1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN" 2 "http://www.w3.org/TR/html4/strict.dtd"> 3 <html> 4 <head> 5 <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> 6 <title>The Often Misunderstood GEP Instruction</title> 7 <link rel="stylesheet" href="llvm.css" type="text/css"> 8 <style type="text/css"> 9 TABLE { text-align: left; border: 1px solid black; border-collapse: collapse; margin: 0 0 0 0; } 10 </style> 11 </head> 12 <body> 13 14 <h1> 15 The Often Misunderstood GEP Instruction 16 </h1> 17 18 <ol> 19 <li><a href="#intro">Introduction</a></li> 20 <li><a href="#addresses">Address Computation</a> 21 <ol> 22 <li><a href="#extra_index">Why is the extra 0 index required?</a></li> 23 <li><a href="#deref">What is dereferenced by GEP?</a></li> 24 <li><a href="#firstptr">Why can you index through the first pointer but not 25 subsequent ones?</a></li> 26 <li><a href="#lead0">Why don't GEP x,0,0,1 and GEP x,1 alias? </a></li> 27 <li><a href="#trail0">Why do GEP x,1,0,0 and GEP x,1 alias? </a></li> 28 <li><a href="#vectors">Can GEP index into vector elements?</a> 29 <li><a href="#addrspace">What effect do address spaces have on GEPs?</a> 30 <li><a href="#int">How is GEP different from ptrtoint, arithmetic, and inttoptr?</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> 32 <li><a href="#vla">How does VLA addressing work with GEPs?</a> 33 </ol></li> 34 <li><a href="#rules">Rules</a> 35 <ol> 36 <li><a href="#bounds">What happens if an array index is out of bounds?</a> 37 <li><a href="#negative">Can array indices be negative?</a> 38 <li><a href="#compare">Can I compare two values computed with GEPs?</a> 39 <li><a href="#types">Can I do GEP with a different pointer type than the type of the underlying object?</a> 40 <li><a href="#null">Can I cast an object's address to integer and add it to null?</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> 42 <li><a href="#tbaa">Can I do type-based alias analysis on LLVM IR?</a> 43 <li><a href="#overflow">What happens if a GEP computation overflows?</a> 44 <li><a href="#check">How can I tell if my front-end is following the rules?</a> 45 </ol></li> 46 <li><a href="#rationale">Rationale</a> 47 <ol> 48 <li><a href="#goals">Why is GEP designed this way?</a></li> 49 <li><a href="#i32">Why do struct member indices always use i32?</a></li> 50 <li><a href="#uglygep">What's an uglygep?</a> 51 </ol></li> 52 <li><a href="#summary">Summary</a></li> 53 </ol> 54 55 <div class="doc_author"> 56 <p>Written by: <a href="mailto:rspencer (a] reidspencer.com">Reid Spencer</a>.</p> 57 </div> 58 59 60 <!-- *********************************************************************** --> 61 <h2><a name="intro">Introduction</a></h2> 62 <!-- *********************************************************************** --> 63 64 <div> 65 <p>This document seeks to dispel the mystery and confusion surrounding LLVM's 66 <a href="LangRef.html#i_getelementptr">GetElementPtr</a> (GEP) instruction. 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> 73 74 <!-- *********************************************************************** --> 75 <h2><a name="addresses">Address Computation</a></h2> 76 <!-- *********************************************************************** --> 77 <div> 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 84 <!-- *********************************************************************** --> 85 <h3> 86 <a name="firstptr">What is the first index of the GEP instruction?</a> 87 </h3> 88 <div> 89 <p>Quick answer: The index stepping through the first operand.</p> 90 <p>The confusion with the first index usually arises from thinking about 91 the GetElementPtr instruction as if it was a C index operator. They aren't the 92 same. For example, when we write, in "C":</p> 93 94 <div class="doc_code"> 95 <pre> 96 AType *Foo; 97 ... 98 X = &Foo->F; 99 </pre> 100 </div> 101 102 <p>it is natural to think that there is only one index, the selection of the 103 field <tt>F</tt>. However, in this example, <tt>Foo</tt> is a pointer. That 104 pointer must be indexed explicitly in LLVM. C, on the other hand, indices 105 through it transparently. To arrive at the same address location as the C 106 code, you would provide the GEP instruction with two index operands. The 107 first operand indexes through the pointer; the second operand indexes the 108 field <tt>F</tt> of the structure, just as if you wrote:</p> 109 110 <div class="doc_code"> 111 <pre> 112 X = &Foo[0].F; 113 </pre> 114 </div> 115 116 <p>Sometimes this question gets rephrased as:</p> 117 <blockquote><p><i>Why is it okay to index through the first pointer, but 118 subsequent pointers won't be dereferenced?</i></p></blockquote> 119 <p>The answer is simply because memory does not have to be accessed to 120 perform the computation. The first operand to the GEP instruction must be a 121 value of a pointer type. The value of the pointer is provided directly to 122 the GEP instruction as an operand without any need for accessing memory. It 123 must, therefore be indexed and requires an index operand. Consider this 124 example:</p> 125 126 <div class="doc_code"> 127 <pre> 128 struct munger_struct { 129 int f1; 130 int f2; 131 }; 132 void munge(struct munger_struct *P) { 133 P[0].f1 = P[1].f1 + P[2].f2; 134 } 135 ... 136 munger_struct Array[3]; 137 ... 138 munge(Array); 139 </pre> 140 </div> 141 142 <p>In this "C" example, the front end compiler (llvm-gcc) will generate three 143 GEP instructions for the three indices through "P" in the assignment 144 statement. The function argument <tt>P</tt> will be the first operand of each 145 of these GEP instructions. The second operand indexes through that pointer. 146 The third operand will be the field offset into the 147 <tt>struct munger_struct</tt> type, for either the <tt>f1</tt> or 148 <tt>f2</tt> field. So, in LLVM assembly the <tt>munge</tt> function looks 149 like:</p> 150 151 <div class="doc_code"> 152 <pre> 153 void %munge(%struct.munger_struct* %P) { 154 entry: 155 %tmp = getelementptr %struct.munger_struct* %P, i32 1, i32 0 156 %tmp = load i32* %tmp 157 %tmp6 = getelementptr %struct.munger_struct* %P, i32 2, i32 1 158 %tmp7 = load i32* %tmp6 159 %tmp8 = add i32 %tmp7, %tmp 160 %tmp9 = getelementptr %struct.munger_struct* %P, i32 0, i32 0 161 store i32 %tmp8, i32* %tmp9 162 ret void 163 } 164 </pre> 165 </div> 166 167 <p>In each case the first operand is the pointer through which the GEP 168 instruction starts. The same is true whether the first operand is an 169 argument, allocated memory, or a global variable. </p> 170 <p>To make this clear, let's consider a more obtuse example:</p> 171 172 <div class="doc_code"> 173 <pre> 174 %MyVar = uninitialized global i32 175 ... 176 %idx1 = getelementptr i32* %MyVar, i64 0 177 %idx2 = getelementptr i32* %MyVar, i64 1 178 %idx3 = getelementptr i32* %MyVar, i64 2 179 </pre> 180 </div> 181 182 <p>These GEP instructions are simply making address computations from the 183 base address of <tt>MyVar</tt>. They compute, as follows (using C syntax): 184 </p> 185 186 <div class="doc_code"> 187 <pre> 188 idx1 = (char*) &MyVar + 0 189 idx2 = (char*) &MyVar + 4 190 idx3 = (char*) &MyVar + 8 191 </pre> 192 </div> 193 194 <p>Since the type <tt>i32</tt> is known to be four bytes long, the indices 195 0, 1 and 2 translate into memory offsets of 0, 4, and 8, respectively. No 196 memory is accessed to make these computations because the address of 197 <tt>%MyVar</tt> is passed directly to the GEP instructions.</p> 198 <p>The obtuse part of this example is in the cases of <tt>%idx2</tt> and 199 <tt>%idx3</tt>. They result in the computation of addresses that point to 200 memory past the end of the <tt>%MyVar</tt> global, which is only one 201 <tt>i32</tt> long, not three <tt>i32</tt>s long. While this is legal in LLVM, 202 it is inadvisable because any load or store with the pointer that results 203 from these GEP instructions would produce undefined results.</p> 204 </div> 205 206 <!-- *********************************************************************** --> 207 <h3> 208 <a name="extra_index">Why is the extra 0 index required?</a> 209 </h3> 210 <!-- *********************************************************************** --> 211 <div> 212 <p>Quick answer: there are no superfluous indices.</p> 213 <p>This question arises most often when the GEP instruction is applied to a 214 global variable which is always a pointer type. For example, consider 215 this:</p> 216 217 <div class="doc_code"> 218 <pre> 219 %MyStruct = uninitialized global { float*, i32 } 220 ... 221 %idx = getelementptr { float*, i32 }* %MyStruct, i64 0, i32 1 222 </pre> 223 </div> 224 225 <p>The GEP above yields an <tt>i32*</tt> by indexing the <tt>i32</tt> typed 226 field of the structure <tt>%MyStruct</tt>. When people first look at it, they 227 wonder why the <tt>i64 0</tt> index is needed. However, a closer inspection 228 of how globals and GEPs work reveals the need. Becoming aware of the following 229 facts will dispel the confusion:</p> 230 <ol> 231 <li>The type of <tt>%MyStruct</tt> is <i>not</i> <tt>{ float*, i32 }</tt> 232 but rather <tt>{ float*, i32 }*</tt>. That is, <tt>%MyStruct</tt> is a 233 pointer to a structure containing a pointer to a <tt>float</tt> and an 234 <tt>i32</tt>.</li> 235 <li>Point #1 is evidenced by noticing the type of the first operand of 236 the GEP instruction (<tt>%MyStruct</tt>) which is 237 <tt>{ float*, i32 }*</tt>.</li> 238 <li>The first index, <tt>i64 0</tt> is required to step over the global 239 variable <tt>%MyStruct</tt>. Since the first argument to the GEP 240 instruction must always be a value of pointer type, the first index 241 steps through that pointer. A value of 0 means 0 elements offset from that 242 pointer.</li> 243 <li>The second index, <tt>i32 1</tt> selects the second field of the 244 structure (the <tt>i32</tt>). </li> 245 </ol> 246 </div> 247 248 <!-- *********************************************************************** --> 249 <h3> 250 <a name="deref">What is dereferenced by GEP?</a> 251 </h3> 252 <div> 253 <p>Quick answer: nothing.</p> 254 <p>The GetElementPtr instruction dereferences nothing. That is, it doesn't 255 access memory in any way. That's what the Load and Store instructions are for. 256 GEP is only involved in the computation of addresses. For example, consider 257 this:</p> 258 259 <div class="doc_code"> 260 <pre> 261 %MyVar = uninitialized global { [40 x i32 ]* } 262 ... 263 %idx = getelementptr { [40 x i32]* }* %MyVar, i64 0, i32 0, i64 0, i64 17 264 </pre> 265 </div> 266 267 <p>In this example, we have a global variable, <tt>%MyVar</tt> that is a 268 pointer to a structure containing a pointer to an array of 40 ints. The 269 GEP instruction seems to be accessing the 18th integer of the structure's 270 array of ints. However, this is actually an illegal GEP instruction. It 271 won't compile. The reason is that the pointer in the structure <i>must</i> 272 be dereferenced in order to index into the array of 40 ints. Since the 273 GEP instruction never accesses memory, it is illegal.</p> 274 <p>In order to access the 18th integer in the array, you would need to do the 275 following:</p> 276 277 <div class="doc_code"> 278 <pre> 279 %idx = getelementptr { [40 x i32]* }* %, i64 0, i32 0 280 %arr = load [40 x i32]** %idx 281 %idx = getelementptr [40 x i32]* %arr, i64 0, i64 17 282 </pre> 283 </div> 284 285 <p>In this case, we have to load the pointer in the structure with a load 286 instruction before we can index into the array. If the example was changed 287 to:</p> 288 289 <div class="doc_code"> 290 <pre> 291 %MyVar = uninitialized global { [40 x i32 ] } 292 ... 293 %idx = getelementptr { [40 x i32] }*, i64 0, i32 0, i64 17 294 </pre> 295 </div> 296 297 <p>then everything works fine. In this case, the structure does not contain a 298 pointer and the GEP instruction can index through the global variable, 299 into the first field of the structure and access the 18th <tt>i32</tt> in the 300 array there.</p> 301 </div> 302 303 <!-- *********************************************************************** --> 304 <h3> 305 <a name="lead0">Why don't GEP x,0,0,1 and GEP x,1 alias?</a> 306 </h3> 307 <div> 308 <p>Quick Answer: They compute different address locations.</p> 309 <p>If you look at the first indices in these GEP 310 instructions you find that they are different (0 and 1), therefore the address 311 computation diverges with that index. Consider this example:</p> 312 313 <div class="doc_code"> 314 <pre> 315 %MyVar = global { [10 x i32 ] } 316 %idx1 = getelementptr { [10 x i32 ] }* %MyVar, i64 0, i32 0, i64 1 317 %idx2 = getelementptr { [10 x i32 ] }* %MyVar, i64 1 318 </pre> 319 </div> 320 321 <p>In this example, <tt>idx1</tt> computes the address of the second integer 322 in the array that is in the structure in <tt>%MyVar</tt>, that is 323 <tt>MyVar+4</tt>. The type of <tt>idx1</tt> is <tt>i32*</tt>. However, 324 <tt>idx2</tt> computes the address of <i>the next</i> structure after 325 <tt>%MyVar</tt>. The type of <tt>idx2</tt> is <tt>{ [10 x i32] }*</tt> and its 326 value is equivalent to <tt>MyVar + 40</tt> because it indexes past the ten 327 4-byte integers in <tt>MyVar</tt>. Obviously, in such a situation, the 328 pointers don't alias.</p> 329 330 </div> 331 332 <!-- *********************************************************************** --> 333 <h3> 334 <a name="trail0">Why do GEP x,1,0,0 and GEP x,1 alias?</a> 335 </h3> 336 <div> 337 <p>Quick Answer: They compute the same address location.</p> 338 <p>These two GEP instructions will compute the same address because indexing 339 through the 0th element does not change the address. However, it does change 340 the type. Consider this example:</p> 341 342 <div class="doc_code"> 343 <pre> 344 %MyVar = global { [10 x i32 ] } 345 %idx1 = getelementptr { [10 x i32 ] }* %MyVar, i64 1, i32 0, i64 0 346 %idx2 = getelementptr { [10 x i32 ] }* %MyVar, i64 1 347 </pre> 348 </div> 349 350 <p>In this example, the value of <tt>%idx1</tt> is <tt>%MyVar+40</tt> and 351 its type is <tt>i32*</tt>. The value of <tt>%idx2</tt> is also 352 <tt>MyVar+40</tt> but its type is <tt>{ [10 x i32] }*</tt>.</p> 353 </div> 354 355 <!-- *********************************************************************** --> 356 357 <h3> 358 <a name="vectors">Can GEP index into vector elements?</a> 359 </h3> 360 <div> 361 <p>This hasn't always been forcefully disallowed, though it's not recommended. 362 It leads to awkward special cases in the optimizers, and fundamental 363 inconsistency in the IR. In the future, it will probably be outright 364 disallowed.</p> 365 366 </div> 367 368 <!-- *********************************************************************** --> 369 370 <h3> 371 <a name="addrspace">What effect do address spaces have on GEPs?</a> 372 </h3> 373 <div> 374 <p>None, except that the address space qualifier on the first operand pointer 375 type always matches the address space qualifier on the result type.</p> 376 377 </div> 378 379 <!-- *********************************************************************** --> 380 381 <h3> 382 <a name="int"> 383 How is GEP different from ptrtoint, arithmetic, and inttoptr? 384 </a> 385 </h3> 386 <div> 387 <p>It's very similar; there are only subtle differences.</p> 388 389 <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. 395 396 <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 400 able to rely on it. See the <a href="#rules">Rules</a> section for more 401 information.</p> 402 403 <p>And, GEP is more concise in common cases.</p> 404 405 <p>However, for the underlying integer computation implied, there 406 is no difference.</p> 407 408 </div> 409 410 <!-- *********************************************************************** --> 411 412 <h3> 413 <a name="be"> 414 I'm writing a backend for a target which needs custom lowering for GEP. 415 How do I do this? 416 </a> 417 </h3> 418 <div> 419 <p>You don't. The integer computation implied by a GEP is target-independent. 420 Typically what you'll need to do is make your backend pattern-match 421 expressions trees involving ADD, MUL, etc., which are what GEP is lowered 422 into. This has the advantage of letting your code work correctly in more 423 cases.</p> 424 425 <p>GEP does use target-dependent parameters for the size and layout of data 426 types, which targets can customize.</p> 427 428 <p>If you require support for addressing units which are not 8 bits, you'll 429 need to fix a lot of code in the backend, with GEP lowering being only a 430 small piece of the overall picture.</p> 431 432 </div> 433 434 <!-- *********************************************************************** --> 435 436 <h3> 437 <a name="vla">How does VLA addressing work with GEPs?</a> 438 </h3> 439 <div> 440 <p>GEPs don't natively support VLAs. LLVM's type system is entirely static, 441 and GEP address computations are guided by an LLVM type.</p> 442 443 <p>VLA indices can be implemented as linearized indices. For example, an 444 expression like X[a][b][c], must be effectively lowered into a form 445 like X[a*m+b*n+c], so that it appears to the GEP as a single-dimensional 446 array reference.</p> 447 448 <p>This means if you want to write an analysis which understands array 449 indices and you want to support VLAs, your code will have to be 450 prepared to reverse-engineer the linearization. One way to solve this 451 problem is to use the ScalarEvolution library, which always presents 452 VLA and non-VLA indexing in the same manner.</p> 453 </div> 454 455 </div> 456 457 <!-- *********************************************************************** --> 458 <h2><a name="rules">Rules</a></h2> 459 <!-- *********************************************************************** --> 460 <div> 461 <!-- *********************************************************************** --> 462 463 <h3> 464 <a name="bounds">What happens if an array index is out of bounds?</a> 465 </h3> 466 <div> 467 <p>There are two senses in which an array index can be out of bounds.</p> 468 469 <p>First, there's the array type which comes from the (static) type of 470 the first operand to the GEP. Indices greater than the number of elements 471 in the corresponding static array type are valid. There is no problem with 472 out of bounds indices in this sense. Indexing into an array only depends 473 on the size of the array element, not the number of elements.</p> 474 475 <p>A common example of how this is used is arrays where the size is not known. 476 It's common to use array types with zero length to represent these. The 477 fact that the static type says there are zero elements is irrelevant; it's 478 perfectly valid to compute arbitrary element indices, as the computation 479 only depends on the size of the array element, not the number of 480 elements. Note that zero-sized arrays are not a special case here.</p> 481 482 <p>This sense is unconnected with <tt>inbounds</tt> keyword. The 483 <tt>inbounds</tt> keyword is designed to describe low-level pointer 484 arithmetic overflow conditions, rather than high-level array 485 indexing rules. 486 487 <p>Analysis passes which wish to understand array indexing should not 488 assume that the static array type bounds are respected.</p> 489 490 <p>The second sense of being out of bounds is computing an address that's 491 beyond the actual underlying allocated object.</p> 492 493 <p>With the <tt>inbounds</tt> keyword, the result value of the GEP is 494 undefined if the address is outside the actual underlying allocated 495 object and not the address one-past-the-end.</p> 496 497 <p>Without the <tt>inbounds</tt> keyword, there are no restrictions 498 on computing out-of-bounds addresses. Obviously, performing a load or 499 a store requires an address of allocated and sufficiently aligned 500 memory. But the GEP itself is only concerned with computing addresses.</p> 501 502 </div> 503 504 <!-- *********************************************************************** --> 505 <h3> 506 <a name="negative">Can array indices be negative?</a> 507 </h3> 508 <div> 509 <p>Yes. This is basically a special case of array indices being out 510 of bounds.</p> 511 512 </div> 513 514 <!-- *********************************************************************** --> 515 <h3> 516 <a name="compare">Can I compare two values computed with GEPs?</a> 517 </h3> 518 <div> 519 <p>Yes. If both addresses are within the same allocated object, or 520 one-past-the-end, you'll get the comparison result you expect. If either 521 is outside of it, integer arithmetic wrapping may occur, so the 522 comparison may not be meaningful.</p> 523 524 </div> 525 526 <!-- *********************************************************************** --> 527 <h3> 528 <a name="types"> 529 Can I do GEP with a different pointer type than the type of 530 the underlying object? 531 </a> 532 </h3> 533 <div> 534 <p>Yes. There are no restrictions on bitcasting a pointer value to an arbitrary 535 pointer type. The types in a GEP serve only to define the parameters for the 536 underlying integer computation. They need not correspond with the actual 537 type of the underlying object.</p> 538 539 <p>Furthermore, loads and stores don't have to use the same types as the type 540 of the underlying object. Types in this context serve only to specify 541 memory size and alignment. Beyond that there are merely a hint to the 542 optimizer indicating how the value will likely be used.</p> 543 544 </div> 545 546 <!-- *********************************************************************** --> 547 <h3> 548 <a name="null"> 549 Can I cast an object's address to integer and add it to null? 550 </a> 551 </h3> 552 <div> 553 <p>You can compute an address that way, but if you use GEP to do the add, 554 you can't use that pointer to actually access the object, unless the 555 object is managed outside of LLVM.</p> 556 557 <p>The underlying integer computation is sufficiently defined; null has a 558 defined value -- zero -- and you can add whatever value you want to it.</p> 559 560 <p>However, it's invalid to access (load from or store to) an LLVM-aware 561 object with such a pointer. This includes GlobalVariables, Allocas, and 562 objects pointed to by noalias pointers.</p> 563 564 <p>If you really need this functionality, you can do the arithmetic with 565 explicit integer instructions, and use inttoptr to convert the result to 566 an address. Most of GEP's special aliasing rules do not apply to pointers 567 computed from ptrtoint, arithmetic, and inttoptr sequences.</p> 568 569 </div> 570 571 <!-- *********************************************************************** --> 572 <h3> 573 <a name="ptrdiff"> 574 Can I compute the distance between two objects, and add 575 that value to one address to compute the other address? 576 </a> 577 </h3> 578 <div> 579 <p>As with arithmetic on null, You can use GEP to compute an address that 580 way, but you can't use that pointer to actually access the object if you 581 do, unless the object is managed outside of LLVM.</p> 582 583 <p>Also as above, ptrtoint and inttoptr provide an alternative way to do this 584 which do not have this restriction.</p> 585 586 </div> 587 588 <!-- *********************************************************************** --> 589 <h3> 590 <a name="tbaa">Can I do type-based alias analysis on LLVM IR?</a> 591 </h3> 592 <div> 593 <p>You can't do type-based alias analysis using LLVM's built-in type system, 594 because LLVM has no restrictions on mixing types in addressing, loads or 595 stores.</p> 596 597 <p>It would be possible to add special annotations to the IR, probably using 598 metadata, to describe a different type system (such as the C type system), 599 and do type-based aliasing on top of that. This is a much bigger 600 undertaking though.</p> 601 602 </div> 603 604 <!-- *********************************************************************** --> 605 606 <h3> 607 <a name="overflow">What happens if a GEP computation overflows?</a> 608 </h3> 609 <div> 610 <p>If the GEP lacks the <tt>inbounds</tt> keyword, the value is the result 611 from evaluating the implied two's complement integer computation. However, 612 since there's no guarantee of where an object will be allocated in the 613 address space, such values have limited meaning.</p> 614 615 <p>If the GEP has the <tt>inbounds</tt> keyword, the result value is 616 undefined (a "<a href="LangRef.html#trapvalues">trap value</a>") if the GEP 617 overflows (i.e. wraps around the end of the address space).</p> 618 619 <p>As such, there are some ramifications of this for inbounds GEPs: scales 620 implied by array/vector/pointer indices are always known to be "nsw" since 621 they are signed values that are scaled by the element size. These values 622 are also allowed to be negative (e.g. "gep i32 *%P, i32 -1") but the 623 pointer itself is logically treated as an unsigned value. This means that 624 GEPs have an asymmetric relation between the pointer base (which is treated 625 as unsigned) and the offset applied to it (which is treated as signed). The 626 result of the additions within the offset calculation cannot have signed 627 overflow, but when applied to the base pointer, there can be signed 628 overflow. 629 </p> 630 631 632 </div> 633 634 <!-- *********************************************************************** --> 635 636 <h3> 637 <a name="check"> 638 How can I tell if my front-end is following the rules? 639 </a> 640 </h3> 641 <div> 642 <p>There is currently no checker for the getelementptr rules. Currently, 643 the only way to do this is to manually check each place in your front-end 644 where GetElementPtr operators are created.</p> 645 646 <p>It's not possible to write a checker which could find all rule 647 violations statically. It would be possible to write a checker which 648 works by instrumenting the code with dynamic checks though. Alternatively, 649 it would be possible to write a static checker which catches a subset of 650 possible problems. However, no such checker exists today.</p> 651 652 </div> 653 654 </div> 655 656 <!-- *********************************************************************** --> 657 <h2><a name="rationale">Rationale</a></h2> 658 <!-- *********************************************************************** --> 659 <div> 660 <!-- *********************************************************************** --> 661 662 <h3> 663 <a name="goals">Why is GEP designed this way?</a> 664 </h3> 665 <div> 666 <p>The design of GEP has the following goals, in rough unofficial 667 order of priority:</p> 668 <ul> 669 <li>Support C, C-like languages, and languages which can be 670 conceptually lowered into C (this covers a lot).</li> 671 <li>Support optimizations such as those that are common in 672 C compilers. In particular, GEP is a cornerstone of LLVM's 673 <a href="LangRef.html#pointeraliasing">pointer aliasing model</a>.</li> 674 <li>Provide a consistent method for computing addresses so that 675 address computations don't need to be a part of load and 676 store instructions in the IR.</li> 677 <li>Support non-C-like languages, to the extent that it doesn't 678 interfere with other goals.</li> 679 <li>Minimize target-specific information in the IR.</li> 680 </ul> 681 </div> 682 683 <!-- *********************************************************************** --> 684 <h3> 685 <a name="i32">Why do struct member indices always use i32?</a> 686 </h3> 687 <div> 688 <p>The specific type i32 is probably just a historical artifact, however it's 689 wide enough for all practical purposes, so there's been no need to change it. 690 It doesn't necessarily imply i32 address arithmetic; it's just an identifier 691 which identifies a field in a struct. Requiring that all struct indices be 692 the same reduces the range of possibilities for cases where two GEPs are 693 effectively the same but have distinct operand types.</p> 694 695 </div> 696 697 <!-- *********************************************************************** --> 698 699 <h3> 700 <a name="uglygep">What's an uglygep?</a> 701 </h3> 702 <div> 703 <p>Some LLVM optimizers operate on GEPs by internally lowering them into 704 more primitive integer expressions, which allows them to be combined 705 with other integer expressions and/or split into multiple separate 706 integer expressions. If they've made non-trivial changes, translating 707 back into LLVM IR can involve reverse-engineering the structure of 708 the addressing in order to fit it into the static type of the original 709 first operand. It isn't always possibly to fully reconstruct this 710 structure; sometimes the underlying addressing doesn't correspond with 711 the static type at all. In such cases the optimizer instead will emit 712 a GEP with the base pointer casted to a simple address-unit pointer, 713 using the name "uglygep". This isn't pretty, but it's just as 714 valid, and it's sufficient to preserve the pointer aliasing guarantees 715 that GEP provides.</p> 716 717 </div> 718 719 </div> 720 721 <!-- *********************************************************************** --> 722 <h2><a name="summary">Summary</a></h2> 723 <!-- *********************************************************************** --> 724 725 <div> 726 <p>In summary, here's some things to always remember about the GetElementPtr 727 instruction:</p> 728 <ol> 729 <li>The GEP instruction never accesses memory, it only provides pointer 730 computations.</li> 731 <li>The first operand to the GEP instruction is always a pointer and it must 732 be indexed.</li> 733 <li>There are no superfluous indices for the GEP instruction.</li> 734 <li>Trailing zero indices are superfluous for pointer aliasing, but not for 735 the types of the pointers.</li> 736 <li>Leading zero indices are not superfluous for pointer aliasing nor the 737 types of the pointers.</li> 738 </ol> 739 </div> 740 741 <!-- *********************************************************************** --> 742 743 <hr> 744 <address> 745 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img 746 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a> 747 <a href="http://validator.w3.org/check/referer"><img 748 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a> 749 <a href="http://llvm.org/">The LLVM Compiler Infrastructure</a><br/> 750 Last modified: $Date$ 751 </address> 752 </body> 753 </html> 754