| blob | 3f16e386c2beac5585b610cd89d758c930a9a7d8 |
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>
69 </ol>
70 </li>
72 <ol>
74 </ol>
75 </li>
77 <ol>
79 <ol>
86 </ol>
87 </li>
89 <ol>
99 </ol>
100 </li>
102 <ol>
109 </ol>
110 </li>
112 <ol>
116 </ol>
117 </li>
119 <ol>
122 </ol>
123 </li>
125 <ol>
132 </ol>
133 </li>
135 <ol>
148 </ol>
149 </li>
151 <ol>
160 </ol>
161 </li>
162 </ol>
163 </li>
165 <ol>
167 <ol>
171 </ol>
172 </li>
174 <ol>
178 </ol>
179 </li>
181 <ol>
189 </ol>
190 </li>
192 <ol>
201 </ol>
202 </li>
204 <ol>
211 </ol>
212 </li>
214 <ol>
221 </ol>
222 </li>
226 <ol>
228 </ol>
229 </li>
231 <ol>
245 </ol>
246 </li>
248 <ol>
257 </ol>
258 </li>
259 </ol>
260 </li>
261 </ol>
266 </div>
268 <!-- *********************************************************************** -->
270 <!-- *********************************************************************** -->
273 <p>This document is a reference manual for the LLVM assembly language.
274 LLVM is a Static Single Assignment (SSA) based representation that provides
275 type safety, low-level operations, flexibility, and the capability of
276 representing 'all' high-level languages cleanly. It is the common code
277 representation used throughout all phases of the LLVM compilation
278 strategy.</p>
279 </div>
281 <!-- *********************************************************************** -->
283 <!-- *********************************************************************** -->
287 <p>The LLVM code representation is designed to be used in three
288 different forms: as an in-memory compiler IR, as an on-disk bitcode
289 representation (suitable for fast loading by a Just-In-Time compiler),
290 and as a human readable assembly language representation. This allows
291 LLVM to provide a powerful intermediate representation for efficient
292 compiler transformations and analysis, while providing a natural means
293 to debug and visualize the transformations. The three different forms
294 of LLVM are all equivalent. This document describes the human readable
295 representation and notation.</p>
297 <p>The LLVM representation aims to be light-weight and low-level
298 while being expressive, typed, and extensible at the same time. It
299 aims to be a "universal IR" of sorts, by being at a low enough level
300 that high-level ideas may be cleanly mapped to it (similar to how
301 microprocessors are "universal IR's", allowing many source languages to
302 be mapped to them). By providing type information, LLVM can be used as
303 the target of optimizations: for example, through pointer analysis, it
304 can be proven that a C automatic variable is never accessed outside of
305 the current function... allowing it to be promoted to a simple SSA
306 value instead of a memory location.</p>
308 </div>
310 <!-- _______________________________________________________________________ -->
315 <p>It is important to note that this document describes 'well formed'
316 LLVM assembly language. There is a difference between what the parser
317 accepts and what is considered 'well formed'. For example, the
318 following instruction is syntactically okay, but not well formed:</p>
321 <pre>
323 </pre>
324 </div>
327 its uses. The LLVM infrastructure provides a verification pass that may
328 be used to verify that an LLVM module is well formed. This pass is
329 automatically run by the parser after parsing input assembly and by
330 the optimizer before it outputs bitcode. The violations pointed out
331 by the verifier pass indicate bugs in transformation passes or input to
332 the parser.</p>
333 </div>
335 <!-- Describe the typesetting conventions here. -->
337 <!-- *********************************************************************** -->
339 <!-- *********************************************************************** -->
343 <p>LLVM identifiers come in two basic types: global and local. Global
344 identifiers (functions, global variables) begin with the @ character. Local
345 identifiers (register names, types) begin with the % character. Additionally,
346 there are three different formats for identifiers, for different purposes:</p>
348 <ol>
349 <li>Named values are represented as a string of characters with their prefix.
350 For example, %foo, @DivisionByZero, %a.really.long.identifier. The actual
352 Identifiers which require other characters in their names can be surrounded
353 with quotes. Special characters may be escaped using "\xx" where xx is the
354 ASCII code for the character in hexadecimal. In this way, any character can
355 be used in a name value, even quotes themselves.
357 <li>Unnamed values are represented as an unsigned numeric value with their
362 </ol>
364 <p>LLVM requires that values start with a prefix for two reasons: Compilers
365 don't need to worry about name clashes with reserved words, and the set of
366 reserved words may be expanded in the future without penalty. Additionally,
367 unnamed identifiers allow a compiler to quickly come up with a temporary
368 variable without having to avoid symbol table conflicts.</p>
370 <p>Reserved words in LLVM are very similar to reserved words in other
371 languages. There are keywords for different opcodes
376 and others. These reserved words cannot conflict with variable names, because
377 none of them start with a prefix character ('%' or '@').</p>
379 <p>Here is an example of LLVM code to multiply the integer variable
385 <pre>
387 </pre>
388 </div>
393 <pre>
395 </pre>
396 </div>
401 <pre>
405 </pre>
406 </div>
409 important lexical features of LLVM:</p>
411 <ol>
414 line.</li>
416 <li>Unnamed temporaries are created when the result of a computation is not
417 assigned to a named value.</li>
421 </ol>
423 <p>...and it also shows a convention that we follow in this document. When
424 demonstrating instructions, we will follow an instruction with a comment that
425 defines the type and name of value produced. Comments are shown in italic
426 text.</p>
428 </div>
430 <!-- *********************************************************************** -->
432 <!-- *********************************************************************** -->
434 <!-- ======================================================================= -->
436 </div>
441 translation unit of the input programs. Each module consists of
442 functions, global variables, and symbol table entries. Modules may be
443 combined together with the LLVM linker, which merges function (and
444 global variable) definitions, resolves forward declarations, and merges
450 href="#globalvars">constant</a> <a href="#t_array">[13 x i8]</a> c"hello world\0A\00" <i>; [13 x i8]*</i>
458 %cast210 = <a
462 <a
464 <a
466 </pre>
467 </div>
474 <p>In general, a module is made up of a list of global values,
475 where both functions and global variables are global values. Global values are
476 represented by a pointer to a memory location (in this case, a pointer to an
477 array of char, and a pointer to a function), and have one of the following <a
480 </div>
482 <!-- ======================================================================= -->
485 </div>
489 <p>
490 All Global Variables and Functions have one of the following types of linkage:
491 </p>
493 <dl>
497 <dd>Global values with private linkage are only directly accessible by
498 objects in the current module. In particular, linking code into a module with
499 an private global value may cause the private to be renamed as necessary to
500 avoid collisions. Because the symbol is private to the module, all
501 references can be updated. This doesn't show up in any symbol table in the
502 object file.
503 </dd>
507 <dd> Similar to private, but the value shows as a local symbol (STB_LOCAL in
508 the case of ELF) in the object file. This corresponds to the notion of the
510 </dd>
513 </dt>
516 into the object file corresponding to the LLVM module. They exist to
517 allow inlining and other optimizations to take place given knowledge of the
518 definition of the global, which is known to be somewhere outside the module.
521 type is only allowed on definitions, not declarations.</dd>
526 the same name when linkage occurs. This is typically used to implement
527 inline functions, templates, or other code which must be generated in each
529 allowed to be discarded.
530 </dd>
536 discarded. This is used for globals that may be emitted in multiple
537 translation units, but that are not guaranteed to be emitted into every
538 translation unit that uses them. One example of this is tentative
539 definitions in C, such as "<tt>int X;</tt>" at global scope.
540 </dd>
545 that some targets may choose to emit different assembly sequences for them
546 for target-dependent reasons. This is used for globals that are declared
547 "weak" in C source code.
548 </dd>
553 pointer to array type. When two global variables with appending linkage are
554 linked together, the two global arrays are appended together. This is the
555 LLVM, typesafe, equivalent of having the system linker append together
556 "sections" with identical names when .o files are linked.
557 </dd>
561 <dd>The semantics of this linkage follow the ELF object file model: the
562 symbol is weak until linked, if not linked, the symbol becomes null instead
563 of being an undefined reference.
564 </dd>
568 <dd>Some languages allow differing globals to be merged, such as two
570 ensure that only equivalent globals are ever merged (the "one definition
573 be merged with equivalent globals. These linkage types are otherwise the
575 </dd>
579 <dd>If none of the above identifiers are used, the global is externally
580 visible, meaning that it participates in linkage and can be used to resolve
581 external symbol references.
582 </dd>
583 </dl>
585 <p>
586 The next two types of linkage are targeted for Microsoft Windows platform
587 only. They are designed to support importing (exporting) symbols from (to)
588 DLLs (Dynamic Link Libraries).
589 </p>
591 <dl>
595 or variable via a global pointer to a pointer that is set up by the DLL
596 exporting the symbol. On Microsoft Windows targets, the pointer name is
598 </dd>
603 pointer to a pointer in a DLL, so that it can be referenced with the
606 name.
607 </dd>
609 </dl>
612 variable is defined to be internal, if another module defined a "<tt>.LC0</tt>"
613 variable and was linked with this one, one of the two would be renamed,
615 external (i.e., lacking any linkage declarations), they are accessible
616 outside of the current module.</p>
622 </div>
624 <!-- ======================================================================= -->
627 </div>
633 specified for the call. The calling convention of any pair of dynamic
634 caller/callee must match, or the behavior of the program is undefined. The
635 following calling conventions are supported by LLVM, and more may be added in
636 the future:</p>
638 <dl>
641 <dd>This calling convention (the default if no other calling convention is
642 specified) matches the target C calling conventions. This calling convention
643 supports varargs function calls and tolerates some mismatch in the declared
644 prototype and implemented declaration of the function (as does normal C).
645 </dd>
649 <dd>This calling convention attempts to make calls as fast as possible
650 (e.g. by passing things in registers). This calling convention allows the
651 target to use whatever tricks it wants to produce fast code for the target,
652 without having to conform to an externally specified ABI (Application Binary
653 Interface). Implementations of this convention should allow arbitrary
655 supported. This calling convention does not support varargs and requires the
656 prototype of all callees to exactly match the prototype of the function
657 definition.
658 </dd>
662 <dd>This calling convention attempts to make code in the caller as efficient
663 as possible under the assumption that the call is not commonly executed. As
664 such, these calls often preserve all registers so that the call does not break
665 any live ranges in the caller side. This calling convention does not support
666 varargs and requires the prototype of all callees to exactly match the
667 prototype of the function definition.
668 </dd>
672 <dd>Any calling convention may be specified by number, allowing
673 target-specific calling conventions to be used. Target specific calling
674 conventions start at 64.
675 </dd>
676 </dl>
678 <p>More calling conventions can be added/defined on an as-needed basis, to
679 support pascal conventions or any other well-known target-independent
680 convention.</p>
682 </div>
684 <!-- ======================================================================= -->
687 </div>
691 <p>
692 All Global Variables and Functions have one of the following visibility styles:
693 </p>
695 <dl>
698 <dd>On targets that use the ELF object file format, default visibility means
699 that the declaration is visible to other
700 modules and, in shared libraries, means that the declared entity may be
701 overridden. On Darwin, default visibility means that the declaration is
702 visible to other modules. Default visibility corresponds to "external
703 linkage" in the language.
704 </dd>
708 <dd>Two declarations of an object with hidden visibility refer to the same
709 object if they are in the same shared object. Usually, hidden visibility
710 indicates that the symbol will not be placed into the dynamic symbol table,
711 so no other module (executable or shared library) can reference it
712 directly.
713 </dd>
717 <dd>On ELF, protected visibility indicates that the symbol will be placed in
718 the dynamic symbol table, but that references within the defining module will
719 bind to the local symbol. That is, the symbol cannot be overridden by another
720 module.
721 </dd>
722 </dl>
724 </div>
726 <!-- ======================================================================= -->
729 </div>
733 <p>LLVM IR allows you to specify name aliases for certain types. This can make
734 it easier to read the IR and make the IR more condensed (particularly when
735 recursive types are involved). An example of a name specification is:
736 </p>
739 <pre>
740 %mytype = type { %mytype*, i32 }
741 </pre>
742 </div>
748 <p>Note that type names are aliases for the structural type that they indicate,
749 and that you can therefore specify multiple names for the same type. This often
750 leads to confusing behavior when dumping out a .ll file. Since LLVM IR uses
751 structural typing, the name is not part of the type. When printing out LLVM IR,
753 shape. This means that if you have code where two different source types end up
754 having the same LLVM type, that the dumper will sometimes print the "wrong" or
755 unexpected type. This is an important design point and isn't going to
756 change.</p>
758 </div>
760 <!-- ======================================================================= -->
763 </div>
767 <p>Global variables define regions of memory allocated at compilation time
768 instead of run-time. Global variables may optionally be initialized, may have
769 an explicit section to be placed in, and may have an optional explicit alignment
770 specified. A variable may be defined as "thread_local", which means that it
771 will not be shared by threads (each thread will have a separated copy of the
772 variable). A variable may be defined as a global "constant," which indicates
774 optimization, allowing the global data to be placed in the read-only section of
775 an executable, etc). Note that variables that need runtime initialization
778 <p>
780 constant, even if the final definition of the global is not. This capability
781 can be used to enable slightly better optimization of the program, but requires
782 the language definition to guarantee that optimizations based on the
783 'constantness' are valid for the translation units that do not include the
784 definition.
785 </p>
787 <p>As SSA values, global variables define pointer values that are in
788 scope (i.e. they dominate) all basic blocks in the program. Global
789 variables always define a pointer to their "content" type because they
790 describe a region of memory, and all memory objects in LLVM are
791 accessed through pointers.</p>
793 <p>A global variable may be declared to reside in a target-specifc numbered
794 address space. For targets that support them, address spaces may affect how
795 optimizations are performed and/or what target instructions are used to access
796 the variable. The default address space is zero. The address space qualifier
797 must precede any other attributes.</p>
799 <p>LLVM allows an explicit section to be specified for globals. If the target
800 supports it, it will emit globals to the section specified.</p>
802 <p>An explicit alignment may be specified for a global. If not present, or if
803 the alignment is set to zero, the alignment of the global is set by the target
804 to whatever it feels convenient. If an explicit alignment is specified, the
805 global is forced to have at least that much alignment. All alignments must be
808 <p>For example, the following defines a global in a numbered address space with
809 an initializer, section, and alignment:</p>
812 <pre>
814 </pre>
815 </div>
817 </div>
820 <!-- ======================================================================= -->
823 </div>
832 name, a (possibly empty) argument list (each with optional
836 an opening curly brace, a list of basic blocks, and a closing curly brace.
838 LLVM function declarations consist of the "<tt>declare</tt>" keyword, an
843 name, a possibly empty list of arguments, an optional alignment, and an optional
846 <p>A function definition contains a list of basic blocks, forming the CFG
847 (Control Flow Graph) for
848 the function. Each basic block may optionally start with a label (giving the
849 basic block a symbol table entry), contains a list of instructions, and ends
851 function return).</p>
853 <p>The first basic block in a function is special in two ways: it is immediately
854 executed on entrance to the function, and it is not allowed to have predecessor
855 basic blocks (i.e. there can not be any branches to the entry block of a
856 function). Because the block can have no predecessors, it also cannot have any
859 <p>LLVM allows an explicit section to be specified for functions. If the target
860 supports it, it will emit functions to the section specified.</p>
862 <p>An explicit alignment may be specified for a function. If not present, or if
863 the alignment is set to zero, the alignment of the function is set by the target
864 to whatever it feels convenient. If an explicit alignment is specified, the
865 function is forced to have at least that much alignment. All alignments must be
871 <tt>
877 </tt>
878 </div>
880 </div>
883 <!-- ======================================================================= -->
886 </div>
889 function, global variable, another alias or bitcast of global value). Aliases
896 <pre>
898 </pre>
899 </div>
901 </div>
905 <!-- ======================================================================= -->
908 <p>The return type and each parameter of a function type may have a set of
910 used to communicate additional information about the result or parameters of
911 a function. Parameter attributes are considered to be part of the function,
912 not of the function type, so functions with different parameter attributes
913 can have the same function type.</p>
915 <p>Parameter attributes are simple keywords that follow the type specified. If
916 multiple parameter attributes are needed, they are space separated. For
917 example:</p>
920 <pre>
921 declare i32 @printf(i8* noalias nocapture, ...)
922 declare i32 @atoi(i8 zeroext)
923 declare signext i8 @returns_signed_char()
924 </pre>
925 </div>
931 <dl>
933 <dd>This indicates to the code generator that the parameter or return value
934 should be zero-extended to a 32-bit value by the caller (for a parameter)
935 or the callee (for a return value).</dd>
938 <dd>This indicates to the code generator that the parameter or return value
939 should be sign-extended to a 32-bit value by the caller (for a parameter)
940 or the callee (for a return value).</dd>
943 <dd>This indicates that this parameter or return value should be treated
944 in a special target-dependent fashion during while emitting code for a
945 function call or return (usually, by putting it in a register as opposed
946 to memory, though some targets use it to distinguish between two different
947 kinds of registers). Use of this attribute is target-specific.</dd>
950 <dd>This indicates that the pointer parameter should really be passed by
951 value to the function. The attribute implies that a hidden copy of the
952 pointee is made between the caller and the callee, so the callee is unable
953 to modify the value in the callee. This attribute is only valid on LLVM
954 pointer arguments. It is generally used to pass structs and arrays by
955 value, but is also valid on pointers to scalars. The copy is considered to
956 belong to the caller not the callee (for example,
959 values. The byval attribute also supports specifying an alignment with the
960 align attribute. This has a target-specific effect on the code generator
961 that usually indicates a desired alignment for the synthesized stack
962 slot.</dd>
965 <dd>This indicates that the pointer parameter specifies the address of a
966 structure that is the return value of the function in the source program.
967 This pointer must be guaranteed by the caller to be valid: loads and stores
968 to the structure may be assumed by the callee to not to trap. This may only
969 be applied to the first parameter. This is not a valid attribute for
970 return values. </dd>
973 <dd>This indicates that the pointer does not alias any global or any other
974 parameter. The caller is responsible for ensuring that this is the
976 that the pointer does not alias any other pointers visible to the
977 caller. For further details, please see the discussion of the NoAlias
978 response in
983 <dd>This indicates that the callee does not make any copies of the pointer
984 that outlive the callee itself. This is not a valid attribute for return
985 values.</dd>
988 <dd>This indicates that the pointer parameter can be excised using the
990 attribute for return values.</dd>
991 </dl>
993 </div>
995 <!-- ======================================================================= -->
998 </div>
1001 <p>Each function may specify a garbage collector name, which is simply a
1002 string.</p>
1008 collector which will cause the compiler to alter its output in order to support
1009 the named garbage collection algorithm.</p>
1010 </div>
1012 <!-- ======================================================================= -->
1015 </div>
1019 <p>Function attributes are set to communicate additional information about
1020 a function. Function attributes are considered to be part of the function,
1021 not of the function type, so functions with different parameter attributes
1022 can have the same function type.</p>
1024 <p>Function attributes are simple keywords that follow the type specified. If
1025 multiple attributes are needed, they are space separated. For
1026 example:</p>
1029 <pre>
1030 define void @f() noinline { ... }
1031 define void @f() alwaysinline { ... }
1032 define void @f() alwaysinline optsize { ... }
1033 define void @f() optsize
1034 </pre>
1035 </div>
1037 <dl>
1039 <dd>This attribute indicates that the inliner should attempt to inline this
1040 function into callers whenever possible, ignoring any active inlining size
1041 threshold for this caller.</dd>
1044 <dd>This attribute indicates that the inliner should never inline this function
1045 in any situation. This attribute may not be used together with the
1049 <dd>This attribute suggests that optimization passes and code generator passes
1050 make choices that keep the code size of this function low, and otherwise do
1051 optimizations specifically to reduce code size.</dd>
1054 <dd>This function attribute indicates that the function never returns normally.
1055 This produces undefined behavior at runtime if the function ever does
1056 dynamically return.</dd>
1059 <dd>This function attribute indicates that the function never returns with an
1060 unwind or exceptional control flow. If the function does unwind, its runtime
1061 behavior is undefined.</dd>
1064 <dd>This attribute indicates that the function computes its result (or decides to
1065 unwind an exception) based strictly on its arguments, without dereferencing any
1066 pointer arguments or otherwise accessing any mutable state (e.g. memory, control
1067 registers, etc) visible to caller functions. It does not write through any
1069 never changes any state visible to callers. This means that it cannot unwind
1074 <dd>This attribute indicates that the function does not write through any
1076 or otherwise modify any state (e.g. memory, control registers, etc) visible to
1077 caller functions. It may dereference pointer arguments and read state that may
1078 be set in the caller. A readonly function always returns the same value (or
1079 unwinds an exception identically) when called with the same set of arguments
1084 <dd>This attribute indicates that the function should emit a stack smashing
1086 stack before the local variables that's checked upon return from the function to
1087 see if it has been overwritten. A heuristic is used to determine if a function
1088 needs stack protectors or not.
1097 function attribute.
1103 </dl>
1105 </div>
1107 <!-- ======================================================================= -->
1110 </div>
1113 <p>
1114 Modules may contain "module-level inline asm" blocks, which corresponds to the
1115 GCC "file scope inline asm" blocks. These blocks are internally concatenated by
1116 LLVM and treated as a single unit, but may be separated in the .ll file if
1117 desired. The syntax is very simple:
1118 </p>
1121 <pre>
1122 module asm "inline asm code goes here"
1123 module asm "more can go here"
1124 </pre>
1125 </div>
1127 <p>The strings can contain any character by escaping non-printable characters.
1129 for the number.
1130 </p>
1132 <p>
1133 The inline asm code is simply printed to the machine code .s file when
1134 assembly code is generated.
1135 </p>
1136 </div>
1138 <!-- ======================================================================= -->
1141 </div>
1144 <p>A module may specify a target specific data layout string that specifies how
1145 data is to be laid out in memory. The syntax for the data layout is simply:</p>
1148 separated by the minus sign character ('-'). Each specification starts with a
1149 letter and may include other information after the letter to define some
1150 aspect of the data layout. The specifications accepted are as follows: </p>
1151 <dl>
1153 <dd>Specifies that the target lays out data in big-endian form. That is, the
1154 bits with the most significance have the lowest address location.</dd>
1156 <dd>Specifies that the target lays out data in little-endian form. That is,
1157 the bits with the least significance have the lowest address location.</dd>
1162 too.</dd>
1164 <dd>This specifies the alignment for an integer type of a given bit
1167 <dd>This specifies the alignment for a vector type of a given bit
1170 <dd>This specifies the alignment for a floating point type of a given bit
1172 (double).</dd>
1174 <dd>This specifies the alignment for an aggregate type of a given bit
1176 </dl>
1177 <p>When constructing the data layout for a given target, LLVM starts with a
1178 default set of specifications which are then (possibly) overriden by the
1180 are given in this list:</p>
1181 <ul>
1195 </ul>
1196 <p>When LLVM is determining the alignment for a given type, it uses the
1197 following rules:</p>
1198 <ol>
1199 <li>If the type sought is an exact match for one of the specifications, that
1200 specification is used.</li>
1201 <li>If no match is found, and the type sought is an integer type, then the
1202 smallest integer type that is larger than the bitwidth of the sought type is
1203 used. If none of the specifications are larger than the bitwidth then the the
1204 largest integer type is used. For example, given the default specifications
1205 above, the i7 type will use the alignment of i8 (next largest) while both
1206 i65 and i256 will use the alignment of i64 (largest specified).</li>
1207 <li>If no match is found, and the type sought is a vector type, then the
1208 largest vector type that is smaller than the sought vector type will be used
1211 </ol>
1212 </div>
1214 <!-- *********************************************************************** -->
1216 <!-- *********************************************************************** -->
1220 <p>The LLVM type system is one of the most important features of the
1221 intermediate representation. Being typed enables a number of
1222 optimizations to be performed on the intermediate representation directly,
1223 without having to do
1224 extra analyses on the side before the transformation. A strong type
1225 system makes it easier to read the generated code and enables novel
1226 analyses and transformations that are not feasible to perform on normal
1227 three address code representations.</p>
1229 </div>
1231 <!-- ======================================================================= -->
1233 Classifications</a> </div>
1235 <p>The types fall into a few useful
1236 classifications:</p>
1239 <tbody>
1241 <tr>
1244 </tr>
1245 <tr>
1248 </tr>
1249 <tr>
1258 </td>
1259 </tr>
1260 <tr>
1265 </tr>
1266 <tr>
1276 </td>
1277 </tr>
1278 </tbody>
1279 </table>
1282 most important. Values of these types are the only ones which can be
1283 produced by instructions, passed as arguments, or used as operands to
1284 instructions.</p>
1285 </div>
1287 <!-- ======================================================================= -->
1291 <p>The primitive types are the fundamental building blocks of the LLVM
1292 system.</p>
1294 </div>
1296 <!-- _______________________________________________________________________ -->
1300 <table>
1301 <tbody>
1308 </tbody>
1309 </table>
1310 </div>
1312 <!-- _______________________________________________________________________ -->
1321 <pre>
1322 void
1323 </pre>
1324 </div>
1326 <!-- _______________________________________________________________________ -->
1335 <pre>
1336 label
1337 </pre>
1338 </div>
1341 <!-- ======================================================================= -->
1346 <p>The real power in LLVM comes from the derived types in the system.
1347 This is what allows a programmer to represent arrays, functions,
1348 pointers, and other useful types. Note that these derived types may be
1349 recursive: For example, it is possible to have a two dimensional array.</p>
1351 </div>
1353 <!-- _______________________________________________________________________ -->
1359 <p>The integer type is a very simple derived type that simply specifies an
1360 arbitrary bit width for the integer type desired. Any bit width from 1 bit to
1365 <pre>
1366 iN
1367 </pre>
1370 value.</p>
1374 <tbody>
1375 <tr>
1378 </tr><tr>
1381 </tr><tr>
1384 </tr>
1385 </tbody>
1386 </table>
1388 <p>Note that the code generator does not yet support large integer types
1389 to be used as function return types. The specific limit on how large a
1390 return type the code generator can currently handle is target-dependent;
1392 targets.</p>
1394 </div>
1396 <!-- _______________________________________________________________________ -->
1403 <p>The array type is a very simple derived type that arranges elements
1404 sequentially in memory. The array type requires a size (number of
1405 elements) and an underlying data type.</p>
1409 <pre>
1411 </pre>
1413 <p>The number of elements is a constant integer value; elementtype may
1414 be any type with a size.</p>
1421 </tr>
1425 </tr>
1429 </tr>
1430 </table>
1436 </tr>
1440 </tr>
1444 </tr>
1445 </table>
1447 <p>Note that 'variable sized arrays' can be implemented in LLVM with a zero
1448 length array. Normally, accesses past the end of an array are undefined in
1450 As a special case, however, zero length arrays are recognized to be variable
1451 length. This allows implementation of 'pascal style arrays' with the LLVM
1454 <p>Note that the code generator does not yet support large aggregate types
1455 to be used as function return types. The specific limit on how large an
1456 aggregate return type the code generator can currently handle is
1457 target-dependent, and also dependent on the aggregate element types.</p>
1459 </div>
1461 <!-- _______________________________________________________________________ -->
1467 <p>The function type can be thought of as a function signature. It
1468 consists of a return type and a list of formal parameter types. The
1469 return type of a function type is a scalar type, a void type, or a struct type.
1470 If the return type is a struct type then all struct elements must be of first
1471 class types, and the struct must have at least one element.</p>
1475 <pre>
1477 </pre>
1481 which indicates that the function takes a variable number of arguments.
1482 Variable argument functions can access their arguments with the <a
1492 </td>
1495 </tt></td>
1500 </td>
1506 LLVM.
1507 </td>
1512 </td>
1513 </tr>
1514 </table>
1516 </div>
1517 <!-- _______________________________________________________________________ -->
1521 <p>The structure type is used to represent a collection of data members
1522 together in memory. The packing of the field types is defined to match
1523 the ABI of the underlying processor. The elements of a structure may
1524 be any type that has a size.</p>
1528 instruction.</p>
1542 </tr>
1543 </table>
1545 <p>Note that the code generator does not yet support large aggregate types
1546 to be used as function return types. The specific limit on how large an
1547 aggregate return type the code generator can currently handle is
1548 target-dependent, and also dependent on the aggregate element types.</p>
1550 </div>
1552 <!-- _______________________________________________________________________ -->
1554 </div>
1557 <p>The packed structure type is used to represent a collection of data members
1558 together in memory. There is no padding between fields. Further, the alignment
1559 of a packed structure is 1 byte. The elements of a packed structure may
1560 be any type that has a size.</p>
1564 instruction.</p>
1579 </tr>
1580 </table>
1581 </div>
1583 <!-- _______________________________________________________________________ -->
1587 <p>As in many languages, the pointer type represents a pointer or
1588 reference to another object, which must live in memory. Pointer types may have
1589 an optional address space attribute defining the target-specific numbered
1590 address space where the pointed-to object resides. The default address space is
1591 zero.</p>
1604 </tr>
1610 </tr>
1615 </tr>
1616 </table>
1617 </div>
1619 <!-- _______________________________________________________________________ -->
1625 <p>A vector type is a simple derived type that represents a vector
1626 of elements. Vector types are used when multiple primitive data
1627 are operated in parallel using a single instruction (SIMD).
1628 A vector type requires a size (number of
1629 elements) and an underlying primitive data type. Vectors must have a power
1635 <pre>
1637 </pre>
1639 <p>The number of elements is a constant integer value; elementtype may
1640 be any integer or floating point type.</p>
1648 </tr>
1652 </tr>
1656 </tr>
1657 </table>
1659 <p>Note that the code generator does not yet support large vector types
1660 to be used as function return types. The specific limit on how large a
1661 vector return type codegen can currently handle is target-dependent;
1662 currently it's often a few times longer than a hardware vector register.</p>
1664 </div>
1666 <!-- _______________________________________________________________________ -->
1672 <p>Opaque types are used to represent unknown types in the system. This
1673 corresponds (for example) to the C notion of a forward declared structure type.
1674 In LLVM, opaque types can eventually be resolved to any type (not just a
1675 structure type).</p>
1679 <pre>
1680 opaque
1681 </pre>
1689 </tr>
1690 </table>
1691 </div>
1693 <!-- ======================================================================= -->
1696 </div>
1700 <p>
1701 An "up reference" allows you to refer to a lexically enclosing type without
1702 requiring it to have a name. For instance, a structure declaration may contain a
1703 pointer to any of the types it is lexically a member of. Example of up
1704 references (with their equivalent as named type declarations) include:</p>
1706 <pre>
1707 { \2 * } %x = type { %x* }
1708 { \2 }* %y = type { %y }*
1709 \1* %z = type %z*
1710 </pre>
1712 <p>
1713 An up reference is needed by the asmprinter for printing out cyclic types when
1714 there is no declared name for a type in the cycle. Because the asmprinter does
1715 not want to print out an infinite type string, it needs a syntax to handle
1716 recursive types that have no names (all names are optional in llvm IR).
1717 </p>
1720 <pre>
1722 </pre>
1724 <p>
1725 The level is the count of the lexical type that is being referred to.
1726 </p>
1734 </tr>
1738 structure.</td>
1739 </tr>
1740 </table>
1741 </div>
1744 <!-- *********************************************************************** -->
1746 <!-- *********************************************************************** -->
1750 <p>LLVM has several different basic types of constants. This section describes
1751 them all and their syntax.</p>
1753 </div>
1755 <!-- ======================================================================= -->
1760 <dl>
1765 </dd>
1771 integer types.
1772 </dd>
1777 exponential notation (e.g. 1.23421e+2), or a more precise hexadecimal
1778 notation (see below). The assembler requires the exact decimal value of
1779 a floating-point constant. For example, the assembler accepts 1.25 but
1788 </dl>
1790 <p>The one non-intuitive notation for constants is the hexadecimal form
1791 of floating point constants. For example, the form '<tt>double
1794 (and the only time that they are generated by the disassembler) is when a
1795 floating point constant must be emitted but it cannot be represented as a
1796 decimal floating point number in a reasonable number of digits. For example,
1797 NaN's, infinities, and other
1798 special values are represented in their IEEE hexadecimal format so that
1799 assembly and disassembly do not cause any bits to change in the constants.</p>
1800 <p>When using the hexadecimal form, constants of types float and double are
1801 represented using the 16-digit form shown above (which matches the IEEE754
1802 representation for double); float values must, however, be exactly representable
1803 as IEE754 single precision.
1804 Hexadecimal format is always used for long
1805 double, and there are three forms of long double. The 80-bit
1807 followed by 20 hexadecimal digits.
1808 The 128-bit format used by PowerPC (two adjacent doubles) is represented
1810 format is represented
1812 target uses this format. Long doubles will only work if they match
1813 the long double format on your target. All hexadecimal formats are big-endian
1814 (sign bit at the left).</p>
1815 </div>
1817 <!-- ======================================================================= -->
1821 </div>
1824 <p>Complex constants are a (potentially recursive) combination of simple
1825 constants and smaller complex constants.</p>
1827 <dl>
1830 <dd>Structure constants are represented with notation similar to structure
1831 type definitions (a comma separated list of elements, surrounded by braces
1835 types of elements must match those specified by the type.
1836 </dd>
1840 <dd>Array constants are represented with notation similar to array type
1841 definitions (a comma separated list of elements, surrounded by square brackets
1844 types of elements must match those specified by the type.
1845 </dd>
1849 <dd>Vector constants are represented with notation similar to vector type
1850 definitions (a comma separated list of elements, surrounded by
1854 match those specified by the type.
1855 </dd>
1861 This is often used to avoid having to print large zero initializers (e.g. for
1862 large arrays) and is always exactly equivalent to using explicit zero
1863 initializers.
1864 </dd>
1868 <dd>A metadata node is a structure-like constant with the type of an empty
1870 constants that are meant to be interpreted as part of the instruction stream,
1871 metadata is a place to attach additional information such as debug info.
1872 </dd>
1873 </dl>
1875 </div>
1877 <!-- ======================================================================= -->
1880 </div>
1886 constants. These constants are explicitly referenced when the <a
1889 file:</p>
1892 <pre>
1893 @X = global i32 17
1894 @Y = global i32 42
1895 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
1896 </pre>
1897 </div>
1899 </div>
1901 <!-- ======================================================================= -->
1905 no specific value. Undefined values may be of any type and be used anywhere
1906 a constant is permitted.</p>
1908 <p>Undefined values indicate to the compiler that the program is well defined
1909 no matter what value is used, giving the compiler more freedom to optimize.
1910 </p>
1911 </div>
1913 <!-- ======================================================================= -->
1915 </div>
1919 <p>Constant expressions are used to allow expressions involving other constants
1920 to be used as constants. Constant expressions may be of any <a
1922 that does not have side effects (e.g. load and call are not supported). The
1923 following is the syntax for constant expressions:</p>
1925 <dl>
1927 <dd>Truncate a constant to another type. The bit size of CST must be larger
1928 than the bit size of TYPE. Both types must be integers.</dd>
1931 <dd>Zero extend a constant to another type. The bit size of CST must be
1932 smaller or equal to the bit size of TYPE. Both types must be integers.</dd>
1935 <dd>Sign extend a constant to another type. The bit size of CST must be
1936 smaller or equal to the bit size of TYPE. Both types must be integers.</dd>
1939 <dd>Truncate a floating point constant to another floating point type. The
1940 size of CST must be larger than the size of TYPE. Both types must be
1941 floating point.</dd>
1944 <dd>Floating point extend a constant to another type. The size of CST must be
1945 smaller or equal to the size of TYPE. Both types must be floating point.</dd>
1948 <dd>Convert a floating point constant to the corresponding unsigned integer
1949 constant. TYPE must be a scalar or vector integer type. CST must be of scalar
1950 or vector floating point type. Both CST and TYPE must be scalars, or vectors
1951 of the same number of elements. If the value won't fit in the integer type,
1952 the results are undefined.</dd>
1955 <dd>Convert a floating point constant to the corresponding signed integer
1956 constant. TYPE must be a scalar or vector integer type. CST must be of scalar
1957 or vector floating point type. Both CST and TYPE must be scalars, or vectors
1958 of the same number of elements. If the value won't fit in the integer type,
1959 the results are undefined.</dd>
1962 <dd>Convert an unsigned integer constant to the corresponding floating point
1963 constant. TYPE must be a scalar or vector floating point type. CST must be of
1964 scalar or vector integer type. Both CST and TYPE must be scalars, or vectors
1965 of the same number of elements. If the value won't fit in the floating point
1966 type, the results are undefined.</dd>
1969 <dd>Convert a signed integer constant to the corresponding floating point
1970 constant. TYPE must be a scalar or vector floating point type. CST must be of
1971 scalar or vector integer type. Both CST and TYPE must be scalars, or vectors
1972 of the same number of elements. If the value won't fit in the floating point
1973 type, the results are undefined.</dd>
1976 <dd>Convert a pointer typed constant to the corresponding integer constant
1977 TYPE must be an integer type. CST must be of pointer type. The CST value is
1978 zero extended, truncated, or unchanged to make it fit in TYPE.</dd>
1981 <dd>Convert a integer constant to a pointer constant. TYPE must be a
1982 pointer type. CST must be of integer type. The CST value is zero extended,
1983 truncated, or unchanged to make it fit in a pointer size. This one is
1987 <dd>Convert a constant, CST, to another TYPE. The constraints of the operands
1995 instruction, the index list may have zero or more indexes, which are required
2001 constants.</dd>
2033 <dd>Perform the specified operation of the LHS and RHS constants. OPCODE may
2035 binary</a> operations. The constraints on operands are the same as those for
2036 the corresponding instruction (e.g. no bitwise operations on floating point
2037 values are allowed).</dd>
2038 </dl>
2039 </div>
2041 <!-- ======================================================================= -->
2043 </div>
2047 <p>Embedded metadata provides a way to attach arbitrary data to the
2048 instruction stream without affecting the behaviour of the program. There are
2049 two metadata primitives, strings and nodes. All metadata has the type of an
2050 empty struct and is identified in syntax by a preceding exclamation point
2052 </p>
2054 <p>A metadata string is a string surrounded by double quotes. It can contain
2057 </p>
2059 <p>Metadata nodes are represented with notation similar to structure constants
2060 (a comma separated list of elements, surrounded by braces and preceeded by an
2062 </p>
2064 <p>A metadata node will attempt to track changes to the values it holds. In
2065 the event that a value is deleted, it will be replaced with a typeless
2068 <p>Optimizations may rely on metadata to provide additional information about
2069 the program that isn't available in the instructions, or that isn't easily
2070 computable. Similarly, the code generator may expect a certain metadata format
2071 to be used to express debugging information.</p>
2072 </div>
2074 <!-- *********************************************************************** -->
2076 <!-- *********************************************************************** -->
2078 <!-- ======================================================================= -->
2081 </div>
2085 <p>
2087 Module-Level Inline Assembly</a>) through the use of a special value. This
2088 value represents the inline assembler as a string (containing the instructions
2089 to emit), a list of operand constraints (stored as a string), and a flag that
2090 indicates whether or not the inline asm expression has side effects. An example
2091 inline assembler expression is:
2092 </p>
2095 <pre>
2097 </pre>
2098 </div>
2100 <p>
2103 </p>
2106 <pre>
2108 </pre>
2109 </div>
2111 <p>
2112 Inline asms with side effects not visible in the constraint list must be marked
2113 as having side effects. This is done through the use of the
2115 </p>
2118 <pre>
2120 </pre>
2121 </div>
2123 <p>TODO: The format of the asm and constraints string still need to be
2124 documented here. Constraints on what can be done (e.g. duplication, moving, etc
2125 need to be documented). This is probably best done by reference to another
2126 document that covers inline asm from a holistic perspective.
2127 </p>
2129 </div>
2131 <!-- *********************************************************************** -->
2133 <!-- *********************************************************************** -->
2137 <p>The LLVM instruction set consists of several different
2144 </div>
2146 <!-- ======================================================================= -->
2148 Instructions</a> </div>
2153 basic block in a program ends with a "Terminator" instruction, which
2154 indicates which block should be executed after the current block is
2156 value: they produce control flow, not values (the one exception being
2165 </div>
2167 <!-- _______________________________________________________________________ -->
2169 Instruction</a> </div>
2172 <pre>
2175 </pre>
2180 optionally a value) from a function back to the caller.</p>
2182 returns a value and then causes control flow, and one that just causes
2183 control flow to occur.</p>
2188 the return value. The type of the return value must be a
2193 instruction with no return value or a return value with a type that
2194 does not match its type, or if it has a void return type and contains
2200 returns back to the calling function's context. If the caller is a "<a
2202 the instruction after the call. If the caller was an "<a
2204 at the beginning of the "normal" destination block. If the instruction
2205 returns a value, that value shall set the call or invoke instruction's
2206 return value.</p>
2210 <pre>
2214 </pre>
2216 <p>Note that the code generator does not yet fully support large
2217 return values. The specific sizes that are currently supported are
2218 dependent on the target. For integers, on 32-bit targets the limit
2220 For aggregate types, the current limits are dependent on the element
2221 types; for example targets are often limited to 2 total integer
2224 </div>
2225 <!-- _______________________________________________________________________ -->
2229 <pre> br i1 <cond>, label <iftrue>, label <iffalse><br> br label <dest> <i>; Unconditional branch</i>
2230 </pre>
2233 transfer to a different basic block in the current function. There are
2234 two forms of this instruction, corresponding to a conditional branch
2235 and an unconditional branch.</p>
2247 <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
2249 </div>
2250 <!-- _______________________________________________________________________ -->
2253 </div>
2258 <pre>
2259 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
2260 </pre>
2266 instruction, allowing a branch to occur to one of many possible
2267 destinations.</p>
2275 table is not allowed to contain duplicate constant entries.</p>
2281 table is searched for the given value. If the value is found, control flow is
2282 transfered to the corresponding destination; otherwise, control flow is
2283 transfered to the default destination.</p>
2287 <p>Depending on properties of the target machine and the particular
2289 ways. For example, it could be generated as a series of chained conditional
2290 branches or with a lookup table.</p>
2294 <pre>
2297 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
2300 switch i32 0, label %dest [ ]
2303 switch i32 %val, label %otherwise [ i32 0, label %onzero
2304 i32 1, label %onone
2305 i32 2, label %ontwo ]
2306 </pre>
2307 </div>
2309 <!-- _______________________________________________________________________ -->
2312 </div>
2318 <pre>
2319 <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>]
2321 </pre>
2326 function, with the possibility of control flow transfer to either the
2338 <ol>
2339 <li>
2341 convention</a> the call should use. If none is specified, the call defaults
2342 to using C calling conventions.
2343 </li>
2350 function value being invoked. In most cases, this is a direct function
2352 an arbitrary pointer to function value.
2353 </li>
2356 function to be invoked. </li>
2359 signature argument types. If the function signature indicates the function
2360 accepts a variable number of arguments, the extra arguments can be
2361 specified. </li>
2372 </ol>
2378 difference is that it establishes an association with a label, which is used by
2379 the runtime library to unwind the stack.</p>
2381 <p>This instruction is used in languages with destructors to ensure that proper
2383 exception. Additionally, this is important for implementation of
2387 <pre>
2388 %retval = invoke i32 @Test(i32 15) to label %Continue
2392 </pre>
2393 </div>
2396 <!-- _______________________________________________________________________ -->
2399 Instruction</a> </div>
2404 <pre>
2405 unwind
2406 </pre>
2411 at the first callee in the dynamic call stack which used an <a
2413 primarily used to implement exception handling.</p>
2418 immediately halt. The dynamic call stack is then searched for the first <a
2420 execution continues at the "exceptional" destination block specified by the
2422 dynamic call chain, undefined behavior results.</p>
2423 </div>
2425 <!-- _______________________________________________________________________ -->
2428 Instruction</a> </div>
2433 <pre>
2434 unreachable
2435 </pre>
2440 instruction is used to inform the optimizer that a particular portion of the
2441 code is not reachable. This can be used to indicate that the code after a
2442 no-return function cannot be reached, and other facts.</p>
2447 </div>
2451 <!-- ======================================================================= -->
2454 <p>Binary operators are used to do most of the computation in a
2455 program. They require two operands of the same type, execute an operation on them, and
2456 produce a single value. The operands might represent
2458 The result value has the same type as its operands.</p>
2460 </div>
2461 <!-- _______________________________________________________________________ -->
2464 </div>
2470 <pre>
2472 </pre>
2483 types.</p>
2487 <p>The value produced is the integer or floating point sum of the two
2488 operands.</p>
2490 <p>If an integer sum has unsigned overflow, the result returned is the
2492 the result.</p>
2494 <p>Because LLVM integers use a two's complement representation, this
2495 instruction is appropriate for both signed and unsigned integers.</p>
2499 <pre>
2501 </pre>
2502 </div>
2503 <!-- _______________________________________________________________________ -->
2506 </div>
2512 <pre>
2514 </pre>
2519 operands.</p>
2523 representations.</p>
2530 types.</p>
2534 <p>The value produced is the integer or floating point difference of
2535 the two operands.</p>
2537 <p>If an integer difference has unsigned overflow, the result returned is the
2539 the result.</p>
2541 <p>Because LLVM integers use a two's complement representation, this
2542 instruction is appropriate for both signed and unsigned integers.</p>
2545 <pre>
2548 </pre>
2549 </div>
2551 <!-- _______________________________________________________________________ -->
2554 </div>
2560 </pre>
2563 operands.</p>
2570 types.</p>
2574 <p>The value produced is the integer or floating point product of the
2575 two operands.</p>
2577 <p>If the result of an integer multiplication has unsigned overflow,
2578 the result returned is the mathematical result modulo
2580 <p>Because LLVM integers use a two's complement representation, and the
2581 result is the same width as the operands, this instruction returns the
2582 correct result for both signed and unsigned integers. If a full product
2584 should be sign-extended or zero-extended as appropriate to the
2585 width of the full product.</p>
2588 </pre>
2589 </div>
2591 <!-- _______________________________________________________________________ -->
2593 </a></div>
2597 </pre>
2600 operands.</p>
2606 values. Both arguments must have identical types.</p>
2611 <p>Note that unsigned integer division and signed integer division are distinct
2616 </pre>
2617 </div>
2618 <!-- _______________________________________________________________________ -->
2620 </a> </div>
2623 <pre>
2625 </pre>
2630 operands.</p>
2636 values. Both arguments must have identical types.</p>
2639 <p>The value produced is the signed integer quotient of the two operands rounded towards zero.</p>
2640 <p>Note that signed integer division and unsigned integer division are distinct
2642 <p>Division by zero leads to undefined behavior. Overflow also leads to
2643 undefined behavior; this is a rare case, but can occur, for example,
2647 </pre>
2648 </div>
2649 <!-- _______________________________________________________________________ -->
2651 Instruction</a> </div>
2654 <pre>
2656 </pre>
2660 operands.</p>
2666 of floating point values. Both arguments must have identical types.</p>
2674 <pre>
2676 </pre>
2677 </div>
2679 <!-- _______________________________________________________________________ -->
2681 </div>
2685 </pre>
2688 unsigned division of its two arguments.</p>
2692 values. Both arguments must have identical types.</p>
2695 This instruction always performs an unsigned division to get the remainder.</p>
2696 <p>Note that unsigned integer remainder and signed integer remainder are
2701 </pre>
2703 </div>
2704 <!-- _______________________________________________________________________ -->
2707 </div>
2713 <pre>
2715 </pre>
2720 signed division of its two operands. This instruction can also take
2722 the elements must be integers.</p>
2728 values. Both arguments must have identical types.</p>
2735 a value. For more information about the difference, see <a
2737 Math Forum</a>. For a table of how this is implemented in various languages,
2740 <p>Note that signed integer remainder and unsigned integer remainder are
2742 <p>Taking the remainder of a division by zero leads to undefined behavior.
2743 Overflow also leads to undefined behavior; this is a rare case, but can occur,
2745 (The remainder doesn't actually overflow, but this rule lets srem be
2746 implemented using instructions that return both the result of the division
2747 and the remainder.)</p>
2750 </pre>
2752 </div>
2753 <!-- _______________________________________________________________________ -->
2761 </pre>
2764 division of its two operands.</p>
2768 of floating point values. Both arguments must have identical types.</p>
2773 The remainder has the same sign as the dividend.</p>
2777 <pre>
2779 </pre>
2780 </div>
2782 <!-- ======================================================================= -->
2784 Operations</a> </div>
2786 <p>Bitwise binary operators are used to do various forms of
2787 bit-twiddling in a program. They are generally very efficient
2788 instructions and can commonly be strength reduced from other
2789 instructions. They require two operands of the same type, execute an operation on them,
2790 and produce a single value. The resulting value is the same type as its operands.</p>
2791 </div>
2793 <!-- _______________________________________________________________________ -->
2795 Instruction</a> </div>
2799 </pre>
2804 the left a specified number of bits.</p>
2825 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> <i>; yields: result=<2 x i32> < i32 2, i32 4></i>
2826 </pre>
2827 </div>
2828 <!-- _______________________________________________________________________ -->
2830 Instruction</a> </div>
2834 </pre>
2838 operand shifted to the right a specified number of bits with zero fill.</p>
2847 <p>This instruction always performs a logical shift right operation. The most
2848 significant bits of the result will be filled with zero bits after the
2855 <pre>
2861 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> <i>; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1></i>
2862 </pre>
2863 </div>
2865 <!-- _______________________________________________________________________ -->
2867 Instruction</a> </div>
2872 </pre>
2876 operand shifted to the right a specified number of bits with sign extension.</p>
2884 <p>This instruction always performs an arithmetic shift right operation,
2885 The most significant bits of the result will be filled with the sign bit
2892 <pre>
2898 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> <i>; yields: result=<2 x i32> < i32 -1, i32 0></i>
2899 </pre>
2900 </div>
2902 <!-- _______________________________________________________________________ -->
2904 Instruction</a> </div>
2910 <pre>
2912 </pre>
2917 its two operands.</p>
2923 values. Both arguments must have identical types.</p>
2927 <p> </p>
2928 <div>
2930 <tbody>
2931 <tr>
2935 </tr>
2936 <tr>
2940 </tr>
2941 <tr>
2945 </tr>
2946 <tr>
2950 </tr>
2951 <tr>
2955 </tr>
2956 </tbody>
2957 </table>
2958 </div>
2960 <pre>
2964 </pre>
2965 </div>
2966 <!-- _______________________________________________________________________ -->
2971 </pre>
2974 or of its two operands.</p>
2979 values. Both arguments must have identical types.</p>
2982 <p> </p>
2983 <div>
2985 <tbody>
2986 <tr>
2990 </tr>
2991 <tr>
2995 </tr>
2996 <tr>
3000 </tr>
3001 <tr>
3005 </tr>
3006 <tr>
3010 </tr>
3011 </tbody>
3012 </table>
3013 </div>
3018 </pre>
3019 </div>
3020 <!-- _______________________________________________________________________ -->
3022 Instruction</a> </div>
3026 </pre>
3034 values. Both arguments must have identical types.</p>
3039 <p> </p>
3040 <div>
3042 <tbody>
3043 <tr>
3047 </tr>
3048 <tr>
3052 </tr>
3053 <tr>
3057 </tr>
3058 <tr>
3062 </tr>
3063 <tr>
3067 </tr>
3068 </tbody>
3069 </table>
3070 </div>
3071 <p> </p>
3077 </pre>
3078 </div>
3080 <!-- ======================================================================= -->
3083 </div>
3087 <p>LLVM supports several instructions to represent vector operations in a
3088 target-independent manner. These instructions cover the element-access and
3089 vector-specific operations needed to process vectors effectively. While LLVM
3090 does directly support these vector operations, many sophisticated algorithms
3091 will want to use target-specific intrinsics to take full advantage of a specific
3092 target.</p>
3094 </div>
3096 <!-- _______________________________________________________________________ -->
3099 </div>
3105 <pre>
3106 <result> = extractelement <n x <ty>> <val>, i32 <idx> <i>; yields <ty></i>
3107 </pre>
3111 <p>
3113 element from a vector at a specified index.
3114 </p>
3119 <p>
3122 an index indicating the position from which to extract the element.
3123 The index may be a variable.</p>
3127 <p>
3128 The result is a scalar of the same type as the element type of
3131 results are undefined.
3132 </p>
3136 <pre>
3138 </pre>
3139 </div>
3142 <!-- _______________________________________________________________________ -->
3145 </div>
3151 <pre>
3152 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> <i>; yields <n x <ty>></i>
3153 </pre>
3157 <p>
3159 element into a vector at a specified index.
3160 </p>
3165 <p>
3168 scalar value whose type must equal the element type of the first
3169 operand. The third operand is an index indicating the position at
3170 which to insert the value. The index may be a variable.</p>
3174 <p>
3179 </p>
3183 <pre>
3185 </pre>
3186 </div>
3188 <!-- _______________________________________________________________________ -->
3191 </div>
3197 <pre>
3198 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> <i>; yields <m x <ty>></i>
3199 </pre>
3203 <p>
3205 from two input vectors, returning a vector with the same element type as
3206 the input and length that is the same as the shuffle mask.
3207 </p>
3211 <p>
3213 with types that match each other. The third argument is a shuffle mask whose
3214 element type is always 'i32'. The result of the instruction is a vector whose
3215 length is the same as the shuffle mask and whose element type is the same as
3216 the element type of the first two operands.
3217 </p>
3219 <p>
3220 The shuffle mask operand is required to be a constant vector with either
3221 constant integer or undef values.
3222 </p>
3226 <p>
3227 The elements of the two input vectors are numbered from left to right across
3228 both of the vectors. The shuffle mask operand specifies, for each element of
3229 the result vector, which element of the two input vectors the result element
3230 gets. The element selector may be undef (meaning "don't care") and the second
3231 operand may be undef if performing a shuffle from only one vector.
3232 </p>
3236 <pre>
3240 <4 x i32> <i32 0, i32 1, i32 2, i32 3> <i>; yields <4 x i32></i> - Identity shuffle.
3244 <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>
3245 </pre>
3246 </div>
3249 <!-- ======================================================================= -->
3252 </div>
3256 <p>LLVM supports several instructions for working with aggregate values.
3257 </p>
3259 </div>
3261 <!-- _______________________________________________________________________ -->
3264 </div>
3270 <pre>
3272 </pre>
3276 <p>
3278 or array element from an aggregate value.
3279 </p>
3284 <p>
3287 type. The operands are constant indices to specify which value to extract
3288 in a similar manner as indices in a
3290 </p>
3294 <p>
3295 The result is the value at the position in the aggregate specified by
3296 the index operands.
3297 </p>
3301 <pre>
3303 </pre>
3304 </div>
3307 <!-- _______________________________________________________________________ -->
3310 </div>
3316 <pre>
3317 <result> = insertvalue <aggregate type> <val>, <ty> <val>, <idx> <i>; yields <n x <ty>></i>
3318 </pre>
3322 <p>
3324 into a struct field or array element in an aggregate.
3325 </p>
3330 <p>
3333 The second operand is a first-class value to insert.
3334 The following operands are constant indices
3335 indicating the position at which to insert the value in a similar manner as
3336 indices in a
3338 The value to insert must have the same type as the value identified
3339 by the indices.
3340 </p>
3344 <p>
3348 </p>
3352 <pre>
3354 </pre>
3355 </div>
3358 <!-- ======================================================================= -->
3361 </div>
3365 <p>A key design point of an SSA-based representation is how it
3366 represents memory. In LLVM, no memory locations are in SSA form, which
3367 makes things very simple. This section describes how to read, write,
3368 allocate, and free memory in LLVM.</p>
3370 </div>
3372 <!-- _______________________________________________________________________ -->
3375 </div>
3381 <pre>
3382 <result> = malloc <type>[, i32 <NumElements>][, align <alignment>] <i>; yields {type*}:result</i>
3383 </pre>
3388 heap and returns a pointer to it. The object is always allocated in the generic
3389 address space (address space zero).</p>
3395 bytes of memory from the operating system and returns a pointer of the
3396 appropriate type to the program. If "NumElements" is specified, it is the
3397 number of elements allocated, otherwise "NumElements" is defaulted to be one.
3398 If a constant alignment is specified, the value result of the allocation is guaranteed to
3399 be aligned to at least that boundary. If not specified, or if zero, the target can
3400 choose to align the allocation on any convenient boundary.</p>
3407 a pointer is returned. The result of a zero byte allocation is undefined. The
3408 result is null if there is insufficient memory available.</p>
3412 <pre>
3420 </pre>
3422 <p>Note that the code generator does not yet respect the
3423 alignment value.</p>
3425 </div>
3427 <!-- _______________________________________________________________________ -->
3430 </div>
3436 <pre>
3438 </pre>
3443 memory heap to be reallocated in the future.</p>
3449 instruction.</p>
3453 <p>Access to the memory pointed to by the pointer is no longer defined
3454 after this instruction executes. If the pointer is null, the operation
3455 is a noop.</p>
3459 <pre>
3461 free [4 x i8]* %array
3462 </pre>
3463 </div>
3465 <!-- _______________________________________________________________________ -->
3468 </div>
3474 <pre>
3475 <result> = alloca <type>[, i32 <NumElements>][, align <alignment>] <i>; yields {type*}:result</i>
3476 </pre>
3481 currently executing function, to be automatically released when this function
3482 returns to its caller. The object is always allocated in the generic address
3483 space (address space zero).</p>
3488 bytes of memory on the runtime stack, returning a pointer of the
3489 appropriate type to the program. If "NumElements" is specified, it is the
3490 number of elements allocated, otherwise "NumElements" is defaulted to be one.
3491 If a constant alignment is specified, the value result of the allocation is guaranteed
3492 to be aligned to at least that boundary. If not specified, or if zero, the target
3493 can choose to align the allocation on any convenient boundary.</p>
3499 <p>Memory is allocated; a pointer is returned. The operation is undefined if
3502 instruction is commonly used to represent automatic variables that must
3503 have an address available. When the function returns (either with the <tt><a
3505 instructions), the memory is reclaimed. Allocating zero bytes
3506 is legal, but the result is undefined.</p>
3510 <pre>
3515 </pre>
3516 </div>
3518 <!-- _______________________________________________________________________ -->
3520 Instruction</a> </div>
3523 <pre> <result> = load <ty>* <pointer>[, align <alignment>]<br> <result> = volatile load <ty>* <pointer>[, align <alignment>]<br></pre>
3528 address from which to load. The pointer must point to a <a
3533 instructions. </p>
3534 <p>
3535 The optional constant "align" argument specifies the alignment of the operation
3536 (that is, the alignment of the memory address). A value of 0 or an
3537 omitted "align" argument means that the operation has the preferential
3538 alignment for the target. It is the responsibility of the code emitter
3539 to ensure that the alignment information is correct. Overestimating
3540 the alignment results in an undefined behavior. Underestimating the
3541 alignment may produce less efficient code. An alignment of 1 is always
3542 safe.
3543 </p>
3545 <p>The location of memory pointed to is loaded. If the value being loaded
3546 is of scalar type then the number of bytes read does not exceed the minimum
3547 number of bytes needed to hold all bits of the type. For example, loading an
3550 is undefined if the value was not originally written using a store of the
3551 same type.</p>
3554 <a
3557 </pre>
3558 </div>
3559 <!-- _______________________________________________________________________ -->
3561 Instruction</a> </div>
3564 <pre> store <ty> <value>, <ty>* <pointer>[, align <alignment>] <i>; yields {void}</i>
3565 volatile store <ty> <value>, <ty>* <pointer>[, align <alignment>] <i>; yields {void}</i>
3566 </pre>
3575 optimizer is not allowed to modify the number or order of execution of
3578 <p>
3579 The optional constant "align" argument specifies the alignment of the operation
3580 (that is, the alignment of the memory address). A value of 0 or an
3581 omitted "align" argument means that the operation has the preferential
3582 alignment for the target. It is the responsibility of the code emitter
3583 to ensure that the alignment information is correct. Overestimating
3584 the alignment results in an undefined behavior. Underestimating the
3585 alignment may produce less efficient code. An alignment of 1 is always
3586 safe.
3587 </p>
3592 written does not exceed the minimum number of bytes needed to hold all
3595 size that is not an integral number of bytes, it is unspecified what
3596 happens to the extra bits that do not belong to the type, but they will
3597 typically be overwritten.</p>
3602 </pre>
3603 </div>
3605 <!-- _______________________________________________________________________ -->
3608 </div>
3612 <pre>
3614 </pre>
3618 <p>
3620 subelement of an aggregate data structure. It performs address calculation only
3621 and does not access memory.</p>
3625 <p>The first argument is always a pointer, and forms the basis of the
3626 calculation. The remaining arguments are indices, that indicate which of the
3627 elements of the aggregate object are indexed. The interpretation of each index
3628 is dependent on the type being indexed into. The first index always indexes the
3629 pointer value given as the first argument, the second index indexes a value of
3630 the type pointed to (not necessarily the value directly pointed to, since the
3631 first index can be non-zero), etc. The first type indexed into must be a pointer
3632 value, subsequent types can be arrays, vectors and structs. Note that subsequent
3633 types being indexed into can never be pointers, since that would require loading
3634 the pointer before continuing calculation.</p>
3636 <p>The type of each index argument depends on the type it is indexing into.
3639 integers of any width are allowed (also non-constants).</p>
3641 <p>For example, let's consider a C code fragment and how it gets
3642 compiled to LLVM:</p>
3645 <pre>
3646 struct RT {
3647 char A;
3649 char C;
3650 };
3651 struct ST {
3652 int X;
3653 double Y;
3654 struct RT Z;
3655 };
3657 int *foo(struct ST *s) {
3659 }
3660 </pre>
3661 </div>
3666 <pre>
3670 define i32* %foo(%ST* %s) {
3671 entry:
3673 ret i32* %reg
3674 }
3675 </pre>
3676 </div>
3682 }</tt>' type, a structure. The second index indexes into the third element of
3684 i8 }</tt>' type, another structure. The third index indexes into the second
3686 array. The two dimensions of the array are subscripted into, yielding an
3690 <p>Note that it is perfectly legal to index partially through a
3691 structure, returning a pointer to an inner element. Because of this,
3692 the LLVM code for the given testcase is equivalent to:</p>
3694 <pre>
3695 define i32* %foo(%ST* %s) {
3701 ret i32* %t5
3702 }
3703 </pre>
3705 <p>Note that it is undefined to access an array out of bounds: array
3706 and pointer indexes must always be within the defined bounds of the
3707 array type when accessed with an instruction that dereferences the
3708 pointer (e.g. a load or store instruction). The one exception for
3709 this rule is zero length arrays. These arrays are defined to be
3710 accessible as variable length arrays, which requires access beyond the
3711 zero'th element.</p>
3713 <p>The getelementptr instruction is often confusing. For some more insight
3719 <pre>
3728 </pre>
3729 </div>
3731 <!-- ======================================================================= -->
3733 </div>
3735 <p>The instructions in this category are the conversion instructions (casting)
3736 which all take a single operand and a type. They perform various bit conversions
3737 on the operand.</p>
3738 </div>
3740 <!-- _______________________________________________________________________ -->
3743 </div>
3747 <pre>
3749 </pre>
3752 <p>
3754 </p>
3757 <p>
3765 <p>
3769 It will always truncate bits.</p>
3772 <pre>
3776 </pre>
3777 </div>
3779 <!-- _______________________________________________________________________ -->
3782 </div>
3786 <pre>
3788 </pre>
3809 <pre>
3812 </pre>
3813 </div>
3815 <!-- _______________________________________________________________________ -->
3818 </div>
3822 <pre>
3824 </pre>
3830 <p>
3838 <p>
3846 <pre>
3849 </pre>
3850 </div>
3852 <!-- _______________________________________________________________________ -->
3855 </div>
3861 <pre>
3863 </pre>
3884 <pre>
3887 </pre>
3888 </div>
3890 <!-- _______________________________________________________________________ -->
3893 </div>
3897 <pre>
3899 </pre>
3903 floating point value.</p>
3909 type must be smaller than the destination type.</p>
3919 <pre>
3922 </pre>
3923 </div>
3925 <!-- _______________________________________________________________________ -->
3928 </div>
3932 <pre>
3934 </pre>
3939 </p>
3952 the results are undefined.</p>
3955 <pre>
3959 </pre>
3960 </div>
3962 <!-- _______________________________________________________________________ -->
3965 </div>
3969 <pre>
3971 </pre>
3976 </p>
3989 the results are undefined.</p>
3992 <pre>
3996 </pre>
3997 </div>
3999 <!-- _______________________________________________________________________ -->
4002 </div>
4006 <pre>
4008 </pre>
4023 integer quantity and converts it to the corresponding floating point value. If
4024 the value cannot fit in the floating point value, the results are undefined.</p>
4027 <pre>
4030 </pre>
4031 </div>
4033 <!-- _______________________________________________________________________ -->
4036 </div>
4040 <pre>
4042 </pre>
4057 integer quantity and converts it to the corresponding floating point value. If
4058 the value cannot fit in the floating point value, the results are undefined.</p>
4061 <pre>
4064 </pre>
4065 </div>
4067 <!-- _______________________________________________________________________ -->
4070 </div>
4074 <pre>
4076 </pre>
4090 truncating or zero extending that value to the size of the integer type. If
4094 change.</p>
4097 <pre>
4100 </pre>
4101 </div>
4103 <!-- _______________________________________________________________________ -->
4106 </div>
4110 <pre>
4112 </pre>
4120 value to cast, and a type to cast it to, which must be a
4128 the size of a pointer then a zero extension is done. If they are the same size,
4132 <pre>
4136 </pre>
4137 </div>
4139 <!-- _______________________________________________________________________ -->
4142 </div>
4146 <pre>
4148 </pre>
4158 a non-aggregate first class value, and a type to cast it to, which must also be
4162 type is a pointer, the destination type must also be a pointer. This
4163 instruction supports bitwise conversion of vectors to integers and to vectors
4164 of other types (as long as they have the same size).</p>
4171 converted to other pointer types with this instruction. To convert pointers to
4176 <pre>
4180 </pre>
4181 </div>
4183 <!-- ======================================================================= -->
4187 instructions, which defy better classification.</p>
4188 </div>
4190 <!-- _______________________________________________________________________ -->
4192 </div>
4195 <pre> <result> = icmp <cond> <ty> <op1>, <op2> <i>; yields {i1} or {<N x i1>}:result</i>
4196 </pre>
4199 a vector of boolean values based on comparison
4200 of its two integer, integer vector, or pointer operands.</p>
4203 the condition code indicating the kind of comparison to perform. It is not
4204 a value, just a keyword. The possible condition code are:
4205 </p>
4206 <ol>
4217 </ol>
4221 They must also be identical types.</p>
4225 yields either an <a href="#t_primitive"><tt>i1</tt></a> or vector of <tt>i1</tt> result, as follows:
4226 </p>
4227 <ol>
4230 </li>
4249 </ol>
4251 values are compared as if they were integers.</p>
4252 <p>If the operands are integer vectors, then they are compared
4254 the same number of elements as the values being compared.
4256 </p>
4265 </pre>
4267 <p>Note that the code generator does not yet support vector types with
4270 </div>
4272 <!-- _______________________________________________________________________ -->
4274 </div>
4277 <pre> <result> = fcmp <cond> <ty> <op1>, <op2> <i>; yields {i1} or {<N x i1>}:result</i>
4278 </pre>
4281 or vector of boolean values based on comparison
4282 of its operands.</p>
4283 <p>
4284 If the operands are floating point scalars, then the result
4286 </p>
4287 <p>If the operands are floating point vectors, then the result type
4288 is a vector of boolean with the same number of elements as the
4289 operands being compared.</p>
4292 the condition code indicating the kind of comparison to perform. It is not
4293 a value, just a keyword. The possible condition code are:</p>
4294 <ol>
4311 </ol>
4317 They must have identical types.</p>
4321 If the operands are vectors, then the vectors are compared
4322 element by element.
4323 Each comparison performed
4325 <ol>
4354 </ol>
4361 </pre>
4363 <p>Note that the code generator does not yet support vector types with
4366 </div>
4368 <!-- _______________________________________________________________________ -->
4371 </div>
4374 <pre> <result> = vicmp <cond> <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4375 </pre>
4378 element-wise comparison of its two integer vector operands.</p>
4381 the condition code indicating the kind of comparison to perform. It is not
4382 a value, just a keyword. The possible condition code are:</p>
4383 <ol>
4394 </ol>
4401 identical type as the values being compared. The most significant bit in each
4403 otherwise. All other bits of the result are undefined. The condition codes
4408 <pre>
4409 <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>
4410 <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>
4411 </pre>
4412 </div>
4414 <!-- _______________________________________________________________________ -->
4417 </div>
4423 element-wise comparison of its two floating point vector operands. The output
4424 elements have the same width as the input elements.</p>
4427 the condition code indicating the kind of comparison to perform. It is not
4428 a value, just a keyword. The possible condition code are:</p>
4429 <ol>
4446 </ol>
4449 types.</p>
4454 an identical number of elements as the values being compared, and each element
4455 having identical with to the width of the floating point elements. The most
4456 significant bit in each element is 1 if the element-wise comparison evaluates to
4457 true, and is 0 otherwise. All other bits of the result are undefined. The
4458 condition codes are evaluated identically to the
4462 <pre>
4464 <result> = vfcmp oeq <2 x float> < float 4, float 0 >, < float 5, float 0 >
4467 <result> = vfcmp ult <2 x double> < double 1, double 2 >, < double 2, double 2>
4468 </pre>
4469 </div>
4471 <!-- _______________________________________________________________________ -->
4474 </div>
4483 the SSA graph representing the function.</p>
4486 <p>The type of the incoming values is specified with the first type
4488 as arguments, with one pair for each predecessor basic block of the
4490 type may be used as the value arguments to the PHI node. Only labels
4491 may be used as the label arguments.</p>
4493 <p>There must be no non-phi instructions between the start of a basic
4494 block and the PHI instructions: i.e. PHI instructions must be first in
4495 a basic block.</p>
4500 specified by the pair corresponding to the predecessor basic block that executed
4501 just prior to the current block.</p>
4504 <pre>
4505 Loop: ; Infinite loop that counts from 0 on up...
4506 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
4507 %nextindvar = add i32 %indvar, 1
4508 br label %Loop
4509 </pre>
4510 </div>
4512 <!-- _______________________________________________________________________ -->
4515 </div>
4521 <pre>
4522 <result> = select <i>selty</i> <cond>, <ty> <val1>, <ty> <val2> <i>; yields ty</i>
4525 </pre>
4529 <p>
4531 condition, without branching.
4532 </p>
4537 <p>
4539 a vector of 'i1' values indicating the
4541 type. If the val1/val2 are vectors and
4542 the condition is a scalar, then entire vectors are selected, not
4543 individual elements.
4544 </p>
4548 <p>
4549 If the condition is an i1 and it evaluates to 1, the instruction returns the first
4550 value argument; otherwise, it returns the second value argument.
4551 </p>
4552 <p>
4553 If the condition is a vector of i1, then the value arguments must
4554 be vectors of the same size, and the selection is done element
4555 by element.
4556 </p>
4560 <pre>
4562 </pre>
4564 <p>Note that the code generator does not yet support conditions
4565 with vector type.</p>
4567 </div>
4570 <!-- _______________________________________________________________________ -->
4573 </div>
4578 <pre>
4579 <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>]
4580 </pre>
4590 <ol>
4591 <li>
4593 any allocas or varargs in the caller. If the "tail" marker is present, the
4594 function call is eligible for tail call optimization. Note that calls may
4597 </li>
4598 <li>
4600 convention</a> the call should use. If none is specified, the call defaults
4601 to using C calling conventions.</p>
4602 </li>
4604 <li>
4608 </li>
4610 <li>
4612 the type of the return value. Functions that return no value are marked
4614 </li>
4615 <li>
4617 value being invoked. The argument types must match the types implied by
4618 this signature. This type can be omitted if the function is not varargs
4619 and if the function type does not return a pointer to a function.</p>
4620 </li>
4621 <li>
4623 be invoked. In most cases, this is a direct function invocation, but
4625 to function value.</p>
4626 </li>
4627 <li>
4629 function signature argument types. All arguments must be of
4631 indicates the function accepts a variable number of arguments, the extra
4632 arguments can be specified.</p>
4633 </li>
4634 <li>
4638 </li>
4639 </ol>
4644 transfer to a specified function, with its incoming arguments bound to
4646 instruction in the called function, control flow continues with the
4647 instruction after the function call, and the return value of the
4648 function is bound to the result argument.</p>
4652 <pre>
4653 %retval = call i32 @test(i32 %argc)
4657 call void %foo(i8 97 signext)
4659 %struct.A = type { i32, i8 }
4665 </pre>
4667 </div>
4669 <!-- _______________________________________________________________________ -->
4672 </div>
4678 <pre>
4680 </pre>
4685 the "variable argument" area of a function call. It is used to implement the
4691 the argument. It returns a value of the specified argument type and
4703 <p>It is legal for this instruction to be called in a function which does not
4705 function.</p>
4709 argument.</p>
4715 <p>Note that the code generator does not yet fully support va_arg
4716 on many targets. Also, it does not currently support va_arg with
4717 aggregate types on any target.</p>
4719 </div>
4721 <!-- *********************************************************************** -->
4723 <!-- *********************************************************************** -->
4728 well known names and semantics and are required to follow certain restrictions.
4729 Overall, these intrinsics represent an extension mechanism for the LLVM
4730 language that does not require changing all of the transformations in LLVM when
4731 adding to the language (or the bitcode reader/writer, the parser, etc...).</p>
4734 prefix is reserved in LLVM for intrinsic names; thus, function names may not
4735 begin with this prefix. Intrinsic functions must always be external functions:
4736 you cannot define the body of intrinsic functions. Intrinsic functions may
4737 only be used in call or invoke instructions: it is illegal to take the address
4738 of an intrinsic function. Additionally, because intrinsic functions are part
4739 of the LLVM language, it is required if any are added that they be documented
4740 here.</p>
4742 <p>Some intrinsic functions can be overloaded, i.e., the intrinsic represents
4743 a family of functions that perform the same operation but on different data
4744 types. Because LLVM can represent over 8 million different integer types,
4745 overloading is used commonly to allow an intrinsic function to operate on any
4746 integer type. One or more of the argument types or the result type can be
4747 overloaded to accept any integer type. Argument types may also be defined as
4748 exactly matching a previous argument's type or the result type. This allows an
4749 intrinsic function which accepts multiple arguments, but needs all of them to
4750 be of the same type, to only be overloaded with respect to a single argument or
4751 the result.</p>
4753 <p>Overloaded intrinsics will have the names of its overloaded argument types
4754 encoded into its function name, each preceded by a period. Only those types
4755 which are overloaded result in a name suffix. Arguments whose type is matched
4757 take an integer of any width and returns an integer of exactly the same integer
4758 width. This leads to a family of functions such as
4760 Only one type, the return type, is overloaded, and only one type suffix is
4761 required. Because the argument's type is matched against the return type, it
4762 does not require its own name suffix.</p>
4764 <p>To learn how to add an intrinsic function, please see the
4766 </p>
4768 </div>
4770 <!-- ======================================================================= -->
4773 </div>
4779 intrinsic functions. These functions are related to the similarly
4782 <p>All of these functions operate on arguments that use a
4783 target-specific value type "<tt>va_list</tt>". The LLVM assembly
4784 language reference manual does not define what this type is, so all
4785 transformations should be prepared to handle these functions regardless of
4786 the type used.</p>
4789 instruction and the variable argument handling intrinsic functions are
4790 used.</p>
4793 <pre>
4794 define i32 @test(i32 %X, ...) {
4795 ; Initialize variable argument processing
4796 %ap = alloca i8*
4797 %ap2 = bitcast i8** %ap to i8*
4798 call void @llvm.va_start(i8* %ap2)
4800 ; Read a single integer argument
4801 %tmp = va_arg i8** %ap, i32
4803 ; Demonstrate usage of llvm.va_copy and llvm.va_end
4804 %aq = alloca i8*
4805 %aq2 = bitcast i8** %aq to i8*
4806 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
4807 call void @llvm.va_end(i8* %aq2)
4809 ; Stop processing of arguments.
4810 call void @llvm.va_end(i8* %ap2)
4811 ret i32 %tmp
4812 }
4814 declare void @llvm.va_start(i8*)
4815 declare void @llvm.va_copy(i8*, i8*)
4816 declare void @llvm.va_end(i8*)
4817 </pre>
4818 </div>
4820 </div>
4822 <!-- _______________________________________________________________________ -->
4825 </div>
4843 macro available in C. In a target-dependent way, it initializes the
4847 last argument of the function as the compiler can figure that out.</p>
4849 </div>
4851 <!-- _______________________________________________________________________ -->
4854 </div>
4872 macro available in C. In a target-dependent way, it destroys the
4878 </div>
4880 <!-- _______________________________________________________________________ -->
4883 </div>
4889 <pre>
4891 </pre>
4896 from the source argument list to the destination argument list.</p>
4907 macro available in C. In a target-dependent way, it copies the source
4910 llvm.va_start</a></tt> intrinsic may be arbitrarily complex and require, for
4911 example, memory allocation.</p>
4913 </div>
4915 <!-- ======================================================================= -->
4918 </div>
4922 <p>
4924 Collection</a> (GC) requires the implementation and generation of these
4925 intrinsics.
4929 Front-ends for type-safe garbage collected languages should generate these
4930 intrinsics to make use of the LLVM garbage collectors. For more details, see <a
4932 </p>
4934 <p>The garbage collection intrinsics only operate on objects in the generic
4935 address space (address space zero).</p>
4937 </div>
4939 <!-- _______________________________________________________________________ -->
4942 </div>
4948 <pre>
4949 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
4950 </pre>
4955 the code generator, and allows some metadata to be associated with it.</p>
4959 <p>The first argument specifies the address of a stack object that contains the
4960 root pointer. The second pointer (which must be either a constant or a global
4961 value address) contains the meta-data to be associated with the root.</p>
4966 location. At compile-time, the code generator generates information to allow
4971 </div>
4974 <!-- _______________________________________________________________________ -->
4977 </div>
4983 <pre>
4984 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
4985 </pre>
4990 locations, allowing garbage collector implementations that require read
4991 barriers.</p>
4995 <p>The second argument is the address to read from, which should be an address
4996 allocated from the garbage collector. The first object is a pointer to the
4997 start of the referenced object, if needed by the language runtime (otherwise
4998 null).</p>
5003 instruction, but may be replaced with substantially more complex code by the
5008 </div>
5011 <!-- _______________________________________________________________________ -->
5014 </div>
5020 <pre>
5021 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
5022 </pre>
5027 locations, allowing garbage collector implementations that require write
5028 barriers (such as generational or reference counting collectors).</p>
5032 <p>The first argument is the reference to store, the second is the start of the
5033 object to store it to, and the third is the address of the field of Obj to
5034 store to. If the runtime does not require a pointer to the object, Obj may be
5035 null.</p>
5040 instruction, but may be replaced with substantially more complex code by the
5045 </div>
5049 <!-- ======================================================================= -->
5052 </div>
5055 <p>
5056 These intrinsics are provided by LLVM to expose special features that may only
5057 be implemented with code generator support.
5058 </p>
5060 </div>
5062 <!-- _______________________________________________________________________ -->
5065 </div>
5070 <pre>
5072 </pre>
5076 <p>
5078 target-specific value indicating the return address of the current function
5079 or one of its callers.
5080 </p>
5084 <p>
5085 The argument to this intrinsic indicates which function to return the address
5086 for. Zero indicates the calling function, one indicates its caller, etc. The
5088 </p>
5092 <p>
5094 the return address of the specified call frame, or zero if it cannot be
5095 identified. The value returned by this intrinsic is likely to be incorrect or 0
5096 for arguments other than zero, so it should only be used for debugging purposes.
5097 </p>
5099 <p>
5100 Note that calling this intrinsic does not prevent function inlining or other
5101 aggressive transformations, so the value returned may not be that of the obvious
5102 source-language caller.
5103 </p>
5104 </div>
5107 <!-- _______________________________________________________________________ -->
5110 </div>
5115 <pre>
5117 </pre>
5121 <p>
5123 target-specific frame pointer value for the specified stack frame.
5124 </p>
5128 <p>
5129 The argument to this intrinsic indicates which function to return the frame
5130 pointer for. Zero indicates the calling function, one indicates its caller,
5132 </p>
5136 <p>
5138 the frame address of the specified call frame, or zero if it cannot be
5139 identified. The value returned by this intrinsic is likely to be incorrect or 0
5140 for arguments other than zero, so it should only be used for debugging purposes.
5141 </p>
5143 <p>
5144 Note that calling this intrinsic does not prevent function inlining or other
5145 aggressive transformations, so the value returned may not be that of the obvious
5146 source-language caller.
5147 </p>
5148 </div>
5150 <!-- _______________________________________________________________________ -->
5153 </div>
5158 <pre>
5159 declare i8 *@llvm.stacksave()
5160 </pre>
5164 <p>
5168 features like scoped automatic variable sized arrays in C99.
5169 </p>
5173 <p>
5174 This intrinsic returns a opaque pointer value that can be passed to <a
5181 </p>
5183 </div>
5185 <!-- _______________________________________________________________________ -->
5188 </div>
5193 <pre>
5194 declare void @llvm.stackrestore(i8 * %ptr)
5195 </pre>
5199 <p>
5201 the function stack to the state it was in when the corresponding <a
5203 useful for implementing language features like scoped automatic variable sized
5204 arrays in C99.
5205 </p>
5209 <p>
5211 </p>
5213 </div>
5216 <!-- _______________________________________________________________________ -->
5219 </div>
5224 <pre>
5226 </pre>
5231 <p>
5233 a prefetch instruction if supported; otherwise, it is a noop. Prefetches have
5234 no
5235 effect on the behavior of the program but can change its performance
5236 characteristics.
5237 </p>
5241 <p>
5247 </p>
5251 <p>
5252 This intrinsic does not modify the behavior of the program. In particular,
5253 prefetches cannot trap and do not produce a value. On targets that support this
5254 intrinsic, the prefetch can provide hints to the processor cache for better
5255 performance.
5256 </p>
5258 </div>
5260 <!-- _______________________________________________________________________ -->
5263 </div>
5268 <pre>
5270 </pre>
5275 <p>
5277 (PC) in a region of
5278 code to simulators and other tools. The method is target specific, but it is
5279 expected that the marker will use exported symbols to transmit the PC of the
5280 marker.
5281 The marker makes no guarantees that it will remain with any specific instruction
5282 after optimizations. It is possible that the presence of a marker will inhibit
5283 optimizations. The intended use is to be inserted after optimizations to allow
5284 correlations of simulation runs.
5285 </p>
5289 <p>
5291 </p>
5295 <p>
5296 This intrinsic does not modify the behavior of the program. Backends that do not
5297 support this intrinisic may ignore it.
5298 </p>
5300 </div>
5302 <!-- _______________________________________________________________________ -->
5305 </div>
5310 <pre>
5311 declare i64 @llvm.readcyclecounter( )
5312 </pre>
5317 <p>
5319 counter register (or similar low latency, high accuracy clocks) on those targets
5320 that support it. On X86, it should map to RDTSC. On Alpha, it should map to RPCC.
5321 As the backing counters overflow quickly (on the order of 9 seconds on alpha), this
5322 should only be used for small timings.
5323 </p>
5327 <p>
5328 When directly supported, reading the cycle counter should not modify any memory.
5329 Implementations are allowed to either return a application specific value or a
5330 system wide value. On backends without support, this is lowered to a constant 0.
5331 </p>
5333 </div>
5335 <!-- ======================================================================= -->
5338 </div>
5341 <p>
5342 LLVM provides intrinsics for a few important standard C library functions.
5343 These intrinsics allow source-language front-ends to pass information about the
5344 alignment of the pointer arguments to the code generator, providing opportunity
5345 for more efficient code generation.
5346 </p>
5348 </div>
5350 <!-- _______________________________________________________________________ -->
5353 </div>
5358 <p>This is an overloaded intrinsic. You can use llvm.memcpy on any integer bit
5359 width. Not all targets support all bit widths however.</p>
5360 <pre>
5369 </pre>
5373 <p>
5375 location to the destination location.
5376 </p>
5378 <p>
5380 intrinsics do not return a value, and takes an extra alignment argument.
5381 </p>
5385 <p>
5386 The first argument is a pointer to the destination, the second is a pointer to
5387 the source. The third argument is an integer argument
5388 specifying the number of bytes to copy, and the fourth argument is the alignment
5389 of the source and destination locations.
5390 </p>
5392 <p>
5394 the caller guarantees that both the source and destination pointers are aligned
5395 to that boundary.
5396 </p>
5400 <p>
5402 location to the destination location, which are not allowed to overlap. It
5403 copies "len" bytes of memory over. If the argument is known to be aligned to
5404 some boundary, this can be specified as the fourth argument, otherwise it should
5406 </p>
5407 </div>
5410 <!-- _______________________________________________________________________ -->
5413 </div>
5418 <p>This is an overloaded intrinsic. You can use llvm.memmove on any integer bit
5419 width. Not all targets support all bit widths however.</p>
5420 <pre>
5429 </pre>
5433 <p>
5435 location to the destination location. It is similar to the
5437 </p>
5439 <p>
5441 intrinsics do not return a value, and takes an extra alignment argument.
5442 </p>
5446 <p>
5447 The first argument is a pointer to the destination, the second is a pointer to
5448 the source. The third argument is an integer argument
5449 specifying the number of bytes to copy, and the fourth argument is the alignment
5450 of the source and destination locations.
5451 </p>
5453 <p>
5455 the caller guarantees that the source and destination pointers are aligned to
5456 that boundary.
5457 </p>
5461 <p>
5463 location to the destination location, which may overlap. It
5464 copies "len" bytes of memory over. If the argument is known to be aligned to
5465 some boundary, this can be specified as the fourth argument, otherwise it should
5467 </p>
5468 </div>
5471 <!-- _______________________________________________________________________ -->
5474 </div>
5479 <p>This is an overloaded intrinsic. You can use llvm.memset on any integer bit
5480 width. Not all targets support all bit widths however.</p>
5481 <pre>
5490 </pre>
5494 <p>
5496 byte value.
5497 </p>
5499 <p>
5501 does not return a value, and takes an extra alignment argument.
5502 </p>
5506 <p>
5507 The first argument is a pointer to the destination to fill, the second is the
5508 byte value to fill it with, the third argument is an integer
5509 argument specifying the number of bytes to fill, and the fourth argument is the
5510 known alignment of destination location.
5511 </p>
5513 <p>
5515 the caller guarantees that the destination pointer is aligned to that boundary.
5516 </p>
5520 <p>
5522 the
5523 destination location. If the argument is known to be aligned to some boundary,
5524 this can be specified as the fourth argument, otherwise it should be set to 0 or
5525 1.
5526 </p>
5527 </div>
5530 <!-- _______________________________________________________________________ -->
5533 </div>
5539 floating point or vector of floating point type. Not all targets support all
5540 types however.</p>
5541 <pre>
5542 declare float @llvm.sqrt.f32(float %Val)
5543 declare double @llvm.sqrt.f64(double %Val)
5544 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
5545 declare fp128 @llvm.sqrt.f128(fp128 %Val)
5546 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
5547 </pre>
5551 <p>
5555 negative numbers other than -0.0 (which allows for better optimization, because
5557 defined to return -0.0 like IEEE sqrt.
5558 </p>
5562 <p>
5563 The argument and return value are floating point numbers of the same type.
5564 </p>
5568 <p>
5569 This function returns the sqrt of the specified operand if it is a nonnegative
5570 floating point number.
5571 </p>
5572 </div>
5574 <!-- _______________________________________________________________________ -->
5577 </div>
5583 floating point or vector of floating point type. Not all targets support all
5584 types however.</p>
5585 <pre>
5586 declare float @llvm.powi.f32(float %Val, i32 %power)
5587 declare double @llvm.powi.f64(double %Val, i32 %power)
5588 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
5589 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
5590 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
5591 </pre>
5595 <p>
5597 specified (positive or negative) power. The order of evaluation of
5598 multiplications is not defined. When a vector of floating point type is
5599 used, the second argument remains a scalar integer value.
5600 </p>
5604 <p>
5605 The second argument is an integer power, and the first is a value to raise to
5606 that power.
5607 </p>
5611 <p>
5612 This function returns the first value raised to the second power with an
5613 unspecified sequence of rounding operations.</p>
5614 </div>
5616 <!-- _______________________________________________________________________ -->
5619 </div>
5625 floating point or vector of floating point type. Not all targets support all
5626 types however.</p>
5627 <pre>
5628 declare float @llvm.sin.f32(float %Val)
5629 declare double @llvm.sin.f64(double %Val)
5630 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
5631 declare fp128 @llvm.sin.f128(fp128 %Val)
5632 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
5633 </pre>
5637 <p>
5639 </p>
5643 <p>
5644 The argument and return value are floating point numbers of the same type.
5645 </p>
5649 <p>
5650 This function returns the sine of the specified operand, returning the
5652 conditions in the same way.</p>
5653 </div>
5655 <!-- _______________________________________________________________________ -->
5658 </div>
5664 floating point or vector of floating point type. Not all targets support all
5665 types however.</p>
5666 <pre>
5667 declare float @llvm.cos.f32(float %Val)
5668 declare double @llvm.cos.f64(double %Val)
5669 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
5670 declare fp128 @llvm.cos.f128(fp128 %Val)
5671 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
5672 </pre>
5676 <p>
5678 </p>
5682 <p>
5683 The argument and return value are floating point numbers of the same type.
5684 </p>
5688 <p>
5689 This function returns the cosine of the specified operand, returning the
5691 conditions in the same way.</p>
5692 </div>
5694 <!-- _______________________________________________________________________ -->
5697 </div>
5703 floating point or vector of floating point type. Not all targets support all
5704 types however.</p>
5705 <pre>
5706 declare float @llvm.pow.f32(float %Val, float %Power)
5707 declare double @llvm.pow.f64(double %Val, double %Power)
5708 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
5709 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
5710 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
5711 </pre>
5715 <p>
5717 specified (positive or negative) power.
5718 </p>
5722 <p>
5723 The second argument is a floating point power, and the first is a value to
5724 raise to that power.
5725 </p>
5729 <p>
5730 This function returns the first value raised to the second power,
5731 returning the
5733 conditions in the same way.</p>
5734 </div>
5737 <!-- ======================================================================= -->
5740 </div>
5743 <p>
5744 LLVM provides intrinsics for a few important bit manipulation operations.
5745 These allow efficient code generation for some algorithms.
5746 </p>
5748 </div>
5750 <!-- _______________________________________________________________________ -->
5753 </div>
5758 <p>This is an overloaded intrinsic function. You can use bswap on any integer
5760 <pre>
5764 </pre>
5768 <p>
5770 values with an even number of bytes (positive multiple of 16 bits). These are
5771 useful for performing operations on data that is not in the target's native
5772 byte order.
5773 </p>
5777 <p>
5780 intrinsic returns an i32 value that has the four bytes of the input i32
5785 </p>
5787 </div>
5789 <!-- _______________________________________________________________________ -->
5792 </div>
5797 <p>This is an overloaded intrinsic. You can use llvm.ctpop on any integer bit
5798 width. Not all targets support all bit widths however.</p>
5799 <pre>
5805 </pre>
5809 <p>
5811 value.
5812 </p>
5816 <p>
5817 The only argument is the value to be counted. The argument may be of any
5818 integer type. The return type must match the argument type.
5819 </p>
5823 <p>
5825 </p>
5826 </div>
5828 <!-- _______________________________________________________________________ -->
5831 </div>
5837 integer bit width. Not all targets support all bit widths however.</p>
5838 <pre>
5844 </pre>
5848 <p>
5850 leading zeros in a variable.
5851 </p>
5855 <p>
5856 The only argument is the value to be counted. The argument may be of any
5857 integer type. The return type must match the argument type.
5858 </p>
5862 <p>
5864 in a variable. If the src == 0 then the result is the size in bits of the type
5866 </p>
5867 </div>
5871 <!-- _______________________________________________________________________ -->
5874 </div>
5880 integer bit width. Not all targets support all bit widths however.</p>
5881 <pre>
5887 </pre>
5891 <p>
5893 trailing zeros.
5894 </p>
5898 <p>
5899 The only argument is the value to be counted. The argument may be of any
5900 integer type. The return type must match the argument type.
5901 </p>
5905 <p>
5907 in a variable. If the src == 0 then the result is the size in bits of the type
5909 </p>
5910 </div>
5912 <!-- _______________________________________________________________________ -->
5915 </div>
5921 on any integer bit width.</p>
5922 <pre>
5923 declare i17 @llvm.part.select.i17 (i17 %val, i32 %loBit, i32 %hiBit)
5924 declare i29 @llvm.part.select.i29 (i29 %val, i32 %loBit, i32 %hiBit)
5925 </pre>
5929 range of bits from an integer value and returns them in the same bit width as
5930 the original value.</p>
5934 any bit width but they must have the same bit width. The second and third
5941 operates in forward mode.</p>
5945 <ol>
5949 to determine the number of bits to retain.</li>
5952 </ol>
5953 <p>In reverse mode, a similar computation is made except that the bits are
5958 </div>
5962 </div>
5968 on any integer bit width.</p>
5969 <pre>
5970 declare i17 @llvm.part.set.i17.i9 (i17 %val, i9 %repl, i32 %lo, i32 %hi)
5971 declare i29 @llvm.part.set.i29.i9 (i29 %val, i9 %repl, i32 %lo, i32 %hi)
5972 </pre>
5976 of bits in an integer value with another integer value. It returns the integer
5977 with the replaced bits.</p>
5984 type since they specify only a bit index.</p>
5990 operates in forward mode.</p>
5993 truncating it down to the size of the replacement area or zero extending it
5994 up to that size.</p>
6001 <p>In reverse mode, a similar computation is made except that the bits are
6007 <pre>
6013 </pre>
6015 </div>
6017 <!-- ======================================================================= -->
6020 </div>
6023 <p>
6024 LLVM provides intrinsics for some arithmetic with overflow operations.
6025 </p>
6027 </div>
6029 <!-- _______________________________________________________________________ -->
6032 </div>
6039 on any integer bit width.</p>
6041 <pre>
6042 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
6043 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
6044 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
6045 </pre>
6050 a signed addition of the two arguments, and indicate whether an overflow
6051 occurred during the signed summation.</p>
6055 <p>The arguments (%a and %b) and the first element of the result structure may
6056 be of integer types of any bit width, but they must have the same bit width. The
6063 a signed addition of the two variables. They return a structure — the
6064 first element of which is the signed summation, and the second element of which
6065 is a bit specifying if the signed summation resulted in an overflow.</p>
6068 <pre>
6069 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
6070 %sum = extractvalue {i32, i1} %res, 0
6071 %obit = extractvalue {i32, i1} %res, 1
6072 br i1 %obit, label %overflow, label %normal
6073 </pre>
6075 </div>
6077 <!-- _______________________________________________________________________ -->
6080 </div>
6087 on any integer bit width.</p>
6089 <pre>
6090 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
6091 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
6092 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
6093 </pre>
6098 an unsigned addition of the two arguments, and indicate whether a carry occurred
6099 during the unsigned summation.</p>
6103 <p>The arguments (%a and %b) and the first element of the result structure may
6104 be of integer types of any bit width, but they must have the same bit width. The
6111 an unsigned addition of the two arguments. They return a structure — the
6112 first element of which is the sum, and the second element of which is a bit
6113 specifying if the unsigned summation resulted in a carry.</p>
6116 <pre>
6117 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
6118 %sum = extractvalue {i32, i1} %res, 0
6119 %obit = extractvalue {i32, i1} %res, 1
6120 br i1 %obit, label %carry, label %normal
6121 </pre>
6123 </div>
6125 <!-- _______________________________________________________________________ -->
6128 </div>
6135 on any integer bit width.</p>
6137 <pre>
6138 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
6139 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
6140 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
6141 </pre>
6146 a signed subtraction of the two arguments, and indicate whether an overflow
6147 occurred during the signed subtraction.</p>
6151 <p>The arguments (%a and %b) and the first element of the result structure may
6152 be of integer types of any bit width, but they must have the same bit width. The
6159 a signed subtraction of the two arguments. They return a structure — the
6160 first element of which is the subtraction, and the second element of which is a bit
6161 specifying if the signed subtraction resulted in an overflow.</p>
6164 <pre>
6165 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
6166 %sum = extractvalue {i32, i1} %res, 0
6167 %obit = extractvalue {i32, i1} %res, 1
6168 br i1 %obit, label %overflow, label %normal
6169 </pre>
6171 </div>
6173 <!-- _______________________________________________________________________ -->
6176 </div>
6183 on any integer bit width.</p>
6185 <pre>
6186 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
6187 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
6188 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
6189 </pre>
6194 an unsigned subtraction of the two arguments, and indicate whether an overflow
6195 occurred during the unsigned subtraction.</p>
6199 <p>The arguments (%a and %b) and the first element of the result structure may
6200 be of integer types of any bit width, but they must have the same bit width. The
6207 an unsigned subtraction of the two arguments. They return a structure — the
6208 first element of which is the subtraction, and the second element of which is a bit
6209 specifying if the unsigned subtraction resulted in an overflow.</p>
6212 <pre>
6213 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
6214 %sum = extractvalue {i32, i1} %res, 0
6215 %obit = extractvalue {i32, i1} %res, 1
6216 br i1 %obit, label %overflow, label %normal
6217 </pre>
6219 </div>
6221 <!-- _______________________________________________________________________ -->
6224 </div>
6231 on any integer bit width.</p>
6233 <pre>
6234 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
6235 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
6236 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
6237 </pre>
6242 a signed multiplication of the two arguments, and indicate whether an overflow
6243 occurred during the signed multiplication.</p>
6247 <p>The arguments (%a and %b) and the first element of the result structure may
6248 be of integer types of any bit width, but they must have the same bit width. The
6255 a signed multiplication of the two arguments. They return a structure —
6256 the first element of which is the multiplication, and the second element of
6257 which is a bit specifying if the signed multiplication resulted in an
6258 overflow.</p>
6261 <pre>
6262 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
6263 %sum = extractvalue {i32, i1} %res, 0
6264 %obit = extractvalue {i32, i1} %res, 1
6265 br i1 %obit, label %overflow, label %normal
6266 </pre>
6268 </div>
6270 <!-- _______________________________________________________________________ -->
6273 </div>
6280 on any integer bit width.</p>
6282 <pre>
6283 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
6284 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
6285 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
6286 </pre>
6291 actively being fixed, but it should not currently be used!</i></p>
6294 a unsigned multiplication of the two arguments, and indicate whether an overflow
6295 occurred during the unsigned multiplication.</p>
6299 <p>The arguments (%a and %b) and the first element of the result structure may
6300 be of integer types of any bit width, but they must have the same bit width. The
6303 multiplication.</p>
6308 an unsigned multiplication of the two arguments. They return a structure —
6309 the first element of which is the multiplication, and the second element of
6310 which is a bit specifying if the unsigned multiplication resulted in an
6311 overflow.</p>
6314 <pre>
6315 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
6316 %sum = extractvalue {i32, i1} %res, 0
6317 %obit = extractvalue {i32, i1} %res, 1
6318 br i1 %obit, label %overflow, label %normal
6319 </pre>
6321 </div>
6323 <!-- ======================================================================= -->
6326 </div>
6329 <p>
6331 are described in the <a
6333 Debugging</a> document.
6334 </p>
6335 </div>
6338 <!-- ======================================================================= -->
6341 </div>
6344 <p> The LLVM exception handling intrinsics (which all start with
6348 </div>
6350 <!-- ======================================================================= -->
6353 </div>
6356 <p>
6357 This intrinsic makes it possible to excise one parameter, marked with
6359 function pointer lacking the nest parameter - the caller does not need
6360 to provide a value for it. Instead, the value to use is stored in
6361 advance in a "trampoline", a block of memory usually allocated
6362 on the stack, which also contains code to splice the nest value into the
6363 argument list. This is used to implement the GCC nested function address
6364 extension.
6365 </p>
6366 <p>
6367 For example, if the function is
6370 <pre>
6373 %p = call i8* @llvm.init.trampoline( i8* %tramp1, i8* bitcast (i32 (i8* nest , i32, i32)* @f to i8*), i8* %nval )
6374 %fp = bitcast i8* %p to i32 (i32, i32)*
6375 </pre>
6378 </div>
6380 <!-- _______________________________________________________________________ -->
6383 </div>
6386 <pre>
6388 </pre>
6390 <p>
6392 and returns a function pointer suitable for executing it.
6393 </p>
6395 <p>
6398 and sufficiently aligned block of memory; this memory is written to by the
6399 intrinsic. Note that the size and the alignment are target-specific - LLVM
6400 currently provides no portable way of determining them, so a front-end that
6401 generates this intrinsic needs to have some target-specific knowledge.
6403 </p>
6405 <p>
6407 dependent code, turning it into a function. A pointer to this function is
6408 returned, but needs to be bitcast to an
6410 before being called. The new function's signature is the same as that of
6413 of pointer type. Calling the new function is equivalent to calling
6417 modified, then the effect of any later call to the returned function pointer is
6418 undefined.
6419 </p>
6420 </div>
6422 <!-- ======================================================================= -->
6425 </div>
6428 <p>
6429 These intrinsic functions expand the "universal IR" of LLVM to represent
6430 hardware constructs for atomic operations and memory synchronization. This
6431 provides an interface to the hardware, not an interface to the programmer. It
6432 is aimed at a low enough level to allow any programming models or APIs
6433 (Application Programming Interfaces) which
6434 need atomic behaviors to map cleanly onto it. It is also modeled primarily on
6435 hardware behavior. Just as hardware provides a "universal IR" for source
6436 languages, it also provides a starting point for developing a "universal"
6437 atomic operation and synchronization IR.
6438 </p>
6439 <p>
6441 software transaction memory systems, atomic primitives, and intrinsic
6442 functions as found in BSD, GNU libc, atomic_ops, APR, and other system and
6443 application libraries. The hardware interface provided by LLVM should allow
6444 a clean implementation of all of these APIs and parallel programming models.
6445 No one model or paradigm should be selected above others unless the hardware
6446 itself ubiquitously does so.
6448 </p>
6449 </div>
6451 <!-- _______________________________________________________________________ -->
6454 </div>
6457 <pre>
6461 </pre>
6463 <p>
6465 specific pairs of memory access types.
6466 </p>
6468 <p>
6470 The first four arguments enables a specific barrier as listed below. The fith
6471 argument specifies that the barrier applies to io or device or uncached memory.
6473 </p>
6474 <ul>
6480 </ul>
6482 <p>
6483 This intrinsic causes the system to enforce some ordering constraints upon
6484 the loads and stores of the program. This barrier does not indicate
6486 which they occur. For any of the specified pairs of load and store operations
6487 (f.ex. load-load, or store-load), all of the first operations preceding the
6488 barrier will complete before any of the second operations succeeding the
6489 barrier begin. Specifically the semantics for each pairing is as follows:
6490 </p>
6491 <ul>
6493 after the barrier begins.</li>
6496 store after the barrier begins.</li>
6498 store after the barrier begins.</li>
6500 load after the barrier begins.</li>
6501 </ul>
6502 <p>
6503 These semantics are applied with a logical "and" behavior when more than one
6504 is enabled in a single memory barrier intrinsic.
6505 </p>
6506 <p>
6507 Backends may implement stronger barriers than those requested when they do not
6508 support as fine grained a barrier as requested. Some architectures do not
6509 need all types of barriers and on such architectures, these become noops.
6510 </p>
6512 <pre>
6513 %ptr = malloc i32
6514 store i32 4, %ptr
6517 call void @llvm.memory.barrier( i1 false, i1 true, i1 false, i1 false )
6520 </pre>
6521 </div>
6523 <!-- _______________________________________________________________________ -->
6526 </div>
6529 <p>
6531 any integer bit width and for different address spaces. Not all targets
6532 support all bit widths however.</p>
6534 <pre>
6536 declare i16 @llvm.atomic.cmp.swap.i16.p0i16( i16* <ptr>, i16 <cmp>, i16 <val> )
6537 declare i32 @llvm.atomic.cmp.swap.i32.p0i32( i32* <ptr>, i32 <cmp>, i32 <val> )
6538 declare i64 @llvm.atomic.cmp.swap.i64.p0i64( i64* <ptr>, i64 <cmp>, i64 <val> )
6540 </pre>
6542 <p>
6543 This loads a value in memory and compares it to a given value. If they are
6544 equal, it stores a new value into the memory.
6545 </p>
6547 <p>
6551 this integer type. While any bit width integer may be used, targets may only
6552 lower representations they support in hardware.
6554 </p>
6556 <p>
6557 This entire intrinsic must be executed atomically. It first loads the value
6560 loaded value is yielded in all cases. This provides the equivalent of an
6561 atomic compare-and-swap operation within the SSA framework.
6562 </p>
6565 <pre>
6566 %ptr = malloc i32
6567 store i32 4, %ptr
6570 %result1 = call i32 @llvm.atomic.cmp.swap.i32.p0i32( i32* %ptr, i32 4, %val1 )
6576 %result2 = call i32 @llvm.atomic.cmp.swap.i32.p0i32( i32* %ptr, i32 5, %val2 )
6581 </pre>
6582 </div>
6584 <!-- _______________________________________________________________________ -->
6587 </div>
6591 <p>
6593 integer bit width. Not all targets support all bit widths however.</p>
6594 <pre>
6600 </pre>
6602 <p>
6606 </p>
6609 <p>
6613 integer type. The targets may only lower integer representations they
6614 support.
6615 </p>
6617 <p>
6620 equivalent of an atomic swap operation within the SSA framework.
6622 </p>
6624 <pre>
6625 %ptr = malloc i32
6626 store i32 4, %ptr
6629 %result1 = call i32 @llvm.atomic.swap.i32.p0i32( i32* %ptr, i32 %val1 )
6635 %result2 = call i32 @llvm.atomic.swap.i32.p0i32( i32* %ptr, i32 %val2 )
6640 </pre>
6641 </div>
6643 <!-- _______________________________________________________________________ -->
6647 </div>
6650 <p>
6652 integer bit width. Not all targets support all bit widths however.</p>
6653 <pre>
6659 </pre>
6661 <p>
6664 </p>
6666 <p>
6668 The intrinsic takes two arguments, the first a pointer to an integer value
6669 and the second an integer value. The result is also an integer value. These
6670 integer types can have any bit width, but they must all have the same bit
6671 width. The targets may only lower integer representations they support.
6672 </p>
6674 <p>
6675 This intrinsic does a series of operations atomically. It first loads the
6678 </p>
6681 <pre>
6682 %ptr = malloc i32
6683 store i32 4, %ptr
6684 %result1 = call i32 @llvm.atomic.load.add.i32.p0i32( i32* %ptr, i32 4 )
6686 %result2 = call i32 @llvm.atomic.load.add.i32.p0i32( i32* %ptr, i32 2 )
6688 %result3 = call i32 @llvm.atomic.load.add.i32.p0i32( i32* %ptr, i32 5 )
6691 </pre>
6692 </div>
6694 <!-- _______________________________________________________________________ -->
6698 </div>
6701 <p>
6703 any integer bit width and for different address spaces. Not all targets
6704 support all bit widths however.</p>
6705 <pre>
6711 </pre>
6713 <p>
6716 </p>
6718 <p>
6720 The intrinsic takes two arguments, the first a pointer to an integer value
6721 and the second an integer value. The result is also an integer value. These
6722 integer types can have any bit width, but they must all have the same bit
6723 width. The targets may only lower integer representations they support.
6724 </p>
6726 <p>
6727 This intrinsic does a series of operations atomically. It first loads the
6730 </p>
6733 <pre>
6734 %ptr = malloc i32
6735 store i32 8, %ptr
6736 %result1 = call i32 @llvm.atomic.load.sub.i32.p0i32( i32* %ptr, i32 4 )
6738 %result2 = call i32 @llvm.atomic.load.sub.i32.p0i32( i32* %ptr, i32 2 )
6740 %result3 = call i32 @llvm.atomic.load.sub.i32.p0i32( i32* %ptr, i32 5 )
6743 </pre>
6744 </div>
6746 <!-- _______________________________________________________________________ -->
6753 </div>
6756 <p>
6760 address spaces. Not all targets support all bit widths however.</p>
6761 <pre>
6767 </pre>
6769 <pre>
6775 </pre>
6777 <pre>
6783 </pre>
6785 <pre>
6791 </pre>
6793 <p>
6797 </p>
6799 <p>
6801 These intrinsics take two arguments, the first a pointer to an integer value
6802 and the second an integer value. The result is also an integer value. These
6803 integer types can have any bit width, but they must all have the same bit
6804 width. The targets may only lower integer representations they support.
6805 </p>
6807 <p>
6808 These intrinsics does a series of operations atomically. They first load the
6812 </p>
6815 <pre>
6816 %ptr = malloc i32
6817 store i32 0x0F0F, %ptr
6818 %result0 = call i32 @llvm.atomic.load.nand.i32.p0i32( i32* %ptr, i32 0xFF )
6820 %result1 = call i32 @llvm.atomic.load.and.i32.p0i32( i32* %ptr, i32 0xFF )
6822 %result2 = call i32 @llvm.atomic.load.or.i32.p0i32( i32* %ptr, i32 0F )
6824 %result3 = call i32 @llvm.atomic.load.xor.i32.p0i32( i32* %ptr, i32 0F )
6827 </pre>
6828 </div>
6831 <!-- _______________________________________________________________________ -->
6838 </div>
6841 <p>
6845 address spaces. Not all targets
6846 support all bit widths however.</p>
6847 <pre>
6853 </pre>
6855 <pre>
6861 </pre>
6863 <pre>
6869 </pre>
6871 <pre>
6877 </pre>
6879 <p>
6880 These intrinsics takes the signed or unsigned minimum or maximum of
6883 </p>
6885 <p>
6887 These intrinsics take two arguments, the first a pointer to an integer value
6888 and the second an integer value. The result is also an integer value. These
6889 integer types can have any bit width, but they must all have the same bit
6890 width. The targets may only lower integer representations they support.
6891 </p>
6893 <p>
6894 These intrinsics does a series of operations atomically. They first load the
6898 </p>
6901 <pre>
6902 %ptr = malloc i32
6903 store i32 7, %ptr
6904 %result0 = call i32 @llvm.atomic.load.min.i32.p0i32( i32* %ptr, i32 -2 )
6906 %result1 = call i32 @llvm.atomic.load.max.i32.p0i32( i32* %ptr, i32 8 )
6908 %result2 = call i32 @llvm.atomic.load.umin.i32.p0i32( i32* %ptr, i32 10 )
6910 %result3 = call i32 @llvm.atomic.load.umax.i32.p0i32( i32* %ptr, i32 30 )
6913 </pre>
6914 </div>
6916 <!-- ======================================================================= -->
6919 </div>
6922 <p> This class of intrinsics is designed to be generic and has
6923 no specific purpose. </p>
6924 </div>
6926 <!-- _______________________________________________________________________ -->
6929 </div>
6934 <pre>
6935 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int> )
6936 </pre>
6940 <p>
6942 </p>
6946 <p>
6947 The first argument is a pointer to a value, the second is a pointer to a
6948 global string, the third is a pointer to a global string which is the source
6949 file name, and the last argument is the line number.
6950 </p>
6954 <p>
6955 This intrinsic allows annotation of local variables with arbitrary strings.
6956 This can be useful for special purpose optimizations that want to look for these
6957 annotations. These have no other defined use, they are ignored by code
6958 generation and optimization.
6959 </p>
6960 </div>
6962 <!-- _______________________________________________________________________ -->
6965 </div>
6971 any integer bit width.
6972 </p>
6973 <pre>
6974 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int> )
6975 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int> )
6976 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int> )
6977 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int> )
6978 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int> )
6979 </pre>
6983 <p>
6985 </p>
6989 <p>
6990 The first argument is an integer value (result of some expression),
6991 the second is a pointer to a global string, the third is a pointer to a global
6992 string which is the source file name, and the last argument is the line number.
6993 It returns the value of the first argument.
6994 </p>
6998 <p>
6999 This intrinsic allows annotations to be put on arbitrary expressions
7000 with arbitrary strings. This can be useful for special purpose optimizations
7001 that want to look for these annotations. These have no other defined use, they
7002 are ignored by code generation and optimization.
7003 </p>
7004 </div>
7006 <!-- _______________________________________________________________________ -->
7009 </div>
7014 <pre>
7015 declare void @llvm.trap()
7016 </pre>
7020 <p>
7022 </p>
7026 <p>
7027 None
7028 </p>
7032 <p>
7033 This intrinsics is lowered to the target dependent trap instruction. If the
7034 target does not have a trap instruction, this intrinsic will be lowered to the
7035 call of the abort() function.
7036 </p>
7037 </div>
7039 <!-- _______________________________________________________________________ -->
7042 </div>
7045 <pre>
7048 </pre>
7050 <p>
7053 it is placed on the stack before local variables.
7054 </p>
7056 <p>
7058 first argument is the value loaded from the stack guard
7060 has enough space to hold the value of the guard.
7061 </p>
7063 <p>
7064 This intrinsic causes the prologue/epilogue inserter to force the position of
7066 stack. This is to ensure that if a local variable on the stack is overwritten,
7067 it will destroy the value of the guard. When the function exits, the guard on
7068 the stack is checked against the original guard. If they're different, then
7070 </p>
7071 </div>
7073 <!-- *********************************************************************** -->
7074 <hr>
7075 <address>
7083 Last modified: $Date$
7084 </address>
7086 </body>
7087 </html>