
76  munger_struct Array[3];
77  ...
78  munge(Array);
79
80In this "C" example, the front end compiler (Clang) will generate three GEP
81instructions for the three indices through "P" in the assignment statement.  The
82function argument ``P`` will be the first operand of each of these GEP
83instructions.  The second operand indexes through that pointer.  The third
84operand will be the field offset into the ``struct munger_struct`` type, for
85either the ``f1`` or ``f2`` field. So, in LLVM assembly the ``munge`` function
86looks like:
87
88.. code-block:: llvm
89
90  void %munge(%struct.munger_struct* %P) {
91  entry:
92    %tmp = getelementptr %struct.munger_struct, %struct.munger_struct* %P, i32 1, i32 0
93    %tmp = load i32* %tmp
94    %tmp6 = getelementptr %struct.munger_struct, %struct.munger_struct* %P, i32 2, i32 1
95    %tmp7 = load i32* %tmp6
96    %tmp8 = add i32 %tmp7, %tmp
97    %tmp9 = getelementptr %struct.munger_struct, %struct.munger_struct* %P, i32 0, i32 0
98    store i32 %tmp8, i32* %tmp9
99    ret void
100  }
101
102In each case the first operand is the pointer through which the GEP instruction
103starts. The same is true whether the first operand is an argument, allocated
104memory, or a global variable.
105
106To make this clear, let's consider a more obtuse example:
107
108.. code-block:: llvm
109
110  %MyVar = uninitialized global i32
111  ...
112  %idx1 = getelementptr i32, i32* %MyVar, i64 0
113  %idx2 = getelementptr i32, i32* %MyVar, i64 1
114  %idx3 = getelementptr i32, i32* %MyVar, i64 2
115
116These GEP instructions are simply making address computations from the base
117address of ``MyVar``.  They compute, as follows (using C syntax):
118
119.. code-block:: c++
120
121  idx1 = (char*) &MyVar + 0
122  idx2 = (char*) &MyVar + 4
123  idx3 = (char*) &MyVar + 8
124
125Since the type ``i32`` is known to be four bytes long, the indices 0, 1 and 2
126translate into memory offsets of 0, 4, and 8, respectively. No memory is
127accessed to make these computations because the address of ``%MyVar`` is passed
128directly to the GEP instructions.
129
130The obtuse part of this example is in the cases of ``%idx2`` and ``%idx3``. They
131result in the computation of addresses that point to memory past the end of the
132``%MyVar`` global, which is only one ``i32`` long, not three ``i32``\s long.
133While this is legal in LLVM, it is inadvisable because any load or store with
134the pointer that results from these GEP instructions would produce undefined
135results.
136
137Why is the extra 0 index required?
138----------------------------------
139
140Quick answer: there are no superfluous indices.
141
142This question arises most often when the GEP instruction is applied to a global
143variable which is always a pointer type. For example, consider this:
144
145.. code-block:: llvm
146
147  %MyStruct = uninitialized global { float*, i32 }
148  ...
149  %idx = getelementptr { float*, i32 }, { float*, i32 }* %MyStruct, i64 0, i32 1
150
151The GEP above yields an ``i32*`` by indexing the ``i32`` typed field of the
152structure ``%MyStruct``. When people first look at it, they wonder why the ``i64
1530`` index is needed. However, a closer inspection of how globals and GEPs work
154reveals the need. Becoming aware of the following facts will dispel the
155confusion:
156
157#. The type of ``%MyStruct`` is *not* ``{ float*, i32 }`` but rather ``{ float*,
158   i32 }*``. That is, ``%MyStruct`` is a pointer to a structure containing a
159   pointer to a ``float`` and an ``i32``.
160
161#. Point #1 is evidenced by noticing the type of the first operand of the GEP
162   instruction (``%MyStruct``) which is ``{ float*, i32 }*``.
163
164#. The first index, ``i64 0`` is required to step over the global variable
165   ``%MyStruct``.  Since the first argument to the GEP instruction must always
166   be a value of pointer type, the first index steps through that pointer. A
167   value of 0 means 0 elements offset from that pointer.
168
169#. The second index, ``i32 1`` selects the second field of the structure (the
170   ``i32``).
171
172What is dereferenced by GEP?
173----------------------------
174
175Quick answer: nothing.
176
177The GetElementPtr instruction dereferences nothing. That is, it doesn't access
178memory in any way. That's what the Load and Store instructions are for.  GEP is
179only involved in the computation of addresses. For example, consider this:
180
181.. code-block:: llvm
182
183  %MyVar = uninitialized global { [40 x i32 ]* }
184  ...
185  %idx = getelementptr { [40 x i32]* }, { [40 x i32]* }* %MyVar, i64 0, i32 0, i64 0, i64 17
186
187In this example, we have a global variable, ``%MyVar`` that is a pointer to a
188structure containing a pointer to an array of 40 ints. The GEP instruction seems
189to be accessing the 18th integer of the structure's array of ints. However, this
190is actually an illegal GEP instruction. It won't compile. The reason is that the
191pointer in the structure *must* be dereferenced in order to index into the
192array of 40 ints. Since the GEP instruction never accesses memory, it is
193illegal.
194
195In order to access the 18th integer in the array, you would need to do the
196following:
197
198.. code-block:: llvm
199
200  %idx = getelementptr { [40 x i32]* }, { [40 x i32]* }* %, i64 0, i32 0
201  %arr = load [40 x i32]** %idx
202  %idx = getelementptr [40 x i32], [40 x i32]* %arr, i64 0, i64 17
203
204In this case, we have to load the pointer in the structure with a load
205instruction before we can index into the array. If the example was changed to:
206
207.. code-block:: llvm
208
209  %MyVar = uninitialized global { [40 x i32 ] }
210  ...
211  %idx = getelementptr { [40 x i32] }, { [40 x i32] }*, i64 0, i32 0, i64 17
212
213then everything works fine. In this case, the structure does not contain a
214pointer and the GEP instruction can index through the global variable, into the
215first field of the structure and access the 18th ``i32`` in the array there.
216
217Why don't GEP x,0,0,1 and GEP x,1 alias?
218----------------------------------------
219
220Quick Answer: They compute different address locations.
221
222If you look at the first indices in these GEP instructions you find that they
223are different (0 and 1), therefore the address computation diverges with that
224index. Consider this example:
225
226.. code-block:: llvm
227
228  %MyVar = global { [10 x i32] }
229  %idx1 = getelementptr { [10 x i32] }, { [10 x i32] }* %MyVar, i64 0, i32 0, i64 1
230  %idx2 = getelementptr { [10 x i32] }, { [10 x i32] }* %MyVar, i64 1
231
232In this example, ``idx1`` computes the address of the second integer in the
233array that is in the structure in ``%MyVar``, that is ``MyVar+4``. The type of
234``idx1`` is ``i32*``. However, ``idx2`` computes the address of *the next*
235structure after ``%MyVar``. The type of ``idx2`` is ``{ [10 x i32] }*`` and its
236value is equivalent to ``MyVar + 40`` because it indexes past the ten 4-byte
237integers in ``MyVar``. Obviously, in such a situation, the pointers don't
238alias.
239
240Why do GEP x,1,0,0 and GEP x,1 alias?
241-------------------------------------
242
243Quick Answer: They compute the same address location.
244
245These two GEP instructions will compute the same address because indexing
246through the 0th element does not change the address. However, it does change the
247type. Consider this example:
248
249.. code-block:: llvm
250
251  %MyVar = global { [10 x i32] }
252  %idx1 = getelementptr { [10 x i32] }, { [10 x i32] }* %MyVar, i64 1, i32 0, i64 0
253  %idx2 = getelementptr { [10 x i32] }, { [10 x i32] }* %MyVar, i64 1
254
255In this example, the value of ``%idx1`` is ``%MyVar+40`` and its type is
256``i32*``. The value of ``%idx2`` is also ``MyVar+40`` but its type is ``{ [10 x
257i32] }*``.
258
259Can GEP index into vector elements?
260-----------------------------------
261
262This hasn't always been forcefully disallowed, though it's not recommended.  It
263leads to awkward special cases in the optimizers, and fundamental inconsistency
264in the IR. In the future, it will probably be outright disallowed.
265
266What effect do address spaces have on GEPs?
267-------------------------------------------
268
269None, except that the address space qualifier on the first operand pointer type
270always matches the address space qualifier on the result type.
271
272How is GEP different from ``ptrtoint``, arithmetic, and ``inttoptr``?
273---------------------------------------------------------------------
274
275It's very similar; there are only subtle differences.
276
277With ptrtoint, you have to pick an integer type. One approach is to pick i64;
278this is safe on everything LLVM supports (LLVM internally assumes pointers are
279never wider than 64 bits in many places), and the optimizer will actually narrow
280the i64 arithmetic down to the actual pointer size on targets which don't
281support 64-bit arithmetic in most cases. However, there are some cases where it
282doesn't do this. With GEP you can avoid this problem.
283
284Also, GEP carries additional pointer aliasing rules. It's invalid to take a GEP
285from one object, address into a different separately allocated object, and
286dereference it. IR producers (front-ends) must follow this rule, and consumers
287(optimizers, specifically alias analysis) benefit from being able to rely on
288it. See the `Rules`_ section for more information.
289
290And, GEP is more concise in common cases.
291
292However, for the underlying integer computation implied, there is no
293difference.
294
295
296I'm writing a backend for a target which needs custom lowering for GEP. How do I do this?
297-----------------------------------------------------------------------------------------
298
299You don't. The integer computation implied by a GEP is target-independent.
300Typically what you'll need to do is make your backend pattern-match expressions
301trees involving ADD, MUL, etc., which are what GEP is lowered into. This has the
302advantage of letting your code work correctly in more cases.
303
304GEP does use target-dependent parameters for the size and layout of data types,
305which targets can customize.
306
307If you require support for addressing units which are not 8 bits, you'll need to
308fix a lot of code in the backend, with GEP lowering being only a small piece of
309the overall picture.
310
311How does VLA addressing work with GEPs?
312---------------------------------------
313
314GEPs don't natively support VLAs. LLVM's type system is entirely static, and GEP
315address computations are guided by an LLVM type.
316
317VLA indices can be implemented as linearized indices. For example, an expression
318like ``X[a][b][c]``, must be effectively lowered into a form like
319``X[a*m+b*n+c]``, so that it appears to the GEP as a single-dimensional array
320reference.
321
322This means if you want to write an analysis which understands array indices and
323you want to support VLAs, your code will have to be prepared to reverse-engineer
324the linearization. One way to solve this problem is to use the ScalarEvolution
325library, which always presents VLA and non-VLA indexing in the same manner.
326
327.. _Rules:
328
329Rules
330=====
331
332What happens if an array index is out of bounds?
333------------------------------------------------
334
335There are two senses in which an array index can be out of bounds.
336
337First, there's the array type which comes from the (static) type of the first
338operand to the GEP. Indices greater than the number of elements in the
339corresponding static array type are valid. There is no problem with out of
340bounds indices in this sense. Indexing into an array only depends on the size of
341the array element, not the number of elements.
342
343A common example of how this is used is arrays where the size is not known.
344It's common to use array types with zero length to represent these. The fact
345that the static type says there are zero elements is irrelevant; it's perfectly
346valid to compute arbitrary element indices, as the computation only depends on
347the size of the array element, not the number of elements. Note that zero-sized
348arrays are not a special case here.
349
350This sense is unconnected with ``inbounds`` keyword. The ``inbounds`` keyword is
351designed to describe low-level pointer arithmetic overflow conditions, rather
352than high-level array indexing rules.
353
354Analysis passes which wish to understand array indexing should not assume that
355the static array type bounds are respected.
356
357The second sense of being out of bounds is computing an address that's beyond
358the actual underlying allocated object.
359
360With the ``inbounds`` keyword, the result value of the GEP is undefined if the
361address is outside the actual underlying allocated object and not the address
362one-past-the-end.
363
364Without the ``inbounds`` keyword, there are no restrictions on computing
365out-of-bounds addresses. Obviously, performing a load or a store requires an
366address of allocated and sufficiently aligned memory. But the GEP itself is only
367concerned with computing addresses.
368
369Can array indices be negative?
370------------------------------
371
372Yes. This is basically a special case of array indices being out of bounds.
373
374Can I compare two values computed with GEPs?
375--------------------------------------------
376
377Yes. If both addresses are within the same allocated object, or
378one-past-the-end, you'll get the comparison result you expect. If either is
379outside of it, integer arithmetic wrapping may occur, so the comparison may not
380be meaningful.
381
382Can I do GEP with a different pointer type than the type of the underlying object?
383----------------------------------------------------------------------------------
384
385Yes. There are no restrictions on bitcasting a pointer value to an arbitrary
386pointer type. The types in a GEP serve only to define the parameters for the
387underlying integer computation. They need not correspond with the actual type of
388the underlying object.
389
390Furthermore, loads and stores don't have to use the same types as the type of
391the underlying object. Types in this context serve only to specify memory size
392and alignment. Beyond that there are merely a hint to the optimizer indicating
393how the value will likely be used.
394
395Can I cast an object's address to integer and add it to null?
396-------------------------------------------------------------
397
398You can compute an address that way, but if you use GEP to do the add, you can't
399use that pointer to actually access the object, unless the object is managed
400outside of LLVM.
401
402The underlying integer computation is sufficiently defined; null has a defined
403value --- zero --- and you can add whatever value you want to it.
404
405However, it's invalid to access (load from or store to) an LLVM-aware object
406with such a pointer. This includes ``GlobalVariables``, ``Allocas``, and objects
407pointed to by noalias pointers.
408
409If you really need this functionality, you can do the arithmetic with explicit
410integer instructions, and use inttoptr to convert the result to an address. Most
411of GEP's special aliasing rules do not apply to pointers computed from ptrtoint,
412arithmetic, and inttoptr sequences.
413
414Can I compute the distance between two objects, and add that value to one address to compute the other address?
415---------------------------------------------------------------------------------------------------------------
416
417As with arithmetic on null, you can use GEP to compute an address that way, but
418you can't use that pointer to actually access the object if you do, unless the
419object is managed outside of LLVM.
420
421Also as above, ptrtoint and inttoptr provide an alternative way to do this which
422do not have this restriction.
423
424Can I do type-based alias analysis on LLVM IR?
425----------------------------------------------
426
427You can't do type-based alias analysis using LLVM's built-in type system,
428because LLVM has no restrictions on mixing types in addressing, loads or stores.
429
430LLVM's type-based alias analysis pass uses metadata to describe a different type
431system (such as the C type system), and performs type-based aliasing on top of
432that.  Further details are in the `language reference <LangRef.html#tbaa>`_.
433
434What happens if a GEP computation overflows?
435--------------------------------------------
436
437If the GEP lacks the ``inbounds`` keyword, the value is the result from
438evaluating the implied two's complement integer computation. However, since
439there's no guarantee of where an object will be allocated in the address space,
440such values have limited meaning.
441
442If the GEP has the ``inbounds`` keyword, the result value is undefined (a "trap
443value") if the GEP overflows (i.e. wraps around the end of the address space).
444
445As such, there are some ramifications of this for inbounds GEPs: scales implied
446by array/vector/pointer indices are always known to be "nsw" since they are
447signed values that are scaled by the element size.  These values are also
448allowed to be negative (e.g. "``gep i32 *%P, i32 -1``") but the pointer itself
449is logically treated as an unsigned value.  This means that GEPs have an
450asymmetric relation between the pointer base (which is treated as unsigned) and
451the offset applied to it (which is treated as signed). The result of the
452additions within the offset calculation cannot have signed overflow, but when
453applied to the base pointer, there can be signed overflow.
454
455How can I tell if my front-end is following the rules?
456------------------------------------------------------
457
458There is currently no checker for the getelementptr rules. Currently, the only
459way to do this is to manually check each place in your front-end where
460GetElementPtr operators are created.
461
462It's not possible to write a checker which could find all rule violations
463statically. It would be possible to write a checker which works by instrumenting
464the code with dynamic checks though. Alternatively, it would be possible to
465write a static checker which catches a subset of possible problems. However, no
466such checker exists today.
467
468Rationale
469=========
470
471Why is GEP designed this way?
472-----------------------------
473
474The design of GEP has the following goals, in rough unofficial order of
475priority:
476
477* Support C, C-like languages, and languages which can be conceptually lowered
478  into C (this covers a lot).
479
480* Support optimizations such as those that are common in C compilers. In
481  particular, GEP is a cornerstone of LLVM's `pointer aliasing
482  model <LangRef.html#pointeraliasing>`_.
483
484* Provide a consistent method for computing addresses so that address
485  computations don't need to be a part of load and store instructions in the IR.
486
487* Support non-C-like languages, to the extent that it doesn't interfere with
488  other goals.
489
490* Minimize target-specific information in the IR.
491
492Why do struct member indices always use ``i32``?
493------------------------------------------------
494
495The specific type i32 is probably just a historical artifact, however it's wide
496enough for all practical purposes, so there's been no need to change it.  It
497doesn't necessarily imply i32 address arithmetic; it's just an identifier which
498identifies a field in a struct. Requiring that all struct indices be the same
499reduces the range of possibilities for cases where two GEPs are effectively the
500same but have distinct operand types.
501
502What's an uglygep?
503------------------
504
505Some LLVM optimizers operate on GEPs by internally lowering them into more
506primitive integer expressions, which allows them to be combined with other
507integer expressions and/or split into multiple separate integer expressions. If
508they've made non-trivial changes, translating back into LLVM IR can involve
509reverse-engineering the structure of the addressing in order to fit it into the
510static type of the original first operand. It isn't always possibly to fully
511reconstruct this structure; sometimes the underlying addressing doesn't
512correspond with the static type at all. In such cases the optimizer instead will
513emit a GEP with the base pointer casted to a simple address-unit pointer, using
514the name "uglygep". This isn't pretty, but it's just as valid, and it's
515sufficient to preserve the pointer aliasing guarantees that GEP provides.
516
517Summary
518=======
519
520In summary, here's some things to always remember about the GetElementPtr
521instruction:
522
523
524#. The GEP instruction never accesses memory, it only provides pointer
525   computations.
526
527#. The first operand to the GEP instruction is always a pointer and it must be
528   indexed.
529
530#. There are no superfluous indices for the GEP instruction.
531
532#. Trailing zero indices are superfluous for pointer aliasing, but not for the
533   types of the pointers.
534
535#. Leading zero indices are not superfluous for pointer aliasing nor the types
536   of the pointers.
537