| blob | 96d4fa1086869383e65b85e926b1bec74120b1e2 |
3 <html>
4 <head>
11 </head>
13 <body>
16 <ol>
21 <ol>
31 </ol>
32 </li>
34 <ol>
36 <ol>
38 </ol>
39 </li>
41 <ol>
49 </ol>
50 </li>
51 </ol>
52 </li>
54 <ol>
60 </ol>
61 </li>
63 <ol>
65 </ol>
66 </li>
68 <ol>
70 <ol>
77 </ol>
78 </li>
80 <ol>
90 </ol>
91 </li>
93 <ol>
100 </ol>
101 </li>
103 <ol>
107 </ol>
108 </li>
110 <ol>
117 </ol>
118 </li>
120 <ol>
133 </ol>
135 <ol>
142 </ol>
143 </li>
144 </ol>
145 </li>
147 <ol>
149 <ol>
153 </ol>
154 </li>
156 <ol>
160 </ol>
161 </li>
163 <ol>
171 </ol>
172 </li>
174 <ol>
180 </ol>
181 </li>
183 <ol>
190 </ol>
191 </li>
195 <ol>
197 Intrinsic</a></li>
198 </ol>
199 </li>
200 </ol>
201 </li>
202 </ol>
207 </div>
209 <!-- *********************************************************************** -->
211 <!-- *********************************************************************** -->
214 <p>This document is a reference manual for the LLVM assembly language.
215 LLVM is an SSA based representation that provides type safety,
216 low-level operations, flexibility, and the capability of representing
217 'all' high-level languages cleanly. It is the common code
218 representation used throughout all phases of the LLVM compilation
219 strategy.</p>
220 </div>
222 <!-- *********************************************************************** -->
224 <!-- *********************************************************************** -->
228 <p>The LLVM code representation is designed to be used in three
229 different forms: as an in-memory compiler IR, as an on-disk bytecode
230 representation (suitable for fast loading by a Just-In-Time compiler),
231 and as a human readable assembly language representation. This allows
232 LLVM to provide a powerful intermediate representation for efficient
233 compiler transformations and analysis, while providing a natural means
234 to debug and visualize the transformations. The three different forms
235 of LLVM are all equivalent. This document describes the human readable
236 representation and notation.</p>
238 <p>The LLVM representation aims to be light-weight and low-level
239 while being expressive, typed, and extensible at the same time. It
240 aims to be a "universal IR" of sorts, by being at a low enough level
241 that high-level ideas may be cleanly mapped to it (similar to how
242 microprocessors are "universal IR's", allowing many source languages to
243 be mapped to them). By providing type information, LLVM can be used as
244 the target of optimizations: for example, through pointer analysis, it
245 can be proven that a C automatic variable is never accessed outside of
246 the current function... allowing it to be promoted to a simple SSA
247 value instead of a memory location.</p>
249 </div>
251 <!-- _______________________________________________________________________ -->
256 <p>It is important to note that this document describes 'well formed'
257 LLVM assembly language. There is a difference between what the parser
258 accepts and what is considered 'well formed'. For example, the
259 following instruction is syntactically okay, but not well formed:</p>
262 <pre>
264 </pre>
265 </div>
268 its uses. The LLVM infrastructure provides a verification pass that may
269 be used to verify that an LLVM module is well formed. This pass is
270 automatically run by the parser after parsing input assembly and by
271 the optimizer before it outputs bytecode. The violations pointed out
272 by the verifier pass indicate bugs in transformation passes or input to
273 the parser.</p>
274 </div>
278 <!-- *********************************************************************** -->
280 <!-- *********************************************************************** -->
284 <p>LLVM uses three different forms of identifiers, for different
285 purposes:</p>
287 <ol>
288 <li>Named values are represented as a string of characters with a '%' prefix.
289 For example, %foo, %DivisionByZero, %a.really.long.identifier. The actual
291 Identifiers which require other characters in their names can be surrounded
293 in a name.</li>
295 <li>Unnamed values are represented as an unsigned numeric value with a '%'
300 </ol>
302 <p>LLVM requires that values start with a '%' sign for two reasons: Compilers
303 don't need to worry about name clashes with reserved words, and the set of
304 reserved words may be expanded in the future without penalty. Additionally,
305 unnamed identifiers allow a compiler to quickly come up with a temporary
306 variable without having to avoid symbol table conflicts.</p>
308 <p>Reserved words in LLVM are very similar to reserved words in other
309 languages. There are keywords for different opcodes
314 and others. These reserved words cannot conflict with variable names, because
315 none of them start with a '%' character.</p>
317 <p>Here is an example of LLVM code to multiply the integer variable
323 <pre>
325 </pre>
326 </div>
331 <pre>
333 </pre>
334 </div>
339 <pre>
343 </pre>
344 </div>
347 important lexical features of LLVM:</p>
349 <ol>
352 line.</li>
354 <li>Unnamed temporaries are created when the result of a computation is not
355 assigned to a named value.</li>
359 </ol>
361 <p>...and it also shows a convention that we follow in this document. When
362 demonstrating instructions, we will follow an instruction with a comment that
363 defines the type and name of value produced. Comments are shown in italic
364 text.</p>
366 </div>
368 <!-- *********************************************************************** -->
370 <!-- *********************************************************************** -->
372 <!-- ======================================================================= -->
374 </div>
379 translation unit of the input programs. Each module consists of
380 functions, global variables, and symbol table entries. Modules may be
381 combined together with the LLVM linker, which merges function (and
382 global variable) definitions, resolves forward declarations, and merges
388 href="#globalvars">constant</a> <a href="#t_array">[13 x i8]</a> c"hello world\0A\00" <i>; [13 x i8]*</i>
396 %cast210 = <a
400 <a
402 <a
404 </pre>
405 </div>
412 <p>In general, a module is made up of a list of global values,
413 where both functions and global variables are global values. Global values are
414 represented by a pointer to a memory location (in this case, a pointer to an
415 array of char, and a pointer to a function), and have one of the following <a
418 </div>
420 <!-- ======================================================================= -->
423 </div>
427 <p>
428 All Global Variables and Functions have one of the following types of linkage:
429 </p>
431 <dl>
435 <dd>Global values with internal linkage are only directly accessible by
436 objects in the current module. In particular, linking code into a module with
437 an internal global value may cause the internal to be renamed as necessary to
438 avoid collisions. Because the symbol is internal to the module, all
439 references can be updated. This corresponds to the notion of the
441 </dd>
446 the same name when linkage occurs. This is typically used to implement
447 inline functions, templates, or other code which must be generated in each
449 allowed to be discarded.
450 </dd>
456 used for globals that may be emitted in multiple translation units, but that
457 are not guaranteed to be emitted into every translation unit that uses them.
458 One example of this are common globals in C, such as "<tt>int X;</tt>" at
459 global scope.
460 </dd>
465 pointer to array type. When two global variables with appending linkage are
466 linked together, the two global arrays are appended together. This is the
467 LLVM, typesafe, equivalent of having the system linker append together
468 "sections" with identical names when .o files are linked.
469 </dd>
472 <dd>The semantics of this linkage follow the ELF model: the symbol is weak
473 until linked, if not linked, the symbol becomes null instead of being an
474 undefined reference.
475 </dd>
479 <dd>If none of the above identifiers are used, the global is externally
480 visible, meaning that it participates in linkage and can be used to resolve
481 external symbol references.
482 </dd>
483 </dl>
485 <p>
486 The next two types of linkage are targeted for Microsoft Windows platform
487 only. They are designed to support importing (exporting) symbols from (to)
488 DLLs.
489 </p>
491 <dl>
495 or variable via a global pointer to a pointer that is set up by the DLL
496 exporting the symbol. On Microsoft Windows targets, the pointer name is
498 </dd>
503 pointer to a pointer in a DLL, so that it can be referenced with the
506 name.
507 </dd>
509 </dl>
512 variable is defined to be internal, if another module defined a "<tt>.LC0</tt>"
513 variable and was linked with this one, one of the two would be renamed,
515 external (i.e., lacking any linkage declarations), they are accessible
516 outside of the current module.</p>
521 linkages.
522 </div>
524 <!-- ======================================================================= -->
527 </div>
533 specified for the call. The calling convention of any pair of dynamic
534 caller/callee must match, or the behavior of the program is undefined. The
535 following calling conventions are supported by LLVM, and more may be added in
536 the future:</p>
538 <dl>
541 <dd>This calling convention (the default if no other calling convention is
542 specified) matches the target C calling conventions. This calling convention
543 supports varargs function calls and tolerates some mismatch in the declared
544 prototype and implemented declaration of the function (as does normal C).
545 </dd>
549 <dd>This calling convention attempts to make calls as fast as possible
550 (e.g. by passing things in registers). This calling convention allows the
551 target to use whatever tricks it wants to produce fast code for the target,
552 without having to conform to an externally specified ABI. Implementations of
553 this convention should allow arbitrary tail call optimization to be supported.
554 This calling convention does not support varargs and requires the prototype of
555 all callees to exactly match the prototype of the function definition.
556 </dd>
560 <dd>This calling convention attempts to make code in the caller as efficient
561 as possible under the assumption that the call is not commonly executed. As
562 such, these calls often preserve all registers so that the call does not break
563 any live ranges in the caller side. This calling convention does not support
564 varargs and requires the prototype of all callees to exactly match the
565 prototype of the function definition.
566 </dd>
570 <dd>Any calling convention may be specified by number, allowing
571 target-specific calling conventions to be used. Target specific calling
572 conventions start at 64.
573 </dd>
574 </dl>
576 <p>More calling conventions can be added/defined on an as-needed basis, to
577 support pascal conventions or any other well-known target-independent
578 convention.</p>
580 </div>
582 <!-- ======================================================================= -->
585 </div>
589 <p>
590 All Global Variables and Functions have one of the following visibility styles:
591 </p>
593 <dl>
596 <dd>On ELF, default visibility means that the declaration is visible to other
597 modules and, in shared libraries, means that the declared entity may be
598 overridden. On Darwin, default visibility means that the declaration is
599 visible to other modules. Default visibility corresponds to "external
600 linkage" in the language.
601 </dd>
605 <dd>Two declarations of an object with hidden visibility refer to the same
606 object if they are in the same shared object. Usually, hidden visibility
607 indicates that the symbol will not be placed into the dynamic symbol table,
608 so no other module (executable or shared library) can reference it
609 directly.
610 </dd>
614 <dd>On ELF, protected visibility indicates that the symbol will be placed in
615 the dynamic symbol table, but that references within the defining module will
616 bind to the local symbol. That is, the symbol cannot be overridden by another
617 module.
618 </dd>
619 </dl>
621 </div>
623 <!-- ======================================================================= -->
626 </div>
630 <p>Global variables define regions of memory allocated at compilation time
631 instead of run-time. Global variables may optionally be initialized, may have
632 an explicit section to be placed in, and may have an optional explicit alignment
633 specified. A variable may be defined as "thread_local", which means that it
634 will not be shared by threads (each thread will have a separated copy of the
635 variable). A variable may be defined as a global "constant," which indicates
637 optimization, allowing the global data to be placed in the read-only section of
638 an executable, etc). Note that variables that need runtime initialization
641 <p>
643 constant, even if the final definition of the global is not. This capability
644 can be used to enable slightly better optimization of the program, but requires
645 the language definition to guarantee that optimizations based on the
646 'constantness' are valid for the translation units that do not include the
647 definition.
648 </p>
650 <p>As SSA values, global variables define pointer values that are in
651 scope (i.e. they dominate) all basic blocks in the program. Global
652 variables always define a pointer to their "content" type because they
653 describe a region of memory, and all memory objects in LLVM are
654 accessed through pointers.</p>
656 <p>LLVM allows an explicit section to be specified for globals. If the target
657 supports it, it will emit globals to the section specified.</p>
659 <p>An explicit alignment may be specified for a global. If not present, or if
660 the alignment is set to zero, the alignment of the global is set by the target
661 to whatever it feels convenient. If an explicit alignment is specified, the
662 global is forced to have at least that much alignment. All alignments must be
665 <p>For example, the following defines a global with an initializer, section,
666 and alignment:</p>
669 <pre>
671 </pre>
672 </div>
674 </div>
677 <!-- ======================================================================= -->
680 </div>
689 name, a (possibly empty) argument list (each with optional
691 optional alignment, an opening curly brace, a list of basic blocks, and a
692 closing curly brace.
694 LLVM function declarations consist of the "<tt>declare</tt>" keyword, an
699 name, a possibly empty list of arguments, and an optional alignment.</p>
701 <p>A function definition contains a list of basic blocks, forming the CFG for
702 the function. Each basic block may optionally start with a label (giving the
703 basic block a symbol table entry), contains a list of instructions, and ends
705 function return).</p>
707 <p>The first basic block in a function is special in two ways: it is immediately
708 executed on entrance to the function, and it is not allowed to have predecessor
709 basic blocks (i.e. there can not be any branches to the entry block of a
710 function). Because the block can have no predecessors, it also cannot have any
713 <p>LLVM functions are identified by their name and type signature. Hence, two
714 functions with the same name but different parameter lists or return values are
715 considered different functions, and LLVM will resolve references to each
716 appropriately.</p>
718 <p>LLVM allows an explicit section to be specified for functions. If the target
719 supports it, it will emit functions to the section specified.</p>
721 <p>An explicit alignment may be specified for a function. If not present, or if
722 the alignment is set to zero, the alignment of the function is set by the target
723 to whatever it feels convenient. If an explicit alignment is specified, the
724 function is forced to have at least that much alignment. All alignments must be
727 </div>
730 <!-- ======================================================================= -->
733 </div>
736 function or global variable or bitcast of global value). Aliases may have an
743 <pre>
745 </pre>
746 </div>
748 </div>
752 <!-- ======================================================================= -->
755 <p>The return type and each parameter of a function type may have a set of
757 used to communicate additional information about the result or parameters of
758 a function. Parameter attributes are considered to be part of the function
759 type so two functions types that differ only by the parameter attributes
760 are different function types.</p>
762 <p>Parameter attributes are simple keywords that follow the type specified. If
763 multiple parameter attributes are needed, they are space separated. For
764 example:</p>
767 <pre>
768 %someFunc = i16 (i8 sext %someParam) zext
769 %someFunc = i16 (i8 zext %someParam) zext
770 </pre>
771 </div>
773 <p>Note that the two function types above are unique because the parameter has
774 a different attribute (sext in the first one, zext in the second). Also note
775 that the attribute for the function result (zext) comes immediately after the
776 argument list.</p>
779 <dl>
781 <dd>This indicates that the parameter should be zero extended just before
782 a call to this function.</dd>
784 <dd>This indicates that the parameter should be sign extended just before
785 a call to this function.</dd>
787 <dd>This indicates that the parameter should be placed in register (if
788 possible) during assembling function call. Support for this attribute is
789 target-specific</dd>
791 <dd>This indicates that the parameter specifies the address of a structure
792 that is the return value of the function in the source program.</dd>
794 <dd>This indicates that the parameter not alias any other object or any
795 other "noalias" objects during the function call.
797 <dd>This function attribute indicates that the function never returns. This
798 indicates to LLVM that every call to this function should be treated as if
801 <dd>This function attribute indicates that the function type does not use
802 the unwind instruction and does not allow stack unwinding to propagate
803 through it.</dd>
804 </dl>
806 </div>
808 <!-- ======================================================================= -->
811 </div>
814 <p>
815 Modules may contain "module-level inline asm" blocks, which corresponds to the
816 GCC "file scope inline asm" blocks. These blocks are internally concatenated by
817 LLVM and treated as a single unit, but may be separated in the .ll file if
818 desired. The syntax is very simple:
819 </p>
822 <pre>
823 module asm "inline asm code goes here"
824 module asm "more can go here"
825 </pre>
826 </div>
828 <p>The strings can contain any character by escaping non-printable characters.
830 for the number.
831 </p>
833 <p>
834 The inline asm code is simply printed to the machine code .s file when
835 assembly code is generated.
836 </p>
837 </div>
839 <!-- ======================================================================= -->
842 </div>
845 <p>A module may specify a target specific data layout string that specifies how
846 data is to be laid out in memory. The syntax for the data layout is simply:</p>
849 separated by the minus sign character ('-'). Each specification starts with a
850 letter and may include other information after the letter to define some
851 aspect of the data layout. The specifications accepted are as follows: </p>
852 <dl>
854 <dd>Specifies that the target lays out data in big-endian form. That is, the
855 bits with the most significance have the lowest address location.</dd>
857 <dd>Specifies that hte target lays out data in little-endian form. That is,
858 the bits with the least significance have the lowest address location.</dd>
863 too.</dd>
865 <dd>This specifies the alignment for an integer type of a given bit
868 <dd>This specifies the alignment for a vector type of a given bit
871 <dd>This specifies the alignment for a floating point type of a given bit
873 (double).</dd>
875 <dd>This specifies the alignment for an aggregate type of a given bit
877 </dl>
878 <p>When constructing the data layout for a given target, LLVM starts with a
879 default set of specifications which are then (possibly) overriden by the
881 are given in this list:</p>
882 <ul>
896 </ul>
897 <p>When llvm is determining the alignment for a given type, it uses the
898 following rules:
899 <ol>
900 <li>If the type sought is an exact match for one of the specifications, that
901 specification is used.</li>
902 <li>If no match is found, and the type sought is an integer type, then the
903 smallest integer type that is larger than the bitwidth of the sought type is
904 used. If none of the specifications are larger than the bitwidth then the the
905 largest integer type is used. For example, given the default specifications
906 above, the i7 type will use the alignment of i8 (next largest) while both
907 i65 and i256 will use the alignment of i64 (largest specified).</li>
908 <li>If no match is found, and the type sought is a vector type, then the
909 largest vector type that is smaller than the sought vector type will be used
912 </ol>
913 </div>
915 <!-- *********************************************************************** -->
917 <!-- *********************************************************************** -->
921 <p>The LLVM type system is one of the most important features of the
922 intermediate representation. Being typed enables a number of
923 optimizations to be performed on the IR directly, without having to do
924 extra analyses on the side before the transformation. A strong type
925 system makes it easier to read the generated code and enables novel
926 analyses and transformations that are not feasible to perform on normal
927 three address code representations.</p>
929 </div>
931 <!-- ======================================================================= -->
934 <p>The primitive types are the fundamental building blocks of the LLVM
935 system. The current set of primitive types is as follows:</p>
940 <table>
941 <tbody>
945 </tbody>
946 </table>
947 </td>
949 <table>
950 <tbody>
954 </tbody>
955 </table>
956 </td>
957 </tr>
958 </table>
959 </div>
961 <!-- _______________________________________________________________________ -->
963 Classifications</a> </div>
965 <p>These different primitive types fall into a few useful
966 classifications:</p>
969 <tbody>
971 <tr>
974 </tr>
975 <tr>
978 </tr>
979 <tr>
983 </td>
984 </tr>
985 </tbody>
986 </table>
989 most important. Values of these types are the only ones which can be
990 produced by instructions, passed as arguments, or used as operands to
991 instructions. This means that all structures and arrays must be
992 manipulated either by pointer or by component.</p>
993 </div>
995 <!-- ======================================================================= -->
1000 <p>The real power in LLVM comes from the derived types in the system.
1001 This is what allows a programmer to represent arrays, functions,
1002 pointers, and other useful types. Note that these derived types may be
1003 recursive: For example, it is possible to have a two dimensional array.</p>
1005 </div>
1007 <!-- _______________________________________________________________________ -->
1013 <p>The integer type is a very simple derived type that simply specifies an
1014 arbitrary bit width for the integer type desired. Any bit width from 1 bit to
1019 <pre>
1020 iN
1021 </pre>
1024 value.</p>
1038 </td>
1045 An integer whose bit width is the answer. <br/>
1048 </td>
1049 </tr>
1050 </table>
1051 </div>
1053 <!-- _______________________________________________________________________ -->
1060 <p>The array type is a very simple derived type that arranges elements
1061 sequentially in memory. The array type requires a size (number of
1062 elements) and an underlying data type.</p>
1066 <pre>
1068 </pre>
1070 <p>The number of elements is a constant integer value; elementtype may
1071 be any type with a size.</p>
1080 </td>
1085 </td>
1086 </tr>
1087 </table>
1095 </td>
1100 </td>
1101 </tr>
1102 </table>
1104 <p>Note that 'variable sized arrays' can be implemented in LLVM with a zero
1105 length array. Normally, accesses past the end of an array are undefined in
1107 As a special case, however, zero length arrays are recognized to be variable
1108 length. This allows implementation of 'pascal style arrays' with the LLVM
1111 </div>
1113 <!-- _______________________________________________________________________ -->
1117 <p>The function type can be thought of as a function signature. It
1118 consists of a return type and a list of formal parameter types.
1119 Function types are usually used to build virtual function tables
1120 (which are structures of pointers to functions), for indirect function
1121 calls, and when defining a function.</p>
1122 <p>
1123 The return type of a function type cannot be an aggregate type.
1124 </p>
1129 which indicates that the function takes a variable number of arguments.
1130 Variable argument functions can access their arguments with the <a
1137 </td>
1140 </tt></td>
1145 </td>
1151 LLVM.
1152 </td>
1153 </tr>
1154 </table>
1156 </div>
1157 <!-- _______________________________________________________________________ -->
1161 <p>The structure type is used to represent a collection of data members
1162 together in memory. The packing of the field types is defined to match
1163 the ABI of the underlying processor. The elements of a structure may
1164 be any type that has a size.</p>
1168 instruction.</p>
1182 </tr>
1183 </table>
1184 </div>
1186 <!-- _______________________________________________________________________ -->
1188 </div>
1191 <p>The packed structure type is used to represent a collection of data members
1192 together in memory. There is no padding between fields. Further, the alignment
1193 of a packed structure is 1 byte. The elements of a packed structure may
1194 be any type that has a size.</p>
1198 instruction.</p>
1207 <td class="left"><tt>< { float, i32 (i32) * } ></tt></td>
1212 </tr>
1213 </table>
1214 </div>
1216 <!-- _______________________________________________________________________ -->
1220 <p>As in many languages, the pointer type represents a pointer or
1221 reference to another object, which must live in memory.</p>
1230 </td>
1237 </td>
1238 </tr>
1239 </table>
1240 </div>
1242 <!-- _______________________________________________________________________ -->
1248 <p>A vector type is a simple derived type that represents a vector
1249 of elements. Vector types are used when multiple primitive data
1250 are operated in parallel using a single instruction (SIMD).
1251 A vector type requires a size (number of
1252 elements) and an underlying primitive data type. Vectors must have a power
1258 <pre>
1260 </pre>
1262 <p>The number of elements is a constant integer value; elementtype may
1263 be any integer or floating point type.</p>
1273 </td>
1278 </td>
1279 </tr>
1280 </table>
1281 </div>
1283 <!-- _______________________________________________________________________ -->
1289 <p>Opaque types are used to represent unknown types in the system. This
1290 corresponds (for example) to the C notion of a foward declared structure type.
1291 In LLVM, opaque types can eventually be resolved to any type (not just a
1292 structure type).</p>
1296 <pre>
1297 opaque
1298 </pre>
1306 </td>
1308 An opaque type.<br/>
1309 </td>
1310 </tr>
1311 </table>
1312 </div>
1315 <!-- *********************************************************************** -->
1317 <!-- *********************************************************************** -->
1321 <p>LLVM has several different basic types of constants. This section describes
1322 them all and their syntax.</p>
1324 </div>
1326 <!-- ======================================================================= -->
1331 <dl>
1336 </dd>
1342 integer types.
1343 </dd>
1348 exponential notation (e.g. 1.23421e+2), or a more precise hexadecimal
1349 notation (see below). Floating point constants must have a <a
1357 </dl>
1359 <p>The one non-intuitive notation for constants is the optional hexadecimal form
1360 of floating point constants. For example, the form '<tt>double
1363 (and the only time that they are generated by the disassembler) is when a
1364 floating point constant must be emitted but it cannot be represented as a
1365 decimal floating point number. For example, NaN's, infinities, and other
1366 special values are represented in their IEEE hexadecimal format so that
1367 assembly and disassembly do not cause any bits to change in the constants.</p>
1369 </div>
1371 <!-- ======================================================================= -->
1373 </div>
1376 <p>Aggregate constants arise from aggregation of simple constants
1377 and smaller aggregate constants.</p>
1379 <dl>
1382 <dd>Structure constants are represented with notation similar to structure
1383 type definitions (a comma separated list of elements, surrounded by braces
1387 types of elements must match those specified by the type.
1388 </dd>
1392 <dd>Array constants are represented with notation similar to array type
1393 definitions (a comma separated list of elements, surrounded by square brackets
1396 types of elements must match those specified by the type.
1397 </dd>
1401 <dd>Vector constants are represented with notation similar to vector type
1402 definitions (a comma separated list of elements, surrounded by
1406 match those specified by the type.
1407 </dd>
1413 This is often used to avoid having to print large zero initializers (e.g. for
1414 large arrays) and is always exactly equivalent to using explicit zero
1415 initializers.
1416 </dd>
1417 </dl>
1419 </div>
1421 <!-- ======================================================================= -->
1424 </div>
1430 constants. These constants are explicitly referenced when the <a
1433 file:</p>
1436 <pre>
1437 @X = global i32 17
1438 @Y = global i32 42
1439 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
1440 </pre>
1441 </div>
1443 </div>
1445 <!-- ======================================================================= -->
1449 no specific value. Undefined values may be of any type and be used anywhere
1450 a constant is permitted.</p>
1452 <p>Undefined values indicate to the compiler that the program is well defined
1453 no matter what value is used, giving the compiler more freedom to optimize.
1454 </p>
1455 </div>
1457 <!-- ======================================================================= -->
1459 </div>
1463 <p>Constant expressions are used to allow expressions involving other constants
1464 to be used as constants. Constant expressions may be of any <a
1466 that does not have side effects (e.g. load and call are not supported). The
1467 following is the syntax for constant expressions:</p>
1469 <dl>
1471 <dd>Truncate a constant to another type. The bit size of CST must be larger
1472 than the bit size of TYPE. Both types must be integers.</dd>
1475 <dd>Zero extend a constant to another type. The bit size of CST must be
1476 smaller or equal to the bit size of TYPE. Both types must be integers.</dd>
1479 <dd>Sign extend a constant to another type. The bit size of CST must be
1480 smaller or equal to the bit size of TYPE. Both types must be integers.</dd>
1483 <dd>Truncate a floating point constant to another floating point type. The
1484 size of CST must be larger than the size of TYPE. Both types must be
1485 floating point.</dd>
1488 <dd>Floating point extend a constant to another type. The size of CST must be
1489 smaller or equal to the size of TYPE. Both types must be floating point.</dd>
1492 <dd>Convert a floating point constant to the corresponding unsigned integer
1493 constant. TYPE must be an integer type. CST must be floating point. If the
1494 value won't fit in the integer type, the results are undefined.</dd>
1497 <dd>Convert a floating point constant to the corresponding signed integer
1498 constant. TYPE must be an integer type. CST must be floating point. If the
1499 value won't fit in the integer type, the results are undefined.</dd>
1502 <dd>Convert an unsigned integer constant to the corresponding floating point
1503 constant. TYPE must be floating point. CST must be of integer type. If the
1504 value won't fit in the floating point type, the results are undefined.</dd>
1507 <dd>Convert a signed integer constant to the corresponding floating point
1508 constant. TYPE must be floating point. CST must be of integer type. If the
1509 value won't fit in the floating point type, the results are undefined.</dd>
1512 <dd>Convert a pointer typed constant to the corresponding integer constant
1513 TYPE must be an integer type. CST must be of pointer type. The CST value is
1514 zero extended, truncated, or unchanged to make it fit in TYPE.</dd>
1517 <dd>Convert a integer constant to a pointer constant. TYPE must be a
1518 pointer type. CST must be of integer type. The CST value is zero extended,
1519 truncated, or unchanged to make it fit in a pointer size. This one is
1523 <dd>Convert a constant, CST, to another TYPE. The size of CST and TYPE must be
1524 identical (same number of bits). The conversion is done as if the CST value
1525 was stored to memory and read back as TYPE. In other words, no bits change
1526 with this operator, just the type. This can be used for conversion of
1527 vector types to any other type, as long as they have the same bit width. For
1528 pointers it is only valid to cast to another pointer type.
1529 </dd>
1535 instruction, the index list may have zero or more indexes, which are required
1541 constants.</dd>
1552 operation</a> on constants.
1567 <dd>Perform the specified operation of the LHS and RHS constants. OPCODE may
1569 binary</a> operations. The constraints on operands are the same as those for
1570 the corresponding instruction (e.g. no bitwise operations on floating point
1571 values are allowed).</dd>
1572 </dl>
1573 </div>
1575 <!-- *********************************************************************** -->
1577 <!-- *********************************************************************** -->
1579 <!-- ======================================================================= -->
1582 </div>
1586 <p>
1588 Module-Level Inline Assembly</a>) through the use of a special value. This
1589 value represents the inline assembler as a string (containing the instructions
1590 to emit), a list of operand constraints (stored as a string), and a flag that
1591 indicates whether or not the inline asm expression has side effects. An example
1592 inline assembler expression is:
1593 </p>
1596 <pre>
1598 </pre>
1599 </div>
1601 <p>
1604 </p>
1607 <pre>
1609 </pre>
1610 </div>
1612 <p>
1613 Inline asms with side effects not visible in the constraint list must be marked
1614 as having side effects. This is done through the use of the
1616 </p>
1619 <pre>
1621 </pre>
1622 </div>
1624 <p>TODO: The format of the asm and constraints string still need to be
1625 documented here. Constraints on what can be done (e.g. duplication, moving, etc
1626 need to be documented).
1627 </p>
1629 </div>
1631 <!-- *********************************************************************** -->
1633 <!-- *********************************************************************** -->
1637 <p>The LLVM instruction set consists of several different
1644 </div>
1646 <!-- ======================================================================= -->
1648 Instructions</a> </div>
1653 basic block in a program ends with a "Terminator" instruction, which
1654 indicates which block should be executed after the current block is
1656 value: they produce control flow, not values (the one exception being
1665 </div>
1667 <!-- _______________________________________________________________________ -->
1669 Instruction</a> </div>
1674 </pre>
1677 value) from a function back to the caller.</p>
1679 returns a value and then causes control flow, and one that just causes
1680 control flow to occur.</p>
1685 instruction inside of the function that returns a value that does not
1686 match the return type of the function.</p>
1689 returns back to the calling function's context. If the caller is a "<a
1691 the instruction after the call. If the caller was an "<a
1693 at the beginning of the "normal" destination block. If the instruction
1694 returns a value, that value shall set the call or invoke instruction's
1695 return value.</p>
1699 </pre>
1700 </div>
1701 <!-- _______________________________________________________________________ -->
1705 <pre> br i1 <cond>, label <iftrue>, label <iffalse><br> br label <dest> <i>; Unconditional branch</i>
1706 </pre>
1709 transfer to a different basic block in the current function. There are
1710 two forms of this instruction, corresponding to a conditional branch
1711 and an unconditional branch.</p>
1723 <pre>Test:<br> %cond = <a href="#i_icmp">icmp</a> eq, i32 %a, %b<br> br i1 %cond, label %IfEqual, label %IfUnequal<br>IfEqual:<br> <a
1725 </div>
1726 <!-- _______________________________________________________________________ -->
1729 </div>
1734 <pre>
1735 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
1736 </pre>
1742 instruction, allowing a branch to occur to one of many possible
1743 destinations.</p>
1751 table is not allowed to contain duplicate constant entries.</p>
1757 table is searched for the given value. If the value is found, control flow is
1758 transfered to the corresponding destination; otherwise, control flow is
1759 transfered to the default destination.</p>
1763 <p>Depending on properties of the target machine and the particular
1765 ways. For example, it could be generated as a series of chained conditional
1766 branches or with a lookup table.</p>
1770 <pre>
1773 switch i32 %Val, label %truedest [i32 0, label %falsedest ]
1776 switch i32 0, label %dest [ ]
1779 switch i32 %val, label %otherwise [ i32 0, label %onzero
1780 i32 1, label %onone
1781 i32 2, label %ontwo ]
1782 </pre>
1783 </div>
1785 <!-- _______________________________________________________________________ -->
1788 </div>
1794 <pre>
1795 <result> = invoke [<a href="#callingconv">cconv</a>] <ptr to function ty> %<function ptr val>(<function args>)
1797 </pre>
1802 function, with the possibility of control flow transfer to either the
1814 <ol>
1815 <li>
1817 convention</a> the call should use. If none is specified, the call defaults
1818 to using C calling conventions.
1819 </li>
1821 function value being invoked. In most cases, this is a direct function
1823 an arbitrary pointer to function value.
1824 </li>
1827 function to be invoked. </li>
1830 signature argument types. If the function signature indicates the function
1831 accepts a variable number of arguments, the extra arguments can be
1832 specified. </li>
1840 </ol>
1846 difference is that it establishes an association with a label, which is used by
1847 the runtime library to unwind the stack.</p>
1849 <p>This instruction is used in languages with destructors to ensure that proper
1851 exception. Additionally, this is important for implementation of
1855 <pre>
1856 %retval = invoke i32 %Test(i32 15) to label %Continue
1860 </pre>
1861 </div>
1864 <!-- _______________________________________________________________________ -->
1867 Instruction</a> </div>
1872 <pre>
1873 unwind
1874 </pre>
1879 at the first callee in the dynamic call stack which used an <a
1881 primarily used to implement exception handling.</p>
1886 immediately halt. The dynamic call stack is then searched for the first <a
1888 execution continues at the "exceptional" destination block specified by the
1890 dynamic call chain, undefined behavior results.</p>
1891 </div>
1893 <!-- _______________________________________________________________________ -->
1896 Instruction</a> </div>
1901 <pre>
1902 unreachable
1903 </pre>
1908 instruction is used to inform the optimizer that a particular portion of the
1909 code is not reachable. This can be used to indicate that the code after a
1910 no-return function cannot be reached, and other facts.</p>
1915 </div>
1919 <!-- ======================================================================= -->
1922 <p>Binary operators are used to do most of the computation in a
1923 program. They require two operands, execute an operation on them, and
1924 produce a single value. The operands might represent
1926 The result value of a binary operator is not
1927 necessarily the same type as its operands.</p>
1929 </div>
1930 <!-- _______________________________________________________________________ -->
1932 Instruction</a> </div>
1936 </pre>
1943 Both arguments must have identical types.</p>
1945 <p>The value produced is the integer or floating point sum of the two
1946 operands.</p>
1949 </pre>
1950 </div>
1951 <!-- _______________________________________________________________________ -->
1953 Instruction</a> </div>
1957 </pre>
1960 operands.</p>
1962 instruction present in most other intermediate representations.</p>
1966 values.
1968 Both arguments must have identical types.</p>
1970 <p>The value produced is the integer or floating point difference of
1971 the two operands.</p>
1973 <pre>
1976 </pre>
1977 </div>
1978 <!-- _______________________________________________________________________ -->
1980 Instruction</a> </div>
1984 </pre>
1987 operands.</p>
1991 values.
1993 Both arguments must have identical types.</p>
1995 <p>The value produced is the integer or floating point product of the
1996 two operands.</p>
1997 <p>Because the operands are the same width, the result of an integer
1998 multiplication is the same whether the operands should be deemed unsigned or
1999 signed.</p>
2002 </pre>
2003 </div>
2004 <!-- _______________________________________________________________________ -->
2006 </a></div>
2010 </pre>
2013 operands.</p>
2018 of the values in which case the elements must be integers.</p>
2020 <p>The value produced is the unsigned integer quotient of the two operands. This
2021 instruction always performs an unsigned division operation, regardless of
2022 whether the arguments are unsigned or not.</p>
2025 </pre>
2026 </div>
2027 <!-- _______________________________________________________________________ -->
2029 </a> </div>
2033 </pre>
2036 operands.</p>
2041 of the values in which case the elements must be integers.</p>
2043 <p>The value produced is the signed integer quotient of the two operands. This
2044 instruction always performs a signed division operation, regardless of whether
2045 the arguments are signed or not.</p>
2048 </pre>
2049 </div>
2050 <!-- _______________________________________________________________________ -->
2052 Instruction</a> </div>
2056 </pre>
2059 operands.</p>
2064 versions of floating point values.</p>
2069 </pre>
2070 </div>
2071 <!-- _______________________________________________________________________ -->
2073 </div>
2077 </pre>
2080 unsigned division of its two arguments.</p>
2084 types.</p>
2087 This instruction always performs an unsigned division to get the remainder,
2088 regardless of whether the arguments are unsigned or not.</p>
2091 </pre>
2093 </div>
2094 <!-- _______________________________________________________________________ -->
2096 Instruction</a> </div>
2100 </pre>
2103 signed division of its two operands.</p>
2107 types.</p>
2112 a value. For more information about the difference, see <a
2114 Math Forum</a>. For a table of how this is implemented in various languages,
2119 </pre>
2121 </div>
2122 <!-- _______________________________________________________________________ -->
2124 Instruction</a> </div>
2128 </pre>
2131 division of its two operands.</p>
2135 identical types.</p>
2140 </pre>
2141 </div>
2143 <!-- ======================================================================= -->
2145 Operations</a> </div>
2147 <p>Bitwise binary operators are used to do various forms of
2148 bit-twiddling in a program. They are generally very efficient
2149 instructions and can commonly be strength reduced from other
2150 instructions. They require two operands, execute an operation on them,
2151 and produce a single value. The resulting value of the bitwise binary
2152 operators is always the same type as its first operand.</p>
2153 </div>
2155 <!-- _______________________________________________________________________ -->
2157 Instruction</a> </div>
2161 </pre>
2164 the left a specified number of bits.</p>
2174 </pre>
2175 </div>
2176 <!-- _______________________________________________________________________ -->
2178 Instruction</a> </div>
2182 </pre>
2186 operand shifted to the right a specified number of bits with zero fill.</p>
2193 <p>This instruction always performs a logical shift right operation. The most
2194 significant bits of the result will be filled with zero bits after the
2195 shift.</p>
2198 <pre>
2203 </pre>
2204 </div>
2206 <!-- _______________________________________________________________________ -->
2208 Instruction</a> </div>
2213 </pre>
2217 operand shifted to the right a specified number of bits with sign extension.</p>
2224 <p>This instruction always performs an arithmetic shift right operation,
2225 The most significant bits of the result will be filled with the sign bit
2229 <pre>
2234 </pre>
2235 </div>
2237 <!-- _______________________________________________________________________ -->
2239 Instruction</a> </div>
2243 </pre>
2246 its two operands.</p>
2250 identical types.</p>
2253 <p> </p>
2256 <tbody>
2257 <tr>
2261 </tr>
2262 <tr>
2266 </tr>
2267 <tr>
2271 </tr>
2272 <tr>
2276 </tr>
2277 <tr>
2281 </tr>
2282 </tbody>
2283 </table>
2284 </div>
2289 </pre>
2290 </div>
2291 <!-- _______________________________________________________________________ -->
2296 </pre>
2299 or of its two operands.</p>
2303 identical types.</p>
2306 <p> </p>
2309 <tbody>
2310 <tr>
2314 </tr>
2315 <tr>
2319 </tr>
2320 <tr>
2324 </tr>
2325 <tr>
2329 </tr>
2330 <tr>
2334 </tr>
2335 </tbody>
2336 </table>
2337 </div>
2342 </pre>
2343 </div>
2344 <!-- _______________________________________________________________________ -->
2346 Instruction</a> </div>
2350 </pre>
2358 identical types.</p>
2361 <p> </p>
2364 <tbody>
2365 <tr>
2369 </tr>
2370 <tr>
2374 </tr>
2375 <tr>
2379 </tr>
2380 <tr>
2384 </tr>
2385 <tr>
2389 </tr>
2390 </tbody>
2391 </table>
2392 </div>
2393 <p> </p>
2399 </pre>
2400 </div>
2402 <!-- ======================================================================= -->
2405 </div>
2409 <p>LLVM supports several instructions to represent vector operations in a
2410 target-independent manner. These instructions cover the element-access and
2411 vector-specific operations needed to process vectors effectively. While LLVM
2412 does directly support these vector operations, many sophisticated algorithms
2413 will want to use target-specific intrinsics to take full advantage of a specific
2414 target.</p>
2416 </div>
2418 <!-- _______________________________________________________________________ -->
2421 </div>
2427 <pre>
2428 <result> = extractelement <n x <ty>> <val>, i32 <idx> <i>; yields <ty></i>
2429 </pre>
2433 <p>
2435 element from a vector at a specified index.
2436 </p>
2441 <p>
2444 an index indicating the position from which to extract the element.
2445 The index may be a variable.</p>
2449 <p>
2450 The result is a scalar of the same type as the element type of
2453 results are undefined.
2454 </p>
2458 <pre>
2460 </pre>
2461 </div>
2464 <!-- _______________________________________________________________________ -->
2467 </div>
2473 <pre>
2474 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> <i>; yields <n x <ty>></i>
2475 </pre>
2479 <p>
2481 element into a vector at a specified index.
2482 </p>
2487 <p>
2490 scalar value whose type must equal the element type of the first
2491 operand. The third operand is an index indicating the position at
2492 which to insert the value. The index may be a variable.</p>
2496 <p>
2501 </p>
2505 <pre>
2507 </pre>
2508 </div>
2510 <!-- _______________________________________________________________________ -->
2513 </div>
2519 <pre>
2520 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <n x i32> <mask> <i>; yields <n x <ty>></i>
2521 </pre>
2525 <p>
2527 from two input vectors, returning a vector of the same type.
2528 </p>
2532 <p>
2534 with types that match each other and types that match the result of the
2535 instruction. The third argument is a shuffle mask, which has the same number
2536 of elements as the other vector type, but whose element type is always 'i32'.
2537 </p>
2539 <p>
2540 The shuffle mask operand is required to be a constant vector with either
2541 constant integer or undef values.
2542 </p>
2546 <p>
2547 The elements of the two input vectors are numbered from left to right across
2548 both of the vectors. The shuffle mask operand specifies, for each element of
2549 the result vector, which element of the two input registers the result element
2550 gets. The element selector may be undef (meaning "don't care") and the second
2551 operand may be undef if performing a shuffle from only one vector.
2552 </p>
2556 <pre>
2560 <4 x i32> <i32 0, i32 1, i32 2, i32 3> <i>; yields <4 x i32></i> - Identity shuffle.
2561 </pre>
2562 </div>
2565 <!-- ======================================================================= -->
2568 </div>
2572 <p>A key design point of an SSA-based representation is how it
2573 represents memory. In LLVM, no memory locations are in SSA form, which
2574 makes things very simple. This section describes how to read, write,
2575 allocate, and free memory in LLVM.</p>
2577 </div>
2579 <!-- _______________________________________________________________________ -->
2582 </div>
2588 <pre>
2589 <result> = malloc <type>[, i32 <NumElements>][, align <alignment>] <i>; yields {type*}:result</i>
2590 </pre>
2595 heap and returns a pointer to it.</p>
2601 bytes of memory from the operating system and returns a pointer of the
2602 appropriate type to the program. If "NumElements" is specified, it is the
2603 number of elements allocated. If an alignment is specified, the value result
2604 of the allocation is guaranteed to be aligned to at least that boundary. If
2605 not specified, or if zero, the target can choose to align the allocation on any
2606 convenient boundary.</p>
2613 a pointer is returned.</p>
2617 <pre>
2625 </pre>
2626 </div>
2628 <!-- _______________________________________________________________________ -->
2631 </div>
2637 <pre>
2639 </pre>
2644 memory heap to be reallocated in the future.</p>
2650 instruction.</p>
2654 <p>Access to the memory pointed to by the pointer is no longer defined
2655 after this instruction executes.</p>
2659 <pre>
2661 free [4 x i8]* %array
2662 </pre>
2663 </div>
2665 <!-- _______________________________________________________________________ -->
2668 </div>
2674 <pre>
2675 <result> = alloca <type>[, i32 <NumElements>][, align <alignment>] <i>; yields {type*}:result</i>
2676 </pre>
2681 currently executing function, to be automatically released when this function
2682 returns to its caller.</p>
2687 bytes of memory on the runtime stack, returning a pointer of the
2688 appropriate type to the program. If "NumElements" is specified, it is the
2689 number of elements allocated. If an alignment is specified, the value result
2690 of the allocation is guaranteed to be aligned to at least that boundary. If
2691 not specified, or if zero, the target can choose to align the allocation on any
2692 convenient boundary.</p>
2700 instruction is commonly used to represent automatic variables that must
2701 have an address available. When the function returns (either with the <tt><a
2703 instructions), the memory is reclaimed.</p>
2707 <pre>
2712 </pre>
2713 </div>
2715 <!-- _______________________________________________________________________ -->
2717 Instruction</a> </div>
2720 <pre> <result> = load <ty>* <pointer>[, align <alignment>]<br> <result> = volatile load <ty>* <pointer>[, align <alignment>]<br></pre>
2725 address from which to load. The pointer must point to a <a
2730 instructions. </p>
2735 <a
2738 </pre>
2739 </div>
2740 <!-- _______________________________________________________________________ -->
2742 Instruction</a> </div>
2745 <pre> store <ty> <value>, <ty>* <pointer>[, align <alignment>] <i>; yields {void}</i>
2746 volatile store <ty> <value>, <ty>* <pointer>[, align <alignment>] <i>; yields {void}</i>
2747 </pre>
2755 optimizer is not allowed to modify the number or order of execution of
2763 <a
2766 </pre>
2767 </div>
2769 <!-- _______________________________________________________________________ -->
2772 </div>
2776 <pre>
2778 </pre>
2782 <p>
2784 subelement of an aggregate data structure.</p>
2788 <p>This instruction takes a list of integer operands that indicate what
2789 elements of the aggregate object to index to. The actual types of the arguments
2790 provided depend on the type of the first pointer argument. The
2792 levels of a structure or to a specific index in an array. When indexing into a
2797 <p>For example, let's consider a C code fragment and how it gets
2798 compiled to LLVM:</p>
2801 <pre>
2802 struct RT {
2803 char A;
2805 char C;
2806 };
2807 struct ST {
2808 int X;
2809 double Y;
2810 struct RT Z;
2811 };
2813 int *foo(struct ST *s) {
2815 }
2816 </pre>
2817 </div>
2822 <pre>
2824 %ST = type { i32, double, %RT }
2826 define i32* %foo(%ST* %s) {
2827 entry:
2829 ret i32* %reg
2830 }
2831 </pre>
2832 </div>
2845 }</tt>' type, a structure. The second index indexes into the third element of
2847 i8 }</tt>' type, another structure. The third index indexes into the second
2849 array. The two dimensions of the array are subscripted into, yielding an
2853 <p>Note that it is perfectly legal to index partially through a
2854 structure, returning a pointer to an inner element. Because of this,
2855 the LLVM code for the given testcase is equivalent to:</p>
2857 <pre>
2858 define i32* %foo(%ST* %s) {
2864 ret i32* %t5
2865 }
2866 </pre>
2868 <p>Note that it is undefined to access an array out of bounds: array and
2869 pointer indexes must always be within the defined bounds of the array type.
2870 The one exception for this rules is zero length arrays. These arrays are
2871 defined to be accessible as variable length arrays, which requires access
2872 beyond the zero'th element.</p>
2874 <p>The getelementptr instruction is often confusing. For some more insight
2880 <pre>
2883 </pre>
2884 </div>
2886 <!-- ======================================================================= -->
2888 </div>
2890 <p>The instructions in this category are the conversion instructions (casting)
2891 which all take a single operand and a type. They perform various bit conversions
2892 on the operand.</p>
2893 </div>
2895 <!-- _______________________________________________________________________ -->
2898 </div>
2902 <pre>
2904 </pre>
2907 <p>
2909 </p>
2912 <p>
2920 <p>
2924 It will always truncate bits.</p>
2927 <pre>
2931 </pre>
2932 </div>
2934 <!-- _______________________________________________________________________ -->
2937 </div>
2941 <pre>
2943 </pre>
2964 <pre>
2967 </pre>
2968 </div>
2970 <!-- _______________________________________________________________________ -->
2973 </div>
2977 <pre>
2979 </pre>
2985 <p>
2993 <p>
3001 <pre>
3004 </pre>
3005 </div>
3007 <!-- _______________________________________________________________________ -->
3010 </div>
3016 <pre>
3018 </pre>
3039 <pre>
3042 </pre>
3043 </div>
3045 <!-- _______________________________________________________________________ -->
3048 </div>
3052 <pre>
3054 </pre>
3058 floating point value.</p>
3064 type must be smaller than the destination type.</p>
3074 <pre>
3077 </pre>
3078 </div>
3080 <!-- _______________________________________________________________________ -->
3083 </div>
3087 <pre>
3089 </pre>
3094 </p>
3105 the results are undefined.</p>
3107 <p>When converting to i1, the conversion is done as a comparison against
3112 <pre>
3116 </pre>
3117 </div>
3119 <!-- _______________________________________________________________________ -->
3122 </div>
3126 <pre>
3128 </pre>
3133 </p>
3145 the results are undefined.</p>
3147 <p>When converting to i1, the conversion is done as a comparison against
3152 <pre>
3156 </pre>
3157 </div>
3159 <!-- _______________________________________________________________________ -->
3162 </div>
3166 <pre>
3168 </pre>
3182 integer quantity and converts it to the corresponding floating point value. If
3183 the value cannot fit in the floating point value, the results are undefined.</p>
3187 <pre>
3190 </pre>
3191 </div>
3193 <!-- _______________________________________________________________________ -->
3196 </div>
3200 <pre>
3202 </pre>
3215 integer quantity and converts it to the corresponding floating point value. If
3216 the value cannot fit in the floating point value, the results are undefined.</p>
3219 <pre>
3222 </pre>
3223 </div>
3225 <!-- _______________________________________________________________________ -->
3228 </div>
3232 <pre>
3234 </pre>
3248 truncating or zero extending that value to the size of the integer type. If
3252 change.</p>
3255 <pre>
3258 </pre>
3259 </div>
3261 <!-- _______________________________________________________________________ -->
3264 </div>
3268 <pre>
3270 </pre>
3278 value to cast, and a type to cast it to, which must be a
3286 the size of a pointer then a zero extension is done. If they are the same size,
3290 <pre>
3294 </pre>
3295 </div>
3297 <!-- _______________________________________________________________________ -->
3300 </div>
3304 <pre>
3306 </pre>
3314 a first class value, and a type to cast it to, which must also be a <a
3317 type is a pointer, the destination type must also be a pointer.</p>
3324 converted to other pointer types with this instruction. To convert pointers to
3329 <pre>
3333 </pre>
3334 </div>
3336 <!-- ======================================================================= -->
3340 instructions, which defy better classification.</p>
3341 </div>
3343 <!-- _______________________________________________________________________ -->
3345 </div>
3348 <pre> <result> = icmp <cond> <ty> <var1>, <var2> <i>; yields {i1}:result</i>
3349 </pre>
3352 of its two integer operands.</p>
3355 the condition code indicating the kind of comparison to perform. It is not
3356 a value, just a keyword. The possible condition code are:
3357 <ol>
3368 </ol>
3375 <ol>
3378 </li>
3397 </ol>
3399 values are compared as if they were integers.</p>
3408 </pre>
3409 </div>
3411 <!-- _______________________________________________________________________ -->
3413 </div>
3416 <pre> <result> = fcmp <cond> <ty> <var1>, <var2> <i>; yields {i1}:result</i>
3417 </pre>
3420 of its floating point operands.</p>
3423 the condition code indicating the kind of comparison to perform. It is not
3424 a value, just a keyword. The possible condition code are:
3425 <ol>
3442 </ol>
3447 types.</p>
3452 <ol>
3481 </ol>
3488 </pre>
3489 </div>
3491 <!-- _______________________________________________________________________ -->
3493 Instruction</a> </div>
3499 the SSA graph representing the function.</p>
3501 <p>The type of the incoming values is specified with the first type
3503 as arguments, with one pair for each predecessor basic block of the
3505 type may be used as the value arguments to the PHI node. Only labels
3506 may be used as the label arguments.</p>
3507 <p>There must be no non-phi instructions between the start of a basic
3508 block and the PHI instructions: i.e. PHI instructions must be first in
3509 a basic block.</p>
3512 specified by the pair corresponding to the predecessor basic block that executed
3513 just prior to the current block.</p>
3515 <pre>Loop: ; Infinite loop that counts from 0 on up...<br> %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]<br> %nextindvar = add i32 %indvar, 1<br> br label %Loop<br></pre>
3516 </div>
3518 <!-- _______________________________________________________________________ -->
3521 </div>
3527 <pre>
3528 <result> = select i1 <cond>, <ty> <val1>, <ty> <val2> <i>; yields ty</i>
3529 </pre>
3533 <p>
3535 condition, without branching.
3536 </p>
3541 <p>
3542 The '<tt>select</tt>' instruction requires a boolean value indicating the condition, and two values of the same <a href="#t_firstclass">first class</a> type.
3543 </p>
3547 <p>
3548 If the boolean condition evaluates to true, the instruction returns the first
3549 value argument; otherwise, it returns the second value argument.
3550 </p>
3554 <pre>
3556 </pre>
3557 </div>
3560 <!-- _______________________________________________________________________ -->
3563 </div>
3568 <pre>
3569 <result> = [tail] call [<a href="#callingconv">cconv</a>] <ty>* <fnptrval>(<param list>)
3570 </pre>
3580 <ol>
3581 <li>
3583 any allocas or varargs in the caller. If the "tail" marker is present, the
3584 function call is eligible for tail call optimization. Note that calls may
3587 </li>
3588 <li>
3590 convention</a> the call should use. If none is specified, the call defaults
3591 to using C calling conventions.
3592 </li>
3593 <li>
3595 being invoked. The argument types must match the types implied by this
3596 signature. This type can be omitted if the function is not varargs and
3597 if the function type does not return a pointer to a function.</p>
3598 </li>
3599 <li>
3601 be invoked. In most cases, this is a direct function invocation, but
3603 to function value.</p>
3604 </li>
3605 <li>
3607 function signature argument types. All arguments must be of
3609 indicates the function accepts a variable number of arguments, the extra
3610 arguments can be specified.</p>
3611 </li>
3612 </ol>
3617 transfer to a specified function, with its incoming arguments bound to
3619 instruction in the called function, control flow continues with the
3620 instruction after the function call, and the return value of the
3621 function is bound to the result argument. This is a simpler case of
3626 <pre>
3627 %retval = call i32 %test(i32 %argc)
3629 %X = tail call i32 %foo()
3631 </pre>
3633 </div>
3635 <!-- _______________________________________________________________________ -->
3638 </div>
3644 <pre>
3646 </pre>
3651 the "variable argument" area of a function call. It is used to implement the
3657 the argument. It returns a value of the specified argument type and
3669 <p>It is legal for this instruction to be called in a function which does not
3671 function.</p>
3675 argument.</p>
3681 </div>
3683 <!-- *********************************************************************** -->
3685 <!-- *********************************************************************** -->
3690 well known names and semantics and are required to follow certain restrictions.
3691 Overall, these intrinsics represent an extension mechanism for the LLVM
3692 language that does not require changing all of the transformations in LLVM when
3693 adding to the language (or the bytecode reader/writer, the parser, etc...).</p>
3696 prefix is reserved in LLVM for intrinsic names; thus, function names may not
3697 begin with this prefix. Intrinsic functions must always be external functions:
3698 you cannot define the body of intrinsic functions. Intrinsic functions may
3699 only be used in call or invoke instructions: it is illegal to take the address
3700 of an intrinsic function. Additionally, because intrinsic functions are part
3701 of the LLVM language, it is required if any are added that they be documented
3702 here.</p>
3704 <p>Some intrinsic functions can be overloaded, i.e., the intrinsic represents
3705 a family of functions that perform the same operation but on different data
3706 types. This is most frequent with the integer types. Since LLVM can represent
3707 over 8 million different integer types, there is a way to declare an intrinsic
3708 that can be overloaded based on its arguments. Such an intrinsic will have the
3709 names of its argument types encoded into its function name, each
3711 integer of any width. This leads to a family of functions such as
3713 </p>
3716 <p>To learn how to add an intrinsic function, please see the
3718 </p>
3720 </div>
3722 <!-- ======================================================================= -->
3725 </div>
3731 intrinsic functions. These functions are related to the similarly
3734 <p>All of these functions operate on arguments that use a
3735 target-specific value type "<tt>va_list</tt>". The LLVM assembly
3736 language reference manual does not define what this type is, so all
3737 transformations should be prepared to handle these functions regardless of
3738 the type used.</p>
3741 instruction and the variable argument handling intrinsic functions are
3742 used.</p>
3745 <pre>
3746 define i32 @test(i32 %X, ...) {
3747 ; Initialize variable argument processing
3748 %ap = alloca i8*
3749 %ap2 = bitcast i8** %ap to i8*
3750 call void @llvm.va_start(i8* %ap2)
3752 ; Read a single integer argument
3753 %tmp = va_arg i8** %ap, i32
3755 ; Demonstrate usage of llvm.va_copy and llvm.va_end
3756 %aq = alloca i8*
3757 %aq2 = bitcast i8** %aq to i8*
3758 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
3759 call void @llvm.va_end(i8* %aq2)
3761 ; Stop processing of arguments.
3762 call void @llvm.va_end(i8* %ap2)
3763 ret i32 %tmp
3764 }
3766 declare void @llvm.va_start(i8*)
3767 declare void @llvm.va_copy(i8*, i8*)
3768 declare void @llvm.va_end(i8*)
3769 </pre>
3770 </div>
3772 </div>
3774 <!-- _______________________________________________________________________ -->
3777 </div>
3795 macro available in C. In a target-dependent way, it initializes the
3799 last argument of the function as the compiler can figure that out.</p>
3801 </div>
3803 <!-- _______________________________________________________________________ -->
3806 </div>
3824 macro available in C. In a target-dependent way, it destroys the
3830 </div>
3832 <!-- _______________________________________________________________________ -->
3835 </div>
3841 <pre>
3843 </pre>
3848 from the source argument list to the destination argument list.</p>
3859 macro available in C. In a target-dependent way, it copies the source
3862 llvm.va_start</a></tt> intrinsic may be arbitrarily complex and require, for
3863 example, memory allocation.</p>
3865 </div>
3867 <!-- ======================================================================= -->
3870 </div>
3874 <p>
3876 Collection</a> requires the implementation and generation of these intrinsics.
3880 Front-ends for type-safe garbage collected languages should generate these
3881 intrinsics to make use of the LLVM garbage collectors. For more details, see <a
3883 </p>
3884 </div>
3886 <!-- _______________________________________________________________________ -->
3889 </div>
3895 <pre>
3897 </pre>
3902 the code generator, and allows some metadata to be associated with it.</p>
3906 <p>The first argument specifies the address of a stack object that contains the
3907 root pointer. The second pointer (which must be either a constant or a global
3908 value address) contains the meta-data to be associated with the root.</p>
3913 location. At compile-time, the code generator generates information to allow
3914 the runtime to find the pointer at GC safe points.
3915 </p>
3917 </div>
3920 <!-- _______________________________________________________________________ -->
3923 </div>
3929 <pre>
3930 declare i8 * @llvm.gcread(i8 * %ObjPtr, i8 ** %Ptr)
3931 </pre>
3936 locations, allowing garbage collector implementations that require read
3937 barriers.</p>
3941 <p>The second argument is the address to read from, which should be an address
3942 allocated from the garbage collector. The first object is a pointer to the
3943 start of the referenced object, if needed by the language runtime (otherwise
3944 null).</p>
3949 instruction, but may be replaced with substantially more complex code by the
3950 garbage collector runtime, as needed.</p>
3952 </div>
3955 <!-- _______________________________________________________________________ -->
3958 </div>
3964 <pre>
3965 declare void @llvm.gcwrite(i8 * %P1, i8 * %Obj, i8 ** %P2)
3966 </pre>
3971 locations, allowing garbage collector implementations that require write
3972 barriers (such as generational or reference counting collectors).</p>
3976 <p>The first argument is the reference to store, the second is the start of the
3977 object to store it to, and the third is the address of the field of Obj to
3978 store to. If the runtime does not require a pointer to the object, Obj may be
3979 null.</p>
3984 instruction, but may be replaced with substantially more complex code by the
3985 garbage collector runtime, as needed.</p>
3987 </div>
3991 <!-- ======================================================================= -->
3994 </div>
3997 <p>
3998 These intrinsics are provided by LLVM to expose special features that may only
3999 be implemented with code generator support.
4000 </p>
4002 </div>
4004 <!-- _______________________________________________________________________ -->
4007 </div>
4012 <pre>
4014 </pre>
4018 <p>
4020 target-specific value indicating the return address of the current function
4021 or one of its callers.
4022 </p>
4026 <p>
4027 The argument to this intrinsic indicates which function to return the address
4028 for. Zero indicates the calling function, one indicates its caller, etc. The
4030 </p>
4034 <p>
4036 the return address of the specified call frame, or zero if it cannot be
4037 identified. The value returned by this intrinsic is likely to be incorrect or 0
4038 for arguments other than zero, so it should only be used for debugging purposes.
4039 </p>
4041 <p>
4042 Note that calling this intrinsic does not prevent function inlining or other
4043 aggressive transformations, so the value returned may not be that of the obvious
4044 source-language caller.
4045 </p>
4046 </div>
4049 <!-- _______________________________________________________________________ -->
4052 </div>
4057 <pre>
4059 </pre>
4063 <p>
4065 target-specific frame pointer value for the specified stack frame.
4066 </p>
4070 <p>
4071 The argument to this intrinsic indicates which function to return the frame
4072 pointer for. Zero indicates the calling function, one indicates its caller,
4074 </p>
4078 <p>
4080 the frame address of the specified call frame, or zero if it cannot be
4081 identified. The value returned by this intrinsic is likely to be incorrect or 0
4082 for arguments other than zero, so it should only be used for debugging purposes.
4083 </p>
4085 <p>
4086 Note that calling this intrinsic does not prevent function inlining or other
4087 aggressive transformations, so the value returned may not be that of the obvious
4088 source-language caller.
4089 </p>
4090 </div>
4092 <!-- _______________________________________________________________________ -->
4095 </div>
4100 <pre>
4101 declare i8 *@llvm.stacksave()
4102 </pre>
4106 <p>
4110 features like scoped automatic variable sized arrays in C99.
4111 </p>
4115 <p>
4116 This intrinsic returns a opaque pointer value that can be passed to <a
4123 </p>
4125 </div>
4127 <!-- _______________________________________________________________________ -->
4130 </div>
4135 <pre>
4136 declare void @llvm.stackrestore(i8 * %ptr)
4137 </pre>
4141 <p>
4143 the function stack to the state it was in when the corresponding <a
4145 useful for implementing language features like scoped automatic variable sized
4146 arrays in C99.
4147 </p>
4151 <p>
4153 </p>
4155 </div>
4158 <!-- _______________________________________________________________________ -->
4161 </div>
4166 <pre>
4169 </pre>
4174 <p>
4176 a prefetch instruction if supported; otherwise, it is a noop. Prefetches have
4177 no
4178 effect on the behavior of the program but can change its performance
4179 characteristics.
4180 </p>
4184 <p>
4190 </p>
4194 <p>
4195 This intrinsic does not modify the behavior of the program. In particular,
4196 prefetches cannot trap and do not produce a value. On targets that support this
4197 intrinsic, the prefetch can provide hints to the processor cache for better
4198 performance.
4199 </p>
4201 </div>
4203 <!-- _______________________________________________________________________ -->
4206 </div>
4211 <pre>
4213 </pre>
4218 <p>
4220 (PC) in a region of
4221 code to simulators and other tools. The method is target specific, but it is
4222 expected that the marker will use exported symbols to transmit the PC of the marker.
4223 The marker makes no guarantees that it will remain with any specific instruction
4224 after optimizations. It is possible that the presence of a marker will inhibit
4225 optimizations. The intended use is to be inserted after optimizations to allow
4226 correlations of simulation runs.
4227 </p>
4231 <p>
4233 </p>
4237 <p>
4238 This intrinsic does not modify the behavior of the program. Backends that do not
4239 support this intrinisic may ignore it.
4240 </p>
4242 </div>
4244 <!-- _______________________________________________________________________ -->
4247 </div>
4252 <pre>
4253 declare i64 @llvm.readcyclecounter( )
4254 </pre>
4259 <p>
4261 counter register (or similar low latency, high accuracy clocks) on those targets
4262 that support it. On X86, it should map to RDTSC. On Alpha, it should map to RPCC.
4263 As the backing counters overflow quickly (on the order of 9 seconds on alpha), this
4264 should only be used for small timings.
4265 </p>
4269 <p>
4270 When directly supported, reading the cycle counter should not modify any memory.
4271 Implementations are allowed to either return a application specific value or a
4272 system wide value. On backends without support, this is lowered to a constant 0.
4273 </p>
4275 </div>
4277 <!-- ======================================================================= -->
4280 </div>
4283 <p>
4284 LLVM provides intrinsics for a few important standard C library functions.
4285 These intrinsics allow source-language front-ends to pass information about the
4286 alignment of the pointer arguments to the code generator, providing opportunity
4287 for more efficient code generation.
4288 </p>
4290 </div>
4292 <!-- _______________________________________________________________________ -->
4295 </div>
4300 <pre>
4305 </pre>
4309 <p>
4311 location to the destination location.
4312 </p>
4314 <p>
4316 intrinsics do not return a value, and takes an extra alignment argument.
4317 </p>
4321 <p>
4322 The first argument is a pointer to the destination, the second is a pointer to
4323 the source. The third argument is an integer argument
4324 specifying the number of bytes to copy, and the fourth argument is the alignment
4325 of the source and destination locations.
4326 </p>
4328 <p>
4330 the caller guarantees that both the source and destination pointers are aligned
4331 to that boundary.
4332 </p>
4336 <p>
4338 location to the destination location, which are not allowed to overlap. It
4339 copies "len" bytes of memory over. If the argument is known to be aligned to
4340 some boundary, this can be specified as the fourth argument, otherwise it should
4342 </p>
4343 </div>
4346 <!-- _______________________________________________________________________ -->
4349 </div>
4354 <pre>
4359 </pre>
4363 <p>
4365 location to the destination location. It is similar to the
4367 </p>
4369 <p>
4371 intrinsics do not return a value, and takes an extra alignment argument.
4372 </p>
4376 <p>
4377 The first argument is a pointer to the destination, the second is a pointer to
4378 the source. The third argument is an integer argument
4379 specifying the number of bytes to copy, and the fourth argument is the alignment
4380 of the source and destination locations.
4381 </p>
4383 <p>
4385 the caller guarantees that the source and destination pointers are aligned to
4386 that boundary.
4387 </p>
4391 <p>
4393 location to the destination location, which may overlap. It
4394 copies "len" bytes of memory over. If the argument is known to be aligned to
4395 some boundary, this can be specified as the fourth argument, otherwise it should
4397 </p>
4398 </div>
4401 <!-- _______________________________________________________________________ -->
4404 </div>
4409 <pre>
4414 </pre>
4418 <p>
4420 byte value.
4421 </p>
4423 <p>
4425 does not return a value, and takes an extra alignment argument.
4426 </p>
4430 <p>
4431 The first argument is a pointer to the destination to fill, the second is the
4432 byte value to fill it with, the third argument is an integer
4433 argument specifying the number of bytes to fill, and the fourth argument is the
4434 known alignment of destination location.
4435 </p>
4437 <p>
4439 the caller guarantees that the destination pointer is aligned to that boundary.
4440 </p>
4444 <p>
4446 the
4447 destination location. If the argument is known to be aligned to some boundary,
4448 this can be specified as the fourth argument, otherwise it should be set to 0 or
4449 1.
4450 </p>
4451 </div>
4454 <!-- _______________________________________________________________________ -->
4457 </div>
4462 <pre>
4463 declare float @llvm.sqrt.f32(float %Val)
4464 declare double @llvm.sqrt.f64(double %Val)
4465 </pre>
4469 <p>
4473 negative numbers (which allows for better optimization).
4474 </p>
4478 <p>
4479 The argument and return value are floating point numbers of the same type.
4480 </p>
4484 <p>
4485 This function returns the sqrt of the specified operand if it is a positive
4486 floating point number.
4487 </p>
4488 </div>
4490 <!-- _______________________________________________________________________ -->
4493 </div>
4498 <pre>
4499 declare float @llvm.powi.f32(float %Val, i32 %power)
4500 declare double @llvm.powi.f64(double %Val, i32 %power)
4501 </pre>
4505 <p>
4507 specified (positive or negative) power. The order of evaluation of
4508 multiplications is not defined.
4509 </p>
4513 <p>
4514 The second argument is an integer power, and the first is a value to raise to
4515 that power.
4516 </p>
4520 <p>
4521 This function returns the first value raised to the second power with an
4522 unspecified sequence of rounding operations.</p>
4523 </div>
4526 <!-- ======================================================================= -->
4529 </div>
4532 <p>
4533 LLVM provides intrinsics for a few important bit manipulation operations.
4534 These allow efficient code generation for some algorithms.
4535 </p>
4537 </div>
4539 <!-- _______________________________________________________________________ -->
4542 </div>
4547 <p>This is an overloaded intrinsic function. You can use bswap on any integer
4549 that includes the type for the result and the operand.
4550 <pre>
4554 </pre>
4558 <p>
4560 values with an even number of bytes (positive multiple of 16 bits). These are
4561 useful for performing operations on data that is not in the target's native
4562 byte order.
4563 </p>
4567 <p>
4570 intrinsic returns an i32 value that has the four bytes of the input i32
4575 </p>
4577 </div>
4579 <!-- _______________________________________________________________________ -->
4582 </div>
4587 <p>This is an overloaded intrinsic. You can use llvm.ctpop on any integer bit
4588 width. Not all targets support all bit widths however.
4589 <pre>
4595 </pre>
4599 <p>
4601 value.
4602 </p>
4606 <p>
4607 The only argument is the value to be counted. The argument may be of any
4608 integer type. The return type must match the argument type.
4609 </p>
4613 <p>
4615 </p>
4616 </div>
4618 <!-- _______________________________________________________________________ -->
4621 </div>
4627 integer bit width. Not all targets support all bit widths however.
4628 <pre>
4634 </pre>
4638 <p>
4640 leading zeros in a variable.
4641 </p>
4645 <p>
4646 The only argument is the value to be counted. The argument may be of any
4647 integer type. The return type must match the argument type.
4648 </p>
4652 <p>
4654 in a variable. If the src == 0 then the result is the size in bits of the type
4656 </p>
4657 </div>
4661 <!-- _______________________________________________________________________ -->
4664 </div>
4670 integer bit width. Not all targets support all bit widths however.
4671 <pre>
4677 </pre>
4681 <p>
4683 trailing zeros.
4684 </p>
4688 <p>
4689 The only argument is the value to be counted. The argument may be of any
4690 integer type. The return type must match the argument type.
4691 </p>
4695 <p>
4697 in a variable. If the src == 0 then the result is the size in bits of the type
4699 </p>
4700 </div>
4702 <!-- _______________________________________________________________________ -->
4705 </div>
4711 on any integer bit width.
4712 <pre>
4713 declare i17 @llvm.part.select.i17.i17 (i17 %val, i32 %loBit, i32 %hiBit)
4714 declare i29 @llvm.part.select.i29.i29 (i29 %val, i32 %loBit, i32 %hiBit)
4715 </pre>
4719 range of bits from an integer value and returns them in the same bit width as
4720 the original value.</p>
4724 any bit width but they must have the same bit width. The second and third
4731 operates in forward mode.</p>
4735 <ol>
4739 to determine the number of bits to retain.</li>
4742 </ol>
4743 <p>In reverse mode, a similar computation is made except that the bits are
4748 </div>
4752 </div>
4758 on any integer bit width.
4759 <pre>
4760 declare i17 @llvm.part.set.i17.i17.i9 (i17 %val, i9 %repl, i32 %lo, i32 %hi)
4761 declare i29 @llvm.part.set.i29.i29.i9 (i29 %val, i9 %repl, i32 %lo, i32 %hi)
4762 </pre>
4766 of bits in an integer value with another integer value. It returns the integer
4767 with the replaced bits.</p>
4774 type since they specify only a bit index.</p>
4780 operates in forward mode.</p>
4782 truncating it down to the size of the replacement area or zero extending it
4783 up to that size.</p>
4788 <p>In reverse mode, a similar computation is made except that the bits are
4792 <pre>
4798 </pre>
4799 </div>
4801 <!-- ======================================================================= -->
4804 </div>
4807 <p>
4809 are described in the <a
4811 Debugging</a> document.
4812 </p>
4813 </div>
4816 <!-- ======================================================================= -->
4819 </div>
4822 <p> The LLVM exception handling intrinsics (which all start with
4826 </div>
4828 <!-- ======================================================================= -->
4831 </div>
4834 <p> This class of intrinsics is designed to be generic and has
4835 no specific purpose. </p>
4836 </div>
4838 <!-- _______________________________________________________________________ -->
4841 </div>
4846 <pre>
4847 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int> )
4848 </pre>
4852 <p>
4854 </p>
4858 <p>
4859 The first argument is a pointer to a value, the second is a pointer to a
4860 global string, the third is a pointer to a global string which is the source
4861 file name, and the last argument is the line number.
4862 </p>
4866 <p>
4867 This intrinsic allows annotation of local variables with arbitrary strings.
4868 This can be useful for special purpose optimizations that want to look for these
4869 annotations. These have no other defined use, they are ignored by code
4870 generation and optimization.
4871 </div>
4874 <!-- *********************************************************************** -->
4875 <hr>
4876 <address>
4884 Last modified: $Date$
4885 </address>
4886 </body>
4887 </html>