| blob | ff707a45b600efab4cfbe9e4e6031704cacaee84 |
3 <html>
4 <head>
11 </head>
13 <body>
16 <ol>
21 <ol>
24 <ol>
40 </ol>
41 </li>
55 </ol>
56 </li>
58 <ol>
61 <ol>
67 </ol>
68 </li>
70 <ol>
72 <ol>
78 </ol>
79 </li>
83 </ol>
84 </li>
86 </ol>
87 </li>
89 <ol>
97 </ol>
98 </li>
100 <ol>
103 </ol>
104 </li>
106 <ol>
109 Global Variable</a></li>
111 Global Variable</a></li>
113 Global Variable</a></li>
114 </ol>
115 </li>
117 <ol>
119 <ol>
127 </ol>
128 </li>
130 <ol>
143 </ol>
144 </li>
146 <ol>
153 </ol>
154 </li>
156 <ol>
160 </ol>
161 </li>
163 <ol>
166 </ol>
167 </li>
169 <ol>
174 </ol>
175 </li>
177 <ol>
190 </ol>
191 </li>
193 <ol>
200 </ol>
201 </li>
202 </ol>
203 </li>
205 <ol>
207 <ol>
211 </ol>
212 </li>
214 <ol>
218 </ol>
219 </li>
221 <ol>
229 </ol>
230 </li>
232 <ol>
241 </ol>
242 </li>
244 <ol>
249 </ol>
250 </li>
252 <ol>
259 </ol>
260 </li>
262 <ol>
265 </ol>
266 </li>
270 <ol>
272 </ol>
273 </li>
275 <ol>
289 </ol>
290 </li>
292 <ol>
297 </ol>
298 </li>
300 <ol>
311 </ol>
312 </li>
313 </ol>
314 </li>
315 </ol>
320 </div>
322 <!-- *********************************************************************** -->
324 <!-- *********************************************************************** -->
328 <p>This document is a reference manual for the LLVM assembly language. LLVM is
329 a Static Single Assignment (SSA) based representation that provides type
330 safety, low-level operations, flexibility, and the capability of representing
331 'all' high-level languages cleanly. It is the common code representation
332 used throughout all phases of the LLVM compilation strategy.</p>
334 </div>
336 <!-- *********************************************************************** -->
338 <!-- *********************************************************************** -->
342 <p>The LLVM code representation is designed to be used in three different forms:
343 as an in-memory compiler IR, as an on-disk bitcode representation (suitable
344 for fast loading by a Just-In-Time compiler), and as a human readable
345 assembly language representation. This allows LLVM to provide a powerful
346 intermediate representation for efficient compiler transformations and
347 analysis, while providing a natural means to debug and visualize the
348 transformations. The three different forms of LLVM are all equivalent. This
349 document describes the human readable representation and notation.</p>
351 <p>The LLVM representation aims to be light-weight and low-level while being
352 expressive, typed, and extensible at the same time. It aims to be a
353 "universal IR" of sorts, by being at a low enough level that high-level ideas
354 may be cleanly mapped to it (similar to how microprocessors are "universal
355 IR's", allowing many source languages to be mapped to them). By providing
356 type information, LLVM can be used as the target of optimizations: for
357 example, through pointer analysis, it can be proven that a C automatic
358 variable is never accessed outside of the current function, allowing it to
359 be promoted to a simple SSA value instead of a memory location.</p>
361 </div>
363 <!-- _______________________________________________________________________ -->
368 <p>It is important to note that this document describes 'well formed' LLVM
369 assembly language. There is a difference between what the parser accepts and
370 what is considered 'well formed'. For example, the following instruction is
371 syntactically okay, but not well formed:</p>
375 </pre>
378 LLVM infrastructure provides a verification pass that may be used to verify
379 that an LLVM module is well formed. This pass is automatically run by the
380 parser after parsing input assembly and by the optimizer before it outputs
381 bitcode. The violations pointed out by the verifier pass indicate bugs in
382 transformation passes or input to the parser.</p>
384 </div>
386 <!-- Describe the typesetting conventions here. -->
388 <!-- *********************************************************************** -->
390 <!-- *********************************************************************** -->
394 <p>LLVM identifiers come in two basic types: global and local. Global
396 character. Local identifiers (register names, types) begin with
398 for identifiers, for different purposes:</p>
400 <ol>
401 <li>Named values are represented as a string of characters with their prefix.
405 other characters in their names can be surrounded with quotes. Special
407 ASCII code for the character in hexadecimal. In this way, any character
408 can be used in a name value, even quotes themselves.</li>
410 <li>Unnamed values are represented as an unsigned numeric value with their
415 </ol>
417 <p>LLVM requires that values start with a prefix for two reasons: Compilers
418 don't need to worry about name clashes with reserved words, and the set of
419 reserved words may be expanded in the future without penalty. Additionally,
420 unnamed identifiers allow a compiler to quickly come up with a temporary
421 variable without having to avoid symbol table conflicts.</p>
423 <p>Reserved words in LLVM are very similar to reserved words in other
424 languages. There are keywords for different opcodes
430 reserved words cannot conflict with variable names, because none of them
433 <p>Here is an example of LLVM code to multiply the integer variable
440 </pre>
446 </pre>
454 </pre>
457 lexical features of LLVM:</p>
459 <ol>
461 line.</li>
463 <li>Unnamed temporaries are created when the result of a computation is not
464 assigned to a named value.</li>
467 </ol>
469 <p>It also shows a convention that we follow in this document. When
470 demonstrating instructions, we will follow an instruction with a comment that
471 defines the type and name of value produced. Comments are shown in italic
472 text.</p>
474 </div>
476 <!-- *********************************************************************** -->
478 <!-- *********************************************************************** -->
480 <!-- ======================================================================= -->
482 </div>
487 of the input programs. Each module consists of functions, global variables,
488 and symbol table entries. Modules may be combined together with the LLVM
489 linker, which merges function (and global variable) definitions, resolves
490 forward declarations, and merges symbol table entries. Here is an example of
495 <a href="#identifiers">@.LC0</a> = <a href="#linkage_internal">internal</a> <a href="#globalvars">constant</a> <a href="#t_array">[13 x i8]</a> c"hello world\0A\00" <i>; [13 x i8]*</i>
503 %cast210 = <a href="#i_getelementptr">getelementptr</a> [13 x i8]* @.LC0, i64 0, i64 0 <i>; i8*</i>
511 !foo = !{!1, null}
512 </pre>
520 <p>In general, a module is made up of a list of global values, where both
521 functions and global variables are global values. Global values are
522 represented by a pointer to a memory location (in this case, a pointer to an
523 array of char, and a pointer to a function), and have one of the
526 </div>
528 <!-- ======================================================================= -->
531 </div>
535 <p>All Global Variables and Functions have one of the following types of
536 linkage:</p>
538 <dl>
541 by objects in the current module. In particular, linking code into a
542 module with an private global value may cause the private to be renamed as
543 necessary to avoid collisions. Because the symbol is private to the
544 module, all references can be updated. This doesn't show up in any symbol
545 table in the object file.</dd>
549 assembler and evaluated by the linker. Unlike normal strong symbols, they
550 are removed by the linker from the final linked image (executable or
551 dynamic library).</dd>
556 linker. The symbols are removed by the linker from the final linked image
557 (executable or dynamic library).</dd>
560 <dd>Similar to private, but the value shows as a local symbol
566 into the object file corresponding to the LLVM module. They exist to
567 allow inlining and other optimizations to take place given knowledge of
568 the definition of the global, which is known to be somewhere outside the
571 This linkage type is only allowed on definitions, not declarations.</dd>
575 the same name when linkage occurs. This can be used to implement
576 some forms of inline functions, templates, or other code which must be
577 generated in each translation unit that uses it, but where the body may
578 be overridden with a more definitive definition later. Unreferenced
581 inline the body of this function into callers because it doesn't know if
582 this definition of the function is the definitive definition within the
583 program or whether it will be overridden by a stronger definition.
584 To enable inlining and other optimizations, use "<tt>linkonce_odr</tt>"
585 linkage.</dd>
595 they are used for tentative definitions in C, such as "<tt>int X;</tt>" at
596 global scope.
597 Symbols with "<tt>common</tt>" linkage are merged in the same way as
600 must have a zero initializer, and may not be marked '<a
602 have common linkage.</dd>
607 pointer to array type. When two global variables with appending linkage
608 are linked together, the two global arrays are appended together. This is
609 the LLVM, typesafe, equivalent of having the system linker append together
613 <dd>The semantics of this linkage follow the ELF object file model: the symbol
614 is weak until linked, if not linked, the symbol becomes null instead of
615 being an undefined reference.</dd>
619 <dd>Some languages allow differing globals to be merged, such as two functions
621 that only equivalent globals are ever merged (the "one definition rule"
624 be merged with equivalent globals. These linkage types are otherwise the
628 <dd>If none of the above identifiers are used, the global is externally
629 visible, meaning that it participates in linkage and can be used to
630 resolve external symbol references.</dd>
631 </dl>
633 <p>The next two types of linkage are targeted for Microsoft Windows platform
634 only. They are designed to support importing (exporting) symbols from (to)
635 DLLs (Dynamic Link Libraries).</p>
637 <dl>
640 or variable via a global pointer to a pointer that is set up by the DLL
641 exporting the symbol. On Microsoft Windows targets, the pointer name is
643 name.</dd>
647 pointer to a pointer in a DLL, so that it can be referenced with the
650 variable name.</dd>
651 </dl>
654 another module defined a "<tt>.LC0</tt>" variable and was linked with this
655 one, one of the two would be renamed, preventing a collision. Since
657 declarations), they are accessible outside of the current module.</p>
666 </div>
668 <!-- ======================================================================= -->
671 </div>
677 convention specified for the call. The calling convention of any pair of
678 dynamic caller/callee must match, or the behavior of the program is
679 undefined. The following calling conventions are supported by LLVM, and more
680 may be added in the future:</p>
682 <dl>
684 <dd>This calling convention (the default if no other calling convention is
685 specified) matches the target C calling conventions. This calling
686 convention supports varargs function calls and tolerates some mismatch in
687 the declared prototype and implemented declaration of the function (as
688 does normal C).</dd>
691 <dd>This calling convention attempts to make calls as fast as possible
692 (e.g. by passing things in registers). This calling convention allows the
693 target to use whatever tricks it wants to produce fast code for the
694 target, without having to conform to an externally specified ABI
695 (Application Binary Interface).
697 when this or the GHC convention is used.</a> This calling convention
698 does not support varargs and requires the prototype of all callees to
699 exactly match the prototype of the function definition.</dd>
702 <dd>This calling convention attempts to make code in the caller as efficient
703 as possible under the assumption that the call is not commonly executed.
704 As such, these calls often preserve all registers so that the call does
705 not break any live ranges in the caller side. This calling convention
706 does not support varargs and requires the prototype of all callees to
707 exactly match the prototype of the function definition.</dd>
710 <dd>This calling convention has been implemented specifically for use by the
712 It passes everything in registers, going to extremes to achieve this by
713 disabling callee save registers. This calling convention should not be
714 used lightly but only for specific situations such as an alternative to
716 implementing functional programming languages.At the moment only X86
717 supports this convention and it has the following limitations:
718 <ul>
720 floating point types are supported.</li>
723 </ul>
724 This calling convention supports
726 requires both the caller and callee are using it.
727 </dd>
730 <dd>Any calling convention may be specified by number, allowing
731 target-specific calling conventions to be used. Target specific calling
733 </dl>
735 <p>More calling conventions can be added/defined on an as-needed basis, to
736 support Pascal conventions or any other well-known target-independent
737 convention.</p>
739 </div>
741 <!-- ======================================================================= -->
744 </div>
748 <p>All Global Variables and Functions have one of the following visibility
749 styles:</p>
751 <dl>
753 <dd>On targets that use the ELF object file format, default visibility means
754 that the declaration is visible to other modules and, in shared libraries,
755 means that the declared entity may be overridden. On Darwin, default
756 visibility means that the declaration is visible to other modules. Default
760 <dd>Two declarations of an object with hidden visibility refer to the same
761 object if they are in the same shared object. Usually, hidden visibility
762 indicates that the symbol will not be placed into the dynamic symbol
763 table, so no other module (executable or shared library) can reference it
764 directly.</dd>
767 <dd>On ELF, protected visibility indicates that the symbol will be placed in
768 the dynamic symbol table, but that references within the defining module
769 will bind to the local symbol. That is, the symbol cannot be overridden by
770 another module.</dd>
771 </dl>
773 </div>
775 <!-- ======================================================================= -->
778 </div>
782 <p>LLVM IR allows you to specify name aliases for certain types. This can make
783 it easier to read the IR and make the IR more condensed (particularly when
784 recursive types are involved). An example of a name specification is:</p>
787 %mytype = type { %mytype*, i32 }
788 </pre>
794 <p>Note that type names are aliases for the structural type that they indicate,
795 and that you can therefore specify multiple names for the same type. This
796 often leads to confusing behavior when dumping out a .ll file. Since LLVM IR
797 uses structural typing, the name is not part of the type. When printing out
799 particular shape. This means that if you have code where two different
800 source types end up having the same LLVM type, that the dumper will sometimes
801 print the "wrong" or unexpected type. This is an important design point and
802 isn't going to change.</p>
804 </div>
806 <!-- ======================================================================= -->
809 </div>
813 <p>Global variables define regions of memory allocated at compilation time
814 instead of run-time. Global variables may optionally be initialized, may
815 have an explicit section to be placed in, and may have an optional explicit
816 alignment specified. A variable may be defined as "thread_local", which
817 means that it will not be shared by threads (each thread will have a
818 separated copy of the variable). A variable may be defined as a global
819 "constant," which indicates that the contents of the variable
821 global data to be placed in the read-only section of an executable, etc).
822 Note that variables that need runtime initialization cannot be marked
826 constant, even if the final definition of the global is not. This capability
827 can be used to enable slightly better optimization of the program, but
828 requires the language definition to guarantee that optimizations based on the
829 'constantness' are valid for the translation units that do not include the
830 definition.</p>
832 <p>As SSA values, global variables define pointer values that are in scope
833 (i.e. they dominate) all basic blocks in the program. Global variables
834 always define a pointer to their "content" type because they describe a
835 region of memory, and all memory objects in LLVM are accessed through
836 pointers.</p>
838 <p>A global variable may be declared to reside in a target-specific numbered
839 address space. For targets that support them, address spaces may affect how
840 optimizations are performed and/or what target instructions are used to
841 access the variable. The default address space is zero. The address space
842 qualifier must precede any other attributes.</p>
844 <p>LLVM allows an explicit section to be specified for globals. If the target
845 supports it, it will emit globals to the section specified.</p>
847 <p>An explicit alignment may be specified for a global, which must be a power
848 of 2. If not present, or if the alignment is set to zero, the alignment of
849 the global is set by the target to whatever it feels convenient. If an
850 explicit alignment is specified, the global is forced to have exactly that
851 alignment. Targets and optimizers are not allowed to over-align the global
852 if the global has an assigned section. In this case, the extra alignment
853 could be observable: for example, code could assume that the globals are
854 densely packed in their section and try to iterate over them as an array,
855 alignment padding would break this iteration.</p>
857 <p>For example, the following defines a global in a numbered address space with
858 an initializer, section, and alignment:</p>
862 </pre>
864 </div>
867 <!-- ======================================================================= -->
870 </div>
879 name, a (possibly empty) argument list (each with optional
883 curly brace, a list of basic blocks, and a closing curly brace.</p>
890 name, a possibly empty list of arguments, an optional alignment, and an
893 <p>A function definition contains a list of basic blocks, forming the CFG
894 (Control Flow Graph) for the function. Each basic block may optionally start
895 with a label (giving the basic block a symbol table entry), contains a list
897 instruction (such as a branch or function return).</p>
899 <p>The first basic block in a function is special in two ways: it is immediately
900 executed on entrance to the function, and it is not allowed to have
901 predecessor basic blocks (i.e. there can not be any branches to the entry
902 block of a function). Because the block can have no predecessors, it also
905 <p>LLVM allows an explicit section to be specified for functions. If the target
906 supports it, it will emit functions to the section specified.</p>
908 <p>An explicit alignment may be specified for a function. If not present, or if
909 the alignment is set to zero, the alignment of the function is set by the
910 target to whatever it feels convenient. If an explicit alignment is
911 specified, the function is forced to have at least that much alignment. All
921 </pre>
923 </div>
925 <!-- ======================================================================= -->
928 </div>
933 function, global variable, another alias or bitcast of global value). Aliases
940 </pre>
942 </div>
944 <!-- ======================================================================= -->
947 </div>
952 nodes</a> (but not metadata strings) are the only valid operands for
953 a named metadata.</p>
957 ; Some unnamed metadata nodes, which are referenced by the named metadata.
961 ; A named metadata.
963 </pre>
965 </div>
967 <!-- ======================================================================= -->
972 <p>The return type and each parameter of a function type may have a set of
974 used to communicate additional information about the result or parameters of
975 a function. Parameter attributes are considered to be part of the function,
976 not of the function type, so functions with different parameter attributes
977 can have the same function type.</p>
979 <p>Parameter attributes are simple keywords that follow the type specified. If
980 multiple parameter attributes are needed, they are space separated. For
981 example:</p>
984 declare i32 @printf(i8* noalias nocapture, ...)
985 declare i32 @atoi(i8 zeroext)
986 declare signext i8 @returns_signed_char()
987 </pre>
994 <dl>
996 <dd>This indicates to the code generator that the parameter or return value
997 should be zero-extended to a 32-bit value by the caller (for a parameter)
998 or the callee (for a return value).</dd>
1001 <dd>This indicates to the code generator that the parameter or return value
1002 should be sign-extended to a 32-bit value by the caller (for a parameter)
1003 or the callee (for a return value).</dd>
1006 <dd>This indicates that this parameter or return value should be treated in a
1007 special target-dependent fashion during while emitting code for a function
1008 call or return (usually, by putting it in a register as opposed to memory,
1009 though some targets use it to distinguish between two different kinds of
1010 registers). Use of this attribute is target-specific.</dd>
1013 <dd>This indicates that the pointer parameter should really be passed by value
1014 to the function. The attribute implies that a hidden copy of the pointee
1015 is made between the caller and the callee, so the callee is unable to
1016 modify the value in the callee. This attribute is only valid on LLVM
1017 pointer arguments. It is generally used to pass structs and arrays by
1018 value, but is also valid on pointers to scalars. The copy is considered
1019 to belong to the caller not the callee (for example,
1022 values. The byval attribute also supports specifying an alignment with
1023 the align attribute. This has a target-specific effect on the code
1024 generator that usually indicates a desired alignment for the synthesized
1025 stack slot.</dd>
1028 <dd>This indicates that the pointer parameter specifies the address of a
1029 structure that is the return value of the function in the source program.
1030 This pointer must be guaranteed by the caller to be valid: loads and
1031 stores to the structure may be assumed by the callee to not to trap. This
1032 may only be applied to the first parameter. This is not a valid attribute
1033 for return values. </dd>
1036 <dd>This indicates that pointer values
1039 ignoring certain "irrelevant" dependencies.
1040 For a call to the parent function, dependencies between memory
1041 references from before or after the call and from those during the call
1043 return value used in that call.
1044 The caller shares the responsibility with the callee for ensuring that
1045 these requirements are met.
1046 For further details, please see the discussion of the NoAlias response in
1048 <br>
1051 arguments, though it is slightly weaker.
1052 <br>
1055 </dd>
1058 <dd>This indicates that the callee does not make any copies of the pointer
1059 that outlive the callee itself. This is not a valid attribute for return
1060 values.</dd>
1063 <dd>This indicates that the pointer parameter can be excised using the
1065 attribute for return values.</dd>
1066 </dl>
1068 </div>
1070 <!-- ======================================================================= -->
1073 </div>
1077 <p>Each function may specify a garbage collector name, which is simply a
1078 string:</p>
1081 define void @f() gc "name" { ... }
1082 </pre>
1085 collector which will cause the compiler to alter its output in order to
1086 support the named garbage collection algorithm.</p>
1088 </div>
1090 <!-- ======================================================================= -->
1093 </div>
1097 <p>Function attributes are set to communicate additional information about a
1098 function. Function attributes are considered to be part of the function, not
1099 of the function type, so functions with different parameter attributes can
1100 have the same function type.</p>
1102 <p>Function attributes are simple keywords that follow the type specified. If
1103 multiple attributes are needed, they are space separated. For example:</p>
1106 define void @f() noinline { ... }
1107 define void @f() alwaysinline { ... }
1108 define void @f() alwaysinline optsize { ... }
1109 define void @f() optsize { ... }
1110 </pre>
1112 <dl>
1114 <dd>This attribute indicates that, when emitting the prologue and epilogue,
1115 the backend should forcibly align the stack pointer. Specify the
1116 desired alignment, which must be a power of two, in parentheses.
1119 <dd>This attribute indicates that the inliner should attempt to inline this
1120 function into callers whenever possible, ignoring any active inlining size
1121 threshold for this caller.</dd>
1124 <dd>This attribute indicates that the source code contained a hint that inlining
1125 this function is desirable (such as the "inline" keyword in C/C++). It
1126 is just a hint; it imposes no requirements on the inliner.</dd>
1129 <dd>This attribute disables prologue / epilogue emission for the function.
1130 This can have very system-specific consequences.</dd>
1136 <dd>This attribute indicates that the inliner should never inline this
1137 function in any situation. This attribute may not be used together with
1141 <dd>This attribute indicates that the code generator should not use a red
1142 zone, even if the target-specific ABI normally permits it.</dd>
1145 <dd>This function attribute indicates that the function never returns
1146 normally. This produces undefined behavior at runtime if the function
1147 ever does dynamically return.</dd>
1150 <dd>This function attribute indicates that the function never returns with an
1151 unwind or exceptional control flow. If the function does unwind, its
1152 runtime behavior is undefined.</dd>
1155 <dd>This attribute suggests that optimization passes and code generator passes
1156 make choices that keep the code size of this function low, and otherwise
1157 do optimizations specifically to reduce code size.</dd>
1160 <dd>This attribute indicates that the function computes its result (or decides
1161 to unwind an exception) based strictly on its arguments, without
1162 dereferencing any pointer arguments or otherwise accessing any mutable
1163 state (e.g. memory, control registers, etc) visible to caller functions.
1164 It does not write through any pointer arguments
1166 changes any state visible to callers. This means that it cannot unwind
1171 <dd>This attribute indicates that the function does not write through any
1173 arguments) or otherwise modify any state (e.g. memory, control registers,
1174 etc) visible to caller functions. It may dereference pointer arguments
1175 and read state that may be set in the caller. A readonly function always
1176 returns the same value (or unwinds an exception identically) when called
1177 with the same set of arguments and global state. It cannot unwind an
1182 <dd>This attribute indicates that the function should emit a stack smashing
1184 the stack before the local variables that's checked upon return from the
1185 function to see if it has been overwritten. A heuristic is used to
1186 determine if a function needs stack protectors or not.<br>
1187 <br>
1194 stack smashing protector. This overrides
1196 <br>
1201 </dl>
1203 </div>
1205 <!-- ======================================================================= -->
1208 </div>
1213 the GCC "file scope inline asm" blocks. These blocks are internally
1214 concatenated by LLVM and treated as a single unit, but may be separated in
1218 module asm "inline asm code goes here"
1219 module asm "more can go here"
1220 </pre>
1222 <p>The strings can contain any character by escaping non-printable characters.
1224 for the number.</p>
1226 <p>The inline asm code is simply printed to the machine code .s file when
1227 assembly code is generated.</p>
1229 </div>
1231 <!-- ======================================================================= -->
1234 </div>
1238 <p>A module may specify a target specific data layout string that specifies how
1239 data is to be laid out in memory. The syntax for the data layout is
1240 simply:</p>
1243 target datalayout = "<i>layout specification</i>"
1244 </pre>
1247 separated by the minus sign character ('-'). Each specification starts with
1248 a letter and may include other information after the letter to define some
1249 aspect of the data layout. The specifications accepted are as follows:</p>
1251 <dl>
1253 <dd>Specifies that the target lays out data in big-endian form. That is, the
1254 bits with the most significance have the lowest address location.</dd>
1257 <dd>Specifies that the target lays out data in little-endian form. That is,
1258 the bits with the least significance have the lowest address
1259 location.</dd>
1268 <dd>This specifies the alignment for an integer type of a given bit
1272 <dd>This specifies the alignment for a vector type of a given bit
1276 <dd>This specifies the alignment for a floating point type of a given bit
1280 targets.
1283 <dd>This specifies the alignment for an aggregate type of a given bit
1287 <dd>This specifies the alignment for a stack object of a given bit
1291 <dd>This specifies a set of native integer widths for the target CPU
1294 this set are considered to support most general arithmetic
1295 operations efficiently.</dd>
1296 </dl>
1298 <p>When constructing the data layout for a given target, LLVM starts with a
1299 default set of specifications which are then (possibly) overridden by the
1301 are given in this list:</p>
1303 <ul>
1318 </ul>
1320 <p>When LLVM is determining the alignment for a given type, it uses the
1321 following rules:</p>
1323 <ol>
1324 <li>If the type sought is an exact match for one of the specifications, that
1325 specification is used.</li>
1327 <li>If no match is found, and the type sought is an integer type, then the
1328 smallest integer type that is larger than the bitwidth of the sought type
1329 is used. If none of the specifications are larger than the bitwidth then
1330 the the largest integer type is used. For example, given the default
1331 specifications above, the i7 type will use the alignment of i8 (next
1332 largest) while both i65 and i256 will use the alignment of i64 (largest
1333 specified).</li>
1335 <li>If no match is found, and the type sought is a vector type, then the
1336 largest vector type that is smaller than the sought vector type will be
1339 </ol>
1341 </div>
1343 <!-- ======================================================================= -->
1346 </div>
1350 <p>Any memory access must be done through a pointer value associated
1351 with an address range of the memory access, otherwise the behavior
1352 is undefined. Pointer values are associated with address ranges
1353 according to the following rules:</p>
1355 <ul>
1356 <li>A pointer value is associated with the addresses associated with
1358 <li>An address of a global variable is associated with the address
1359 range of the variable's storage.</li>
1360 <li>The result value of an allocation instruction is associated with
1361 the address range of the allocated storage.</li>
1362 <li>A null pointer in the default address-space is associated with
1363 no address.</li>
1364 <li>An integer constant other than zero or a pointer value returned
1365 from a function not defined within LLVM may be associated with address
1366 ranges allocated through mechanisms other than those provided by
1367 LLVM. Such ranges shall not overlap with any ranges of addresses
1368 allocated by mechanisms provided by LLVM.</li>
1369 </ul>
1372 to the following rules:</p>
1374 <ul>
1375 <li>A pointer value formed from a
1378 <li>The result value of a
1381 <li>A pointer value formed by an
1383 pointer values that contribute (directly or indirectly) to the
1384 computation of the pointer's value.</li>
1386 </ul>
1390 slightly weaker.</p>
1392 <p>LLVM IR does not associate types with memory. The result type of a
1394 alignment of the memory from which to load, as well as the
1395 interpretation of the value. The first operand type of a
1397 and alignment of the store.</p>
1399 <p>Consequently, type-based alias analysis, aka TBAA, aka
1402 additional information which specialized optimization passes may use
1403 to implement type-based alias analysis.</p>
1405 </div>
1407 <!-- ======================================================================= -->
1410 </div>
1417 The optimizers must not change the number of volatile operations or change their
1418 order of execution relative to other volatile operations. The optimizers
1420 operations. This is not Java's "volatile" and has no cross-thread
1421 synchronization behavior.</p>
1423 </div>
1425 <!-- *********************************************************************** -->
1427 <!-- *********************************************************************** -->
1431 <p>The LLVM type system is one of the most important features of the
1432 intermediate representation. Being typed enables a number of optimizations
1433 to be performed on the intermediate representation directly, without having
1434 to do extra analyses on the side before the transformation. A strong type
1435 system makes it easier to read the generated code and enables novel analyses
1436 and transformations that are not feasible to perform on normal three address
1437 code representations.</p>
1439 </div>
1441 <!-- ======================================================================= -->
1443 Classifications</a> </div>
1450 <tbody>
1452 <tr>
1455 </tr>
1456 <tr>
1459 </tr>
1460 <tr>
1471 </td>
1472 </tr>
1473 <tr>
1479 </tr>
1480 <tr>
1490 </td>
1491 </tr>
1492 </tbody>
1493 </table>
1496 important. Values of these types are the only ones which can be produced by
1497 instructions.</p>
1499 </div>
1501 <!-- ======================================================================= -->
1506 <p>The primitive types are the fundamental building blocks of the LLVM
1507 system.</p>
1509 </div>
1511 <!-- _______________________________________________________________________ -->
1517 <p>The integer type is a very simple type that simply specifies an arbitrary
1518 bit width for the integer type desired. Any bit width from 1 bit to
1522 <pre>
1523 iN
1524 </pre>
1527 value.</p>
1534 </tr>
1538 </tr>
1542 </tr>
1543 </table>
1545 </div>
1547 <!-- _______________________________________________________________________ -->
1552 <table>
1553 <tbody>
1560 </tbody>
1561 </table>
1563 </div>
1565 <!-- _______________________________________________________________________ -->
1574 <pre>
1575 void
1576 </pre>
1578 </div>
1580 <!-- _______________________________________________________________________ -->
1589 <pre>
1590 label
1591 </pre>
1593 </div>
1595 <!-- _______________________________________________________________________ -->
1601 <p>The metadata type represents embedded metadata. No derived types may be
1603 arguments.
1606 <pre>
1607 metadata
1608 </pre>
1610 </div>
1613 <!-- ======================================================================= -->
1618 <p>The real power in LLVM comes from the derived types in the system. This is
1619 what allows a programmer to represent arrays, functions, pointers, and other
1620 useful types. Each of these types contain one or more element types which
1621 may be a primitive type, or another derived type. For example, it is
1622 possible to have a two dimensional array, using an array as the element type
1623 of another array.</p>
1626 </div>
1628 <!-- _______________________________________________________________________ -->
1633 <p>Aggregate Types are a subset of derived types that can contain multiple
1638 </div>
1640 <!-- _______________________________________________________________________ -->
1646 <p>The array type is a very simple derived type that arranges elements
1647 sequentially in memory. The array type requires a size (number of elements)
1648 and an underlying data type.</p>
1651 <pre>
1653 </pre>
1656 be any type with a size.</p>
1663 </tr>
1667 </tr>
1671 </tr>
1672 </table>
1678 </tr>
1682 </tr>
1686 </tr>
1687 </table>
1689 <p>There is no restriction on indexing beyond the end of the array implied by
1690 a static type (though there are restrictions on indexing beyond the bounds
1691 of an allocated object in some cases). This means that single-dimension
1692 'variable sized array' addressing can be implemented in LLVM with a zero
1693 length array type. An implementation of 'pascal style arrays' in LLVM could
1696 </div>
1698 <!-- _______________________________________________________________________ -->
1704 <p>The function type can be thought of as a function signature. It consists of
1705 a return type and a list of formal parameter types. The return type of a
1706 function type is a scalar type, a void type, a struct type, or a union
1707 type. If the return type is a struct type then all struct elements must be
1708 of first class types, and the struct must have at least one element.</p>
1711 <pre>
1713 </pre>
1717 which indicates that the function takes a variable number of arguments.
1718 Variable argument functions can access their arguments with
1728 </td>
1731 </tt></td>
1735 </td>
1741 LLVM.
1742 </td>
1747 </td>
1748 </tr>
1749 </table>
1751 </div>
1753 <!-- _______________________________________________________________________ -->
1759 <p>The structure type is used to represent a collection of data members together
1760 in memory. The packing of the field types is defined to match the ABI of the
1761 underlying processor. The elements of a structure may be any type that has a
1762 size.</p>
1767 Structures in registers are accessed using the
1771 <pre>
1773 </pre>
1786 </tr>
1787 </table>
1789 </div>
1791 <!-- _______________________________________________________________________ -->
1793 </div>
1798 <p>The packed structure type is used to represent a collection of data members
1799 together in memory. There is no padding between fields. Further, the
1800 alignment of a packed structure is 1 byte. The elements of a packed
1801 structure may be any type that has a size.</p>
1808 <pre>
1810 </pre>
1824 </tr>
1825 </table>
1827 </div>
1829 <!-- _______________________________________________________________________ -->
1835 <p>A union type describes an object with size and alignment suitable for
1836 an object of any one of a given set of types (also known as an "untagged"
1837 union). It is similar in concept and usage to a
1839 have an offset of zero. The elements of a union may be any type that has a
1840 size. Unions must have at least one member - empty unions are not allowed.
1841 </p>
1843 <p>The size of the union as a whole will be the size of its largest member,
1844 and the alignment requirements of the union as a whole will be the largest
1845 alignment requirement of any member.</p>
1850 Since all members are at offset zero, the getelementptr instruction does
1851 not affect the address, only the type of the resulting pointer.</p>
1854 <pre>
1856 </pre>
1871 </tr>
1872 </table>
1874 </div>
1876 <!-- _______________________________________________________________________ -->
1882 <p>The pointer type is used to specify memory locations.
1883 Pointers are commonly used to reference objects in memory.</p>
1885 <p>Pointer types may have an optional address space attribute defining the
1886 numbered address space where the pointed-to object resides. The default
1887 address space is number zero. The semantics of non-zero address
1888 spaces are target-specific.</p>
1894 <pre>
1896 </pre>
1904 </tr>
1910 </tr>
1915 </tr>
1916 </table>
1918 </div>
1920 <!-- _______________________________________________________________________ -->
1926 <p>A vector type is a simple derived type that represents a vector of elements.
1927 Vector types are used when multiple primitive data are operated in parallel
1928 using a single instruction (SIMD). A vector type requires a size (number of
1929 elements) and an underlying primitive data type. Vector types are considered
1933 <pre>
1935 </pre>
1937 <p>The number of elements is a constant integer value; elementtype may be any
1938 integer or floating point type.</p>
1945 </tr>
1949 </tr>
1953 </tr>
1954 </table>
1956 </div>
1958 <!-- _______________________________________________________________________ -->
1963 <p>Opaque types are used to represent unknown types in the system. This
1964 corresponds (for example) to the C notion of a forward declared structure
1965 type. In LLVM, opaque types can eventually be resolved to any type (not just
1966 a structure type).</p>
1969 <pre>
1970 opaque
1971 </pre>
1978 </tr>
1979 </table>
1981 </div>
1983 <!-- ======================================================================= -->
1986 </div>
1992 requiring it to have a name. For instance, a structure declaration may
1993 contain a pointer to any of the types it is lexically a member of. Example
1994 of up references (with their equivalent as named type declarations)
1995 include:</p>
1997 <pre>
1998 { \2 * } %x = type { %x* }
1999 { \2 }* %y = type { %y }*
2000 \1* %z = type %z*
2001 </pre>
2003 <p>An up reference is needed by the asmprinter for printing out cyclic types
2004 when there is no declared name for a type in the cycle. Because the
2005 asmprinter does not want to print out an infinite type string, it needs a
2006 syntax to handle recursive types that have no names (all names are optional
2007 in llvm IR).</p>
2010 <pre>
2012 </pre>
2021 </tr>
2025 structure.</td>
2026 </tr>
2027 </table>
2029 </div>
2031 <!-- *********************************************************************** -->
2033 <!-- *********************************************************************** -->
2037 <p>LLVM has several different basic types of constants. This section describes
2038 them all and their syntax.</p>
2040 </div>
2042 <!-- ======================================================================= -->
2047 <dl>
2055 with integer types.</dd>
2059 exponential notation (e.g. 1.23421e+2), or a more precise hexadecimal
2060 notation (see below). The assembler requires the exact decimal value of a
2061 floating-point constant. For example, the assembler accepts 1.25 but
2068 </dl>
2070 <p>The one non-intuitive notation for constants is the hexadecimal form of
2071 floating point constants. For example, the form '<tt>double
2074 constants are required (and the only time that they are generated by the
2075 disassembler) is when a floating point constant must be emitted but it cannot
2076 be represented as a decimal floating point number in a reasonable number of
2077 digits. For example, NaN's, infinities, and other special values are
2078 represented in their IEEE hexadecimal format so that assembly and disassembly
2079 do not cause any bits to change in the constants.</p>
2081 <p>When using the hexadecimal form, constants of types float and double are
2082 represented using the 16-digit form shown above (which matches the IEEE754
2083 representation for double); float values must, however, be exactly
2084 representable as IEE754 single precision. Hexadecimal format is always used
2085 for long double, and there are three forms of long double. The 80-bit format
2087 The 128-bit format used by PowerPC (two adjacent doubles) is represented
2090 currently supported target uses this format. Long doubles will only work if
2091 they match the long double format on your target. All hexadecimal formats
2092 are big-endian (sign bit at the left).</p>
2094 </div>
2096 <!-- ======================================================================= -->
2100 </div>
2104 <p>Complex constants are a (potentially recursive) combination of simple
2105 constants and smaller complex constants.</p>
2107 <dl>
2109 <dd>Structure constants are represented with notation similar to structure
2110 type definitions (a comma separated list of elements, surrounded by braces
2114 the number and types of elements must match those specified by the
2115 type.</dd>
2118 <dd>Union constants are represented with notation similar to a structure with
2119 a single element - that is, a single typed element surrounded
2122 struct as long as the type of the struct element matches the type of
2123 one of the union members.</dd>
2126 <dd>Array constants are represented with notation similar to array type
2127 definitions (a comma separated list of elements, surrounded by square
2130 the number and types of elements must match those specified by the
2131 type.</dd>
2134 <dd>Vector constants are represented with notation similar to vector type
2135 definitions (a comma separated list of elements, surrounded by
2137 42, i32 11, i32 74, i32 100 ></tt>". Vector constants must
2139 elements must match those specified by the type.</dd>
2145 This is often used to avoid having to print large zero initializers
2146 (e.g. for large arrays) and is always exactly equivalent to using explicit
2147 zero initializers.</dd>
2150 <dd>A metadata node is a structure-like constant with
2153 be interpreted as part of the instruction stream, metadata is a place to
2154 attach additional information such as debug info.</dd>
2155 </dl>
2157 </div>
2159 <!-- ======================================================================= -->
2162 </div>
2168 (link-time) constants. These constants are explicitly referenced when
2171 legal LLVM file:</p>
2174 @X = global i32 17
2175 @Y = global i32 42
2176 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2177 </pre>
2179 </div>
2181 <!-- ======================================================================= -->
2186 indicates that the user of the value may receive an unspecified bit-pattern.
2187 Undefined values may be of any type (other than label or void) and be used
2188 anywhere a constant is permitted.</p>
2190 <p>Undefined values are useful because they indicate to the compiler that the
2191 program is well defined no matter what value is used. This gives the
2192 compiler more freedom to optimize. Here are some examples of (potentially
2193 surprising) transformations that are valid (in pseudo IR):</p>
2197 %A = add %X, undef
2198 %B = sub %X, undef
2199 %C = xor %X, undef
2200 Safe:
2201 %A = undef
2202 %B = undef
2203 %C = undef
2204 </pre>
2206 <p>This is safe because all of the output bits are affected by the undef bits.
2207 Any output bit can have a zero or one depending on the input bits.</p>
2210 %A = or %X, undef
2211 %B = and %X, undef
2212 Safe:
2213 %A = -1
2214 %B = 0
2215 Unsafe:
2216 %A = undef
2217 %B = undef
2218 </pre>
2220 <p>These logical operations have bits that are not always affected by the input.
2221 For example, if "%X" has a zero bit, then the output of the 'and' operation will
2222 always be a zero, no matter what the corresponding bit from the undef is. As
2223 such, it is unsafe to optimize or assume that the result of the and is undef.
2224 However, it is safe to assume that all bits of the undef could be 0, and
2225 optimize the and to 0. Likewise, it is safe to assume that all the bits of
2226 the undef operand to the or could be set, allowing the or to be folded to
2230 %A = select undef, %X, %Y
2231 %B = select undef, 42, %Y
2232 %C = select %X, %Y, undef
2233 Safe:
2234 %A = %X (or %Y)
2235 %B = 42 (or %Y)
2236 %C = %Y
2237 Unsafe:
2238 %A = undef
2239 %B = undef
2240 %C = undef
2241 </pre>
2243 <p>This set of examples show that undefined select (and conditional branch)
2244 conditions can go "either way" but they have to come from one of the two
2245 operands. In the %A example, if %X and %Y were both known to have a clear low
2246 bit, then %A would have to have a cleared low bit. However, in the %C example,
2247 the optimizer is allowed to assume that the undef operand could be the same as
2248 %Y, allowing the whole select to be eliminated.</p>
2252 %A = xor undef, undef
2254 %B = undef
2255 %C = xor %B, %B
2257 %D = undef
2258 %E = icmp lt %D, 4
2259 %F = icmp gte %D, 4
2261 Safe:
2262 %A = undef
2263 %B = undef
2264 %C = undef
2265 %D = undef
2266 %E = undef
2267 %F = undef
2268 </pre>
2270 <p>This example points out that two undef operands are not necessarily the same.
2271 This can be surprising to people (and also matches C semantics) where they
2272 assume that "X^X" is always zero, even if X is undef. This isn't true for a
2273 number of reasons, but the short answer is that an undef "variable" can
2274 arbitrarily change its value over its "live range". This is true because the
2276 logically read from arbitrary registers that happen to be around when needed,
2277 so the value is not necessarily consistent over time. In fact, %A and %C need
2278 to have the same semantics or the core LLVM "replace all uses with" concept
2279 would not hold.</p>
2282 %A = fdiv undef, %X
2283 %B = fdiv %X, undef
2284 Safe:
2285 %A = undef
2286 b: unreachable
2287 </pre>
2291 allowed to have an arbitrary bit-pattern. This means that the %A operation
2292 can be constant folded to undef because the undef could be an SNaN, and fdiv is
2293 not (currently) defined on SNaN's. However, in the second example, we can make
2294 a more aggressive assumption: because the undef is allowed to be an arbitrary
2295 value, we are allowed to assume that it could be zero. Since a divide by zero
2297 does not execute at all. This allows us to delete the divide and all code after
2298 it: since the undefined operation "can't happen", the optimizer can assume that
2299 it occurs in dead code.
2300 </p>
2303 a: store undef -> %X
2304 b: store %X -> undef
2305 Safe:
2307 b: unreachable
2308 </pre>
2311 can be assumed to not have any effect: we can assume that the value is
2312 overwritten with bits that happen to match what was already there. However, a
2313 store "to" an undefined location could clobber arbitrary memory, therefore, it
2314 has undefined behavior.</p>
2316 </div>
2318 <!-- ======================================================================= -->
2323 instead of representing an unspecified bit pattern, they represent the
2324 fact that an instruction or constant expression which cannot evoke side
2325 effects has nevertheless detected a condition which results in undefined
2326 behavior.</p>
2328 <p>There is currently no way of representing a trap value in the IR; they
2329 only exist when produced by operations such as
2334 <ul>
2336 their operands.</li>
2339 to their dynamic predecessor basic block.</li>
2341 <li>Function arguments depend on the corresponding actual argument values in
2342 the dynamic callers of their functions.</li>
2346 control back to them.</li>
2350 or exception-throwing call instructions that dynamically transfer control
2351 back to them.</li>
2353 <li>Non-volatile loads and stores depend on the most recent stores to all of the
2354 referenced memory addresses, following the order in the IR
2355 (including loads and stores implied by intrinsics such as
2358 <!-- TODO: In the case of multiple threads, this only applies if the store
2359 "happens-before" the load or store. -->
2361 <!-- TODO: floating-point exception state -->
2363 <li>An instruction with externally visible side effects depends on the most
2364 recent preceding instruction with externally visible side effects, following
2365 the order in the IR. (This includes
2370 if the terminator instruction has multiple successors and the instruction
2371 is always executed when control transfers to one of the successors, and
2372 may not be executed when control is transfered to another.</li>
2376 </ul>
2378 <p>Whenever a trap value is generated, all values which depend on it evaluate
2379 to trap. If they have side effects, the evoke their side effects as if each
2380 operand with a trap value were undef. If they have externally-visible side
2381 effects, the behavior is undefined.</p>
2386 entry:
2389 %trap_yet_again = getelementptr i32* @h, i32 %still_trap
2390 store i32 0, i32* %trap_yet_again ; undefined behavior
2392 store i32 %trap, i32* @g ; Trap value conceptually stored to memory.
2393 %trap2 = load i32* @g ; Returns a trap value, not just undef.
2395 volatile store i32 %trap, i32* @g ; External observation; undefined behavior.
2397 %narrowaddr = bitcast i32* @g to i16*
2398 %wideaddr = bitcast i32* @g to i64*
2399 %trap3 = load 16* %narrowaddr ; Returns a trap value.
2400 %trap4 = load i64* %widaddr ; Returns a trap value.
2402 %cmp = icmp i32 slt %trap, 0 ; Returns a trap value.
2403 %br i1 %cmp, %true, %end ; Branch to either destination.
2405 true:
2406 volatile store i32 0, i32* @g ; This is control-dependent on %cmp, so
2407 ; it has undefined behavior.
2408 br label %end
2410 end:
2412 ; Both edges into this PHI are
2413 ; control-dependent on %cmp, so this
2414 ; always results in a trap value.
2416 volatile store i32 0, i32* @g ; %end is control-equivalent to %entry
2417 ; so this is defined (ignoring earlier
2418 ; undefined behavior in this example).
2419 </pre>
2421 </div>
2423 <!-- ======================================================================= -->
2425 Blocks</a></div>
2431 basic block in the specified function, and always has an i8* type. Taking
2432 the address of the entry block is illegal.</p>
2434 <p>This value only has defined behavior when used as an operand to the
2436 against null. Pointer equality tests between labels addresses is undefined
2437 behavior - though, again, comparison against null is ok, and no label is
2438 equal to the null pointer. This may also be passed around as an opaque
2439 pointer sized value as long as the bits are not inspected. This allows
2443 <p>Finally, some targets may provide defined semantics when
2444 using the value as the operand to an inline assembly, but that is target
2445 specific.
2446 </p>
2448 </div>
2451 <!-- ======================================================================= -->
2453 </div>
2457 <p>Constant expressions are used to allow expressions involving other constants
2458 to be used as constants. Constant expressions may be of
2460 operation that does not have side effects (e.g. load and call are not
2461 supported). The following is the syntax for constant expressions:</p>
2463 <dl>
2465 <dd>Truncate a constant to another type. The bit size of CST must be larger
2466 than the bit size of TYPE. Both types must be integers.</dd>
2469 <dd>Zero extend a constant to another type. The bit size of CST must be
2470 smaller than the bit size of TYPE. Both types must be integers.</dd>
2473 <dd>Sign extend a constant to another type. The bit size of CST must be
2474 smaller than the bit size of TYPE. Both types must be integers.</dd>
2477 <dd>Truncate a floating point constant to another floating point type. The
2478 size of CST must be larger than the size of TYPE. Both types must be
2479 floating point.</dd>
2482 <dd>Floating point extend a constant to another type. The size of CST must be
2483 smaller or equal to the size of TYPE. Both types must be floating
2484 point.</dd>
2487 <dd>Convert a floating point constant to the corresponding unsigned integer
2488 constant. TYPE must be a scalar or vector integer type. CST must be of
2489 scalar or vector floating point type. Both CST and TYPE must be scalars,
2490 or vectors of the same number of elements. If the value won't fit in the
2491 integer type, the results are undefined.</dd>
2494 <dd>Convert a floating point constant to the corresponding signed integer
2495 constant. TYPE must be a scalar or vector integer type. CST must be of
2496 scalar or vector floating point type. Both CST and TYPE must be scalars,
2497 or vectors of the same number of elements. If the value won't fit in the
2498 integer type, the results are undefined.</dd>
2501 <dd>Convert an unsigned integer constant to the corresponding floating point
2502 constant. TYPE must be a scalar or vector floating point type. CST must be
2503 of scalar or vector integer type. Both CST and TYPE must be scalars, or
2504 vectors of the same number of elements. If the value won't fit in the
2505 floating point type, the results are undefined.</dd>
2508 <dd>Convert a signed integer constant to the corresponding floating point
2509 constant. TYPE must be a scalar or vector floating point type. CST must be
2510 of scalar or vector integer type. Both CST and TYPE must be scalars, or
2511 vectors of the same number of elements. If the value won't fit in the
2512 floating point type, the results are undefined.</dd>
2515 <dd>Convert a pointer typed constant to the corresponding integer constant
2521 <dd>Convert a integer constant to a pointer constant. TYPE must be a pointer
2522 type. CST must be of integer type. The CST value is zero extended,
2523 truncated, or unchanged to make it fit in a pointer size. This one is
2527 <dd>Convert a constant, CST, to another TYPE. The constraints of the operands
2535 instruction, the index list may have zero or more indexes, which are
2549 constants.</dd>
2553 constants.</dd>
2557 constants.</dd>
2561 constants. The index list is interpreted in a similar manner as indices in
2563 index value must be specified.</dd>
2567 constants. The index list is interpreted in a similar manner as indices in
2569 index value must be specified.</dd>
2572 <dd>Perform the specified operation of the LHS and RHS constants. OPCODE may
2575 on operands are the same as those for the corresponding instruction
2576 (e.g. no bitwise operations on floating point values are allowed).</dd>
2577 </dl>
2579 </div>
2581 <!-- *********************************************************************** -->
2583 <!-- *********************************************************************** -->
2585 <!-- ======================================================================= -->
2588 </div>
2592 <p>LLVM supports inline assembler expressions (as opposed
2594 a special value. This value represents the inline assembler as a string
2595 (containing the instructions to emit), a list of operand constraints (stored
2596 as a string), a flag that indicates whether or not the inline asm
2597 expression has side effects, and a flag indicating whether the function
2598 containing the asm needs to align its stack conservatively. An example
2599 inline assembler expression is:</p>
2603 </pre>
2607 have:</p>
2611 </pre>
2613 <p>Inline asms with side effects not visible in the constraint list must be
2614 marked as having side effects. This is done through the use of the
2619 </pre>
2621 <p>In some cases inline asms will contain code that will not work unless the
2622 stack is aligned in some way, such as calls or SSE instructions on x86,
2623 yet will not contain code that does that alignment within the asm.
2624 The compiler should make conservative assumptions about what the asm might
2625 contain and should generate its usual stack alignment code in the prologue
2630 </pre>
2633 first.</p>
2635 <p>TODO: The format of the asm and constraints string still need to be
2636 documented here. Constraints on what can be done (e.g. duplication, moving,
2637 etc need to be documented). This is probably best done by reference to
2638 another document that covers inline asm from a holistic perspective.</p>
2639 </div>
2643 </div>
2648 attached to it that contains a constant integer. If present, the code
2649 generator will use the integer as the location cookie value when report
2650 errors through the LLVMContext error reporting mechanisms. This allows a
2651 front-end to correlate backend errors that occur with inline asm back to the
2652 source code that produced it. For example:</p>
2656 ...
2658 </pre>
2660 <p>It is up to the front-end to make sense of the magic numbers it places in the
2661 IR.</p>
2663 </div>
2665 <!-- ======================================================================= -->
2667 Strings</a>
2668 </div>
2672 <p>LLVM IR allows metadata to be attached to instructions in the program that
2673 can convey extra information about the code to the optimizers and code
2674 generator. One example application of metadata is source-level debug
2675 information. There are two metadata primitives: strings and nodes. All
2679 <p>A metadata string is a string surrounded by double quotes. It can contain
2683 <p>Metadata nodes are represented with notation similar to structure constants
2684 (a comma separated list of elements, surrounded by braces and preceded by an
2689 metadata nodes, which can be looked up in the module symbol table. For
2690 example: "<tt>!foo = metadata !{!4, !3}</tt>".
2693 function is using two metadata arguments.</p>
2697 </pre>
2704 </pre>
2705 </div>
2708 <!-- *********************************************************************** -->
2711 </div>
2712 <!-- *********************************************************************** -->
2715 code generation or other IR semantics. These are documented here. All globals
2716 of this sort should have a section specified as "<tt>llvm.metadata</tt>". This
2717 section and all globals that start with "<tt>llvm.</tt>" are reserved for use
2718 by LLVM.</p>
2720 <!-- ======================================================================= -->
2723 </div>
2729 pointers to global variables and functions which may optionally have a pointer
2730 cast formed of bitcast or getelementptr. For example, a legal use of it is:</p>
2732 <pre>
2733 @X = global i8 4
2734 @Y = global i32 123
2736 @llvm.used = appending global [2 x i8*] [
2737 i8* @X,
2738 i8* bitcast (i32* @Y to i8*)
2739 ], section "llvm.metadata"
2740 </pre>
2743 compiler, assembler, and linker are required to treat the symbol as if there is
2744 a reference to the global that it cannot see. For example, if a variable has
2746 list, it cannot be deleted. This is commonly used to represent references from
2747 inline asms and other things the compiler cannot "see", and corresponds to
2750 <p>On some targets, the code generator must emit a directive to the assembler or
2751 object file to prevent the assembler and linker from molesting the symbol.</p>
2753 </div>
2755 <!-- ======================================================================= -->
2758 </div>
2764 touching the symbol. On targets that support it, this allows an intelligent
2765 linker to optimize references to the symbol without being impeded as it would be
2768 <p>This is a rare construct that should only be used in rare circumstances, and
2769 should not be exposed to source languages.</p>
2771 </div>
2773 <!-- ======================================================================= -->
2776 </div>
2779 <pre>
2780 %0 = type { i32, void ()* }
2782 </pre>
2783 <p>The <tt>@llvm.global_ctors</tt> array contains a list of constructor functions and associated priorities. The functions referenced by this array will be called in ascending order of priority (i.e. lowest first) when the module is loaded. The order of functions with the same priority is not defined.
2784 </p>
2786 </div>
2788 <!-- ======================================================================= -->
2791 </div>
2794 <pre>
2795 %0 = type { i32, void ()* }
2797 </pre>
2799 <p>The <tt>@llvm.global_dtors</tt> array contains a list of destructor functions and associated priorities. The functions referenced by this array will be called in descending order of priority (i.e. highest first) when the module is loaded. The order of functions with the same priority is not defined.
2800 </p>
2802 </div>
2805 <!-- *********************************************************************** -->
2807 <!-- *********************************************************************** -->
2811 <p>The LLVM instruction set consists of several different classifications of
2818 </div>
2820 <!-- ======================================================================= -->
2822 Instructions</a> </div>
2827 in a program ends with a "Terminator" instruction, which indicates which
2828 block should be executed after the current block is finished. These
2830 control flow, not values (the one exception being the
2833 <p>There are seven different terminator instructions: the
2842 </div>
2844 <!-- _______________________________________________________________________ -->
2846 Instruction</a> </div>
2851 <pre>
2854 </pre>
2858 a value) from a function back to the caller.</p>
2861 value and then causes control flow, and one that just causes control flow to
2862 occur.</p>
2866 return value. The type of the return value must be a
2871 value or a return value with a type that does not match its type, or if it
2873 return value.</p>
2877 the calling function's context. If the caller is a
2879 instruction after the call. If the caller was an
2881 the beginning of the "normal" destination block. If the instruction returns
2882 a value, that value shall set the call or invoke instruction's return
2883 value.</p>
2886 <pre>
2890 </pre>
2892 </div>
2893 <!-- _______________________________________________________________________ -->
2899 <pre>
2900 br i1 <cond>, label <iftrue>, label <iffalse><br> br label <dest> <i>; Unconditional branch</i>
2901 </pre>
2905 different basic block in the current function. There are two forms of this
2906 instruction, corresponding to a conditional branch and an unconditional
2907 branch.</p>
2913 target.</p>
2922 <pre>
2923 Test:
2925 br i1 %cond, label %IfEqual, label %IfUnequal
2926 IfEqual:
2928 IfUnequal:
2930 </pre>
2932 </div>
2934 <!-- _______________________________________________________________________ -->
2937 </div>
2942 <pre>
2943 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
2944 </pre>
2949 instruction, allowing a branch to occur to one of many possible
2950 destinations.</p>
2956 The table is not allowed to contain duplicate constant entries.</p>
2961 is searched for the given value. If the value is found, control flow is
2962 transferred to the corresponding destination; otherwise, control flow is
2963 transferred to the default destination.</p>
2966 <p>Depending on properties of the target machine and the particular
2968 different ways. For example, it could be generated as a series of chained
2969 conditional branches or with a lookup table.</p>
2972 <pre>
2975 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
2978 switch i32 0, label %dest [ ]
2981 switch i32 %val, label %otherwise [ i32 0, label %onzero
2982 i32 1, label %onone
2983 i32 2, label %ontwo ]
2984 </pre>
2986 </div>
2989 <!-- _______________________________________________________________________ -->
2992 </div>
2997 <pre>
2999 </pre>
3004 within the current function, whose address is specified by
3011 rest of the arguments indicate the full set of possible destinations that the
3012 address may point to. Blocks are allowed to occur multiple times in the
3013 destination list, though this isn't particularly useful.</p>
3015 <p>This destination list is required so that dataflow analysis has an accurate
3016 understanding of the CFG.</p>
3020 <p>Control transfers to the block specified in the address argument. All
3021 possible destination blocks must be listed in the label list, otherwise this
3022 instruction has undefined behavior. This implies that jumps to labels
3023 defined in other functions have undefined behavior as well.</p>
3030 <pre>
3031 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3032 </pre>
3034 </div>
3037 <!-- _______________________________________________________________________ -->
3040 </div>
3045 <pre>
3046 <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>]
3048 </pre>
3052 function, with the possibility of control flow transfer to either the
3055 control flow will return to the "normal" label. If the callee (or any
3057 instruction, control is interrupted and continued at the dynamically nearest
3063 <ol>
3065 convention</a> the call should use. If none is specified, the call
3066 defaults to using C calling conventions.</li>
3073 function value being invoked. In most cases, this is a direct function
3075 off an arbitrary pointer to function value.</li>
3078 function to be invoked. </li>
3081 signature argument types and parameter attributes. All arguments must be
3083 signature indicates the function accepts a variable number of arguments,
3084 the extra arguments can be specified.</li>
3095 </ol>
3098 <p>This instruction is designed to operate as a standard
3100 primary difference is that it establishes an association with a label, which
3101 is used by the runtime library to unwind the stack.</p>
3103 <p>This instruction is used in languages with destructors to ensure that proper
3105 exception. Additionally, this is important for implementation of
3108 <p>For the purposes of the SSA form, the definition of the value returned by the
3110 block to the "normal" label. If the callee unwinds then no return value is
3111 available.</p>
3113 <p>Note that the code generator does not yet completely support unwind, and
3114 that the invoke/unwind semantics are likely to change in future versions.</p>
3117 <pre>
3118 %retval = invoke i32 @Test(i32 15) to label %Continue
3122 </pre>
3124 </div>
3126 <!-- _______________________________________________________________________ -->
3129 Instruction</a> </div>
3134 <pre>
3135 unwind
3136 </pre>
3140 at the first callee in the dynamic call stack which used
3142 This is primarily used to implement exception handling.</p>
3146 immediately halt. The dynamic call stack is then searched for the
3148 Once found, execution continues at the "exceptional" destination block
3150 instruction in the dynamic call chain, undefined behavior results.</p>
3152 <p>Note that the code generator does not yet completely support unwind, and
3153 that the invoke/unwind semantics are likely to change in future versions.</p>
3155 </div>
3157 <!-- _______________________________________________________________________ -->
3160 Instruction</a> </div>
3165 <pre>
3166 unreachable
3167 </pre>
3171 instruction is used to inform the optimizer that a particular portion of the
3172 code is not reachable. This can be used to indicate that the code after a
3173 no-return function cannot be reached, and other facts.</p>
3178 </div>
3180 <!-- ======================================================================= -->
3185 <p>Binary operators are used to do most of the computation in a program. They
3186 require two operands of the same type, execute an operation on them, and
3187 produce a single value. The operands might represent multiple data, as is
3189 has the same type as its operands.</p>
3193 </div>
3195 <!-- _______________________________________________________________________ -->
3198 </div>
3203 <pre>
3208 </pre>
3216 integer values. Both arguments must have identical types.</p>
3221 <p>If the sum has unsigned overflow, the result returned is the mathematical
3224 <p>Because LLVM integers use a two's complement representation, this instruction
3225 is appropriate for both signed and unsigned integers.</p>
3231 respectively, occurs.</p>
3234 <pre>
3236 </pre>
3238 </div>
3240 <!-- _______________________________________________________________________ -->
3243 </div>
3248 <pre>
3250 </pre>
3258 floating point values. Both arguments must have identical types.</p>
3264 <pre>
3266 </pre>
3268 </div>
3270 <!-- _______________________________________________________________________ -->
3273 </div>
3278 <pre>
3283 </pre>
3287 operands.</p>
3291 representations.</p>
3296 integer values. Both arguments must have identical types.</p>
3301 <p>If the difference has unsigned overflow, the result returned is the
3303 result.</p>
3305 <p>Because LLVM integers use a two's complement representation, this instruction
3306 is appropriate for both signed and unsigned integers.</p>
3312 respectively, occurs.</p>
3315 <pre>
3318 </pre>
3320 </div>
3322 <!-- _______________________________________________________________________ -->
3325 </div>
3330 <pre>
3332 </pre>
3336 operands.</p>
3340 representations.</p>
3345 floating point values. Both arguments must have identical types.</p>
3351 <pre>
3354 </pre>
3356 </div>
3358 <!-- _______________________________________________________________________ -->
3361 </div>
3366 <pre>
3371 </pre>
3379 integer values. Both arguments must have identical types.</p>
3384 <p>If the result of the multiplication has unsigned overflow, the result
3386 width of the result.</p>
3388 <p>Because LLVM integers use a two's complement representation, and the result
3389 is the same width as the operands, this instruction returns the correct
3390 result for both signed and unsigned integers. If a full product
3392 be sign-extended or zero-extended as appropriate to the width of the full
3393 product.</p>
3399 respectively, occurs.</p>
3402 <pre>
3404 </pre>
3406 </div>
3408 <!-- _______________________________________________________________________ -->
3411 </div>
3416 <pre>
3418 </pre>
3426 floating point values. Both arguments must have identical types.</p>
3432 <pre>
3434 </pre>
3436 </div>
3438 <!-- _______________________________________________________________________ -->
3440 </a></div>
3445 <pre>
3447 </pre>
3455 values. Both arguments must have identical types.</p>
3460 <p>Note that unsigned integer division and signed integer division are distinct
3466 <pre>
3468 </pre>
3470 </div>
3472 <!-- _______________________________________________________________________ -->
3474 </a> </div>
3479 <pre>
3482 </pre>
3490 values. Both arguments must have identical types.</p>
3493 <p>The value produced is the signed integer quotient of the two operands rounded
3494 towards zero.</p>
3496 <p>Note that signed integer division and unsigned integer division are distinct
3499 <p>Division by zero leads to undefined behavior. Overflow also leads to
3500 undefined behavior; this is a rare case, but can occur, for example, by doing
3505 be rounded.</p>
3508 <pre>
3510 </pre>
3512 </div>
3514 <!-- _______________________________________________________________________ -->
3516 Instruction</a> </div>
3521 <pre>
3523 </pre>
3531 floating point values. Both arguments must have identical types.</p>
3537 <pre>
3539 </pre>
3541 </div>
3543 <!-- _______________________________________________________________________ -->
3545 </div>
3550 <pre>
3552 </pre>
3556 division of its two arguments.</p>
3561 values. Both arguments must have identical types.</p>
3565 This instruction always performs an unsigned division to get the
3566 remainder.</p>
3568 <p>Note that unsigned integer remainder and signed integer remainder are
3574 <pre>
3576 </pre>
3578 </div>
3580 <!-- _______________________________________________________________________ -->
3583 </div>
3588 <pre>
3590 </pre>
3594 division of its two operands. This instruction can also take
3596 elements must be integers.</p>
3601 values. Both arguments must have identical types.</p>
3607 a value. For more information about the difference,
3609 Math Forum</a>. For a table of how this is implemented in various languages,
3613 <p>Note that signed integer remainder and unsigned integer remainder are
3616 <p>Taking the remainder of a division by zero leads to undefined behavior.
3617 Overflow also leads to undefined behavior; this is a rare case, but can
3618 occur, for example, by taking the remainder of a 32-bit division of
3620 lets srem be implemented using instructions that return both the result of
3621 the division and the remainder.)</p>
3624 <pre>
3626 </pre>
3628 </div>
3630 <!-- _______________________________________________________________________ -->
3637 <pre>
3639 </pre>
3643 its two operands.</p>
3648 floating point values. Both arguments must have identical types.</p>
3652 has the same sign as the dividend.</p>
3655 <pre>
3657 </pre>
3659 </div>
3661 <!-- ======================================================================= -->
3663 Operations</a> </div>
3667 <p>Bitwise binary operators are used to do various forms of bit-twiddling in a
3668 program. They are generally very efficient instructions and can commonly be
3669 strength reduced from other instructions. They require two operands of the
3670 same type, execute an operation on them, and produce a single value. The
3671 resulting value is the same type as its operands.</p>
3673 </div>
3675 <!-- _______________________________________________________________________ -->
3677 Instruction</a> </div>
3682 <pre>
3684 </pre>
3688 a specified number of bits.</p>
3698 is (statically or dynamically) negative or equal to or larger than the number
3704 <pre>
3709 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> <i>; yields: result=<2 x i32> < i32 2, i32 4></i>
3710 </pre>
3712 </div>
3714 <!-- _______________________________________________________________________ -->
3716 Instruction</a> </div>
3721 <pre>
3723 </pre>
3727 operand shifted to the right a specified number of bits with zero fill.</p>
3735 <p>This instruction always performs a logical shift right operation. The most
3736 significant bits of the result will be filled with zero bits after the shift.
3743 <pre>
3749 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> <i>; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1></i>
3750 </pre>
3752 </div>
3754 <!-- _______________________________________________________________________ -->
3756 Instruction</a> </div>
3760 <pre>
3762 </pre>
3766 operand shifted to the right a specified number of bits with sign
3767 extension.</p>
3775 <p>This instruction always performs an arithmetic shift right operation, The
3776 most significant bits of the result will be filled with the sign bit
3783 <pre>
3789 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> <i>; yields: result=<2 x i32> < i32 -1, i32 0></i>
3790 </pre>
3792 </div>
3794 <!-- _______________________________________________________________________ -->
3796 Instruction</a> </div>
3801 <pre>
3803 </pre>
3807 operands.</p>
3812 values. Both arguments must have identical types.</p>
3818 <tbody>
3819 <tr>
3823 </tr>
3824 <tr>
3828 </tr>
3829 <tr>
3833 </tr>
3834 <tr>
3838 </tr>
3839 <tr>
3843 </tr>
3844 </tbody>
3845 </table>
3848 <pre>
3852 </pre>
3853 </div>
3854 <!-- _______________________________________________________________________ -->
3860 <pre>
3862 </pre>
3866 two operands.</p>
3871 values. Both arguments must have identical types.</p>
3877 <tbody>
3878 <tr>
3882 </tr>
3883 <tr>
3887 </tr>
3888 <tr>
3892 </tr>
3893 <tr>
3897 </tr>
3898 <tr>
3902 </tr>
3903 </tbody>
3904 </table>
3907 <pre>
3911 </pre>
3913 </div>
3915 <!-- _______________________________________________________________________ -->
3917 Instruction</a> </div>
3922 <pre>
3924 </pre>
3934 values. Both arguments must have identical types.</p>
3940 <tbody>
3941 <tr>
3945 </tr>
3946 <tr>
3950 </tr>
3951 <tr>
3955 </tr>
3956 <tr>
3960 </tr>
3961 <tr>
3965 </tr>
3966 </tbody>
3967 </table>
3970 <pre>
3975 </pre>
3977 </div>
3979 <!-- ======================================================================= -->
3982 </div>
3986 <p>LLVM supports several instructions to represent vector operations in a
3987 target-independent manner. These instructions cover the element-access and
3988 vector-specific operations needed to process vectors effectively. While LLVM
3989 does directly support these vector operations, many sophisticated algorithms
3990 will want to use target-specific intrinsics to take full advantage of a
3991 specific target.</p>
3993 </div>
3995 <!-- _______________________________________________________________________ -->
3998 </div>
4003 <pre>
4004 <result> = extractelement <n x <ty>> <val>, i32 <idx> <i>; yields <ty></i>
4005 </pre>
4009 from a vector at a specified index.</p>
4015 indicating the position from which to extract the element. The index may be
4016 a variable.</p>
4019 <p>The result is a scalar of the same type as the element type of
4022 results are undefined.</p>
4025 <pre>
4027 </pre>
4029 </div>
4031 <!-- _______________________________________________________________________ -->
4034 </div>
4039 <pre>
4040 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> <i>; yields <n x <ty>></i>
4041 </pre>
4045 vector at a specified index.</p>
4050 whose type must equal the element type of the first operand. The third
4051 operand is an index indicating the position at which to insert the value.
4052 The index may be a variable.</p>
4058 results are undefined.</p>
4061 <pre>
4062 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 <i>; yields <4 x i32></i>
4063 </pre>
4065 </div>
4067 <!-- _______________________________________________________________________ -->
4070 </div>
4075 <pre>
4076 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> <i>; yields <m x <ty>></i>
4077 </pre>
4081 from two input vectors, returning a vector with the same element type as the
4082 input and length that is the same as the shuffle mask.</p>
4086 with types that match each other. The third argument is a shuffle mask whose
4087 element type is always 'i32'. The result of the instruction is a vector
4088 whose length is the same as the shuffle mask and whose element type is the
4089 same as the element type of the first two operands.</p>
4091 <p>The shuffle mask operand is required to be a constant vector with either
4092 constant integer or undef values.</p>
4095 <p>The elements of the two input vectors are numbered from left to right across
4096 both of the vectors. The shuffle mask operand specifies, for each element of
4097 the result vector, which element of the two input vectors the result element
4098 gets. The element selector may be undef (meaning "don't care") and the
4099 second operand may be undef if performing a shuffle from only one vector.</p>
4102 <pre>
4106 <4 x i32> <i32 0, i32 1, i32 2, i32 3> <i>; yields <4 x i32></i> - Identity shuffle.
4110 <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>
4111 </pre>
4113 </div>
4115 <!-- ======================================================================= -->
4118 </div>
4122 <p>LLVM supports several instructions for working with
4125 </div>
4127 <!-- _______________________________________________________________________ -->
4130 </div>
4135 <pre>
4137 </pre>
4147 specify which value to extract in a similar manner as indices in a
4151 <p>The result is the value at the position in the aggregate specified by the
4152 index operands.</p>
4155 <pre>
4157 </pre>
4159 </div>
4161 <!-- _______________________________________________________________________ -->
4164 </div>
4169 <pre>
4170 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx> <i>; yields <aggregate type></i>
4171 </pre>
4181 value to insert. The following operands are constant indices indicating
4182 the position at which to insert the value in a similar manner as indices in a
4184 value to insert must have the same type as the value identified by the
4185 indices.</p>
4193 <pre>
4196 </pre>
4198 </div>
4201 <!-- ======================================================================= -->
4204 </div>
4208 <p>A key design point of an SSA-based representation is how it represents
4209 memory. In LLVM, no memory locations are in SSA form, which makes things
4210 very simple. This section describes how to read, write, and allocate
4211 memory in LLVM.</p>
4213 </div>
4215 <!-- _______________________________________________________________________ -->
4218 </div>
4223 <pre>
4224 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] <i>; yields {type*}:result</i>
4225 </pre>
4229 currently executing function, to be automatically released when this function
4230 returns to its caller. The object is always allocated in the generic address
4231 space (address space zero).</p>
4236 runtime stack, returning a pointer of the appropriate type to the program.
4237 If "NumElements" is specified, it is the number of elements allocated,
4238 otherwise "NumElements" is defaulted to be one. If a constant alignment is
4239 specified, the value result of the allocation is guaranteed to be aligned to
4240 at least that boundary. If not specified, or if zero, the target can choose
4241 to align the allocation on any convenient boundary compatible with the
4242 type.</p>
4247 <p>Memory is allocated; a pointer is returned. The operation is undefined if
4249 memory is automatically released when the function returns. The
4251 variables that must have an address available. When the function returns
4254 reclaimed. Allocating zero bytes is legal, but the result is undefined.</p>
4257 <pre>
4262 </pre>
4264 </div>
4266 <!-- _______________________________________________________________________ -->
4268 Instruction</a> </div>
4273 <pre>
4274 <result> = load <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>]
4275 <result> = volatile load <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>]
4277 </pre>
4284 from which to load. The pointer must point to
4291 operation (that is, the alignment of the memory address). A value of 0 or an
4293 alignment for the target. It is the responsibility of the code emitter to
4294 ensure that the alignment information is correct. Overestimating the
4295 alignment results in undefined behavior. Underestimating the alignment may
4302 and code generator that this load is not expected to be reused in the cache.
4303 The code generator may select special instructions to save cache bandwidth,
4307 <p>The location of memory pointed to is loaded. If the value being loaded is of
4308 scalar type then the number of bytes read does not exceed the minimum number
4309 of bytes needed to hold all bits of the type. For example, loading an
4312 is undefined if the value was not originally written using a store of the
4313 same type.</p>
4316 <pre>
4320 </pre>
4322 </div>
4324 <!-- _______________________________________________________________________ -->
4326 Instruction</a> </div>
4331 <pre>
4332 store <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] <i>; yields {void}</i>
4333 volatile store <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] <i>; yields {void}</i>
4334 </pre>
4341 and an address at which to store it. The type of the
4350 operation (that is, the alignment of the memory address). A value of 0 or an
4351 omitted "align" argument means that the operation has the preferential
4352 alignment for the target. It is the responsibility of the code emitter to
4353 ensure that the alignment information is correct. Overestimating the
4354 alignment results in an undefined behavior. Underestimating the alignment may
4357 <p>The optional !nontemporal metadata must reference a single metatadata
4359 value 1. The existence of the !nontemporal metatadata on the
4360 instruction tells the optimizer and code generator that this load is
4361 not expected to be reused in the cache. The code generator may
4362 select special instructions to save cache bandwidth, such as the
4363 MOVNT instruction on x86.</p>
4370 does not exceed the minimum number of bytes needed to hold all bits of the
4373 integral number of bytes, it is unspecified what happens to the extra bits
4374 that do not belong to the type, but they will typically be overwritten.</p>
4377 <pre>
4381 </pre>
4383 </div>
4385 <!-- _______________________________________________________________________ -->
4388 </div>
4393 <pre>
4396 </pre>
4401 It performs address calculation only and does not access memory.</p>
4404 <p>The first argument is always a pointer, and forms the basis of the
4405 calculation. The remaining arguments are indices that indicate which of the
4406 elements of the aggregate object are indexed. The interpretation of each
4407 index is dependent on the type being indexed into. The first index always
4408 indexes the pointer value given as the first argument, the second index
4409 indexes a value of the type pointed to (not necessarily the value directly
4410 pointed to, since the first index can be non-zero), etc. The first type
4411 indexed into must be a pointer value, subsequent types can be arrays,
4412 vectors, structs and unions. Note that subsequent types being indexed into
4413 can never be pointers, since that would require loading the pointer before
4414 continuing calculation.</p>
4416 <p>The type of each index argument depends on the type it is indexing into.
4419 or vector, integers of any width are allowed, and they are not required to be
4420 constant.</p>
4422 <p>For example, let's consider a C code fragment and how it gets compiled to
4423 LLVM:</p>
4426 struct RT {
4427 char A;
4429 char C;
4430 };
4431 struct ST {
4432 int X;
4433 double Y;
4434 struct RT Z;
4435 };
4437 int *foo(struct ST *s) {
4439 }
4440 </pre>
4448 define i32* @foo(%ST* %s) {
4449 entry:
4451 ret i32* %reg
4452 }
4453 </pre>
4458 }</tt>' type, a structure. The second index indexes into the third element
4460 i8 }</tt>' type, another structure. The third index indexes into the second
4462 array. The two dimensions of the array are subscripted into, yielding an
4466 <p>Note that it is perfectly legal to index partially through a structure,
4467 returning a pointer to an inner element. Because of this, the LLVM code for
4468 the given testcase is equivalent to:</p>
4470 <pre>
4471 define i32* @foo(%ST* %s) {
4477 ret i32* %t5
4478 }
4479 </pre>
4484 or if any of the addresses that would be formed by successive addition of
4485 the offsets implied by the indices to the base address with infinitely
4488 the addresses that point into the object, plus the address one byte past
4489 the end.</p>
4492 the base address with silently-wrapping two's complement arithmetic, and
4494 pointed to by the base pointer. The result value may not necessarily be
4495 used to access memory though, even if it happens to point into allocated
4497 section for more information.</p>
4499 <p>The getelementptr instruction is often confusing. For some more insight into
4503 <pre>
4512 </pre>
4514 </div>
4516 <!-- ======================================================================= -->
4518 </div>
4522 <p>The instructions in this category are the conversion instructions (casting)
4523 which all take a single operand and a type. They perform various bit
4524 conversions on the operand.</p>
4526 </div>
4528 <!-- _______________________________________________________________________ -->
4531 </div>
4535 <pre>
4537 </pre>
4546 size and type of the result, which must be
4549 allowed.</p>
4558 <pre>
4562 </pre>
4564 </div>
4566 <!-- _______________________________________________________________________ -->
4569 </div>
4573 <pre>
4575 </pre>
4596 <pre>
4599 </pre>
4601 </div>
4603 <!-- _______________________________________________________________________ -->
4606 </div>
4610 <pre>
4612 </pre>
4632 <pre>
4635 </pre>
4637 </div>
4639 <!-- _______________________________________________________________________ -->
4642 </div>
4647 <pre>
4649 </pre>
4667 undefined.</p>
4670 <pre>
4673 </pre>
4675 </div>
4677 <!-- _______________________________________________________________________ -->
4680 </div>
4684 <pre>
4686 </pre>
4690 floating point value.</p>
4696 type must be smaller than the destination type.</p>
4706 <pre>
4709 </pre>
4711 </div>
4713 <!-- _______________________________________________________________________ -->
4716 </div>
4720 <pre>
4722 </pre>
4738 towards zero) unsigned integer value. If the value cannot fit
4742 <pre>
4746 </pre>
4748 </div>
4750 <!-- _______________________________________________________________________ -->
4753 </div>
4757 <pre>
4759 </pre>
4777 the results are undefined.</p>
4780 <pre>
4784 </pre>
4786 </div>
4788 <!-- _______________________________________________________________________ -->
4791 </div>
4795 <pre>
4797 </pre>
4812 integer quantity and converts it to the corresponding floating point
4813 value. If the value cannot fit in the floating point value, the results are
4814 undefined.</p>
4817 <pre>
4820 </pre>
4822 </div>
4824 <!-- _______________________________________________________________________ -->
4827 </div>
4831 <pre>
4833 </pre>
4848 quantity and converts it to the corresponding floating point value. If the
4849 value cannot fit in the floating point value, the results are undefined.</p>
4852 <pre>
4855 </pre>
4857 </div>
4859 <!-- _______________________________________________________________________ -->
4862 </div>
4866 <pre>
4868 </pre>
4882 truncating or zero extending that value to the size of the integer type. If
4886 change.</p>
4889 <pre>
4892 </pre>
4894 </div>
4896 <!-- _______________________________________________________________________ -->
4899 </div>
4903 <pre>
4905 </pre>
4913 value to cast, and a type to cast it to, which must be a
4921 than the size of a pointer then a zero extension is done. If they are the
4925 <pre>
4929 </pre>
4931 </div>
4933 <!-- _______________________________________________________________________ -->
4936 </div>
4940 <pre>
4942 </pre>
4950 non-aggregate first class value, and a type to cast it to, which must also be
4953 identical. If the source type is a pointer, the destination type must also be
4954 a pointer. This instruction supports bitwise conversion of vectors to
4955 integers and to vectors of other types (as long as they have the same
4956 size).</p>
4963 be converted to other pointer types with this instruction. To convert
4968 <pre>
4972 </pre>
4974 </div>
4976 <!-- ======================================================================= -->
4982 defy better classification.</p>
4984 </div>
4986 <!-- _______________________________________________________________________ -->
4988 </div>
4993 <pre>
4994 <result> = icmp <cond> <ty> <op1>, <op2> <i>; yields {i1} or {<N x i1>}:result</i>
4995 </pre>
4999 boolean values based on comparison of its two integer, integer vector, or
5000 pointer operands.</p>
5004 the condition code indicating the kind of comparison to perform. It is not a
5005 value, just a keyword. The possible condition code are:</p>
5007 <ol>
5018 </ol>
5022 typed. They must also be identical types.</p>
5028 result, as follows:</p>
5030 <ol>
5033 performed.</li>
5037 performed.</li>
5064 </ol>
5067 values are compared as if they were integers.</p>
5069 <p>If the operands are integer vectors, then they are compared element by
5074 <pre>
5081 </pre>
5083 <p>Note that the code generator does not yet support vector types with
5086 </div>
5088 <!-- _______________________________________________________________________ -->
5090 </div>
5095 <pre>
5096 <result> = fcmp <cond> <ty> <op1>, <op2> <i>; yields {i1} or {<N x i1>}:result</i>
5097 </pre>
5101 values based on comparison of its operands.</p>
5103 <p>If the operands are floating point scalars, then the result type is a boolean
5106 <p>If the operands are floating point vectors, then the result type is a vector
5107 of boolean with the same number of elements as the operands being
5108 compared.</p>
5112 the condition code indicating the kind of comparison to perform. It is not a
5113 value, just a keyword. The possible condition code are:</p>
5115 <ol>
5132 </ol>
5140 identical types.</p>
5145 vectors, then the vectors are compared element by element. Each comparison
5147 follows:</p>
5149 <ol>
5193 </ol>
5196 <pre>
5201 </pre>
5203 <p>Note that the code generator does not yet support vector types with
5206 </div>
5208 <!-- _______________________________________________________________________ -->
5211 </div>
5216 <pre>
5218 </pre>
5222 SSA graph representing the function.</p>
5225 <p>The type of the incoming values is specified with the first type field. After
5227 one pair for each predecessor basic block of the current block. Only values
5229 arguments to the PHI node. Only labels may be used as the label
5230 arguments.</p>
5232 <p>There must be no non-phi instructions between the start of a basic block and
5233 the PHI instructions: i.e. PHI instructions must be first in a basic
5234 block.</p>
5236 <p>For the purposes of the SSA form, the use of each incoming value is deemed to
5237 occur on the edge from the corresponding predecessor block to the current
5239 value on the same edge).</p>
5243 specified by the pair corresponding to the predecessor basic block that
5244 executed just prior to the current block.</p>
5247 <pre>
5248 Loop: ; Infinite loop that counts from 0 on up...
5249 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
5250 %nextindvar = add i32 %indvar, 1
5251 br label %Loop
5252 </pre>
5254 </div>
5256 <!-- _______________________________________________________________________ -->
5259 </div>
5264 <pre>
5265 <result> = select <i>selty</i> <cond>, <ty> <val1>, <ty> <val2> <i>; yields ty</i>
5268 </pre>
5272 condition, without branching.</p>
5277 values indicating the condition, and two values of the
5279 vectors and the condition is a scalar, then entire vectors are selected, not
5280 individual elements.</p>
5284 first value argument; otherwise, it returns the second value argument.</p>
5286 <p>If the condition is a vector of i1, then the value arguments must be vectors
5287 of the same size, and the selection is done element by element.</p>
5290 <pre>
5292 </pre>
5294 <p>Note that the code generator does not yet support conditions
5295 with vector type.</p>
5297 </div>
5299 <!-- _______________________________________________________________________ -->
5302 </div>
5307 <pre>
5308 <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>]
5309 </pre>
5317 <ol>
5319 access any allocas or varargs in the caller. Note that calls may be
5320 marked "tail" even if they do not occur before
5322 present, the function call is eligible for tail call optimization,
5324 optimized into a jump</a>. The code generator may optimize calls marked
5326 sibling call optimization</a> when the caller and callee have
5327 matching signatures, or 2) forced tail call optimization when the
5328 following extra requirements are met:
5329 <ul>
5330 <li>Caller and callee both have the calling
5332 <li>The call is in tail position (ret immediately follows call and ret
5333 uses value of call or is void).</li>
5337 constraints are met.</a></li>
5338 </ul>
5339 </li>
5342 convention</a> the call should use. If none is specified, the call
5343 defaults to using C calling conventions. The calling convention of the
5344 call must match the calling convention of the target function, or else the
5345 behavior is undefined.</li>
5352 type of the return value. Functions that return no value are marked
5356 being invoked. The argument types must match the types implied by this
5357 signature. This type can be omitted if the function is not varargs and if
5358 the function type does not return a pointer to a function.</li>
5361 be invoked. In most cases, this is a direct function invocation, but
5363 to function value.</li>
5366 signature argument types and parameter attributes. All arguments must be
5368 signature indicates the function accepts a variable number of arguments,
5369 the extra arguments can be specified.</li>
5374 </ol>
5378 a specified function, with its incoming arguments bound to the specified
5380 function, control flow continues with the instruction after the function
5381 call, and the return value of the function is bound to the result
5382 argument.</p>
5385 <pre>
5386 %retval = call i32 @test(i32 %argc)
5390 call void %foo(i8 97 signext)
5392 %struct.A = type { i32, i8 }
5398 </pre>
5400 <p>llvm treats calls to some functions with names and arguments that match the
5401 standard C99 library as being the C99 library functions, and may perform
5402 optimizations or generate code for them under that assumption. This is
5403 something we'd like to change in the future to provide better support for
5404 freestanding environments and non-C-based languages.</p>
5406 </div>
5408 <!-- _______________________________________________________________________ -->
5411 </div>
5416 <pre>
5418 </pre>
5422 the "variable argument" area of a function call. It is used to implement the
5427 argument. It returns a value of the specified argument type and increments
5434 to the next argument. For more information, see the variable argument
5437 <p>It is legal for this instruction to be called in a function which does not
5439 function.</p>
5443 argument.</p>
5448 <p>Note that the code generator does not yet fully support va_arg on many
5449 targets. Also, it does not currently support va_arg with aggregate types on
5450 any target.</p>
5452 </div>
5454 <!-- *********************************************************************** -->
5456 <!-- *********************************************************************** -->
5461 well known names and semantics and are required to follow certain
5462 restrictions. Overall, these intrinsics represent an extension mechanism for
5463 the LLVM language that does not require changing all of the transformations
5464 in LLVM when adding to the language (or the bitcode reader/writer, the
5465 parser, etc...).</p>
5468 prefix is reserved in LLVM for intrinsic names; thus, function names may not
5469 begin with this prefix. Intrinsic functions must always be external
5470 functions: you cannot define the body of intrinsic functions. Intrinsic
5471 functions may only be used in call or invoke instructions: it is illegal to
5472 take the address of an intrinsic function. Additionally, because intrinsic
5473 functions are part of the LLVM language, it is required if any are added that
5474 they be documented here.</p>
5476 <p>Some intrinsic functions can be overloaded, i.e., the intrinsic represents a
5477 family of functions that perform the same operation but on different data
5478 types. Because LLVM can represent over 8 million different integer types,
5479 overloading is used commonly to allow an intrinsic function to operate on any
5480 integer type. One or more of the argument types or the result type can be
5481 overloaded to accept any integer type. Argument types may also be defined as
5482 exactly matching a previous argument's type or the result type. This allows
5483 an intrinsic function which accepts multiple arguments, but needs all of them
5484 to be of the same type, to only be overloaded with respect to a single
5485 argument or the result.</p>
5487 <p>Overloaded intrinsics will have the names of its overloaded argument types
5488 encoded into its function name, each preceded by a period. Only those types
5489 which are overloaded result in a name suffix. Arguments whose type is matched
5491 can take an integer of any width and returns an integer of exactly the same
5492 integer width. This leads to a family of functions such as
5494 %val)</tt>. Only one type, the return type, is overloaded, and only one type
5495 suffix is required. Because the argument's type is matched against the return
5496 type, it does not require its own name suffix.</p>
5498 <p>To learn how to add an intrinsic function, please see the
5501 </div>
5503 <!-- ======================================================================= -->
5506 </div>
5510 <p>Variable argument support is defined in LLVM with
5512 intrinsic functions. These functions are related to the similarly named
5515 <p>All of these functions operate on arguments that use a target-specific value
5516 type "<tt>va_list</tt>". The LLVM assembly language reference manual does
5517 not define what this type is, so all transformations should be prepared to
5518 handle these functions regardless of the type used.</p>
5521 instruction and the variable argument handling intrinsic functions are
5522 used.</p>
5525 define i32 @test(i32 %X, ...) {
5526 ; Initialize variable argument processing
5527 %ap = alloca i8*
5528 %ap2 = bitcast i8** %ap to i8*
5529 call void @llvm.va_start(i8* %ap2)
5531 ; Read a single integer argument
5532 %tmp = va_arg i8** %ap, i32
5534 ; Demonstrate usage of llvm.va_copy and llvm.va_end
5535 %aq = alloca i8*
5536 %aq2 = bitcast i8** %aq to i8*
5537 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
5538 call void @llvm.va_end(i8* %aq2)
5540 ; Stop processing of arguments.
5541 call void @llvm.va_end(i8* %ap2)
5542 ret i32 %tmp
5543 }
5545 declare void @llvm.va_start(i8*)
5546 declare void @llvm.va_copy(i8*, i8*)
5547 declare void @llvm.va_end(i8*)
5548 </pre>
5550 </div>
5552 <!-- _______________________________________________________________________ -->
5555 </div>
5561 <pre>
5563 </pre>
5574 macro available in C. In a target-dependent way, it initializes
5578 need to know the last argument of the function as the compiler can figure
5579 that out.</p>
5581 </div>
5583 <!-- _______________________________________________________________________ -->
5586 </div>
5591 <pre>
5593 </pre>
5597 which has been initialized previously
5606 macro available in C. In a target-dependent way, it destroys
5612 </div>
5614 <!-- _______________________________________________________________________ -->
5617 </div>
5622 <pre>
5624 </pre>
5628 from the source argument list to the destination argument list.</p>
5633 from.</p>
5637 macro available in C. In a target-dependent way, it copies the
5639 element. This intrinsic is necessary because
5641 arbitrarily complex and require, for example, memory allocation.</p>
5643 </div>
5645 <!-- ======================================================================= -->
5648 </div>
5653 Collection</a> (GC) requires the implementation and generation of these
5655 roots on the stack</a>, as well as garbage collector implementations that
5657 barriers. Front-ends for type-safe garbage collected languages should generate
5658 these intrinsics to make use of the LLVM garbage collectors. For more details,
5662 <p>The garbage collection intrinsics only operate on objects in the generic
5663 address space (address space zero).</p>
5665 </div>
5667 <!-- _______________________________________________________________________ -->
5670 </div>
5675 <pre>
5676 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
5677 </pre>
5681 the code generator, and allows some metadata to be associated with it.</p>
5684 <p>The first argument specifies the address of a stack object that contains the
5685 root pointer. The second pointer (which must be either a constant or a
5686 global value address) contains the meta-data to be associated with the
5687 root.</p>
5691 location. At compile-time, the code generator generates information to allow
5696 </div>
5698 <!-- _______________________________________________________________________ -->
5701 </div>
5706 <pre>
5707 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
5708 </pre>
5712 locations, allowing garbage collector implementations that require read
5713 barriers.</p>
5716 <p>The second argument is the address to read from, which should be an address
5717 allocated from the garbage collector. The first object is a pointer to the
5718 start of the referenced object, if needed by the language runtime (otherwise
5719 null).</p>
5723 instruction, but may be replaced with substantially more complex code by the
5728 </div>
5730 <!-- _______________________________________________________________________ -->
5733 </div>
5738 <pre>
5739 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
5740 </pre>
5744 locations, allowing garbage collector implementations that require write
5745 barriers (such as generational or reference counting collectors).</p>
5748 <p>The first argument is the reference to store, the second is the start of the
5749 object to store it to, and the third is the address of the field of Obj to
5750 store to. If the runtime does not require a pointer to the object, Obj may
5751 be null.</p>
5755 instruction, but may be replaced with substantially more complex code by the
5760 </div>
5762 <!-- ======================================================================= -->
5765 </div>
5769 <p>These intrinsics are provided by LLVM to expose special features that may
5770 only be implemented with code generator support.</p>
5772 </div>
5774 <!-- _______________________________________________________________________ -->
5777 </div>
5782 <pre>
5784 </pre>
5788 target-specific value indicating the return address of the current function
5789 or one of its callers.</p>
5792 <p>The argument to this intrinsic indicates which function to return the address
5793 for. Zero indicates the calling function, one indicates its caller, etc.
5798 indicating the return address of the specified call frame, or zero if it
5799 cannot be identified. The value returned by this intrinsic is likely to be
5800 incorrect or 0 for arguments other than zero, so it should only be used for
5801 debugging purposes.</p>
5803 <p>Note that calling this intrinsic does not prevent function inlining or other
5804 aggressive transformations, so the value returned may not be that of the
5805 obvious source-language caller.</p>
5807 </div>
5809 <!-- _______________________________________________________________________ -->
5812 </div>
5817 <pre>
5819 </pre>
5823 target-specific frame pointer value for the specified stack frame.</p>
5826 <p>The argument to this intrinsic indicates which function to return the frame
5827 pointer for. Zero indicates the calling function, one indicates its caller,
5832 indicating the frame address of the specified call frame, or zero if it
5833 cannot be identified. The value returned by this intrinsic is likely to be
5834 incorrect or 0 for arguments other than zero, so it should only be used for
5835 debugging purposes.</p>
5837 <p>Note that calling this intrinsic does not prevent function inlining or other
5838 aggressive transformations, so the value returned may not be that of the
5839 obvious source-language caller.</p>
5841 </div>
5843 <!-- _______________________________________________________________________ -->
5846 </div>
5851 <pre>
5852 declare i8* @llvm.stacksave()
5853 </pre>
5857 of the function stack, for use
5859 useful for implementing language features like scoped automatic variable
5860 sized arrays in C99.</p>
5863 <p>This intrinsic returns a opaque pointer value that can be passed
5871 </div>
5873 <!-- _______________________________________________________________________ -->
5876 </div>
5881 <pre>
5882 declare void @llvm.stackrestore(i8* %ptr)
5883 </pre>
5887 the function stack to the state it was in when the
5889 executed. This is useful for implementing language features like scoped
5890 automatic variable sized arrays in C99.</p>
5893 <p>See the description
5896 </div>
5898 <!-- _______________________________________________________________________ -->
5901 </div>
5906 <pre>
5908 </pre>
5912 insert a prefetch instruction if supported; otherwise, it is a noop.
5913 Prefetches have no effect on the behavior of the program but can change its
5914 performance characteristics.</p>
5924 <p>This intrinsic does not modify the behavior of the program. In particular,
5925 prefetches cannot trap and do not produce a value. On targets that support
5926 this intrinsic, the prefetch can provide hints to the processor cache for
5927 better performance.</p>
5929 </div>
5931 <!-- _______________________________________________________________________ -->
5934 </div>
5939 <pre>
5941 </pre>
5945 Counter (PC) in a region of code to simulators and other tools. The method
5946 is target specific, but it is expected that the marker will use exported
5947 symbols to transmit the PC of the marker. The marker makes no guarantees
5948 that it will remain with any specific instruction after optimizations. It is
5949 possible that the presence of a marker will inhibit optimizations. The
5950 intended use is to be inserted after optimizations to allow correlations of
5951 simulation runs.</p>
5957 <p>This intrinsic does not modify the behavior of the program. Backends that do
5958 not support this intrinsic may ignore it.</p>
5960 </div>
5962 <!-- _______________________________________________________________________ -->
5965 </div>
5970 <pre>
5971 declare i64 @llvm.readcyclecounter()
5972 </pre>
5976 counter register (or similar low latency, high accuracy clocks) on those
5977 targets that support it. On X86, it should map to RDTSC. On Alpha, it
5978 should map to RPCC. As the backing counters overflow quickly (on the order
5982 <p>When directly supported, reading the cycle counter should not modify any
5983 memory. Implementations are allowed to either return a application specific
5984 value or a system wide value. On backends without support, this is lowered
5987 </div>
5989 <!-- ======================================================================= -->
5992 </div>
5996 <p>LLVM provides intrinsics for a few important standard C library functions.
5997 These intrinsics allow source-language front-ends to pass information about
5998 the alignment of the pointer arguments to the code generator, providing
5999 opportunity for more efficient code generation.</p>
6001 </div>
6003 <!-- _______________________________________________________________________ -->
6006 </div>
6012 integer bit width and for different address spaces. Not all targets support
6013 all bit widths however.</p>
6015 <pre>
6020 </pre>
6024 source location to the destination location.</p>
6027 intrinsics do not return a value, takes extra alignment/isvolatile arguments
6028 and the pointers can be in specified address spaces.</p>
6032 <p>The first argument is a pointer to the destination, the second is a pointer
6033 to the source. The third argument is an integer argument specifying the
6034 number of bytes to copy, the fourth argument is the alignment of the
6035 source and destination locations, and the fifth is a boolean indicating a
6036 volatile access.</p>
6039 then the caller guarantees that both the source and destination pointers are
6040 aligned to that boundary.</p>
6044 The detailed access behavior is not very cleanly specified and it is unwise
6045 to depend on it.</p>
6050 source location to the destination location, which are not allowed to
6051 overlap. It copies "len" bytes of memory over. If the argument is known to
6052 be aligned to some boundary, this can be specified as the fourth argument,
6055 </div>
6057 <!-- _______________________________________________________________________ -->
6060 </div>
6065 <p>This is an overloaded intrinsic. You can use llvm.memmove on any integer bit
6066 width and for different address space. Not all targets support all bit
6067 widths however.</p>
6069 <pre>
6074 </pre>
6078 source location to the destination location. It is similar to the
6080 overlap.</p>
6083 intrinsics do not return a value, takes extra alignment/isvolatile arguments
6084 and the pointers can be in specified address spaces.</p>
6088 <p>The first argument is a pointer to the destination, the second is a pointer
6089 to the source. The third argument is an integer argument specifying the
6090 number of bytes to copy, the fourth argument is the alignment of the
6091 source and destination locations, and the fifth is a boolean indicating a
6092 volatile access.</p>
6095 then the caller guarantees that the source and destination pointers are
6096 aligned to that boundary.</p>
6100 The detailed access behavior is not very cleanly specified and it is unwise
6101 to depend on it.</p>
6106 source location to the destination location, which may overlap. It copies
6107 "len" bytes of memory over. If the argument is known to be aligned to some
6108 boundary, this can be specified as the fourth argument, otherwise it should
6111 </div>
6113 <!-- _______________________________________________________________________ -->
6116 </div>
6121 <p>This is an overloaded intrinsic. You can use llvm.memset on any integer bit
6122 width and for different address spaces. However, not all targets support all
6123 bit widths.</p>
6125 <pre>
6130 </pre>
6134 particular byte value.</p>
6137 intrinsic does not return a value and takes extra alignment/volatile
6138 arguments. Also, the destination can be in an arbitrary address space.</p>
6141 <p>The first argument is a pointer to the destination to fill, the second is the
6142 byte value with which to fill it, the third argument is an integer argument
6143 specifying the number of bytes to fill, and the fourth argument is the known
6144 alignment of the destination location.</p>
6147 then the caller guarantees that the destination pointer is aligned to that
6148 boundary.</p>
6152 The detailed access behavior is not very cleanly specified and it is unwise
6153 to depend on it.</p>
6157 at the destination location. If the argument is known to be aligned to some
6158 boundary, this can be specified as the fourth argument, otherwise it should
6161 </div>
6163 <!-- _______________________________________________________________________ -->
6166 </div>
6172 floating point or vector of floating point type. Not all targets support all
6173 types however.</p>
6175 <pre>
6176 declare float @llvm.sqrt.f32(float %Val)
6177 declare double @llvm.sqrt.f64(double %Val)
6178 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
6179 declare fp128 @llvm.sqrt.f128(fp128 %Val)
6180 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
6181 </pre>
6187 behavior for negative numbers other than -0.0 (which allows for better
6188 optimization, because there is no need to worry about errno being
6192 <p>The argument and return value are floating point numbers of the same
6193 type.</p>
6196 <p>This function returns the sqrt of the specified operand if it is a
6197 nonnegative floating point number.</p>
6199 </div>
6201 <!-- _______________________________________________________________________ -->
6204 </div>
6210 floating point or vector of floating point type. Not all targets support all
6211 types however.</p>
6213 <pre>
6214 declare float @llvm.powi.f32(float %Val, i32 %power)
6215 declare double @llvm.powi.f64(double %Val, i32 %power)
6216 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
6217 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
6218 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
6219 </pre>
6223 specified (positive or negative) power. The order of evaluation of
6224 multiplications is not defined. When a vector of floating point type is
6225 used, the second argument remains a scalar integer value.</p>
6228 <p>The second argument is an integer power, and the first is a value to raise to
6229 that power.</p>
6232 <p>This function returns the first value raised to the second power with an
6233 unspecified sequence of rounding operations.</p>
6235 </div>
6237 <!-- _______________________________________________________________________ -->
6240 </div>
6246 floating point or vector of floating point type. Not all targets support all
6247 types however.</p>
6249 <pre>
6250 declare float @llvm.sin.f32(float %Val)
6251 declare double @llvm.sin.f64(double %Val)
6252 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
6253 declare fp128 @llvm.sin.f128(fp128 %Val)
6254 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
6255 </pre>
6261 <p>The argument and return value are floating point numbers of the same
6262 type.</p>
6265 <p>This function returns the sine of the specified operand, returning the same
6267 in the same way.</p>
6269 </div>
6271 <!-- _______________________________________________________________________ -->
6274 </div>
6280 floating point or vector of floating point type. Not all targets support all
6281 types however.</p>
6283 <pre>
6284 declare float @llvm.cos.f32(float %Val)
6285 declare double @llvm.cos.f64(double %Val)
6286 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
6287 declare fp128 @llvm.cos.f128(fp128 %Val)
6288 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
6289 </pre>
6295 <p>The argument and return value are floating point numbers of the same
6296 type.</p>
6299 <p>This function returns the cosine of the specified operand, returning the same
6301 in the same way.</p>
6303 </div>
6305 <!-- _______________________________________________________________________ -->
6308 </div>
6314 floating point or vector of floating point type. Not all targets support all
6315 types however.</p>
6317 <pre>
6318 declare float @llvm.pow.f32(float %Val, float %Power)
6319 declare double @llvm.pow.f64(double %Val, double %Power)
6320 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
6321 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
6322 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
6323 </pre>
6327 specified (positive or negative) power.</p>
6330 <p>The second argument is a floating point power, and the first is a value to
6331 raise to that power.</p>
6334 <p>This function returns the first value raised to the second power, returning
6336 conditions in the same way.</p>
6338 </div>
6340 <!-- ======================================================================= -->
6343 </div>
6347 <p>LLVM provides intrinsics for a few important bit manipulation operations.
6348 These allow efficient code generation for some algorithms.</p>
6350 </div>
6352 <!-- _______________________________________________________________________ -->
6355 </div>
6360 <p>This is an overloaded intrinsic function. You can use bswap on any integer
6363 <pre>
6367 </pre>
6371 values with an even number of bytes (positive multiple of 16 bits). These
6372 are useful for performing operations on data that is not in the target's
6373 native byte order.</p>
6377 and low byte of the input i16 swapped. Similarly,
6383 more, respectively).</p>
6385 </div>
6387 <!-- _______________________________________________________________________ -->
6390 </div>
6395 <p>This is an overloaded intrinsic. You can use llvm.ctpop on any integer bit
6396 width. Not all targets support all bit widths however.</p>
6398 <pre>
6404 </pre>
6408 in a value.</p>
6411 <p>The only argument is the value to be counted. The argument may be of any
6412 integer type. The return type must match the argument type.</p>
6417 </div>
6419 <!-- _______________________________________________________________________ -->
6422 </div>
6428 integer bit width. Not all targets support all bit widths however.</p>
6430 <pre>
6436 </pre>
6440 leading zeros in a variable.</p>
6443 <p>The only argument is the value to be counted. The argument may be of any
6444 integer type. The return type must match the argument type.</p>
6448 zeros in a variable. If the src == 0 then the result is the size in bits of
6451 </div>
6453 <!-- _______________________________________________________________________ -->
6456 </div>
6462 integer bit width. Not all targets support all bit widths however.</p>
6464 <pre>
6470 </pre>
6474 trailing zeros.</p>
6477 <p>The only argument is the value to be counted. The argument may be of any
6478 integer type. The return type must match the argument type.</p>
6482 zeros in a variable. If the src == 0 then the result is the size in bits of
6485 </div>
6487 <!-- ======================================================================= -->
6490 </div>
6496 </div>
6498 <!-- _______________________________________________________________________ -->
6501 </div>
6507 on any integer bit width.</p>
6509 <pre>
6510 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
6511 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
6512 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
6513 </pre>
6517 a signed addition of the two arguments, and indicate whether an overflow
6518 occurred during the signed summation.</p>
6521 <p>The arguments (%a and %b) and the first element of the result structure may
6522 be of integer types of any bit width, but they must have the same bit
6523 width. The second element of the result structure must be of
6525 undergo signed addition.</p>
6529 a signed addition of the two variables. They return a structure — the
6530 first element of which is the signed summation, and the second element of
6531 which is a bit specifying if the signed summation resulted in an
6532 overflow.</p>
6535 <pre>
6536 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
6537 %sum = extractvalue {i32, i1} %res, 0
6538 %obit = extractvalue {i32, i1} %res, 1
6539 br i1 %obit, label %overflow, label %normal
6540 </pre>
6542 </div>
6544 <!-- _______________________________________________________________________ -->
6547 </div>
6553 on any integer bit width.</p>
6555 <pre>
6556 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
6557 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
6558 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
6559 </pre>
6563 an unsigned addition of the two arguments, and indicate whether a carry
6564 occurred during the unsigned summation.</p>
6567 <p>The arguments (%a and %b) and the first element of the result structure may
6568 be of integer types of any bit width, but they must have the same bit
6569 width. The second element of the result structure must be of
6571 undergo unsigned addition.</p>
6575 an unsigned addition of the two arguments. They return a structure —
6576 the first element of which is the sum, and the second element of which is a
6577 bit specifying if the unsigned summation resulted in a carry.</p>
6580 <pre>
6581 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
6582 %sum = extractvalue {i32, i1} %res, 0
6583 %obit = extractvalue {i32, i1} %res, 1
6584 br i1 %obit, label %carry, label %normal
6585 </pre>
6587 </div>
6589 <!-- _______________________________________________________________________ -->
6592 </div>
6598 on any integer bit width.</p>
6600 <pre>
6601 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
6602 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
6603 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
6604 </pre>
6608 a signed subtraction of the two arguments, and indicate whether an overflow
6609 occurred during the signed subtraction.</p>
6612 <p>The arguments (%a and %b) and the first element of the result structure may
6613 be of integer types of any bit width, but they must have the same bit
6614 width. The second element of the result structure must be of
6616 undergo signed subtraction.</p>
6620 a signed subtraction of the two arguments. They return a structure —
6621 the first element of which is the subtraction, and the second element of
6622 which is a bit specifying if the signed subtraction resulted in an
6623 overflow.</p>
6626 <pre>
6627 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
6628 %sum = extractvalue {i32, i1} %res, 0
6629 %obit = extractvalue {i32, i1} %res, 1
6630 br i1 %obit, label %overflow, label %normal
6631 </pre>
6633 </div>
6635 <!-- _______________________________________________________________________ -->
6638 </div>
6644 on any integer bit width.</p>
6646 <pre>
6647 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
6648 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
6649 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
6650 </pre>
6654 an unsigned subtraction of the two arguments, and indicate whether an
6655 overflow occurred during the unsigned subtraction.</p>
6658 <p>The arguments (%a and %b) and the first element of the result structure may
6659 be of integer types of any bit width, but they must have the same bit
6660 width. The second element of the result structure must be of
6662 undergo unsigned subtraction.</p>
6666 an unsigned subtraction of the two arguments. They return a structure —
6667 the first element of which is the subtraction, and the second element of
6668 which is a bit specifying if the unsigned subtraction resulted in an
6669 overflow.</p>
6672 <pre>
6673 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
6674 %sum = extractvalue {i32, i1} %res, 0
6675 %obit = extractvalue {i32, i1} %res, 1
6676 br i1 %obit, label %overflow, label %normal
6677 </pre>
6679 </div>
6681 <!-- _______________________________________________________________________ -->
6684 </div>
6690 on any integer bit width.</p>
6692 <pre>
6693 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
6694 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
6695 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
6696 </pre>
6701 a signed multiplication of the two arguments, and indicate whether an
6702 overflow occurred during the signed multiplication.</p>
6705 <p>The arguments (%a and %b) and the first element of the result structure may
6706 be of integer types of any bit width, but they must have the same bit
6707 width. The second element of the result structure must be of
6709 undergo signed multiplication.</p>
6713 a signed multiplication of the two arguments. They return a structure —
6714 the first element of which is the multiplication, and the second element of
6715 which is a bit specifying if the signed multiplication resulted in an
6716 overflow.</p>
6719 <pre>
6720 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
6721 %sum = extractvalue {i32, i1} %res, 0
6722 %obit = extractvalue {i32, i1} %res, 1
6723 br i1 %obit, label %overflow, label %normal
6724 </pre>
6726 </div>
6728 <!-- _______________________________________________________________________ -->
6731 </div>
6737 on any integer bit width.</p>
6739 <pre>
6740 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
6741 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
6742 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
6743 </pre>
6747 a unsigned multiplication of the two arguments, and indicate whether an
6748 overflow occurred during the unsigned multiplication.</p>
6751 <p>The arguments (%a and %b) and the first element of the result structure may
6752 be of integer types of any bit width, but they must have the same bit
6753 width. The second element of the result structure must be of
6755 undergo unsigned multiplication.</p>
6759 an unsigned multiplication of the two arguments. They return a structure
6760 — the first element of which is the multiplication, and the second
6761 element of which is a bit specifying if the unsigned multiplication resulted
6762 in an overflow.</p>
6765 <pre>
6766 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
6767 %sum = extractvalue {i32, i1} %res, 0
6768 %obit = extractvalue {i32, i1} %res, 1
6769 br i1 %obit, label %overflow, label %normal
6770 </pre>
6772 </div>
6774 <!-- ======================================================================= -->
6777 </div>
6781 <p>Half precision floating point is a storage-only format. This means that it is
6782 a dense encoding (in memory) but does not support computation in the
6783 format.</p>
6785 <p>This means that code must first load the half-precision floating point
6786 value as an i16, then convert it to float with <a
6788 Computation can then be performed on the float value (including extending to
6789 double etc). To store the value back to memory, it is first converted to
6790 float if needed, then converted to i16 with
6792 storing as an i16 value.</p>
6793 </div>
6795 <!-- _______________________________________________________________________ -->
6798 </div>
6803 <pre>
6804 declare i16 @llvm.convert.to.fp16(f32 %a)
6805 </pre>
6809 a conversion from single precision floating point format to half precision
6810 floating point format.</p>
6813 <p>The intrinsic function contains single argument - the value to be
6814 converted.</p>
6818 a conversion from single precision floating point format to half precision
6820 contains the converted number.</p>
6823 <pre>
6824 %res = call i16 @llvm.convert.to.fp16(f32 %a)
6825 store i16 %res, i16* @x, align 2
6826 </pre>
6828 </div>
6830 <!-- _______________________________________________________________________ -->
6833 </div>
6838 <pre>
6839 declare f32 @llvm.convert.from.fp16(i16 %a)
6840 </pre>
6844 a conversion from half precision floating point format to single precision
6845 floating point format.</p>
6848 <p>The intrinsic function contains single argument - the value to be
6849 converted.</p>
6853 conversion from half single precision floating point format to single
6854 precision floating point format. The input half-float value is represented by
6858 <pre>
6859 %a = load i16* @x, align 2
6860 %res = call f32 @llvm.convert.from.fp16(i16 %a)
6861 </pre>
6863 </div>
6865 <!-- ======================================================================= -->
6868 </div>
6873 prefix), are described in
6877 </div>
6879 <!-- ======================================================================= -->
6882 </div>
6886 <p>The LLVM exception handling intrinsics (which all start with
6891 </div>
6893 <!-- ======================================================================= -->
6896 </div>
6900 <p>This intrinsic makes it possible to excise one parameter, marked with
6902 The result is a callable
6903 function pointer lacking the nest parameter - the caller does not need to
6904 provide a value for it. Instead, the value to use is stored in advance in a
6905 "trampoline", a block of memory usually allocated on the stack, which also
6906 contains code to splice the nest value into the argument list. This is used
6907 to implement the GCC nested function address extension.</p>
6909 <p>For example, if the function is
6912 follows:</p>
6917 %p = call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8* nest , i32, i32)* @f to i8*), i8* %nval)
6918 %fp = bitcast i8* %p to i32 (i32, i32)*
6919 </pre>
6924 </div>
6926 <!-- _______________________________________________________________________ -->
6929 </div>
6934 <pre>
6936 </pre>
6940 function pointer suitable for executing it.</p>
6945 sufficiently aligned block of memory; this memory is written to by the
6946 intrinsic. Note that the size and the alignment are target-specific - LLVM
6947 currently provides no portable way of determining them, so a front-end that
6948 generates this intrinsic needs to have some target-specific knowledge.
6954 dependent code, turning it into a function. A pointer to this function is
6956 function pointer type</a> before being called. The new function's signature
6959 is allowed, and it must be of pointer type. Calling the new function is
6964 returned function pointer is undefined.</p>
6966 </div>
6968 <!-- ======================================================================= -->
6971 </div>
6976 hardware constructs for atomic operations and memory synchronization. This
6977 provides an interface to the hardware, not an interface to the programmer. It
6978 is aimed at a low enough level to allow any programming models or APIs
6979 (Application Programming Interfaces) which need atomic behaviors to map
6980 cleanly onto it. It is also modeled primarily on hardware behavior. Just as
6981 hardware provides a "universal IR" for source languages, it also provides a
6982 starting point for developing a "universal" atomic operation and
6983 synchronization IR.</p>
6986 software transaction memory systems, atomic primitives, and intrinsic
6987 functions as found in BSD, GNU libc, atomic_ops, APR, and other system and
6988 application libraries. The hardware interface provided by LLVM should allow
6989 a clean implementation of all of these APIs and parallel programming models.
6990 No one model or paradigm should be selected above others unless the hardware
6991 itself ubiquitously does so.</p>
6993 </div>
6995 <!-- _______________________________________________________________________ -->
6998 </div>
7001 <pre>
7002 declare void @llvm.memory.barrier(i1 <ll>, i1 <ls>, i1 <sl>, i1 <ss>, i1 <device>)
7003 </pre>
7007 specific pairs of memory access types.</p>
7011 The first four arguments enables a specific barrier as listed below. The
7012 fifth argument specifies that the barrier applies to io or device or uncached
7013 memory.</p>
7015 <ul>
7021 </ul>
7024 <p>This intrinsic causes the system to enforce some ordering constraints upon
7025 the loads and stores of the program. This barrier does not
7028 and store operations (f.ex. load-load, or store-load), all of the first
7029 operations preceding the barrier will complete before any of the second
7030 operations succeeding the barrier begin. Specifically the semantics for each
7031 pairing is as follows:</p>
7033 <ul>
7035 after the barrier begins.</li>
7037 store after the barrier begins.</li>
7039 store after the barrier begins.</li>
7041 load after the barrier begins.</li>
7042 </ul>
7045 is enabled in a single memory barrier intrinsic.</p>
7047 <p>Backends may implement stronger barriers than those requested when they do
7048 not support as fine grained a barrier as requested. Some architectures do
7049 not need all types of barriers and on such architectures, these become
7050 noops.</p>
7053 <pre>
7054 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7055 %ptr = bitcast i8* %mallocP to i32*
7056 store i32 4, %ptr
7059 call void @llvm.memory.barrier(i1 false, i1 true, i1 false, i1 false)
7062 </pre>
7064 </div>
7066 <!-- _______________________________________________________________________ -->
7069 </div>
7075 any integer bit width and for different address spaces. Not all targets
7076 support all bit widths however.</p>
7078 <pre>
7080 declare i16 @llvm.atomic.cmp.swap.i16.p0i16(i16* <ptr>, i16 <cmp>, i16 <val>)
7081 declare i32 @llvm.atomic.cmp.swap.i32.p0i32(i32* <ptr>, i32 <cmp>, i32 <val>)
7082 declare i64 @llvm.atomic.cmp.swap.i64.p0i64(i64* <ptr>, i64 <cmp>, i64 <val>)
7083 </pre>
7086 <p>This loads a value in memory and compares it to a given value. If they are
7087 equal, it stores a new value into the memory.</p>
7093 this integer type. While any bit width integer may be used, targets may only
7094 lower representations they support in hardware.</p>
7097 <p>This entire intrinsic must be executed atomically. It first loads the value
7100 memory. The loaded value is yielded in all cases. This provides the
7101 equivalent of an atomic compare-and-swap operation within the SSA
7102 framework.</p>
7105 <pre>
7106 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7107 %ptr = bitcast i8* %mallocP to i32*
7108 store i32 4, %ptr
7111 %result1 = call i32 @llvm.atomic.cmp.swap.i32.p0i32(i32* %ptr, i32 4, %val1)
7117 %result2 = call i32 @llvm.atomic.cmp.swap.i32.p0i32(i32* %ptr, i32 5, %val2)
7122 </pre>
7124 </div>
7126 <!-- _______________________________________________________________________ -->
7129 </div>
7134 integer bit width. Not all targets support all bit widths however.</p>
7136 <pre>
7141 </pre>
7152 integer type. The targets may only lower integer representations they
7153 support.</p>
7158 equivalent of an atomic swap operation within the SSA framework.</p>
7161 <pre>
7162 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7163 %ptr = bitcast i8* %mallocP to i32*
7164 store i32 4, %ptr
7167 %result1 = call i32 @llvm.atomic.swap.i32.p0i32(i32* %ptr, i32 %val1)
7173 %result2 = call i32 @llvm.atomic.swap.i32.p0i32(i32* %ptr, i32 %val2)
7178 </pre>
7180 </div>
7182 <!-- _______________________________________________________________________ -->
7186 </div>
7192 any integer bit width. Not all targets support all bit widths however.</p>
7194 <pre>
7199 </pre>
7206 <p>The intrinsic takes two arguments, the first a pointer to an integer value
7207 and the second an integer value. The result is also an integer value. These
7208 integer types can have any bit width, but they must all have the same bit
7209 width. The targets may only lower integer representations they support.</p>
7212 <p>This intrinsic does a series of operations atomically. It first loads the
7217 <pre>
7218 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7219 %ptr = bitcast i8* %mallocP to i32*
7220 store i32 4, %ptr
7221 %result1 = call i32 @llvm.atomic.load.add.i32.p0i32(i32* %ptr, i32 4)
7223 %result2 = call i32 @llvm.atomic.load.add.i32.p0i32(i32* %ptr, i32 2)
7225 %result3 = call i32 @llvm.atomic.load.add.i32.p0i32(i32* %ptr, i32 5)
7228 </pre>
7230 </div>
7232 <!-- _______________________________________________________________________ -->
7236 </div>
7242 any integer bit width and for different address spaces. Not all targets
7243 support all bit widths however.</p>
7245 <pre>
7250 </pre>
7257 <p>The intrinsic takes two arguments, the first a pointer to an integer value
7258 and the second an integer value. The result is also an integer value. These
7259 integer types can have any bit width, but they must all have the same bit
7260 width. The targets may only lower integer representations they support.</p>
7263 <p>This intrinsic does a series of operations atomically. It first loads the
7269 <pre>
7270 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7271 %ptr = bitcast i8* %mallocP to i32*
7272 store i32 8, %ptr
7273 %result1 = call i32 @llvm.atomic.load.sub.i32.p0i32(i32* %ptr, i32 4)
7275 %result2 = call i32 @llvm.atomic.load.sub.i32.p0i32(i32* %ptr, i32 2)
7277 %result3 = call i32 @llvm.atomic.load.sub.i32.p0i32(i32* %ptr, i32 5)
7280 </pre>
7282 </div>
7284 <!-- _______________________________________________________________________ -->
7290 </div>
7295 <p>These are overloaded intrinsics. You can
7298 bit width and for different address spaces. Not all targets support all bit
7299 widths however.</p>
7301 <pre>
7306 </pre>
7308 <pre>
7313 </pre>
7315 <pre>
7320 </pre>
7322 <pre>
7327 </pre>
7335 <p>These intrinsics take two arguments, the first a pointer to an integer value
7336 and the second an integer value. The result is also an integer value. These
7337 integer types can have any bit width, but they must all have the same bit
7338 width. The targets may only lower integer representations they support.</p>
7341 <p>These intrinsics does a series of operations atomically. They first load the
7347 <pre>
7348 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7349 %ptr = bitcast i8* %mallocP to i32*
7350 store i32 0x0F0F, %ptr
7351 %result0 = call i32 @llvm.atomic.load.nand.i32.p0i32(i32* %ptr, i32 0xFF)
7353 %result1 = call i32 @llvm.atomic.load.and.i32.p0i32(i32* %ptr, i32 0xFF)
7355 %result2 = call i32 @llvm.atomic.load.or.i32.p0i32(i32* %ptr, i32 0F)
7357 %result3 = call i32 @llvm.atomic.load.xor.i32.p0i32(i32* %ptr, i32 0F)
7360 </pre>
7362 </div>
7364 <!-- _______________________________________________________________________ -->
7370 </div>
7378 address spaces. Not all targets support all bit widths however.</p>
7380 <pre>
7385 </pre>
7387 <pre>
7392 </pre>
7394 <pre>
7399 </pre>
7401 <pre>
7406 </pre>
7409 <p>These intrinsics takes the signed or unsigned minimum or maximum of
7414 <p>These intrinsics take two arguments, the first a pointer to an integer value
7415 and the second an integer value. The result is also an integer value. These
7416 integer types can have any bit width, but they must all have the same bit
7417 width. The targets may only lower integer representations they support.</p>
7420 <p>These intrinsics does a series of operations atomically. They first load the
7426 <pre>
7427 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7428 %ptr = bitcast i8* %mallocP to i32*
7429 store i32 7, %ptr
7430 %result0 = call i32 @llvm.atomic.load.min.i32.p0i32(i32* %ptr, i32 -2)
7432 %result1 = call i32 @llvm.atomic.load.max.i32.p0i32(i32* %ptr, i32 8)
7434 %result2 = call i32 @llvm.atomic.load.umin.i32.p0i32(i32* %ptr, i32 10)
7436 %result3 = call i32 @llvm.atomic.load.umax.i32.p0i32(i32* %ptr, i32 30)
7439 </pre>
7441 </div>
7444 <!-- ======================================================================= -->
7447 </div>
7451 <p>This class of intrinsics exists to information about the lifetime of memory
7452 objects and ranges where variables are immutable.</p>
7454 </div>
7456 <!-- _______________________________________________________________________ -->
7459 </div>
7464 <pre>
7466 </pre>
7470 object's lifetime.</p>
7473 <p>The first argument is a constant integer representing the size of the
7474 object, or -1 if it is variable sized. The second argument is a pointer to
7475 the object.</p>
7478 <p>This intrinsic indicates that before this point in the code, the value of the
7480 never be used and has an undefined value. A load from the pointer that
7481 precedes this intrinsic can be replaced with
7484 </div>
7486 <!-- _______________________________________________________________________ -->
7489 </div>
7494 <pre>
7496 </pre>
7500 object's lifetime.</p>
7503 <p>The first argument is a constant integer representing the size of the
7504 object, or -1 if it is variable sized. The second argument is a pointer to
7505 the object.</p>
7508 <p>This intrinsic indicates that after this point in the code, the value of the
7510 never be used and has an undefined value. Any stores into the memory object
7511 following this intrinsic may be removed as dead.
7513 </div>
7515 <!-- _______________________________________________________________________ -->
7518 </div>
7523 <pre>
7525 </pre>
7529 a memory object will not change.</p>
7532 <p>The first argument is a constant integer representing the size of the
7533 object, or -1 if it is variable sized. The second argument is a pointer to
7534 the object.</p>
7538 the return value, the referenced memory location is constant and
7539 unchanging.</p>
7541 </div>
7543 <!-- _______________________________________________________________________ -->
7546 </div>
7551 <pre>
7552 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
7553 </pre>
7557 a memory object are mutable.</p>
7561 The second argument is a constant integer representing the size of the
7562 object, or -1 if it is variable sized and the third argument is a pointer
7563 to the object.</p>
7568 </div>
7570 <!-- ======================================================================= -->
7573 </div>
7577 <p>This class of intrinsics is designed to be generic and has no specific
7578 purpose.</p>
7580 </div>
7582 <!-- _______________________________________________________________________ -->
7585 </div>
7590 <pre>
7591 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
7592 </pre>
7598 <p>The first argument is a pointer to a value, the second is a pointer to a
7599 global string, the third is a pointer to a global string which is the source
7600 file name, and the last argument is the line number.</p>
7603 <p>This intrinsic allows annotation of local variables with arbitrary strings.
7604 This can be useful for special purpose optimizations that want to look for
7605 these annotations. These have no other defined use, they are ignored by code
7606 generation and optimization.</p>
7608 </div>
7610 <!-- _______________________________________________________________________ -->
7613 </div>
7619 any integer bit width.</p>
7621 <pre>
7622 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
7623 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
7624 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
7625 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
7626 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
7627 </pre>
7633 <p>The first argument is an integer value (result of some expression), the
7634 second is a pointer to a global string, the third is a pointer to a global
7635 string which is the source file name, and the last argument is the line
7636 number. It returns the value of the first argument.</p>
7639 <p>This intrinsic allows annotations to be put on arbitrary expressions with
7640 arbitrary strings. This can be useful for special purpose optimizations that
7641 want to look for these annotations. These have no other defined use, they
7642 are ignored by code generation and optimization.</p>
7644 </div>
7646 <!-- _______________________________________________________________________ -->
7649 </div>
7654 <pre>
7655 declare void @llvm.trap()
7656 </pre>
7665 <p>This intrinsics is lowered to the target dependent trap instruction. If the
7666 target does not have a trap instruction, this intrinsic will be lowered to
7669 </div>
7671 <!-- _______________________________________________________________________ -->
7674 </div>
7679 <pre>
7681 </pre>
7686 ensure that it is placed on the stack before local variables.</p>
7690 arguments. The first argument is the value loaded from the stack
7692 that has enough space to hold the value of the guard.</p>
7695 <p>This intrinsic causes the prologue/epilogue inserter to force the position of
7697 stack. This is to ensure that if a local variable on the stack is
7698 overwritten, it will destroy the value of the guard. When the function exits,
7699 the guard on the stack is checked against the original guard. If they're
7701 function.</p>
7703 </div>
7705 <!-- _______________________________________________________________________ -->
7708 </div>
7713 <pre>
7716 </pre>
7720 to the optimizers to discover at compile time either a) when an
7721 operation like memcpy will either overflow a buffer that corresponds to
7722 an object, or b) to determine that a runtime check for overflow isn't
7723 necessary. An object in this context means an allocation of a
7724 specific class, structure, array, or other object.</p>
7737 at compile time.</p>
7739 </div>
7741 <!-- *********************************************************************** -->
7742 <hr>
7743 <address>
7751 Last modified: $Date$
7752 </address>
7754 </body>
7755 </html>