| blob | 41c45cab12d678caa551bf3f65c904fcc1c0792c |
3 <html>
4 <head>
10 </style>
11 </head>
12 <body>
15 The Often Misunderstood GEP Instruction
16 </div>
18 <ol>
21 <ol>
25 subsequent ones?</a></li>
31 <li><a href="#be">I'm writing a backend for a target which needs custom lowering for GEP. How do I do this?</a>
33 </ol></li>
35 <ol>
39 <li><a href="#types">Can I do GEP with a different pointer type than the type of the underlying object?</a>
41 <li><a href="#ptrdiff">Can I compute the distance between two objects, and add that value to one address to compute the other address?</a>
45 </ol></li>
47 <ol>
51 </ol></li>
53 </ol>
57 </div>
60 <!-- *********************************************************************** -->
62 <!-- *********************************************************************** -->
65 <p>This document seeks to dispel the mystery and confusion surrounding LLVM's
67 Questions about the wily GEP instruction are
68 probably the most frequently occurring questions once a developer gets down to
69 coding with LLVM. Here we lay out the sources of confusion and show that the
70 GEP instruction is really quite simple.
71 </p>
72 </div>
74 <!-- *********************************************************************** -->
76 <!-- *********************************************************************** -->
78 <p>When people are first confronted with the GEP instruction, they tend to
79 relate it to known concepts from other programming paradigms, most notably C
80 array indexing and field selection. GEP closely resembles C array indexing
81 and field selection, however it's is a little different and this leads to
82 the following questions.</p>
83 </div>
85 <!-- *********************************************************************** -->
88 </div>
91 <p>The confusion with the first index usually arises from thinking about
92 the GetElementPtr instruction as if it was a C index operator. They aren't the
96 <pre>
97 AType *Foo;
98 ...
100 </pre>
101 </div>
103 <p>it is natural to think that there is only one index, the selection of the
105 pointer must be indexed explicitly in LLVM. C, on the other hand, indices
106 through it transparently. To arrive at the same address location as the C
107 code, you would provide the GEP instruction with two index operands. The
108 first operand indexes through the pointer; the second operand indexes the
112 <pre>
114 </pre>
115 </div>
118 <blockquote><p><i>Why is it okay to index through the first pointer, but
119 subsequent pointers won't be dereferenced?</i></p></blockquote>
120 <p>The answer is simply because memory does not have to be accessed to
121 perform the computation. The first operand to the GEP instruction must be a
122 value of a pointer type. The value of the pointer is provided directly to
123 the GEP instruction as an operand without any need for accessing memory. It
124 must, therefore be indexed and requires an index operand. Consider this
125 example:</p>
128 <pre>
129 struct munger_struct {
130 int f1;
131 int f2;
132 };
133 void munge(struct munger_struct *P) {
135 }
136 ...
137 munger_struct Array[3];
138 ...
139 munge(Array);
140 </pre>
141 </div>
144 GEP instructions for the three indices through "P" in the assignment
146 of these GEP instructions. The second operand indexes through that pointer.
147 The third operand will be the field offset into the
150 like:</p>
153 <pre>
154 void %munge(%struct.munger_struct* %P) {
155 entry:
157 %tmp = load i32* %tmp
159 %tmp7 = load i32* %tmp6
160 %tmp8 = add i32 %tmp7, %tmp
162 store i32 %tmp8, i32* %tmp9
163 ret void
164 }
165 </pre>
166 </div>
168 <p>In each case the first operand is the pointer through which the GEP
169 instruction starts. The same is true whether the first operand is an
170 argument, allocated memory, or a global variable. </p>
174 <pre>
175 %MyVar = uninitialized global i32
176 ...
177 %idx1 = getelementptr i32* %MyVar, i64 0
178 %idx2 = getelementptr i32* %MyVar, i64 1
179 %idx3 = getelementptr i32* %MyVar, i64 2
180 </pre>
181 </div>
183 <p>These GEP instructions are simply making address computations from the
185 </p>
188 <pre>
192 </pre>
193 </div>
197 memory is accessed to make these computations because the address of
203 it is inadvisable because any load or store with the pointer that results
204 from these GEP instructions would produce undefined results.</p>
205 </div>
207 <!-- *********************************************************************** -->
210 </div>
211 <!-- *********************************************************************** -->
214 <p>This question arises most often when the GEP instruction is applied to a
215 global variable which is always a pointer type. For example, consider
216 this:</p>
219 <pre>
220 %MyStruct = uninitialized global { float*, i32 }
221 ...
223 </pre>
224 </div>
229 of how globals and GEPs work reveals the need. Becoming aware of the following
230 facts will dispel the confusion:</p>
231 <ol>
241 instruction must always be a value of pointer type, the first index
243 pointer.</li>
246 </ol>
247 </div>
249 <!-- *********************************************************************** -->
252 </div>
255 <p>The GetElementPtr instruction dereferences nothing. That is, it doesn't
256 access memory in any way. That's what the Load and Store instructions are for.
257 GEP is only involved in the computation of addresses. For example, consider
258 this:</p>
261 <pre>
262 %MyVar = uninitialized global { [40 x i32 ]* }
263 ...
265 </pre>
266 </div>
269 pointer to a structure containing a pointer to an array of 40 ints. The
270 GEP instruction seems to be accessing the 18th integer of the structure's
271 array of ints. However, this is actually an illegal GEP instruction. It
273 be dereferenced in order to index into the array of 40 ints. Since the
274 GEP instruction never accesses memory, it is illegal.</p>
276 following:</p>
279 <pre>
281 %arr = load [40 x i32]** %idx
283 </pre>
284 </div>
286 <p>In this case, we have to load the pointer in the structure with a load
287 instruction before we can index into the array. If the example was changed
288 to:</p>
291 <pre>
292 %MyVar = uninitialized global { [40 x i32 ] }
293 ...
295 </pre>
296 </div>
298 <p>then everything works fine. In this case, the structure does not contain a
299 pointer and the GEP instruction can index through the global variable,
301 array there.</p>
302 </div>
304 <!-- *********************************************************************** -->
307 </div>
310 <p>If you look at the first indices in these GEP
312 computation diverges with that index. Consider this example:</p>
315 <pre>
316 %MyVar = global { [10 x i32 ] }
319 </pre>
320 </div>
329 pointers don't alias.</p>
331 </div>
333 <!-- *********************************************************************** -->
336 </div>
339 <p>These two GEP instructions will compute the same address because indexing
340 through the 0th element does not change the address. However, it does change
341 the type. Consider this example:</p>
344 <pre>
345 %MyVar = global { [10 x i32 ] }
348 </pre>
349 </div>
354 </div>
356 <!-- *********************************************************************** -->
360 </div>
362 <p>This hasn't always been forcefully disallowed, though it's not recommended.
363 It leads to awkward special cases in the optimizers, and fundamental
364 inconsistency in the IR. In the future, it will probably be outright
365 disallowed.</p>
367 </div>
369 <!-- *********************************************************************** -->
373 </div>
375 <p>None, except that the address space qualifier on the first operand pointer
376 type always matches the address space qualifier on the result type.</p>
378 </div>
380 <!-- *********************************************************************** -->
384 and inttoptr?</b></a>
385 </div>
389 <p>With ptrtoint, you have to pick an integer type. One approach is to pick i64;
390 this is safe on everything LLVM supports (LLVM internally assumes pointers
391 are never wider than 64 bits in many places), and the optimizer will actually
392 narrow the i64 arithmetic down to the actual pointer size on targets which
393 don't support 64-bit arithmetic in most cases. However, there are some cases
394 where it doesn't do this. With GEP you can avoid this problem.
396 <p>Also, GEP carries additional pointer aliasing rules. It's invalid to take a
397 GEP from one object, address into a different separately allocated
398 object, and dereference it. IR producers (front-ends) must follow this rule,
399 and consumers (optimizers, specifically alias analysis) benefit from being
401 information.</p>
405 <p>However, for the underlying integer computation implied, there
406 is no difference.</p>
408 </div>
410 <!-- *********************************************************************** -->
414 lowering for GEP. How do I do this?</b></a>
415 </div>
417 <p>You don't. The integer computation implied by a GEP is target-independent.
418 Typically what you'll need to do is make your backend pattern-match
419 expressions trees involving ADD, MUL, etc., which are what GEP is lowered
420 into. This has the advantage of letting your code work correctly in more
421 cases.</p>
423 <p>GEP does use target-dependent parameters for the size and layout of data
424 types, which targets can customize.</p>
427 need to fix a lot of code in the backend, with GEP lowering being only a
428 small piece of the overall picture.</p>
430 </div>
432 <!-- *********************************************************************** -->
436 </div>
438 <p>GEPs don't natively support VLAs. LLVM's type system is entirely static,
439 and GEP address computations are guided by an LLVM type.</p>
441 <p>VLA indices can be implemented as linearized indices. For example, an
442 expression like X[a][b][c], must be effectively lowered into a form
443 like X[a*m+b*n+c], so that it appears to the GEP as a single-dimensional
444 array reference.</p>
446 <p>This means if you want to write an analysis which understands array
447 indices and you want to support VLAs, your code will have to be
448 prepared to reverse-engineer the linearization. One way to solve this
449 problem is to use the ScalarEvolution library, which always presents
450 VLA and non-VLA indexing in the same manner.</p>
451 </div>
453 <!-- *********************************************************************** -->
455 <!-- *********************************************************************** -->
457 <!-- *********************************************************************** -->
461 </div>
465 <p>First, there's the array type which comes from the (static) type of
466 the first operand to the GEP. Indices greater than the number of elements
467 in the corresponding static array type are valid. There is no problem with
468 out of bounds indices in this sense. Indexing into an array only depends
469 on the size of the array element, not the number of elements.</p>
471 <p>A common example of how this is used is arrays where the size is not known.
472 It's common to use array types with zero length to represent these. The
473 fact that the static type says there are zero elements is irrelevant; it's
474 perfectly valid to compute arbitrary element indices, as the computation
475 only depends on the size of the array element, not the number of
476 elements. Note that zero-sized arrays are not a special case here.</p>
480 arithmetic overflow conditions, rather than high-level array
481 indexing rules.
483 <p>Analysis passes which wish to understand array indexing should not
484 assume that the static array type bounds are respected.</p>
486 <p>The second sense of being out of bounds is computing an address that's
487 beyond the actual underlying allocated object.</p>
490 undefined if the address is outside the actual underlying allocated
491 object and not the address one-past-the-end.</p>
494 on computing out-of-bounds addresses. Obviously, performing a load or
495 a store requires an address of allocated and sufficiently aligned
496 memory. But the GEP itself is only concerned with computing addresses.</p>
498 </div>
500 <!-- *********************************************************************** -->
503 </div>
505 <p>Yes. This is basically a special case of array indices being out
506 of bounds.</p>
508 </div>
510 <!-- *********************************************************************** -->
513 </div>
515 <p>Yes. If both addresses are within the same allocated object, or
516 one-past-the-end, you'll get the comparison result you expect. If either
517 is outside of it, integer arithmetic wrapping may occur, so the
518 comparison may not be meaningful.</p>
520 </div>
522 <!-- *********************************************************************** -->
525 the underlying object?</b></a>
526 </div>
528 <p>Yes. There are no restrictions on bitcasting a pointer value to an arbitrary
529 pointer type. The types in a GEP serve only to define the parameters for the
530 underlying integer computation. They need not correspond with the actual
531 type of the underlying object.</p>
533 <p>Furthermore, loads and stores don't have to use the same types as the type
534 of the underlying object. Types in this context serve only to specify
535 memory size and alignment. Beyond that there are merely a hint to the
536 optimizer indicating how the value will likely be used.</p>
538 </div>
540 <!-- *********************************************************************** -->
543 to null?</b></a>
544 </div>
546 <p>You can compute an address that way, but if you use GEP to do the add,
547 you can't use that pointer to actually access the object, unless the
548 object is managed outside of LLVM.</p>
550 <p>The underlying integer computation is sufficiently defined; null has a
551 defined value -- zero -- and you can add whatever value you want to it.</p>
553 <p>However, it's invalid to access (load from or store to) an LLVM-aware
554 object with such a pointer. This includes GlobalVariables, Allocas, and
555 objects pointed to by noalias pointers.</p>
557 <p>If you really need this functionality, you can do the arithmetic with
558 explicit integer instructions, and use inttoptr to convert the result to
559 an address. Most of GEP's special aliasing rules do not apply to pointers
560 computed from ptrtoint, arithmetic, and inttoptr sequences.</p>
562 </div>
564 <!-- *********************************************************************** -->
567 that value to one address to compute the other address?</b></a>
568 </div>
570 <p>As with arithmetic on null, You can use GEP to compute an address that
571 way, but you can't use that pointer to actually access the object if you
572 do, unless the object is managed outside of LLVM.</p>
574 <p>Also as above, ptrtoint and inttoptr provide an alternative way to do this
575 which do not have this restriction.</p>
577 </div>
579 <!-- *********************************************************************** -->
582 </div>
584 <p>You can't do type-based alias analysis using LLVM's built-in type system,
585 because LLVM has no restrictions on mixing types in addressing, loads or
586 stores.</p>
588 <p>It would be possible to add special annotations to the IR, probably using
589 metadata, to describe a different type system (such as the C type system),
590 and do type-based aliasing on top of that. This is a much bigger
591 undertaking though.</p>
593 </div>
595 <!-- *********************************************************************** -->
599 </div>
602 from evaluating the implied two's complement integer computation. However,
603 since there's no guarantee of where an object will be allocated in the
604 address space, such values have limited meaning.</p>
608 overflows (i.e. wraps around the end of the address space).</p>
610 <p>As such, there are some ramifications of this for inbounds GEPs: scales
611 implied by array/vector/pointer indices are always known to be "nsw" since
612 they are signed values that are scaled by the element size. These values
613 are also allowed to be negative (e.g. "gep i32 *%P, i32 -1") but the
614 pointer itself is logically treated as an unsigned value. This means that
615 GEPs have an asymmetric relation between the pointer base (which is treated
616 as unsigned) and the offset applied to it (which is treated as signed). The
617 result of the additions within the offset calculation cannot have signed
618 overflow, but when applied to the base pointer, there can be signed
619 overflow.
620 </p>
623 </div>
625 <!-- *********************************************************************** -->
629 rules?</b></a>
630 </div>
632 <p>There is currently no checker for the getelementptr rules. Currently,
633 the only way to do this is to manually check each place in your front-end
634 where GetElementPtr operators are created.</p>
636 <p>It's not possible to write a checker which could find all rule
637 violations statically. It would be possible to write a checker which
638 works by instrumenting the code with dynamic checks though. Alternatively,
639 it would be possible to write a static checker which catches a subset of
640 possible problems. However, no such checker exists today.</p>
642 </div>
644 <!-- *********************************************************************** -->
646 <!-- *********************************************************************** -->
648 <!-- *********************************************************************** -->
652 </div>
654 <p>The design of GEP has the following goals, in rough unofficial
655 order of priority:</p>
656 <ul>
657 <li>Support C, C-like languages, and languages which can be
658 conceptually lowered into C (this covers a lot).</li>
659 <li>Support optimizations such as those that are common in
660 C compilers. In particular, GEP is a cornerstone of LLVM's
662 <li>Provide a consistent method for computing addresses so that
663 address computations don't need to be a part of load and
664 store instructions in the IR.</li>
665 <li>Support non-C-like languages, to the extent that it doesn't
666 interfere with other goals.</li>
668 </ul>
669 </div>
671 <!-- *********************************************************************** -->
674 </div>
676 <p>The specific type i32 is probably just a historical artifact, however it's
677 wide enough for all practical purposes, so there's been no need to change it.
678 It doesn't necessarily imply i32 address arithmetic; it's just an identifier
679 which identifies a field in a struct. Requiring that all struct indices be
680 the same reduces the range of possibilities for cases where two GEPs are
681 effectively the same but have distinct operand types.</p>
683 </div>
685 <!-- *********************************************************************** -->
689 </div>
691 <p>Some LLVM optimizers operate on GEPs by internally lowering them into
692 more primitive integer expressions, which allows them to be combined
693 with other integer expressions and/or split into multiple separate
694 integer expressions. If they've made non-trivial changes, translating
695 back into LLVM IR can involve reverse-engineering the structure of
696 the addressing in order to fit it into the static type of the original
697 first operand. It isn't always possibly to fully reconstruct this
698 structure; sometimes the underlying addressing doesn't correspond with
699 the static type at all. In such cases the optimizer instead will emit
700 a GEP with the base pointer casted to a simple address-unit pointer,
701 using the name "uglygep". This isn't pretty, but it's just as
702 valid, and it's sufficient to preserve the pointer aliasing guarantees
703 that GEP provides.</p>
705 </div>
707 <!-- *********************************************************************** -->
709 <!-- *********************************************************************** -->
712 <p>In summary, here's some things to always remember about the GetElementPtr
713 instruction:</p>
714 <ol>
715 <li>The GEP instruction never accesses memory, it only provides pointer
716 computations.</li>
717 <li>The first operand to the GEP instruction is always a pointer and it must
718 be indexed.</li>
720 <li>Trailing zero indices are superfluous for pointer aliasing, but not for
721 the types of the pointers.</li>
722 <li>Leading zero indices are not superfluous for pointer aliasing nor the
723 types of the pointers.</li>
724 </ol>
725 </div>
727 <!-- *********************************************************************** -->
729 <hr>
730 <address>
736 Last modified: $Date$
737 </address>
738 </body>
739 </html>