| blob | 8c5ba804c745762e126811d8ff5d401806fb6356 |
3 <html>
4 <head>
11 </head>
13 <body>
16 <ol>
21 <ol>
24 <ol>
28 <li><a href="#linkage_linker_private_weak_def_auto">'<tt>linker_private_weak_def_auto</tt>' Linkage</a></li>
41 </ol>
42 </li>
56 </ol>
57 </li>
59 <ol>
62 <ol>
69 </ol>
70 </li>
72 <ol>
74 <ol>
79 </ol>
80 </li>
84 </ol>
85 </li>
87 </ol>
88 </li>
90 <ol>
98 </ol>
99 </li>
101 <ol>
104 </ol>
105 </li>
107 <ol>
110 Global Variable</a></li>
112 Global Variable</a></li>
114 Global Variable</a></li>
115 </ol>
116 </li>
118 <ol>
120 <ol>
128 </ol>
129 </li>
131 <ol>
144 </ol>
145 </li>
147 <ol>
154 </ol>
155 </li>
157 <ol>
161 </ol>
162 </li>
164 <ol>
167 </ol>
168 </li>
170 <ol>
175 </ol>
176 </li>
178 <ol>
191 </ol>
192 </li>
194 <ol>
201 </ol>
202 </li>
203 </ol>
204 </li>
206 <ol>
208 <ol>
212 </ol>
213 </li>
215 <ol>
219 </ol>
220 </li>
222 <ol>
230 </ol>
231 </li>
233 <ol>
242 </ol>
243 </li>
245 <ol>
250 </ol>
251 </li>
253 <ol>
260 </ol>
261 </li>
263 <ol>
266 </ol>
267 </li>
271 <ol>
273 </ol>
274 </li>
276 <ol>
290 </ol>
291 </li>
293 <ol>
298 </ol>
299 </li>
301 <ol>
312 </ol>
313 </li>
314 </ol>
315 </li>
316 </ol>
321 </div>
323 <!-- *********************************************************************** -->
325 <!-- *********************************************************************** -->
329 <p>This document is a reference manual for the LLVM assembly language. LLVM is
330 a Static Single Assignment (SSA) based representation that provides type
331 safety, low-level operations, flexibility, and the capability of representing
332 'all' high-level languages cleanly. It is the common code representation
333 used throughout all phases of the LLVM compilation strategy.</p>
335 </div>
337 <!-- *********************************************************************** -->
339 <!-- *********************************************************************** -->
343 <p>The LLVM code representation is designed to be used in three different forms:
344 as an in-memory compiler IR, as an on-disk bitcode representation (suitable
345 for fast loading by a Just-In-Time compiler), and as a human readable
346 assembly language representation. This allows LLVM to provide a powerful
347 intermediate representation for efficient compiler transformations and
348 analysis, while providing a natural means to debug and visualize the
349 transformations. The three different forms of LLVM are all equivalent. This
350 document describes the human readable representation and notation.</p>
352 <p>The LLVM representation aims to be light-weight and low-level while being
353 expressive, typed, and extensible at the same time. It aims to be a
354 "universal IR" of sorts, by being at a low enough level that high-level ideas
355 may be cleanly mapped to it (similar to how microprocessors are "universal
356 IR's", allowing many source languages to be mapped to them). By providing
357 type information, LLVM can be used as the target of optimizations: for
358 example, through pointer analysis, it can be proven that a C automatic
359 variable is never accessed outside of the current function, allowing it to
360 be promoted to a simple SSA value instead of a memory location.</p>
362 </div>
364 <!-- _______________________________________________________________________ -->
369 <p>It is important to note that this document describes 'well formed' LLVM
370 assembly language. There is a difference between what the parser accepts and
371 what is considered 'well formed'. For example, the following instruction is
372 syntactically okay, but not well formed:</p>
376 </pre>
379 LLVM infrastructure provides a verification pass that may be used to verify
380 that an LLVM module is well formed. This pass is automatically run by the
381 parser after parsing input assembly and by the optimizer before it outputs
382 bitcode. The violations pointed out by the verifier pass indicate bugs in
383 transformation passes or input to the parser.</p>
385 </div>
387 <!-- Describe the typesetting conventions here. -->
389 <!-- *********************************************************************** -->
391 <!-- *********************************************************************** -->
395 <p>LLVM identifiers come in two basic types: global and local. Global
397 character. Local identifiers (register names, types) begin with
399 for identifiers, for different purposes:</p>
401 <ol>
402 <li>Named values are represented as a string of characters with their prefix.
406 other characters in their names can be surrounded with quotes. Special
408 ASCII code for the character in hexadecimal. In this way, any character
409 can be used in a name value, even quotes themselves.</li>
411 <li>Unnamed values are represented as an unsigned numeric value with their
416 </ol>
418 <p>LLVM requires that values start with a prefix for two reasons: Compilers
419 don't need to worry about name clashes with reserved words, and the set of
420 reserved words may be expanded in the future without penalty. Additionally,
421 unnamed identifiers allow a compiler to quickly come up with a temporary
422 variable without having to avoid symbol table conflicts.</p>
424 <p>Reserved words in LLVM are very similar to reserved words in other
425 languages. There are keywords for different opcodes
431 reserved words cannot conflict with variable names, because none of them
434 <p>Here is an example of LLVM code to multiply the integer variable
441 </pre>
447 </pre>
455 </pre>
458 lexical features of LLVM:</p>
460 <ol>
462 line.</li>
464 <li>Unnamed temporaries are created when the result of a computation is not
465 assigned to a named value.</li>
468 </ol>
470 <p>It also shows a convention that we follow in this document. When
471 demonstrating instructions, we will follow an instruction with a comment that
472 defines the type and name of value produced. Comments are shown in italic
473 text.</p>
475 </div>
477 <!-- *********************************************************************** -->
479 <!-- *********************************************************************** -->
481 <!-- ======================================================================= -->
483 </div>
488 of the input programs. Each module consists of functions, global variables,
489 and symbol table entries. Modules may be combined together with the LLVM
490 linker, which merges function (and global variable) definitions, resolves
491 forward declarations, and merges symbol table entries. Here is an example of
496 <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>
504 %cast210 = <a href="#i_getelementptr">getelementptr</a> [13 x i8]* @.LC0, i64 0, i64 0 <i>; i8*</i>
509 }
513 !foo = !{!1, null}
514 </pre>
522 <p>In general, a module is made up of a list of global values, where both
523 functions and global variables are global values. Global values are
524 represented by a pointer to a memory location (in this case, a pointer to an
525 array of char, and a pointer to a function), and have one of the
528 </div>
530 <!-- ======================================================================= -->
533 </div>
537 <p>All Global Variables and Functions have one of the following types of
538 linkage:</p>
540 <dl>
543 by objects in the current module. In particular, linking code into a
544 module with an private global value may cause the private to be renamed as
545 necessary to avoid collisions. Because the symbol is private to the
546 module, all references can be updated. This doesn't show up in any symbol
547 table in the object file.</dd>
551 assembler and evaluated by the linker. Unlike normal strong symbols, they
552 are removed by the linker from the final linked image (executable or
553 dynamic library).</dd>
558 linker. The symbols are removed by the linker from the final linked image
559 (executable or dynamic library).</dd>
561 <dt><tt><b><a name="linkage_linker_private_weak_def_auto">linker_private_weak_def_auto</a></b></tt></dt>
563 of the object is not taken. For instance, functions that had an inline
564 definition, but the compiler decided not to inline it. Note,
567 visibility. The symbols are removed by the linker from the final linked
568 image (executable or dynamic library).</dd>
571 <dd>Similar to private, but the value shows as a local symbol
577 into the object file corresponding to the LLVM module. They exist to
578 allow inlining and other optimizations to take place given knowledge of
579 the definition of the global, which is known to be somewhere outside the
582 This linkage type is only allowed on definitions, not declarations.</dd>
586 the same name when linkage occurs. This can be used to implement
587 some forms of inline functions, templates, or other code which must be
588 generated in each translation unit that uses it, but where the body may
589 be overridden with a more definitive definition later. Unreferenced
592 inline the body of this function into callers because it doesn't know if
593 this definition of the function is the definitive definition within the
594 program or whether it will be overridden by a stronger definition.
595 To enable inlining and other optimizations, use "<tt>linkonce_odr</tt>"
596 linkage.</dd>
606 they are used for tentative definitions in C, such as "<tt>int X;</tt>" at
607 global scope.
608 Symbols with "<tt>common</tt>" linkage are merged in the same way as
611 must have a zero initializer, and may not be marked '<a
613 have common linkage.</dd>
618 pointer to array type. When two global variables with appending linkage
619 are linked together, the two global arrays are appended together. This is
620 the LLVM, typesafe, equivalent of having the system linker append together
624 <dd>The semantics of this linkage follow the ELF object file model: the symbol
625 is weak until linked, if not linked, the symbol becomes null instead of
626 being an undefined reference.</dd>
630 <dd>Some languages allow differing globals to be merged, such as two functions
632 that only equivalent globals are ever merged (the "one definition rule"
635 be merged with equivalent globals. These linkage types are otherwise the
639 <dd>If none of the above identifiers are used, the global is externally
640 visible, meaning that it participates in linkage and can be used to
641 resolve external symbol references.</dd>
642 </dl>
644 <p>The next two types of linkage are targeted for Microsoft Windows platform
645 only. They are designed to support importing (exporting) symbols from (to)
646 DLLs (Dynamic Link Libraries).</p>
648 <dl>
651 or variable via a global pointer to a pointer that is set up by the DLL
652 exporting the symbol. On Microsoft Windows targets, the pointer name is
654 name.</dd>
658 pointer to a pointer in a DLL, so that it can be referenced with the
661 variable name.</dd>
662 </dl>
665 another module defined a "<tt>.LC0</tt>" variable and was linked with this
666 one, one of the two would be renamed, preventing a collision. Since
668 declarations), they are accessible outside of the current module.</p>
677 </div>
679 <!-- ======================================================================= -->
682 </div>
688 convention specified for the call. The calling convention of any pair of
689 dynamic caller/callee must match, or the behavior of the program is
690 undefined. The following calling conventions are supported by LLVM, and more
691 may be added in the future:</p>
693 <dl>
695 <dd>This calling convention (the default if no other calling convention is
696 specified) matches the target C calling conventions. This calling
697 convention supports varargs function calls and tolerates some mismatch in
698 the declared prototype and implemented declaration of the function (as
699 does normal C).</dd>
702 <dd>This calling convention attempts to make calls as fast as possible
703 (e.g. by passing things in registers). This calling convention allows the
704 target to use whatever tricks it wants to produce fast code for the
705 target, without having to conform to an externally specified ABI
706 (Application Binary Interface).
708 when this or the GHC convention is used.</a> This calling convention
709 does not support varargs and requires the prototype of all callees to
710 exactly match the prototype of the function definition.</dd>
713 <dd>This calling convention attempts to make code in the caller as efficient
714 as possible under the assumption that the call is not commonly executed.
715 As such, these calls often preserve all registers so that the call does
716 not break any live ranges in the caller side. This calling convention
717 does not support varargs and requires the prototype of all callees to
718 exactly match the prototype of the function definition.</dd>
721 <dd>This calling convention has been implemented specifically for use by the
723 It passes everything in registers, going to extremes to achieve this by
724 disabling callee save registers. This calling convention should not be
725 used lightly but only for specific situations such as an alternative to
727 implementing functional programming languages.At the moment only X86
728 supports this convention and it has the following limitations:
729 <ul>
731 floating point types are supported.</li>
734 </ul>
735 This calling convention supports
737 requires both the caller and callee are using it.
738 </dd>
741 <dd>Any calling convention may be specified by number, allowing
742 target-specific calling conventions to be used. Target specific calling
744 </dl>
746 <p>More calling conventions can be added/defined on an as-needed basis, to
747 support Pascal conventions or any other well-known target-independent
748 convention.</p>
750 </div>
752 <!-- ======================================================================= -->
755 </div>
759 <p>All Global Variables and Functions have one of the following visibility
760 styles:</p>
762 <dl>
764 <dd>On targets that use the ELF object file format, default visibility means
765 that the declaration is visible to other modules and, in shared libraries,
766 means that the declared entity may be overridden. On Darwin, default
767 visibility means that the declaration is visible to other modules. Default
771 <dd>Two declarations of an object with hidden visibility refer to the same
772 object if they are in the same shared object. Usually, hidden visibility
773 indicates that the symbol will not be placed into the dynamic symbol
774 table, so no other module (executable or shared library) can reference it
775 directly.</dd>
778 <dd>On ELF, protected visibility indicates that the symbol will be placed in
779 the dynamic symbol table, but that references within the defining module
780 will bind to the local symbol. That is, the symbol cannot be overridden by
781 another module.</dd>
782 </dl>
784 </div>
786 <!-- ======================================================================= -->
789 </div>
793 <p>LLVM IR allows you to specify name aliases for certain types. This can make
794 it easier to read the IR and make the IR more condensed (particularly when
795 recursive types are involved). An example of a name specification is:</p>
798 %mytype = type { %mytype*, i32 }
799 </pre>
805 <p>Note that type names are aliases for the structural type that they indicate,
806 and that you can therefore specify multiple names for the same type. This
807 often leads to confusing behavior when dumping out a .ll file. Since LLVM IR
808 uses structural typing, the name is not part of the type. When printing out
810 particular shape. This means that if you have code where two different
811 source types end up having the same LLVM type, that the dumper will sometimes
812 print the "wrong" or unexpected type. This is an important design point and
813 isn't going to change.</p>
815 </div>
817 <!-- ======================================================================= -->
820 </div>
824 <p>Global variables define regions of memory allocated at compilation time
825 instead of run-time. Global variables may optionally be initialized, may
826 have an explicit section to be placed in, and may have an optional explicit
827 alignment specified. A variable may be defined as "thread_local", which
828 means that it will not be shared by threads (each thread will have a
829 separated copy of the variable). A variable may be defined as a global
830 "constant," which indicates that the contents of the variable
832 global data to be placed in the read-only section of an executable, etc).
833 Note that variables that need runtime initialization cannot be marked
837 constant, even if the final definition of the global is not. This capability
838 can be used to enable slightly better optimization of the program, but
839 requires the language definition to guarantee that optimizations based on the
840 'constantness' are valid for the translation units that do not include the
841 definition.</p>
843 <p>As SSA values, global variables define pointer values that are in scope
844 (i.e. they dominate) all basic blocks in the program. Global variables
845 always define a pointer to their "content" type because they describe a
846 region of memory, and all memory objects in LLVM are accessed through
847 pointers.</p>
850 that the address is not significant, only the content. Constants marked
851 like this can be merged with other constants if they have the same
854 constant whose address is significant.</p>
856 <p>A global variable may be declared to reside in a target-specific numbered
857 address space. For targets that support them, address spaces may affect how
858 optimizations are performed and/or what target instructions are used to
859 access the variable. The default address space is zero. The address space
860 qualifier must precede any other attributes.</p>
862 <p>LLVM allows an explicit section to be specified for globals. If the target
863 supports it, it will emit globals to the section specified.</p>
865 <p>An explicit alignment may be specified for a global, which must be a power
866 of 2. If not present, or if the alignment is set to zero, the alignment of
867 the global is set by the target to whatever it feels convenient. If an
868 explicit alignment is specified, the global is forced to have exactly that
869 alignment. Targets and optimizers are not allowed to over-align the global
870 if the global has an assigned section. In this case, the extra alignment
871 could be observable: for example, code could assume that the globals are
872 densely packed in their section and try to iterate over them as an array,
873 alignment padding would break this iteration.</p>
875 <p>For example, the following defines a global in a numbered address space with
876 an initializer, section, and alignment:</p>
880 </pre>
882 </div>
885 <!-- ======================================================================= -->
888 </div>
898 name, a (possibly empty) argument list (each with optional
902 curly brace, a list of basic blocks, and a closing curly brace.</p>
910 name, a possibly empty list of arguments, an optional alignment, and an
913 <p>A function definition contains a list of basic blocks, forming the CFG
914 (Control Flow Graph) for the function. Each basic block may optionally start
915 with a label (giving the basic block a symbol table entry), contains a list
917 instruction (such as a branch or function return).</p>
919 <p>The first basic block in a function is special in two ways: it is immediately
920 executed on entrance to the function, and it is not allowed to have
921 predecessor basic blocks (i.e. there can not be any branches to the entry
922 block of a function). Because the block can have no predecessors, it also
925 <p>LLVM allows an explicit section to be specified for functions. If the target
926 supports it, it will emit functions to the section specified.</p>
928 <p>An explicit alignment may be specified for a function. If not present, or if
929 the alignment is set to zero, the alignment of the function is set by the
930 target to whatever it feels convenient. If an explicit alignment is
931 specified, the function is forced to have at least that much alignment. All
935 be significant and two identical functions can be merged</p>.
944 </pre>
946 </div>
948 <!-- ======================================================================= -->
951 </div>
956 function, global variable, another alias or bitcast of global value). Aliases
963 </pre>
965 </div>
967 <!-- ======================================================================= -->
970 </div>
975 nodes</a> (but not metadata strings) are the only valid operands for
976 a named metadata.</p>
980 ; Some unnamed metadata nodes, which are referenced by the named metadata.
984 ; A named metadata.
986 </pre>
988 </div>
990 <!-- ======================================================================= -->
995 <p>The return type and each parameter of a function type may have a set of
997 used to communicate additional information about the result or parameters of
998 a function. Parameter attributes are considered to be part of the function,
999 not of the function type, so functions with different parameter attributes
1000 can have the same function type.</p>
1002 <p>Parameter attributes are simple keywords that follow the type specified. If
1003 multiple parameter attributes are needed, they are space separated. For
1004 example:</p>
1007 declare i32 @printf(i8* noalias nocapture, ...)
1008 declare i32 @atoi(i8 zeroext)
1009 declare signext i8 @returns_signed_char()
1010 </pre>
1017 <dl>
1019 <dd>This indicates to the code generator that the parameter or return value
1020 should be zero-extended to the extent required by the target's ABI (which
1022 parameter) or the callee (for a return value).</dd>
1025 <dd>This indicates to the code generator that the parameter or return value
1026 should be sign-extended to the extent required by the target's ABI (which
1027 is usually 32-bits) by the caller (for a parameter) or the callee (for a
1028 return value).</dd>
1031 <dd>This indicates that this parameter or return value should be treated in a
1032 special target-dependent fashion during while emitting code for a function
1033 call or return (usually, by putting it in a register as opposed to memory,
1034 though some targets use it to distinguish between two different kinds of
1035 registers). Use of this attribute is target-specific.</dd>
1038 <dd><p>This indicates that the pointer parameter should really be passed by
1039 value to the function. The attribute implies that a hidden copy of the
1040 pointee
1041 is made between the caller and the callee, so the callee is unable to
1042 modify the value in the callee. This attribute is only valid on LLVM
1043 pointer arguments. It is generally used to pass structs and arrays by
1044 value, but is also valid on pointers to scalars. The copy is considered
1045 to belong to the caller not the callee (for example,
1048 values.</p>
1050 <p>The byval attribute also supports specifying an alignment with
1051 the align attribute. It indicates the alignment of the stack slot to
1052 form and the known alignment of the pointer specified to the call site. If
1053 the alignment is not specified, then the code generator makes a
1054 target-specific assumption.</p></dd>
1057 <dd>This indicates that the pointer parameter specifies the address of a
1058 structure that is the return value of the function in the source program.
1059 This pointer must be guaranteed by the caller to be valid: loads and
1060 stores to the structure may be assumed by the callee to not to trap. This
1061 may only be applied to the first parameter. This is not a valid attribute
1062 for return values. </dd>
1065 <dd>This indicates that pointer values
1068 ignoring certain "irrelevant" dependencies.
1069 For a call to the parent function, dependencies between memory
1070 references from before or after the call and from those during the call
1072 return value used in that call.
1073 The caller shares the responsibility with the callee for ensuring that
1074 these requirements are met.
1075 For further details, please see the discussion of the NoAlias response in
1077 <br>
1080 arguments, though it is slightly weaker.
1081 <br>
1084 </dd>
1087 <dd>This indicates that the callee does not make any copies of the pointer
1088 that outlive the callee itself. This is not a valid attribute for return
1089 values.</dd>
1092 <dd>This indicates that the pointer parameter can be excised using the
1094 attribute for return values.</dd>
1095 </dl>
1097 </div>
1099 <!-- ======================================================================= -->
1102 </div>
1106 <p>Each function may specify a garbage collector name, which is simply a
1107 string:</p>
1110 define void @f() gc "name" { ... }
1111 </pre>
1114 collector which will cause the compiler to alter its output in order to
1115 support the named garbage collection algorithm.</p>
1117 </div>
1119 <!-- ======================================================================= -->
1122 </div>
1126 <p>Function attributes are set to communicate additional information about a
1127 function. Function attributes are considered to be part of the function, not
1128 of the function type, so functions with different parameter attributes can
1129 have the same function type.</p>
1131 <p>Function attributes are simple keywords that follow the type specified. If
1132 multiple attributes are needed, they are space separated. For example:</p>
1135 define void @f() noinline { ... }
1136 define void @f() alwaysinline { ... }
1137 define void @f() alwaysinline optsize { ... }
1138 define void @f() optsize { ... }
1139 </pre>
1141 <dl>
1143 <dd>This attribute indicates that, when emitting the prologue and epilogue,
1144 the backend should forcibly align the stack pointer. Specify the
1145 desired alignment, which must be a power of two, in parentheses.
1148 <dd>This attribute indicates that the inliner should attempt to inline this
1149 function into callers whenever possible, ignoring any active inlining size
1150 threshold for this caller.</dd>
1153 <dd>This attribute indicates that the function should be 'hotpatchable',
1154 meaning the function can be patched and/or hooked even while it is
1155 loaded into memory. On x86, the function prologue will be preceded
1156 by six bytes of padding and will begin with a two-byte instruction.
1157 Most of the functions in the Windows system DLLs in Windows XP SP2 or
1158 higher were compiled in this fashion.</dd>
1161 <dd>This attribute indicates that the source code contained a hint that inlining
1162 this function is desirable (such as the "inline" keyword in C/C++). It
1163 is just a hint; it imposes no requirements on the inliner.</dd>
1166 <dd>This attribute disables prologue / epilogue emission for the function.
1167 This can have very system-specific consequences.</dd>
1173 <dd>This attribute indicates that the inliner should never inline this
1174 function in any situation. This attribute may not be used together with
1178 <dd>This attribute indicates that the code generator should not use a red
1179 zone, even if the target-specific ABI normally permits it.</dd>
1182 <dd>This function attribute indicates that the function never returns
1183 normally. This produces undefined behavior at runtime if the function
1184 ever does dynamically return.</dd>
1187 <dd>This function attribute indicates that the function never returns with an
1188 unwind or exceptional control flow. If the function does unwind, its
1189 runtime behavior is undefined.</dd>
1192 <dd>This attribute suggests that optimization passes and code generator passes
1193 make choices that keep the code size of this function low, and otherwise
1194 do optimizations specifically to reduce code size.</dd>
1197 <dd>This attribute indicates that the function computes its result (or decides
1198 to unwind an exception) based strictly on its arguments, without
1199 dereferencing any pointer arguments or otherwise accessing any mutable
1200 state (e.g. memory, control registers, etc) visible to caller functions.
1201 It does not write through any pointer arguments
1203 changes any state visible to callers. This means that it cannot unwind
1208 <dd>This attribute indicates that the function does not write through any
1210 arguments) or otherwise modify any state (e.g. memory, control registers,
1211 etc) visible to caller functions. It may dereference pointer arguments
1212 and read state that may be set in the caller. A readonly function always
1213 returns the same value (or unwinds an exception identically) when called
1214 with the same set of arguments and global state. It cannot unwind an
1219 <dd>This attribute indicates that the function should emit a stack smashing
1221 the stack before the local variables that's checked upon return from the
1222 function to see if it has been overwritten. A heuristic is used to
1223 determine if a function needs stack protectors or not.<br>
1224 <br>
1231 stack smashing protector. This overrides
1233 <br>
1238 </dl>
1240 </div>
1242 <!-- ======================================================================= -->
1245 </div>
1250 the GCC "file scope inline asm" blocks. These blocks are internally
1251 concatenated by LLVM and treated as a single unit, but may be separated in
1255 module asm "inline asm code goes here"
1256 module asm "more can go here"
1257 </pre>
1259 <p>The strings can contain any character by escaping non-printable characters.
1261 for the number.</p>
1263 <p>The inline asm code is simply printed to the machine code .s file when
1264 assembly code is generated.</p>
1266 </div>
1268 <!-- ======================================================================= -->
1271 </div>
1275 <p>A module may specify a target specific data layout string that specifies how
1276 data is to be laid out in memory. The syntax for the data layout is
1277 simply:</p>
1280 target datalayout = "<i>layout specification</i>"
1281 </pre>
1284 separated by the minus sign character ('-'). Each specification starts with
1285 a letter and may include other information after the letter to define some
1286 aspect of the data layout. The specifications accepted are as follows:</p>
1288 <dl>
1290 <dd>Specifies that the target lays out data in big-endian form. That is, the
1291 bits with the most significance have the lowest address location.</dd>
1294 <dd>Specifies that the target lays out data in little-endian form. That is,
1295 the bits with the least significance have the lowest address
1296 location.</dd>
1305 <dd>This specifies the alignment for an integer type of a given bit
1309 <dd>This specifies the alignment for a vector type of a given bit
1313 <dd>This specifies the alignment for a floating point type of a given bit
1317 targets.
1320 <dd>This specifies the alignment for an aggregate type of a given bit
1324 <dd>This specifies the alignment for a stack object of a given bit
1328 <dd>This specifies a set of native integer widths for the target CPU
1331 this set are considered to support most general arithmetic
1332 operations efficiently.</dd>
1333 </dl>
1335 <p>When constructing the data layout for a given target, LLVM starts with a
1336 default set of specifications which are then (possibly) overridden by the
1338 are given in this list:</p>
1340 <ul>
1355 </ul>
1357 <p>When LLVM is determining the alignment for a given type, it uses the
1358 following rules:</p>
1360 <ol>
1361 <li>If the type sought is an exact match for one of the specifications, that
1362 specification is used.</li>
1364 <li>If no match is found, and the type sought is an integer type, then the
1365 smallest integer type that is larger than the bitwidth of the sought type
1366 is used. If none of the specifications are larger than the bitwidth then
1367 the the largest integer type is used. For example, given the default
1368 specifications above, the i7 type will use the alignment of i8 (next
1369 largest) while both i65 and i256 will use the alignment of i64 (largest
1370 specified).</li>
1372 <li>If no match is found, and the type sought is a vector type, then the
1373 largest vector type that is smaller than the sought vector type will be
1376 </ol>
1378 </div>
1380 <!-- ======================================================================= -->
1383 </div>
1387 <p>Any memory access must be done through a pointer value associated
1388 with an address range of the memory access, otherwise the behavior
1389 is undefined. Pointer values are associated with address ranges
1390 according to the following rules:</p>
1392 <ul>
1393 <li>A pointer value is associated with the addresses associated with
1395 <li>An address of a global variable is associated with the address
1396 range of the variable's storage.</li>
1397 <li>The result value of an allocation instruction is associated with
1398 the address range of the allocated storage.</li>
1399 <li>A null pointer in the default address-space is associated with
1400 no address.</li>
1401 <li>An integer constant other than zero or a pointer value returned
1402 from a function not defined within LLVM may be associated with address
1403 ranges allocated through mechanisms other than those provided by
1404 LLVM. Such ranges shall not overlap with any ranges of addresses
1405 allocated by mechanisms provided by LLVM.</li>
1406 </ul>
1409 to the following rules:</p>
1411 <ul>
1412 <li>A pointer value formed from a
1415 <li>The result value of a
1418 <li>A pointer value formed by an
1420 pointer values that contribute (directly or indirectly) to the
1421 computation of the pointer's value.</li>
1423 </ul>
1427 slightly weaker.</p>
1429 <p>LLVM IR does not associate types with memory. The result type of a
1431 alignment of the memory from which to load, as well as the
1432 interpretation of the value. The first operand type of a
1434 and alignment of the store.</p>
1436 <p>Consequently, type-based alias analysis, aka TBAA, aka
1439 additional information which specialized optimization passes may use
1440 to implement type-based alias analysis.</p>
1442 </div>
1444 <!-- ======================================================================= -->
1447 </div>
1454 The optimizers must not change the number of volatile operations or change their
1455 order of execution relative to other volatile operations. The optimizers
1457 operations. This is not Java's "volatile" and has no cross-thread
1458 synchronization behavior.</p>
1460 </div>
1462 <!-- *********************************************************************** -->
1464 <!-- *********************************************************************** -->
1468 <p>The LLVM type system is one of the most important features of the
1469 intermediate representation. Being typed enables a number of optimizations
1470 to be performed on the intermediate representation directly, without having
1471 to do extra analyses on the side before the transformation. A strong type
1472 system makes it easier to read the generated code and enables novel analyses
1473 and transformations that are not feasible to perform on normal three address
1474 code representations.</p>
1476 </div>
1478 <!-- ======================================================================= -->
1480 Classifications</a> </div>
1487 <tbody>
1489 <tr>
1492 </tr>
1493 <tr>
1496 </tr>
1497 <tr>
1507 </td>
1508 </tr>
1509 <tr>
1517 </tr>
1518 <tr>
1527 </td>
1528 </tr>
1529 </tbody>
1530 </table>
1533 important. Values of these types are the only ones which can be produced by
1534 instructions.</p>
1536 </div>
1538 <!-- ======================================================================= -->
1543 <p>The primitive types are the fundamental building blocks of the LLVM
1544 system.</p>
1546 </div>
1548 <!-- _______________________________________________________________________ -->
1554 <p>The integer type is a very simple type that simply specifies an arbitrary
1555 bit width for the integer type desired. Any bit width from 1 bit to
1559 <pre>
1560 iN
1561 </pre>
1564 value.</p>
1571 </tr>
1575 </tr>
1579 </tr>
1580 </table>
1582 </div>
1584 <!-- _______________________________________________________________________ -->
1589 <table>
1590 <tbody>
1597 </tbody>
1598 </table>
1600 </div>
1602 <!-- _______________________________________________________________________ -->
1608 <p>The x86mmx type represents a value held in an MMX register on an x86 machine. The operations allowed on it are quite limited: parameters and return values, load and store, and bitcast. User-specified MMX instructions are represented as intrinsic or asm calls with arguments and/or results of this type. There are no arrays, vectors or constants of this type.</p>
1611 <pre>
1612 x86mmx
1613 </pre>
1615 </div>
1617 <!-- _______________________________________________________________________ -->
1626 <pre>
1627 void
1628 </pre>
1630 </div>
1632 <!-- _______________________________________________________________________ -->
1641 <pre>
1642 label
1643 </pre>
1645 </div>
1647 <!-- _______________________________________________________________________ -->
1653 <p>The metadata type represents embedded metadata. No derived types may be
1655 arguments.
1658 <pre>
1659 metadata
1660 </pre>
1662 </div>
1665 <!-- ======================================================================= -->
1670 <p>The real power in LLVM comes from the derived types in the system. This is
1671 what allows a programmer to represent arrays, functions, pointers, and other
1672 useful types. Each of these types contain one or more element types which
1673 may be a primitive type, or another derived type. For example, it is
1674 possible to have a two dimensional array, using an array as the element type
1675 of another array.</p>
1678 </div>
1680 <!-- _______________________________________________________________________ -->
1685 <p>Aggregate Types are a subset of derived types that can contain multiple
1688 aggregate types.</p>
1690 </div>
1692 <!-- _______________________________________________________________________ -->
1698 <p>The array type is a very simple derived type that arranges elements
1699 sequentially in memory. The array type requires a size (number of elements)
1700 and an underlying data type.</p>
1703 <pre>
1705 </pre>
1708 be any type with a size.</p>
1715 </tr>
1719 </tr>
1723 </tr>
1724 </table>
1730 </tr>
1734 </tr>
1738 </tr>
1739 </table>
1741 <p>There is no restriction on indexing beyond the end of the array implied by
1742 a static type (though there are restrictions on indexing beyond the bounds
1743 of an allocated object in some cases). This means that single-dimension
1744 'variable sized array' addressing can be implemented in LLVM with a zero
1745 length array type. An implementation of 'pascal style arrays' in LLVM could
1748 </div>
1750 <!-- _______________________________________________________________________ -->
1756 <p>The function type can be thought of as a function signature. It consists of
1757 a return type and a list of formal parameter types. The return type of a
1758 function type is a first class type or a void type.</p>
1761 <pre>
1763 </pre>
1767 which indicates that the function takes a variable number of arguments.
1768 Variable argument functions can access their arguments with
1778 </td>
1781 </tt></td>
1785 </td>
1791 LLVM.
1792 </td>
1797 </td>
1798 </tr>
1799 </table>
1801 </div>
1803 <!-- _______________________________________________________________________ -->
1809 <p>The structure type is used to represent a collection of data members together
1810 in memory. The packing of the field types is defined to match the ABI of the
1811 underlying processor. The elements of a structure may be any type that has a
1812 size.</p>
1817 Structures in registers are accessed using the
1821 <pre>
1823 </pre>
1836 </tr>
1837 </table>
1839 </div>
1841 <!-- _______________________________________________________________________ -->
1843 </div>
1848 <p>The packed structure type is used to represent a collection of data members
1849 together in memory. There is no padding between fields. Further, the
1850 alignment of a packed structure is 1 byte. The elements of a packed
1851 structure may be any type that has a size.</p>
1858 <pre>
1860 </pre>
1874 </tr>
1875 </table>
1877 </div>
1879 <!-- _______________________________________________________________________ -->
1885 <p>The pointer type is used to specify memory locations.
1886 Pointers are commonly used to reference objects in memory.</p>
1888 <p>Pointer types may have an optional address space attribute defining the
1889 numbered address space where the pointed-to object resides. The default
1890 address space is number zero. The semantics of non-zero address
1891 spaces are target-specific.</p>
1897 <pre>
1899 </pre>
1907 </tr>
1913 </tr>
1918 </tr>
1919 </table>
1921 </div>
1923 <!-- _______________________________________________________________________ -->
1929 <p>A vector type is a simple derived type that represents a vector of elements.
1930 Vector types are used when multiple primitive data are operated in parallel
1931 using a single instruction (SIMD). A vector type requires a size (number of
1932 elements) and an underlying primitive data type. Vector types are considered
1936 <pre>
1938 </pre>
1941 may be any integer or floating point type. Vectors of size zero are not
1942 allowed, and pointers are not allowed as the element type.</p>
1949 </tr>
1953 </tr>
1957 </tr>
1958 </table>
1960 </div>
1962 <!-- _______________________________________________________________________ -->
1967 <p>Opaque types are used to represent unknown types in the system. This
1968 corresponds (for example) to the C notion of a forward declared structure
1969 type. In LLVM, opaque types can eventually be resolved to any type (not just
1970 a structure type).</p>
1973 <pre>
1974 opaque
1975 </pre>
1982 </tr>
1983 </table>
1985 </div>
1987 <!-- ======================================================================= -->
1990 </div>
1996 requiring it to have a name. For instance, a structure declaration may
1997 contain a pointer to any of the types it is lexically a member of. Example
1998 of up references (with their equivalent as named type declarations)
1999 include:</p>
2001 <pre>
2002 { \2 * } %x = type { %x* }
2003 { \2 }* %y = type { %y }*
2004 \1* %z = type %z*
2005 </pre>
2007 <p>An up reference is needed by the asmprinter for printing out cyclic types
2008 when there is no declared name for a type in the cycle. Because the
2009 asmprinter does not want to print out an infinite type string, it needs a
2010 syntax to handle recursive types that have no names (all names are optional
2011 in llvm IR).</p>
2014 <pre>
2016 </pre>
2025 </tr>
2029 structure.</td>
2030 </tr>
2031 </table>
2033 </div>
2035 <!-- *********************************************************************** -->
2037 <!-- *********************************************************************** -->
2041 <p>LLVM has several different basic types of constants. This section describes
2042 them all and their syntax.</p>
2044 </div>
2046 <!-- ======================================================================= -->
2051 <dl>
2059 with integer types.</dd>
2063 exponential notation (e.g. 1.23421e+2), or a more precise hexadecimal
2064 notation (see below). The assembler requires the exact decimal value of a
2065 floating-point constant. For example, the assembler accepts 1.25 but
2072 </dl>
2074 <p>The one non-intuitive notation for constants is the hexadecimal form of
2075 floating point constants. For example, the form '<tt>double
2078 constants are required (and the only time that they are generated by the
2079 disassembler) is when a floating point constant must be emitted but it cannot
2080 be represented as a decimal floating point number in a reasonable number of
2081 digits. For example, NaN's, infinities, and other special values are
2082 represented in their IEEE hexadecimal format so that assembly and disassembly
2083 do not cause any bits to change in the constants.</p>
2085 <p>When using the hexadecimal form, constants of types float and double are
2086 represented using the 16-digit form shown above (which matches the IEEE754
2087 representation for double); float values must, however, be exactly
2088 representable as IEE754 single precision. Hexadecimal format is always used
2089 for long double, and there are three forms of long double. The 80-bit format
2091 The 128-bit format used by PowerPC (two adjacent doubles) is represented
2094 currently supported target uses this format. Long doubles will only work if
2095 they match the long double format on your target. All hexadecimal formats
2096 are big-endian (sign bit at the left).</p>
2099 </div>
2101 <!-- ======================================================================= -->
2105 </div>
2109 <p>Complex constants are a (potentially recursive) combination of simple
2110 constants and smaller complex constants.</p>
2112 <dl>
2114 <dd>Structure constants are represented with notation similar to structure
2115 type definitions (a comma separated list of elements, surrounded by braces
2119 the number and types of elements must match those specified by the
2120 type.</dd>
2123 <dd>Array constants are represented with notation similar to array type
2124 definitions (a comma separated list of elements, surrounded by square
2127 the number and types of elements must match those specified by the
2128 type.</dd>
2131 <dd>Vector constants are represented with notation similar to vector type
2132 definitions (a comma separated list of elements, surrounded by
2134 42, i32 11, i32 74, i32 100 ></tt>". Vector constants must
2136 elements must match those specified by the type.</dd>
2142 This is often used to avoid having to print large zero initializers
2143 (e.g. for large arrays) and is always exactly equivalent to using explicit
2144 zero initializers.</dd>
2147 <dd>A metadata node is a structure-like constant with
2150 be interpreted as part of the instruction stream, metadata is a place to
2151 attach additional information such as debug info.</dd>
2152 </dl>
2154 </div>
2156 <!-- ======================================================================= -->
2159 </div>
2165 (link-time) constants. These constants are explicitly referenced when
2168 legal LLVM file:</p>
2171 @X = global i32 17
2172 @Y = global i32 42
2173 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2174 </pre>
2176 </div>
2178 <!-- ======================================================================= -->
2183 indicates that the user of the value may receive an unspecified bit-pattern.
2187 <p>Undefined values are useful because they indicate to the compiler that the
2188 program is well defined no matter what value is used. This gives the
2189 compiler more freedom to optimize. Here are some examples of (potentially
2190 surprising) transformations that are valid (in pseudo IR):</p>
2194 %A = add %X, undef
2195 %B = sub %X, undef
2196 %C = xor %X, undef
2197 Safe:
2198 %A = undef
2199 %B = undef
2200 %C = undef
2201 </pre>
2203 <p>This is safe because all of the output bits are affected by the undef bits.
2204 Any output bit can have a zero or one depending on the input bits.</p>
2207 %A = or %X, undef
2208 %B = and %X, undef
2209 Safe:
2210 %A = -1
2211 %B = 0
2212 Unsafe:
2213 %A = undef
2214 %B = undef
2215 </pre>
2217 <p>These logical operations have bits that are not always affected by the input.
2228 %A = select undef, %X, %Y
2229 %B = select undef, 42, %Y
2230 %C = select %X, %Y, undef
2231 Safe:
2232 %A = %X (or %Y)
2233 %B = 42 (or %Y)
2234 %C = %Y
2235 Unsafe:
2236 %A = undef
2237 %B = undef
2238 %C = undef
2239 </pre>
2248 eliminated.</p>
2251 %A = xor undef, undef
2253 %B = undef
2254 %C = xor %B, %B
2256 %D = undef
2257 %E = icmp lt %D, 4
2258 %F = icmp gte %D, 4
2260 Safe:
2261 %A = undef
2262 %B = undef
2263 %C = undef
2264 %D = undef
2265 %E = undef
2266 %F = undef
2267 </pre>
2270 necessarily the same. This can be surprising to people (and also matches C
2271 semantics) where they assume that "<tt>X^X</tt>" is always zero, even
2274 its value over its "live range". This is true because the variable doesn't
2276 from arbitrary registers that happen to be around when needed, so the value
2278 need to have the same semantics or the core LLVM "replace all uses with"
2279 concept would not hold.</p>
2282 %A = fdiv undef, %X
2283 %B = fdiv %X, undef
2284 Safe:
2285 %A = undef
2286 b: unreachable
2287 </pre>
2294 defined on SNaN's. However, in the second example, we can make a more
2296 arbitrary value, we are allowed to assume that it could be zero. Since a
2298 the operation does not execute at all. This allows us to delete the divide and
2299 all code after it. Because the undefined operation "can't happen", the
2300 optimizer can assume that it occurs in dead code.</p>
2303 a: store undef -> %X
2304 b: store %X -> undef
2305 Safe:
2307 b: unreachable
2308 </pre>
2311 undefined value can be assumed to not have any effect; we can assume that the
2312 value is overwritten with bits that happen to match what was already there.
2314 memory, therefore, it 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 transferred to another.</li>
2375 instruction if the set of instructions it otherwise depends on would be
2376 different if the terminator had transferred control to a different
2377 successor.</li>
2381 </ul>
2383 <p>Whenever a trap value is generated, all values which depend on it evaluate
2384 to trap. If they have side effects, the evoke their side effects as if each
2385 operand with a trap value were undef. If they have externally-visible side
2386 effects, the behavior is undefined.</p>
2391 entry:
2394 %trap_yet_again = getelementptr i32* @h, i32 %still_trap
2395 store i32 0, i32* %trap_yet_again ; undefined behavior
2397 store i32 %trap, i32* @g ; Trap value conceptually stored to memory.
2398 %trap2 = load i32* @g ; Returns a trap value, not just undef.
2400 volatile store i32 %trap, i32* @g ; External observation; undefined behavior.
2402 %narrowaddr = bitcast i32* @g to i16*
2403 %wideaddr = bitcast i32* @g to i64*
2404 %trap3 = load 16* %narrowaddr ; Returns a trap value.
2405 %trap4 = load i64* %widaddr ; Returns a trap value.
2407 %cmp = icmp i32 slt %trap, 0 ; Returns a trap value.
2408 %br i1 %cmp, %true, %end ; Branch to either destination.
2410 true:
2411 volatile store i32 0, i32* @g ; This is control-dependent on %cmp, so
2412 ; it has undefined behavior.
2413 br label %end
2415 end:
2417 ; Both edges into this PHI are
2418 ; control-dependent on %cmp, so this
2419 ; always results in a trap value.
2421 volatile store i32 0, i32* @g ; This would depend on the store in %true
2422 ; if %cmp is true, or the store in %entry
2423 ; otherwise, so this is undefined behavior.
2425 %br i1 %cmp, %second_true, %second_end
2426 ; The same branch again, but this time the
2427 ; true block doesn't have side effects.
2429 second_true:
2430 ; No side effects!
2431 br label %end
2433 second_end:
2434 volatile store i32 0, i32* @g ; This time, the instruction always depends
2435 ; on the store in %end. Also, it is
2436 ; control-equivalent to %end, so this is
2437 ; well- defined (again, ignoring earlier
2438 ; undefined behavior in this example).
2439 </pre>
2441 </div>
2443 <!-- ======================================================================= -->
2445 Blocks</a></div>
2451 basic block in the specified function, and always has an i8* type. Taking
2452 the address of the entry block is illegal.</p>
2454 <p>This value only has defined behavior when used as an operand to the
2456 comparisons against null. Pointer equality tests between labels addresses
2457 results in undefined behavior — though, again, comparison against null
2458 is ok, and no label is equal to the null pointer. This may be passed around
2459 as an opaque pointer sized value as long as the bits are not inspected. This
2462 instruction.</p>
2464 <p>Finally, some targets may provide defined semantics when using the value as
2465 the operand to an inline assembly, but that is target specific.</p>
2467 </div>
2470 <!-- ======================================================================= -->
2472 </div>
2476 <p>Constant expressions are used to allow expressions involving other constants
2477 to be used as constants. Constant expressions may be of
2479 operation that does not have side effects (e.g. load and call are not
2480 supported). The following is the syntax for constant expressions:</p>
2482 <dl>
2484 <dd>Truncate a constant to another type. The bit size of CST must be larger
2485 than the bit size of TYPE. Both types must be integers.</dd>
2488 <dd>Zero extend a constant to another type. The bit size of CST must be
2489 smaller than the bit size of TYPE. Both types must be integers.</dd>
2492 <dd>Sign extend a constant to another type. The bit size of CST must be
2493 smaller than the bit size of TYPE. Both types must be integers.</dd>
2496 <dd>Truncate a floating point constant to another floating point type. The
2497 size of CST must be larger than the size of TYPE. Both types must be
2498 floating point.</dd>
2501 <dd>Floating point extend a constant to another type. The size of CST must be
2502 smaller or equal to the size of TYPE. Both types must be floating
2503 point.</dd>
2506 <dd>Convert a floating point constant to the corresponding unsigned integer
2507 constant. TYPE must be a scalar or vector integer type. CST must be of
2508 scalar or vector floating point type. Both CST and TYPE must be scalars,
2509 or vectors of the same number of elements. If the value won't fit in the
2510 integer type, the results are undefined.</dd>
2513 <dd>Convert a floating point constant to the corresponding signed integer
2514 constant. TYPE must be a scalar or vector integer type. CST must be of
2515 scalar or vector floating point type. Both CST and TYPE must be scalars,
2516 or vectors of the same number of elements. If the value won't fit in the
2517 integer type, the results are undefined.</dd>
2520 <dd>Convert an unsigned integer constant to the corresponding floating point
2521 constant. TYPE must be a scalar or vector floating point type. CST must be
2522 of scalar or vector integer type. Both CST and TYPE must be scalars, or
2523 vectors of the same number of elements. If the value won't fit in the
2524 floating point type, the results are undefined.</dd>
2527 <dd>Convert a signed integer constant to the corresponding floating point
2528 constant. TYPE must be a scalar or vector floating point type. CST must be
2529 of scalar or vector integer type. Both CST and TYPE must be scalars, or
2530 vectors of the same number of elements. If the value won't fit in the
2531 floating point type, the results are undefined.</dd>
2534 <dd>Convert a pointer typed constant to the corresponding integer constant
2540 <dd>Convert a integer constant to a pointer constant. TYPE must be a pointer
2541 type. CST must be of integer type. The CST value is zero extended,
2542 truncated, or unchanged to make it fit in a pointer size. This one is
2546 <dd>Convert a constant, CST, to another TYPE. The constraints of the operands
2554 instruction, the index list may have zero or more indexes, which are
2568 constants.</dd>
2572 constants.</dd>
2576 constants.</dd>
2580 constants. The index list is interpreted in a similar manner as indices in
2582 index value must be specified.</dd>
2586 constants. The index list is interpreted in a similar manner as indices in
2588 index value must be specified.</dd>
2591 <dd>Perform the specified operation of the LHS and RHS constants. OPCODE may
2594 on operands are the same as those for the corresponding instruction
2595 (e.g. no bitwise operations on floating point values are allowed).</dd>
2596 </dl>
2598 </div>
2600 <!-- *********************************************************************** -->
2602 <!-- *********************************************************************** -->
2604 <!-- ======================================================================= -->
2607 </div>
2611 <p>LLVM supports inline assembler expressions (as opposed
2613 a special value. This value represents the inline assembler as a string
2614 (containing the instructions to emit), a list of operand constraints (stored
2615 as a string), a flag that indicates whether or not the inline asm
2616 expression has side effects, and a flag indicating whether the function
2617 containing the asm needs to align its stack conservatively. An example
2618 inline assembler expression is:</p>
2622 </pre>
2626 have:</p>
2630 </pre>
2632 <p>Inline asms with side effects not visible in the constraint list must be
2633 marked as having side effects. This is done through the use of the
2638 </pre>
2640 <p>In some cases inline asms will contain code that will not work unless the
2641 stack is aligned in some way, such as calls or SSE instructions on x86,
2642 yet will not contain code that does that alignment within the asm.
2643 The compiler should make conservative assumptions about what the asm might
2644 contain and should generate its usual stack alignment code in the prologue
2649 </pre>
2652 first.</p>
2654 <p>TODO: The format of the asm and constraints string still need to be
2655 documented here. Constraints on what can be done (e.g. duplication, moving,
2656 etc need to be documented). This is probably best done by reference to
2657 another document that covers inline asm from a holistic perspective.</p>
2658 </div>
2662 </div>
2667 attached to it that contains a list of constant integers. If present, the
2668 code generator will use the integer as the location cookie value when report
2669 errors through the LLVMContext error reporting mechanisms. This allows a
2670 front-end to correlate backend errors that occur with inline asm back to the
2671 source code that produced it. For example:</p>
2675 ...
2677 </pre>
2679 <p>It is up to the front-end to make sense of the magic numbers it places in the
2680 IR. If the MDNode contains multiple constants, the code generator will use
2681 the one that corresponds to the line of the asm that the error occurs on.</p>
2683 </div>
2685 <!-- ======================================================================= -->
2687 Strings</a>
2688 </div>
2692 <p>LLVM IR allows metadata to be attached to instructions in the program that
2693 can convey extra information about the code to the optimizers and code
2694 generator. One example application of metadata is source-level debug
2695 information. There are two metadata primitives: strings and nodes. All
2699 <p>A metadata string is a string surrounded by double quotes. It can contain
2703 <p>Metadata nodes are represented with notation similar to structure constants
2704 (a comma separated list of elements, surrounded by braces and preceded by an
2709 metadata nodes, which can be looked up in the module symbol table. For
2710 example: "<tt>!foo = metadata !{!4, !3}</tt>".
2713 function is using two metadata arguments.</p>
2716 <pre>
2718 </pre>
2719 </div>
2725 <pre>
2727 </pre>
2728 </div>
2730 </div>
2733 <!-- *********************************************************************** -->
2736 </div>
2737 <!-- *********************************************************************** -->
2740 code generation or other IR semantics. These are documented here. All globals
2741 of this sort should have a section specified as "<tt>llvm.metadata</tt>". This
2742 section and all globals that start with "<tt>llvm.</tt>" are reserved for use
2743 by LLVM.</p>
2745 <!-- ======================================================================= -->
2748 </div>
2754 pointers to global variables and functions which may optionally have a pointer
2755 cast formed of bitcast or getelementptr. For example, a legal use of it is:</p>
2757 <pre>
2758 @X = global i8 4
2759 @Y = global i32 123
2761 @llvm.used = appending global [2 x i8*] [
2762 i8* @X,
2763 i8* bitcast (i32* @Y to i8*)
2764 ], section "llvm.metadata"
2765 </pre>
2768 compiler, assembler, and linker are required to treat the symbol as if there is
2769 a reference to the global that it cannot see. For example, if a variable has
2771 list, it cannot be deleted. This is commonly used to represent references from
2772 inline asms and other things the compiler cannot "see", and corresponds to
2775 <p>On some targets, the code generator must emit a directive to the assembler or
2776 object file to prevent the assembler and linker from molesting the symbol.</p>
2778 </div>
2780 <!-- ======================================================================= -->
2783 </div>
2789 touching the symbol. On targets that support it, this allows an intelligent
2790 linker to optimize references to the symbol without being impeded as it would be
2793 <p>This is a rare construct that should only be used in rare circumstances, and
2794 should not be exposed to source languages.</p>
2796 </div>
2798 <!-- ======================================================================= -->
2801 </div>
2804 <pre>
2805 %0 = type { i32, void ()* }
2807 </pre>
2808 <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.
2809 </p>
2811 </div>
2813 <!-- ======================================================================= -->
2816 </div>
2819 <pre>
2820 %0 = type { i32, void ()* }
2822 </pre>
2824 <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.
2825 </p>
2827 </div>
2830 <!-- *********************************************************************** -->
2832 <!-- *********************************************************************** -->
2836 <p>The LLVM instruction set consists of several different classifications of
2843 </div>
2845 <!-- ======================================================================= -->
2847 Instructions</a> </div>
2852 in a program ends with a "Terminator" instruction, which indicates which
2853 block should be executed after the current block is finished. These
2855 control flow, not values (the one exception being the
2858 <p>There are seven different terminator instructions: the
2867 </div>
2869 <!-- _______________________________________________________________________ -->
2871 Instruction</a> </div>
2876 <pre>
2879 </pre>
2883 a value) from a function back to the caller.</p>
2886 value and then causes control flow, and one that just causes control flow to
2887 occur.</p>
2891 return value. The type of the return value must be a
2896 value or a return value with a type that does not match its type, or if it
2898 return value.</p>
2902 the calling function's context. If the caller is a
2904 instruction after the call. If the caller was an
2906 the beginning of the "normal" destination block. If the instruction returns
2907 a value, that value shall set the call or invoke instruction's return
2908 value.</p>
2911 <pre>
2915 </pre>
2917 </div>
2918 <!-- _______________________________________________________________________ -->
2924 <pre>
2925 br i1 <cond>, label <iftrue>, label <iffalse><br> br label <dest> <i>; Unconditional branch</i>
2926 </pre>
2930 different basic block in the current function. There are two forms of this
2931 instruction, corresponding to a conditional branch and an unconditional
2932 branch.</p>
2938 target.</p>
2947 <pre>
2948 Test:
2950 br i1 %cond, label %IfEqual, label %IfUnequal
2951 IfEqual:
2953 IfUnequal:
2955 </pre>
2957 </div>
2959 <!-- _______________________________________________________________________ -->
2962 </div>
2967 <pre>
2968 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
2969 </pre>
2974 instruction, allowing a branch to occur to one of many possible
2975 destinations.</p>
2981 The table is not allowed to contain duplicate constant entries.</p>
2986 is searched for the given value. If the value is found, control flow is
2987 transferred to the corresponding destination; otherwise, control flow is
2988 transferred to the default destination.</p>
2991 <p>Depending on properties of the target machine and the particular
2993 different ways. For example, it could be generated as a series of chained
2994 conditional branches or with a lookup table.</p>
2997 <pre>
3000 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3003 switch i32 0, label %dest [ ]
3006 switch i32 %val, label %otherwise [ i32 0, label %onzero
3007 i32 1, label %onone
3008 i32 2, label %ontwo ]
3009 </pre>
3011 </div>
3014 <!-- _______________________________________________________________________ -->
3017 </div>
3022 <pre>
3024 </pre>
3029 within the current function, whose address is specified by
3036 rest of the arguments indicate the full set of possible destinations that the
3037 address may point to. Blocks are allowed to occur multiple times in the
3038 destination list, though this isn't particularly useful.</p>
3040 <p>This destination list is required so that dataflow analysis has an accurate
3041 understanding of the CFG.</p>
3045 <p>Control transfers to the block specified in the address argument. All
3046 possible destination blocks must be listed in the label list, otherwise this
3047 instruction has undefined behavior. This implies that jumps to labels
3048 defined in other functions have undefined behavior as well.</p>
3055 <pre>
3056 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3057 </pre>
3059 </div>
3062 <!-- _______________________________________________________________________ -->
3065 </div>
3070 <pre>
3071 <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>]
3073 </pre>
3077 function, with the possibility of control flow transfer to either the
3080 control flow will return to the "normal" label. If the callee (or any
3082 instruction, control is interrupted and continued at the dynamically nearest
3088 <ol>
3090 convention</a> the call should use. If none is specified, the call
3091 defaults to using C calling conventions.</li>
3098 function value being invoked. In most cases, this is a direct function
3100 off an arbitrary pointer to function value.</li>
3103 function to be invoked. </li>
3106 signature argument types and parameter attributes. All arguments must be
3108 signature indicates the function accepts a variable number of arguments,
3109 the extra arguments can be specified.</li>
3120 </ol>
3123 <p>This instruction is designed to operate as a standard
3125 primary difference is that it establishes an association with a label, which
3126 is used by the runtime library to unwind the stack.</p>
3128 <p>This instruction is used in languages with destructors to ensure that proper
3130 exception. Additionally, this is important for implementation of
3133 <p>For the purposes of the SSA form, the definition of the value returned by the
3135 block to the "normal" label. If the callee unwinds then no return value is
3136 available.</p>
3138 <p>Note that the code generator does not yet completely support unwind, and
3139 that the invoke/unwind semantics are likely to change in future versions.</p>
3142 <pre>
3143 %retval = invoke i32 @Test(i32 15) to label %Continue
3147 </pre>
3149 </div>
3151 <!-- _______________________________________________________________________ -->
3154 Instruction</a> </div>
3159 <pre>
3160 unwind
3161 </pre>
3165 at the first callee in the dynamic call stack which used
3167 This is primarily used to implement exception handling.</p>
3171 immediately halt. The dynamic call stack is then searched for the
3173 Once found, execution continues at the "exceptional" destination block
3175 instruction in the dynamic call chain, undefined behavior results.</p>
3177 <p>Note that the code generator does not yet completely support unwind, and
3178 that the invoke/unwind semantics are likely to change in future versions.</p>
3180 </div>
3182 <!-- _______________________________________________________________________ -->
3185 Instruction</a> </div>
3190 <pre>
3191 unreachable
3192 </pre>
3196 instruction is used to inform the optimizer that a particular portion of the
3197 code is not reachable. This can be used to indicate that the code after a
3198 no-return function cannot be reached, and other facts.</p>
3203 </div>
3205 <!-- ======================================================================= -->
3210 <p>Binary operators are used to do most of the computation in a program. They
3211 require two operands of the same type, execute an operation on them, and
3212 produce a single value. The operands might represent multiple data, as is
3214 has the same type as its operands.</p>
3218 </div>
3220 <!-- _______________________________________________________________________ -->
3223 </div>
3228 <pre>
3233 </pre>
3241 integer values. Both arguments must have identical types.</p>
3246 <p>If the sum has unsigned overflow, the result returned is the mathematical
3249 <p>Because LLVM integers use a two's complement representation, this instruction
3250 is appropriate for both signed and unsigned integers.</p>
3256 respectively, occurs.</p>
3259 <pre>
3261 </pre>
3263 </div>
3265 <!-- _______________________________________________________________________ -->
3268 </div>
3273 <pre>
3275 </pre>
3283 floating point values. Both arguments must have identical types.</p>
3289 <pre>
3291 </pre>
3293 </div>
3295 <!-- _______________________________________________________________________ -->
3298 </div>
3303 <pre>
3308 </pre>
3312 operands.</p>
3316 representations.</p>
3321 integer values. Both arguments must have identical types.</p>
3326 <p>If the difference has unsigned overflow, the result returned is the
3328 result.</p>
3330 <p>Because LLVM integers use a two's complement representation, this instruction
3331 is appropriate for both signed and unsigned integers.</p>
3337 respectively, occurs.</p>
3340 <pre>
3343 </pre>
3345 </div>
3347 <!-- _______________________________________________________________________ -->
3350 </div>
3355 <pre>
3357 </pre>
3361 operands.</p>
3365 representations.</p>
3370 floating point values. Both arguments must have identical types.</p>
3376 <pre>
3379 </pre>
3381 </div>
3383 <!-- _______________________________________________________________________ -->
3386 </div>
3391 <pre>
3396 </pre>
3404 integer values. Both arguments must have identical types.</p>
3409 <p>If the result of the multiplication has unsigned overflow, the result
3411 width of the result.</p>
3413 <p>Because LLVM integers use a two's complement representation, and the result
3414 is the same width as the operands, this instruction returns the correct
3415 result for both signed and unsigned integers. If a full product
3417 be sign-extended or zero-extended as appropriate to the width of the full
3418 product.</p>
3424 respectively, occurs.</p>
3427 <pre>
3429 </pre>
3431 </div>
3433 <!-- _______________________________________________________________________ -->
3436 </div>
3441 <pre>
3443 </pre>
3451 floating point values. Both arguments must have identical types.</p>
3457 <pre>
3459 </pre>
3461 </div>
3463 <!-- _______________________________________________________________________ -->
3465 </a></div>
3470 <pre>
3473 </pre>
3481 values. Both arguments must have identical types.</p>
3486 <p>Note that unsigned integer division and signed integer division are distinct
3497 <pre>
3499 </pre>
3501 </div>
3503 <!-- _______________________________________________________________________ -->
3505 </a> </div>
3510 <pre>
3513 </pre>
3521 values. Both arguments must have identical types.</p>
3524 <p>The value produced is the signed integer quotient of the two operands rounded
3525 towards zero.</p>
3527 <p>Note that signed integer division and unsigned integer division are distinct
3530 <p>Division by zero leads to undefined behavior. Overflow also leads to
3531 undefined behavior; this is a rare case, but can occur, for example, by doing
3536 be rounded.</p>
3539 <pre>
3541 </pre>
3543 </div>
3545 <!-- _______________________________________________________________________ -->
3547 Instruction</a> </div>
3552 <pre>
3554 </pre>
3562 floating point values. Both arguments must have identical types.</p>
3568 <pre>
3570 </pre>
3572 </div>
3574 <!-- _______________________________________________________________________ -->
3576 </div>
3581 <pre>
3583 </pre>
3587 division of its two arguments.</p>
3592 values. Both arguments must have identical types.</p>
3596 This instruction always performs an unsigned division to get the
3597 remainder.</p>
3599 <p>Note that unsigned integer remainder and signed integer remainder are
3605 <pre>
3607 </pre>
3609 </div>
3611 <!-- _______________________________________________________________________ -->
3614 </div>
3619 <pre>
3621 </pre>
3625 division of its two operands. This instruction can also take
3627 elements must be integers.</p>
3632 values. Both arguments must have identical types.</p>
3639 For more information about the difference,
3641 Math Forum</a>. For a table of how this is implemented in various languages,
3645 <p>Note that signed integer remainder and unsigned integer remainder are
3648 <p>Taking the remainder of a division by zero leads to undefined behavior.
3649 Overflow also leads to undefined behavior; this is a rare case, but can
3650 occur, for example, by taking the remainder of a 32-bit division of
3652 lets srem be implemented using instructions that return both the result of
3653 the division and the remainder.)</p>
3656 <pre>
3658 </pre>
3660 </div>
3662 <!-- _______________________________________________________________________ -->
3669 <pre>
3671 </pre>
3675 its two operands.</p>
3680 floating point values. Both arguments must have identical types.</p>
3684 has the same sign as the dividend.</p>
3687 <pre>
3689 </pre>
3691 </div>
3693 <!-- ======================================================================= -->
3695 Operations</a> </div>
3699 <p>Bitwise binary operators are used to do various forms of bit-twiddling in a
3700 program. They are generally very efficient instructions and can commonly be
3701 strength reduced from other instructions. They require two operands of the
3702 same type, execute an operation on them, and produce a single value. The
3703 resulting value is the same type as its operands.</p>
3705 </div>
3707 <!-- _______________________________________________________________________ -->
3709 Instruction</a> </div>
3714 <pre>
3719 </pre>
3723 a specified number of bits.</p>
3733 is (statically or dynamically) negative or equal to or larger than the number
3742 with the resultant sign bit. As such, NUW/NSW have the same semantics as
3743 they would if the shift were expressed as a mul instruction with the same
3747 <pre>
3752 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> <i>; yields: result=<2 x i32> < i32 2, i32 4></i>
3753 </pre>
3755 </div>
3757 <!-- _______________________________________________________________________ -->
3759 Instruction</a> </div>
3764 <pre>
3767 </pre>
3771 operand shifted to the right a specified number of bits with zero fill.</p>
3779 <p>This instruction always performs a logical shift right operation. The most
3780 significant bits of the result will be filled with zero bits after the shift.
3788 shifted out are non-zero.</p>
3792 <pre>
3798 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> <i>; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1></i>
3799 </pre>
3801 </div>
3803 <!-- _______________________________________________________________________ -->
3805 Instruction</a> </div>
3809 <pre>
3812 </pre>
3816 operand shifted to the right a specified number of bits with sign
3817 extension.</p>
3825 <p>This instruction always performs an arithmetic shift right operation, The
3826 most significant bits of the result will be filled with the sign bit
3834 shifted out are non-zero.</p>
3837 <pre>
3843 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> <i>; yields: result=<2 x i32> < i32 -1, i32 0></i>
3844 </pre>
3846 </div>
3848 <!-- _______________________________________________________________________ -->
3850 Instruction</a> </div>
3855 <pre>
3857 </pre>
3861 operands.</p>
3866 values. Both arguments must have identical types.</p>
3872 <tbody>
3873 <tr>
3877 </tr>
3878 <tr>
3882 </tr>
3883 <tr>
3887 </tr>
3888 <tr>
3892 </tr>
3893 <tr>
3897 </tr>
3898 </tbody>
3899 </table>
3902 <pre>
3906 </pre>
3907 </div>
3908 <!-- _______________________________________________________________________ -->
3914 <pre>
3916 </pre>
3920 two operands.</p>
3925 values. Both arguments must have identical types.</p>
3931 <tbody>
3932 <tr>
3936 </tr>
3937 <tr>
3941 </tr>
3942 <tr>
3946 </tr>
3947 <tr>
3951 </tr>
3952 <tr>
3956 </tr>
3957 </tbody>
3958 </table>
3961 <pre>
3965 </pre>
3967 </div>
3969 <!-- _______________________________________________________________________ -->
3971 Instruction</a> </div>
3976 <pre>
3978 </pre>
3988 values. Both arguments must have identical types.</p>
3994 <tbody>
3995 <tr>
3999 </tr>
4000 <tr>
4004 </tr>
4005 <tr>
4009 </tr>
4010 <tr>
4014 </tr>
4015 <tr>
4019 </tr>
4020 </tbody>
4021 </table>
4024 <pre>
4029 </pre>
4031 </div>
4033 <!-- ======================================================================= -->
4036 </div>
4040 <p>LLVM supports several instructions to represent vector operations in a
4041 target-independent manner. These instructions cover the element-access and
4042 vector-specific operations needed to process vectors effectively. While LLVM
4043 does directly support these vector operations, many sophisticated algorithms
4044 will want to use target-specific intrinsics to take full advantage of a
4045 specific target.</p>
4047 </div>
4049 <!-- _______________________________________________________________________ -->
4052 </div>
4057 <pre>
4058 <result> = extractelement <n x <ty>> <val>, i32 <idx> <i>; yields <ty></i>
4059 </pre>
4063 from a vector at a specified index.</p>
4069 indicating the position from which to extract the element. The index may be
4070 a variable.</p>
4073 <p>The result is a scalar of the same type as the element type of
4076 results are undefined.</p>
4079 <pre>
4081 </pre>
4083 </div>
4085 <!-- _______________________________________________________________________ -->
4088 </div>
4093 <pre>
4094 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> <i>; yields <n x <ty>></i>
4095 </pre>
4099 vector at a specified index.</p>
4104 whose type must equal the element type of the first operand. The third
4105 operand is an index indicating the position at which to insert the value.
4106 The index may be a variable.</p>
4112 results are undefined.</p>
4115 <pre>
4116 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 <i>; yields <4 x i32></i>
4117 </pre>
4119 </div>
4121 <!-- _______________________________________________________________________ -->
4124 </div>
4129 <pre>
4130 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> <i>; yields <m x <ty>></i>
4131 </pre>
4135 from two input vectors, returning a vector with the same element type as the
4136 input and length that is the same as the shuffle mask.</p>
4140 with types that match each other. The third argument is a shuffle mask whose
4141 element type is always 'i32'. The result of the instruction is a vector
4142 whose length is the same as the shuffle mask and whose element type is the
4143 same as the element type of the first two operands.</p>
4145 <p>The shuffle mask operand is required to be a constant vector with either
4146 constant integer or undef values.</p>
4149 <p>The elements of the two input vectors are numbered from left to right across
4150 both of the vectors. The shuffle mask operand specifies, for each element of
4151 the result vector, which element of the two input vectors the result element
4152 gets. The element selector may be undef (meaning "don't care") and the
4153 second operand may be undef if performing a shuffle from only one vector.</p>
4156 <pre>
4160 <4 x i32> <i32 0, i32 1, i32 2, i32 3> <i>; yields <4 x i32></i> - Identity shuffle.
4164 <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>
4165 </pre>
4167 </div>
4169 <!-- ======================================================================= -->
4172 </div>
4176 <p>LLVM supports several instructions for working with
4179 </div>
4181 <!-- _______________________________________________________________________ -->
4184 </div>
4189 <pre>
4191 </pre>
4201 specify which value to extract in a similar manner as indices in a
4204 <ul>
4205 <li>Since the value being indexed is not a pointer, the first index is
4206 omitted and assumed to be zero.</li>
4208 <li>Not only struct indices but also array indices must be in
4209 bounds.</li>
4210 </ul>
4213 <p>The result is the value at the position in the aggregate specified by the
4214 index operands.</p>
4217 <pre>
4219 </pre>
4221 </div>
4223 <!-- _______________________________________________________________________ -->
4226 </div>
4231 <pre>
4232 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx> <i>; yields <aggregate type></i>
4233 </pre>
4243 value to insert. The following operands are constant indices indicating
4244 the position at which to insert the value in a similar manner as indices in a
4246 value to insert must have the same type as the value identified by the
4247 indices.</p>
4255 <pre>
4258 </pre>
4260 </div>
4263 <!-- ======================================================================= -->
4266 </div>
4270 <p>A key design point of an SSA-based representation is how it represents
4271 memory. In LLVM, no memory locations are in SSA form, which makes things
4272 very simple. This section describes how to read, write, and allocate
4273 memory in LLVM.</p>
4275 </div>
4277 <!-- _______________________________________________________________________ -->
4280 </div>
4285 <pre>
4286 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] <i>; yields {type*}:result</i>
4287 </pre>
4291 currently executing function, to be automatically released when this function
4292 returns to its caller. The object is always allocated in the generic address
4293 space (address space zero).</p>
4298 runtime stack, returning a pointer of the appropriate type to the program.
4299 If "NumElements" is specified, it is the number of elements allocated,
4300 otherwise "NumElements" is defaulted to be one. If a constant alignment is
4301 specified, the value result of the allocation is guaranteed to be aligned to
4302 at least that boundary. If not specified, or if zero, the target can choose
4303 to align the allocation on any convenient boundary compatible with the
4304 type.</p>
4309 <p>Memory is allocated; a pointer is returned. The operation is undefined if
4311 memory is automatically released when the function returns. The
4313 variables that must have an address available. When the function returns
4316 reclaimed. Allocating zero bytes is legal, but the result is undefined.</p>
4319 <pre>
4324 </pre>
4326 </div>
4328 <!-- _______________________________________________________________________ -->
4330 Instruction</a> </div>
4335 <pre>
4336 <result> = load <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>]
4337 <result> = volatile load <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>]
4339 </pre>
4346 from which to load. The pointer must point to
4353 operation (that is, the alignment of the memory address). A value of 0 or an
4355 alignment for the target. It is the responsibility of the code emitter to
4356 ensure that the alignment information is correct. Overestimating the
4357 alignment results in undefined behavior. Underestimating the alignment may
4364 and code generator that this load is not expected to be reused in the cache.
4365 The code generator may select special instructions to save cache bandwidth,
4369 <p>The location of memory pointed to is loaded. If the value being loaded is of
4370 scalar type then the number of bytes read does not exceed the minimum number
4371 of bytes needed to hold all bits of the type. For example, loading an
4374 is undefined if the value was not originally written using a store of the
4375 same type.</p>
4378 <pre>
4382 </pre>
4384 </div>
4386 <!-- _______________________________________________________________________ -->
4388 Instruction</a> </div>
4393 <pre>
4394 store <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] <i>; yields {void}</i>
4395 volatile store <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] <i>; yields {void}</i>
4396 </pre>
4403 and an address at which to store it. The type of the
4412 operation (that is, the alignment of the memory address). A value of 0 or an
4413 omitted "align" argument means that the operation has the preferential
4414 alignment for the target. It is the responsibility of the code emitter to
4415 ensure that the alignment information is correct. Overestimating the
4416 alignment results in an undefined behavior. Underestimating the alignment may
4419 <p>The optional !nontemporal metadata must reference a single metatadata
4421 value 1. The existence of the !nontemporal metatadata on the
4422 instruction tells the optimizer and code generator that this load is
4423 not expected to be reused in the cache. The code generator may
4424 select special instructions to save cache bandwidth, such as the
4425 MOVNT instruction on x86.</p>
4432 does not exceed the minimum number of bytes needed to hold all bits of the
4435 integral number of bytes, it is unspecified what happens to the extra bits
4436 that do not belong to the type, but they will typically be overwritten.</p>
4439 <pre>
4443 </pre>
4445 </div>
4447 <!-- _______________________________________________________________________ -->
4450 </div>
4455 <pre>
4458 </pre>
4463 It performs address calculation only and does not access memory.</p>
4466 <p>The first argument is always a pointer, and forms the basis of the
4467 calculation. The remaining arguments are indices that indicate which of the
4468 elements of the aggregate object are indexed. The interpretation of each
4469 index is dependent on the type being indexed into. The first index always
4470 indexes the pointer value given as the first argument, the second index
4471 indexes a value of the type pointed to (not necessarily the value directly
4472 pointed to, since the first index can be non-zero), etc. The first type
4473 indexed into must be a pointer value, subsequent types can be arrays,
4474 vectors, and structs. Note that subsequent types being indexed into
4475 can never be pointers, since that would require loading the pointer before
4476 continuing calculation.</p>
4478 <p>The type of each index argument depends on the type it is indexing into.
4481 or vector, integers of any width are allowed, and they are not required to be
4482 constant.</p>
4484 <p>For example, let's consider a C code fragment and how it gets compiled to
4485 LLVM:</p>
4488 struct RT {
4489 char A;
4491 char C;
4492 };
4493 struct ST {
4494 int X;
4495 double Y;
4496 struct RT Z;
4497 };
4499 int *foo(struct ST *s) {
4501 }
4502 </pre>
4510 define i32* @foo(%ST* %s) {
4511 entry:
4513 ret i32* %reg
4514 }
4515 </pre>
4520 }</tt>' type, a structure. The second index indexes into the third element
4522 i8 }</tt>' type, another structure. The third index indexes into the second
4524 array. The two dimensions of the array are subscripted into, yielding an
4528 <p>Note that it is perfectly legal to index partially through a structure,
4529 returning a pointer to an inner element. Because of this, the LLVM code for
4530 the given testcase is equivalent to:</p>
4532 <pre>
4533 define i32* @foo(%ST* %s) {
4539 ret i32* %t5
4540 }
4541 </pre>
4546 or if any of the addresses that would be formed by successive addition of
4547 the offsets implied by the indices to the base address with infinitely
4550 the addresses that point into the object, plus the address one byte past
4551 the end.</p>
4554 the base address with silently-wrapping two's complement arithmetic, and
4556 pointed to by the base pointer. The result value may not necessarily be
4557 used to access memory though, even if it happens to point into allocated
4559 section for more information.</p>
4561 <p>The getelementptr instruction is often confusing. For some more insight into
4565 <pre>
4574 </pre>
4576 </div>
4578 <!-- ======================================================================= -->
4580 </div>
4584 <p>The instructions in this category are the conversion instructions (casting)
4585 which all take a single operand and a type. They perform various bit
4586 conversions on the operand.</p>
4588 </div>
4590 <!-- _______________________________________________________________________ -->
4593 </div>
4597 <pre>
4599 </pre>
4608 of the same number of integers.
4611 Equal sized types are not allowed.</p>
4620 <pre>
4624 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> <i>; yields <i8 8, i8 7></i>
4625 </pre>
4627 </div>
4629 <!-- _______________________________________________________________________ -->
4632 </div>
4636 <pre>
4638 </pre>
4648 of the same number of integers.
4650 the bit size of the destination type,
4660 <pre>
4663 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> <i>; yields <i32 8, i32 7></i>
4664 </pre>
4666 </div>
4668 <!-- _______________________________________________________________________ -->
4671 </div>
4675 <pre>
4677 </pre>
4685 of the same number of integers.
4687 the bit size of the destination type,
4698 <pre>
4701 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> <i>; yields <i32 8, i32 7></i>
4702 </pre>
4704 </div>
4706 <!-- _______________________________________________________________________ -->
4709 </div>
4714 <pre>
4716 </pre>
4734 undefined.</p>
4737 <pre>
4740 </pre>
4742 </div>
4744 <!-- _______________________________________________________________________ -->
4747 </div>
4751 <pre>
4753 </pre>
4757 floating point value.</p>
4763 type must be smaller than the destination type.</p>
4773 <pre>
4776 </pre>
4778 </div>
4780 <!-- _______________________________________________________________________ -->
4783 </div>
4787 <pre>
4789 </pre>
4805 towards zero) unsigned integer value. If the value cannot fit
4809 <pre>
4813 </pre>
4815 </div>
4817 <!-- _______________________________________________________________________ -->
4820 </div>
4824 <pre>
4826 </pre>
4844 the results are undefined.</p>
4847 <pre>
4851 </pre>
4853 </div>
4855 <!-- _______________________________________________________________________ -->
4858 </div>
4862 <pre>
4864 </pre>
4879 integer quantity and converts it to the corresponding floating point
4880 value. If the value cannot fit in the floating point value, the results are
4881 undefined.</p>
4884 <pre>
4887 </pre>
4889 </div>
4891 <!-- _______________________________________________________________________ -->
4894 </div>
4898 <pre>
4900 </pre>
4915 quantity and converts it to the corresponding floating point value. If the
4916 value cannot fit in the floating point value, the results are undefined.</p>
4919 <pre>
4922 </pre>
4924 </div>
4926 <!-- _______________________________________________________________________ -->
4929 </div>
4933 <pre>
4935 </pre>
4949 truncating or zero extending that value to the size of the integer type. If
4953 change.</p>
4956 <pre>
4959 </pre>
4961 </div>
4963 <!-- _______________________________________________________________________ -->
4966 </div>
4970 <pre>
4972 </pre>
4980 value to cast, and a type to cast it to, which must be a
4988 than the size of a pointer then a zero extension is done. If they are the
4992 <pre>
4996 </pre>
4998 </div>
5000 <!-- _______________________________________________________________________ -->
5003 </div>
5007 <pre>
5009 </pre>
5017 non-aggregate first class value, and a type to cast it to, which must also be
5020 identical. If the source type is a pointer, the destination type must also be
5021 a pointer. This instruction supports bitwise conversion of vectors to
5022 integers and to vectors of other types (as long as they have the same
5023 size).</p>
5030 be converted to other pointer types with this instruction. To convert
5035 <pre>
5039 </pre>
5041 </div>
5043 <!-- ======================================================================= -->
5049 defy better classification.</p>
5051 </div>
5053 <!-- _______________________________________________________________________ -->
5055 </div>
5060 <pre>
5061 <result> = icmp <cond> <ty> <op1>, <op2> <i>; yields {i1} or {<N x i1>}:result</i>
5062 </pre>
5066 boolean values based on comparison of its two integer, integer vector, or
5067 pointer operands.</p>
5071 the condition code indicating the kind of comparison to perform. It is not a
5072 value, just a keyword. The possible condition code are:</p>
5074 <ol>
5085 </ol>
5089 typed. They must also be identical types.</p>
5095 result, as follows:</p>
5097 <ol>
5100 performed.</li>
5104 performed.</li>
5131 </ol>
5134 values are compared as if they were integers.</p>
5136 <p>If the operands are integer vectors, then they are compared element by
5141 <pre>
5148 </pre>
5150 <p>Note that the code generator does not yet support vector types with
5153 </div>
5155 <!-- _______________________________________________________________________ -->
5157 </div>
5162 <pre>
5163 <result> = fcmp <cond> <ty> <op1>, <op2> <i>; yields {i1} or {<N x i1>}:result</i>
5164 </pre>
5168 values based on comparison of its operands.</p>
5170 <p>If the operands are floating point scalars, then the result type is a boolean
5173 <p>If the operands are floating point vectors, then the result type is a vector
5174 of boolean with the same number of elements as the operands being
5175 compared.</p>
5179 the condition code indicating the kind of comparison to perform. It is not a
5180 value, just a keyword. The possible condition code are:</p>
5182 <ol>
5199 </ol>
5207 identical types.</p>
5212 vectors, then the vectors are compared element by element. Each comparison
5214 follows:</p>
5216 <ol>
5260 </ol>
5263 <pre>
5268 </pre>
5270 <p>Note that the code generator does not yet support vector types with
5273 </div>
5275 <!-- _______________________________________________________________________ -->
5278 </div>
5283 <pre>
5285 </pre>
5289 SSA graph representing the function.</p>
5292 <p>The type of the incoming values is specified with the first type field. After
5294 one pair for each predecessor basic block of the current block. Only values
5296 arguments to the PHI node. Only labels may be used as the label
5297 arguments.</p>
5299 <p>There must be no non-phi instructions between the start of a basic block and
5300 the PHI instructions: i.e. PHI instructions must be first in a basic
5301 block.</p>
5303 <p>For the purposes of the SSA form, the use of each incoming value is deemed to
5304 occur on the edge from the corresponding predecessor block to the current
5306 value on the same edge).</p>
5310 specified by the pair corresponding to the predecessor basic block that
5311 executed just prior to the current block.</p>
5314 <pre>
5315 Loop: ; Infinite loop that counts from 0 on up...
5316 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
5317 %nextindvar = add i32 %indvar, 1
5318 br label %Loop
5319 </pre>
5321 </div>
5323 <!-- _______________________________________________________________________ -->
5326 </div>
5331 <pre>
5332 <result> = select <i>selty</i> <cond>, <ty> <val1>, <ty> <val2> <i>; yields ty</i>
5335 </pre>
5339 condition, without branching.</p>
5344 values indicating the condition, and two values of the
5346 vectors and the condition is a scalar, then entire vectors are selected, not
5347 individual elements.</p>
5351 first value argument; otherwise, it returns the second value argument.</p>
5353 <p>If the condition is a vector of i1, then the value arguments must be vectors
5354 of the same size, and the selection is done element by element.</p>
5357 <pre>
5359 </pre>
5361 <p>Note that the code generator does not yet support conditions
5362 with vector type.</p>
5364 </div>
5366 <!-- _______________________________________________________________________ -->
5369 </div>
5374 <pre>
5375 <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>]
5376 </pre>
5384 <ol>
5386 access any allocas or varargs in the caller. Note that calls may be
5387 marked "tail" even if they do not occur before
5389 present, the function call is eligible for tail call optimization,
5391 optimized into a jump</a>. The code generator may optimize calls marked
5393 sibling call optimization</a> when the caller and callee have
5394 matching signatures, or 2) forced tail call optimization when the
5395 following extra requirements are met:
5396 <ul>
5397 <li>Caller and callee both have the calling
5399 <li>The call is in tail position (ret immediately follows call and ret
5400 uses value of call or is void).</li>
5404 constraints are met.</a></li>
5405 </ul>
5406 </li>
5409 convention</a> the call should use. If none is specified, the call
5410 defaults to using C calling conventions. The calling convention of the
5411 call must match the calling convention of the target function, or else the
5412 behavior is undefined.</li>
5419 type of the return value. Functions that return no value are marked
5423 being invoked. The argument types must match the types implied by this
5424 signature. This type can be omitted if the function is not varargs and if
5425 the function type does not return a pointer to a function.</li>
5428 be invoked. In most cases, this is a direct function invocation, but
5430 to function value.</li>
5433 signature argument types and parameter attributes. All arguments must be
5435 signature indicates the function accepts a variable number of arguments,
5436 the extra arguments can be specified.</li>
5441 </ol>
5445 a specified function, with its incoming arguments bound to the specified
5447 function, control flow continues with the instruction after the function
5448 call, and the return value of the function is bound to the result
5449 argument.</p>
5452 <pre>
5453 %retval = call i32 @test(i32 %argc)
5457 call void %foo(i8 97 signext)
5459 %struct.A = type { i32, i8 }
5465 </pre>
5467 <p>llvm treats calls to some functions with names and arguments that match the
5468 standard C99 library as being the C99 library functions, and may perform
5469 optimizations or generate code for them under that assumption. This is
5470 something we'd like to change in the future to provide better support for
5471 freestanding environments and non-C-based languages.</p>
5473 </div>
5475 <!-- _______________________________________________________________________ -->
5478 </div>
5483 <pre>
5485 </pre>
5489 the "variable argument" area of a function call. It is used to implement the
5494 argument. It returns a value of the specified argument type and increments
5501 to the next argument. For more information, see the variable argument
5504 <p>It is legal for this instruction to be called in a function which does not
5506 function.</p>
5510 argument.</p>
5515 <p>Note that the code generator does not yet fully support va_arg on many
5516 targets. Also, it does not currently support va_arg with aggregate types on
5517 any target.</p>
5519 </div>
5521 <!-- *********************************************************************** -->
5523 <!-- *********************************************************************** -->
5528 well known names and semantics and are required to follow certain
5529 restrictions. Overall, these intrinsics represent an extension mechanism for
5530 the LLVM language that does not require changing all of the transformations
5531 in LLVM when adding to the language (or the bitcode reader/writer, the
5532 parser, etc...).</p>
5535 prefix is reserved in LLVM for intrinsic names; thus, function names may not
5536 begin with this prefix. Intrinsic functions must always be external
5537 functions: you cannot define the body of intrinsic functions. Intrinsic
5538 functions may only be used in call or invoke instructions: it is illegal to
5539 take the address of an intrinsic function. Additionally, because intrinsic
5540 functions are part of the LLVM language, it is required if any are added that
5541 they be documented here.</p>
5543 <p>Some intrinsic functions can be overloaded, i.e., the intrinsic represents a
5544 family of functions that perform the same operation but on different data
5545 types. Because LLVM can represent over 8 million different integer types,
5546 overloading is used commonly to allow an intrinsic function to operate on any
5547 integer type. One or more of the argument types or the result type can be
5548 overloaded to accept any integer type. Argument types may also be defined as
5549 exactly matching a previous argument's type or the result type. This allows
5550 an intrinsic function which accepts multiple arguments, but needs all of them
5551 to be of the same type, to only be overloaded with respect to a single
5552 argument or the result.</p>
5554 <p>Overloaded intrinsics will have the names of its overloaded argument types
5555 encoded into its function name, each preceded by a period. Only those types
5556 which are overloaded result in a name suffix. Arguments whose type is matched
5558 can take an integer of any width and returns an integer of exactly the same
5559 integer width. This leads to a family of functions such as
5561 %val)</tt>. Only one type, the return type, is overloaded, and only one type
5562 suffix is required. Because the argument's type is matched against the return
5563 type, it does not require its own name suffix.</p>
5565 <p>To learn how to add an intrinsic function, please see the
5568 </div>
5570 <!-- ======================================================================= -->
5573 </div>
5577 <p>Variable argument support is defined in LLVM with
5579 intrinsic functions. These functions are related to the similarly named
5582 <p>All of these functions operate on arguments that use a target-specific value
5583 type "<tt>va_list</tt>". The LLVM assembly language reference manual does
5584 not define what this type is, so all transformations should be prepared to
5585 handle these functions regardless of the type used.</p>
5588 instruction and the variable argument handling intrinsic functions are
5589 used.</p>
5592 define i32 @test(i32 %X, ...) {
5593 ; Initialize variable argument processing
5594 %ap = alloca i8*
5595 %ap2 = bitcast i8** %ap to i8*
5596 call void @llvm.va_start(i8* %ap2)
5598 ; Read a single integer argument
5599 %tmp = va_arg i8** %ap, i32
5601 ; Demonstrate usage of llvm.va_copy and llvm.va_end
5602 %aq = alloca i8*
5603 %aq2 = bitcast i8** %aq to i8*
5604 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
5605 call void @llvm.va_end(i8* %aq2)
5607 ; Stop processing of arguments.
5608 call void @llvm.va_end(i8* %ap2)
5609 ret i32 %tmp
5610 }
5612 declare void @llvm.va_start(i8*)
5613 declare void @llvm.va_copy(i8*, i8*)
5614 declare void @llvm.va_end(i8*)
5615 </pre>
5617 </div>
5619 <!-- _______________________________________________________________________ -->
5622 </div>
5628 <pre>
5630 </pre>
5641 macro available in C. In a target-dependent way, it initializes
5645 need to know the last argument of the function as the compiler can figure
5646 that out.</p>
5648 </div>
5650 <!-- _______________________________________________________________________ -->
5653 </div>
5658 <pre>
5660 </pre>
5664 which has been initialized previously
5673 macro available in C. In a target-dependent way, it destroys
5679 </div>
5681 <!-- _______________________________________________________________________ -->
5684 </div>
5689 <pre>
5691 </pre>
5695 from the source argument list to the destination argument list.</p>
5700 from.</p>
5704 macro available in C. In a target-dependent way, it copies the
5706 element. This intrinsic is necessary because
5708 arbitrarily complex and require, for example, memory allocation.</p>
5710 </div>
5712 <!-- ======================================================================= -->
5715 </div>
5720 Collection</a> (GC) requires the implementation and generation of these
5722 roots on the stack</a>, as well as garbage collector implementations that
5724 barriers. Front-ends for type-safe garbage collected languages should generate
5725 these intrinsics to make use of the LLVM garbage collectors. For more details,
5729 <p>The garbage collection intrinsics only operate on objects in the generic
5730 address space (address space zero).</p>
5732 </div>
5734 <!-- _______________________________________________________________________ -->
5737 </div>
5742 <pre>
5743 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
5744 </pre>
5748 the code generator, and allows some metadata to be associated with it.</p>
5751 <p>The first argument specifies the address of a stack object that contains the
5752 root pointer. The second pointer (which must be either a constant or a
5753 global value address) contains the meta-data to be associated with the
5754 root.</p>
5758 location. At compile-time, the code generator generates information to allow
5763 </div>
5765 <!-- _______________________________________________________________________ -->
5768 </div>
5773 <pre>
5774 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
5775 </pre>
5779 locations, allowing garbage collector implementations that require read
5780 barriers.</p>
5783 <p>The second argument is the address to read from, which should be an address
5784 allocated from the garbage collector. The first object is a pointer to the
5785 start of the referenced object, if needed by the language runtime (otherwise
5786 null).</p>
5790 instruction, but may be replaced with substantially more complex code by the
5795 </div>
5797 <!-- _______________________________________________________________________ -->
5800 </div>
5805 <pre>
5806 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
5807 </pre>
5811 locations, allowing garbage collector implementations that require write
5812 barriers (such as generational or reference counting collectors).</p>
5815 <p>The first argument is the reference to store, the second is the start of the
5816 object to store it to, and the third is the address of the field of Obj to
5817 store to. If the runtime does not require a pointer to the object, Obj may
5818 be null.</p>
5822 instruction, but may be replaced with substantially more complex code by the
5827 </div>
5829 <!-- ======================================================================= -->
5832 </div>
5836 <p>These intrinsics are provided by LLVM to expose special features that may
5837 only be implemented with code generator support.</p>
5839 </div>
5841 <!-- _______________________________________________________________________ -->
5844 </div>
5849 <pre>
5851 </pre>
5855 target-specific value indicating the return address of the current function
5856 or one of its callers.</p>
5859 <p>The argument to this intrinsic indicates which function to return the address
5860 for. Zero indicates the calling function, one indicates its caller, etc.
5865 indicating the return address of the specified call frame, or zero if it
5866 cannot be identified. The value returned by this intrinsic is likely to be
5867 incorrect or 0 for arguments other than zero, so it should only be used for
5868 debugging purposes.</p>
5870 <p>Note that calling this intrinsic does not prevent function inlining or other
5871 aggressive transformations, so the value returned may not be that of the
5872 obvious source-language caller.</p>
5874 </div>
5876 <!-- _______________________________________________________________________ -->
5879 </div>
5884 <pre>
5886 </pre>
5890 target-specific frame pointer value for the specified stack frame.</p>
5893 <p>The argument to this intrinsic indicates which function to return the frame
5894 pointer for. Zero indicates the calling function, one indicates its caller,
5899 indicating the frame address of the specified call frame, or zero if it
5900 cannot be identified. The value returned by this intrinsic is likely to be
5901 incorrect or 0 for arguments other than zero, so it should only be used for
5902 debugging purposes.</p>
5904 <p>Note that calling this intrinsic does not prevent function inlining or other
5905 aggressive transformations, so the value returned may not be that of the
5906 obvious source-language caller.</p>
5908 </div>
5910 <!-- _______________________________________________________________________ -->
5913 </div>
5918 <pre>
5919 declare i8* @llvm.stacksave()
5920 </pre>
5924 of the function stack, for use
5926 useful for implementing language features like scoped automatic variable
5927 sized arrays in C99.</p>
5930 <p>This intrinsic returns a opaque pointer value that can be passed
5938 </div>
5940 <!-- _______________________________________________________________________ -->
5943 </div>
5948 <pre>
5949 declare void @llvm.stackrestore(i8* %ptr)
5950 </pre>
5954 the function stack to the state it was in when the
5956 executed. This is useful for implementing language features like scoped
5957 automatic variable sized arrays in C99.</p>
5960 <p>See the description
5963 </div>
5965 <!-- _______________________________________________________________________ -->
5968 </div>
5973 <pre>
5975 </pre>
5979 insert a prefetch instruction if supported; otherwise, it is a noop.
5980 Prefetches have no effect on the behavior of the program but can change its
5981 performance characteristics.</p>
5991 <p>This intrinsic does not modify the behavior of the program. In particular,
5992 prefetches cannot trap and do not produce a value. On targets that support
5993 this intrinsic, the prefetch can provide hints to the processor cache for
5994 better performance.</p>
5996 </div>
5998 <!-- _______________________________________________________________________ -->
6001 </div>
6006 <pre>
6008 </pre>
6012 Counter (PC) in a region of code to simulators and other tools. The method
6013 is target specific, but it is expected that the marker will use exported
6014 symbols to transmit the PC of the marker. The marker makes no guarantees
6015 that it will remain with any specific instruction after optimizations. It is
6016 possible that the presence of a marker will inhibit optimizations. The
6017 intended use is to be inserted after optimizations to allow correlations of
6018 simulation runs.</p>
6024 <p>This intrinsic does not modify the behavior of the program. Backends that do
6025 not support this intrinsic may ignore it.</p>
6027 </div>
6029 <!-- _______________________________________________________________________ -->
6032 </div>
6037 <pre>
6038 declare i64 @llvm.readcyclecounter()
6039 </pre>
6043 counter register (or similar low latency, high accuracy clocks) on those
6044 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6045 should map to RPCC. As the backing counters overflow quickly (on the order
6049 <p>When directly supported, reading the cycle counter should not modify any
6050 memory. Implementations are allowed to either return a application specific
6051 value or a system wide value. On backends without support, this is lowered
6054 </div>
6056 <!-- ======================================================================= -->
6059 </div>
6063 <p>LLVM provides intrinsics for a few important standard C library functions.
6064 These intrinsics allow source-language front-ends to pass information about
6065 the alignment of the pointer arguments to the code generator, providing
6066 opportunity for more efficient code generation.</p>
6068 </div>
6070 <!-- _______________________________________________________________________ -->
6073 </div>
6079 integer bit width and for different address spaces. Not all targets support
6080 all bit widths however.</p>
6082 <pre>
6087 </pre>
6091 source location to the destination location.</p>
6094 intrinsics do not return a value, takes extra alignment/isvolatile arguments
6095 and the pointers can be in specified address spaces.</p>
6099 <p>The first argument is a pointer to the destination, the second is a pointer
6100 to the source. The third argument is an integer argument specifying the
6101 number of bytes to copy, the fourth argument is the alignment of the
6102 source and destination locations, and the fifth is a boolean indicating a
6103 volatile access.</p>
6106 then the caller guarantees that both the source and destination pointers are
6107 aligned to that boundary.</p>
6111 The detailed access behavior is not very cleanly specified and it is unwise
6112 to depend on it.</p>
6117 source location to the destination location, which are not allowed to
6118 overlap. It copies "len" bytes of memory over. If the argument is known to
6119 be aligned to some boundary, this can be specified as the fourth argument,
6122 </div>
6124 <!-- _______________________________________________________________________ -->
6127 </div>
6132 <p>This is an overloaded intrinsic. You can use llvm.memmove on any integer bit
6133 width and for different address space. Not all targets support all bit
6134 widths however.</p>
6136 <pre>
6141 </pre>
6145 source location to the destination location. It is similar to the
6147 overlap.</p>
6150 intrinsics do not return a value, takes extra alignment/isvolatile arguments
6151 and the pointers can be in specified address spaces.</p>
6155 <p>The first argument is a pointer to the destination, the second is a pointer
6156 to the source. The third argument is an integer argument specifying the
6157 number of bytes to copy, the fourth argument is the alignment of the
6158 source and destination locations, and the fifth is a boolean indicating a
6159 volatile access.</p>
6162 then the caller guarantees that the source and destination pointers are
6163 aligned to that boundary.</p>
6167 The detailed access behavior is not very cleanly specified and it is unwise
6168 to depend on it.</p>
6173 source location to the destination location, which may overlap. It copies
6174 "len" bytes of memory over. If the argument is known to be aligned to some
6175 boundary, this can be specified as the fourth argument, otherwise it should
6178 </div>
6180 <!-- _______________________________________________________________________ -->
6183 </div>
6188 <p>This is an overloaded intrinsic. You can use llvm.memset on any integer bit
6189 width and for different address spaces. However, not all targets support all
6190 bit widths.</p>
6192 <pre>
6197 </pre>
6201 particular byte value.</p>
6204 intrinsic does not return a value and takes extra alignment/volatile
6205 arguments. Also, the destination can be in an arbitrary address space.</p>
6208 <p>The first argument is a pointer to the destination to fill, the second is the
6209 byte value with which to fill it, the third argument is an integer argument
6210 specifying the number of bytes to fill, and the fourth argument is the known
6211 alignment of the destination location.</p>
6214 then the caller guarantees that the destination pointer is aligned to that
6215 boundary.</p>
6219 The detailed access behavior is not very cleanly specified and it is unwise
6220 to depend on it.</p>
6224 at the destination location. If the argument is known to be aligned to some
6225 boundary, this can be specified as the fourth argument, otherwise it should
6228 </div>
6230 <!-- _______________________________________________________________________ -->
6233 </div>
6239 floating point or vector of floating point type. Not all targets support all
6240 types however.</p>
6242 <pre>
6243 declare float @llvm.sqrt.f32(float %Val)
6244 declare double @llvm.sqrt.f64(double %Val)
6245 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
6246 declare fp128 @llvm.sqrt.f128(fp128 %Val)
6247 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
6248 </pre>
6254 behavior for negative numbers other than -0.0 (which allows for better
6255 optimization, because there is no need to worry about errno being
6259 <p>The argument and return value are floating point numbers of the same
6260 type.</p>
6263 <p>This function returns the sqrt of the specified operand if it is a
6264 nonnegative floating point number.</p>
6266 </div>
6268 <!-- _______________________________________________________________________ -->
6271 </div>
6277 floating point or vector of floating point type. Not all targets support all
6278 types however.</p>
6280 <pre>
6281 declare float @llvm.powi.f32(float %Val, i32 %power)
6282 declare double @llvm.powi.f64(double %Val, i32 %power)
6283 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
6284 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
6285 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
6286 </pre>
6290 specified (positive or negative) power. The order of evaluation of
6291 multiplications is not defined. When a vector of floating point type is
6292 used, the second argument remains a scalar integer value.</p>
6295 <p>The second argument is an integer power, and the first is a value to raise to
6296 that power.</p>
6299 <p>This function returns the first value raised to the second power with an
6300 unspecified sequence of rounding operations.</p>
6302 </div>
6304 <!-- _______________________________________________________________________ -->
6307 </div>
6313 floating point or vector of floating point type. Not all targets support all
6314 types however.</p>
6316 <pre>
6317 declare float @llvm.sin.f32(float %Val)
6318 declare double @llvm.sin.f64(double %Val)
6319 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
6320 declare fp128 @llvm.sin.f128(fp128 %Val)
6321 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
6322 </pre>
6328 <p>The argument and return value are floating point numbers of the same
6329 type.</p>
6332 <p>This function returns the sine of the specified operand, returning the same
6334 in the same way.</p>
6336 </div>
6338 <!-- _______________________________________________________________________ -->
6341 </div>
6347 floating point or vector of floating point type. Not all targets support all
6348 types however.</p>
6350 <pre>
6351 declare float @llvm.cos.f32(float %Val)
6352 declare double @llvm.cos.f64(double %Val)
6353 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
6354 declare fp128 @llvm.cos.f128(fp128 %Val)
6355 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
6356 </pre>
6362 <p>The argument and return value are floating point numbers of the same
6363 type.</p>
6366 <p>This function returns the cosine of the specified operand, returning the same
6368 in the same way.</p>
6370 </div>
6372 <!-- _______________________________________________________________________ -->
6375 </div>
6381 floating point or vector of floating point type. Not all targets support all
6382 types however.</p>
6384 <pre>
6385 declare float @llvm.pow.f32(float %Val, float %Power)
6386 declare double @llvm.pow.f64(double %Val, double %Power)
6387 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
6388 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
6389 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
6390 </pre>
6394 specified (positive or negative) power.</p>
6397 <p>The second argument is a floating point power, and the first is a value to
6398 raise to that power.</p>
6401 <p>This function returns the first value raised to the second power, returning
6403 conditions in the same way.</p>
6405 </div>
6407 <!-- ======================================================================= -->
6410 </div>
6414 <p>LLVM provides intrinsics for a few important bit manipulation operations.
6415 These allow efficient code generation for some algorithms.</p>
6417 </div>
6419 <!-- _______________________________________________________________________ -->
6422 </div>
6427 <p>This is an overloaded intrinsic function. You can use bswap on any integer
6430 <pre>
6434 </pre>
6438 values with an even number of bytes (positive multiple of 16 bits). These
6439 are useful for performing operations on data that is not in the target's
6440 native byte order.</p>
6444 and low byte of the input i16 swapped. Similarly,
6450 more, respectively).</p>
6452 </div>
6454 <!-- _______________________________________________________________________ -->
6457 </div>
6462 <p>This is an overloaded intrinsic. You can use llvm.ctpop on any integer bit
6463 width. Not all targets support all bit widths however.</p>
6465 <pre>
6471 </pre>
6475 in a value.</p>
6478 <p>The only argument is the value to be counted. The argument may be of any
6479 integer type. The return type must match the argument type.</p>
6484 </div>
6486 <!-- _______________________________________________________________________ -->
6489 </div>
6495 integer bit width. Not all targets support all bit widths however.</p>
6497 <pre>
6503 </pre>
6507 leading zeros in a variable.</p>
6510 <p>The only argument is the value to be counted. The argument may be of any
6511 integer type. The return type must match the argument type.</p>
6515 zeros in a variable. If the src == 0 then the result is the size in bits of
6518 </div>
6520 <!-- _______________________________________________________________________ -->
6523 </div>
6529 integer bit width. Not all targets support all bit widths however.</p>
6531 <pre>
6537 </pre>
6541 trailing zeros.</p>
6544 <p>The only argument is the value to be counted. The argument may be of any
6545 integer type. The return type must match the argument type.</p>
6549 zeros in a variable. If the src == 0 then the result is the size in bits of
6552 </div>
6554 <!-- ======================================================================= -->
6557 </div>
6563 </div>
6565 <!-- _______________________________________________________________________ -->
6568 </div>
6574 on any integer bit width.</p>
6576 <pre>
6577 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
6578 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
6579 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
6580 </pre>
6584 a signed addition of the two arguments, and indicate whether an overflow
6585 occurred during the signed summation.</p>
6588 <p>The arguments (%a and %b) and the first element of the result structure may
6589 be of integer types of any bit width, but they must have the same bit
6590 width. The second element of the result structure must be of
6592 undergo signed addition.</p>
6596 a signed addition of the two variables. They return a structure — the
6597 first element of which is the signed summation, and the second element of
6598 which is a bit specifying if the signed summation resulted in an
6599 overflow.</p>
6602 <pre>
6603 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
6604 %sum = extractvalue {i32, i1} %res, 0
6605 %obit = extractvalue {i32, i1} %res, 1
6606 br i1 %obit, label %overflow, label %normal
6607 </pre>
6609 </div>
6611 <!-- _______________________________________________________________________ -->
6614 </div>
6620 on any integer bit width.</p>
6622 <pre>
6623 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
6624 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
6625 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
6626 </pre>
6630 an unsigned addition of the two arguments, and indicate whether a carry
6631 occurred during the unsigned summation.</p>
6634 <p>The arguments (%a and %b) and the first element of the result structure may
6635 be of integer types of any bit width, but they must have the same bit
6636 width. The second element of the result structure must be of
6638 undergo unsigned addition.</p>
6642 an unsigned addition of the two arguments. They return a structure —
6643 the first element of which is the sum, and the second element of which is a
6644 bit specifying if the unsigned summation resulted in a carry.</p>
6647 <pre>
6648 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
6649 %sum = extractvalue {i32, i1} %res, 0
6650 %obit = extractvalue {i32, i1} %res, 1
6651 br i1 %obit, label %carry, label %normal
6652 </pre>
6654 </div>
6656 <!-- _______________________________________________________________________ -->
6659 </div>
6665 on any integer bit width.</p>
6667 <pre>
6668 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
6669 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
6670 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
6671 </pre>
6675 a signed subtraction of the two arguments, and indicate whether an overflow
6676 occurred during the signed subtraction.</p>
6679 <p>The arguments (%a and %b) and the first element of the result structure may
6680 be of integer types of any bit width, but they must have the same bit
6681 width. The second element of the result structure must be of
6683 undergo signed subtraction.</p>
6687 a signed subtraction of the two arguments. They return a structure —
6688 the first element of which is the subtraction, and the second element of
6689 which is a bit specifying if the signed subtraction resulted in an
6690 overflow.</p>
6693 <pre>
6694 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
6695 %sum = extractvalue {i32, i1} %res, 0
6696 %obit = extractvalue {i32, i1} %res, 1
6697 br i1 %obit, label %overflow, label %normal
6698 </pre>
6700 </div>
6702 <!-- _______________________________________________________________________ -->
6705 </div>
6711 on any integer bit width.</p>
6713 <pre>
6714 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
6715 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
6716 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
6717 </pre>
6721 an unsigned subtraction of the two arguments, and indicate whether an
6722 overflow occurred during the unsigned subtraction.</p>
6725 <p>The arguments (%a and %b) and the first element of the result structure may
6726 be of integer types of any bit width, but they must have the same bit
6727 width. The second element of the result structure must be of
6729 undergo unsigned subtraction.</p>
6733 an unsigned subtraction of the two arguments. They return a structure —
6734 the first element of which is the subtraction, and the second element of
6735 which is a bit specifying if the unsigned subtraction resulted in an
6736 overflow.</p>
6739 <pre>
6740 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
6741 %sum = extractvalue {i32, i1} %res, 0
6742 %obit = extractvalue {i32, i1} %res, 1
6743 br i1 %obit, label %overflow, label %normal
6744 </pre>
6746 </div>
6748 <!-- _______________________________________________________________________ -->
6751 </div>
6757 on any integer bit width.</p>
6759 <pre>
6760 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
6761 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
6762 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
6763 </pre>
6768 a signed multiplication of the two arguments, and indicate whether an
6769 overflow occurred during the signed multiplication.</p>
6772 <p>The arguments (%a and %b) and the first element of the result structure may
6773 be of integer types of any bit width, but they must have the same bit
6774 width. The second element of the result structure must be of
6776 undergo signed multiplication.</p>
6780 a signed multiplication of the two arguments. They return a structure —
6781 the first element of which is the multiplication, and the second element of
6782 which is a bit specifying if the signed multiplication resulted in an
6783 overflow.</p>
6786 <pre>
6787 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
6788 %sum = extractvalue {i32, i1} %res, 0
6789 %obit = extractvalue {i32, i1} %res, 1
6790 br i1 %obit, label %overflow, label %normal
6791 </pre>
6793 </div>
6795 <!-- _______________________________________________________________________ -->
6798 </div>
6804 on any integer bit width.</p>
6806 <pre>
6807 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
6808 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
6809 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
6810 </pre>
6814 a unsigned multiplication of the two arguments, and indicate whether an
6815 overflow occurred during the unsigned multiplication.</p>
6818 <p>The arguments (%a and %b) and the first element of the result structure may
6819 be of integer types of any bit width, but they must have the same bit
6820 width. The second element of the result structure must be of
6822 undergo unsigned multiplication.</p>
6826 an unsigned multiplication of the two arguments. They return a structure
6827 — the first element of which is the multiplication, and the second
6828 element of which is a bit specifying if the unsigned multiplication resulted
6829 in an overflow.</p>
6832 <pre>
6833 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
6834 %sum = extractvalue {i32, i1} %res, 0
6835 %obit = extractvalue {i32, i1} %res, 1
6836 br i1 %obit, label %overflow, label %normal
6837 </pre>
6839 </div>
6841 <!-- ======================================================================= -->
6844 </div>
6848 <p>Half precision floating point is a storage-only format. This means that it is
6849 a dense encoding (in memory) but does not support computation in the
6850 format.</p>
6852 <p>This means that code must first load the half-precision floating point
6853 value as an i16, then convert it to float with <a
6855 Computation can then be performed on the float value (including extending to
6856 double etc). To store the value back to memory, it is first converted to
6857 float if needed, then converted to i16 with
6859 storing as an i16 value.</p>
6860 </div>
6862 <!-- _______________________________________________________________________ -->
6865 </div>
6870 <pre>
6871 declare i16 @llvm.convert.to.fp16(f32 %a)
6872 </pre>
6876 a conversion from single precision floating point format to half precision
6877 floating point format.</p>
6880 <p>The intrinsic function contains single argument - the value to be
6881 converted.</p>
6885 a conversion from single precision floating point format to half precision
6887 contains the converted number.</p>
6890 <pre>
6891 %res = call i16 @llvm.convert.to.fp16(f32 %a)
6892 store i16 %res, i16* @x, align 2
6893 </pre>
6895 </div>
6897 <!-- _______________________________________________________________________ -->
6900 </div>
6905 <pre>
6906 declare f32 @llvm.convert.from.fp16(i16 %a)
6907 </pre>
6911 a conversion from half precision floating point format to single precision
6912 floating point format.</p>
6915 <p>The intrinsic function contains single argument - the value to be
6916 converted.</p>
6920 conversion from half single precision floating point format to single
6921 precision floating point format. The input half-float value is represented by
6925 <pre>
6926 %a = load i16* @x, align 2
6927 %res = call f32 @llvm.convert.from.fp16(i16 %a)
6928 </pre>
6930 </div>
6932 <!-- ======================================================================= -->
6935 </div>
6940 prefix), are described in
6944 </div>
6946 <!-- ======================================================================= -->
6949 </div>
6953 <p>The LLVM exception handling intrinsics (which all start with
6958 </div>
6960 <!-- ======================================================================= -->
6963 </div>
6967 <p>This intrinsic makes it possible to excise one parameter, marked with
6969 The result is a callable
6970 function pointer lacking the nest parameter - the caller does not need to
6971 provide a value for it. Instead, the value to use is stored in advance in a
6972 "trampoline", a block of memory usually allocated on the stack, which also
6973 contains code to splice the nest value into the argument list. This is used
6974 to implement the GCC nested function address extension.</p>
6976 <p>For example, if the function is
6979 follows:</p>
6984 %p = call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8* nest , i32, i32)* @f to i8*), i8* %nval)
6985 %fp = bitcast i8* %p to i32 (i32, i32)*
6986 </pre>
6991 </div>
6993 <!-- _______________________________________________________________________ -->
6996 </div>
7001 <pre>
7003 </pre>
7007 function pointer suitable for executing it.</p>
7012 sufficiently aligned block of memory; this memory is written to by the
7013 intrinsic. Note that the size and the alignment are target-specific - LLVM
7014 currently provides no portable way of determining them, so a front-end that
7015 generates this intrinsic needs to have some target-specific knowledge.
7021 dependent code, turning it into a function. A pointer to this function is
7023 function pointer type</a> before being called. The new function's signature
7026 is allowed, and it must be of pointer type. Calling the new function is
7031 returned function pointer is undefined.</p>
7033 </div>
7035 <!-- ======================================================================= -->
7038 </div>
7043 hardware constructs for atomic operations and memory synchronization. This
7044 provides an interface to the hardware, not an interface to the programmer. It
7045 is aimed at a low enough level to allow any programming models or APIs
7046 (Application Programming Interfaces) which need atomic behaviors to map
7047 cleanly onto it. It is also modeled primarily on hardware behavior. Just as
7048 hardware provides a "universal IR" for source languages, it also provides a
7049 starting point for developing a "universal" atomic operation and
7050 synchronization IR.</p>
7053 software transaction memory systems, atomic primitives, and intrinsic
7054 functions as found in BSD, GNU libc, atomic_ops, APR, and other system and
7055 application libraries. The hardware interface provided by LLVM should allow
7056 a clean implementation of all of these APIs and parallel programming models.
7057 No one model or paradigm should be selected above others unless the hardware
7058 itself ubiquitously does so.</p>
7060 </div>
7062 <!-- _______________________________________________________________________ -->
7065 </div>
7068 <pre>
7069 declare void @llvm.memory.barrier(i1 <ll>, i1 <ls>, i1 <sl>, i1 <ss>, i1 <device>)
7070 </pre>
7074 specific pairs of memory access types.</p>
7078 The first four arguments enables a specific barrier as listed below. The
7079 fifth argument specifies that the barrier applies to io or device or uncached
7080 memory.</p>
7082 <ul>
7088 </ul>
7091 <p>This intrinsic causes the system to enforce some ordering constraints upon
7092 the loads and stores of the program. This barrier does not
7095 and store operations (f.ex. load-load, or store-load), all of the first
7096 operations preceding the barrier will complete before any of the second
7097 operations succeeding the barrier begin. Specifically the semantics for each
7098 pairing is as follows:</p>
7100 <ul>
7102 after the barrier begins.</li>
7104 store after the barrier begins.</li>
7106 store after the barrier begins.</li>
7108 load after the barrier begins.</li>
7109 </ul>
7112 is enabled in a single memory barrier intrinsic.</p>
7114 <p>Backends may implement stronger barriers than those requested when they do
7115 not support as fine grained a barrier as requested. Some architectures do
7116 not need all types of barriers and on such architectures, these become
7117 noops.</p>
7120 <pre>
7121 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7122 %ptr = bitcast i8* %mallocP to i32*
7123 store i32 4, %ptr
7126 call void @llvm.memory.barrier(i1 false, i1 true, i1 false, i1 false)
7129 </pre>
7131 </div>
7133 <!-- _______________________________________________________________________ -->
7136 </div>
7142 any integer bit width and for different address spaces. Not all targets
7143 support all bit widths however.</p>
7145 <pre>
7147 declare i16 @llvm.atomic.cmp.swap.i16.p0i16(i16* <ptr>, i16 <cmp>, i16 <val>)
7148 declare i32 @llvm.atomic.cmp.swap.i32.p0i32(i32* <ptr>, i32 <cmp>, i32 <val>)
7149 declare i64 @llvm.atomic.cmp.swap.i64.p0i64(i64* <ptr>, i64 <cmp>, i64 <val>)
7150 </pre>
7153 <p>This loads a value in memory and compares it to a given value. If they are
7154 equal, it stores a new value into the memory.</p>
7160 this integer type. While any bit width integer may be used, targets may only
7161 lower representations they support in hardware.</p>
7164 <p>This entire intrinsic must be executed atomically. It first loads the value
7167 memory. The loaded value is yielded in all cases. This provides the
7168 equivalent of an atomic compare-and-swap operation within the SSA
7169 framework.</p>
7172 <pre>
7173 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7174 %ptr = bitcast i8* %mallocP to i32*
7175 store i32 4, %ptr
7178 %result1 = call i32 @llvm.atomic.cmp.swap.i32.p0i32(i32* %ptr, i32 4, %val1)
7184 %result2 = call i32 @llvm.atomic.cmp.swap.i32.p0i32(i32* %ptr, i32 5, %val2)
7189 </pre>
7191 </div>
7193 <!-- _______________________________________________________________________ -->
7196 </div>
7201 integer bit width. Not all targets support all bit widths however.</p>
7203 <pre>
7208 </pre>
7219 integer type. The targets may only lower integer representations they
7220 support.</p>
7225 equivalent of an atomic swap operation within the SSA framework.</p>
7228 <pre>
7229 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7230 %ptr = bitcast i8* %mallocP to i32*
7231 store i32 4, %ptr
7234 %result1 = call i32 @llvm.atomic.swap.i32.p0i32(i32* %ptr, i32 %val1)
7240 %result2 = call i32 @llvm.atomic.swap.i32.p0i32(i32* %ptr, i32 %val2)
7245 </pre>
7247 </div>
7249 <!-- _______________________________________________________________________ -->
7253 </div>
7259 any integer bit width. Not all targets support all bit widths however.</p>
7261 <pre>
7266 </pre>
7273 <p>The intrinsic takes two arguments, the first a pointer to an integer value
7274 and the second an integer value. The result is also an integer value. These
7275 integer types can have any bit width, but they must all have the same bit
7276 width. The targets may only lower integer representations they support.</p>
7279 <p>This intrinsic does a series of operations atomically. It first loads the
7284 <pre>
7285 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7286 %ptr = bitcast i8* %mallocP to i32*
7287 store i32 4, %ptr
7288 %result1 = call i32 @llvm.atomic.load.add.i32.p0i32(i32* %ptr, i32 4)
7290 %result2 = call i32 @llvm.atomic.load.add.i32.p0i32(i32* %ptr, i32 2)
7292 %result3 = call i32 @llvm.atomic.load.add.i32.p0i32(i32* %ptr, i32 5)
7295 </pre>
7297 </div>
7299 <!-- _______________________________________________________________________ -->
7303 </div>
7309 any integer bit width and for different address spaces. Not all targets
7310 support all bit widths however.</p>
7312 <pre>
7317 </pre>
7324 <p>The intrinsic takes two arguments, the first a pointer to an integer value
7325 and the second an integer value. The result is also an integer value. These
7326 integer types can have any bit width, but they must all have the same bit
7327 width. The targets may only lower integer representations they support.</p>
7330 <p>This intrinsic does a series of operations atomically. It first loads the
7336 <pre>
7337 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7338 %ptr = bitcast i8* %mallocP to i32*
7339 store i32 8, %ptr
7340 %result1 = call i32 @llvm.atomic.load.sub.i32.p0i32(i32* %ptr, i32 4)
7342 %result2 = call i32 @llvm.atomic.load.sub.i32.p0i32(i32* %ptr, i32 2)
7344 %result3 = call i32 @llvm.atomic.load.sub.i32.p0i32(i32* %ptr, i32 5)
7347 </pre>
7349 </div>
7351 <!-- _______________________________________________________________________ -->
7357 </div>
7362 <p>These are overloaded intrinsics. You can
7365 bit width and for different address spaces. Not all targets support all bit
7366 widths however.</p>
7368 <pre>
7373 </pre>
7375 <pre>
7380 </pre>
7382 <pre>
7387 </pre>
7389 <pre>
7394 </pre>
7402 <p>These intrinsics take two arguments, the first a pointer to an integer value
7403 and the second an integer value. The result is also an integer value. These
7404 integer types can have any bit width, but they must all have the same bit
7405 width. The targets may only lower integer representations they support.</p>
7408 <p>These intrinsics does a series of operations atomically. They first load the
7414 <pre>
7415 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7416 %ptr = bitcast i8* %mallocP to i32*
7417 store i32 0x0F0F, %ptr
7418 %result0 = call i32 @llvm.atomic.load.nand.i32.p0i32(i32* %ptr, i32 0xFF)
7420 %result1 = call i32 @llvm.atomic.load.and.i32.p0i32(i32* %ptr, i32 0xFF)
7422 %result2 = call i32 @llvm.atomic.load.or.i32.p0i32(i32* %ptr, i32 0F)
7424 %result3 = call i32 @llvm.atomic.load.xor.i32.p0i32(i32* %ptr, i32 0F)
7427 </pre>
7429 </div>
7431 <!-- _______________________________________________________________________ -->
7437 </div>
7445 address spaces. Not all targets support all bit widths however.</p>
7447 <pre>
7452 </pre>
7454 <pre>
7459 </pre>
7461 <pre>
7466 </pre>
7468 <pre>
7473 </pre>
7476 <p>These intrinsics takes the signed or unsigned minimum or maximum of
7481 <p>These intrinsics take two arguments, the first a pointer to an integer value
7482 and the second an integer value. The result is also an integer value. These
7483 integer types can have any bit width, but they must all have the same bit
7484 width. The targets may only lower integer representations they support.</p>
7487 <p>These intrinsics does a series of operations atomically. They first load the
7493 <pre>
7494 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7495 %ptr = bitcast i8* %mallocP to i32*
7496 store i32 7, %ptr
7497 %result0 = call i32 @llvm.atomic.load.min.i32.p0i32(i32* %ptr, i32 -2)
7499 %result1 = call i32 @llvm.atomic.load.max.i32.p0i32(i32* %ptr, i32 8)
7501 %result2 = call i32 @llvm.atomic.load.umin.i32.p0i32(i32* %ptr, i32 10)
7503 %result3 = call i32 @llvm.atomic.load.umax.i32.p0i32(i32* %ptr, i32 30)
7506 </pre>
7508 </div>
7511 <!-- ======================================================================= -->
7514 </div>
7518 <p>This class of intrinsics exists to information about the lifetime of memory
7519 objects and ranges where variables are immutable.</p>
7521 </div>
7523 <!-- _______________________________________________________________________ -->
7526 </div>
7531 <pre>
7533 </pre>
7537 object's lifetime.</p>
7540 <p>The first argument is a constant integer representing the size of the
7541 object, or -1 if it is variable sized. The second argument is a pointer to
7542 the object.</p>
7545 <p>This intrinsic indicates that before this point in the code, the value of the
7547 never be used and has an undefined value. A load from the pointer that
7548 precedes this intrinsic can be replaced with
7551 </div>
7553 <!-- _______________________________________________________________________ -->
7556 </div>
7561 <pre>
7563 </pre>
7567 object's lifetime.</p>
7570 <p>The first argument is a constant integer representing the size of the
7571 object, or -1 if it is variable sized. The second argument is a pointer to
7572 the object.</p>
7575 <p>This intrinsic indicates that after this point in the code, the value of the
7577 never be used and has an undefined value. Any stores into the memory object
7578 following this intrinsic may be removed as dead.
7580 </div>
7582 <!-- _______________________________________________________________________ -->
7585 </div>
7590 <pre>
7592 </pre>
7596 a memory object will not change.</p>
7599 <p>The first argument is a constant integer representing the size of the
7600 object, or -1 if it is variable sized. The second argument is a pointer to
7601 the object.</p>
7605 the return value, the referenced memory location is constant and
7606 unchanging.</p>
7608 </div>
7610 <!-- _______________________________________________________________________ -->
7613 </div>
7618 <pre>
7619 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
7620 </pre>
7624 a memory object are mutable.</p>
7628 The second argument is a constant integer representing the size of the
7629 object, or -1 if it is variable sized and the third argument is a pointer
7630 to the object.</p>
7635 </div>
7637 <!-- ======================================================================= -->
7640 </div>
7644 <p>This class of intrinsics is designed to be generic and has no specific
7645 purpose.</p>
7647 </div>
7649 <!-- _______________________________________________________________________ -->
7652 </div>
7657 <pre>
7658 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
7659 </pre>
7665 <p>The first argument is a pointer to a value, the second is a pointer to a
7666 global string, the third is a pointer to a global string which is the source
7667 file name, and the last argument is the line number.</p>
7670 <p>This intrinsic allows annotation of local variables with arbitrary strings.
7671 This can be useful for special purpose optimizations that want to look for
7672 these annotations. These have no other defined use, they are ignored by code
7673 generation and optimization.</p>
7675 </div>
7677 <!-- _______________________________________________________________________ -->
7680 </div>
7686 any integer bit width.</p>
7688 <pre>
7689 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
7690 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
7691 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
7692 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
7693 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
7694 </pre>
7700 <p>The first argument is an integer value (result of some expression), the
7701 second is a pointer to a global string, the third is a pointer to a global
7702 string which is the source file name, and the last argument is the line
7703 number. It returns the value of the first argument.</p>
7706 <p>This intrinsic allows annotations to be put on arbitrary expressions with
7707 arbitrary strings. This can be useful for special purpose optimizations that
7708 want to look for these annotations. These have no other defined use, they
7709 are ignored by code generation and optimization.</p>
7711 </div>
7713 <!-- _______________________________________________________________________ -->
7716 </div>
7721 <pre>
7722 declare void @llvm.trap()
7723 </pre>
7732 <p>This intrinsics is lowered to the target dependent trap instruction. If the
7733 target does not have a trap instruction, this intrinsic will be lowered to
7736 </div>
7738 <!-- _______________________________________________________________________ -->
7741 </div>
7746 <pre>
7748 </pre>
7753 ensure that it is placed on the stack before local variables.</p>
7757 arguments. The first argument is the value loaded from the stack
7759 that has enough space to hold the value of the guard.</p>
7762 <p>This intrinsic causes the prologue/epilogue inserter to force the position of
7764 stack. This is to ensure that if a local variable on the stack is
7765 overwritten, it will destroy the value of the guard. When the function exits,
7766 the guard on the stack is checked against the original guard. If they are
7768 function.</p>
7770 </div>
7772 <!-- _______________________________________________________________________ -->
7775 </div>
7780 <pre>
7783 </pre>
7787 the optimizers to determine at compile time whether a) an operation (like
7788 memcpy) will overflow a buffer that corresponds to an object, or b) that a
7789 runtime check for overflow isn't necessary. An object in this context means
7790 an allocation of a specific class, structure, array, or other object.</p>
7803 compile time.</p>
7805 </div>
7807 <!-- *********************************************************************** -->
7808 <hr>
7809 <address>
7817 Last modified: $Date$
7818 </address>
7820 </body>
7821 </html>