| blob | 416b105cd7286d436092b25f3099b7205482933e |
3 <html>
4 <head>
11 </head>
13 <body>
16 <ol>
21 <ol>
34 </ol>
35 </li>
37 <ol>
40 <ol>
45 </ol>
46 </li>
48 <ol>
57 </ol>
58 </li>
60 </ol>
61 </li>
63 <ol>
70 </ol>
71 </li>
73 <ol>
75 </ol>
76 </li>
78 <ol>
80 <ol>
87 </ol>
88 </li>
90 <ol>
103 </ol>
104 </li>
106 <ol>
113 </ol>
114 </li>
116 <ol>
120 </ol>
121 </li>
123 <ol>
126 </ol>
127 </li>
129 <ol>
136 </ol>
137 </li>
139 <ol>
152 </ol>
153 </li>
155 <ol>
162 </ol>
163 </li>
164 </ol>
165 </li>
167 <ol>
169 <ol>
173 </ol>
174 </li>
176 <ol>
180 </ol>
181 </li>
183 <ol>
191 </ol>
192 </li>
194 <ol>
203 </ol>
204 </li>
206 <ol>
213 </ol>
214 </li>
216 <ol>
223 </ol>
224 </li>
228 <ol>
230 </ol>
231 </li>
233 <ol>
247 </ol>
248 </li>
250 <ol>
259 </ol>
260 </li>
261 </ol>
262 </li>
263 </ol>
268 </div>
270 <!-- *********************************************************************** -->
272 <!-- *********************************************************************** -->
275 <p>This document is a reference manual for the LLVM assembly language.
276 LLVM is a Static Single Assignment (SSA) based representation that provides
277 type safety, low-level operations, flexibility, and the capability of
278 representing 'all' high-level languages cleanly. It is the common code
279 representation used throughout all phases of the LLVM compilation
280 strategy.</p>
281 </div>
283 <!-- *********************************************************************** -->
285 <!-- *********************************************************************** -->
289 <p>The LLVM code representation is designed to be used in three
290 different forms: as an in-memory compiler IR, as an on-disk bitcode
291 representation (suitable for fast loading by a Just-In-Time compiler),
292 and as a human readable assembly language representation. This allows
293 LLVM to provide a powerful intermediate representation for efficient
294 compiler transformations and analysis, while providing a natural means
295 to debug and visualize the transformations. The three different forms
296 of LLVM are all equivalent. This document describes the human readable
297 representation and notation.</p>
299 <p>The LLVM representation aims to be light-weight and low-level
300 while being expressive, typed, and extensible at the same time. It
301 aims to be a "universal IR" of sorts, by being at a low enough level
302 that high-level ideas may be cleanly mapped to it (similar to how
303 microprocessors are "universal IR's", allowing many source languages to
304 be mapped to them). By providing type information, LLVM can be used as
305 the target of optimizations: for example, through pointer analysis, it
306 can be proven that a C automatic variable is never accessed outside of
307 the current function... allowing it to be promoted to a simple SSA
308 value instead of a memory location.</p>
310 </div>
312 <!-- _______________________________________________________________________ -->
317 <p>It is important to note that this document describes 'well formed'
318 LLVM assembly language. There is a difference between what the parser
319 accepts and what is considered 'well formed'. For example, the
320 following instruction is syntactically okay, but not well formed:</p>
323 <pre>
325 </pre>
326 </div>
329 its uses. The LLVM infrastructure provides a verification pass that may
330 be used to verify that an LLVM module is well formed. This pass is
331 automatically run by the parser after parsing input assembly and by
332 the optimizer before it outputs bitcode. The violations pointed out
333 by the verifier pass indicate bugs in transformation passes or input to
334 the parser.</p>
335 </div>
337 <!-- Describe the typesetting conventions here. -->
339 <!-- *********************************************************************** -->
341 <!-- *********************************************************************** -->
345 <p>LLVM identifiers come in two basic types: global and local. Global
346 identifiers (functions, global variables) begin with the @ character. Local
347 identifiers (register names, types) begin with the % character. Additionally,
348 there are three different formats for identifiers, for different purposes:</p>
350 <ol>
351 <li>Named values are represented as a string of characters with their prefix.
352 For example, %foo, @DivisionByZero, %a.really.long.identifier. The actual
354 Identifiers which require other characters in their names can be surrounded
355 with quotes. Special characters may be escaped using "\xx" where xx is the
356 ASCII code for the character in hexadecimal. In this way, any character can
357 be used in a name value, even quotes themselves.
359 <li>Unnamed values are represented as an unsigned numeric value with their
364 </ol>
366 <p>LLVM requires that values start with a prefix for two reasons: Compilers
367 don't need to worry about name clashes with reserved words, and the set of
368 reserved words may be expanded in the future without penalty. Additionally,
369 unnamed identifiers allow a compiler to quickly come up with a temporary
370 variable without having to avoid symbol table conflicts.</p>
372 <p>Reserved words in LLVM are very similar to reserved words in other
373 languages. There are keywords for different opcodes
378 and others. These reserved words cannot conflict with variable names, because
379 none of them start with a prefix character ('%' or '@').</p>
381 <p>Here is an example of LLVM code to multiply the integer variable
387 <pre>
389 </pre>
390 </div>
395 <pre>
397 </pre>
398 </div>
403 <pre>
407 </pre>
408 </div>
411 important lexical features of LLVM:</p>
413 <ol>
416 line.</li>
418 <li>Unnamed temporaries are created when the result of a computation is not
419 assigned to a named value.</li>
423 </ol>
425 <p>...and it also shows a convention that we follow in this document. When
426 demonstrating instructions, we will follow an instruction with a comment that
427 defines the type and name of value produced. Comments are shown in italic
428 text.</p>
430 </div>
432 <!-- *********************************************************************** -->
434 <!-- *********************************************************************** -->
436 <!-- ======================================================================= -->
438 </div>
443 translation unit of the input programs. Each module consists of
444 functions, global variables, and symbol table entries. Modules may be
445 combined together with the LLVM linker, which merges function (and
446 global variable) definitions, resolves forward declarations, and merges
452 href="#globalvars">constant</a> <a href="#t_array">[13 x i8]</a> c"hello world\0A\00" <i>; [13 x i8]*</i>
460 %cast210 = <a
464 <a
466 <a
468 </pre>
469 </div>
476 <p>In general, a module is made up of a list of global values,
477 where both functions and global variables are global values. Global values are
478 represented by a pointer to a memory location (in this case, a pointer to an
479 array of char, and a pointer to a function), and have one of the following <a
482 </div>
484 <!-- ======================================================================= -->
487 </div>
491 <p>
492 All Global Variables and Functions have one of the following types of linkage:
493 </p>
495 <dl>
499 <dd>Global values with private linkage are only directly accessible by
500 objects in the current module. In particular, linking code into a module with
501 an private global value may cause the private to be renamed as necessary to
502 avoid collisions. Because the symbol is private to the module, all
503 references can be updated. This doesn't show up in any symbol table in the
504 object file.
505 </dd>
509 <dd> Similar to private, but the value shows as a local symbol (STB_LOCAL in
510 the case of ELF) in the object file. This corresponds to the notion of the
512 </dd>
515 </dt>
518 into the object file corresponding to the LLVM module. They exist to
519 allow inlining and other optimizations to take place given knowledge of the
520 definition of the global, which is known to be somewhere outside the module.
523 type is only allowed on definitions, not declarations.</dd>
528 the same name when linkage occurs. This is typically used to implement
529 inline functions, templates, or other code which must be generated in each
531 allowed to be discarded.
532 </dd>
538 discarded. This is used for globals that may be emitted in multiple
539 translation units, but that are not guaranteed to be emitted into every
540 translation unit that uses them. One example of this is tentative
541 definitions in C, such as "<tt>int X;</tt>" at global scope.
542 </dd>
547 that some targets may choose to emit different assembly sequences for them
548 for target-dependent reasons. This is used for globals that are declared
549 "weak" in C source code.
550 </dd>
555 pointer to array type. When two global variables with appending linkage are
556 linked together, the two global arrays are appended together. This is the
557 LLVM, typesafe, equivalent of having the system linker append together
558 "sections" with identical names when .o files are linked.
559 </dd>
563 <dd>The semantics of this linkage follow the ELF object file model: the
564 symbol is weak until linked, if not linked, the symbol becomes null instead
565 of being an undefined reference.
566 </dd>
570 <dd>Some languages allow differing globals to be merged, such as two
572 ensure that only equivalent globals are ever merged (the "one definition
575 be merged with equivalent globals. These linkage types are otherwise the
577 </dd>
581 <dd>If none of the above identifiers are used, the global is externally
582 visible, meaning that it participates in linkage and can be used to resolve
583 external symbol references.
584 </dd>
585 </dl>
587 <p>
588 The next two types of linkage are targeted for Microsoft Windows platform
589 only. They are designed to support importing (exporting) symbols from (to)
590 DLLs (Dynamic Link Libraries).
591 </p>
593 <dl>
597 or variable via a global pointer to a pointer that is set up by the DLL
598 exporting the symbol. On Microsoft Windows targets, the pointer name is
600 </dd>
605 pointer to a pointer in a DLL, so that it can be referenced with the
608 name.
609 </dd>
611 </dl>
614 variable is defined to be internal, if another module defined a "<tt>.LC0</tt>"
615 variable and was linked with this one, one of the two would be renamed,
617 external (i.e., lacking any linkage declarations), they are accessible
618 outside of the current module.</p>
624 </div>
626 <!-- ======================================================================= -->
629 </div>
635 specified for the call. The calling convention of any pair of dynamic
636 caller/callee must match, or the behavior of the program is undefined. The
637 following calling conventions are supported by LLVM, and more may be added in
638 the future:</p>
640 <dl>
643 <dd>This calling convention (the default if no other calling convention is
644 specified) matches the target C calling conventions. This calling convention
645 supports varargs function calls and tolerates some mismatch in the declared
646 prototype and implemented declaration of the function (as does normal C).
647 </dd>
651 <dd>This calling convention attempts to make calls as fast as possible
652 (e.g. by passing things in registers). This calling convention allows the
653 target to use whatever tricks it wants to produce fast code for the target,
654 without having to conform to an externally specified ABI (Application Binary
655 Interface). Implementations of this convention should allow arbitrary
657 supported. This calling convention does not support varargs and requires the
658 prototype of all callees to exactly match the prototype of the function
659 definition.
660 </dd>
664 <dd>This calling convention attempts to make code in the caller as efficient
665 as possible under the assumption that the call is not commonly executed. As
666 such, these calls often preserve all registers so that the call does not break
667 any live ranges in the caller side. This calling convention does not support
668 varargs and requires the prototype of all callees to exactly match the
669 prototype of the function definition.
670 </dd>
674 <dd>Any calling convention may be specified by number, allowing
675 target-specific calling conventions to be used. Target specific calling
676 conventions start at 64.
677 </dd>
678 </dl>
680 <p>More calling conventions can be added/defined on an as-needed basis, to
681 support pascal conventions or any other well-known target-independent
682 convention.</p>
684 </div>
686 <!-- ======================================================================= -->
689 </div>
693 <p>
694 All Global Variables and Functions have one of the following visibility styles:
695 </p>
697 <dl>
700 <dd>On targets that use the ELF object file format, default visibility means
701 that the declaration is visible to other
702 modules and, in shared libraries, means that the declared entity may be
703 overridden. On Darwin, default visibility means that the declaration is
704 visible to other modules. Default visibility corresponds to "external
705 linkage" in the language.
706 </dd>
710 <dd>Two declarations of an object with hidden visibility refer to the same
711 object if they are in the same shared object. Usually, hidden visibility
712 indicates that the symbol will not be placed into the dynamic symbol table,
713 so no other module (executable or shared library) can reference it
714 directly.
715 </dd>
719 <dd>On ELF, protected visibility indicates that the symbol will be placed in
720 the dynamic symbol table, but that references within the defining module will
721 bind to the local symbol. That is, the symbol cannot be overridden by another
722 module.
723 </dd>
724 </dl>
726 </div>
728 <!-- ======================================================================= -->
731 </div>
735 <p>LLVM IR allows you to specify name aliases for certain types. This can make
736 it easier to read the IR and make the IR more condensed (particularly when
737 recursive types are involved). An example of a name specification is:
738 </p>
741 <pre>
742 %mytype = type { %mytype*, i32 }
743 </pre>
744 </div>
750 <p>Note that type names are aliases for the structural type that they indicate,
751 and that you can therefore specify multiple names for the same type. This often
752 leads to confusing behavior when dumping out a .ll file. Since LLVM IR uses
753 structural typing, the name is not part of the type. When printing out LLVM IR,
755 shape. This means that if you have code where two different source types end up
756 having the same LLVM type, that the dumper will sometimes print the "wrong" or
757 unexpected type. This is an important design point and isn't going to
758 change.</p>
760 </div>
762 <!-- ======================================================================= -->
765 </div>
769 <p>Global variables define regions of memory allocated at compilation time
770 instead of run-time. Global variables may optionally be initialized, may have
771 an explicit section to be placed in, and may have an optional explicit alignment
772 specified. A variable may be defined as "thread_local", which means that it
773 will not be shared by threads (each thread will have a separated copy of the
774 variable). A variable may be defined as a global "constant," which indicates
776 optimization, allowing the global data to be placed in the read-only section of
777 an executable, etc). Note that variables that need runtime initialization
780 <p>
782 constant, even if the final definition of the global is not. This capability
783 can be used to enable slightly better optimization of the program, but requires
784 the language definition to guarantee that optimizations based on the
785 'constantness' are valid for the translation units that do not include the
786 definition.
787 </p>
789 <p>As SSA values, global variables define pointer values that are in
790 scope (i.e. they dominate) all basic blocks in the program. Global
791 variables always define a pointer to their "content" type because they
792 describe a region of memory, and all memory objects in LLVM are
793 accessed through pointers.</p>
795 <p>A global variable may be declared to reside in a target-specifc numbered
796 address space. For targets that support them, address spaces may affect how
797 optimizations are performed and/or what target instructions are used to access
798 the variable. The default address space is zero. The address space qualifier
799 must precede any other attributes.</p>
801 <p>LLVM allows an explicit section to be specified for globals. If the target
802 supports it, it will emit globals to the section specified.</p>
804 <p>An explicit alignment may be specified for a global. If not present, or if
805 the alignment is set to zero, the alignment of the global is set by the target
806 to whatever it feels convenient. If an explicit alignment is specified, the
807 global is forced to have at least that much alignment. All alignments must be
810 <p>For example, the following defines a global in a numbered address space with
811 an initializer, section, and alignment:</p>
814 <pre>
816 </pre>
817 </div>
819 </div>
822 <!-- ======================================================================= -->
825 </div>
834 name, a (possibly empty) argument list (each with optional
838 an opening curly brace, a list of basic blocks, and a closing curly brace.
840 LLVM function declarations consist of the "<tt>declare</tt>" keyword, an
845 name, a possibly empty list of arguments, an optional alignment, and an optional
848 <p>A function definition contains a list of basic blocks, forming the CFG
849 (Control Flow Graph) for
850 the function. Each basic block may optionally start with a label (giving the
851 basic block a symbol table entry), contains a list of instructions, and ends
853 function return).</p>
855 <p>The first basic block in a function is special in two ways: it is immediately
856 executed on entrance to the function, and it is not allowed to have predecessor
857 basic blocks (i.e. there can not be any branches to the entry block of a
858 function). Because the block can have no predecessors, it also cannot have any
861 <p>LLVM allows an explicit section to be specified for functions. If the target
862 supports it, it will emit functions to the section specified.</p>
864 <p>An explicit alignment may be specified for a function. If not present, or if
865 the alignment is set to zero, the alignment of the function is set by the target
866 to whatever it feels convenient. If an explicit alignment is specified, the
867 function is forced to have at least that much alignment. All alignments must be
873 <tt>
879 </tt>
880 </div>
882 </div>
885 <!-- ======================================================================= -->
888 </div>
891 function, global variable, another alias or bitcast of global value). Aliases
898 <pre>
900 </pre>
901 </div>
903 </div>
907 <!-- ======================================================================= -->
910 <p>The return type and each parameter of a function type may have a set of
912 used to communicate additional information about the result or parameters of
913 a function. Parameter attributes are considered to be part of the function,
914 not of the function type, so functions with different parameter attributes
915 can have the same function type.</p>
917 <p>Parameter attributes are simple keywords that follow the type specified. If
918 multiple parameter attributes are needed, they are space separated. For
919 example:</p>
922 <pre>
923 declare i32 @printf(i8* noalias nocapture, ...)
924 declare i32 @atoi(i8 zeroext)
925 declare signext i8 @returns_signed_char()
926 </pre>
927 </div>
933 <dl>
935 <dd>This indicates to the code generator that the parameter or return value
936 should be zero-extended to a 32-bit value by the caller (for a parameter)
937 or the callee (for a return value).</dd>
940 <dd>This indicates to the code generator that the parameter or return value
941 should be sign-extended to a 32-bit value by the caller (for a parameter)
942 or the callee (for a return value).</dd>
945 <dd>This indicates that this parameter or return value should be treated
946 in a special target-dependent fashion during while emitting code for a
947 function call or return (usually, by putting it in a register as opposed
948 to memory, though some targets use it to distinguish between two different
949 kinds of registers). Use of this attribute is target-specific.</dd>
952 <dd>This indicates that the pointer parameter should really be passed by
953 value to the function. The attribute implies that a hidden copy of the
954 pointee is made between the caller and the callee, so the callee is unable
955 to modify the value in the callee. This attribute is only valid on LLVM
956 pointer arguments. It is generally used to pass structs and arrays by
957 value, but is also valid on pointers to scalars. The copy is considered to
958 belong to the caller not the callee (for example,
961 values. The byval attribute also supports specifying an alignment with the
962 align attribute. This has a target-specific effect on the code generator
963 that usually indicates a desired alignment for the synthesized stack
964 slot.</dd>
967 <dd>This indicates that the pointer parameter specifies the address of a
968 structure that is the return value of the function in the source program.
969 This pointer must be guaranteed by the caller to be valid: loads and stores
970 to the structure may be assumed by the callee to not to trap. This may only
971 be applied to the first parameter. This is not a valid attribute for
972 return values. </dd>
975 <dd>This indicates that the pointer does not alias any global or any other
976 parameter. The caller is responsible for ensuring that this is the
978 that the pointer does not alias any other pointers visible to the
979 caller. For further details, please see the discussion of the NoAlias
980 response in
985 <dd>This indicates that the callee does not make any copies of the pointer
986 that outlive the callee itself. This is not a valid attribute for return
987 values.</dd>
990 <dd>This indicates that the pointer parameter can be excised using the
992 attribute for return values.</dd>
993 </dl>
995 </div>
997 <!-- ======================================================================= -->
1000 </div>
1003 <p>Each function may specify a garbage collector name, which is simply a
1004 string.</p>
1010 collector which will cause the compiler to alter its output in order to support
1011 the named garbage collection algorithm.</p>
1012 </div>
1014 <!-- ======================================================================= -->
1017 </div>
1021 <p>Function attributes are set to communicate additional information about
1022 a function. Function attributes are considered to be part of the function,
1023 not of the function type, so functions with different parameter attributes
1024 can have the same function type.</p>
1026 <p>Function attributes are simple keywords that follow the type specified. If
1027 multiple attributes are needed, they are space separated. For
1028 example:</p>
1031 <pre>
1032 define void @f() noinline { ... }
1033 define void @f() alwaysinline { ... }
1034 define void @f() alwaysinline optsize { ... }
1035 define void @f() optsize
1036 </pre>
1037 </div>
1039 <dl>
1041 <dd>This attribute indicates that the inliner should attempt to inline this
1042 function into callers whenever possible, ignoring any active inlining size
1043 threshold for this caller.</dd>
1046 <dd>This attribute indicates that the inliner should never inline this function
1047 in any situation. This attribute may not be used together with the
1051 <dd>This attribute suggests that optimization passes and code generator passes
1052 make choices that keep the code size of this function low, and otherwise do
1053 optimizations specifically to reduce code size.</dd>
1056 <dd>This function attribute indicates that the function never returns normally.
1057 This produces undefined behavior at runtime if the function ever does
1058 dynamically return.</dd>
1061 <dd>This function attribute indicates that the function never returns with an
1062 unwind or exceptional control flow. If the function does unwind, its runtime
1063 behavior is undefined.</dd>
1066 <dd>This attribute indicates that the function computes its result (or decides to
1067 unwind an exception) based strictly on its arguments, without dereferencing any
1068 pointer arguments or otherwise accessing any mutable state (e.g. memory, control
1069 registers, etc) visible to caller functions. It does not write through any
1071 never changes any state visible to callers. This means that it cannot unwind
1076 <dd>This attribute indicates that the function does not write through any
1078 or otherwise modify any state (e.g. memory, control registers, etc) visible to
1079 caller functions. It may dereference pointer arguments and read state that may
1080 be set in the caller. A readonly function always returns the same value (or
1081 unwinds an exception identically) when called with the same set of arguments
1086 <dd>This attribute indicates that the function should emit a stack smashing
1088 stack before the local variables that's checked upon return from the function to
1089 see if it has been overwritten. A heuristic is used to determine if a function
1090 needs stack protectors or not.
1099 function attribute.
1107 <dd>This attribute indicates that the code generator should not use a
1108 red zone, even if the target-specific ABI normally permits it.
1109 </dd>
1114 </dl>
1116 </div>
1118 <!-- ======================================================================= -->
1121 </div>
1124 <p>
1125 Modules may contain "module-level inline asm" blocks, which corresponds to the
1126 GCC "file scope inline asm" blocks. These blocks are internally concatenated by
1127 LLVM and treated as a single unit, but may be separated in the .ll file if
1128 desired. The syntax is very simple:
1129 </p>
1132 <pre>
1133 module asm "inline asm code goes here"
1134 module asm "more can go here"
1135 </pre>
1136 </div>
1138 <p>The strings can contain any character by escaping non-printable characters.
1140 for the number.
1141 </p>
1143 <p>
1144 The inline asm code is simply printed to the machine code .s file when
1145 assembly code is generated.
1146 </p>
1147 </div>
1149 <!-- ======================================================================= -->
1152 </div>
1155 <p>A module may specify a target specific data layout string that specifies how
1156 data is to be laid out in memory. The syntax for the data layout is simply:</p>
1159 separated by the minus sign character ('-'). Each specification starts with a
1160 letter and may include other information after the letter to define some
1161 aspect of the data layout. The specifications accepted are as follows: </p>
1162 <dl>
1164 <dd>Specifies that the target lays out data in big-endian form. That is, the
1165 bits with the most significance have the lowest address location.</dd>
1167 <dd>Specifies that the target lays out data in little-endian form. That is,
1168 the bits with the least significance have the lowest address location.</dd>
1173 too.</dd>
1175 <dd>This specifies the alignment for an integer type of a given bit
1178 <dd>This specifies the alignment for a vector type of a given bit
1181 <dd>This specifies the alignment for a floating point type of a given bit
1183 (double).</dd>
1185 <dd>This specifies the alignment for an aggregate type of a given bit
1188 <dd>This specifies the alignment for a stack object of a given bit
1190 </dl>
1191 <p>When constructing the data layout for a given target, LLVM starts with a
1192 default set of specifications which are then (possibly) overriden by the
1194 are given in this list:</p>
1195 <ul>
1210 </ul>
1211 <p>When LLVM is determining the alignment for a given type, it uses the
1212 following rules:</p>
1213 <ol>
1214 <li>If the type sought is an exact match for one of the specifications, that
1215 specification is used.</li>
1216 <li>If no match is found, and the type sought is an integer type, then the
1217 smallest integer type that is larger than the bitwidth of the sought type is
1218 used. If none of the specifications are larger than the bitwidth then the the
1219 largest integer type is used. For example, given the default specifications
1220 above, the i7 type will use the alignment of i8 (next largest) while both
1221 i65 and i256 will use the alignment of i64 (largest specified).</li>
1222 <li>If no match is found, and the type sought is a vector type, then the
1223 largest vector type that is smaller than the sought vector type will be used
1226 </ol>
1227 </div>
1229 <!-- *********************************************************************** -->
1231 <!-- *********************************************************************** -->
1235 <p>The LLVM type system is one of the most important features of the
1236 intermediate representation. Being typed enables a number of
1237 optimizations to be performed on the intermediate representation directly,
1238 without having to do
1239 extra analyses on the side before the transformation. A strong type
1240 system makes it easier to read the generated code and enables novel
1241 analyses and transformations that are not feasible to perform on normal
1242 three address code representations.</p>
1244 </div>
1246 <!-- ======================================================================= -->
1248 Classifications</a> </div>
1250 <p>The types fall into a few useful
1251 classifications:</p>
1254 <tbody>
1256 <tr>
1259 </tr>
1260 <tr>
1263 </tr>
1264 <tr>
1274 </td>
1275 </tr>
1276 <tr>
1282 </tr>
1283 <tr>
1293 </td>
1294 </tr>
1295 </tbody>
1296 </table>
1299 most important. Values of these types are the only ones which can be
1300 produced by instructions, passed as arguments, or used as operands to
1301 instructions.</p>
1302 </div>
1304 <!-- ======================================================================= -->
1308 <p>The primitive types are the fundamental building blocks of the LLVM
1309 system.</p>
1311 </div>
1313 <!-- _______________________________________________________________________ -->
1317 <table>
1318 <tbody>
1325 </tbody>
1326 </table>
1327 </div>
1329 <!-- _______________________________________________________________________ -->
1338 <pre>
1339 void
1340 </pre>
1341 </div>
1343 <!-- _______________________________________________________________________ -->
1352 <pre>
1353 label
1354 </pre>
1355 </div>
1357 <!-- _______________________________________________________________________ -->
1362 <p>The metadata type represents embedded metadata. The only derived type that
1364 takes metadata typed parameters, but not pointer to metadata types.</p>
1368 <pre>
1369 metadata
1370 </pre>
1371 </div>
1374 <!-- ======================================================================= -->
1379 <p>The real power in LLVM comes from the derived types in the system.
1380 This is what allows a programmer to represent arrays, functions,
1381 pointers, and other useful types. Note that these derived types may be
1382 recursive: For example, it is possible to have a two dimensional array.</p>
1384 </div>
1386 <!-- _______________________________________________________________________ -->
1392 <p>The integer type is a very simple derived type that simply specifies an
1393 arbitrary bit width for the integer type desired. Any bit width from 1 bit to
1398 <pre>
1399 iN
1400 </pre>
1403 value.</p>
1410 </tr>
1414 </tr>
1418 </tr>
1419 </table>
1421 <p>Note that the code generator does not yet support large integer types
1422 to be used as function return types. The specific limit on how large a
1423 return type the code generator can currently handle is target-dependent;
1425 targets.</p>
1427 </div>
1429 <!-- _______________________________________________________________________ -->
1436 <p>The array type is a very simple derived type that arranges elements
1437 sequentially in memory. The array type requires a size (number of
1438 elements) and an underlying data type.</p>
1442 <pre>
1444 </pre>
1446 <p>The number of elements is a constant integer value; elementtype may
1447 be any type with a size.</p>
1454 </tr>
1458 </tr>
1462 </tr>
1463 </table>
1469 </tr>
1473 </tr>
1477 </tr>
1478 </table>
1480 <p>Note that 'variable sized arrays' can be implemented in LLVM with a zero
1481 length array. Normally, accesses past the end of an array are undefined in
1483 As a special case, however, zero length arrays are recognized to be variable
1484 length. This allows implementation of 'pascal style arrays' with the LLVM
1487 <p>Note that the code generator does not yet support large aggregate types
1488 to be used as function return types. The specific limit on how large an
1489 aggregate return type the code generator can currently handle is
1490 target-dependent, and also dependent on the aggregate element types.</p>
1492 </div>
1494 <!-- _______________________________________________________________________ -->
1500 <p>The function type can be thought of as a function signature. It
1501 consists of a return type and a list of formal parameter types. The
1502 return type of a function type is a scalar type, a void type, or a struct type.
1503 If the return type is a struct type then all struct elements must be of first
1504 class types, and the struct must have at least one element.</p>
1508 <pre>
1510 </pre>
1514 which indicates that the function takes a variable number of arguments.
1515 Variable argument functions can access their arguments with the <a
1525 </td>
1528 </tt></td>
1533 </td>
1539 LLVM.
1540 </td>
1545 </td>
1546 </tr>
1547 </table>
1549 </div>
1550 <!-- _______________________________________________________________________ -->
1554 <p>The structure type is used to represent a collection of data members
1555 together in memory. The packing of the field types is defined to match
1556 the ABI of the underlying processor. The elements of a structure may
1557 be any type that has a size.</p>
1561 instruction.</p>
1575 </tr>
1576 </table>
1578 <p>Note that the code generator does not yet support large aggregate types
1579 to be used as function return types. The specific limit on how large an
1580 aggregate return type the code generator can currently handle is
1581 target-dependent, and also dependent on the aggregate element types.</p>
1583 </div>
1585 <!-- _______________________________________________________________________ -->
1587 </div>
1590 <p>The packed structure type is used to represent a collection of data members
1591 together in memory. There is no padding between fields. Further, the alignment
1592 of a packed structure is 1 byte. The elements of a packed structure may
1593 be any type that has a size.</p>
1597 instruction.</p>
1612 </tr>
1613 </table>
1614 </div>
1616 <!-- _______________________________________________________________________ -->
1620 <p>As in many languages, the pointer type represents a pointer or
1621 reference to another object, which must live in memory. Pointer types may have
1622 an optional address space attribute defining the target-specific numbered
1623 address space where the pointed-to object resides. The default address space is
1624 zero.</p>
1637 </tr>
1643 </tr>
1648 </tr>
1649 </table>
1650 </div>
1652 <!-- _______________________________________________________________________ -->
1658 <p>A vector type is a simple derived type that represents a vector
1659 of elements. Vector types are used when multiple primitive data
1660 are operated in parallel using a single instruction (SIMD).
1661 A vector type requires a size (number of
1662 elements) and an underlying primitive data type. Vectors must have a power
1668 <pre>
1670 </pre>
1672 <p>The number of elements is a constant integer value; elementtype may
1673 be any integer or floating point type.</p>
1681 </tr>
1685 </tr>
1689 </tr>
1690 </table>
1692 <p>Note that the code generator does not yet support large vector types
1693 to be used as function return types. The specific limit on how large a
1694 vector return type codegen can currently handle is target-dependent;
1695 currently it's often a few times longer than a hardware vector register.</p>
1697 </div>
1699 <!-- _______________________________________________________________________ -->
1705 <p>Opaque types are used to represent unknown types in the system. This
1706 corresponds (for example) to the C notion of a forward declared structure type.
1707 In LLVM, opaque types can eventually be resolved to any type (not just a
1708 structure type).</p>
1712 <pre>
1713 opaque
1714 </pre>
1722 </tr>
1723 </table>
1724 </div>
1726 <!-- ======================================================================= -->
1729 </div>
1733 <p>
1734 An "up reference" allows you to refer to a lexically enclosing type without
1735 requiring it to have a name. For instance, a structure declaration may contain a
1736 pointer to any of the types it is lexically a member of. Example of up
1737 references (with their equivalent as named type declarations) include:</p>
1739 <pre>
1740 { \2 * } %x = type { %x* }
1741 { \2 }* %y = type { %y }*
1742 \1* %z = type %z*
1743 </pre>
1745 <p>
1746 An up reference is needed by the asmprinter for printing out cyclic types when
1747 there is no declared name for a type in the cycle. Because the asmprinter does
1748 not want to print out an infinite type string, it needs a syntax to handle
1749 recursive types that have no names (all names are optional in llvm IR).
1750 </p>
1753 <pre>
1755 </pre>
1757 <p>
1758 The level is the count of the lexical type that is being referred to.
1759 </p>
1767 </tr>
1771 structure.</td>
1772 </tr>
1773 </table>
1774 </div>
1777 <!-- *********************************************************************** -->
1779 <!-- *********************************************************************** -->
1783 <p>LLVM has several different basic types of constants. This section describes
1784 them all and their syntax.</p>
1786 </div>
1788 <!-- ======================================================================= -->
1793 <dl>
1798 </dd>
1804 integer types.
1805 </dd>
1810 exponential notation (e.g. 1.23421e+2), or a more precise hexadecimal
1811 notation (see below). The assembler requires the exact decimal value of
1812 a floating-point constant. For example, the assembler accepts 1.25 but
1821 </dl>
1823 <p>The one non-intuitive notation for constants is the hexadecimal form
1824 of floating point constants. For example, the form '<tt>double
1827 (and the only time that they are generated by the disassembler) is when a
1828 floating point constant must be emitted but it cannot be represented as a
1829 decimal floating point number in a reasonable number of digits. For example,
1830 NaN's, infinities, and other
1831 special values are represented in their IEEE hexadecimal format so that
1832 assembly and disassembly do not cause any bits to change in the constants.</p>
1833 <p>When using the hexadecimal form, constants of types float and double are
1834 represented using the 16-digit form shown above (which matches the IEEE754
1835 representation for double); float values must, however, be exactly representable
1836 as IEE754 single precision.
1837 Hexadecimal format is always used for long
1838 double, and there are three forms of long double. The 80-bit
1840 followed by 20 hexadecimal digits.
1841 The 128-bit format used by PowerPC (two adjacent doubles) is represented
1843 format is represented
1845 target uses this format. Long doubles will only work if they match
1846 the long double format on your target. All hexadecimal formats are big-endian
1847 (sign bit at the left).</p>
1848 </div>
1850 <!-- ======================================================================= -->
1854 </div>
1857 <p>Complex constants are a (potentially recursive) combination of simple
1858 constants and smaller complex constants.</p>
1860 <dl>
1863 <dd>Structure constants are represented with notation similar to structure
1864 type definitions (a comma separated list of elements, surrounded by braces
1868 types of elements must match those specified by the type.
1869 </dd>
1873 <dd>Array constants are represented with notation similar to array type
1874 definitions (a comma separated list of elements, surrounded by square brackets
1877 types of elements must match those specified by the type.
1878 </dd>
1882 <dd>Vector constants are represented with notation similar to vector type
1883 definitions (a comma separated list of elements, surrounded by
1887 match those specified by the type.
1888 </dd>
1894 This is often used to avoid having to print large zero initializers (e.g. for
1895 large arrays) and is always exactly equivalent to using explicit zero
1896 initializers.
1897 </dd>
1901 <dd>A metadata node is a structure-like constant with
1904 that are meant to be interpreted as part of the instruction stream, metadata
1905 is a place to attach additional information such as debug info.
1906 </dd>
1907 </dl>
1909 </div>
1911 <!-- ======================================================================= -->
1914 </div>
1920 constants. These constants are explicitly referenced when the <a
1923 file:</p>
1926 <pre>
1927 @X = global i32 17
1928 @Y = global i32 42
1929 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
1930 </pre>
1931 </div>
1933 </div>
1935 <!-- ======================================================================= -->
1939 no specific value. Undefined values may be of any type and be used anywhere
1940 a constant is permitted.</p>
1942 <p>Undefined values indicate to the compiler that the program is well defined
1943 no matter what value is used, giving the compiler more freedom to optimize.
1944 </p>
1945 </div>
1947 <!-- ======================================================================= -->
1949 </div>
1953 <p>Constant expressions are used to allow expressions involving other constants
1954 to be used as constants. Constant expressions may be of any <a
1956 that does not have side effects (e.g. load and call are not supported). The
1957 following is the syntax for constant expressions:</p>
1959 <dl>
1961 <dd>Truncate a constant to another type. The bit size of CST must be larger
1962 than the bit size of TYPE. Both types must be integers.</dd>
1965 <dd>Zero extend a constant to another type. The bit size of CST must be
1966 smaller or equal to the bit size of TYPE. Both types must be integers.</dd>
1969 <dd>Sign extend a constant to another type. The bit size of CST must be
1970 smaller or equal to the bit size of TYPE. Both types must be integers.</dd>
1973 <dd>Truncate a floating point constant to another floating point type. The
1974 size of CST must be larger than the size of TYPE. Both types must be
1975 floating point.</dd>
1978 <dd>Floating point extend a constant to another type. The size of CST must be
1979 smaller or equal to the size of TYPE. Both types must be floating point.</dd>
1982 <dd>Convert a floating point constant to the corresponding unsigned integer
1983 constant. TYPE must be a scalar or vector integer type. CST must be of scalar
1984 or vector floating point type. Both CST and TYPE must be scalars, or vectors
1985 of the same number of elements. If the value won't fit in the integer type,
1986 the results are undefined.</dd>
1989 <dd>Convert a floating point constant to the corresponding signed integer
1990 constant. TYPE must be a scalar or vector integer type. CST must be of scalar
1991 or vector floating point type. Both CST and TYPE must be scalars, or vectors
1992 of the same number of elements. If the value won't fit in the integer type,
1993 the results are undefined.</dd>
1996 <dd>Convert an unsigned integer constant to the corresponding floating point
1997 constant. TYPE must be a scalar or vector floating point type. CST must be of
1998 scalar or vector integer type. Both CST and TYPE must be scalars, or vectors
1999 of the same number of elements. If the value won't fit in the floating point
2000 type, the results are undefined.</dd>
2003 <dd>Convert a signed integer constant to the corresponding floating point
2004 constant. TYPE must be a scalar or vector floating point type. CST must be of
2005 scalar or vector integer type. Both CST and TYPE must be scalars, or vectors
2006 of the same number of elements. If the value won't fit in the floating point
2007 type, the results are undefined.</dd>
2010 <dd>Convert a pointer typed constant to the corresponding integer constant
2011 TYPE must be an integer type. CST must be of pointer type. The CST value is
2012 zero extended, truncated, or unchanged to make it fit in TYPE.</dd>
2015 <dd>Convert a integer constant to a pointer constant. TYPE must be a
2016 pointer type. CST must be of integer type. The CST value is zero extended,
2017 truncated, or unchanged to make it fit in a pointer size. This one is
2021 <dd>Convert a constant, CST, to another TYPE. The constraints of the operands
2029 instruction, the index list may have zero or more indexes, which are required
2035 constants.</dd>
2061 <dd>Perform the specified operation of the LHS and RHS constants. OPCODE may
2063 binary</a> operations. The constraints on operands are the same as those for
2064 the corresponding instruction (e.g. no bitwise operations on floating point
2065 values are allowed).</dd>
2066 </dl>
2067 </div>
2069 <!-- ======================================================================= -->
2071 </div>
2075 <p>Embedded metadata provides a way to attach arbitrary data to the
2076 instruction stream without affecting the behaviour of the program. There are
2077 two metadata primitives, strings and nodes. All metadata has the
2080 </p>
2082 <p>A metadata string is a string surrounded by double quotes. It can contain
2085 </p>
2087 <p>Metadata nodes are represented with notation similar to structure constants
2088 (a comma separated list of elements, surrounded by braces and preceeded by an
2090 </p>
2092 <p>A metadata node will attempt to track changes to the values it holds. In
2093 the event that a value is deleted, it will be replaced with a typeless
2096 <p>Optimizations may rely on metadata to provide additional information about
2097 the program that isn't available in the instructions, or that isn't easily
2098 computable. Similarly, the code generator may expect a certain metadata format
2099 to be used to express debugging information.</p>
2100 </div>
2102 <!-- *********************************************************************** -->
2104 <!-- *********************************************************************** -->
2106 <!-- ======================================================================= -->
2109 </div>
2113 <p>
2115 Module-Level Inline Assembly</a>) through the use of a special value. This
2116 value represents the inline assembler as a string (containing the instructions
2117 to emit), a list of operand constraints (stored as a string), and a flag that
2118 indicates whether or not the inline asm expression has side effects. An example
2119 inline assembler expression is:
2120 </p>
2123 <pre>
2125 </pre>
2126 </div>
2128 <p>
2131 </p>
2134 <pre>
2136 </pre>
2137 </div>
2139 <p>
2140 Inline asms with side effects not visible in the constraint list must be marked
2141 as having side effects. This is done through the use of the
2143 </p>
2146 <pre>
2148 </pre>
2149 </div>
2151 <p>TODO: The format of the asm and constraints string still need to be
2152 documented here. Constraints on what can be done (e.g. duplication, moving, etc
2153 need to be documented). This is probably best done by reference to another
2154 document that covers inline asm from a holistic perspective.
2155 </p>
2157 </div>
2159 <!-- *********************************************************************** -->
2161 <!-- *********************************************************************** -->
2165 <p>The LLVM instruction set consists of several different
2172 </div>
2174 <!-- ======================================================================= -->
2176 Instructions</a> </div>
2181 basic block in a program ends with a "Terminator" instruction, which
2182 indicates which block should be executed after the current block is
2184 value: they produce control flow, not values (the one exception being
2193 </div>
2195 <!-- _______________________________________________________________________ -->
2197 Instruction</a> </div>
2200 <pre>
2203 </pre>
2208 optionally a value) from a function back to the caller.</p>
2210 returns a value and then causes control flow, and one that just causes
2211 control flow to occur.</p>
2216 the return value. The type of the return value must be a
2221 instruction with no return value or a return value with a type that
2222 does not match its type, or if it has a void return type and contains
2228 returns back to the calling function's context. If the caller is a "<a
2230 the instruction after the call. If the caller was an "<a
2232 at the beginning of the "normal" destination block. If the instruction
2233 returns a value, that value shall set the call or invoke instruction's
2234 return value.</p>
2238 <pre>
2242 </pre>
2244 <p>Note that the code generator does not yet fully support large
2245 return values. The specific sizes that are currently supported are
2246 dependent on the target. For integers, on 32-bit targets the limit
2248 For aggregate types, the current limits are dependent on the element
2249 types; for example targets are often limited to 2 total integer
2252 </div>
2253 <!-- _______________________________________________________________________ -->
2257 <pre> br i1 <cond>, label <iftrue>, label <iffalse><br> br label <dest> <i>; Unconditional branch</i>
2258 </pre>
2261 transfer to a different basic block in the current function. There are
2262 two forms of this instruction, corresponding to a conditional branch
2263 and an unconditional branch.</p>
2275 <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
2277 </div>
2278 <!-- _______________________________________________________________________ -->
2281 </div>
2286 <pre>
2287 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
2288 </pre>
2294 instruction, allowing a branch to occur to one of many possible
2295 destinations.</p>
2303 table is not allowed to contain duplicate constant entries.</p>
2309 table is searched for the given value. If the value is found, control flow is
2310 transfered to the corresponding destination; otherwise, control flow is
2311 transfered to the default destination.</p>
2315 <p>Depending on properties of the target machine and the particular
2317 ways. For example, it could be generated as a series of chained conditional
2318 branches or with a lookup table.</p>
2322 <pre>
2325 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
2328 switch i32 0, label %dest [ ]
2331 switch i32 %val, label %otherwise [ i32 0, label %onzero
2332 i32 1, label %onone
2333 i32 2, label %ontwo ]
2334 </pre>
2335 </div>
2337 <!-- _______________________________________________________________________ -->
2340 </div>
2346 <pre>
2347 <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>]
2349 </pre>
2354 function, with the possibility of control flow transfer to either the
2366 <ol>
2367 <li>
2369 convention</a> the call should use. If none is specified, the call defaults
2370 to using C calling conventions.
2371 </li>
2378 function value being invoked. In most cases, this is a direct function
2380 an arbitrary pointer to function value.
2381 </li>
2384 function to be invoked. </li>
2387 signature argument types. If the function signature indicates the function
2388 accepts a variable number of arguments, the extra arguments can be
2389 specified. </li>
2400 </ol>
2406 difference is that it establishes an association with a label, which is used by
2407 the runtime library to unwind the stack.</p>
2409 <p>This instruction is used in languages with destructors to ensure that proper
2411 exception. Additionally, this is important for implementation of
2414 <p>For the purposes of the SSA form, the definition of the value
2416 the edge from the current block to the "normal" label. If the callee
2417 unwinds then no return value is available.</p>
2420 <pre>
2421 %retval = invoke i32 @Test(i32 15) to label %Continue
2425 </pre>
2426 </div>
2429 <!-- _______________________________________________________________________ -->
2432 Instruction</a> </div>
2437 <pre>
2438 unwind
2439 </pre>
2444 at the first callee in the dynamic call stack which used an <a
2446 primarily used to implement exception handling.</p>
2451 immediately halt. The dynamic call stack is then searched for the first <a
2453 execution continues at the "exceptional" destination block specified by the
2455 dynamic call chain, undefined behavior results.</p>
2456 </div>
2458 <!-- _______________________________________________________________________ -->
2461 Instruction</a> </div>
2466 <pre>
2467 unreachable
2468 </pre>
2473 instruction is used to inform the optimizer that a particular portion of the
2474 code is not reachable. This can be used to indicate that the code after a
2475 no-return function cannot be reached, and other facts.</p>
2480 </div>
2484 <!-- ======================================================================= -->
2487 <p>Binary operators are used to do most of the computation in a
2488 program. They require two operands of the same type, execute an operation on them, and
2489 produce a single value. The operands might represent
2491 The result value has the same type as its operands.</p>
2493 </div>
2494 <!-- _______________________________________________________________________ -->
2497 </div>
2503 <pre>
2505 </pre>
2516 have identical types.</p>
2522 <p>If the sum has unsigned overflow, the result returned is the
2524 the result.</p>
2526 <p>Because LLVM integers use a two's complement representation, this
2527 instruction is appropriate for both signed and unsigned integers.</p>
2531 <pre>
2533 </pre>
2534 </div>
2535 <!-- _______________________________________________________________________ -->
2538 </div>
2544 <pre>
2546 </pre>
2556 floating point values. Both arguments must have identical types.</p>
2564 <pre>
2566 </pre>
2567 </div>
2568 <!-- _______________________________________________________________________ -->
2571 </div>
2577 <pre>
2579 </pre>
2584 operands.</p>
2588 representations.</p>
2594 integer values. Both arguments must have identical types.</p>
2600 <p>If the difference has unsigned overflow, the result returned is the
2602 the result.</p>
2604 <p>Because LLVM integers use a two's complement representation, this
2605 instruction is appropriate for both signed and unsigned integers.</p>
2608 <pre>
2611 </pre>
2612 </div>
2614 <!-- _______________________________________________________________________ -->
2617 </div>
2623 <pre>
2625 </pre>
2630 operands.</p>
2634 representations.</p>
2640 of floating point values. Both arguments must have identical types.</p>
2647 <pre>
2650 </pre>
2651 </div>
2653 <!-- _______________________________________________________________________ -->
2656 </div>
2662 </pre>
2665 operands.</p>
2671 values. Both arguments must have identical types.</p>
2677 <p>If the result of the multiplication has unsigned overflow,
2678 the result returned is the mathematical result modulo
2680 <p>Because LLVM integers use a two's complement representation, and the
2681 result is the same width as the operands, this instruction returns the
2682 correct result for both signed and unsigned integers. If a full product
2684 should be sign-extended or zero-extended as appropriate to the
2685 width of the full product.</p>
2688 </pre>
2689 </div>
2691 <!-- _______________________________________________________________________ -->
2694 </div>
2700 </pre>
2703 operands.</p>
2709 of floating point values. Both arguments must have identical types.</p>
2717 </pre>
2718 </div>
2720 <!-- _______________________________________________________________________ -->
2722 </a></div>
2726 </pre>
2729 operands.</p>
2735 values. Both arguments must have identical types.</p>
2740 <p>Note that unsigned integer division and signed integer division are distinct
2745 </pre>
2746 </div>
2747 <!-- _______________________________________________________________________ -->
2749 </a> </div>
2752 <pre>
2754 </pre>
2759 operands.</p>
2765 values. Both arguments must have identical types.</p>
2768 <p>The value produced is the signed integer quotient of the two operands rounded towards zero.</p>
2769 <p>Note that signed integer division and unsigned integer division are distinct
2771 <p>Division by zero leads to undefined behavior. Overflow also leads to
2772 undefined behavior; this is a rare case, but can occur, for example,
2776 </pre>
2777 </div>
2778 <!-- _______________________________________________________________________ -->
2780 Instruction</a> </div>
2783 <pre>
2785 </pre>
2789 operands.</p>
2795 of floating point values. Both arguments must have identical types.</p>
2803 <pre>
2805 </pre>
2806 </div>
2808 <!-- _______________________________________________________________________ -->
2810 </div>
2814 </pre>
2817 unsigned division of its two arguments.</p>
2821 values. Both arguments must have identical types.</p>
2824 This instruction always performs an unsigned division to get the remainder.</p>
2825 <p>Note that unsigned integer remainder and signed integer remainder are
2830 </pre>
2832 </div>
2833 <!-- _______________________________________________________________________ -->
2836 </div>
2842 <pre>
2844 </pre>
2849 signed division of its two operands. This instruction can also take
2851 the elements must be integers.</p>
2857 values. Both arguments must have identical types.</p>
2864 a value. For more information about the difference, see <a
2866 Math Forum</a>. For a table of how this is implemented in various languages,
2869 <p>Note that signed integer remainder and unsigned integer remainder are
2871 <p>Taking the remainder of a division by zero leads to undefined behavior.
2872 Overflow also leads to undefined behavior; this is a rare case, but can occur,
2874 (The remainder doesn't actually overflow, but this rule lets srem be
2875 implemented using instructions that return both the result of the division
2876 and the remainder.)</p>
2879 </pre>
2881 </div>
2882 <!-- _______________________________________________________________________ -->
2890 </pre>
2893 division of its two operands.</p>
2897 of floating point values. Both arguments must have identical types.</p>
2902 The remainder has the same sign as the dividend.</p>
2906 <pre>
2908 </pre>
2909 </div>
2911 <!-- ======================================================================= -->
2913 Operations</a> </div>
2915 <p>Bitwise binary operators are used to do various forms of
2916 bit-twiddling in a program. They are generally very efficient
2917 instructions and can commonly be strength reduced from other
2918 instructions. They require two operands of the same type, execute an operation on them,
2919 and produce a single value. The resulting value is the same type as its operands.</p>
2920 </div>
2922 <!-- _______________________________________________________________________ -->
2924 Instruction</a> </div>
2928 </pre>
2933 the left a specified number of bits.</p>
2954 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> <i>; yields: result=<2 x i32> < i32 2, i32 4></i>
2955 </pre>
2956 </div>
2957 <!-- _______________________________________________________________________ -->
2959 Instruction</a> </div>
2963 </pre>
2967 operand shifted to the right a specified number of bits with zero fill.</p>
2976 <p>This instruction always performs a logical shift right operation. The most
2977 significant bits of the result will be filled with zero bits after the
2984 <pre>
2990 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> <i>; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1></i>
2991 </pre>
2992 </div>
2994 <!-- _______________________________________________________________________ -->
2996 Instruction</a> </div>
3001 </pre>
3005 operand shifted to the right a specified number of bits with sign extension.</p>
3013 <p>This instruction always performs an arithmetic shift right operation,
3014 The most significant bits of the result will be filled with the sign bit
3021 <pre>
3027 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> <i>; yields: result=<2 x i32> < i32 -1, i32 0></i>
3028 </pre>
3029 </div>
3031 <!-- _______________________________________________________________________ -->
3033 Instruction</a> </div>
3039 <pre>
3041 </pre>
3046 its two operands.</p>
3052 values. Both arguments must have identical types.</p>
3056 <p> </p>
3057 <div>
3059 <tbody>
3060 <tr>
3064 </tr>
3065 <tr>
3069 </tr>
3070 <tr>
3074 </tr>
3075 <tr>
3079 </tr>
3080 <tr>
3084 </tr>
3085 </tbody>
3086 </table>
3087 </div>
3089 <pre>
3093 </pre>
3094 </div>
3095 <!-- _______________________________________________________________________ -->
3100 </pre>
3103 or of its two operands.</p>
3108 values. Both arguments must have identical types.</p>
3111 <p> </p>
3112 <div>
3114 <tbody>
3115 <tr>
3119 </tr>
3120 <tr>
3124 </tr>
3125 <tr>
3129 </tr>
3130 <tr>
3134 </tr>
3135 <tr>
3139 </tr>
3140 </tbody>
3141 </table>
3142 </div>
3147 </pre>
3148 </div>
3149 <!-- _______________________________________________________________________ -->
3151 Instruction</a> </div>
3155 </pre>
3163 values. Both arguments must have identical types.</p>
3168 <p> </p>
3169 <div>
3171 <tbody>
3172 <tr>
3176 </tr>
3177 <tr>
3181 </tr>
3182 <tr>
3186 </tr>
3187 <tr>
3191 </tr>
3192 <tr>
3196 </tr>
3197 </tbody>
3198 </table>
3199 </div>
3200 <p> </p>
3206 </pre>
3207 </div>
3209 <!-- ======================================================================= -->
3212 </div>
3216 <p>LLVM supports several instructions to represent vector operations in a
3217 target-independent manner. These instructions cover the element-access and
3218 vector-specific operations needed to process vectors effectively. While LLVM
3219 does directly support these vector operations, many sophisticated algorithms
3220 will want to use target-specific intrinsics to take full advantage of a specific
3221 target.</p>
3223 </div>
3225 <!-- _______________________________________________________________________ -->
3228 </div>
3234 <pre>
3235 <result> = extractelement <n x <ty>> <val>, i32 <idx> <i>; yields <ty></i>
3236 </pre>
3240 <p>
3242 element from a vector at a specified index.
3243 </p>
3248 <p>
3251 an index indicating the position from which to extract the element.
3252 The index may be a variable.</p>
3256 <p>
3257 The result is a scalar of the same type as the element type of
3260 results are undefined.
3261 </p>
3265 <pre>
3267 </pre>
3268 </div>
3271 <!-- _______________________________________________________________________ -->
3274 </div>
3280 <pre>
3281 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> <i>; yields <n x <ty>></i>
3282 </pre>
3286 <p>
3288 element into a vector at a specified index.
3289 </p>
3294 <p>
3297 scalar value whose type must equal the element type of the first
3298 operand. The third operand is an index indicating the position at
3299 which to insert the value. The index may be a variable.</p>
3303 <p>
3308 </p>
3312 <pre>
3314 </pre>
3315 </div>
3317 <!-- _______________________________________________________________________ -->
3320 </div>
3326 <pre>
3327 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> <i>; yields <m x <ty>></i>
3328 </pre>
3332 <p>
3334 from two input vectors, returning a vector with the same element type as
3335 the input and length that is the same as the shuffle mask.
3336 </p>
3340 <p>
3342 with types that match each other. The third argument is a shuffle mask whose
3343 element type is always 'i32'. The result of the instruction is a vector whose
3344 length is the same as the shuffle mask and whose element type is the same as
3345 the element type of the first two operands.
3346 </p>
3348 <p>
3349 The shuffle mask operand is required to be a constant vector with either
3350 constant integer or undef values.
3351 </p>
3355 <p>
3356 The elements of the two input vectors are numbered from left to right across
3357 both of the vectors. The shuffle mask operand specifies, for each element of
3358 the result vector, which element of the two input vectors the result element
3359 gets. The element selector may be undef (meaning "don't care") and the second
3360 operand may be undef if performing a shuffle from only one vector.
3361 </p>
3365 <pre>
3369 <4 x i32> <i32 0, i32 1, i32 2, i32 3> <i>; yields <4 x i32></i> - Identity shuffle.
3373 <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>
3374 </pre>
3375 </div>
3378 <!-- ======================================================================= -->
3381 </div>
3385 <p>LLVM supports several instructions for working with aggregate values.
3386 </p>
3388 </div>
3390 <!-- _______________________________________________________________________ -->
3393 </div>
3399 <pre>
3401 </pre>
3405 <p>
3407 or array element from an aggregate value.
3408 </p>
3413 <p>
3416 type. The operands are constant indices to specify which value to extract
3417 in a similar manner as indices in a
3419 </p>
3423 <p>
3424 The result is the value at the position in the aggregate specified by
3425 the index operands.
3426 </p>
3430 <pre>
3432 </pre>
3433 </div>
3436 <!-- _______________________________________________________________________ -->
3439 </div>
3445 <pre>
3446 <result> = insertvalue <aggregate type> <val>, <ty> <val>, <idx> <i>; yields <n x <ty>></i>
3447 </pre>
3451 <p>
3453 into a struct field or array element in an aggregate.
3454 </p>
3459 <p>
3462 The second operand is a first-class value to insert.
3463 The following operands are constant indices
3464 indicating the position at which to insert the value in a similar manner as
3465 indices in a
3467 The value to insert must have the same type as the value identified
3468 by the indices.
3469 </p>
3473 <p>
3477 </p>
3481 <pre>
3483 </pre>
3484 </div>
3487 <!-- ======================================================================= -->
3490 </div>
3494 <p>A key design point of an SSA-based representation is how it
3495 represents memory. In LLVM, no memory locations are in SSA form, which
3496 makes things very simple. This section describes how to read, write,
3497 allocate, and free memory in LLVM.</p>
3499 </div>
3501 <!-- _______________________________________________________________________ -->
3504 </div>
3510 <pre>
3511 <result> = malloc <type>[, i32 <NumElements>][, align <alignment>] <i>; yields {type*}:result</i>
3512 </pre>
3517 heap and returns a pointer to it. The object is always allocated in the generic
3518 address space (address space zero).</p>
3524 bytes of memory from the operating system and returns a pointer of the
3525 appropriate type to the program. If "NumElements" is specified, it is the
3526 number of elements allocated, otherwise "NumElements" is defaulted to be one.
3527 If a constant alignment is specified, the value result of the allocation is
3528 guaranteed to be aligned to at least that boundary. If not specified, or if
3529 zero, the target can choose to align the allocation on any convenient boundary
3530 compatible with the type.</p>
3537 a pointer is returned. The result of a zero byte allocation is undefined. The
3538 result is null if there is insufficient memory available.</p>
3542 <pre>
3550 </pre>
3552 <p>Note that the code generator does not yet respect the
3553 alignment value.</p>
3555 </div>
3557 <!-- _______________________________________________________________________ -->
3560 </div>
3566 <pre>
3568 </pre>
3573 memory heap to be reallocated in the future.</p>
3579 instruction.</p>
3583 <p>Access to the memory pointed to by the pointer is no longer defined
3584 after this instruction executes. If the pointer is null, the operation
3585 is a noop.</p>
3589 <pre>
3591 free [4 x i8]* %array
3592 </pre>
3593 </div>
3595 <!-- _______________________________________________________________________ -->
3598 </div>
3604 <pre>
3605 <result> = alloca <type>[, i32 <NumElements>][, align <alignment>] <i>; yields {type*}:result</i>
3606 </pre>
3611 currently executing function, to be automatically released when this function
3612 returns to its caller. The object is always allocated in the generic address
3613 space (address space zero).</p>
3618 bytes of memory on the runtime stack, returning a pointer of the
3619 appropriate type to the program. If "NumElements" is specified, it is the
3620 number of elements allocated, otherwise "NumElements" is defaulted to be one.
3621 If a constant alignment is specified, the value result of the allocation is
3622 guaranteed to be aligned to at least that boundary. If not specified, or if
3623 zero, the target can choose to align the allocation on any convenient boundary
3624 compatible with the type.</p>
3630 <p>Memory is allocated; a pointer is returned. The operation is undefined if
3633 instruction is commonly used to represent automatic variables that must
3634 have an address available. When the function returns (either with the <tt><a
3636 instructions), the memory is reclaimed. Allocating zero bytes
3637 is legal, but the result is undefined.</p>
3641 <pre>
3646 </pre>
3647 </div>
3649 <!-- _______________________________________________________________________ -->
3651 Instruction</a> </div>
3654 <pre> <result> = load <ty>* <pointer>[, align <alignment>]<br> <result> = volatile load <ty>* <pointer>[, align <alignment>]<br></pre>
3659 address from which to load. The pointer must point to a <a
3664 instructions. </p>
3665 <p>
3666 The optional constant "align" argument specifies the alignment of the operation
3667 (that is, the alignment of the memory address). A value of 0 or an
3668 omitted "align" argument means that the operation has the preferential
3669 alignment for the target. It is the responsibility of the code emitter
3670 to ensure that the alignment information is correct. Overestimating
3671 the alignment results in an undefined behavior. Underestimating the
3672 alignment may produce less efficient code. An alignment of 1 is always
3673 safe.
3674 </p>
3676 <p>The location of memory pointed to is loaded. If the value being loaded
3677 is of scalar type then the number of bytes read does not exceed the minimum
3678 number of bytes needed to hold all bits of the type. For example, loading an
3681 is undefined if the value was not originally written using a store of the
3682 same type.</p>
3685 <a
3688 </pre>
3689 </div>
3690 <!-- _______________________________________________________________________ -->
3692 Instruction</a> </div>
3695 <pre> store <ty> <value>, <ty>* <pointer>[, align <alignment>] <i>; yields {void}</i>
3696 volatile store <ty> <value>, <ty>* <pointer>[, align <alignment>] <i>; yields {void}</i>
3697 </pre>
3706 optimizer is not allowed to modify the number or order of execution of
3709 <p>
3710 The optional constant "align" argument specifies the alignment of the operation
3711 (that is, the alignment of the memory address). A value of 0 or an
3712 omitted "align" argument means that the operation has the preferential
3713 alignment for the target. It is the responsibility of the code emitter
3714 to ensure that the alignment information is correct. Overestimating
3715 the alignment results in an undefined behavior. Underestimating the
3716 alignment may produce less efficient code. An alignment of 1 is always
3717 safe.
3718 </p>
3723 written does not exceed the minimum number of bytes needed to hold all
3726 size that is not an integral number of bytes, it is unspecified what
3727 happens to the extra bits that do not belong to the type, but they will
3728 typically be overwritten.</p>
3733 </pre>
3734 </div>
3736 <!-- _______________________________________________________________________ -->
3739 </div>
3743 <pre>
3745 </pre>
3749 <p>
3751 subelement of an aggregate data structure. It performs address calculation only
3752 and does not access memory.</p>
3756 <p>The first argument is always a pointer, and forms the basis of the
3757 calculation. The remaining arguments are indices, that indicate which of the
3758 elements of the aggregate object are indexed. The interpretation of each index
3759 is dependent on the type being indexed into. The first index always indexes the
3760 pointer value given as the first argument, the second index indexes a value of
3761 the type pointed to (not necessarily the value directly pointed to, since the
3762 first index can be non-zero), etc. The first type indexed into must be a pointer
3763 value, subsequent types can be arrays, vectors and structs. Note that subsequent
3764 types being indexed into can never be pointers, since that would require loading
3765 the pointer before continuing calculation.</p>
3767 <p>The type of each index argument depends on the type it is indexing into.
3770 integers of any width are allowed (also non-constants).</p>
3772 <p>For example, let's consider a C code fragment and how it gets
3773 compiled to LLVM:</p>
3776 <pre>
3777 struct RT {
3778 char A;
3780 char C;
3781 };
3782 struct ST {
3783 int X;
3784 double Y;
3785 struct RT Z;
3786 };
3788 int *foo(struct ST *s) {
3790 }
3791 </pre>
3792 </div>
3797 <pre>
3801 define i32* %foo(%ST* %s) {
3802 entry:
3804 ret i32* %reg
3805 }
3806 </pre>
3807 </div>
3813 }</tt>' type, a structure. The second index indexes into the third element of
3815 i8 }</tt>' type, another structure. The third index indexes into the second
3817 array. The two dimensions of the array are subscripted into, yielding an
3821 <p>Note that it is perfectly legal to index partially through a
3822 structure, returning a pointer to an inner element. Because of this,
3823 the LLVM code for the given testcase is equivalent to:</p>
3825 <pre>
3826 define i32* %foo(%ST* %s) {
3832 ret i32* %t5
3833 }
3834 </pre>
3836 <p>Note that it is undefined to access an array out of bounds: array
3837 and pointer indexes must always be within the defined bounds of the
3838 array type when accessed with an instruction that dereferences the
3839 pointer (e.g. a load or store instruction). The one exception for
3840 this rule is zero length arrays. These arrays are defined to be
3841 accessible as variable length arrays, which requires access beyond the
3842 zero'th element.</p>
3844 <p>The getelementptr instruction is often confusing. For some more insight
3850 <pre>
3859 </pre>
3860 </div>
3862 <!-- ======================================================================= -->
3864 </div>
3866 <p>The instructions in this category are the conversion instructions (casting)
3867 which all take a single operand and a type. They perform various bit conversions
3868 on the operand.</p>
3869 </div>
3871 <!-- _______________________________________________________________________ -->
3874 </div>
3878 <pre>
3880 </pre>
3883 <p>
3885 </p>
3888 <p>
3896 <p>
3900 It will always truncate bits.</p>
3903 <pre>
3907 </pre>
3908 </div>
3910 <!-- _______________________________________________________________________ -->
3913 </div>
3917 <pre>
3919 </pre>
3940 <pre>
3943 </pre>
3944 </div>
3946 <!-- _______________________________________________________________________ -->
3949 </div>
3953 <pre>
3955 </pre>
3961 <p>
3969 <p>
3977 <pre>
3980 </pre>
3981 </div>
3983 <!-- _______________________________________________________________________ -->
3986 </div>
3992 <pre>
3994 </pre>
4015 <pre>
4018 </pre>
4019 </div>
4021 <!-- _______________________________________________________________________ -->
4024 </div>
4028 <pre>
4030 </pre>
4034 floating point value.</p>
4040 type must be smaller than the destination type.</p>
4050 <pre>
4053 </pre>
4054 </div>
4056 <!-- _______________________________________________________________________ -->
4059 </div>
4063 <pre>
4065 </pre>
4070 </p>
4083 the results are undefined.</p>
4086 <pre>
4090 </pre>
4091 </div>
4093 <!-- _______________________________________________________________________ -->
4096 </div>
4100 <pre>
4102 </pre>
4107 </p>
4120 the results are undefined.</p>
4123 <pre>
4127 </pre>
4128 </div>
4130 <!-- _______________________________________________________________________ -->
4133 </div>
4137 <pre>
4139 </pre>
4154 integer quantity and converts it to the corresponding floating point value. If
4155 the value cannot fit in the floating point value, the results are undefined.</p>
4158 <pre>
4161 </pre>
4162 </div>
4164 <!-- _______________________________________________________________________ -->
4167 </div>
4171 <pre>
4173 </pre>
4188 integer quantity and converts it to the corresponding floating point value. If
4189 the value cannot fit in the floating point value, the results are undefined.</p>
4192 <pre>
4195 </pre>
4196 </div>
4198 <!-- _______________________________________________________________________ -->
4201 </div>
4205 <pre>
4207 </pre>
4221 truncating or zero extending that value to the size of the integer type. If
4225 change.</p>
4228 <pre>
4231 </pre>
4232 </div>
4234 <!-- _______________________________________________________________________ -->
4237 </div>
4241 <pre>
4243 </pre>
4251 value to cast, and a type to cast it to, which must be a
4259 the size of a pointer then a zero extension is done. If they are the same size,
4263 <pre>
4267 </pre>
4268 </div>
4270 <!-- _______________________________________________________________________ -->
4273 </div>
4277 <pre>
4279 </pre>
4289 a non-aggregate first class value, and a type to cast it to, which must also be
4293 type is a pointer, the destination type must also be a pointer. This
4294 instruction supports bitwise conversion of vectors to integers and to vectors
4295 of other types (as long as they have the same size).</p>
4302 converted to other pointer types with this instruction. To convert pointers to
4307 <pre>
4311 </pre>
4312 </div>
4314 <!-- ======================================================================= -->
4318 instructions, which defy better classification.</p>
4319 </div>
4321 <!-- _______________________________________________________________________ -->
4323 </div>
4326 <pre> <result> = icmp <cond> <ty> <op1>, <op2> <i>; yields {i1} or {<N x i1>}:result</i>
4327 </pre>
4330 a vector of boolean values based on comparison
4331 of its two integer, integer vector, or pointer operands.</p>
4334 the condition code indicating the kind of comparison to perform. It is not
4335 a value, just a keyword. The possible condition code are:
4336 </p>
4337 <ol>
4348 </ol>
4352 They must also be identical types.</p>
4356 yields either an <a href="#t_primitive"><tt>i1</tt></a> or vector of <tt>i1</tt> result, as follows:
4357 </p>
4358 <ol>
4361 </li>
4380 </ol>
4382 values are compared as if they were integers.</p>
4383 <p>If the operands are integer vectors, then they are compared
4385 the same number of elements as the values being compared.
4387 </p>
4396 </pre>
4398 <p>Note that the code generator does not yet support vector types with
4401 </div>
4403 <!-- _______________________________________________________________________ -->
4405 </div>
4408 <pre> <result> = fcmp <cond> <ty> <op1>, <op2> <i>; yields {i1} or {<N x i1>}:result</i>
4409 </pre>
4412 or vector of boolean values based on comparison
4413 of its operands.</p>
4414 <p>
4415 If the operands are floating point scalars, then the result
4417 </p>
4418 <p>If the operands are floating point vectors, then the result type
4419 is a vector of boolean with the same number of elements as the
4420 operands being compared.</p>
4423 the condition code indicating the kind of comparison to perform. It is not
4424 a value, just a keyword. The possible condition code are:</p>
4425 <ol>
4442 </ol>
4448 They must have identical types.</p>
4452 If the operands are vectors, then the vectors are compared
4453 element by element.
4454 Each comparison performed
4456 <ol>
4485 </ol>
4492 </pre>
4494 <p>Note that the code generator does not yet support vector types with
4497 </div>
4499 <!-- _______________________________________________________________________ -->
4502 </div>
4511 the SSA graph representing the function.</p>
4514 <p>The type of the incoming values is specified with the first type
4516 as arguments, with one pair for each predecessor basic block of the
4518 type may be used as the value arguments to the PHI node. Only labels
4519 may be used as the label arguments.</p>
4521 <p>There must be no non-phi instructions between the start of a basic
4522 block and the PHI instructions: i.e. PHI instructions must be first in
4523 a basic block.</p>
4525 <p>For the purposes of the SSA form, the use of each incoming value is
4526 deemed to occur on the edge from the corresponding predecessor block
4528 instruction's return value on the same edge).</p>
4533 specified by the pair corresponding to the predecessor basic block that executed
4534 just prior to the current block.</p>
4537 <pre>
4538 Loop: ; Infinite loop that counts from 0 on up...
4539 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
4540 %nextindvar = add i32 %indvar, 1
4541 br label %Loop
4542 </pre>
4543 </div>
4545 <!-- _______________________________________________________________________ -->
4548 </div>
4554 <pre>
4555 <result> = select <i>selty</i> <cond>, <ty> <val1>, <ty> <val2> <i>; yields ty</i>
4558 </pre>
4562 <p>
4564 condition, without branching.
4565 </p>
4570 <p>
4572 a vector of 'i1' values indicating the
4574 type. If the val1/val2 are vectors and
4575 the condition is a scalar, then entire vectors are selected, not
4576 individual elements.
4577 </p>
4581 <p>
4582 If the condition is an i1 and it evaluates to 1, the instruction returns the first
4583 value argument; otherwise, it returns the second value argument.
4584 </p>
4585 <p>
4586 If the condition is a vector of i1, then the value arguments must
4587 be vectors of the same size, and the selection is done element
4588 by element.
4589 </p>
4593 <pre>
4595 </pre>
4597 <p>Note that the code generator does not yet support conditions
4598 with vector type.</p>
4600 </div>
4603 <!-- _______________________________________________________________________ -->
4606 </div>
4611 <pre>
4612 <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>]
4613 </pre>
4623 <ol>
4624 <li>
4626 any allocas or varargs in the caller. If the "tail" marker is present, the
4627 function call is eligible for tail call optimization. Note that calls may
4630 </li>
4631 <li>
4633 convention</a> the call should use. If none is specified, the call defaults
4634 to using C calling conventions.</p>
4635 </li>
4637 <li>
4641 </li>
4643 <li>
4645 the type of the return value. Functions that return no value are marked
4647 </li>
4648 <li>
4650 value being invoked. The argument types must match the types implied by
4651 this signature. This type can be omitted if the function is not varargs
4652 and if the function type does not return a pointer to a function.</p>
4653 </li>
4654 <li>
4656 be invoked. In most cases, this is a direct function invocation, but
4658 to function value.</p>
4659 </li>
4660 <li>
4662 function signature argument types. All arguments must be of
4664 indicates the function accepts a variable number of arguments, the extra
4665 arguments can be specified.</p>
4666 </li>
4667 <li>
4671 </li>
4672 </ol>
4677 transfer to a specified function, with its incoming arguments bound to
4679 instruction in the called function, control flow continues with the
4680 instruction after the function call, and the return value of the
4681 function is bound to the result argument.</p>
4685 <pre>
4686 %retval = call i32 @test(i32 %argc)
4690 call void %foo(i8 97 signext)
4692 %struct.A = type { i32, i8 }
4698 </pre>
4700 </div>
4702 <!-- _______________________________________________________________________ -->
4705 </div>
4711 <pre>
4713 </pre>
4718 the "variable argument" area of a function call. It is used to implement the
4724 the argument. It returns a value of the specified argument type and
4736 <p>It is legal for this instruction to be called in a function which does not
4738 function.</p>
4742 argument.</p>
4748 <p>Note that the code generator does not yet fully support va_arg
4749 on many targets. Also, it does not currently support va_arg with
4750 aggregate types on any target.</p>
4752 </div>
4754 <!-- *********************************************************************** -->
4756 <!-- *********************************************************************** -->
4761 well known names and semantics and are required to follow certain restrictions.
4762 Overall, these intrinsics represent an extension mechanism for the LLVM
4763 language that does not require changing all of the transformations in LLVM when
4764 adding to the language (or the bitcode reader/writer, the parser, etc...).</p>
4767 prefix is reserved in LLVM for intrinsic names; thus, function names may not
4768 begin with this prefix. Intrinsic functions must always be external functions:
4769 you cannot define the body of intrinsic functions. Intrinsic functions may
4770 only be used in call or invoke instructions: it is illegal to take the address
4771 of an intrinsic function. Additionally, because intrinsic functions are part
4772 of the LLVM language, it is required if any are added that they be documented
4773 here.</p>
4775 <p>Some intrinsic functions can be overloaded, i.e., the intrinsic represents
4776 a family of functions that perform the same operation but on different data
4777 types. Because LLVM can represent over 8 million different integer types,
4778 overloading is used commonly to allow an intrinsic function to operate on any
4779 integer type. One or more of the argument types or the result type can be
4780 overloaded to accept any integer type. Argument types may also be defined as
4781 exactly matching a previous argument's type or the result type. This allows an
4782 intrinsic function which accepts multiple arguments, but needs all of them to
4783 be of the same type, to only be overloaded with respect to a single argument or
4784 the result.</p>
4786 <p>Overloaded intrinsics will have the names of its overloaded argument types
4787 encoded into its function name, each preceded by a period. Only those types
4788 which are overloaded result in a name suffix. Arguments whose type is matched
4790 take an integer of any width and returns an integer of exactly the same integer
4791 width. This leads to a family of functions such as
4793 Only one type, the return type, is overloaded, and only one type suffix is
4794 required. Because the argument's type is matched against the return type, it
4795 does not require its own name suffix.</p>
4797 <p>To learn how to add an intrinsic function, please see the
4799 </p>
4801 </div>
4803 <!-- ======================================================================= -->
4806 </div>
4812 intrinsic functions. These functions are related to the similarly
4815 <p>All of these functions operate on arguments that use a
4816 target-specific value type "<tt>va_list</tt>". The LLVM assembly
4817 language reference manual does not define what this type is, so all
4818 transformations should be prepared to handle these functions regardless of
4819 the type used.</p>
4822 instruction and the variable argument handling intrinsic functions are
4823 used.</p>
4826 <pre>
4827 define i32 @test(i32 %X, ...) {
4828 ; Initialize variable argument processing
4829 %ap = alloca i8*
4830 %ap2 = bitcast i8** %ap to i8*
4831 call void @llvm.va_start(i8* %ap2)
4833 ; Read a single integer argument
4834 %tmp = va_arg i8** %ap, i32
4836 ; Demonstrate usage of llvm.va_copy and llvm.va_end
4837 %aq = alloca i8*
4838 %aq2 = bitcast i8** %aq to i8*
4839 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
4840 call void @llvm.va_end(i8* %aq2)
4842 ; Stop processing of arguments.
4843 call void @llvm.va_end(i8* %ap2)
4844 ret i32 %tmp
4845 }
4847 declare void @llvm.va_start(i8*)
4848 declare void @llvm.va_copy(i8*, i8*)
4849 declare void @llvm.va_end(i8*)
4850 </pre>
4851 </div>
4853 </div>
4855 <!-- _______________________________________________________________________ -->
4858 </div>
4876 macro available in C. In a target-dependent way, it initializes the
4880 last argument of the function as the compiler can figure that out.</p>
4882 </div>
4884 <!-- _______________________________________________________________________ -->
4887 </div>
4905 macro available in C. In a target-dependent way, it destroys the
4911 </div>
4913 <!-- _______________________________________________________________________ -->
4916 </div>
4922 <pre>
4924 </pre>
4929 from the source argument list to the destination argument list.</p>
4940 macro available in C. In a target-dependent way, it copies the source
4943 llvm.va_start</a></tt> intrinsic may be arbitrarily complex and require, for
4944 example, memory allocation.</p>
4946 </div>
4948 <!-- ======================================================================= -->
4951 </div>
4955 <p>
4957 Collection</a> (GC) requires the implementation and generation of these
4958 intrinsics.
4962 Front-ends for type-safe garbage collected languages should generate these
4963 intrinsics to make use of the LLVM garbage collectors. For more details, see <a
4965 </p>
4967 <p>The garbage collection intrinsics only operate on objects in the generic
4968 address space (address space zero).</p>
4970 </div>
4972 <!-- _______________________________________________________________________ -->
4975 </div>
4981 <pre>
4982 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
4983 </pre>
4988 the code generator, and allows some metadata to be associated with it.</p>
4992 <p>The first argument specifies the address of a stack object that contains the
4993 root pointer. The second pointer (which must be either a constant or a global
4994 value address) contains the meta-data to be associated with the root.</p>
4999 location. At compile-time, the code generator generates information to allow
5004 </div>
5007 <!-- _______________________________________________________________________ -->
5010 </div>
5016 <pre>
5017 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
5018 </pre>
5023 locations, allowing garbage collector implementations that require read
5024 barriers.</p>
5028 <p>The second argument is the address to read from, which should be an address
5029 allocated from the garbage collector. The first object is a pointer to the
5030 start of the referenced object, if needed by the language runtime (otherwise
5031 null).</p>
5036 instruction, but may be replaced with substantially more complex code by the
5041 </div>
5044 <!-- _______________________________________________________________________ -->
5047 </div>
5053 <pre>
5054 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
5055 </pre>
5060 locations, allowing garbage collector implementations that require write
5061 barriers (such as generational or reference counting collectors).</p>
5065 <p>The first argument is the reference to store, the second is the start of the
5066 object to store it to, and the third is the address of the field of Obj to
5067 store to. If the runtime does not require a pointer to the object, Obj may be
5068 null.</p>
5073 instruction, but may be replaced with substantially more complex code by the
5078 </div>
5082 <!-- ======================================================================= -->
5085 </div>
5088 <p>
5089 These intrinsics are provided by LLVM to expose special features that may only
5090 be implemented with code generator support.
5091 </p>
5093 </div>
5095 <!-- _______________________________________________________________________ -->
5098 </div>
5103 <pre>
5105 </pre>
5109 <p>
5111 target-specific value indicating the return address of the current function
5112 or one of its callers.
5113 </p>
5117 <p>
5118 The argument to this intrinsic indicates which function to return the address
5119 for. Zero indicates the calling function, one indicates its caller, etc. The
5121 </p>
5125 <p>
5127 the return address of the specified call frame, or zero if it cannot be
5128 identified. The value returned by this intrinsic is likely to be incorrect or 0
5129 for arguments other than zero, so it should only be used for debugging purposes.
5130 </p>
5132 <p>
5133 Note that calling this intrinsic does not prevent function inlining or other
5134 aggressive transformations, so the value returned may not be that of the obvious
5135 source-language caller.
5136 </p>
5137 </div>
5140 <!-- _______________________________________________________________________ -->
5143 </div>
5148 <pre>
5150 </pre>
5154 <p>
5156 target-specific frame pointer value for the specified stack frame.
5157 </p>
5161 <p>
5162 The argument to this intrinsic indicates which function to return the frame
5163 pointer for. Zero indicates the calling function, one indicates its caller,
5165 </p>
5169 <p>
5171 the frame address of the specified call frame, or zero if it cannot be
5172 identified. The value returned by this intrinsic is likely to be incorrect or 0
5173 for arguments other than zero, so it should only be used for debugging purposes.
5174 </p>
5176 <p>
5177 Note that calling this intrinsic does not prevent function inlining or other
5178 aggressive transformations, so the value returned may not be that of the obvious
5179 source-language caller.
5180 </p>
5181 </div>
5183 <!-- _______________________________________________________________________ -->
5186 </div>
5191 <pre>
5192 declare i8 *@llvm.stacksave()
5193 </pre>
5197 <p>
5201 features like scoped automatic variable sized arrays in C99.
5202 </p>
5206 <p>
5207 This intrinsic returns a opaque pointer value that can be passed to <a
5214 </p>
5216 </div>
5218 <!-- _______________________________________________________________________ -->
5221 </div>
5226 <pre>
5227 declare void @llvm.stackrestore(i8 * %ptr)
5228 </pre>
5232 <p>
5234 the function stack to the state it was in when the corresponding <a
5236 useful for implementing language features like scoped automatic variable sized
5237 arrays in C99.
5238 </p>
5242 <p>
5244 </p>
5246 </div>
5249 <!-- _______________________________________________________________________ -->
5252 </div>
5257 <pre>
5259 </pre>
5264 <p>
5266 a prefetch instruction if supported; otherwise, it is a noop. Prefetches have
5267 no
5268 effect on the behavior of the program but can change its performance
5269 characteristics.
5270 </p>
5274 <p>
5280 </p>
5284 <p>
5285 This intrinsic does not modify the behavior of the program. In particular,
5286 prefetches cannot trap and do not produce a value. On targets that support this
5287 intrinsic, the prefetch can provide hints to the processor cache for better
5288 performance.
5289 </p>
5291 </div>
5293 <!-- _______________________________________________________________________ -->
5296 </div>
5301 <pre>
5303 </pre>
5308 <p>
5310 (PC) in a region of
5311 code to simulators and other tools. The method is target specific, but it is
5312 expected that the marker will use exported symbols to transmit the PC of the
5313 marker.
5314 The marker makes no guarantees that it will remain with any specific instruction
5315 after optimizations. It is possible that the presence of a marker will inhibit
5316 optimizations. The intended use is to be inserted after optimizations to allow
5317 correlations of simulation runs.
5318 </p>
5322 <p>
5324 </p>
5328 <p>
5329 This intrinsic does not modify the behavior of the program. Backends that do not
5330 support this intrinisic may ignore it.
5331 </p>
5333 </div>
5335 <!-- _______________________________________________________________________ -->
5338 </div>
5343 <pre>
5344 declare i64 @llvm.readcyclecounter( )
5345 </pre>
5350 <p>
5352 counter register (or similar low latency, high accuracy clocks) on those targets
5353 that support it. On X86, it should map to RDTSC. On Alpha, it should map to RPCC.
5354 As the backing counters overflow quickly (on the order of 9 seconds on alpha), this
5355 should only be used for small timings.
5356 </p>
5360 <p>
5361 When directly supported, reading the cycle counter should not modify any memory.
5362 Implementations are allowed to either return a application specific value or a
5363 system wide value. On backends without support, this is lowered to a constant 0.
5364 </p>
5366 </div>
5368 <!-- ======================================================================= -->
5371 </div>
5374 <p>
5375 LLVM provides intrinsics for a few important standard C library functions.
5376 These intrinsics allow source-language front-ends to pass information about the
5377 alignment of the pointer arguments to the code generator, providing opportunity
5378 for more efficient code generation.
5379 </p>
5381 </div>
5383 <!-- _______________________________________________________________________ -->
5386 </div>
5391 <p>This is an overloaded intrinsic. You can use llvm.memcpy on any integer bit
5392 width. Not all targets support all bit widths however.</p>
5393 <pre>
5402 </pre>
5406 <p>
5408 location to the destination location.
5409 </p>
5411 <p>
5413 intrinsics do not return a value, and takes an extra alignment argument.
5414 </p>
5418 <p>
5419 The first argument is a pointer to the destination, the second is a pointer to
5420 the source. The third argument is an integer argument
5421 specifying the number of bytes to copy, and the fourth argument is the alignment
5422 of the source and destination locations.
5423 </p>
5425 <p>
5427 the caller guarantees that both the source and destination pointers are aligned
5428 to that boundary.
5429 </p>
5433 <p>
5435 location to the destination location, which are not allowed to overlap. It
5436 copies "len" bytes of memory over. If the argument is known to be aligned to
5437 some boundary, this can be specified as the fourth argument, otherwise it should
5439 </p>
5440 </div>
5443 <!-- _______________________________________________________________________ -->
5446 </div>
5451 <p>This is an overloaded intrinsic. You can use llvm.memmove on any integer bit
5452 width. Not all targets support all bit widths however.</p>
5453 <pre>
5462 </pre>
5466 <p>
5468 location to the destination location. It is similar to the
5470 </p>
5472 <p>
5474 intrinsics do not return a value, and takes an extra alignment argument.
5475 </p>
5479 <p>
5480 The first argument is a pointer to the destination, the second is a pointer to
5481 the source. The third argument is an integer argument
5482 specifying the number of bytes to copy, and the fourth argument is the alignment
5483 of the source and destination locations.
5484 </p>
5486 <p>
5488 the caller guarantees that the source and destination pointers are aligned to
5489 that boundary.
5490 </p>
5494 <p>
5496 location to the destination location, which may overlap. It
5497 copies "len" bytes of memory over. If the argument is known to be aligned to
5498 some boundary, this can be specified as the fourth argument, otherwise it should
5500 </p>
5501 </div>
5504 <!-- _______________________________________________________________________ -->
5507 </div>
5512 <p>This is an overloaded intrinsic. You can use llvm.memset on any integer bit
5513 width. Not all targets support all bit widths however.</p>
5514 <pre>
5523 </pre>
5527 <p>
5529 byte value.
5530 </p>
5532 <p>
5534 does not return a value, and takes an extra alignment argument.
5535 </p>
5539 <p>
5540 The first argument is a pointer to the destination to fill, the second is the
5541 byte value to fill it with, the third argument is an integer
5542 argument specifying the number of bytes to fill, and the fourth argument is the
5543 known alignment of destination location.
5544 </p>
5546 <p>
5548 the caller guarantees that the destination pointer is aligned to that boundary.
5549 </p>
5553 <p>
5555 the
5556 destination location. If the argument is known to be aligned to some boundary,
5557 this can be specified as the fourth argument, otherwise it should be set to 0 or
5558 1.
5559 </p>
5560 </div>
5563 <!-- _______________________________________________________________________ -->
5566 </div>
5572 floating point or vector of floating point type. Not all targets support all
5573 types however.</p>
5574 <pre>
5575 declare float @llvm.sqrt.f32(float %Val)
5576 declare double @llvm.sqrt.f64(double %Val)
5577 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
5578 declare fp128 @llvm.sqrt.f128(fp128 %Val)
5579 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
5580 </pre>
5584 <p>
5588 negative numbers other than -0.0 (which allows for better optimization, because
5590 defined to return -0.0 like IEEE sqrt.
5591 </p>
5595 <p>
5596 The argument and return value are floating point numbers of the same type.
5597 </p>
5601 <p>
5602 This function returns the sqrt of the specified operand if it is a nonnegative
5603 floating point number.
5604 </p>
5605 </div>
5607 <!-- _______________________________________________________________________ -->
5610 </div>
5616 floating point or vector of floating point type. Not all targets support all
5617 types however.</p>
5618 <pre>
5619 declare float @llvm.powi.f32(float %Val, i32 %power)
5620 declare double @llvm.powi.f64(double %Val, i32 %power)
5621 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
5622 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
5623 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
5624 </pre>
5628 <p>
5630 specified (positive or negative) power. The order of evaluation of
5631 multiplications is not defined. When a vector of floating point type is
5632 used, the second argument remains a scalar integer value.
5633 </p>
5637 <p>
5638 The second argument is an integer power, and the first is a value to raise to
5639 that power.
5640 </p>
5644 <p>
5645 This function returns the first value raised to the second power with an
5646 unspecified sequence of rounding operations.</p>
5647 </div>
5649 <!-- _______________________________________________________________________ -->
5652 </div>
5658 floating point or vector of floating point type. Not all targets support all
5659 types however.</p>
5660 <pre>
5661 declare float @llvm.sin.f32(float %Val)
5662 declare double @llvm.sin.f64(double %Val)
5663 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
5664 declare fp128 @llvm.sin.f128(fp128 %Val)
5665 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
5666 </pre>
5670 <p>
5672 </p>
5676 <p>
5677 The argument and return value are floating point numbers of the same type.
5678 </p>
5682 <p>
5683 This function returns the sine of the specified operand, returning the
5685 conditions in the same way.</p>
5686 </div>
5688 <!-- _______________________________________________________________________ -->
5691 </div>
5697 floating point or vector of floating point type. Not all targets support all
5698 types however.</p>
5699 <pre>
5700 declare float @llvm.cos.f32(float %Val)
5701 declare double @llvm.cos.f64(double %Val)
5702 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
5703 declare fp128 @llvm.cos.f128(fp128 %Val)
5704 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
5705 </pre>
5709 <p>
5711 </p>
5715 <p>
5716 The argument and return value are floating point numbers of the same type.
5717 </p>
5721 <p>
5722 This function returns the cosine of the specified operand, returning the
5724 conditions in the same way.</p>
5725 </div>
5727 <!-- _______________________________________________________________________ -->
5730 </div>
5736 floating point or vector of floating point type. Not all targets support all
5737 types however.</p>
5738 <pre>
5739 declare float @llvm.pow.f32(float %Val, float %Power)
5740 declare double @llvm.pow.f64(double %Val, double %Power)
5741 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
5742 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
5743 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
5744 </pre>
5748 <p>
5750 specified (positive or negative) power.
5751 </p>
5755 <p>
5756 The second argument is a floating point power, and the first is a value to
5757 raise to that power.
5758 </p>
5762 <p>
5763 This function returns the first value raised to the second power,
5764 returning the
5766 conditions in the same way.</p>
5767 </div>
5770 <!-- ======================================================================= -->
5773 </div>
5776 <p>
5777 LLVM provides intrinsics for a few important bit manipulation operations.
5778 These allow efficient code generation for some algorithms.
5779 </p>
5781 </div>
5783 <!-- _______________________________________________________________________ -->
5786 </div>
5791 <p>This is an overloaded intrinsic function. You can use bswap on any integer
5793 <pre>
5797 </pre>
5801 <p>
5803 values with an even number of bytes (positive multiple of 16 bits). These are
5804 useful for performing operations on data that is not in the target's native
5805 byte order.
5806 </p>
5810 <p>
5813 intrinsic returns an i32 value that has the four bytes of the input i32
5818 </p>
5820 </div>
5822 <!-- _______________________________________________________________________ -->
5825 </div>
5830 <p>This is an overloaded intrinsic. You can use llvm.ctpop on any integer bit
5831 width. Not all targets support all bit widths however.</p>
5832 <pre>
5838 </pre>
5842 <p>
5844 value.
5845 </p>
5849 <p>
5850 The only argument is the value to be counted. The argument may be of any
5851 integer type. The return type must match the argument type.
5852 </p>
5856 <p>
5858 </p>
5859 </div>
5861 <!-- _______________________________________________________________________ -->
5864 </div>
5870 integer bit width. Not all targets support all bit widths however.</p>
5871 <pre>
5877 </pre>
5881 <p>
5883 leading zeros in a variable.
5884 </p>
5888 <p>
5889 The only argument is the value to be counted. The argument may be of any
5890 integer type. The return type must match the argument type.
5891 </p>
5895 <p>
5897 in a variable. If the src == 0 then the result is the size in bits of the type
5899 </p>
5900 </div>
5904 <!-- _______________________________________________________________________ -->
5907 </div>
5913 integer bit width. Not all targets support all bit widths however.</p>
5914 <pre>
5920 </pre>
5924 <p>
5926 trailing zeros.
5927 </p>
5931 <p>
5932 The only argument is the value to be counted. The argument may be of any
5933 integer type. The return type must match the argument type.
5934 </p>
5938 <p>
5940 in a variable. If the src == 0 then the result is the size in bits of the type
5942 </p>
5943 </div>
5945 <!-- _______________________________________________________________________ -->
5948 </div>
5954 on any integer bit width.</p>
5955 <pre>
5956 declare i17 @llvm.part.select.i17 (i17 %val, i32 %loBit, i32 %hiBit)
5957 declare i29 @llvm.part.select.i29 (i29 %val, i32 %loBit, i32 %hiBit)
5958 </pre>
5962 range of bits from an integer value and returns them in the same bit width as
5963 the original value.</p>
5967 any bit width but they must have the same bit width. The second and third
5974 operates in forward mode.</p>
5978 <ol>
5982 to determine the number of bits to retain.</li>
5985 </ol>
5986 <p>In reverse mode, a similar computation is made except that the bits are
5991 </div>
5995 </div>
6001 on any integer bit width.</p>
6002 <pre>
6003 declare i17 @llvm.part.set.i17.i9 (i17 %val, i9 %repl, i32 %lo, i32 %hi)
6004 declare i29 @llvm.part.set.i29.i9 (i29 %val, i9 %repl, i32 %lo, i32 %hi)
6005 </pre>
6009 of bits in an integer value with another integer value. It returns the integer
6010 with the replaced bits.</p>
6017 type since they specify only a bit index.</p>
6023 operates in forward mode.</p>
6026 truncating it down to the size of the replacement area or zero extending it
6027 up to that size.</p>
6034 <p>In reverse mode, a similar computation is made except that the bits are
6040 <pre>
6046 </pre>
6048 </div>
6050 <!-- ======================================================================= -->
6053 </div>
6056 <p>
6057 LLVM provides intrinsics for some arithmetic with overflow operations.
6058 </p>
6060 </div>
6062 <!-- _______________________________________________________________________ -->
6065 </div>
6072 on any integer bit width.</p>
6074 <pre>
6075 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
6076 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
6077 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
6078 </pre>
6083 a signed addition of the two arguments, and indicate whether an overflow
6084 occurred during the signed summation.</p>
6088 <p>The arguments (%a and %b) and the first element of the result structure may
6089 be of integer types of any bit width, but they must have the same bit width. The
6096 a signed addition of the two variables. They return a structure — the
6097 first element of which is the signed summation, and the second element of which
6098 is a bit specifying if the signed summation resulted in an overflow.</p>
6101 <pre>
6102 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
6103 %sum = extractvalue {i32, i1} %res, 0
6104 %obit = extractvalue {i32, i1} %res, 1
6105 br i1 %obit, label %overflow, label %normal
6106 </pre>
6108 </div>
6110 <!-- _______________________________________________________________________ -->
6113 </div>
6120 on any integer bit width.</p>
6122 <pre>
6123 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
6124 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
6125 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
6126 </pre>
6131 an unsigned addition of the two arguments, and indicate whether a carry occurred
6132 during the unsigned summation.</p>
6136 <p>The arguments (%a and %b) and the first element of the result structure may
6137 be of integer types of any bit width, but they must have the same bit width. The
6144 an unsigned addition of the two arguments. They return a structure — the
6145 first element of which is the sum, and the second element of which is a bit
6146 specifying if the unsigned summation resulted in a carry.</p>
6149 <pre>
6150 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
6151 %sum = extractvalue {i32, i1} %res, 0
6152 %obit = extractvalue {i32, i1} %res, 1
6153 br i1 %obit, label %carry, label %normal
6154 </pre>
6156 </div>
6158 <!-- _______________________________________________________________________ -->
6161 </div>
6168 on any integer bit width.</p>
6170 <pre>
6171 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
6172 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
6173 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
6174 </pre>
6179 a signed subtraction of the two arguments, and indicate whether an overflow
6180 occurred during the signed subtraction.</p>
6184 <p>The arguments (%a and %b) and the first element of the result structure may
6185 be of integer types of any bit width, but they must have the same bit width. The
6192 a signed subtraction of the two arguments. They return a structure — the
6193 first element of which is the subtraction, and the second element of which is a bit
6194 specifying if the signed subtraction resulted in an overflow.</p>
6197 <pre>
6198 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
6199 %sum = extractvalue {i32, i1} %res, 0
6200 %obit = extractvalue {i32, i1} %res, 1
6201 br i1 %obit, label %overflow, label %normal
6202 </pre>
6204 </div>
6206 <!-- _______________________________________________________________________ -->
6209 </div>
6216 on any integer bit width.</p>
6218 <pre>
6219 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
6220 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
6221 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
6222 </pre>
6227 an unsigned subtraction of the two arguments, and indicate whether an overflow
6228 occurred during the unsigned subtraction.</p>
6232 <p>The arguments (%a and %b) and the first element of the result structure may
6233 be of integer types of any bit width, but they must have the same bit width. The
6240 an unsigned subtraction of the two arguments. They return a structure — the
6241 first element of which is the subtraction, and the second element of which is a bit
6242 specifying if the unsigned subtraction resulted in an overflow.</p>
6245 <pre>
6246 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
6247 %sum = extractvalue {i32, i1} %res, 0
6248 %obit = extractvalue {i32, i1} %res, 1
6249 br i1 %obit, label %overflow, label %normal
6250 </pre>
6252 </div>
6254 <!-- _______________________________________________________________________ -->
6257 </div>
6264 on any integer bit width.</p>
6266 <pre>
6267 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
6268 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
6269 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
6270 </pre>
6275 a signed multiplication of the two arguments, and indicate whether an overflow
6276 occurred during the signed multiplication.</p>
6280 <p>The arguments (%a and %b) and the first element of the result structure may
6281 be of integer types of any bit width, but they must have the same bit width. The
6288 a signed multiplication of the two arguments. They return a structure —
6289 the first element of which is the multiplication, and the second element of
6290 which is a bit specifying if the signed multiplication resulted in an
6291 overflow.</p>
6294 <pre>
6295 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
6296 %sum = extractvalue {i32, i1} %res, 0
6297 %obit = extractvalue {i32, i1} %res, 1
6298 br i1 %obit, label %overflow, label %normal
6299 </pre>
6301 </div>
6303 <!-- _______________________________________________________________________ -->
6306 </div>
6313 on any integer bit width.</p>
6315 <pre>
6316 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
6317 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
6318 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
6319 </pre>
6324 a unsigned multiplication of the two arguments, and indicate whether an overflow
6325 occurred during the unsigned multiplication.</p>
6329 <p>The arguments (%a and %b) and the first element of the result structure may
6330 be of integer types of any bit width, but they must have the same bit width. The
6333 multiplication.</p>
6338 an unsigned multiplication of the two arguments. They return a structure —
6339 the first element of which is the multiplication, and the second element of
6340 which is a bit specifying if the unsigned multiplication resulted in an
6341 overflow.</p>
6344 <pre>
6345 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
6346 %sum = extractvalue {i32, i1} %res, 0
6347 %obit = extractvalue {i32, i1} %res, 1
6348 br i1 %obit, label %overflow, label %normal
6349 </pre>
6351 </div>
6353 <!-- ======================================================================= -->
6356 </div>
6359 <p>
6361 are described in the <a
6363 Debugging</a> document.
6364 </p>
6365 </div>
6368 <!-- ======================================================================= -->
6371 </div>
6374 <p> The LLVM exception handling intrinsics (which all start with
6378 </div>
6380 <!-- ======================================================================= -->
6383 </div>
6386 <p>
6387 This intrinsic makes it possible to excise one parameter, marked with
6389 function pointer lacking the nest parameter - the caller does not need
6390 to provide a value for it. Instead, the value to use is stored in
6391 advance in a "trampoline", a block of memory usually allocated
6392 on the stack, which also contains code to splice the nest value into the
6393 argument list. This is used to implement the GCC nested function address
6394 extension.
6395 </p>
6396 <p>
6397 For example, if the function is
6400 <pre>
6403 %p = call i8* @llvm.init.trampoline( i8* %tramp1, i8* bitcast (i32 (i8* nest , i32, i32)* @f to i8*), i8* %nval )
6404 %fp = bitcast i8* %p to i32 (i32, i32)*
6405 </pre>
6408 </div>
6410 <!-- _______________________________________________________________________ -->
6413 </div>
6416 <pre>
6418 </pre>
6420 <p>
6422 and returns a function pointer suitable for executing it.
6423 </p>
6425 <p>
6428 and sufficiently aligned block of memory; this memory is written to by the
6429 intrinsic. Note that the size and the alignment are target-specific - LLVM
6430 currently provides no portable way of determining them, so a front-end that
6431 generates this intrinsic needs to have some target-specific knowledge.
6433 </p>
6435 <p>
6437 dependent code, turning it into a function. A pointer to this function is
6438 returned, but needs to be bitcast to an
6440 before being called. The new function's signature is the same as that of
6443 of pointer type. Calling the new function is equivalent to calling
6447 modified, then the effect of any later call to the returned function pointer is
6448 undefined.
6449 </p>
6450 </div>
6452 <!-- ======================================================================= -->
6455 </div>
6458 <p>
6459 These intrinsic functions expand the "universal IR" of LLVM to represent
6460 hardware constructs for atomic operations and memory synchronization. This
6461 provides an interface to the hardware, not an interface to the programmer. It
6462 is aimed at a low enough level to allow any programming models or APIs
6463 (Application Programming Interfaces) which
6464 need atomic behaviors to map cleanly onto it. It is also modeled primarily on
6465 hardware behavior. Just as hardware provides a "universal IR" for source
6466 languages, it also provides a starting point for developing a "universal"
6467 atomic operation and synchronization IR.
6468 </p>
6469 <p>
6471 software transaction memory systems, atomic primitives, and intrinsic
6472 functions as found in BSD, GNU libc, atomic_ops, APR, and other system and
6473 application libraries. The hardware interface provided by LLVM should allow
6474 a clean implementation of all of these APIs and parallel programming models.
6475 No one model or paradigm should be selected above others unless the hardware
6476 itself ubiquitously does so.
6478 </p>
6479 </div>
6481 <!-- _______________________________________________________________________ -->
6484 </div>
6487 <pre>
6491 </pre>
6493 <p>
6495 specific pairs of memory access types.
6496 </p>
6498 <p>
6500 The first four arguments enables a specific barrier as listed below. The fith
6501 argument specifies that the barrier applies to io or device or uncached memory.
6503 </p>
6504 <ul>
6510 </ul>
6512 <p>
6513 This intrinsic causes the system to enforce some ordering constraints upon
6514 the loads and stores of the program. This barrier does not indicate
6516 which they occur. For any of the specified pairs of load and store operations
6517 (f.ex. load-load, or store-load), all of the first operations preceding the
6518 barrier will complete before any of the second operations succeeding the
6519 barrier begin. Specifically the semantics for each pairing is as follows:
6520 </p>
6521 <ul>
6523 after the barrier begins.</li>
6526 store after the barrier begins.</li>
6528 store after the barrier begins.</li>
6530 load after the barrier begins.</li>
6531 </ul>
6532 <p>
6533 These semantics are applied with a logical "and" behavior when more than one
6534 is enabled in a single memory barrier intrinsic.
6535 </p>
6536 <p>
6537 Backends may implement stronger barriers than those requested when they do not
6538 support as fine grained a barrier as requested. Some architectures do not
6539 need all types of barriers and on such architectures, these become noops.
6540 </p>
6542 <pre>
6543 %ptr = malloc i32
6544 store i32 4, %ptr
6547 call void @llvm.memory.barrier( i1 false, i1 true, i1 false, i1 false )
6550 </pre>
6551 </div>
6553 <!-- _______________________________________________________________________ -->
6556 </div>
6559 <p>
6561 any integer bit width and for different address spaces. Not all targets
6562 support all bit widths however.</p>
6564 <pre>
6566 declare i16 @llvm.atomic.cmp.swap.i16.p0i16( i16* <ptr>, i16 <cmp>, i16 <val> )
6567 declare i32 @llvm.atomic.cmp.swap.i32.p0i32( i32* <ptr>, i32 <cmp>, i32 <val> )
6568 declare i64 @llvm.atomic.cmp.swap.i64.p0i64( i64* <ptr>, i64 <cmp>, i64 <val> )
6570 </pre>
6572 <p>
6573 This loads a value in memory and compares it to a given value. If they are
6574 equal, it stores a new value into the memory.
6575 </p>
6577 <p>
6581 this integer type. While any bit width integer may be used, targets may only
6582 lower representations they support in hardware.
6584 </p>
6586 <p>
6587 This entire intrinsic must be executed atomically. It first loads the value
6590 loaded value is yielded in all cases. This provides the equivalent of an
6591 atomic compare-and-swap operation within the SSA framework.
6592 </p>
6595 <pre>
6596 %ptr = malloc i32
6597 store i32 4, %ptr
6600 %result1 = call i32 @llvm.atomic.cmp.swap.i32.p0i32( i32* %ptr, i32 4, %val1 )
6606 %result2 = call i32 @llvm.atomic.cmp.swap.i32.p0i32( i32* %ptr, i32 5, %val2 )
6611 </pre>
6612 </div>
6614 <!-- _______________________________________________________________________ -->
6617 </div>
6621 <p>
6623 integer bit width. Not all targets support all bit widths however.</p>
6624 <pre>
6630 </pre>
6632 <p>
6636 </p>
6639 <p>
6643 integer type. The targets may only lower integer representations they
6644 support.
6645 </p>
6647 <p>
6650 equivalent of an atomic swap operation within the SSA framework.
6652 </p>
6654 <pre>
6655 %ptr = malloc i32
6656 store i32 4, %ptr
6659 %result1 = call i32 @llvm.atomic.swap.i32.p0i32( i32* %ptr, i32 %val1 )
6665 %result2 = call i32 @llvm.atomic.swap.i32.p0i32( i32* %ptr, i32 %val2 )
6670 </pre>
6671 </div>
6673 <!-- _______________________________________________________________________ -->
6677 </div>
6680 <p>
6682 integer bit width. Not all targets support all bit widths however.</p>
6683 <pre>
6689 </pre>
6691 <p>
6694 </p>
6696 <p>
6698 The intrinsic takes two arguments, the first a pointer to an integer value
6699 and the second an integer value. The result is also an integer value. These
6700 integer types can have any bit width, but they must all have the same bit
6701 width. The targets may only lower integer representations they support.
6702 </p>
6704 <p>
6705 This intrinsic does a series of operations atomically. It first loads the
6708 </p>
6711 <pre>
6712 %ptr = malloc i32
6713 store i32 4, %ptr
6714 %result1 = call i32 @llvm.atomic.load.add.i32.p0i32( i32* %ptr, i32 4 )
6716 %result2 = call i32 @llvm.atomic.load.add.i32.p0i32( i32* %ptr, i32 2 )
6718 %result3 = call i32 @llvm.atomic.load.add.i32.p0i32( i32* %ptr, i32 5 )
6721 </pre>
6722 </div>
6724 <!-- _______________________________________________________________________ -->
6728 </div>
6731 <p>
6733 any integer bit width and for different address spaces. Not all targets
6734 support all bit widths however.</p>
6735 <pre>
6741 </pre>
6743 <p>
6746 </p>
6748 <p>
6750 The intrinsic takes two arguments, the first a pointer to an integer value
6751 and the second an integer value. The result is also an integer value. These
6752 integer types can have any bit width, but they must all have the same bit
6753 width. The targets may only lower integer representations they support.
6754 </p>
6756 <p>
6757 This intrinsic does a series of operations atomically. It first loads the
6760 </p>
6763 <pre>
6764 %ptr = malloc i32
6765 store i32 8, %ptr
6766 %result1 = call i32 @llvm.atomic.load.sub.i32.p0i32( i32* %ptr, i32 4 )
6768 %result2 = call i32 @llvm.atomic.load.sub.i32.p0i32( i32* %ptr, i32 2 )
6770 %result3 = call i32 @llvm.atomic.load.sub.i32.p0i32( i32* %ptr, i32 5 )
6773 </pre>
6774 </div>
6776 <!-- _______________________________________________________________________ -->
6783 </div>
6786 <p>
6790 address spaces. Not all targets support all bit widths however.</p>
6791 <pre>
6797 </pre>
6799 <pre>
6805 </pre>
6807 <pre>
6813 </pre>
6815 <pre>
6821 </pre>
6823 <p>
6827 </p>
6829 <p>
6831 These intrinsics take two arguments, the first a pointer to an integer value
6832 and the second an integer value. The result is also an integer value. These
6833 integer types can have any bit width, but they must all have the same bit
6834 width. The targets may only lower integer representations they support.
6835 </p>
6837 <p>
6838 These intrinsics does a series of operations atomically. They first load the
6842 </p>
6845 <pre>
6846 %ptr = malloc i32
6847 store i32 0x0F0F, %ptr
6848 %result0 = call i32 @llvm.atomic.load.nand.i32.p0i32( i32* %ptr, i32 0xFF )
6850 %result1 = call i32 @llvm.atomic.load.and.i32.p0i32( i32* %ptr, i32 0xFF )
6852 %result2 = call i32 @llvm.atomic.load.or.i32.p0i32( i32* %ptr, i32 0F )
6854 %result3 = call i32 @llvm.atomic.load.xor.i32.p0i32( i32* %ptr, i32 0F )
6857 </pre>
6858 </div>
6861 <!-- _______________________________________________________________________ -->
6868 </div>
6871 <p>
6875 address spaces. Not all targets
6876 support all bit widths however.</p>
6877 <pre>
6883 </pre>
6885 <pre>
6891 </pre>
6893 <pre>
6899 </pre>
6901 <pre>
6907 </pre>
6909 <p>
6910 These intrinsics takes the signed or unsigned minimum or maximum of
6913 </p>
6915 <p>
6917 These intrinsics take two arguments, the first a pointer to an integer value
6918 and the second an integer value. The result is also an integer value. These
6919 integer types can have any bit width, but they must all have the same bit
6920 width. The targets may only lower integer representations they support.
6921 </p>
6923 <p>
6924 These intrinsics does a series of operations atomically. They first load the
6928 </p>
6931 <pre>
6932 %ptr = malloc i32
6933 store i32 7, %ptr
6934 %result0 = call i32 @llvm.atomic.load.min.i32.p0i32( i32* %ptr, i32 -2 )
6936 %result1 = call i32 @llvm.atomic.load.max.i32.p0i32( i32* %ptr, i32 8 )
6938 %result2 = call i32 @llvm.atomic.load.umin.i32.p0i32( i32* %ptr, i32 10 )
6940 %result3 = call i32 @llvm.atomic.load.umax.i32.p0i32( i32* %ptr, i32 30 )
6943 </pre>
6944 </div>
6946 <!-- ======================================================================= -->
6949 </div>
6952 <p> This class of intrinsics is designed to be generic and has
6953 no specific purpose. </p>
6954 </div>
6956 <!-- _______________________________________________________________________ -->
6959 </div>
6964 <pre>
6965 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int> )
6966 </pre>
6970 <p>
6972 </p>
6976 <p>
6977 The first argument is a pointer to a value, the second is a pointer to a
6978 global string, the third is a pointer to a global string which is the source
6979 file name, and the last argument is the line number.
6980 </p>
6984 <p>
6985 This intrinsic allows annotation of local variables with arbitrary strings.
6986 This can be useful for special purpose optimizations that want to look for these
6987 annotations. These have no other defined use, they are ignored by code
6988 generation and optimization.
6989 </p>
6990 </div>
6992 <!-- _______________________________________________________________________ -->
6995 </div>
7001 any integer bit width.
7002 </p>
7003 <pre>
7004 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int> )
7005 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int> )
7006 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int> )
7007 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int> )
7008 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int> )
7009 </pre>
7013 <p>
7015 </p>
7019 <p>
7020 The first argument is an integer value (result of some expression),
7021 the second is a pointer to a global string, the third is a pointer to a global
7022 string which is the source file name, and the last argument is the line number.
7023 It returns the value of the first argument.
7024 </p>
7028 <p>
7029 This intrinsic allows annotations to be put on arbitrary expressions
7030 with arbitrary strings. This can be useful for special purpose optimizations
7031 that want to look for these annotations. These have no other defined use, they
7032 are ignored by code generation and optimization.
7033 </p>
7034 </div>
7036 <!-- _______________________________________________________________________ -->
7039 </div>
7044 <pre>
7045 declare void @llvm.trap()
7046 </pre>
7050 <p>
7052 </p>
7056 <p>
7057 None
7058 </p>
7062 <p>
7063 This intrinsics is lowered to the target dependent trap instruction. If the
7064 target does not have a trap instruction, this intrinsic will be lowered to the
7065 call of the abort() function.
7066 </p>
7067 </div>
7069 <!-- _______________________________________________________________________ -->
7072 </div>
7075 <pre>
7078 </pre>
7080 <p>
7083 it is placed on the stack before local variables.
7084 </p>
7086 <p>
7088 first argument is the value loaded from the stack guard
7090 has enough space to hold the value of the guard.
7091 </p>
7093 <p>
7094 This intrinsic causes the prologue/epilogue inserter to force the position of
7096 stack. This is to ensure that if a local variable on the stack is overwritten,
7097 it will destroy the value of the guard. When the function exits, the guard on
7098 the stack is checked against the original guard. If they're different, then
7100 </p>
7101 </div>
7103 <!-- *********************************************************************** -->
7104 <hr>
7105 <address>
7113 Last modified: $Date$
7114 </address>
7116 </body>
7117 </html>