| blob | 83ea026c6b7a472ba5fdb6daa84795eb3a593184 |
3 <html>
4 <head>
11 </head>
13 <body>
16 <ol>
21 <ol>
34 </ol>
35 </li>
37 <ol>
40 <ol>
44 </ol>
45 </li>
47 <ol>
56 </ol>
57 </li>
59 </ol>
60 </li>
62 <ol>
68 </ol>
69 </li>
71 <ol>
73 </ol>
74 </li>
76 <ol>
78 <ol>
85 </ol>
86 </li>
88 <ol>
98 </ol>
99 </li>
101 <ol>
108 </ol>
109 </li>
111 <ol>
115 </ol>
116 </li>
118 <ol>
121 </ol>
122 </li>
124 <ol>
131 </ol>
132 </li>
134 <ol>
147 </ol>
148 </li>
150 <ol>
159 </ol>
160 </li>
161 </ol>
162 </li>
164 <ol>
166 <ol>
170 </ol>
171 </li>
173 <ol>
177 </ol>
178 </li>
180 <ol>
188 </ol>
189 </li>
191 <ol>
200 </ol>
201 </li>
203 <ol>
210 </ol>
211 </li>
213 <ol>
220 </ol>
221 </li>
225 <ol>
227 </ol>
228 </li>
230 <ol>
244 </ol>
245 </li>
247 <ol>
256 </ol>
257 </li>
258 </ol>
259 </li>
260 </ol>
265 </div>
267 <!-- *********************************************************************** -->
269 <!-- *********************************************************************** -->
272 <p>This document is a reference manual for the LLVM assembly language.
273 LLVM is a Static Single Assignment (SSA) based representation that provides
274 type safety, low-level operations, flexibility, and the capability of
275 representing 'all' high-level languages cleanly. It is the common code
276 representation used throughout all phases of the LLVM compilation
277 strategy.</p>
278 </div>
280 <!-- *********************************************************************** -->
282 <!-- *********************************************************************** -->
286 <p>The LLVM code representation is designed to be used in three
287 different forms: as an in-memory compiler IR, as an on-disk bitcode
288 representation (suitable for fast loading by a Just-In-Time compiler),
289 and as a human readable assembly language representation. This allows
290 LLVM to provide a powerful intermediate representation for efficient
291 compiler transformations and analysis, while providing a natural means
292 to debug and visualize the transformations. The three different forms
293 of LLVM are all equivalent. This document describes the human readable
294 representation and notation.</p>
296 <p>The LLVM representation aims to be light-weight and low-level
297 while being expressive, typed, and extensible at the same time. It
298 aims to be a "universal IR" of sorts, by being at a low enough level
299 that high-level ideas may be cleanly mapped to it (similar to how
300 microprocessors are "universal IR's", allowing many source languages to
301 be mapped to them). By providing type information, LLVM can be used as
302 the target of optimizations: for example, through pointer analysis, it
303 can be proven that a C automatic variable is never accessed outside of
304 the current function... allowing it to be promoted to a simple SSA
305 value instead of a memory location.</p>
307 </div>
309 <!-- _______________________________________________________________________ -->
314 <p>It is important to note that this document describes 'well formed'
315 LLVM assembly language. There is a difference between what the parser
316 accepts and what is considered 'well formed'. For example, the
317 following instruction is syntactically okay, but not well formed:</p>
320 <pre>
322 </pre>
323 </div>
326 its uses. The LLVM infrastructure provides a verification pass that may
327 be used to verify that an LLVM module is well formed. This pass is
328 automatically run by the parser after parsing input assembly and by
329 the optimizer before it outputs bitcode. The violations pointed out
330 by the verifier pass indicate bugs in transformation passes or input to
331 the parser.</p>
332 </div>
334 <!-- Describe the typesetting conventions here. -->
336 <!-- *********************************************************************** -->
338 <!-- *********************************************************************** -->
342 <p>LLVM identifiers come in two basic types: global and local. Global
343 identifiers (functions, global variables) begin with the @ character. Local
344 identifiers (register names, types) begin with the % character. Additionally,
345 there are three different formats for identifiers, for different purposes:</p>
347 <ol>
348 <li>Named values are represented as a string of characters with their prefix.
349 For example, %foo, @DivisionByZero, %a.really.long.identifier. The actual
351 Identifiers which require other characters in their names can be surrounded
352 with quotes. Special characters may be escaped using "\xx" where xx is the
353 ASCII code for the character in hexadecimal. In this way, any character can
354 be used in a name value, even quotes themselves.
356 <li>Unnamed values are represented as an unsigned numeric value with their
361 </ol>
363 <p>LLVM requires that values start with a prefix for two reasons: Compilers
364 don't need to worry about name clashes with reserved words, and the set of
365 reserved words may be expanded in the future without penalty. Additionally,
366 unnamed identifiers allow a compiler to quickly come up with a temporary
367 variable without having to avoid symbol table conflicts.</p>
369 <p>Reserved words in LLVM are very similar to reserved words in other
370 languages. There are keywords for different opcodes
375 and others. These reserved words cannot conflict with variable names, because
376 none of them start with a prefix character ('%' or '@').</p>
378 <p>Here is an example of LLVM code to multiply the integer variable
384 <pre>
386 </pre>
387 </div>
392 <pre>
394 </pre>
395 </div>
400 <pre>
404 </pre>
405 </div>
408 important lexical features of LLVM:</p>
410 <ol>
413 line.</li>
415 <li>Unnamed temporaries are created when the result of a computation is not
416 assigned to a named value.</li>
420 </ol>
422 <p>...and it also shows a convention that we follow in this document. When
423 demonstrating instructions, we will follow an instruction with a comment that
424 defines the type and name of value produced. Comments are shown in italic
425 text.</p>
427 </div>
429 <!-- *********************************************************************** -->
431 <!-- *********************************************************************** -->
433 <!-- ======================================================================= -->
435 </div>
440 translation unit of the input programs. Each module consists of
441 functions, global variables, and symbol table entries. Modules may be
442 combined together with the LLVM linker, which merges function (and
443 global variable) definitions, resolves forward declarations, and merges
449 href="#globalvars">constant</a> <a href="#t_array">[13 x i8]</a> c"hello world\0A\00" <i>; [13 x i8]*</i>
457 %cast210 = <a
461 <a
463 <a
465 </pre>
466 </div>
473 <p>In general, a module is made up of a list of global values,
474 where both functions and global variables are global values. Global values are
475 represented by a pointer to a memory location (in this case, a pointer to an
476 array of char, and a pointer to a function), and have one of the following <a
479 </div>
481 <!-- ======================================================================= -->
484 </div>
488 <p>
489 All Global Variables and Functions have one of the following types of linkage:
490 </p>
492 <dl>
496 <dd>Global values with private linkage are only directly accessible by
497 objects in the current module. In particular, linking code into a module with
498 an private global value may cause the private to be renamed as necessary to
499 avoid collisions. Because the symbol is private to the module, all
500 references can be updated. This doesn't show up in any symbol table in the
501 object file.
502 </dd>
506 <dd> Similar to private, but the value shows as a local symbol (STB_LOCAL in
507 the case of ELF) in the object file. This corresponds to the notion of the
509 </dd>
514 the same name when linkage occurs. This is typically used to implement
515 inline functions, templates, or other code which must be generated in each
517 allowed to be discarded.
518 </dd>
524 discarded. This is used for globals that may be emitted in multiple
525 translation units, but that are not guaranteed to be emitted into every
526 translation unit that uses them. One example of this is tentative
527 definitions in C, such as "<tt>int X;</tt>" at global scope.
528 </dd>
533 that some targets may choose to emit different assembly sequences for them
534 for target-dependent reasons. This is used for globals that are declared
535 "weak" in C source code.
536 </dd>
541 pointer to array type. When two global variables with appending linkage are
542 linked together, the two global arrays are appended together. This is the
543 LLVM, typesafe, equivalent of having the system linker append together
544 "sections" with identical names when .o files are linked.
545 </dd>
548 <dd>The semantics of this linkage follow the ELF object file model: the
549 symbol is weak until linked, if not linked, the symbol becomes null instead
550 of being an undefined reference.
551 </dd>
555 <dd>If none of the above identifiers are used, the global is externally
556 visible, meaning that it participates in linkage and can be used to resolve
557 external symbol references.
558 </dd>
559 </dl>
561 <p>
562 The next two types of linkage are targeted for Microsoft Windows platform
563 only. They are designed to support importing (exporting) symbols from (to)
564 DLLs (Dynamic Link Libraries).
565 </p>
567 <dl>
571 or variable via a global pointer to a pointer that is set up by the DLL
572 exporting the symbol. On Microsoft Windows targets, the pointer name is
574 </dd>
579 pointer to a pointer in a DLL, so that it can be referenced with the
582 name.
583 </dd>
585 </dl>
588 variable is defined to be internal, if another module defined a "<tt>.LC0</tt>"
589 variable and was linked with this one, one of the two would be renamed,
591 external (i.e., lacking any linkage declarations), they are accessible
592 outside of the current module.</p>
597 linkages.</p>
598 </div>
600 <!-- ======================================================================= -->
603 </div>
609 specified for the call. The calling convention of any pair of dynamic
610 caller/callee must match, or the behavior of the program is undefined. The
611 following calling conventions are supported by LLVM, and more may be added in
612 the future:</p>
614 <dl>
617 <dd>This calling convention (the default if no other calling convention is
618 specified) matches the target C calling conventions. This calling convention
619 supports varargs function calls and tolerates some mismatch in the declared
620 prototype and implemented declaration of the function (as does normal C).
621 </dd>
625 <dd>This calling convention attempts to make calls as fast as possible
626 (e.g. by passing things in registers). This calling convention allows the
627 target to use whatever tricks it wants to produce fast code for the target,
628 without having to conform to an externally specified ABI (Application Binary
629 Interface). Implementations of this convention should allow arbitrary
631 supported. This calling convention does not support varargs and requires the
632 prototype of all callees to exactly match the prototype of the function
633 definition.
634 </dd>
638 <dd>This calling convention attempts to make code in the caller as efficient
639 as possible under the assumption that the call is not commonly executed. As
640 such, these calls often preserve all registers so that the call does not break
641 any live ranges in the caller side. This calling convention does not support
642 varargs and requires the prototype of all callees to exactly match the
643 prototype of the function definition.
644 </dd>
648 <dd>Any calling convention may be specified by number, allowing
649 target-specific calling conventions to be used. Target specific calling
650 conventions start at 64.
651 </dd>
652 </dl>
654 <p>More calling conventions can be added/defined on an as-needed basis, to
655 support pascal conventions or any other well-known target-independent
656 convention.</p>
658 </div>
660 <!-- ======================================================================= -->
663 </div>
667 <p>
668 All Global Variables and Functions have one of the following visibility styles:
669 </p>
671 <dl>
674 <dd>On targets that use the ELF object file format, default visibility means
675 that the declaration is visible to other
676 modules and, in shared libraries, means that the declared entity may be
677 overridden. On Darwin, default visibility means that the declaration is
678 visible to other modules. Default visibility corresponds to "external
679 linkage" in the language.
680 </dd>
684 <dd>Two declarations of an object with hidden visibility refer to the same
685 object if they are in the same shared object. Usually, hidden visibility
686 indicates that the symbol will not be placed into the dynamic symbol table,
687 so no other module (executable or shared library) can reference it
688 directly.
689 </dd>
693 <dd>On ELF, protected visibility indicates that the symbol will be placed in
694 the dynamic symbol table, but that references within the defining module will
695 bind to the local symbol. That is, the symbol cannot be overridden by another
696 module.
697 </dd>
698 </dl>
700 </div>
702 <!-- ======================================================================= -->
705 </div>
709 <p>LLVM IR allows you to specify name aliases for certain types. This can make
710 it easier to read the IR and make the IR more condensed (particularly when
711 recursive types are involved). An example of a name specification is:
712 </p>
715 <pre>
716 %mytype = type { %mytype*, i32 }
717 </pre>
718 </div>
724 <p>Note that type names are aliases for the structural type that they indicate,
725 and that you can therefore specify multiple names for the same type. This often
726 leads to confusing behavior when dumping out a .ll file. Since LLVM IR uses
727 structural typing, the name is not part of the type. When printing out LLVM IR,
729 shape. This means that if you have code where two different source types end up
730 having the same LLVM type, that the dumper will sometimes print the "wrong" or
731 unexpected type. This is an important design point and isn't going to
732 change.</p>
734 </div>
736 <!-- ======================================================================= -->
739 </div>
743 <p>Global variables define regions of memory allocated at compilation time
744 instead of run-time. Global variables may optionally be initialized, may have
745 an explicit section to be placed in, and may have an optional explicit alignment
746 specified. A variable may be defined as "thread_local", which means that it
747 will not be shared by threads (each thread will have a separated copy of the
748 variable). A variable may be defined as a global "constant," which indicates
750 optimization, allowing the global data to be placed in the read-only section of
751 an executable, etc). Note that variables that need runtime initialization
754 <p>
756 constant, even if the final definition of the global is not. This capability
757 can be used to enable slightly better optimization of the program, but requires
758 the language definition to guarantee that optimizations based on the
759 'constantness' are valid for the translation units that do not include the
760 definition.
761 </p>
763 <p>As SSA values, global variables define pointer values that are in
764 scope (i.e. they dominate) all basic blocks in the program. Global
765 variables always define a pointer to their "content" type because they
766 describe a region of memory, and all memory objects in LLVM are
767 accessed through pointers.</p>
769 <p>A global variable may be declared to reside in a target-specifc numbered
770 address space. For targets that support them, address spaces may affect how
771 optimizations are performed and/or what target instructions are used to access
772 the variable. The default address space is zero. The address space qualifier
773 must precede any other attributes.</p>
775 <p>LLVM allows an explicit section to be specified for globals. If the target
776 supports it, it will emit globals to the section specified.</p>
778 <p>An explicit alignment may be specified for a global. If not present, or if
779 the alignment is set to zero, the alignment of the global is set by the target
780 to whatever it feels convenient. If an explicit alignment is specified, the
781 global is forced to have at least that much alignment. All alignments must be
784 <p>For example, the following defines a global in a numbered address space with
785 an initializer, section, and alignment:</p>
788 <pre>
790 </pre>
791 </div>
793 </div>
796 <!-- ======================================================================= -->
799 </div>
808 name, a (possibly empty) argument list (each with optional
812 an opening curly brace, a list of basic blocks, and a closing curly brace.
814 LLVM function declarations consist of the "<tt>declare</tt>" keyword, an
819 name, a possibly empty list of arguments, an optional alignment, and an optional
822 <p>A function definition contains a list of basic blocks, forming the CFG
823 (Control Flow Graph) for
824 the function. Each basic block may optionally start with a label (giving the
825 basic block a symbol table entry), contains a list of instructions, and ends
827 function return).</p>
829 <p>The first basic block in a function is special in two ways: it is immediately
830 executed on entrance to the function, and it is not allowed to have predecessor
831 basic blocks (i.e. there can not be any branches to the entry block of a
832 function). Because the block can have no predecessors, it also cannot have any
835 <p>LLVM allows an explicit section to be specified for functions. If the target
836 supports it, it will emit functions to the section specified.</p>
838 <p>An explicit alignment may be specified for a function. If not present, or if
839 the alignment is set to zero, the alignment of the function is set by the target
840 to whatever it feels convenient. If an explicit alignment is specified, the
841 function is forced to have at least that much alignment. All alignments must be
847 <tt>
853 </tt>
854 </div>
856 </div>
859 <!-- ======================================================================= -->
862 </div>
865 function, global variable, another alias or bitcast of global value). Aliases
872 <pre>
874 </pre>
875 </div>
877 </div>
881 <!-- ======================================================================= -->
884 <p>The return type and each parameter of a function type may have a set of
886 used to communicate additional information about the result or parameters of
887 a function. Parameter attributes are considered to be part of the function,
888 not of the function type, so functions with different parameter attributes
889 can have the same function type.</p>
891 <p>Parameter attributes are simple keywords that follow the type specified. If
892 multiple parameter attributes are needed, they are space separated. For
893 example:</p>
896 <pre>
897 declare i32 @printf(i8* noalias nocapture, ...)
898 declare i32 @atoi(i8 zeroext)
899 declare signext i8 @returns_signed_char()
900 </pre>
901 </div>
907 <dl>
909 <dd>This indicates to the code generator that the parameter or return value
910 should be zero-extended to a 32-bit value by the caller (for a parameter)
911 or the callee (for a return value).</dd>
914 <dd>This indicates to the code generator that the parameter or return value
915 should be sign-extended to a 32-bit value by the caller (for a parameter)
916 or the callee (for a return value).</dd>
919 <dd>This indicates that this parameter or return value should be treated
920 in a special target-dependent fashion during while emitting code for a
921 function call or return (usually, by putting it in a register as opposed
922 to memory, though some targets use it to distinguish between two different
923 kinds of registers). Use of this attribute is target-specific.</dd>
926 <dd>This indicates that the pointer parameter should really be passed by
927 value to the function. The attribute implies that a hidden copy of the
928 pointee is made between the caller and the callee, so the callee is unable
929 to modify the value in the callee. This attribute is only valid on LLVM
930 pointer arguments. It is generally used to pass structs and arrays by
931 value, but is also valid on pointers to scalars. The copy is considered to
932 belong to the caller not the callee (for example,
935 values. The byval attribute also supports specifying an alignment with the
936 align attribute. This has a target-specific effect on the code generator
937 that usually indicates a desired alignment for the synthesized stack
938 slot.</dd>
941 <dd>This indicates that the pointer parameter specifies the address of a
942 structure that is the return value of the function in the source program.
943 This pointer must be guaranteed by the caller to be valid: loads and stores
944 to the structure may be assumed by the callee to not to trap. This may only
945 be applied to the first parameter. This is not a valid attribute for
946 return values. </dd>
949 <dd>This indicates that the pointer does not alias any global or any other
950 parameter. The caller is responsible for ensuring that this is the
952 that the pointer does not alias any other pointers visible to the
953 caller. For further details, please see the discussion of the NoAlias
954 response in
959 <dd>This indicates that the callee does not make any copies of the pointer
960 that outlive the callee itself. This is not a valid attribute for return
961 values.</dd>
964 <dd>This indicates that the pointer parameter can be excised using the
966 attribute for return values.</dd>
967 </dl>
969 </div>
971 <!-- ======================================================================= -->
974 </div>
977 <p>Each function may specify a garbage collector name, which is simply a
978 string.</p>
984 collector which will cause the compiler to alter its output in order to support
985 the named garbage collection algorithm.</p>
986 </div>
988 <!-- ======================================================================= -->
991 </div>
995 <p>Function attributes are set to communicate additional information about
996 a function. Function attributes are considered to be part of the function,
997 not of the function type, so functions with different parameter attributes
998 can have the same function type.</p>
1000 <p>Function attributes are simple keywords that follow the type specified. If
1001 multiple attributes are needed, they are space separated. For
1002 example:</p>
1005 <pre>
1006 define void @f() noinline { ... }
1007 define void @f() alwaysinline { ... }
1008 define void @f() alwaysinline optsize { ... }
1009 define void @f() optsize
1010 </pre>
1011 </div>
1013 <dl>
1015 <dd>This attribute indicates that the inliner should attempt to inline this
1016 function into callers whenever possible, ignoring any active inlining size
1017 threshold for this caller.</dd>
1020 <dd>This attribute indicates that the inliner should never inline this function
1021 in any situation. This attribute may not be used together with the
1025 <dd>This attribute suggests that optimization passes and code generator passes
1026 make choices that keep the code size of this function low, and otherwise do
1027 optimizations specifically to reduce code size.</dd>
1030 <dd>This function attribute indicates that the function never returns normally.
1031 This produces undefined behavior at runtime if the function ever does
1032 dynamically return.</dd>
1035 <dd>This function attribute indicates that the function never returns with an
1036 unwind or exceptional control flow. If the function does unwind, its runtime
1037 behavior is undefined.</dd>
1040 <dd>This attribute indicates that the function computes its result (or the
1041 exception it throws) based strictly on its arguments, without dereferencing any
1042 pointer arguments or otherwise accessing any mutable state (e.g. memory, control
1043 registers, etc) visible to caller functions. It does not write through any
1045 never changes any state visible to callers.</dd>
1048 <dd>This attribute indicates that the function does not write through any
1050 or otherwise modify any state (e.g. memory, control registers, etc) visible to
1051 caller functions. It may dereference pointer arguments and read state that may
1052 be set in the caller. A readonly function always returns the same value (or
1053 throws the same exception) when called with the same set of arguments and global
1054 state.</dd>
1057 <dd>This attribute indicates that the function should emit a stack smashing
1059 stack before the local variables that's checked upon return from the function to
1060 see if it has been overwritten. A heuristic is used to determine if a function
1061 needs stack protectors or not.
1070 function attribute.
1076 </dl>
1078 </div>
1080 <!-- ======================================================================= -->
1083 </div>
1086 <p>
1087 Modules may contain "module-level inline asm" blocks, which corresponds to the
1088 GCC "file scope inline asm" blocks. These blocks are internally concatenated by
1089 LLVM and treated as a single unit, but may be separated in the .ll file if
1090 desired. The syntax is very simple:
1091 </p>
1094 <pre>
1095 module asm "inline asm code goes here"
1096 module asm "more can go here"
1097 </pre>
1098 </div>
1100 <p>The strings can contain any character by escaping non-printable characters.
1102 for the number.
1103 </p>
1105 <p>
1106 The inline asm code is simply printed to the machine code .s file when
1107 assembly code is generated.
1108 </p>
1109 </div>
1111 <!-- ======================================================================= -->
1114 </div>
1117 <p>A module may specify a target specific data layout string that specifies how
1118 data is to be laid out in memory. The syntax for the data layout is simply:</p>
1121 separated by the minus sign character ('-'). Each specification starts with a
1122 letter and may include other information after the letter to define some
1123 aspect of the data layout. The specifications accepted are as follows: </p>
1124 <dl>
1126 <dd>Specifies that the target lays out data in big-endian form. That is, the
1127 bits with the most significance have the lowest address location.</dd>
1129 <dd>Specifies that the target lays out data in little-endian form. That is,
1130 the bits with the least significance have the lowest address location.</dd>
1135 too.</dd>
1137 <dd>This specifies the alignment for an integer type of a given bit
1140 <dd>This specifies the alignment for a vector type of a given bit
1143 <dd>This specifies the alignment for a floating point type of a given bit
1145 (double).</dd>
1147 <dd>This specifies the alignment for an aggregate type of a given bit
1149 </dl>
1150 <p>When constructing the data layout for a given target, LLVM starts with a
1151 default set of specifications which are then (possibly) overriden by the
1153 are given in this list:</p>
1154 <ul>
1168 </ul>
1169 <p>When LLVM is determining the alignment for a given type, it uses the
1170 following rules:</p>
1171 <ol>
1172 <li>If the type sought is an exact match for one of the specifications, that
1173 specification is used.</li>
1174 <li>If no match is found, and the type sought is an integer type, then the
1175 smallest integer type that is larger than the bitwidth of the sought type is
1176 used. If none of the specifications are larger than the bitwidth then the the
1177 largest integer type is used. For example, given the default specifications
1178 above, the i7 type will use the alignment of i8 (next largest) while both
1179 i65 and i256 will use the alignment of i64 (largest specified).</li>
1180 <li>If no match is found, and the type sought is a vector type, then the
1181 largest vector type that is smaller than the sought vector type will be used
1184 </ol>
1185 </div>
1187 <!-- *********************************************************************** -->
1189 <!-- *********************************************************************** -->
1193 <p>The LLVM type system is one of the most important features of the
1194 intermediate representation. Being typed enables a number of
1195 optimizations to be performed on the intermediate representation directly,
1196 without having to do
1197 extra analyses on the side before the transformation. A strong type
1198 system makes it easier to read the generated code and enables novel
1199 analyses and transformations that are not feasible to perform on normal
1200 three address code representations.</p>
1202 </div>
1204 <!-- ======================================================================= -->
1206 Classifications</a> </div>
1208 <p>The types fall into a few useful
1209 classifications:</p>
1212 <tbody>
1214 <tr>
1217 </tr>
1218 <tr>
1221 </tr>
1222 <tr>
1231 </td>
1232 </tr>
1233 <tr>
1238 </tr>
1239 <tr>
1249 </td>
1250 </tr>
1251 </tbody>
1252 </table>
1255 most important. Values of these types are the only ones which can be
1256 produced by instructions, passed as arguments, or used as operands to
1257 instructions.</p>
1258 </div>
1260 <!-- ======================================================================= -->
1264 <p>The primitive types are the fundamental building blocks of the LLVM
1265 system.</p>
1267 </div>
1269 <!-- _______________________________________________________________________ -->
1273 <table>
1274 <tbody>
1281 </tbody>
1282 </table>
1283 </div>
1285 <!-- _______________________________________________________________________ -->
1294 <pre>
1295 void
1296 </pre>
1297 </div>
1299 <!-- _______________________________________________________________________ -->
1308 <pre>
1309 label
1310 </pre>
1311 </div>
1314 <!-- ======================================================================= -->
1319 <p>The real power in LLVM comes from the derived types in the system.
1320 This is what allows a programmer to represent arrays, functions,
1321 pointers, and other useful types. Note that these derived types may be
1322 recursive: For example, it is possible to have a two dimensional array.</p>
1324 </div>
1326 <!-- _______________________________________________________________________ -->
1332 <p>The integer type is a very simple derived type that simply specifies an
1333 arbitrary bit width for the integer type desired. Any bit width from 1 bit to
1338 <pre>
1339 iN
1340 </pre>
1343 value.</p>
1347 <tbody>
1348 <tr>
1351 </tr><tr>
1354 </tr><tr>
1357 </tr>
1358 </tbody>
1359 </table>
1361 <p>Note that the code generator does not yet support large integer types
1362 to be used as function return types. The specific limit on how large a
1363 return type the code generator can currently handle is target-dependent;
1365 targets.</p>
1367 </div>
1369 <!-- _______________________________________________________________________ -->
1376 <p>The array type is a very simple derived type that arranges elements
1377 sequentially in memory. The array type requires a size (number of
1378 elements) and an underlying data type.</p>
1382 <pre>
1384 </pre>
1386 <p>The number of elements is a constant integer value; elementtype may
1387 be any type with a size.</p>
1394 </tr>
1398 </tr>
1402 </tr>
1403 </table>
1409 </tr>
1413 </tr>
1417 </tr>
1418 </table>
1420 <p>Note that 'variable sized arrays' can be implemented in LLVM with a zero
1421 length array. Normally, accesses past the end of an array are undefined in
1423 As a special case, however, zero length arrays are recognized to be variable
1424 length. This allows implementation of 'pascal style arrays' with the LLVM
1427 <p>Note that the code generator does not yet support large aggregate types
1428 to be used as function return types. The specific limit on how large an
1429 aggregate return type the code generator can currently handle is
1430 target-dependent, and also dependent on the aggregate element types.</p>
1432 </div>
1434 <!-- _______________________________________________________________________ -->
1440 <p>The function type can be thought of as a function signature. It
1441 consists of a return type and a list of formal parameter types. The
1442 return type of a function type is a scalar type, a void type, or a struct type.
1443 If the return type is a struct type then all struct elements must be of first
1444 class types, and the struct must have at least one element.</p>
1448 <pre>
1450 </pre>
1454 which indicates that the function takes a variable number of arguments.
1455 Variable argument functions can access their arguments with the <a
1465 </td>
1468 </tt></td>
1473 </td>
1479 LLVM.
1480 </td>
1485 </td>
1486 </tr>
1487 </table>
1489 </div>
1490 <!-- _______________________________________________________________________ -->
1494 <p>The structure type is used to represent a collection of data members
1495 together in memory. The packing of the field types is defined to match
1496 the ABI of the underlying processor. The elements of a structure may
1497 be any type that has a size.</p>
1501 instruction.</p>
1515 </tr>
1516 </table>
1518 <p>Note that the code generator does not yet support large aggregate types
1519 to be used as function return types. The specific limit on how large an
1520 aggregate return type the code generator can currently handle is
1521 target-dependent, and also dependent on the aggregate element types.</p>
1523 </div>
1525 <!-- _______________________________________________________________________ -->
1527 </div>
1530 <p>The packed structure type is used to represent a collection of data members
1531 together in memory. There is no padding between fields. Further, the alignment
1532 of a packed structure is 1 byte. The elements of a packed structure may
1533 be any type that has a size.</p>
1537 instruction.</p>
1552 </tr>
1553 </table>
1554 </div>
1556 <!-- _______________________________________________________________________ -->
1560 <p>As in many languages, the pointer type represents a pointer or
1561 reference to another object, which must live in memory. Pointer types may have
1562 an optional address space attribute defining the target-specific numbered
1563 address space where the pointed-to object resides. The default address space is
1564 zero.</p>
1577 </tr>
1583 </tr>
1588 </tr>
1589 </table>
1590 </div>
1592 <!-- _______________________________________________________________________ -->
1598 <p>A vector type is a simple derived type that represents a vector
1599 of elements. Vector types are used when multiple primitive data
1600 are operated in parallel using a single instruction (SIMD).
1601 A vector type requires a size (number of
1602 elements) and an underlying primitive data type. Vectors must have a power
1608 <pre>
1610 </pre>
1612 <p>The number of elements is a constant integer value; elementtype may
1613 be any integer or floating point type.</p>
1621 </tr>
1625 </tr>
1629 </tr>
1630 </table>
1632 <p>Note that the code generator does not yet support large vector types
1633 to be used as function return types. The specific limit on how large a
1634 vector return type codegen can currently handle is target-dependent;
1635 currently it's often a few times longer than a hardware vector register.</p>
1637 </div>
1639 <!-- _______________________________________________________________________ -->
1645 <p>Opaque types are used to represent unknown types in the system. This
1646 corresponds (for example) to the C notion of a forward declared structure type.
1647 In LLVM, opaque types can eventually be resolved to any type (not just a
1648 structure type).</p>
1652 <pre>
1653 opaque
1654 </pre>
1662 </tr>
1663 </table>
1664 </div>
1666 <!-- ======================================================================= -->
1669 </div>
1673 <p>
1674 An "up reference" allows you to refer to a lexically enclosing type without
1675 requiring it to have a name. For instance, a structure declaration may contain a
1676 pointer to any of the types it is lexically a member of. Example of up
1677 references (with their equivalent as named type declarations) include:</p>
1679 <pre>
1680 { \2 * } %x = type { %x* }
1681 { \2 }* %y = type { %y }*
1682 \1* %z = type %z*
1683 </pre>
1685 <p>
1686 An up reference is needed by the asmprinter for printing out cyclic types when
1687 there is no declared name for a type in the cycle. Because the asmprinter does
1688 not want to print out an infinite type string, it needs a syntax to handle
1689 recursive types that have no names (all names are optional in llvm IR).
1690 </p>
1693 <pre>
1695 </pre>
1697 <p>
1698 The level is the count of the lexical type that is being referred to.
1699 </p>
1707 </tr>
1711 structure.</td>
1712 </tr>
1713 </table>
1714 </div>
1717 <!-- *********************************************************************** -->
1719 <!-- *********************************************************************** -->
1723 <p>LLVM has several different basic types of constants. This section describes
1724 them all and their syntax.</p>
1726 </div>
1728 <!-- ======================================================================= -->
1733 <dl>
1738 </dd>
1744 integer types.
1745 </dd>
1750 exponential notation (e.g. 1.23421e+2), or a more precise hexadecimal
1751 notation (see below). The assembler requires the exact decimal value of
1752 a floating-point constant. For example, the assembler accepts 1.25 but
1761 </dl>
1763 <p>The one non-intuitive notation for constants is the hexadecimal form
1764 of floating point constants. For example, the form '<tt>double
1767 (and the only time that they are generated by the disassembler) is when a
1768 floating point constant must be emitted but it cannot be represented as a
1769 decimal floating point number in a reasonable number of digits. For example,
1770 NaN's, infinities, and other
1771 special values are represented in their IEEE hexadecimal format so that
1772 assembly and disassembly do not cause any bits to change in the constants.</p>
1773 <p>When using the hexadecimal form, constants of types float and double are
1774 represented using the 16-digit form shown above (which matches the IEEE754
1775 representation for double); float values must, however, be exactly representable
1776 as IEE754 single precision.
1777 Hexadecimal format is always used for long
1778 double, and there are three forms of long double. The 80-bit
1780 followed by 20 hexadecimal digits.
1781 The 128-bit format used by PowerPC (two adjacent doubles) is represented
1783 format is represented
1785 target uses this format. Long doubles will only work if they match
1786 the long double format on your target. All hexadecimal formats are big-endian
1787 (sign bit at the left).</p>
1788 </div>
1790 <!-- ======================================================================= -->
1794 </div>
1797 <p>Complex constants are a (potentially recursive) combination of simple
1798 constants and smaller complex constants.</p>
1800 <dl>
1803 <dd>Structure constants are represented with notation similar to structure
1804 type definitions (a comma separated list of elements, surrounded by braces
1808 types of elements must match those specified by the type.
1809 </dd>
1813 <dd>Array constants are represented with notation similar to array type
1814 definitions (a comma separated list of elements, surrounded by square brackets
1817 types of elements must match those specified by the type.
1818 </dd>
1822 <dd>Vector constants are represented with notation similar to vector type
1823 definitions (a comma separated list of elements, surrounded by
1827 match those specified by the type.
1828 </dd>
1834 This is often used to avoid having to print large zero initializers (e.g. for
1835 large arrays) and is always exactly equivalent to using explicit zero
1836 initializers.
1837 </dd>
1838 </dl>
1840 </div>
1842 <!-- ======================================================================= -->
1845 </div>
1851 constants. These constants are explicitly referenced when the <a
1854 file:</p>
1857 <pre>
1858 @X = global i32 17
1859 @Y = global i32 42
1860 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
1861 </pre>
1862 </div>
1864 </div>
1866 <!-- ======================================================================= -->
1870 no specific value. Undefined values may be of any type and be used anywhere
1871 a constant is permitted.</p>
1873 <p>Undefined values indicate to the compiler that the program is well defined
1874 no matter what value is used, giving the compiler more freedom to optimize.
1875 </p>
1876 </div>
1878 <!-- ======================================================================= -->
1880 </div>
1884 <p>Constant expressions are used to allow expressions involving other constants
1885 to be used as constants. Constant expressions may be of any <a
1887 that does not have side effects (e.g. load and call are not supported). The
1888 following is the syntax for constant expressions:</p>
1890 <dl>
1892 <dd>Truncate a constant to another type. The bit size of CST must be larger
1893 than the bit size of TYPE. Both types must be integers.</dd>
1896 <dd>Zero extend a constant to another type. The bit size of CST must be
1897 smaller or equal to the bit size of TYPE. Both types must be integers.</dd>
1900 <dd>Sign extend a constant to another type. The bit size of CST must be
1901 smaller or equal to the bit size of TYPE. Both types must be integers.</dd>
1904 <dd>Truncate a floating point constant to another floating point type. The
1905 size of CST must be larger than the size of TYPE. Both types must be
1906 floating point.</dd>
1909 <dd>Floating point extend a constant to another type. The size of CST must be
1910 smaller or equal to the size of TYPE. Both types must be floating point.</dd>
1913 <dd>Convert a floating point constant to the corresponding unsigned integer
1914 constant. TYPE must be a scalar or vector integer type. CST must be of scalar
1915 or vector floating point type. Both CST and TYPE must be scalars, or vectors
1916 of the same number of elements. If the value won't fit in the integer type,
1917 the results are undefined.</dd>
1920 <dd>Convert a floating point constant to the corresponding signed integer
1921 constant. TYPE must be a scalar or vector integer type. CST must be of scalar
1922 or vector floating point type. Both CST and TYPE must be scalars, or vectors
1923 of the same number of elements. If the value won't fit in the integer type,
1924 the results are undefined.</dd>
1927 <dd>Convert an unsigned integer constant to the corresponding floating point
1928 constant. TYPE must be a scalar or vector floating point type. CST must be of
1929 scalar or vector integer type. Both CST and TYPE must be scalars, or vectors
1930 of the same number of elements. If the value won't fit in the floating point
1931 type, the results are undefined.</dd>
1934 <dd>Convert a signed integer constant to the corresponding floating point
1935 constant. TYPE must be a scalar or vector floating point type. CST must be of
1936 scalar or vector integer type. Both CST and TYPE must be scalars, or vectors
1937 of the same number of elements. If the value won't fit in the floating point
1938 type, the results are undefined.</dd>
1941 <dd>Convert a pointer typed constant to the corresponding integer constant
1942 TYPE must be an integer type. CST must be of pointer type. The CST value is
1943 zero extended, truncated, or unchanged to make it fit in TYPE.</dd>
1946 <dd>Convert a integer constant to a pointer constant. TYPE must be a
1947 pointer type. CST must be of integer type. The CST value is zero extended,
1948 truncated, or unchanged to make it fit in a pointer size. This one is
1952 <dd>Convert a constant, CST, to another TYPE. The constraints of the operands
1960 instruction, the index list may have zero or more indexes, which are required
1966 constants.</dd>
1998 <dd>Perform the specified operation of the LHS and RHS constants. OPCODE may
2000 binary</a> operations. The constraints on operands are the same as those for
2001 the corresponding instruction (e.g. no bitwise operations on floating point
2002 values are allowed).</dd>
2003 </dl>
2004 </div>
2006 <!-- *********************************************************************** -->
2008 <!-- *********************************************************************** -->
2010 <!-- ======================================================================= -->
2013 </div>
2017 <p>
2019 Module-Level Inline Assembly</a>) through the use of a special value. This
2020 value represents the inline assembler as a string (containing the instructions
2021 to emit), a list of operand constraints (stored as a string), and a flag that
2022 indicates whether or not the inline asm expression has side effects. An example
2023 inline assembler expression is:
2024 </p>
2027 <pre>
2029 </pre>
2030 </div>
2032 <p>
2035 </p>
2038 <pre>
2040 </pre>
2041 </div>
2043 <p>
2044 Inline asms with side effects not visible in the constraint list must be marked
2045 as having side effects. This is done through the use of the
2047 </p>
2050 <pre>
2052 </pre>
2053 </div>
2055 <p>TODO: The format of the asm and constraints string still need to be
2056 documented here. Constraints on what can be done (e.g. duplication, moving, etc
2057 need to be documented). This is probably best done by reference to another
2058 document that covers inline asm from a holistic perspective.
2059 </p>
2061 </div>
2063 <!-- *********************************************************************** -->
2065 <!-- *********************************************************************** -->
2069 <p>The LLVM instruction set consists of several different
2076 </div>
2078 <!-- ======================================================================= -->
2080 Instructions</a> </div>
2085 basic block in a program ends with a "Terminator" instruction, which
2086 indicates which block should be executed after the current block is
2088 value: they produce control flow, not values (the one exception being
2097 </div>
2099 <!-- _______________________________________________________________________ -->
2101 Instruction</a> </div>
2104 <pre>
2107 </pre>
2112 optionally a value) from a function back to the caller.</p>
2114 returns a value and then causes control flow, and one that just causes
2115 control flow to occur.</p>
2120 the return value. The type of the return value must be a
2125 instruction with no return value or a return value with a type that
2126 does not match its type, or if it has a void return type and contains
2132 returns back to the calling function's context. If the caller is a "<a
2134 the instruction after the call. If the caller was an "<a
2136 at the beginning of the "normal" destination block. If the instruction
2137 returns a value, that value shall set the call or invoke instruction's
2138 return value.</p>
2142 <pre>
2146 </pre>
2148 <p>Note that the code generator does not yet fully support large
2149 return values. The specific sizes that are currently supported are
2150 dependent on the target. For integers, on 32-bit targets the limit
2152 For aggregate types, the current limits are dependent on the element
2153 types; for example targets are often limited to 2 total integer
2156 </div>
2157 <!-- _______________________________________________________________________ -->
2161 <pre> br i1 <cond>, label <iftrue>, label <iffalse><br> br label <dest> <i>; Unconditional branch</i>
2162 </pre>
2165 transfer to a different basic block in the current function. There are
2166 two forms of this instruction, corresponding to a conditional branch
2167 and an unconditional branch.</p>
2179 <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
2181 </div>
2182 <!-- _______________________________________________________________________ -->
2185 </div>
2190 <pre>
2191 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
2192 </pre>
2198 instruction, allowing a branch to occur to one of many possible
2199 destinations.</p>
2207 table is not allowed to contain duplicate constant entries.</p>
2213 table is searched for the given value. If the value is found, control flow is
2214 transfered to the corresponding destination; otherwise, control flow is
2215 transfered to the default destination.</p>
2219 <p>Depending on properties of the target machine and the particular
2221 ways. For example, it could be generated as a series of chained conditional
2222 branches or with a lookup table.</p>
2226 <pre>
2229 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
2232 switch i32 0, label %dest [ ]
2235 switch i32 %val, label %otherwise [ i32 0, label %onzero
2236 i32 1, label %onone
2237 i32 2, label %ontwo ]
2238 </pre>
2239 </div>
2241 <!-- _______________________________________________________________________ -->
2244 </div>
2250 <pre>
2251 <result> = invoke [<a href="#callingconv">cconv</a>] [<a href="#paramattrs">ret attrs</a>] <ptr to function ty> <function ptr val>(<function args>) [<a href="#fnattrs">fn attrs</a>]
2253 </pre>
2258 function, with the possibility of control flow transfer to either the
2270 <ol>
2271 <li>
2273 convention</a> the call should use. If none is specified, the call defaults
2274 to using C calling conventions.
2275 </li>
2282 function value being invoked. In most cases, this is a direct function
2284 an arbitrary pointer to function value.
2285 </li>
2288 function to be invoked. </li>
2291 signature argument types. If the function signature indicates the function
2292 accepts a variable number of arguments, the extra arguments can be
2293 specified. </li>
2304 </ol>
2310 difference is that it establishes an association with a label, which is used by
2311 the runtime library to unwind the stack.</p>
2313 <p>This instruction is used in languages with destructors to ensure that proper
2315 exception. Additionally, this is important for implementation of
2319 <pre>
2320 %retval = invoke i32 @Test(i32 15) to label %Continue
2324 </pre>
2325 </div>
2328 <!-- _______________________________________________________________________ -->
2331 Instruction</a> </div>
2336 <pre>
2337 unwind
2338 </pre>
2343 at the first callee in the dynamic call stack which used an <a
2345 primarily used to implement exception handling.</p>
2350 immediately halt. The dynamic call stack is then searched for the first <a
2352 execution continues at the "exceptional" destination block specified by the
2354 dynamic call chain, undefined behavior results.</p>
2355 </div>
2357 <!-- _______________________________________________________________________ -->
2360 Instruction</a> </div>
2365 <pre>
2366 unreachable
2367 </pre>
2372 instruction is used to inform the optimizer that a particular portion of the
2373 code is not reachable. This can be used to indicate that the code after a
2374 no-return function cannot be reached, and other facts.</p>
2379 </div>
2383 <!-- ======================================================================= -->
2386 <p>Binary operators are used to do most of the computation in a
2387 program. They require two operands of the same type, execute an operation on them, and
2388 produce a single value. The operands might represent
2390 The result value has the same type as its operands.</p>
2392 </div>
2393 <!-- _______________________________________________________________________ -->
2396 </div>
2402 <pre>
2404 </pre>
2415 types.</p>
2419 <p>The value produced is the integer or floating point sum of the two
2420 operands.</p>
2422 <p>If an integer sum has unsigned overflow, the result returned is the
2424 the result.</p>
2426 <p>Because LLVM integers use a two's complement representation, this
2427 instruction is appropriate for both signed and unsigned integers.</p>
2431 <pre>
2433 </pre>
2434 </div>
2435 <!-- _______________________________________________________________________ -->
2438 </div>
2444 <pre>
2446 </pre>
2451 operands.</p>
2455 representations.</p>
2462 types.</p>
2466 <p>The value produced is the integer or floating point difference of
2467 the two operands.</p>
2469 <p>If an integer difference has unsigned overflow, the result returned is the
2471 the result.</p>
2473 <p>Because LLVM integers use a two's complement representation, this
2474 instruction is appropriate for both signed and unsigned integers.</p>
2477 <pre>
2480 </pre>
2481 </div>
2483 <!-- _______________________________________________________________________ -->
2486 </div>
2492 </pre>
2495 operands.</p>
2502 types.</p>
2506 <p>The value produced is the integer or floating point product of the
2507 two operands.</p>
2509 <p>If the result of an integer multiplication has unsigned overflow,
2510 the result returned is the mathematical result modulo
2512 <p>Because LLVM integers use a two's complement representation, and the
2513 result is the same width as the operands, this instruction returns the
2514 correct result for both signed and unsigned integers. If a full product
2516 should be sign-extended or zero-extended as appropriate to the
2517 width of the full product.</p>
2520 </pre>
2521 </div>
2523 <!-- _______________________________________________________________________ -->
2525 </a></div>
2529 </pre>
2532 operands.</p>
2538 values. Both arguments must have identical types.</p>
2543 <p>Note that unsigned integer division and signed integer division are distinct
2548 </pre>
2549 </div>
2550 <!-- _______________________________________________________________________ -->
2552 </a> </div>
2555 <pre>
2557 </pre>
2562 operands.</p>
2568 values. Both arguments must have identical types.</p>
2571 <p>The value produced is the signed integer quotient of the two operands rounded towards zero.</p>
2572 <p>Note that signed integer division and unsigned integer division are distinct
2574 <p>Division by zero leads to undefined behavior. Overflow also leads to
2575 undefined behavior; this is a rare case, but can occur, for example,
2579 </pre>
2580 </div>
2581 <!-- _______________________________________________________________________ -->
2583 Instruction</a> </div>
2586 <pre>
2588 </pre>
2592 operands.</p>
2598 of floating point values. Both arguments must have identical types.</p>
2606 <pre>
2608 </pre>
2609 </div>
2611 <!-- _______________________________________________________________________ -->
2613 </div>
2617 </pre>
2620 unsigned division of its two arguments.</p>
2624 values. Both arguments must have identical types.</p>
2627 This instruction always performs an unsigned division to get the remainder.</p>
2628 <p>Note that unsigned integer remainder and signed integer remainder are
2633 </pre>
2635 </div>
2636 <!-- _______________________________________________________________________ -->
2639 </div>
2645 <pre>
2647 </pre>
2652 signed division of its two operands. This instruction can also take
2654 the elements must be integers.</p>
2660 values. Both arguments must have identical types.</p>
2667 a value. For more information about the difference, see <a
2669 Math Forum</a>. For a table of how this is implemented in various languages,
2672 <p>Note that signed integer remainder and unsigned integer remainder are
2674 <p>Taking the remainder of a division by zero leads to undefined behavior.
2675 Overflow also leads to undefined behavior; this is a rare case, but can occur,
2677 (The remainder doesn't actually overflow, but this rule lets srem be
2678 implemented using instructions that return both the result of the division
2679 and the remainder.)</p>
2682 </pre>
2684 </div>
2685 <!-- _______________________________________________________________________ -->
2693 </pre>
2696 division of its two operands.</p>
2700 of floating point values. Both arguments must have identical types.</p>
2705 The remainder has the same sign as the dividend.</p>
2709 <pre>
2711 </pre>
2712 </div>
2714 <!-- ======================================================================= -->
2716 Operations</a> </div>
2718 <p>Bitwise binary operators are used to do various forms of
2719 bit-twiddling in a program. They are generally very efficient
2720 instructions and can commonly be strength reduced from other
2721 instructions. They require two operands of the same type, execute an operation on them,
2722 and produce a single value. The resulting value is the same type as its operands.</p>
2723 </div>
2725 <!-- _______________________________________________________________________ -->
2727 Instruction</a> </div>
2731 </pre>
2736 the left a specified number of bits.</p>
2757 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> <i>; yields: result=<2 x i32> < i32 2, i32 4></i>
2758 </pre>
2759 </div>
2760 <!-- _______________________________________________________________________ -->
2762 Instruction</a> </div>
2766 </pre>
2770 operand shifted to the right a specified number of bits with zero fill.</p>
2779 <p>This instruction always performs a logical shift right operation. The most
2780 significant bits of the result will be filled with zero bits after the
2787 <pre>
2793 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> <i>; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1></i>
2794 </pre>
2795 </div>
2797 <!-- _______________________________________________________________________ -->
2799 Instruction</a> </div>
2804 </pre>
2808 operand shifted to the right a specified number of bits with sign extension.</p>
2816 <p>This instruction always performs an arithmetic shift right operation,
2817 The most significant bits of the result will be filled with the sign bit
2824 <pre>
2830 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> <i>; yields: result=<2 x i32> < i32 -1, i32 0></i>
2831 </pre>
2832 </div>
2834 <!-- _______________________________________________________________________ -->
2836 Instruction</a> </div>
2842 <pre>
2844 </pre>
2849 its two operands.</p>
2855 values. Both arguments must have identical types.</p>
2859 <p> </p>
2860 <div>
2862 <tbody>
2863 <tr>
2867 </tr>
2868 <tr>
2872 </tr>
2873 <tr>
2877 </tr>
2878 <tr>
2882 </tr>
2883 <tr>
2887 </tr>
2888 </tbody>
2889 </table>
2890 </div>
2892 <pre>
2896 </pre>
2897 </div>
2898 <!-- _______________________________________________________________________ -->
2903 </pre>
2906 or of its two operands.</p>
2911 values. Both arguments must have identical types.</p>
2914 <p> </p>
2915 <div>
2917 <tbody>
2918 <tr>
2922 </tr>
2923 <tr>
2927 </tr>
2928 <tr>
2932 </tr>
2933 <tr>
2937 </tr>
2938 <tr>
2942 </tr>
2943 </tbody>
2944 </table>
2945 </div>
2950 </pre>
2951 </div>
2952 <!-- _______________________________________________________________________ -->
2954 Instruction</a> </div>
2958 </pre>
2966 values. Both arguments must have identical types.</p>
2971 <p> </p>
2972 <div>
2974 <tbody>
2975 <tr>
2979 </tr>
2980 <tr>
2984 </tr>
2985 <tr>
2989 </tr>
2990 <tr>
2994 </tr>
2995 <tr>
2999 </tr>
3000 </tbody>
3001 </table>
3002 </div>
3003 <p> </p>
3009 </pre>
3010 </div>
3012 <!-- ======================================================================= -->
3015 </div>
3019 <p>LLVM supports several instructions to represent vector operations in a
3020 target-independent manner. These instructions cover the element-access and
3021 vector-specific operations needed to process vectors effectively. While LLVM
3022 does directly support these vector operations, many sophisticated algorithms
3023 will want to use target-specific intrinsics to take full advantage of a specific
3024 target.</p>
3026 </div>
3028 <!-- _______________________________________________________________________ -->
3031 </div>
3037 <pre>
3038 <result> = extractelement <n x <ty>> <val>, i32 <idx> <i>; yields <ty></i>
3039 </pre>
3043 <p>
3045 element from a vector at a specified index.
3046 </p>
3051 <p>
3054 an index indicating the position from which to extract the element.
3055 The index may be a variable.</p>
3059 <p>
3060 The result is a scalar of the same type as the element type of
3063 results are undefined.
3064 </p>
3068 <pre>
3070 </pre>
3071 </div>
3074 <!-- _______________________________________________________________________ -->
3077 </div>
3083 <pre>
3084 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> <i>; yields <n x <ty>></i>
3085 </pre>
3089 <p>
3091 element into a vector at a specified index.
3092 </p>
3097 <p>
3100 scalar value whose type must equal the element type of the first
3101 operand. The third operand is an index indicating the position at
3102 which to insert the value. The index may be a variable.</p>
3106 <p>
3111 </p>
3115 <pre>
3117 </pre>
3118 </div>
3120 <!-- _______________________________________________________________________ -->
3123 </div>
3129 <pre>
3130 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> <i>; yields <m x <ty>></i>
3131 </pre>
3135 <p>
3137 from two input vectors, returning a vector with the same element type as
3138 the input and length that is the same as the shuffle mask.
3139 </p>
3143 <p>
3145 with types that match each other. The third argument is a shuffle mask whose
3146 element type is always 'i32'. The result of the instruction is a vector whose
3147 length is the same as the shuffle mask and whose element type is the same as
3148 the element type of the first two operands.
3149 </p>
3151 <p>
3152 The shuffle mask operand is required to be a constant vector with either
3153 constant integer or undef values.
3154 </p>
3158 <p>
3159 The elements of the two input vectors are numbered from left to right across
3160 both of the vectors. The shuffle mask operand specifies, for each element of
3161 the result vector, which element of the two input vectors the result element
3162 gets. The element selector may be undef (meaning "don't care") and the second
3163 operand may be undef if performing a shuffle from only one vector.
3164 </p>
3168 <pre>
3172 <4 x i32> <i32 0, i32 1, i32 2, i32 3> <i>; yields <4 x i32></i> - Identity shuffle.
3176 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > <i>; yields <8 x i32></i>
3177 </pre>
3178 </div>
3181 <!-- ======================================================================= -->
3184 </div>
3188 <p>LLVM supports several instructions for working with aggregate values.
3189 </p>
3191 </div>
3193 <!-- _______________________________________________________________________ -->
3196 </div>
3202 <pre>
3204 </pre>
3208 <p>
3210 or array element from an aggregate value.
3211 </p>
3216 <p>
3219 type. The operands are constant indices to specify which value to extract
3220 in a similar manner as indices in a
3222 </p>
3226 <p>
3227 The result is the value at the position in the aggregate specified by
3228 the index operands.
3229 </p>
3233 <pre>
3235 </pre>
3236 </div>
3239 <!-- _______________________________________________________________________ -->
3242 </div>
3248 <pre>
3249 <result> = insertvalue <aggregate type> <val>, <ty> <val>, <idx> <i>; yields <n x <ty>></i>
3250 </pre>
3254 <p>
3256 into a struct field or array element in an aggregate.
3257 </p>
3262 <p>
3265 The second operand is a first-class value to insert.
3266 The following operands are constant indices
3267 indicating the position at which to insert the value in a similar manner as
3268 indices in a
3270 The value to insert must have the same type as the value identified
3271 by the indices.
3272 </p>
3276 <p>
3280 </p>
3284 <pre>
3286 </pre>
3287 </div>
3290 <!-- ======================================================================= -->
3293 </div>
3297 <p>A key design point of an SSA-based representation is how it
3298 represents memory. In LLVM, no memory locations are in SSA form, which
3299 makes things very simple. This section describes how to read, write,
3300 allocate, and free memory in LLVM.</p>
3302 </div>
3304 <!-- _______________________________________________________________________ -->
3307 </div>
3313 <pre>
3314 <result> = malloc <type>[, i32 <NumElements>][, align <alignment>] <i>; yields {type*}:result</i>
3315 </pre>
3320 heap and returns a pointer to it. The object is always allocated in the generic
3321 address space (address space zero).</p>
3327 bytes of memory from the operating system and returns a pointer of the
3328 appropriate type to the program. If "NumElements" is specified, it is the
3329 number of elements allocated, otherwise "NumElements" is defaulted to be one.
3330 If a constant alignment is specified, the value result of the allocation is guaranteed to
3331 be aligned to at least that boundary. If not specified, or if zero, the target can
3332 choose to align the allocation on any convenient boundary.</p>
3339 a pointer is returned. The result of a zero byte allocation is undefined. The
3340 result is null if there is insufficient memory available.</p>
3344 <pre>
3352 </pre>
3354 <p>Note that the code generator does not yet respect the
3355 alignment value.</p>
3357 </div>
3359 <!-- _______________________________________________________________________ -->
3362 </div>
3368 <pre>
3370 </pre>
3375 memory heap to be reallocated in the future.</p>
3381 instruction.</p>
3385 <p>Access to the memory pointed to by the pointer is no longer defined
3386 after this instruction executes. If the pointer is null, the operation
3387 is a noop.</p>
3391 <pre>
3393 free [4 x i8]* %array
3394 </pre>
3395 </div>
3397 <!-- _______________________________________________________________________ -->
3400 </div>
3406 <pre>
3407 <result> = alloca <type>[, i32 <NumElements>][, align <alignment>] <i>; yields {type*}:result</i>
3408 </pre>
3413 currently executing function, to be automatically released when this function
3414 returns to its caller. The object is always allocated in the generic address
3415 space (address space zero).</p>
3420 bytes of memory on the runtime stack, returning a pointer of the
3421 appropriate type to the program. If "NumElements" is specified, it is the
3422 number of elements allocated, otherwise "NumElements" is defaulted to be one.
3423 If a constant alignment is specified, the value result of the allocation is guaranteed
3424 to be aligned to at least that boundary. If not specified, or if zero, the target
3425 can choose to align the allocation on any convenient boundary.</p>
3431 <p>Memory is allocated; a pointer is returned. The operation is undefiend if
3434 instruction is commonly used to represent automatic variables that must
3435 have an address available. When the function returns (either with the <tt><a
3437 instructions), the memory is reclaimed. Allocating zero bytes
3438 is legal, but the result is undefined.</p>
3442 <pre>
3447 </pre>
3448 </div>
3450 <!-- _______________________________________________________________________ -->
3452 Instruction</a> </div>
3455 <pre> <result> = load <ty>* <pointer>[, align <alignment>]<br> <result> = volatile load <ty>* <pointer>[, align <alignment>]<br></pre>
3460 address from which to load. The pointer must point to a <a
3465 instructions. </p>
3466 <p>
3467 The optional constant "align" argument specifies the alignment of the operation
3468 (that is, the alignment of the memory address). A value of 0 or an
3469 omitted "align" argument means that the operation has the preferential
3470 alignment for the target. It is the responsibility of the code emitter
3471 to ensure that the alignment information is correct. Overestimating
3472 the alignment results in an undefined behavior. Underestimating the
3473 alignment may produce less efficient code. An alignment of 1 is always
3474 safe.
3475 </p>
3480 <a
3483 </pre>
3484 </div>
3485 <!-- _______________________________________________________________________ -->
3487 Instruction</a> </div>
3490 <pre> store <ty> <value>, <ty>* <pointer>[, align <alignment>] <i>; yields {void}</i>
3491 volatile store <ty> <value>, <ty>* <pointer>[, align <alignment>] <i>; yields {void}</i>
3492 </pre>
3501 optimizer is not allowed to modify the number or order of execution of
3504 <p>
3505 The optional constant "align" argument specifies the alignment of the operation
3506 (that is, the alignment of the memory address). A value of 0 or an
3507 omitted "align" argument means that the operation has the preferential
3508 alignment for the target. It is the responsibility of the code emitter
3509 to ensure that the alignment information is correct. Overestimating
3510 the alignment results in an undefined behavior. Underestimating the
3511 alignment may produce less efficient code. An alignment of 1 is always
3512 safe.
3513 </p>
3521 </pre>
3522 </div>
3524 <!-- _______________________________________________________________________ -->
3527 </div>
3531 <pre>
3533 </pre>
3537 <p>
3539 subelement of an aggregate data structure. It performs address calculation only
3540 and does not access memory.</p>
3544 <p>The first argument is always a pointer, and forms the basis of the
3545 calculation. The remaining arguments are indices, that indicate which of the
3546 elements of the aggregate object are indexed. The interpretation of each index
3547 is dependent on the type being indexed into. The first index always indexes the
3548 pointer value given as the first argument, the second index indexes a value of
3549 the type pointed to (not necessarily the value directly pointed to, since the
3550 first index can be non-zero), etc. The first type indexed into must be a pointer
3551 value, subsequent types can be arrays, vectors and structs. Note that subsequent
3552 types being indexed into can never be pointers, since that would require loading
3553 the pointer before continuing calculation.</p>
3555 <p>The type of each index argument depends on the type it is indexing into.
3561 <p>For example, let's consider a C code fragment and how it gets
3562 compiled to LLVM:</p>
3565 <pre>
3566 struct RT {
3567 char A;
3569 char C;
3570 };
3571 struct ST {
3572 int X;
3573 double Y;
3574 struct RT Z;
3575 };
3577 int *foo(struct ST *s) {
3579 }
3580 </pre>
3581 </div>
3586 <pre>
3590 define i32* %foo(%ST* %s) {
3591 entry:
3593 ret i32* %reg
3594 }
3595 </pre>
3596 </div>
3602 }</tt>' type, a structure. The second index indexes into the third element of
3604 i8 }</tt>' type, another structure. The third index indexes into the second
3606 array. The two dimensions of the array are subscripted into, yielding an
3610 <p>Note that it is perfectly legal to index partially through a
3611 structure, returning a pointer to an inner element. Because of this,
3612 the LLVM code for the given testcase is equivalent to:</p>
3614 <pre>
3615 define i32* %foo(%ST* %s) {
3621 ret i32* %t5
3622 }
3623 </pre>
3625 <p>Note that it is undefined to access an array out of bounds: array and
3626 pointer indexes must always be within the defined bounds of the array type.
3627 The one exception for this rule is zero length arrays. These arrays are
3628 defined to be accessible as variable length arrays, which requires access
3629 beyond the zero'th element.</p>
3631 <p>The getelementptr instruction is often confusing. For some more insight
3637 <pre>
3644 </pre>
3645 </div>
3647 <!-- ======================================================================= -->
3649 </div>
3651 <p>The instructions in this category are the conversion instructions (casting)
3652 which all take a single operand and a type. They perform various bit conversions
3653 on the operand.</p>
3654 </div>
3656 <!-- _______________________________________________________________________ -->
3659 </div>
3663 <pre>
3665 </pre>
3668 <p>
3670 </p>
3673 <p>
3681 <p>
3685 It will always truncate bits.</p>
3688 <pre>
3692 </pre>
3693 </div>
3695 <!-- _______________________________________________________________________ -->
3698 </div>
3702 <pre>
3704 </pre>
3725 <pre>
3728 </pre>
3729 </div>
3731 <!-- _______________________________________________________________________ -->
3734 </div>
3738 <pre>
3740 </pre>
3746 <p>
3754 <p>
3762 <pre>
3765 </pre>
3766 </div>
3768 <!-- _______________________________________________________________________ -->
3771 </div>
3777 <pre>
3779 </pre>
3800 <pre>
3803 </pre>
3804 </div>
3806 <!-- _______________________________________________________________________ -->
3809 </div>
3813 <pre>
3815 </pre>
3819 floating point value.</p>
3825 type must be smaller than the destination type.</p>
3835 <pre>
3838 </pre>
3839 </div>
3841 <!-- _______________________________________________________________________ -->
3844 </div>
3848 <pre>
3850 </pre>
3855 </p>
3868 the results are undefined.</p>
3871 <pre>
3875 </pre>
3876 </div>
3878 <!-- _______________________________________________________________________ -->
3881 </div>
3885 <pre>
3887 </pre>
3892 </p>
3905 the results are undefined.</p>
3908 <pre>
3912 </pre>
3913 </div>
3915 <!-- _______________________________________________________________________ -->
3918 </div>
3922 <pre>
3924 </pre>
3939 integer quantity and converts it to the corresponding floating point value. If
3940 the value cannot fit in the floating point value, the results are undefined.</p>
3943 <pre>
3946 </pre>
3947 </div>
3949 <!-- _______________________________________________________________________ -->
3952 </div>
3956 <pre>
3958 </pre>
3973 integer quantity and converts it to the corresponding floating point value. If
3974 the value cannot fit in the floating point value, the results are undefined.</p>
3977 <pre>
3980 </pre>
3981 </div>
3983 <!-- _______________________________________________________________________ -->
3986 </div>
3990 <pre>
3992 </pre>
4006 truncating or zero extending that value to the size of the integer type. If
4010 change.</p>
4013 <pre>
4016 </pre>
4017 </div>
4019 <!-- _______________________________________________________________________ -->
4022 </div>
4026 <pre>
4028 </pre>
4036 value to cast, and a type to cast it to, which must be a
4044 the size of a pointer then a zero extension is done. If they are the same size,
4048 <pre>
4052 </pre>
4053 </div>
4055 <!-- _______________________________________________________________________ -->
4058 </div>
4062 <pre>
4064 </pre>
4074 a non-aggregate first class value, and a type to cast it to, which must also be
4078 type is a pointer, the destination type must also be a pointer. This
4079 instruction supports bitwise conversion of vectors to integers and to vectors
4080 of other types (as long as they have the same size).</p>
4087 converted to other pointer types with this instruction. To convert pointers to
4092 <pre>
4096 </pre>
4097 </div>
4099 <!-- ======================================================================= -->
4103 instructions, which defy better classification.</p>
4104 </div>
4106 <!-- _______________________________________________________________________ -->
4108 </div>
4111 <pre> <result> = icmp <cond> <ty> <op1>, <op2> <i>; yields {i1} or {<N x i1>}:result</i>
4112 </pre>
4115 a vector of boolean values based on comparison
4116 of its two integer, integer vector, or pointer operands.</p>
4119 the condition code indicating the kind of comparison to perform. It is not
4120 a value, just a keyword. The possible condition code are:
4121 </p>
4122 <ol>
4133 </ol>
4137 They must also be identical types.</p>
4141 yields either an <a href="#t_primitive"><tt>i1</tt></a> or vector of <tt>i1</tt> result, as follows:
4142 </p>
4143 <ol>
4146 </li>
4165 </ol>
4167 values are compared as if they were integers.</p>
4168 <p>If the operands are integer vectors, then they are compared
4170 the same number of elements as the values being compared.
4172 </p>
4181 </pre>
4183 <p>Note that the code generator does not yet support vector types with
4186 </div>
4188 <!-- _______________________________________________________________________ -->
4190 </div>
4193 <pre> <result> = fcmp <cond> <ty> <op1>, <op2> <i>; yields {i1} or {<N x i1>}:result</i>
4194 </pre>
4197 or vector of boolean values based on comparison
4198 of its operands.</p>
4199 <p>
4200 If the operands are floating point scalars, then the result
4202 </p>
4203 <p>If the operands are floating point vectors, then the result type
4204 is a vector of boolean with the same number of elements as the
4205 operands being compared.</p>
4208 the condition code indicating the kind of comparison to perform. It is not
4209 a value, just a keyword. The possible condition code are:</p>
4210 <ol>
4227 </ol>
4233 They must have identical types.</p>
4237 If the operands are vectors, then the vectors are compared
4238 element by element.
4239 Each comparison performed
4241 <ol>
4270 </ol>
4277 </pre>
4279 <p>Note that the code generator does not yet support vector types with
4282 </div>
4284 <!-- _______________________________________________________________________ -->
4287 </div>
4290 <pre> <result> = vicmp <cond> <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4291 </pre>
4294 element-wise comparison of its two integer vector operands.</p>
4297 the condition code indicating the kind of comparison to perform. It is not
4298 a value, just a keyword. The possible condition code are:</p>
4299 <ol>
4310 </ol>
4317 identical type as the values being compared. The most significant bit in each
4319 otherwise. All other bits of the result are undefined. The condition codes
4324 <pre>
4325 <result> = vicmp eq <2 x i32> < i32 4, i32 0>, < i32 5, i32 0> <i>; yields: result=<2 x i32> < i32 0, i32 -1 ></i>
4326 <result> = vicmp ult <2 x i8 > < i8 1, i8 2>, < i8 2, i8 2 > <i>; yields: result=<2 x i8> < i8 -1, i8 0 ></i>
4327 </pre>
4328 </div>
4330 <!-- _______________________________________________________________________ -->
4333 </div>
4339 element-wise comparison of its two floating point vector operands. The output
4340 elements have the same width as the input elements.</p>
4343 the condition code indicating the kind of comparison to perform. It is not
4344 a value, just a keyword. The possible condition code are:</p>
4345 <ol>
4362 </ol>
4365 types.</p>
4370 an identical number of elements as the values being compared, and each element
4371 having identical with to the width of the floating point elements. The most
4372 significant bit in each element is 1 if the element-wise comparison evaluates to
4373 true, and is 0 otherwise. All other bits of the result are undefined. The
4374 condition codes are evaluated identically to the
4378 <pre>
4380 <result> = vfcmp oeq <2 x float> < float 4, float 0 >, < float 5, float 0 >
4383 <result> = vfcmp ult <2 x double> < double 1, double 2 >, < double 2, double 2>
4384 </pre>
4385 </div>
4387 <!-- _______________________________________________________________________ -->
4390 </div>
4399 the SSA graph representing the function.</p>
4402 <p>The type of the incoming values is specified with the first type
4404 as arguments, with one pair for each predecessor basic block of the
4406 type may be used as the value arguments to the PHI node. Only labels
4407 may be used as the label arguments.</p>
4409 <p>There must be no non-phi instructions between the start of a basic
4410 block and the PHI instructions: i.e. PHI instructions must be first in
4411 a basic block.</p>
4416 specified by the pair corresponding to the predecessor basic block that executed
4417 just prior to the current block.</p>
4420 <pre>
4421 Loop: ; Infinite loop that counts from 0 on up...
4422 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
4423 %nextindvar = add i32 %indvar, 1
4424 br label %Loop
4425 </pre>
4426 </div>
4428 <!-- _______________________________________________________________________ -->
4431 </div>
4437 <pre>
4438 <result> = select <i>selty</i> <cond>, <ty> <val1>, <ty> <val2> <i>; yields ty</i>
4441 </pre>
4445 <p>
4447 condition, without branching.
4448 </p>
4453 <p>
4455 a vector of 'i1' values indicating the
4457 type. If the val1/val2 are vectors and
4458 the condition is a scalar, then entire vectors are selected, not
4459 individual elements.
4460 </p>
4464 <p>
4465 If the condition is an i1 and it evaluates to 1, the instruction returns the first
4466 value argument; otherwise, it returns the second value argument.
4467 </p>
4468 <p>
4469 If the condition is a vector of i1, then the value arguments must
4470 be vectors of the same size, and the selection is done element
4471 by element.
4472 </p>
4476 <pre>
4478 </pre>
4480 <p>Note that the code generator does not yet support conditions
4481 with vector type.</p>
4483 </div>
4486 <!-- _______________________________________________________________________ -->
4489 </div>
4494 <pre>
4495 <result> = [tail] call [<a href="#callingconv">cconv</a>] [<a href="#paramattrs">ret attrs</a>] <ty> [<fnty>*] <fnptrval>(<function args>) [<a href="#fnattrs">fn attrs</a>]
4496 </pre>
4506 <ol>
4507 <li>
4509 any allocas or varargs in the caller. If the "tail" marker is present, the
4510 function call is eligible for tail call optimization. Note that calls may
4513 </li>
4514 <li>
4516 convention</a> the call should use. If none is specified, the call defaults
4517 to using C calling conventions.</p>
4518 </li>
4520 <li>
4524 </li>
4526 <li>
4528 the type of the return value. Functions that return no value are marked
4530 </li>
4531 <li>
4533 value being invoked. The argument types must match the types implied by
4534 this signature. This type can be omitted if the function is not varargs
4535 and if the function type does not return a pointer to a function.</p>
4536 </li>
4537 <li>
4539 be invoked. In most cases, this is a direct function invocation, but
4541 to function value.</p>
4542 </li>
4543 <li>
4545 function signature argument types. All arguments must be of
4547 indicates the function accepts a variable number of arguments, the extra
4548 arguments can be specified.</p>
4549 </li>
4550 <li>
4554 </li>
4555 </ol>
4560 transfer to a specified function, with its incoming arguments bound to
4562 instruction in the called function, control flow continues with the
4563 instruction after the function call, and the return value of the
4564 function is bound to the result argument.</p>
4568 <pre>
4569 %retval = call i32 @test(i32 %argc)
4573 call void %foo(i8 97 signext)
4575 %struct.A = type { i32, i8 }
4581 </pre>
4583 </div>
4585 <!-- _______________________________________________________________________ -->
4588 </div>
4594 <pre>
4596 </pre>
4601 the "variable argument" area of a function call. It is used to implement the
4607 the argument. It returns a value of the specified argument type and
4619 <p>It is legal for this instruction to be called in a function which does not
4621 function.</p>
4625 argument.</p>
4631 <p>Note that the code generator does not yet fully support va_arg
4632 on many targets. Also, it does not currently support va_arg with
4633 aggregate types on any target.</p>
4635 </div>
4637 <!-- *********************************************************************** -->
4639 <!-- *********************************************************************** -->
4644 well known names and semantics and are required to follow certain restrictions.
4645 Overall, these intrinsics represent an extension mechanism for the LLVM
4646 language that does not require changing all of the transformations in LLVM when
4647 adding to the language (or the bitcode reader/writer, the parser, etc...).</p>
4650 prefix is reserved in LLVM for intrinsic names; thus, function names may not
4651 begin with this prefix. Intrinsic functions must always be external functions:
4652 you cannot define the body of intrinsic functions. Intrinsic functions may
4653 only be used in call or invoke instructions: it is illegal to take the address
4654 of an intrinsic function. Additionally, because intrinsic functions are part
4655 of the LLVM language, it is required if any are added that they be documented
4656 here.</p>
4658 <p>Some intrinsic functions can be overloaded, i.e., the intrinsic represents
4659 a family of functions that perform the same operation but on different data
4660 types. Because LLVM can represent over 8 million different integer types,
4661 overloading is used commonly to allow an intrinsic function to operate on any
4662 integer type. One or more of the argument types or the result type can be
4663 overloaded to accept any integer type. Argument types may also be defined as
4664 exactly matching a previous argument's type or the result type. This allows an
4665 intrinsic function which accepts multiple arguments, but needs all of them to
4666 be of the same type, to only be overloaded with respect to a single argument or
4667 the result.</p>
4669 <p>Overloaded intrinsics will have the names of its overloaded argument types
4670 encoded into its function name, each preceded by a period. Only those types
4671 which are overloaded result in a name suffix. Arguments whose type is matched
4673 take an integer of any width and returns an integer of exactly the same integer
4674 width. This leads to a family of functions such as
4676 Only one type, the return type, is overloaded, and only one type suffix is
4677 required. Because the argument's type is matched against the return type, it
4678 does not require its own name suffix.</p>
4680 <p>To learn how to add an intrinsic function, please see the
4682 </p>
4684 </div>
4686 <!-- ======================================================================= -->
4689 </div>
4695 intrinsic functions. These functions are related to the similarly
4698 <p>All of these functions operate on arguments that use a
4699 target-specific value type "<tt>va_list</tt>". The LLVM assembly
4700 language reference manual does not define what this type is, so all
4701 transformations should be prepared to handle these functions regardless of
4702 the type used.</p>
4705 instruction and the variable argument handling intrinsic functions are
4706 used.</p>
4709 <pre>
4710 define i32 @test(i32 %X, ...) {
4711 ; Initialize variable argument processing
4712 %ap = alloca i8*
4713 %ap2 = bitcast i8** %ap to i8*
4714 call void @llvm.va_start(i8* %ap2)
4716 ; Read a single integer argument
4717 %tmp = va_arg i8** %ap, i32
4719 ; Demonstrate usage of llvm.va_copy and llvm.va_end
4720 %aq = alloca i8*
4721 %aq2 = bitcast i8** %aq to i8*
4722 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
4723 call void @llvm.va_end(i8* %aq2)
4725 ; Stop processing of arguments.
4726 call void @llvm.va_end(i8* %ap2)
4727 ret i32 %tmp
4728 }
4730 declare void @llvm.va_start(i8*)
4731 declare void @llvm.va_copy(i8*, i8*)
4732 declare void @llvm.va_end(i8*)
4733 </pre>
4734 </div>
4736 </div>
4738 <!-- _______________________________________________________________________ -->
4741 </div>
4759 macro available in C. In a target-dependent way, it initializes the
4763 last argument of the function as the compiler can figure that out.</p>
4765 </div>
4767 <!-- _______________________________________________________________________ -->
4770 </div>
4788 macro available in C. In a target-dependent way, it destroys the
4794 </div>
4796 <!-- _______________________________________________________________________ -->
4799 </div>
4805 <pre>
4807 </pre>
4812 from the source argument list to the destination argument list.</p>
4823 macro available in C. In a target-dependent way, it copies the source
4826 llvm.va_start</a></tt> intrinsic may be arbitrarily complex and require, for
4827 example, memory allocation.</p>
4829 </div>
4831 <!-- ======================================================================= -->
4834 </div>
4838 <p>
4840 Collection</a> (GC) requires the implementation and generation of these
4841 intrinsics.
4845 Front-ends for type-safe garbage collected languages should generate these
4846 intrinsics to make use of the LLVM garbage collectors. For more details, see <a
4848 </p>
4850 <p>The garbage collection intrinsics only operate on objects in the generic
4851 address space (address space zero).</p>
4853 </div>
4855 <!-- _______________________________________________________________________ -->
4858 </div>
4864 <pre>
4865 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
4866 </pre>
4871 the code generator, and allows some metadata to be associated with it.</p>
4875 <p>The first argument specifies the address of a stack object that contains the
4876 root pointer. The second pointer (which must be either a constant or a global
4877 value address) contains the meta-data to be associated with the root.</p>
4882 location. At compile-time, the code generator generates information to allow
4887 </div>
4890 <!-- _______________________________________________________________________ -->
4893 </div>
4899 <pre>
4900 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
4901 </pre>
4906 locations, allowing garbage collector implementations that require read
4907 barriers.</p>
4911 <p>The second argument is the address to read from, which should be an address
4912 allocated from the garbage collector. The first object is a pointer to the
4913 start of the referenced object, if needed by the language runtime (otherwise
4914 null).</p>
4919 instruction, but may be replaced with substantially more complex code by the
4924 </div>
4927 <!-- _______________________________________________________________________ -->
4930 </div>
4936 <pre>
4937 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
4938 </pre>
4943 locations, allowing garbage collector implementations that require write
4944 barriers (such as generational or reference counting collectors).</p>
4948 <p>The first argument is the reference to store, the second is the start of the
4949 object to store it to, and the third is the address of the field of Obj to
4950 store to. If the runtime does not require a pointer to the object, Obj may be
4951 null.</p>
4956 instruction, but may be replaced with substantially more complex code by the
4961 </div>
4965 <!-- ======================================================================= -->
4968 </div>
4971 <p>
4972 These intrinsics are provided by LLVM to expose special features that may only
4973 be implemented with code generator support.
4974 </p>
4976 </div>
4978 <!-- _______________________________________________________________________ -->
4981 </div>
4986 <pre>
4988 </pre>
4992 <p>
4994 target-specific value indicating the return address of the current function
4995 or one of its callers.
4996 </p>
5000 <p>
5001 The argument to this intrinsic indicates which function to return the address
5002 for. Zero indicates the calling function, one indicates its caller, etc. The
5004 </p>
5008 <p>
5010 the return address of the specified call frame, or zero if it cannot be
5011 identified. The value returned by this intrinsic is likely to be incorrect or 0
5012 for arguments other than zero, so it should only be used for debugging purposes.
5013 </p>
5015 <p>
5016 Note that calling this intrinsic does not prevent function inlining or other
5017 aggressive transformations, so the value returned may not be that of the obvious
5018 source-language caller.
5019 </p>
5020 </div>
5023 <!-- _______________________________________________________________________ -->
5026 </div>
5031 <pre>
5033 </pre>
5037 <p>
5039 target-specific frame pointer value for the specified stack frame.
5040 </p>
5044 <p>
5045 The argument to this intrinsic indicates which function to return the frame
5046 pointer for. Zero indicates the calling function, one indicates its caller,
5048 </p>
5052 <p>
5054 the frame address of the specified call frame, or zero if it cannot be
5055 identified. The value returned by this intrinsic is likely to be incorrect or 0
5056 for arguments other than zero, so it should only be used for debugging purposes.
5057 </p>
5059 <p>
5060 Note that calling this intrinsic does not prevent function inlining or other
5061 aggressive transformations, so the value returned may not be that of the obvious
5062 source-language caller.
5063 </p>
5064 </div>
5066 <!-- _______________________________________________________________________ -->
5069 </div>
5074 <pre>
5075 declare i8 *@llvm.stacksave()
5076 </pre>
5080 <p>
5084 features like scoped automatic variable sized arrays in C99.
5085 </p>
5089 <p>
5090 This intrinsic returns a opaque pointer value that can be passed to <a
5097 </p>
5099 </div>
5101 <!-- _______________________________________________________________________ -->
5104 </div>
5109 <pre>
5110 declare void @llvm.stackrestore(i8 * %ptr)
5111 </pre>
5115 <p>
5117 the function stack to the state it was in when the corresponding <a
5119 useful for implementing language features like scoped automatic variable sized
5120 arrays in C99.
5121 </p>
5125 <p>
5127 </p>
5129 </div>
5132 <!-- _______________________________________________________________________ -->
5135 </div>
5140 <pre>
5142 </pre>
5147 <p>
5149 a prefetch instruction if supported; otherwise, it is a noop. Prefetches have
5150 no
5151 effect on the behavior of the program but can change its performance
5152 characteristics.
5153 </p>
5157 <p>
5163 </p>
5167 <p>
5168 This intrinsic does not modify the behavior of the program. In particular,
5169 prefetches cannot trap and do not produce a value. On targets that support this
5170 intrinsic, the prefetch can provide hints to the processor cache for better
5171 performance.
5172 </p>
5174 </div>
5176 <!-- _______________________________________________________________________ -->
5179 </div>
5184 <pre>
5186 </pre>
5191 <p>
5193 (PC) in a region of
5194 code to simulators and other tools. The method is target specific, but it is
5195 expected that the marker will use exported symbols to transmit the PC of the
5196 marker.
5197 The marker makes no guarantees that it will remain with any specific instruction
5198 after optimizations. It is possible that the presence of a marker will inhibit
5199 optimizations. The intended use is to be inserted after optimizations to allow
5200 correlations of simulation runs.
5201 </p>
5205 <p>
5207 </p>
5211 <p>
5212 This intrinsic does not modify the behavior of the program. Backends that do not
5213 support this intrinisic may ignore it.
5214 </p>
5216 </div>
5218 <!-- _______________________________________________________________________ -->
5221 </div>
5226 <pre>
5227 declare i64 @llvm.readcyclecounter( )
5228 </pre>
5233 <p>
5235 counter register (or similar low latency, high accuracy clocks) on those targets
5236 that support it. On X86, it should map to RDTSC. On Alpha, it should map to RPCC.
5237 As the backing counters overflow quickly (on the order of 9 seconds on alpha), this
5238 should only be used for small timings.
5239 </p>
5243 <p>
5244 When directly supported, reading the cycle counter should not modify any memory.
5245 Implementations are allowed to either return a application specific value or a
5246 system wide value. On backends without support, this is lowered to a constant 0.
5247 </p>
5249 </div>
5251 <!-- ======================================================================= -->
5254 </div>
5257 <p>
5258 LLVM provides intrinsics for a few important standard C library functions.
5259 These intrinsics allow source-language front-ends to pass information about the
5260 alignment of the pointer arguments to the code generator, providing opportunity
5261 for more efficient code generation.
5262 </p>
5264 </div>
5266 <!-- _______________________________________________________________________ -->
5269 </div>
5274 <p>This is an overloaded intrinsic. You can use llvm.memcpy on any integer bit
5275 width. Not all targets support all bit widths however.</p>
5276 <pre>
5285 </pre>
5289 <p>
5291 location to the destination location.
5292 </p>
5294 <p>
5296 intrinsics do not return a value, and takes an extra alignment argument.
5297 </p>
5301 <p>
5302 The first argument is a pointer to the destination, the second is a pointer to
5303 the source. The third argument is an integer argument
5304 specifying the number of bytes to copy, and the fourth argument is the alignment
5305 of the source and destination locations.
5306 </p>
5308 <p>
5310 the caller guarantees that both the source and destination pointers are aligned
5311 to that boundary.
5312 </p>
5316 <p>
5318 location to the destination location, which are not allowed to overlap. It
5319 copies "len" bytes of memory over. If the argument is known to be aligned to
5320 some boundary, this can be specified as the fourth argument, otherwise it should
5322 </p>
5323 </div>
5326 <!-- _______________________________________________________________________ -->
5329 </div>
5334 <p>This is an overloaded intrinsic. You can use llvm.memmove on any integer bit
5335 width. Not all targets support all bit widths however.</p>
5336 <pre>
5345 </pre>
5349 <p>
5351 location to the destination location. It is similar to the
5353 </p>
5355 <p>
5357 intrinsics do not return a value, and takes an extra alignment argument.
5358 </p>
5362 <p>
5363 The first argument is a pointer to the destination, the second is a pointer to
5364 the source. The third argument is an integer argument
5365 specifying the number of bytes to copy, and the fourth argument is the alignment
5366 of the source and destination locations.
5367 </p>
5369 <p>
5371 the caller guarantees that the source and destination pointers are aligned to
5372 that boundary.
5373 </p>
5377 <p>
5379 location to the destination location, which may overlap. It
5380 copies "len" bytes of memory over. If the argument is known to be aligned to
5381 some boundary, this can be specified as the fourth argument, otherwise it should
5383 </p>
5384 </div>
5387 <!-- _______________________________________________________________________ -->
5390 </div>
5395 <p>This is an overloaded intrinsic. You can use llvm.memset on any integer bit
5396 width. Not all targets support all bit widths however.</p>
5397 <pre>
5406 </pre>
5410 <p>
5412 byte value.
5413 </p>
5415 <p>
5417 does not return a value, and takes an extra alignment argument.
5418 </p>
5422 <p>
5423 The first argument is a pointer to the destination to fill, the second is the
5424 byte value to fill it with, the third argument is an integer
5425 argument specifying the number of bytes to fill, and the fourth argument is the
5426 known alignment of destination location.
5427 </p>
5429 <p>
5431 the caller guarantees that the destination pointer is aligned to that boundary.
5432 </p>
5436 <p>
5438 the
5439 destination location. If the argument is known to be aligned to some boundary,
5440 this can be specified as the fourth argument, otherwise it should be set to 0 or
5441 1.
5442 </p>
5443 </div>
5446 <!-- _______________________________________________________________________ -->
5449 </div>
5455 floating point or vector of floating point type. Not all targets support all
5456 types however.</p>
5457 <pre>
5458 declare float @llvm.sqrt.f32(float %Val)
5459 declare double @llvm.sqrt.f64(double %Val)
5460 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
5461 declare fp128 @llvm.sqrt.f128(fp128 %Val)
5462 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
5463 </pre>
5467 <p>
5471 negative numbers other than -0.0 (which allows for better optimization, because
5473 defined to return -0.0 like IEEE sqrt.
5474 </p>
5478 <p>
5479 The argument and return value are floating point numbers of the same type.
5480 </p>
5484 <p>
5485 This function returns the sqrt of the specified operand if it is a nonnegative
5486 floating point number.
5487 </p>
5488 </div>
5490 <!-- _______________________________________________________________________ -->
5493 </div>
5499 floating point or vector of floating point type. Not all targets support all
5500 types however.</p>
5501 <pre>
5502 declare float @llvm.powi.f32(float %Val, i32 %power)
5503 declare double @llvm.powi.f64(double %Val, i32 %power)
5504 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
5505 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
5506 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
5507 </pre>
5511 <p>
5513 specified (positive or negative) power. The order of evaluation of
5514 multiplications is not defined. When a vector of floating point type is
5515 used, the second argument remains a scalar integer value.
5516 </p>
5520 <p>
5521 The second argument is an integer power, and the first is a value to raise to
5522 that power.
5523 </p>
5527 <p>
5528 This function returns the first value raised to the second power with an
5529 unspecified sequence of rounding operations.</p>
5530 </div>
5532 <!-- _______________________________________________________________________ -->
5535 </div>
5541 floating point or vector of floating point type. Not all targets support all
5542 types however.</p>
5543 <pre>
5544 declare float @llvm.sin.f32(float %Val)
5545 declare double @llvm.sin.f64(double %Val)
5546 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
5547 declare fp128 @llvm.sin.f128(fp128 %Val)
5548 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
5549 </pre>
5553 <p>
5555 </p>
5559 <p>
5560 The argument and return value are floating point numbers of the same type.
5561 </p>
5565 <p>
5566 This function returns the sine of the specified operand, returning the
5568 conditions in the same way.</p>
5569 </div>
5571 <!-- _______________________________________________________________________ -->
5574 </div>
5580 floating point or vector of floating point type. Not all targets support all
5581 types however.</p>
5582 <pre>
5583 declare float @llvm.cos.f32(float %Val)
5584 declare double @llvm.cos.f64(double %Val)
5585 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
5586 declare fp128 @llvm.cos.f128(fp128 %Val)
5587 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
5588 </pre>
5592 <p>
5594 </p>
5598 <p>
5599 The argument and return value are floating point numbers of the same type.
5600 </p>
5604 <p>
5605 This function returns the cosine of the specified operand, returning the
5607 conditions in the same way.</p>
5608 </div>
5610 <!-- _______________________________________________________________________ -->
5613 </div>
5619 floating point or vector of floating point type. Not all targets support all
5620 types however.</p>
5621 <pre>
5622 declare float @llvm.pow.f32(float %Val, float %Power)
5623 declare double @llvm.pow.f64(double %Val, double %Power)
5624 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
5625 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
5626 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
5627 </pre>
5631 <p>
5633 specified (positive or negative) power.
5634 </p>
5638 <p>
5639 The second argument is a floating point power, and the first is a value to
5640 raise to that power.
5641 </p>
5645 <p>
5646 This function returns the first value raised to the second power,
5647 returning the
5649 conditions in the same way.</p>
5650 </div>
5653 <!-- ======================================================================= -->
5656 </div>
5659 <p>
5660 LLVM provides intrinsics for a few important bit manipulation operations.
5661 These allow efficient code generation for some algorithms.
5662 </p>
5664 </div>
5666 <!-- _______________________________________________________________________ -->
5669 </div>
5674 <p>This is an overloaded intrinsic function. You can use bswap on any integer
5676 <pre>
5680 </pre>
5684 <p>
5686 values with an even number of bytes (positive multiple of 16 bits). These are
5687 useful for performing operations on data that is not in the target's native
5688 byte order.
5689 </p>
5693 <p>
5696 intrinsic returns an i32 value that has the four bytes of the input i32
5701 </p>
5703 </div>
5705 <!-- _______________________________________________________________________ -->
5708 </div>
5713 <p>This is an overloaded intrinsic. You can use llvm.ctpop on any integer bit
5714 width. Not all targets support all bit widths however.</p>
5715 <pre>
5721 </pre>
5725 <p>
5727 value.
5728 </p>
5732 <p>
5733 The only argument is the value to be counted. The argument may be of any
5734 integer type. The return type must match the argument type.
5735 </p>
5739 <p>
5741 </p>
5742 </div>
5744 <!-- _______________________________________________________________________ -->
5747 </div>
5753 integer bit width. Not all targets support all bit widths however.</p>
5754 <pre>
5760 </pre>
5764 <p>
5766 leading zeros in a variable.
5767 </p>
5771 <p>
5772 The only argument is the value to be counted. The argument may be of any
5773 integer type. The return type must match the argument type.
5774 </p>
5778 <p>
5780 in a variable. If the src == 0 then the result is the size in bits of the type
5782 </p>
5783 </div>
5787 <!-- _______________________________________________________________________ -->
5790 </div>
5796 integer bit width. Not all targets support all bit widths however.</p>
5797 <pre>
5803 </pre>
5807 <p>
5809 trailing zeros.
5810 </p>
5814 <p>
5815 The only argument is the value to be counted. The argument may be of any
5816 integer type. The return type must match the argument type.
5817 </p>
5821 <p>
5823 in a variable. If the src == 0 then the result is the size in bits of the type
5825 </p>
5826 </div>
5828 <!-- _______________________________________________________________________ -->
5831 </div>
5837 on any integer bit width.</p>
5838 <pre>
5839 declare i17 @llvm.part.select.i17 (i17 %val, i32 %loBit, i32 %hiBit)
5840 declare i29 @llvm.part.select.i29 (i29 %val, i32 %loBit, i32 %hiBit)
5841 </pre>
5845 range of bits from an integer value and returns them in the same bit width as
5846 the original value.</p>
5850 any bit width but they must have the same bit width. The second and third
5857 operates in forward mode.</p>
5861 <ol>
5865 to determine the number of bits to retain.</li>
5868 </ol>
5869 <p>In reverse mode, a similar computation is made except that the bits are
5874 </div>
5878 </div>
5884 on any integer bit width.</p>
5885 <pre>
5886 declare i17 @llvm.part.set.i17.i9 (i17 %val, i9 %repl, i32 %lo, i32 %hi)
5887 declare i29 @llvm.part.set.i29.i9 (i29 %val, i9 %repl, i32 %lo, i32 %hi)
5888 </pre>
5892 of bits in an integer value with another integer value. It returns the integer
5893 with the replaced bits.</p>
5900 type since they specify only a bit index.</p>
5906 operates in forward mode.</p>
5909 truncating it down to the size of the replacement area or zero extending it
5910 up to that size.</p>
5917 <p>In reverse mode, a similar computation is made except that the bits are
5923 <pre>
5929 </pre>
5931 </div>
5933 <!-- ======================================================================= -->
5936 </div>
5939 <p>
5940 LLVM provides intrinsics for some arithmetic with overflow operations.
5941 </p>
5943 </div>
5945 <!-- _______________________________________________________________________ -->
5948 </div>
5955 on any integer bit width.</p>
5957 <pre>
5958 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
5959 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
5960 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
5961 </pre>
5966 a signed addition of the two arguments, and indicate whether an overflow
5967 occurred during the signed summation.</p>
5971 <p>The arguments (%a and %b) and the first element of the result structure may
5972 be of integer types of any bit width, but they must have the same bit width. The
5979 a signed addition of the two variables. They return a structure — the
5980 first element of which is the signed summation, and the second element of which
5981 is a bit specifying if the signed summation resulted in an overflow.</p>
5984 <pre>
5985 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
5986 %sum = extractvalue {i32, i1} %res, 0
5987 %obit = extractvalue {i32, i1} %res, 1
5988 br i1 %obit, label %overflow, label %normal
5989 </pre>
5991 </div>
5993 <!-- _______________________________________________________________________ -->
5996 </div>
6003 on any integer bit width.</p>
6005 <pre>
6006 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
6007 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
6008 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
6009 </pre>
6014 an unsigned addition of the two arguments, and indicate whether a carry occurred
6015 during the unsigned summation.</p>
6019 <p>The arguments (%a and %b) and the first element of the result structure may
6020 be of integer types of any bit width, but they must have the same bit width. The
6027 an unsigned addition of the two arguments. They return a structure — the
6028 first element of which is the sum, and the second element of which is a bit
6029 specifying if the unsigned summation resulted in a carry.</p>
6032 <pre>
6033 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
6034 %sum = extractvalue {i32, i1} %res, 0
6035 %obit = extractvalue {i32, i1} %res, 1
6036 br i1 %obit, label %carry, label %normal
6037 </pre>
6039 </div>
6041 <!-- _______________________________________________________________________ -->
6044 </div>
6051 on any integer bit width.</p>
6053 <pre>
6054 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
6055 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
6056 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
6057 </pre>
6062 a signed subtraction of the two arguments, and indicate whether an overflow
6063 occurred during the signed subtraction.</p>
6067 <p>The arguments (%a and %b) and the first element of the result structure may
6068 be of integer types of any bit width, but they must have the same bit width. The
6075 a signed subtraction of the two arguments. They return a structure — the
6076 first element of which is the subtraction, and the second element of which is a bit
6077 specifying if the signed subtraction resulted in an overflow.</p>
6080 <pre>
6081 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
6082 %sum = extractvalue {i32, i1} %res, 0
6083 %obit = extractvalue {i32, i1} %res, 1
6084 br i1 %obit, label %overflow, label %normal
6085 </pre>
6087 </div>
6089 <!-- _______________________________________________________________________ -->
6092 </div>
6099 on any integer bit width.</p>
6101 <pre>
6102 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
6103 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
6104 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
6105 </pre>
6110 an unsigned subtraction of the two arguments, and indicate whether an overflow
6111 occurred during the unsigned subtraction.</p>
6115 <p>The arguments (%a and %b) and the first element of the result structure may
6116 be of integer types of any bit width, but they must have the same bit width. The
6123 an unsigned subtraction of the two arguments. They return a structure — the
6124 first element of which is the subtraction, and the second element of which is a bit
6125 specifying if the unsigned subtraction resulted in an overflow.</p>
6128 <pre>
6129 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
6130 %sum = extractvalue {i32, i1} %res, 0
6131 %obit = extractvalue {i32, i1} %res, 1
6132 br i1 %obit, label %overflow, label %normal
6133 </pre>
6135 </div>
6137 <!-- _______________________________________________________________________ -->
6140 </div>
6147 on any integer bit width.</p>
6149 <pre>
6150 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
6151 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
6152 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
6153 </pre>
6158 a signed multiplication of the two arguments, and indicate whether an overflow
6159 occurred during the signed multiplication.</p>
6163 <p>The arguments (%a and %b) and the first element of the result structure may
6164 be of integer types of any bit width, but they must have the same bit width. The
6171 a signed multiplication of the two arguments. They return a structure —
6172 the first element of which is the multiplication, and the second element of
6173 which is a bit specifying if the signed multiplication resulted in an
6174 overflow.</p>
6177 <pre>
6178 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
6179 %sum = extractvalue {i32, i1} %res, 0
6180 %obit = extractvalue {i32, i1} %res, 1
6181 br i1 %obit, label %overflow, label %normal
6182 </pre>
6184 </div>
6186 <!-- _______________________________________________________________________ -->
6189 </div>
6196 on any integer bit width.</p>
6198 <pre>
6199 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
6200 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
6201 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
6202 </pre>
6207 actively being fixed, but it should not currently be used!</i></p>
6210 a unsigned multiplication of the two arguments, and indicate whether an overflow
6211 occurred during the unsigned multiplication.</p>
6215 <p>The arguments (%a and %b) and the first element of the result structure may
6216 be of integer types of any bit width, but they must have the same bit width. The
6219 multiplication.</p>
6224 an unsigned multiplication of the two arguments. They return a structure —
6225 the first element of which is the multiplication, and the second element of
6226 which is a bit specifying if the unsigned multiplication resulted in an
6227 overflow.</p>
6230 <pre>
6231 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
6232 %sum = extractvalue {i32, i1} %res, 0
6233 %obit = extractvalue {i32, i1} %res, 1
6234 br i1 %obit, label %overflow, label %normal
6235 </pre>
6237 </div>
6239 <!-- ======================================================================= -->
6242 </div>
6245 <p>
6247 are described in the <a
6249 Debugging</a> document.
6250 </p>
6251 </div>
6254 <!-- ======================================================================= -->
6257 </div>
6260 <p> The LLVM exception handling intrinsics (which all start with
6264 </div>
6266 <!-- ======================================================================= -->
6269 </div>
6272 <p>
6273 This intrinsic makes it possible to excise one parameter, marked with
6275 function pointer lacking the nest parameter - the caller does not need
6276 to provide a value for it. Instead, the value to use is stored in
6277 advance in a "trampoline", a block of memory usually allocated
6278 on the stack, which also contains code to splice the nest value into the
6279 argument list. This is used to implement the GCC nested function address
6280 extension.
6281 </p>
6282 <p>
6283 For example, if the function is
6286 <pre>
6289 %p = call i8* @llvm.init.trampoline( i8* %tramp1, i8* bitcast (i32 (i8* nest , i32, i32)* @f to i8*), i8* %nval )
6290 %fp = bitcast i8* %p to i32 (i32, i32)*
6291 </pre>
6294 </div>
6296 <!-- _______________________________________________________________________ -->
6299 </div>
6302 <pre>
6304 </pre>
6306 <p>
6308 and returns a function pointer suitable for executing it.
6309 </p>
6311 <p>
6314 and sufficiently aligned block of memory; this memory is written to by the
6315 intrinsic. Note that the size and the alignment are target-specific - LLVM
6316 currently provides no portable way of determining them, so a front-end that
6317 generates this intrinsic needs to have some target-specific knowledge.
6319 </p>
6321 <p>
6323 dependent code, turning it into a function. A pointer to this function is
6324 returned, but needs to be bitcast to an
6326 before being called. The new function's signature is the same as that of
6329 of pointer type. Calling the new function is equivalent to calling
6333 modified, then the effect of any later call to the returned function pointer is
6334 undefined.
6335 </p>
6336 </div>
6338 <!-- ======================================================================= -->
6341 </div>
6344 <p>
6345 These intrinsic functions expand the "universal IR" of LLVM to represent
6346 hardware constructs for atomic operations and memory synchronization. This
6347 provides an interface to the hardware, not an interface to the programmer. It
6348 is aimed at a low enough level to allow any programming models or APIs
6349 (Application Programming Interfaces) which
6350 need atomic behaviors to map cleanly onto it. It is also modeled primarily on
6351 hardware behavior. Just as hardware provides a "universal IR" for source
6352 languages, it also provides a starting point for developing a "universal"
6353 atomic operation and synchronization IR.
6354 </p>
6355 <p>
6357 software transaction memory systems, atomic primitives, and intrinsic
6358 functions as found in BSD, GNU libc, atomic_ops, APR, and other system and
6359 application libraries. The hardware interface provided by LLVM should allow
6360 a clean implementation of all of these APIs and parallel programming models.
6361 No one model or paradigm should be selected above others unless the hardware
6362 itself ubiquitously does so.
6364 </p>
6365 </div>
6367 <!-- _______________________________________________________________________ -->
6370 </div>
6373 <pre>
6377 </pre>
6379 <p>
6381 specific pairs of memory access types.
6382 </p>
6384 <p>
6386 The first four arguments enables a specific barrier as listed below. The fith
6387 argument specifies that the barrier applies to io or device or uncached memory.
6389 </p>
6390 <ul>
6396 </ul>
6398 <p>
6399 This intrinsic causes the system to enforce some ordering constraints upon
6400 the loads and stores of the program. This barrier does not indicate
6402 which they occur. For any of the specified pairs of load and store operations
6403 (f.ex. load-load, or store-load), all of the first operations preceding the
6404 barrier will complete before any of the second operations succeeding the
6405 barrier begin. Specifically the semantics for each pairing is as follows:
6406 </p>
6407 <ul>
6409 after the barrier begins.</li>
6412 store after the barrier begins.</li>
6414 store after the barrier begins.</li>
6416 load after the barrier begins.</li>
6417 </ul>
6418 <p>
6419 These semantics are applied with a logical "and" behavior when more than one
6420 is enabled in a single memory barrier intrinsic.
6421 </p>
6422 <p>
6423 Backends may implement stronger barriers than those requested when they do not
6424 support as fine grained a barrier as requested. Some architectures do not
6425 need all types of barriers and on such architectures, these become noops.
6426 </p>
6428 <pre>
6429 %ptr = malloc i32
6430 store i32 4, %ptr
6433 call void @llvm.memory.barrier( i1 false, i1 true, i1 false, i1 false )
6436 </pre>
6437 </div>
6439 <!-- _______________________________________________________________________ -->
6442 </div>
6445 <p>
6447 any integer bit width and for different address spaces. Not all targets
6448 support all bit widths however.</p>
6450 <pre>
6452 declare i16 @llvm.atomic.cmp.swap.i16.p0i16( i16* <ptr>, i16 <cmp>, i16 <val> )
6453 declare i32 @llvm.atomic.cmp.swap.i32.p0i32( i32* <ptr>, i32 <cmp>, i32 <val> )
6454 declare i64 @llvm.atomic.cmp.swap.i64.p0i64( i64* <ptr>, i64 <cmp>, i64 <val> )
6456 </pre>
6458 <p>
6459 This loads a value in memory and compares it to a given value. If they are
6460 equal, it stores a new value into the memory.
6461 </p>
6463 <p>
6467 this integer type. While any bit width integer may be used, targets may only
6468 lower representations they support in hardware.
6470 </p>
6472 <p>
6473 This entire intrinsic must be executed atomically. It first loads the value
6476 loaded value is yielded in all cases. This provides the equivalent of an
6477 atomic compare-and-swap operation within the SSA framework.
6478 </p>
6481 <pre>
6482 %ptr = malloc i32
6483 store i32 4, %ptr
6486 %result1 = call i32 @llvm.atomic.cmp.swap.i32.p0i32( i32* %ptr, i32 4, %val1 )
6492 %result2 = call i32 @llvm.atomic.cmp.swap.i32.p0i32( i32* %ptr, i32 5, %val2 )
6497 </pre>
6498 </div>
6500 <!-- _______________________________________________________________________ -->
6503 </div>
6507 <p>
6509 integer bit width. Not all targets support all bit widths however.</p>
6510 <pre>
6516 </pre>
6518 <p>
6522 </p>
6525 <p>
6529 integer type. The targets may only lower integer representations they
6530 support.
6531 </p>
6533 <p>
6536 equivalent of an atomic swap operation within the SSA framework.
6538 </p>
6540 <pre>
6541 %ptr = malloc i32
6542 store i32 4, %ptr
6545 %result1 = call i32 @llvm.atomic.swap.i32.p0i32( i32* %ptr, i32 %val1 )
6551 %result2 = call i32 @llvm.atomic.swap.i32.p0i32( i32* %ptr, i32 %val2 )
6556 </pre>
6557 </div>
6559 <!-- _______________________________________________________________________ -->
6563 </div>
6566 <p>
6568 integer bit width. Not all targets support all bit widths however.</p>
6569 <pre>
6575 </pre>
6577 <p>
6580 </p>
6582 <p>
6584 The intrinsic takes two arguments, the first a pointer to an integer value
6585 and the second an integer value. The result is also an integer value. These
6586 integer types can have any bit width, but they must all have the same bit
6587 width. The targets may only lower integer representations they support.
6588 </p>
6590 <p>
6591 This intrinsic does a series of operations atomically. It first loads the
6594 </p>
6597 <pre>
6598 %ptr = malloc i32
6599 store i32 4, %ptr
6600 %result1 = call i32 @llvm.atomic.load.add.i32.p0i32( i32* %ptr, i32 4 )
6602 %result2 = call i32 @llvm.atomic.load.add.i32.p0i32( i32* %ptr, i32 2 )
6604 %result3 = call i32 @llvm.atomic.load.add.i32.p0i32( i32* %ptr, i32 5 )
6607 </pre>
6608 </div>
6610 <!-- _______________________________________________________________________ -->
6614 </div>
6617 <p>
6619 any integer bit width and for different address spaces. Not all targets
6620 support all bit widths however.</p>
6621 <pre>
6627 </pre>
6629 <p>
6632 </p>
6634 <p>
6636 The intrinsic takes two arguments, the first a pointer to an integer value
6637 and the second an integer value. The result is also an integer value. These
6638 integer types can have any bit width, but they must all have the same bit
6639 width. The targets may only lower integer representations they support.
6640 </p>
6642 <p>
6643 This intrinsic does a series of operations atomically. It first loads the
6646 </p>
6649 <pre>
6650 %ptr = malloc i32
6651 store i32 8, %ptr
6652 %result1 = call i32 @llvm.atomic.load.sub.i32.p0i32( i32* %ptr, i32 4 )
6654 %result2 = call i32 @llvm.atomic.load.sub.i32.p0i32( i32* %ptr, i32 2 )
6656 %result3 = call i32 @llvm.atomic.load.sub.i32.p0i32( i32* %ptr, i32 5 )
6659 </pre>
6660 </div>
6662 <!-- _______________________________________________________________________ -->
6669 </div>
6672 <p>
6676 address spaces. Not all targets support all bit widths however.</p>
6677 <pre>
6683 </pre>
6685 <pre>
6691 </pre>
6693 <pre>
6699 </pre>
6701 <pre>
6707 </pre>
6709 <p>
6713 </p>
6715 <p>
6717 These intrinsics take two arguments, the first a pointer to an integer value
6718 and the second an integer value. The result is also an integer value. These
6719 integer types can have any bit width, but they must all have the same bit
6720 width. The targets may only lower integer representations they support.
6721 </p>
6723 <p>
6724 These intrinsics does a series of operations atomically. They first load the
6728 </p>
6731 <pre>
6732 %ptr = malloc i32
6733 store i32 0x0F0F, %ptr
6734 %result0 = call i32 @llvm.atomic.load.nand.i32.p0i32( i32* %ptr, i32 0xFF )
6736 %result1 = call i32 @llvm.atomic.load.and.i32.p0i32( i32* %ptr, i32 0xFF )
6738 %result2 = call i32 @llvm.atomic.load.or.i32.p0i32( i32* %ptr, i32 0F )
6740 %result3 = call i32 @llvm.atomic.load.xor.i32.p0i32( i32* %ptr, i32 0F )
6743 </pre>
6744 </div>
6747 <!-- _______________________________________________________________________ -->
6754 </div>
6757 <p>
6761 address spaces. Not all targets
6762 support all bit widths however.</p>
6763 <pre>
6769 </pre>
6771 <pre>
6777 </pre>
6779 <pre>
6785 </pre>
6787 <pre>
6793 </pre>
6795 <p>
6796 These intrinsics takes the signed or unsigned minimum or maximum of
6799 </p>
6801 <p>
6803 These intrinsics take two arguments, the first a pointer to an integer value
6804 and the second an integer value. The result is also an integer value. These
6805 integer types can have any bit width, but they must all have the same bit
6806 width. The targets may only lower integer representations they support.
6807 </p>
6809 <p>
6810 These intrinsics does a series of operations atomically. They first load the
6814 </p>
6817 <pre>
6818 %ptr = malloc i32
6819 store i32 7, %ptr
6820 %result0 = call i32 @llvm.atomic.load.min.i32.p0i32( i32* %ptr, i32 -2 )
6822 %result1 = call i32 @llvm.atomic.load.max.i32.p0i32( i32* %ptr, i32 8 )
6824 %result2 = call i32 @llvm.atomic.load.umin.i32.p0i32( i32* %ptr, i32 10 )
6826 %result3 = call i32 @llvm.atomic.load.umax.i32.p0i32( i32* %ptr, i32 30 )
6829 </pre>
6830 </div>
6832 <!-- ======================================================================= -->
6835 </div>
6838 <p> This class of intrinsics is designed to be generic and has
6839 no specific purpose. </p>
6840 </div>
6842 <!-- _______________________________________________________________________ -->
6845 </div>
6850 <pre>
6851 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int> )
6852 </pre>
6856 <p>
6858 </p>
6862 <p>
6863 The first argument is a pointer to a value, the second is a pointer to a
6864 global string, the third is a pointer to a global string which is the source
6865 file name, and the last argument is the line number.
6866 </p>
6870 <p>
6871 This intrinsic allows annotation of local variables with arbitrary strings.
6872 This can be useful for special purpose optimizations that want to look for these
6873 annotations. These have no other defined use, they are ignored by code
6874 generation and optimization.
6875 </p>
6876 </div>
6878 <!-- _______________________________________________________________________ -->
6881 </div>
6887 any integer bit width.
6888 </p>
6889 <pre>
6890 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int> )
6891 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int> )
6892 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int> )
6893 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int> )
6894 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int> )
6895 </pre>
6899 <p>
6901 </p>
6905 <p>
6906 The first argument is an integer value (result of some expression),
6907 the second is a pointer to a global string, the third is a pointer to a global
6908 string which is the source file name, and the last argument is the line number.
6909 It returns the value of the first argument.
6910 </p>
6914 <p>
6915 This intrinsic allows annotations to be put on arbitrary expressions
6916 with arbitrary strings. This can be useful for special purpose optimizations
6917 that want to look for these annotations. These have no other defined use, they
6918 are ignored by code generation and optimization.
6919 </p>
6920 </div>
6922 <!-- _______________________________________________________________________ -->
6925 </div>
6930 <pre>
6931 declare void @llvm.trap()
6932 </pre>
6936 <p>
6938 </p>
6942 <p>
6943 None
6944 </p>
6948 <p>
6949 This intrinsics is lowered to the target dependent trap instruction. If the
6950 target does not have a trap instruction, this intrinsic will be lowered to the
6951 call of the abort() function.
6952 </p>
6953 </div>
6955 <!-- _______________________________________________________________________ -->
6958 </div>
6961 <pre>
6964 </pre>
6966 <p>
6969 it is placed on the stack before local variables.
6970 </p>
6972 <p>
6974 first argument is the value loaded from the stack guard
6976 has enough space to hold the value of the guard.
6977 </p>
6979 <p>
6980 This intrinsic causes the prologue/epilogue inserter to force the position of
6982 stack. This is to ensure that if a local variable on the stack is overwritten,
6983 it will destroy the value of the guard. When the function exits, the guard on
6984 the stack is checked against the original guard. If they're different, then
6986 </p>
6987 </div>
6989 <!-- *********************************************************************** -->
6990 <hr>
6991 <address>
6999 Last modified: $Date$
7000 </address>
7002 </body>
7003 </html>