| blob | 3bf38a366d02c95928effd78bc79ba8567be2082 |
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>
32 <li><a href="#be">I'm writing a backend for a target which needs custom lowering for GEP. How do I do this?</a>
34 </ol></li>
36 <ol>
40 <li><a href="#types">Can I do GEP with a different pointer type than the type of the underlying object?</a>
42 <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>
46 </ol></li>
48 <ol>
52 </ol></li>
54 </ol>
58 </div>
61 <!-- *********************************************************************** -->
63 <!-- *********************************************************************** -->
66 <p>This document seeks to dispel the mystery and confusion surrounding LLVM's
68 Questions about the wily GEP instruction are
69 probably the most frequently occurring questions once a developer gets down to
70 coding with LLVM. Here we lay out the sources of confusion and show that the
71 GEP instruction is really quite simple.
72 </p>
73 </div>
75 <!-- *********************************************************************** -->
77 <!-- *********************************************************************** -->
79 <p>When people are first confronted with the GEP instruction, they tend to
80 relate it to known concepts from other programming paradigms, most notably C
81 array indexing and field selection. GEP closely resembles C array indexing
82 and field selection, however it's is a little different and this leads to
83 the following questions.</p>
84 </div>
86 <!-- *********************************************************************** -->
89 </div>
92 <p>The confusion with the first index usually arises from thinking about
93 the GetElementPtr instruction as if it was a C index operator. They aren't the
97 <pre>
98 AType *Foo;
99 ...
101 </pre>
102 </div>
104 <p>it is natural to think that there is only one index, the selection of the
106 pointer must be indexed explicitly in LLVM. C, on the other hand, indices
107 through it transparently. To arrive at the same address location as the C
108 code, you would provide the GEP instruction with two index operands. The
109 first operand indexes through the pointer; the second operand indexes the
113 <pre>
115 </pre>
116 </div>
119 <blockquote><p><i>Why is it okay to index through the first pointer, but
120 subsequent pointers won't be dereferenced?</i></p></blockquote>
121 <p>The answer is simply because memory does not have to be accessed to
122 perform the computation. The first operand to the GEP instruction must be a
123 value of a pointer type. The value of the pointer is provided directly to
124 the GEP instruction as an operand without any need for accessing memory. It
125 must, therefore be indexed and requires an index operand. Consider this
126 example:</p>
129 <pre>
130 struct munger_struct {
131 int f1;
132 int f2;
133 };
134 void munge(struct munger_struct *P) {
136 }
137 ...
138 munger_struct Array[3];
139 ...
140 munge(Array);
141 </pre>
142 </div>
145 GEP instructions for the three indices through "P" in the assignment
147 of these GEP instructions. The second operand indexes through that pointer.
148 The third operand will be the field offset into the
151 like:</p>
154 <pre>
155 void %munge(%struct.munger_struct* %P) {
156 entry:
158 %tmp = load i32* %tmp
160 %tmp7 = load i32* %tmp6
161 %tmp8 = add i32 %tmp7, %tmp
163 store i32 %tmp8, i32* %tmp9
164 ret void
165 }
166 </pre>
167 </div>
169 <p>In each case the first operand is the pointer through which the GEP
170 instruction starts. The same is true whether the first operand is an
171 argument, allocated memory, or a global variable. </p>
175 <pre>
176 %MyVar = uninitialized global i32
177 ...
178 %idx1 = getelementptr i32* %MyVar, i64 0
179 %idx2 = getelementptr i32* %MyVar, i64 1
180 %idx3 = getelementptr i32* %MyVar, i64 2
181 </pre>
182 </div>
184 <p>These GEP instructions are simply making address computations from the
186 </p>
189 <pre>
193 </pre>
194 </div>
198 memory is accessed to make these computations because the address of
204 it is inadvisable because any load or store with the pointer that results
205 from these GEP instructions would produce undefined results.</p>
206 </div>
208 <!-- *********************************************************************** -->
211 </div>
212 <!-- *********************************************************************** -->
215 <p>This question arises most often when the GEP instruction is applied to a
216 global variable which is always a pointer type. For example, consider
217 this:</p>
220 <pre>
221 %MyStruct = uninitialized global { float*, i32 }
222 ...
224 </pre>
225 </div>
230 of how globals and GEPs work reveals the need. Becoming aware of the following
231 facts will dispel the confusion:</p>
232 <ol>
242 instruction must always be a value of pointer type, the first index
244 pointer.</li>
247 </ol>
248 </div>
250 <!-- *********************************************************************** -->
253 </div>
256 <p>The GetElementPtr instruction dereferences nothing. That is, it doesn't
257 access memory in any way. That's what the Load and Store instructions are for.
258 GEP is only involved in the computation of addresses. For example, consider
259 this:</p>
262 <pre>
263 %MyVar = uninitialized global { [40 x i32 ]* }
264 ...
266 </pre>
267 </div>
270 pointer to a structure containing a pointer to an array of 40 ints. The
271 GEP instruction seems to be accessing the 18th integer of the structure's
272 array of ints. However, this is actually an illegal GEP instruction. It
274 be dereferenced in order to index into the array of 40 ints. Since the
275 GEP instruction never accesses memory, it is illegal.</p>
277 following:</p>
280 <pre>
282 %arr = load [40 x i32]** %idx
284 </pre>
285 </div>
287 <p>In this case, we have to load the pointer in the structure with a load
288 instruction before we can index into the array. If the example was changed
289 to:</p>
292 <pre>
293 %MyVar = uninitialized global { [40 x i32 ] }
294 ...
296 </pre>
297 </div>
299 <p>then everything works fine. In this case, the structure does not contain a
300 pointer and the GEP instruction can index through the global variable,
302 array there.</p>
303 </div>
305 <!-- *********************************************************************** -->
308 </div>
311 <p>If you look at the first indices in these GEP
313 computation diverges with that index. Consider this example:</p>
316 <pre>
317 %MyVar = global { [10 x i32 ] }
320 </pre>
321 </div>
330 pointers don't alias.</p>
332 </div>
334 <!-- *********************************************************************** -->
337 </div>
340 <p>These two GEP instructions will compute the same address because indexing
341 through the 0th element does not change the address. However, it does change
342 the type. Consider this example:</p>
345 <pre>
346 %MyVar = global { [10 x i32 ] }
349 </pre>
350 </div>
355 </div>
357 <!-- *********************************************************************** -->
361 </div>
363 <p>This hasn't always been forcefully disallowed, though it's not recommended.
364 It leads to awkward special cases in the optimizers, and fundamental
365 inconsistency in the IR. In the future, it will probably be outright
366 disallowed.</p>
368 </div>
370 <!-- *********************************************************************** -->
374 </div>
378 </div>
380 <!-- *********************************************************************** -->
384 </div>
386 <p>None, except that the address space qualifier on the first operand pointer
387 type always matches the address space qualifier on the result type.</p>
389 </div>
391 <!-- *********************************************************************** -->
395 and inttoptr?</b></a>
396 </div>
400 <p>With ptrtoint, you have to pick an integer type. One approach is to pick i64;
401 this is safe on everything LLVM supports (LLVM internally assumes pointers
402 are never wider than 64 bits in many places), and the optimizer will actually
403 narrow the i64 arithmetic down to the actual pointer size on targets which
404 don't support 64-bit arithmetic in most cases. However, there are some cases
405 where it doesn't do this. With GEP you can avoid this problem.
407 <p>Also, GEP carries additional pointer aliasing rules. It's invalid to take a
408 GEP from one object, address into a different separately allocated
409 object, and dereference it. IR producers (front-ends) must follow this rule,
410 and consumers (optimizers, specifically alias analysis) benefit from being
412 information.</p>
416 <p>However, for the underlying integer computation implied, there
417 is no difference.</p>
419 </div>
421 <!-- *********************************************************************** -->
425 lowering for GEP. How do I do this?</b></a>
426 </div>
428 <p>You don't. The integer computation implied by a GEP is target-independent.
429 Typically what you'll need to do is make your backend pattern-match
430 expressions trees involving ADD, MUL, etc., which are what GEP is lowered
431 into. This has the advantage of letting your code work correctly in more
432 cases.</p>
434 <p>GEP does use target-dependent parameters for the size and layout of data
435 types, which targets can customize.</p>
438 need to fix a lot of code in the backend, with GEP lowering being only a
439 small piece of the overall picture.</p>
441 </div>
443 <!-- *********************************************************************** -->
447 </div>
449 <p>GEPs don't natively support VLAs. LLVM's type system is entirely static,
450 and GEP address computations are guided by an LLVM type.</p>
452 <p>VLA indices can be implemented as linearized indices. For example, an
453 expression like X[a][b][c], must be effectively lowered into a form
454 like X[a*m+b*n+c], so that it appears to the GEP as a single-dimensional
455 array reference.</p>
457 <p>This means if you want to write an analysis which understands array
458 indices and you want to support VLAs, your code will have to be
459 prepared to reverse-engineer the linearization. One way to solve this
460 problem is to use the ScalarEvolution library, which always presents
461 VLA and non-VLA indexing in the same manner.</p>
462 </div>
464 <!-- *********************************************************************** -->
466 <!-- *********************************************************************** -->
468 <!-- *********************************************************************** -->
472 </div>
476 <p>First, there's the array type which comes from the (static) type of
477 the first operand to the GEP. Indices greater than the number of elements
478 in the corresponding static array type are valid. There is no problem with
479 out of bounds indices in this sense. Indexing into an array only depends
480 on the size of the array element, not the number of elements.</p>
482 <p>A common example of how this is used is arrays where the size is not known.
483 It's common to use array types with zero length to represent these. The
484 fact that the static type says there are zero elements is irrelevant; it's
485 perfectly valid to compute arbitrary element indices, as the computation
486 only depends on the size of the array element, not the number of
487 elements. Note that zero-sized arrays are not a special case here.</p>
491 arithmetic overflow conditions, rather than high-level array
492 indexing rules.
494 <p>Analysis passes which wish to understand array indexing should not
495 assume that the static array type bounds are respected.</p>
497 <p>The second sense of being out of bounds is computing an address that's
498 beyond the actual underlying allocated object.</p>
501 undefined if the address is outside the actual underlying allocated
502 object and not the address one-past-the-end.</p>
505 on computing out-of-bounds addresses. Obviously, performing a load or
506 a store requires an address of allocated and sufficiently aligned
507 memory. But the GEP itself is only concerned with computing addresses.</p>
509 </div>
511 <!-- *********************************************************************** -->
514 </div>
516 <p>Yes. This is basically a special case of array indices being out
517 of bounds.</p>
519 </div>
521 <!-- *********************************************************************** -->
524 </div>
526 <p>Yes. If both addresses are within the same allocated object, or
527 one-past-the-end, you'll get the comparison result you expect. If either
528 is outside of it, integer arithmetic wrapping may occur, so the
529 comparison may not be meaningful.</p>
531 </div>
533 <!-- *********************************************************************** -->
536 the underlying object?</b></a>
537 </div>
539 <p>Yes. There are no restrictions on bitcasting a pointer value to an arbitrary
540 pointer type. The types in a GEP serve only to define the parameters for the
541 underlying integer computation. They need not correspond with the actual
542 type of the underlying object.</p>
544 <p>Furthermore, loads and stores don't have to use the same types as the type
545 of the underlying object. Types in this context serve only to specify
546 memory size and alignment. Beyond that there are merely a hint to the
547 optimizer indicating how the value will likely be used.</p>
549 </div>
551 <!-- *********************************************************************** -->
554 to null?</b></a>
555 </div>
557 <p>You can compute an address that way, but if you use GEP to do the add,
558 you can't use that pointer to actually access the object, unless the
559 object is managed outside of LLVM.</p>
561 <p>The underlying integer computation is sufficiently defined; null has a
562 defined value -- zero -- and you can add whatever value you want to it.</p>
564 <p>However, it's invalid to access (load from or store to) an LLVM-aware
565 object with such a pointer. This includes GlobalVariables, Allocas, and
566 objects pointed to by noalias pointers.</p>
568 <p>If you really need this functionality, you can do the arithmetic with
569 explicit integer instructions, and use inttoptr to convert the result to
570 an address. Most of GEP's special aliasing rules do not apply to pointers
571 computed from ptrtoint, arithmetic, and inttoptr sequences.</p>
573 </div>
575 <!-- *********************************************************************** -->
578 that value to one address to compute the other address?</b></a>
579 </div>
581 <p>As with arithmetic on null, You can use GEP to compute an address that
582 way, but you can't use that pointer to actually access the object if you
583 do, unless the object is managed outside of LLVM.</p>
585 <p>Also as above, ptrtoint and inttoptr provide an alternative way to do this
586 which do not have this restriction.</p>
588 </div>
590 <!-- *********************************************************************** -->
593 </div>
595 <p>You can't do type-based alias analysis using LLVM's built-in type system,
596 because LLVM has no restrictions on mixing types in addressing, loads or
597 stores.</p>
599 <p>It would be possible to add special annotations to the IR, probably using
600 metadata, to describe a different type system (such as the C type system),
601 and do type-based aliasing on top of that. This is a much bigger
602 undertaking though.</p>
604 </div>
606 <!-- *********************************************************************** -->
610 </div>
613 undefined.</p>
615 <p>Otherwise, the result value is the result from evaluating the implied
616 two's complement integer computation. However, since there's no
617 guarantee of where an object will be allocated in the address space,
618 such values have limited meaning.</p>
620 </div>
622 <!-- *********************************************************************** -->
626 rules?</b></a>
627 </div>
629 <p>There is currently no checker for the getelementptr rules. Currently,
630 the only way to do this is to manually check each place in your front-end
631 where GetElementPtr operators are created.</p>
633 <p>It's not possible to write a checker which could find all rule
634 violations statically. It would be possible to write a checker which
635 works by instrumenting the code with dynamic checks though. Alternatively,
636 it would be possible to write a static checker which catches a subset of
637 possible problems. However, no such checker exists today.</p>
639 </div>
641 <!-- *********************************************************************** -->
643 <!-- *********************************************************************** -->
645 <!-- *********************************************************************** -->
649 </div>
651 <p>The design of GEP has the following goals, in rough unofficial
652 order of priority:</p>
653 <ul>
654 <li>Support C, C-like languages, and languages which can be
655 conceptually lowered into C (this covers a lot).</li>
656 <li>Support optimizations such as those that are common in
657 C compilers. In particular, GEP is a cornerstone of LLVM's
659 <li>Provide a consistent method for computing addresses so that
660 address computations don't need to be a part of load and
661 store instructions in the IR.</li>
662 <li>Support non-C-like languages, to the extent that it doesn't
663 interfere with other goals.</li>
665 </ul>
666 </div>
668 <!-- *********************************************************************** -->
671 </div>
673 <p>The specific type i32 is probably just a historical artifact, however it's
674 wide enough for all practical purposes, so there's been no need to change it.
675 It doesn't necessarily imply i32 address arithmetic; it's just an identifier
676 which identifies a field in a struct. Requiring that all struct indices be
677 the same reduces the range of possibilities for cases where two GEPs are
678 effectively the same but have distinct operand types.</p>
680 </div>
682 <!-- *********************************************************************** -->
686 </div>
688 <p>Some LLVM optimizers operate on GEPs by internally lowering them into
689 more primitive integer expressions, which allows them to be combined
690 with other integer expressions and/or split into multiple separate
691 integer expressions. If they've made non-trivial changes, translating
692 back into LLVM IR can involve reverse-engineering the structure of
693 the addressing in order to fit it into the static type of the original
694 first operand. It isn't always possibly to fully reconstruct this
695 structure; sometimes the underlying addressing doesn't correspond with
696 the static type at all. In such cases the optimizer instead will emit
697 a GEP with the base pointer casted to a simple address-unit pointer,
698 using the name "uglygep". This isn't pretty, but it's just as
699 valid, and it's sufficient to preserve the pointer aliasing guarantees
700 that GEP provides.</p>
702 </div>
704 <!-- *********************************************************************** -->
706 <!-- *********************************************************************** -->
709 <p>In summary, here's some things to always remember about the GetElementPtr
710 instruction:</p>
711 <ol>
712 <li>The GEP instruction never accesses memory, it only provides pointer
713 computations.</li>
714 <li>The first operand to the GEP instruction is always a pointer and it must
715 be indexed.</li>
717 <li>Trailing zero indices are superfluous for pointer aliasing, but not for
718 the types of the pointers.</li>
719 <li>Leading zero indices are not superfluous for pointer aliasing nor the
720 types of the pointers.</li>
721 </ol>
722 </div>
724 <!-- *********************************************************************** -->
726 <hr>
727 <address>
733 Last modified: $Date$
734 </address>
735 </body>
736 </html>