| blob | 381fed5c6d3debcc1cdc02f0544632b25763521e |
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 <!-- *********************************************************************** -->
327 <div>
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 <!-- *********************************************************************** -->
341 <div>
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 <!-- _______________________________________________________________________ -->
363 <h4>
365 </h4>
367 <div>
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 </div>
389 <!-- Describe the typesetting conventions here. -->
391 <!-- *********************************************************************** -->
393 <!-- *********************************************************************** -->
395 <div>
397 <p>LLVM identifiers come in two basic types: global and local. Global
399 character. Local identifiers (register names, types) begin with
401 for identifiers, for different purposes:</p>
403 <ol>
404 <li>Named values are represented as a string of characters with their prefix.
408 other characters in their names can be surrounded with quotes. Special
410 ASCII code for the character in hexadecimal. In this way, any character
411 can be used in a name value, even quotes themselves.</li>
413 <li>Unnamed values are represented as an unsigned numeric value with their
418 </ol>
420 <p>LLVM requires that values start with a prefix for two reasons: Compilers
421 don't need to worry about name clashes with reserved words, and the set of
422 reserved words may be expanded in the future without penalty. Additionally,
423 unnamed identifiers allow a compiler to quickly come up with a temporary
424 variable without having to avoid symbol table conflicts.</p>
426 <p>Reserved words in LLVM are very similar to reserved words in other
427 languages. There are keywords for different opcodes
433 reserved words cannot conflict with variable names, because none of them
436 <p>Here is an example of LLVM code to multiply the integer variable
443 </pre>
449 </pre>
457 </pre>
460 lexical features of LLVM:</p>
462 <ol>
464 line.</li>
466 <li>Unnamed temporaries are created when the result of a computation is not
467 assigned to a named value.</li>
470 </ol>
472 <p>It also shows a convention that we follow in this document. When
473 demonstrating instructions, we will follow an instruction with a comment that
474 defines the type and name of value produced. Comments are shown in italic
475 text.</p>
477 </div>
479 <!-- *********************************************************************** -->
481 <!-- *********************************************************************** -->
482 <div>
483 <!-- ======================================================================= -->
484 <h3>
486 </h3>
488 <div>
491 of the input programs. Each module consists of functions, global variables,
492 and symbol table entries. Modules may be combined together with the LLVM
493 linker, which merges function (and global variable) definitions, resolves
494 forward declarations, and merges symbol table entries. Here is an example of
499 <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>
507 %cast210 = <a href="#i_getelementptr">getelementptr</a> [13 x i8]* @.LC0, i64 0, i64 0 <i>; i8*</i>
512 }
516 !foo = !{!1, null}
517 </pre>
525 <p>In general, a module is made up of a list of global values, where both
526 functions and global variables are global values. Global values are
527 represented by a pointer to a memory location (in this case, a pointer to an
528 array of char, and a pointer to a function), and have one of the
531 </div>
533 <!-- ======================================================================= -->
534 <h3>
536 </h3>
538 <div>
540 <p>All Global Variables and Functions have one of the following types of
541 linkage:</p>
543 <dl>
546 by objects in the current module. In particular, linking code into a
547 module with an private global value may cause the private to be renamed as
548 necessary to avoid collisions. Because the symbol is private to the
549 module, all references can be updated. This doesn't show up in any symbol
550 table in the object file.</dd>
554 assembler and evaluated by the linker. Unlike normal strong symbols, they
555 are removed by the linker from the final linked image (executable or
556 dynamic library).</dd>
561 linker. The symbols are removed by the linker from the final linked image
562 (executable or dynamic library).</dd>
564 <dt><tt><b><a name="linkage_linker_private_weak_def_auto">linker_private_weak_def_auto</a></b></tt></dt>
566 of the object is not taken. For instance, functions that had an inline
567 definition, but the compiler decided not to inline it. Note,
570 visibility. The symbols are removed by the linker from the final linked
571 image (executable or dynamic library).</dd>
574 <dd>Similar to private, but the value shows as a local symbol
580 into the object file corresponding to the LLVM module. They exist to
581 allow inlining and other optimizations to take place given knowledge of
582 the definition of the global, which is known to be somewhere outside the
585 This linkage type is only allowed on definitions, not declarations.</dd>
589 the same name when linkage occurs. This can be used to implement
590 some forms of inline functions, templates, or other code which must be
591 generated in each translation unit that uses it, but where the body may
592 be overridden with a more definitive definition later. Unreferenced
595 inline the body of this function into callers because it doesn't know if
596 this definition of the function is the definitive definition within the
597 program or whether it will be overridden by a stronger definition.
598 To enable inlining and other optimizations, use "<tt>linkonce_odr</tt>"
599 linkage.</dd>
609 they are used for tentative definitions in C, such as "<tt>int X;</tt>" at
610 global scope.
611 Symbols with "<tt>common</tt>" linkage are merged in the same way as
614 must have a zero initializer, and may not be marked '<a
616 have common linkage.</dd>
621 pointer to array type. When two global variables with appending linkage
622 are linked together, the two global arrays are appended together. This is
623 the LLVM, typesafe, equivalent of having the system linker append together
627 <dd>The semantics of this linkage follow the ELF object file model: the symbol
628 is weak until linked, if not linked, the symbol becomes null instead of
629 being an undefined reference.</dd>
633 <dd>Some languages allow differing globals to be merged, such as two functions
635 that only equivalent globals are ever merged (the "one definition rule"
638 be merged with equivalent globals. These linkage types are otherwise the
642 <dd>If none of the above identifiers are used, the global is externally
643 visible, meaning that it participates in linkage and can be used to
644 resolve external symbol references.</dd>
645 </dl>
647 <p>The next two types of linkage are targeted for Microsoft Windows platform
648 only. They are designed to support importing (exporting) symbols from (to)
649 DLLs (Dynamic Link Libraries).</p>
651 <dl>
654 or variable via a global pointer to a pointer that is set up by the DLL
655 exporting the symbol. On Microsoft Windows targets, the pointer name is
657 name.</dd>
661 pointer to a pointer in a DLL, so that it can be referenced with the
664 variable name.</dd>
665 </dl>
668 another module defined a "<tt>.LC0</tt>" variable and was linked with this
669 one, one of the two would be renamed, preventing a collision. Since
671 declarations), they are accessible outside of the current module.</p>
680 </div>
682 <!-- ======================================================================= -->
683 <h3>
685 </h3>
687 <div>
691 convention specified for the call. The calling convention of any pair of
692 dynamic caller/callee must match, or the behavior of the program is
693 undefined. The following calling conventions are supported by LLVM, and more
694 may be added in the future:</p>
696 <dl>
698 <dd>This calling convention (the default if no other calling convention is
699 specified) matches the target C calling conventions. This calling
700 convention supports varargs function calls and tolerates some mismatch in
701 the declared prototype and implemented declaration of the function (as
702 does normal C).</dd>
705 <dd>This calling convention attempts to make calls as fast as possible
706 (e.g. by passing things in registers). This calling convention allows the
707 target to use whatever tricks it wants to produce fast code for the
708 target, without having to conform to an externally specified ABI
709 (Application Binary Interface).
711 when this or the GHC convention is used.</a> This calling convention
712 does not support varargs and requires the prototype of all callees to
713 exactly match the prototype of the function definition.</dd>
716 <dd>This calling convention attempts to make code in the caller as efficient
717 as possible under the assumption that the call is not commonly executed.
718 As such, these calls often preserve all registers so that the call does
719 not break any live ranges in the caller side. This calling convention
720 does not support varargs and requires the prototype of all callees to
721 exactly match the prototype of the function definition.</dd>
724 <dd>This calling convention has been implemented specifically for use by the
726 It passes everything in registers, going to extremes to achieve this by
727 disabling callee save registers. This calling convention should not be
728 used lightly but only for specific situations such as an alternative to
730 implementing functional programming languages.At the moment only X86
731 supports this convention and it has the following limitations:
732 <ul>
734 floating point types are supported.</li>
737 </ul>
738 This calling convention supports
740 requires both the caller and callee are using it.
741 </dd>
744 <dd>Any calling convention may be specified by number, allowing
745 target-specific calling conventions to be used. Target specific calling
747 </dl>
749 <p>More calling conventions can be added/defined on an as-needed basis, to
750 support Pascal conventions or any other well-known target-independent
751 convention.</p>
753 </div>
755 <!-- ======================================================================= -->
756 <h3>
758 </h3>
760 <div>
762 <p>All Global Variables and Functions have one of the following visibility
763 styles:</p>
765 <dl>
767 <dd>On targets that use the ELF object file format, default visibility means
768 that the declaration is visible to other modules and, in shared libraries,
769 means that the declared entity may be overridden. On Darwin, default
770 visibility means that the declaration is visible to other modules. Default
774 <dd>Two declarations of an object with hidden visibility refer to the same
775 object if they are in the same shared object. Usually, hidden visibility
776 indicates that the symbol will not be placed into the dynamic symbol
777 table, so no other module (executable or shared library) can reference it
778 directly.</dd>
781 <dd>On ELF, protected visibility indicates that the symbol will be placed in
782 the dynamic symbol table, but that references within the defining module
783 will bind to the local symbol. That is, the symbol cannot be overridden by
784 another module.</dd>
785 </dl>
787 </div>
789 <!-- ======================================================================= -->
790 <h3>
792 </h3>
794 <div>
796 <p>LLVM IR allows you to specify name aliases for certain types. This can make
797 it easier to read the IR and make the IR more condensed (particularly when
798 recursive types are involved). An example of a name specification is:</p>
801 %mytype = type { %mytype*, i32 }
802 </pre>
808 <p>Note that type names are aliases for the structural type that they indicate,
809 and that you can therefore specify multiple names for the same type. This
810 often leads to confusing behavior when dumping out a .ll file. Since LLVM IR
811 uses structural typing, the name is not part of the type. When printing out
813 particular shape. This means that if you have code where two different
814 source types end up having the same LLVM type, that the dumper will sometimes
815 print the "wrong" or unexpected type. This is an important design point and
816 isn't going to change.</p>
818 </div>
820 <!-- ======================================================================= -->
821 <h3>
823 </h3>
825 <div>
827 <p>Global variables define regions of memory allocated at compilation time
828 instead of run-time. Global variables may optionally be initialized, may
829 have an explicit section to be placed in, and may have an optional explicit
830 alignment specified. A variable may be defined as "thread_local", which
831 means that it will not be shared by threads (each thread will have a
832 separated copy of the variable). A variable may be defined as a global
833 "constant," which indicates that the contents of the variable
835 global data to be placed in the read-only section of an executable, etc).
836 Note that variables that need runtime initialization cannot be marked
840 constant, even if the final definition of the global is not. This capability
841 can be used to enable slightly better optimization of the program, but
842 requires the language definition to guarantee that optimizations based on the
843 'constantness' are valid for the translation units that do not include the
844 definition.</p>
846 <p>As SSA values, global variables define pointer values that are in scope
847 (i.e. they dominate) all basic blocks in the program. Global variables
848 always define a pointer to their "content" type because they describe a
849 region of memory, and all memory objects in LLVM are accessed through
850 pointers.</p>
853 that the address is not significant, only the content. Constants marked
854 like this can be merged with other constants if they have the same
857 constant whose address is significant.</p>
859 <p>A global variable may be declared to reside in a target-specific numbered
860 address space. For targets that support them, address spaces may affect how
861 optimizations are performed and/or what target instructions are used to
862 access the variable. The default address space is zero. The address space
863 qualifier must precede any other attributes.</p>
865 <p>LLVM allows an explicit section to be specified for globals. If the target
866 supports it, it will emit globals to the section specified.</p>
868 <p>An explicit alignment may be specified for a global, which must be a power
869 of 2. If not present, or if the alignment is set to zero, the alignment of
870 the global is set by the target to whatever it feels convenient. If an
871 explicit alignment is specified, the global is forced to have exactly that
872 alignment. Targets and optimizers are not allowed to over-align the global
873 if the global has an assigned section. In this case, the extra alignment
874 could be observable: for example, code could assume that the globals are
875 densely packed in their section and try to iterate over them as an array,
876 alignment padding would break this iteration.</p>
878 <p>For example, the following defines a global in a numbered address space with
879 an initializer, section, and alignment:</p>
883 </pre>
885 </div>
888 <!-- ======================================================================= -->
889 <h3>
891 </h3>
893 <div>
901 name, a (possibly empty) argument list (each with optional
905 curly brace, a list of basic blocks, and a closing curly brace.</p>
913 name, a possibly empty list of arguments, an optional alignment, and an
916 <p>A function definition contains a list of basic blocks, forming the CFG
917 (Control Flow Graph) for the function. Each basic block may optionally start
918 with a label (giving the basic block a symbol table entry), contains a list
920 instruction (such as a branch or function return).</p>
922 <p>The first basic block in a function is special in two ways: it is immediately
923 executed on entrance to the function, and it is not allowed to have
924 predecessor basic blocks (i.e. there can not be any branches to the entry
925 block of a function). Because the block can have no predecessors, it also
928 <p>LLVM allows an explicit section to be specified for functions. If the target
929 supports it, it will emit functions to the section specified.</p>
931 <p>An explicit alignment may be specified for a function. If not present, or if
932 the alignment is set to zero, the alignment of the function is set by the
933 target to whatever it feels convenient. If an explicit alignment is
934 specified, the function is forced to have at least that much alignment. All
938 be significant and two identical functions can be merged</p>.
947 </pre>
949 </div>
951 <!-- ======================================================================= -->
952 <h3>
954 </h3>
956 <div>
959 function, global variable, another alias or bitcast of global value). Aliases
966 </pre>
968 </div>
970 <!-- ======================================================================= -->
971 <h3>
973 </h3>
975 <div>
978 nodes</a> (but not metadata strings) are the only valid operands for
979 a named metadata.</p>
983 ; Some unnamed metadata nodes, which are referenced by the named metadata.
987 ; A named metadata.
989 </pre>
991 </div>
993 <!-- ======================================================================= -->
994 <h3>
996 </h3>
998 <div>
1000 <p>The return type and each parameter of a function type may have a set of
1002 used to communicate additional information about the result or parameters of
1003 a function. Parameter attributes are considered to be part of the function,
1004 not of the function type, so functions with different parameter attributes
1005 can have the same function type.</p>
1007 <p>Parameter attributes are simple keywords that follow the type specified. If
1008 multiple parameter attributes are needed, they are space separated. For
1009 example:</p>
1012 declare i32 @printf(i8* noalias nocapture, ...)
1013 declare i32 @atoi(i8 zeroext)
1014 declare signext i8 @returns_signed_char()
1015 </pre>
1022 <dl>
1024 <dd>This indicates to the code generator that the parameter or return value
1025 should be zero-extended to the extent required by the target's ABI (which
1027 parameter) or the callee (for a return value).</dd>
1030 <dd>This indicates to the code generator that the parameter or return value
1031 should be sign-extended to the extent required by the target's ABI (which
1032 is usually 32-bits) by the caller (for a parameter) or the callee (for a
1033 return value).</dd>
1036 <dd>This indicates that this parameter or return value should be treated in a
1037 special target-dependent fashion during while emitting code for a function
1038 call or return (usually, by putting it in a register as opposed to memory,
1039 though some targets use it to distinguish between two different kinds of
1040 registers). Use of this attribute is target-specific.</dd>
1043 <dd><p>This indicates that the pointer parameter should really be passed by
1044 value to the function. The attribute implies that a hidden copy of the
1045 pointee
1046 is made between the caller and the callee, so the callee is unable to
1047 modify the value in the callee. This attribute is only valid on LLVM
1048 pointer arguments. It is generally used to pass structs and arrays by
1049 value, but is also valid on pointers to scalars. The copy is considered
1050 to belong to the caller not the callee (for example,
1053 values.</p>
1055 <p>The byval attribute also supports specifying an alignment with
1056 the align attribute. It indicates the alignment of the stack slot to
1057 form and the known alignment of the pointer specified to the call site. If
1058 the alignment is not specified, then the code generator makes a
1059 target-specific assumption.</p></dd>
1062 <dd>This indicates that the pointer parameter specifies the address of a
1063 structure that is the return value of the function in the source program.
1064 This pointer must be guaranteed by the caller to be valid: loads and
1065 stores to the structure may be assumed by the callee to not to trap. This
1066 may only be applied to the first parameter. This is not a valid attribute
1067 for return values. </dd>
1070 <dd>This indicates that pointer values
1073 ignoring certain "irrelevant" dependencies.
1074 For a call to the parent function, dependencies between memory
1075 references from before or after the call and from those during the call
1077 return value used in that call.
1078 The caller shares the responsibility with the callee for ensuring that
1079 these requirements are met.
1080 For further details, please see the discussion of the NoAlias response in
1082 <br>
1085 arguments, though it is slightly weaker.
1086 <br>
1089 </dd>
1092 <dd>This indicates that the callee does not make any copies of the pointer
1093 that outlive the callee itself. This is not a valid attribute for return
1094 values.</dd>
1097 <dd>This indicates that the pointer parameter can be excised using the
1099 attribute for return values.</dd>
1100 </dl>
1102 </div>
1104 <!-- ======================================================================= -->
1105 <h3>
1107 </h3>
1109 <div>
1111 <p>Each function may specify a garbage collector name, which is simply a
1112 string:</p>
1115 define void @f() gc "name" { ... }
1116 </pre>
1119 collector which will cause the compiler to alter its output in order to
1120 support the named garbage collection algorithm.</p>
1122 </div>
1124 <!-- ======================================================================= -->
1125 <h3>
1127 </h3>
1129 <div>
1131 <p>Function attributes are set to communicate additional information about a
1132 function. Function attributes are considered to be part of the function, not
1133 of the function type, so functions with different parameter attributes can
1134 have the same function type.</p>
1136 <p>Function attributes are simple keywords that follow the type specified. If
1137 multiple attributes are needed, they are space separated. For example:</p>
1140 define void @f() noinline { ... }
1141 define void @f() alwaysinline { ... }
1142 define void @f() alwaysinline optsize { ... }
1143 define void @f() optsize { ... }
1144 </pre>
1146 <dl>
1148 <dd>This attribute indicates that, when emitting the prologue and epilogue,
1149 the backend should forcibly align the stack pointer. Specify the
1150 desired alignment, which must be a power of two, in parentheses.
1153 <dd>This attribute indicates that the inliner should attempt to inline this
1154 function into callers whenever possible, ignoring any active inlining size
1155 threshold for this caller.</dd>
1158 <dd>This attribute indicates that the function should be 'hotpatchable',
1159 meaning the function can be patched and/or hooked even while it is
1160 loaded into memory. On x86, the function prologue will be preceded
1161 by six bytes of padding and will begin with a two-byte instruction.
1162 Most of the functions in the Windows system DLLs in Windows XP SP2 or
1163 higher were compiled in this fashion.</dd>
1166 <dd>This attribute indicates that the source code contained a hint that inlining
1167 this function is desirable (such as the "inline" keyword in C/C++). It
1168 is just a hint; it imposes no requirements on the inliner.</dd>
1171 <dd>This attribute disables prologue / epilogue emission for the function.
1172 This can have very system-specific consequences.</dd>
1178 <dd>This attribute indicates that the inliner should never inline this
1179 function in any situation. This attribute may not be used together with
1183 <dd>This attribute indicates that the code generator should not use a red
1184 zone, even if the target-specific ABI normally permits it.</dd>
1187 <dd>This function attribute indicates that the function never returns
1188 normally. This produces undefined behavior at runtime if the function
1189 ever does dynamically return.</dd>
1192 <dd>This function attribute indicates that the function never returns with an
1193 unwind or exceptional control flow. If the function does unwind, its
1194 runtime behavior is undefined.</dd>
1197 <dd>This attribute suggests that optimization passes and code generator passes
1198 make choices that keep the code size of this function low, and otherwise
1199 do optimizations specifically to reduce code size.</dd>
1202 <dd>This attribute indicates that the function computes its result (or decides
1203 to unwind an exception) based strictly on its arguments, without
1204 dereferencing any pointer arguments or otherwise accessing any mutable
1205 state (e.g. memory, control registers, etc) visible to caller functions.
1206 It does not write through any pointer arguments
1208 changes any state visible to callers. This means that it cannot unwind
1213 <dd>This attribute indicates that the function does not write through any
1215 arguments) or otherwise modify any state (e.g. memory, control registers,
1216 etc) visible to caller functions. It may dereference pointer arguments
1217 and read state that may be set in the caller. A readonly function always
1218 returns the same value (or unwinds an exception identically) when called
1219 with the same set of arguments and global state. It cannot unwind an
1224 <dd>This attribute indicates that the function should emit a stack smashing
1226 the stack before the local variables that's checked upon return from the
1227 function to see if it has been overwritten. A heuristic is used to
1228 determine if a function needs stack protectors or not.<br>
1229 <br>
1236 stack smashing protector. This overrides
1238 <br>
1243 </dl>
1245 </div>
1247 <!-- ======================================================================= -->
1248 <h3>
1250 </h3>
1252 <div>
1255 the GCC "file scope inline asm" blocks. These blocks are internally
1256 concatenated by LLVM and treated as a single unit, but may be separated in
1260 module asm "inline asm code goes here"
1261 module asm "more can go here"
1262 </pre>
1264 <p>The strings can contain any character by escaping non-printable characters.
1266 for the number.</p>
1268 <p>The inline asm code is simply printed to the machine code .s file when
1269 assembly code is generated.</p>
1271 </div>
1273 <!-- ======================================================================= -->
1274 <h3>
1276 </h3>
1278 <div>
1280 <p>A module may specify a target specific data layout string that specifies how
1281 data is to be laid out in memory. The syntax for the data layout is
1282 simply:</p>
1285 target datalayout = "<i>layout specification</i>"
1286 </pre>
1289 separated by the minus sign character ('-'). Each specification starts with
1290 a letter and may include other information after the letter to define some
1291 aspect of the data layout. The specifications accepted are as follows:</p>
1293 <dl>
1295 <dd>Specifies that the target lays out data in big-endian form. That is, the
1296 bits with the most significance have the lowest address location.</dd>
1299 <dd>Specifies that the target lays out data in little-endian form. That is,
1300 the bits with the least significance have the lowest address
1301 location.</dd>
1310 <dd>This specifies the alignment for an integer type of a given bit
1314 <dd>This specifies the alignment for a vector type of a given bit
1318 <dd>This specifies the alignment for a floating point type of a given bit
1322 targets.
1325 <dd>This specifies the alignment for an aggregate type of a given bit
1329 <dd>This specifies the alignment for a stack object of a given bit
1333 <dd>This specifies a set of native integer widths for the target CPU
1336 this set are considered to support most general arithmetic
1337 operations efficiently.</dd>
1338 </dl>
1340 <p>When constructing the data layout for a given target, LLVM starts with a
1341 default set of specifications which are then (possibly) overridden by the
1343 are given in this list:</p>
1345 <ul>
1360 </ul>
1362 <p>When LLVM is determining the alignment for a given type, it uses the
1363 following rules:</p>
1365 <ol>
1366 <li>If the type sought is an exact match for one of the specifications, that
1367 specification is used.</li>
1369 <li>If no match is found, and the type sought is an integer type, then the
1370 smallest integer type that is larger than the bitwidth of the sought type
1371 is used. If none of the specifications are larger than the bitwidth then
1372 the the largest integer type is used. For example, given the default
1373 specifications above, the i7 type will use the alignment of i8 (next
1374 largest) while both i65 and i256 will use the alignment of i64 (largest
1375 specified).</li>
1377 <li>If no match is found, and the type sought is a vector type, then the
1378 largest vector type that is smaller than the sought vector type will be
1381 </ol>
1383 </div>
1385 <!-- ======================================================================= -->
1386 <h3>
1388 </h3>
1390 <div>
1392 <p>Any memory access must be done through a pointer value associated
1393 with an address range of the memory access, otherwise the behavior
1394 is undefined. Pointer values are associated with address ranges
1395 according to the following rules:</p>
1397 <ul>
1398 <li>A pointer value is associated with the addresses associated with
1400 <li>An address of a global variable is associated with the address
1401 range of the variable's storage.</li>
1402 <li>The result value of an allocation instruction is associated with
1403 the address range of the allocated storage.</li>
1404 <li>A null pointer in the default address-space is associated with
1405 no address.</li>
1406 <li>An integer constant other than zero or a pointer value returned
1407 from a function not defined within LLVM may be associated with address
1408 ranges allocated through mechanisms other than those provided by
1409 LLVM. Such ranges shall not overlap with any ranges of addresses
1410 allocated by mechanisms provided by LLVM.</li>
1411 </ul>
1414 to the following rules:</p>
1416 <ul>
1417 <li>A pointer value formed from a
1420 <li>The result value of a
1423 <li>A pointer value formed by an
1425 pointer values that contribute (directly or indirectly) to the
1426 computation of the pointer's value.</li>
1428 </ul>
1432 slightly weaker.</p>
1434 <p>LLVM IR does not associate types with memory. The result type of a
1436 alignment of the memory from which to load, as well as the
1437 interpretation of the value. The first operand type of a
1439 and alignment of the store.</p>
1441 <p>Consequently, type-based alias analysis, aka TBAA, aka
1444 additional information which specialized optimization passes may use
1445 to implement type-based alias analysis.</p>
1447 </div>
1449 <!-- ======================================================================= -->
1450 <h3>
1452 </h3>
1454 <div>
1459 The optimizers must not change the number of volatile operations or change their
1460 order of execution relative to other volatile operations. The optimizers
1462 operations. This is not Java's "volatile" and has no cross-thread
1463 synchronization behavior.</p>
1465 </div>
1467 </div>
1469 <!-- *********************************************************************** -->
1471 <!-- *********************************************************************** -->
1473 <div>
1475 <p>The LLVM type system is one of the most important features of the
1476 intermediate representation. Being typed enables a number of optimizations
1477 to be performed on the intermediate representation directly, without having
1478 to do extra analyses on the side before the transformation. A strong type
1479 system makes it easier to read the generated code and enables novel analyses
1480 and transformations that are not feasible to perform on normal three address
1481 code representations.</p>
1483 <!-- ======================================================================= -->
1484 <h3>
1486 </h3>
1488 <div>
1493 <tbody>
1495 <tr>
1498 </tr>
1499 <tr>
1502 </tr>
1503 <tr>
1513 </td>
1514 </tr>
1515 <tr>
1523 </tr>
1524 <tr>
1533 </td>
1534 </tr>
1535 </tbody>
1536 </table>
1539 important. Values of these types are the only ones which can be produced by
1540 instructions.</p>
1542 </div>
1544 <!-- ======================================================================= -->
1545 <h3>
1547 </h3>
1549 <div>
1551 <p>The primitive types are the fundamental building blocks of the LLVM
1552 system.</p>
1554 <!-- _______________________________________________________________________ -->
1555 <h4>
1557 </h4>
1559 <div>
1562 <p>The integer type is a very simple type that simply specifies an arbitrary
1563 bit width for the integer type desired. Any bit width from 1 bit to
1567 <pre>
1568 iN
1569 </pre>
1572 value.</p>
1579 </tr>
1583 </tr>
1587 </tr>
1588 </table>
1590 </div>
1592 <!-- _______________________________________________________________________ -->
1593 <h4>
1595 </h4>
1597 <div>
1599 <table>
1600 <tbody>
1607 </tbody>
1608 </table>
1610 </div>
1612 <!-- _______________________________________________________________________ -->
1613 <h4>
1615 </h4>
1617 <div>
1620 <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>
1623 <pre>
1624 x86mmx
1625 </pre>
1627 </div>
1629 <!-- _______________________________________________________________________ -->
1630 <h4>
1632 </h4>
1634 <div>
1640 <pre>
1641 void
1642 </pre>
1644 </div>
1646 <!-- _______________________________________________________________________ -->
1647 <h4>
1649 </h4>
1651 <div>
1657 <pre>
1658 label
1659 </pre>
1661 </div>
1663 <!-- _______________________________________________________________________ -->
1664 <h4>
1666 </h4>
1668 <div>
1671 <p>The metadata type represents embedded metadata. No derived types may be
1673 arguments.
1676 <pre>
1677 metadata
1678 </pre>
1680 </div>
1682 </div>
1684 <!-- ======================================================================= -->
1685 <h3>
1687 </h3>
1689 <div>
1691 <p>The real power in LLVM comes from the derived types in the system. This is
1692 what allows a programmer to represent arrays, functions, pointers, and other
1693 useful types. Each of these types contain one or more element types which
1694 may be a primitive type, or another derived type. For example, it is
1695 possible to have a two dimensional array, using an array as the element type
1696 of another array.</p>
1699 <!-- _______________________________________________________________________ -->
1700 <h4>
1702 </h4>
1704 <div>
1706 <p>Aggregate Types are a subset of derived types that can contain multiple
1709 aggregate types.</p>
1711 </div>
1713 <!-- _______________________________________________________________________ -->
1714 <h4>
1716 </h4>
1718 <div>
1721 <p>The array type is a very simple derived type that arranges elements
1722 sequentially in memory. The array type requires a size (number of elements)
1723 and an underlying data type.</p>
1726 <pre>
1728 </pre>
1731 be any type with a size.</p>
1738 </tr>
1742 </tr>
1746 </tr>
1747 </table>
1753 </tr>
1757 </tr>
1761 </tr>
1762 </table>
1764 <p>There is no restriction on indexing beyond the end of the array implied by
1765 a static type (though there are restrictions on indexing beyond the bounds
1766 of an allocated object in some cases). This means that single-dimension
1767 'variable sized array' addressing can be implemented in LLVM with a zero
1768 length array type. An implementation of 'pascal style arrays' in LLVM could
1771 </div>
1773 <!-- _______________________________________________________________________ -->
1774 <h4>
1776 </h4>
1778 <div>
1781 <p>The function type can be thought of as a function signature. It consists of
1782 a return type and a list of formal parameter types. The return type of a
1783 function type is a first class type or a void type.</p>
1786 <pre>
1788 </pre>
1792 which indicates that the function takes a variable number of arguments.
1793 Variable argument functions can access their arguments with
1803 </td>
1806 </tt></td>
1810 </td>
1816 LLVM.
1817 </td>
1822 </td>
1823 </tr>
1824 </table>
1826 </div>
1828 <!-- _______________________________________________________________________ -->
1829 <h4>
1831 </h4>
1833 <div>
1836 <p>The structure type is used to represent a collection of data members together
1837 in memory. The packing of the field types is defined to match the ABI of the
1838 underlying processor. The elements of a structure may be any type that has a
1839 size.</p>
1844 Structures in registers are accessed using the
1848 <pre>
1850 </pre>
1863 </tr>
1864 </table>
1866 </div>
1868 <!-- _______________________________________________________________________ -->
1869 <h4>
1871 </h4>
1873 <div>
1876 <p>The packed structure type is used to represent a collection of data members
1877 together in memory. There is no padding between fields. Further, the
1878 alignment of a packed structure is 1 byte. The elements of a packed
1879 structure may be any type that has a size.</p>
1886 <pre>
1888 </pre>
1902 </tr>
1903 </table>
1905 </div>
1907 <!-- _______________________________________________________________________ -->
1908 <h4>
1910 </h4>
1912 <div>
1915 <p>The pointer type is used to specify memory locations.
1916 Pointers are commonly used to reference objects in memory.</p>
1918 <p>Pointer types may have an optional address space attribute defining the
1919 numbered address space where the pointed-to object resides. The default
1920 address space is number zero. The semantics of non-zero address
1921 spaces are target-specific.</p>
1927 <pre>
1929 </pre>
1937 </tr>
1943 </tr>
1948 </tr>
1949 </table>
1951 </div>
1953 <!-- _______________________________________________________________________ -->
1954 <h4>
1956 </h4>
1958 <div>
1961 <p>A vector type is a simple derived type that represents a vector of elements.
1962 Vector types are used when multiple primitive data are operated in parallel
1963 using a single instruction (SIMD). A vector type requires a size (number of
1964 elements) and an underlying primitive data type. Vector types are considered
1968 <pre>
1970 </pre>
1973 may be any integer or floating point type. Vectors of size zero are not
1974 allowed, and pointers are not allowed as the element type.</p>
1981 </tr>
1985 </tr>
1989 </tr>
1990 </table>
1992 </div>
1994 <!-- _______________________________________________________________________ -->
1995 <h4>
1997 </h4>
1999 <div>
2002 <p>Opaque types are used to represent unknown types in the system. This
2003 corresponds (for example) to the C notion of a forward declared structure
2004 type. In LLVM, opaque types can eventually be resolved to any type (not just
2005 a structure type).</p>
2008 <pre>
2009 opaque
2010 </pre>
2017 </tr>
2018 </table>
2020 </div>
2022 </div>
2024 <!-- ======================================================================= -->
2025 <h3>
2027 </h3>
2029 <div>
2033 requiring it to have a name. For instance, a structure declaration may
2034 contain a pointer to any of the types it is lexically a member of. Example
2035 of up references (with their equivalent as named type declarations)
2036 include:</p>
2038 <pre>
2039 { \2 * } %x = type { %x* }
2040 { \2 }* %y = type { %y }*
2041 \1* %z = type %z*
2042 </pre>
2044 <p>An up reference is needed by the asmprinter for printing out cyclic types
2045 when there is no declared name for a type in the cycle. Because the
2046 asmprinter does not want to print out an infinite type string, it needs a
2047 syntax to handle recursive types that have no names (all names are optional
2048 in llvm IR).</p>
2051 <pre>
2053 </pre>
2062 </tr>
2066 structure.</td>
2067 </tr>
2068 </table>
2070 </div>
2072 </div>
2074 <!-- *********************************************************************** -->
2076 <!-- *********************************************************************** -->
2078 <div>
2080 <p>LLVM has several different basic types of constants. This section describes
2081 them all and their syntax.</p>
2083 <!-- ======================================================================= -->
2084 <h3>
2086 </h3>
2088 <div>
2090 <dl>
2098 with integer types.</dd>
2102 exponential notation (e.g. 1.23421e+2), or a more precise hexadecimal
2103 notation (see below). The assembler requires the exact decimal value of a
2104 floating-point constant. For example, the assembler accepts 1.25 but
2111 </dl>
2113 <p>The one non-intuitive notation for constants is the hexadecimal form of
2114 floating point constants. For example, the form '<tt>double
2117 constants are required (and the only time that they are generated by the
2118 disassembler) is when a floating point constant must be emitted but it cannot
2119 be represented as a decimal floating point number in a reasonable number of
2120 digits. For example, NaN's, infinities, and other special values are
2121 represented in their IEEE hexadecimal format so that assembly and disassembly
2122 do not cause any bits to change in the constants.</p>
2124 <p>When using the hexadecimal form, constants of types float and double are
2125 represented using the 16-digit form shown above (which matches the IEEE754
2126 representation for double); float values must, however, be exactly
2127 representable as IEE754 single precision. Hexadecimal format is always used
2128 for long double, and there are three forms of long double. The 80-bit format
2130 The 128-bit format used by PowerPC (two adjacent doubles) is represented
2133 currently supported target uses this format. Long doubles will only work if
2134 they match the long double format on your target. All hexadecimal formats
2135 are big-endian (sign bit at the left).</p>
2138 </div>
2140 <!-- ======================================================================= -->
2141 <h3>
2144 </h3>
2146 <div>
2148 <p>Complex constants are a (potentially recursive) combination of simple
2149 constants and smaller complex constants.</p>
2151 <dl>
2153 <dd>Structure constants are represented with notation similar to structure
2154 type definitions (a comma separated list of elements, surrounded by braces
2158 the number and types of elements must match those specified by the
2159 type.</dd>
2162 <dd>Array constants are represented with notation similar to array type
2163 definitions (a comma separated list of elements, surrounded by square
2166 the number and types of elements must match those specified by the
2167 type.</dd>
2170 <dd>Vector constants are represented with notation similar to vector type
2171 definitions (a comma separated list of elements, surrounded by
2173 42, i32 11, i32 74, i32 100 ></tt>". Vector constants must
2175 elements must match those specified by the type.</dd>
2181 This is often used to avoid having to print large zero initializers
2182 (e.g. for large arrays) and is always exactly equivalent to using explicit
2183 zero initializers.</dd>
2186 <dd>A metadata node is a structure-like constant with
2189 be interpreted as part of the instruction stream, metadata is a place to
2190 attach additional information such as debug info.</dd>
2191 </dl>
2193 </div>
2195 <!-- ======================================================================= -->
2196 <h3>
2198 </h3>
2200 <div>
2204 (link-time) constants. These constants are explicitly referenced when
2207 legal LLVM file:</p>
2210 @X = global i32 17
2211 @Y = global i32 42
2212 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2213 </pre>
2215 </div>
2217 <!-- ======================================================================= -->
2218 <h3>
2220 </h3>
2222 <div>
2225 indicates that the user of the value may receive an unspecified bit-pattern.
2229 <p>Undefined values are useful because they indicate to the compiler that the
2230 program is well defined no matter what value is used. This gives the
2231 compiler more freedom to optimize. Here are some examples of (potentially
2232 surprising) transformations that are valid (in pseudo IR):</p>
2236 %A = add %X, undef
2237 %B = sub %X, undef
2238 %C = xor %X, undef
2239 Safe:
2240 %A = undef
2241 %B = undef
2242 %C = undef
2243 </pre>
2245 <p>This is safe because all of the output bits are affected by the undef bits.
2246 Any output bit can have a zero or one depending on the input bits.</p>
2249 %A = or %X, undef
2250 %B = and %X, undef
2251 Safe:
2252 %A = -1
2253 %B = 0
2254 Unsafe:
2255 %A = undef
2256 %B = undef
2257 </pre>
2259 <p>These logical operations have bits that are not always affected by the input.
2270 %A = select undef, %X, %Y
2271 %B = select undef, 42, %Y
2272 %C = select %X, %Y, undef
2273 Safe:
2274 %A = %X (or %Y)
2275 %B = 42 (or %Y)
2276 %C = %Y
2277 Unsafe:
2278 %A = undef
2279 %B = undef
2280 %C = undef
2281 </pre>
2290 eliminated.</p>
2293 %A = xor undef, undef
2295 %B = undef
2296 %C = xor %B, %B
2298 %D = undef
2299 %E = icmp lt %D, 4
2300 %F = icmp gte %D, 4
2302 Safe:
2303 %A = undef
2304 %B = undef
2305 %C = undef
2306 %D = undef
2307 %E = undef
2308 %F = undef
2309 </pre>
2312 necessarily the same. This can be surprising to people (and also matches C
2313 semantics) where they assume that "<tt>X^X</tt>" is always zero, even
2316 its value over its "live range". This is true because the variable doesn't
2318 from arbitrary registers that happen to be around when needed, so the value
2320 need to have the same semantics or the core LLVM "replace all uses with"
2321 concept would not hold.</p>
2324 %A = fdiv undef, %X
2325 %B = fdiv %X, undef
2326 Safe:
2327 %A = undef
2328 b: unreachable
2329 </pre>
2336 defined on SNaN's. However, in the second example, we can make a more
2338 arbitrary value, we are allowed to assume that it could be zero. Since a
2340 the operation does not execute at all. This allows us to delete the divide and
2341 all code after it. Because the undefined operation "can't happen", the
2342 optimizer can assume that it occurs in dead code.</p>
2345 a: store undef -> %X
2346 b: store %X -> undef
2347 Safe:
2349 b: unreachable
2350 </pre>
2353 undefined value can be assumed to not have any effect; we can assume that the
2354 value is overwritten with bits that happen to match what was already there.
2356 memory, therefore, it has undefined behavior.</p>
2358 </div>
2360 <!-- ======================================================================= -->
2361 <h3>
2363 </h3>
2365 <div>
2368 instead of representing an unspecified bit pattern, they represent the
2369 fact that an instruction or constant expression which cannot evoke side
2370 effects has nevertheless detected a condition which results in undefined
2371 behavior.</p>
2373 <p>There is currently no way of representing a trap value in the IR; they
2374 only exist when produced by operations such as
2379 <ul>
2381 their operands.</li>
2384 to their dynamic predecessor basic block.</li>
2386 <li>Function arguments depend on the corresponding actual argument values in
2387 the dynamic callers of their functions.</li>
2391 control back to them.</li>
2395 or exception-throwing call instructions that dynamically transfer control
2396 back to them.</li>
2398 <li>Non-volatile loads and stores depend on the most recent stores to all of the
2399 referenced memory addresses, following the order in the IR
2400 (including loads and stores implied by intrinsics such as
2403 <!-- TODO: In the case of multiple threads, this only applies if the store
2404 "happens-before" the load or store. -->
2406 <!-- TODO: floating-point exception state -->
2408 <li>An instruction with externally visible side effects depends on the most
2409 recent preceding instruction with externally visible side effects, following
2410 the order in the IR. (This includes
2415 if the terminator instruction has multiple successors and the instruction
2416 is always executed when control transfers to one of the successors, and
2417 may not be executed when control is transferred to another.</li>
2420 instruction if the set of instructions it otherwise depends on would be
2421 different if the terminator had transferred control to a different
2422 successor.</li>
2426 </ul>
2428 <p>Whenever a trap value is generated, all values which depend on it evaluate
2429 to trap. If they have side effects, the evoke their side effects as if each
2430 operand with a trap value were undef. If they have externally-visible side
2431 effects, the behavior is undefined.</p>
2436 entry:
2439 %trap_yet_again = getelementptr i32* @h, i32 %still_trap
2440 store i32 0, i32* %trap_yet_again ; undefined behavior
2442 store i32 %trap, i32* @g ; Trap value conceptually stored to memory.
2443 %trap2 = load i32* @g ; Returns a trap value, not just undef.
2445 volatile store i32 %trap, i32* @g ; External observation; undefined behavior.
2447 %narrowaddr = bitcast i32* @g to i16*
2448 %wideaddr = bitcast i32* @g to i64*
2449 %trap3 = load 16* %narrowaddr ; Returns a trap value.
2450 %trap4 = load i64* %widaddr ; Returns a trap value.
2452 %cmp = icmp i32 slt %trap, 0 ; Returns a trap value.
2453 %br i1 %cmp, %true, %end ; Branch to either destination.
2455 true:
2456 volatile store i32 0, i32* @g ; This is control-dependent on %cmp, so
2457 ; it has undefined behavior.
2458 br label %end
2460 end:
2462 ; Both edges into this PHI are
2463 ; control-dependent on %cmp, so this
2464 ; always results in a trap value.
2466 volatile store i32 0, i32* @g ; This would depend on the store in %true
2467 ; if %cmp is true, or the store in %entry
2468 ; otherwise, so this is undefined behavior.
2470 %br i1 %cmp, %second_true, %second_end
2471 ; The same branch again, but this time the
2472 ; true block doesn't have side effects.
2474 second_true:
2475 ; No side effects!
2476 br label %end
2478 second_end:
2479 volatile store i32 0, i32* @g ; This time, the instruction always depends
2480 ; on the store in %end. Also, it is
2481 ; control-equivalent to %end, so this is
2482 ; well- defined (again, ignoring earlier
2483 ; undefined behavior in this example).
2484 </pre>
2486 </div>
2488 <!-- ======================================================================= -->
2489 <h3>
2491 </h3>
2493 <div>
2498 basic block in the specified function, and always has an i8* type. Taking
2499 the address of the entry block is illegal.</p>
2501 <p>This value only has defined behavior when used as an operand to the
2503 comparisons against null. Pointer equality tests between labels addresses
2504 results in undefined behavior — though, again, comparison against null
2505 is ok, and no label is equal to the null pointer. This may be passed around
2506 as an opaque pointer sized value as long as the bits are not inspected. This
2509 instruction.</p>
2511 <p>Finally, some targets may provide defined semantics when using the value as
2512 the operand to an inline assembly, but that is target specific.</p>
2514 </div>
2517 <!-- ======================================================================= -->
2518 <h3>
2520 </h3>
2522 <div>
2524 <p>Constant expressions are used to allow expressions involving other constants
2525 to be used as constants. Constant expressions may be of
2527 operation that does not have side effects (e.g. load and call are not
2528 supported). The following is the syntax for constant expressions:</p>
2530 <dl>
2532 <dd>Truncate a constant to another type. The bit size of CST must be larger
2533 than the bit size of TYPE. Both types must be integers.</dd>
2536 <dd>Zero extend a constant to another type. The bit size of CST must be
2537 smaller than the bit size of TYPE. Both types must be integers.</dd>
2540 <dd>Sign extend a constant to another type. The bit size of CST must be
2541 smaller than the bit size of TYPE. Both types must be integers.</dd>
2544 <dd>Truncate a floating point constant to another floating point type. The
2545 size of CST must be larger than the size of TYPE. Both types must be
2546 floating point.</dd>
2549 <dd>Floating point extend a constant to another type. The size of CST must be
2550 smaller or equal to the size of TYPE. Both types must be floating
2551 point.</dd>
2554 <dd>Convert a floating point constant to the corresponding unsigned integer
2555 constant. TYPE must be a scalar or vector integer type. CST must be of
2556 scalar or vector floating point type. Both CST and TYPE must be scalars,
2557 or vectors of the same number of elements. If the value won't fit in the
2558 integer type, the results are undefined.</dd>
2561 <dd>Convert a floating point constant to the corresponding signed integer
2562 constant. TYPE must be a scalar or vector integer type. CST must be of
2563 scalar or vector floating point type. Both CST and TYPE must be scalars,
2564 or vectors of the same number of elements. If the value won't fit in the
2565 integer type, the results are undefined.</dd>
2568 <dd>Convert an unsigned integer constant to the corresponding floating point
2569 constant. TYPE must be a scalar or vector floating point type. CST must be
2570 of scalar or vector integer type. Both CST and TYPE must be scalars, or
2571 vectors of the same number of elements. If the value won't fit in the
2572 floating point type, the results are undefined.</dd>
2575 <dd>Convert a signed integer constant to the corresponding floating point
2576 constant. TYPE must be a scalar or vector floating point type. CST must be
2577 of scalar or vector integer type. Both CST and TYPE must be scalars, or
2578 vectors of the same number of elements. If the value won't fit in the
2579 floating point type, the results are undefined.</dd>
2582 <dd>Convert a pointer typed constant to the corresponding integer constant
2588 <dd>Convert a integer constant to a pointer constant. TYPE must be a pointer
2589 type. CST must be of integer type. The CST value is zero extended,
2590 truncated, or unchanged to make it fit in a pointer size. This one is
2594 <dd>Convert a constant, CST, to another TYPE. The constraints of the operands
2602 instruction, the index list may have zero or more indexes, which are
2616 constants.</dd>
2620 constants.</dd>
2624 constants.</dd>
2628 constants. The index list is interpreted in a similar manner as indices in
2630 index value must be specified.</dd>
2634 constants. The index list is interpreted in a similar manner as indices in
2636 index value must be specified.</dd>
2639 <dd>Perform the specified operation of the LHS and RHS constants. OPCODE may
2642 on operands are the same as those for the corresponding instruction
2643 (e.g. no bitwise operations on floating point values are allowed).</dd>
2644 </dl>
2646 </div>
2648 </div>
2650 <!-- *********************************************************************** -->
2652 <!-- *********************************************************************** -->
2653 <div>
2654 <!-- ======================================================================= -->
2655 <h3>
2657 </h3>
2659 <div>
2661 <p>LLVM supports inline assembler expressions (as opposed
2663 a special value. This value represents the inline assembler as a string
2664 (containing the instructions to emit), a list of operand constraints (stored
2665 as a string), a flag that indicates whether or not the inline asm
2666 expression has side effects, and a flag indicating whether the function
2667 containing the asm needs to align its stack conservatively. An example
2668 inline assembler expression is:</p>
2672 </pre>
2676 have:</p>
2680 </pre>
2682 <p>Inline asms with side effects not visible in the constraint list must be
2683 marked as having side effects. This is done through the use of the
2688 </pre>
2690 <p>In some cases inline asms will contain code that will not work unless the
2691 stack is aligned in some way, such as calls or SSE instructions on x86,
2692 yet will not contain code that does that alignment within the asm.
2693 The compiler should make conservative assumptions about what the asm might
2694 contain and should generate its usual stack alignment code in the prologue
2699 </pre>
2702 first.</p>
2704 <p>TODO: The format of the asm and constraints string still need to be
2705 documented here. Constraints on what can be done (e.g. duplication, moving,
2706 etc need to be documented). This is probably best done by reference to
2707 another document that covers inline asm from a holistic perspective.</p>
2709 <h4>
2711 </h4>
2713 <div>
2716 attached to it that contains a list of constant integers. If present, the
2717 code generator will use the integer as the location cookie value when report
2718 errors through the LLVMContext error reporting mechanisms. This allows a
2719 front-end to correlate backend errors that occur with inline asm back to the
2720 source code that produced it. For example:</p>
2724 ...
2726 </pre>
2728 <p>It is up to the front-end to make sense of the magic numbers it places in the
2729 IR. If the MDNode contains multiple constants, the code generator will use
2730 the one that corresponds to the line of the asm that the error occurs on.</p>
2732 </div>
2734 </div>
2736 <!-- ======================================================================= -->
2737 <h3>
2739 </h3>
2741 <div>
2743 <p>LLVM IR allows metadata to be attached to instructions in the program that
2744 can convey extra information about the code to the optimizers and code
2745 generator. One example application of metadata is source-level debug
2746 information. There are two metadata primitives: strings and nodes. All
2750 <p>A metadata string is a string surrounded by double quotes. It can contain
2754 <p>Metadata nodes are represented with notation similar to structure constants
2755 (a comma separated list of elements, surrounded by braces and preceded by an
2760 metadata nodes, which can be looked up in the module symbol table. For
2761 example: "<tt>!foo = metadata !{!4, !3}</tt>".
2764 function is using two metadata arguments.</p>
2767 <pre>
2769 </pre>
2770 </div>
2776 <pre>
2778 </pre>
2779 </div>
2781 </div>
2783 </div>
2785 <!-- *********************************************************************** -->
2786 <h2>
2788 </h2>
2789 <!-- *********************************************************************** -->
2790 <div>
2792 code generation or other IR semantics. These are documented here. All globals
2793 of this sort should have a section specified as "<tt>llvm.metadata</tt>". This
2794 section and all globals that start with "<tt>llvm.</tt>" are reserved for use
2795 by LLVM.</p>
2797 <!-- ======================================================================= -->
2798 <h3>
2800 </h3>
2802 <div>
2806 pointers to global variables and functions which may optionally have a pointer
2807 cast formed of bitcast or getelementptr. For example, a legal use of it is:</p>
2809 <pre>
2810 @X = global i8 4
2811 @Y = global i32 123
2813 @llvm.used = appending global [2 x i8*] [
2814 i8* @X,
2815 i8* bitcast (i32* @Y to i8*)
2816 ], section "llvm.metadata"
2817 </pre>
2820 compiler, assembler, and linker are required to treat the symbol as if there is
2821 a reference to the global that it cannot see. For example, if a variable has
2823 list, it cannot be deleted. This is commonly used to represent references from
2824 inline asms and other things the compiler cannot "see", and corresponds to
2827 <p>On some targets, the code generator must emit a directive to the assembler or
2828 object file to prevent the assembler and linker from molesting the symbol.</p>
2830 </div>
2832 <!-- ======================================================================= -->
2833 <h3>
2836 </a>
2837 </h3>
2839 <div>
2843 touching the symbol. On targets that support it, this allows an intelligent
2844 linker to optimize references to the symbol without being impeded as it would be
2847 <p>This is a rare construct that should only be used in rare circumstances, and
2848 should not be exposed to source languages.</p>
2850 </div>
2852 <!-- ======================================================================= -->
2853 <h3>
2855 </h3>
2857 <div>
2858 <pre>
2859 %0 = type { i32, void ()* }
2861 </pre>
2862 <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.
2863 </p>
2865 </div>
2867 <!-- ======================================================================= -->
2868 <h3>
2870 </h3>
2872 <div>
2873 <pre>
2874 %0 = type { i32, void ()* }
2876 </pre>
2878 <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.
2879 </p>
2881 </div>
2883 </div>
2885 <!-- *********************************************************************** -->
2887 <!-- *********************************************************************** -->
2889 <div>
2891 <p>The LLVM instruction set consists of several different classifications of
2898 <!-- ======================================================================= -->
2899 <h3>
2901 </h3>
2903 <div>
2906 in a program ends with a "Terminator" instruction, which indicates which
2907 block should be executed after the current block is finished. These
2909 control flow, not values (the one exception being the
2912 <p>There are seven different terminator instructions: the
2921 <!-- _______________________________________________________________________ -->
2922 <h4>
2924 </h4>
2926 <div>
2929 <pre>
2932 </pre>
2936 a value) from a function back to the caller.</p>
2939 value and then causes control flow, and one that just causes control flow to
2940 occur.</p>
2944 return value. The type of the return value must be a
2949 value or a return value with a type that does not match its type, or if it
2951 return value.</p>
2955 the calling function's context. If the caller is a
2957 instruction after the call. If the caller was an
2959 the beginning of the "normal" destination block. If the instruction returns
2960 a value, that value shall set the call or invoke instruction's return
2961 value.</p>
2964 <pre>
2968 </pre>
2970 </div>
2971 <!-- _______________________________________________________________________ -->
2972 <h4>
2974 </h4>
2976 <div>
2979 <pre>
2980 br i1 <cond>, label <iftrue>, label <iffalse><br> br label <dest> <i>; Unconditional branch</i>
2981 </pre>
2985 different basic block in the current function. There are two forms of this
2986 instruction, corresponding to a conditional branch and an unconditional
2987 branch.</p>
2993 target.</p>
3002 <pre>
3003 Test:
3005 br i1 %cond, label %IfEqual, label %IfUnequal
3006 IfEqual:
3008 IfUnequal:
3010 </pre>
3012 </div>
3014 <!-- _______________________________________________________________________ -->
3015 <h4>
3017 </h4>
3019 <div>
3022 <pre>
3023 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3024 </pre>
3029 instruction, allowing a branch to occur to one of many possible
3030 destinations.</p>
3036 The table is not allowed to contain duplicate constant entries.</p>
3041 is searched for the given value. If the value is found, control flow is
3042 transferred to the corresponding destination; otherwise, control flow is
3043 transferred to the default destination.</p>
3046 <p>Depending on properties of the target machine and the particular
3048 different ways. For example, it could be generated as a series of chained
3049 conditional branches or with a lookup table.</p>
3052 <pre>
3055 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3058 switch i32 0, label %dest [ ]
3061 switch i32 %val, label %otherwise [ i32 0, label %onzero
3062 i32 1, label %onone
3063 i32 2, label %ontwo ]
3064 </pre>
3066 </div>
3069 <!-- _______________________________________________________________________ -->
3070 <h4>
3072 </h4>
3074 <div>
3077 <pre>
3079 </pre>
3084 within the current function, whose address is specified by
3091 rest of the arguments indicate the full set of possible destinations that the
3092 address may point to. Blocks are allowed to occur multiple times in the
3093 destination list, though this isn't particularly useful.</p>
3095 <p>This destination list is required so that dataflow analysis has an accurate
3096 understanding of the CFG.</p>
3100 <p>Control transfers to the block specified in the address argument. All
3101 possible destination blocks must be listed in the label list, otherwise this
3102 instruction has undefined behavior. This implies that jumps to labels
3103 defined in other functions have undefined behavior as well.</p>
3110 <pre>
3111 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3112 </pre>
3114 </div>
3117 <!-- _______________________________________________________________________ -->
3118 <h4>
3120 </h4>
3122 <div>
3125 <pre>
3126 <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>]
3128 </pre>
3132 function, with the possibility of control flow transfer to either the
3135 control flow will return to the "normal" label. If the callee (or any
3137 instruction, control is interrupted and continued at the dynamically nearest
3143 <ol>
3145 convention</a> the call should use. If none is specified, the call
3146 defaults to using C calling conventions.</li>
3153 function value being invoked. In most cases, this is a direct function
3155 off an arbitrary pointer to function value.</li>
3158 function to be invoked. </li>
3161 signature argument types and parameter attributes. All arguments must be
3163 signature indicates the function accepts a variable number of arguments,
3164 the extra arguments can be specified.</li>
3175 </ol>
3178 <p>This instruction is designed to operate as a standard
3180 primary difference is that it establishes an association with a label, which
3181 is used by the runtime library to unwind the stack.</p>
3183 <p>This instruction is used in languages with destructors to ensure that proper
3185 exception. Additionally, this is important for implementation of
3188 <p>For the purposes of the SSA form, the definition of the value returned by the
3190 block to the "normal" label. If the callee unwinds then no return value is
3191 available.</p>
3193 <p>Note that the code generator does not yet completely support unwind, and
3194 that the invoke/unwind semantics are likely to change in future versions.</p>
3197 <pre>
3198 %retval = invoke i32 @Test(i32 15) to label %Continue
3202 </pre>
3204 </div>
3206 <!-- _______________________________________________________________________ -->
3208 <h4>
3210 </h4>
3212 <div>
3215 <pre>
3216 unwind
3217 </pre>
3221 at the first callee in the dynamic call stack which used
3223 This is primarily used to implement exception handling.</p>
3227 immediately halt. The dynamic call stack is then searched for the
3229 Once found, execution continues at the "exceptional" destination block
3231 instruction in the dynamic call chain, undefined behavior results.</p>
3233 <p>Note that the code generator does not yet completely support unwind, and
3234 that the invoke/unwind semantics are likely to change in future versions.</p>
3236 </div>
3238 <!-- _______________________________________________________________________ -->
3240 <h4>
3242 </h4>
3244 <div>
3247 <pre>
3248 unreachable
3249 </pre>
3253 instruction is used to inform the optimizer that a particular portion of the
3254 code is not reachable. This can be used to indicate that the code after a
3255 no-return function cannot be reached, and other facts.</p>
3260 </div>
3262 </div>
3264 <!-- ======================================================================= -->
3265 <h3>
3267 </h3>
3269 <div>
3271 <p>Binary operators are used to do most of the computation in a program. They
3272 require two operands of the same type, execute an operation on them, and
3273 produce a single value. The operands might represent multiple data, as is
3275 has the same type as its operands.</p>
3279 <!-- _______________________________________________________________________ -->
3280 <h4>
3282 </h4>
3284 <div>
3287 <pre>
3292 </pre>
3300 integer values. Both arguments must have identical types.</p>
3305 <p>If the sum has unsigned overflow, the result returned is the mathematical
3308 <p>Because LLVM integers use a two's complement representation, this instruction
3309 is appropriate for both signed and unsigned integers.</p>
3315 respectively, occurs.</p>
3318 <pre>
3320 </pre>
3322 </div>
3324 <!-- _______________________________________________________________________ -->
3325 <h4>
3327 </h4>
3329 <div>
3332 <pre>
3334 </pre>
3342 floating point values. Both arguments must have identical types.</p>
3348 <pre>
3350 </pre>
3352 </div>
3354 <!-- _______________________________________________________________________ -->
3355 <h4>
3357 </h4>
3359 <div>
3362 <pre>
3367 </pre>
3371 operands.</p>
3375 representations.</p>
3380 integer values. Both arguments must have identical types.</p>
3385 <p>If the difference has unsigned overflow, the result returned is the
3387 result.</p>
3389 <p>Because LLVM integers use a two's complement representation, this instruction
3390 is appropriate for both signed and unsigned integers.</p>
3396 respectively, occurs.</p>
3399 <pre>
3402 </pre>
3404 </div>
3406 <!-- _______________________________________________________________________ -->
3407 <h4>
3409 </h4>
3411 <div>
3414 <pre>
3416 </pre>
3420 operands.</p>
3424 representations.</p>
3429 floating point values. Both arguments must have identical types.</p>
3435 <pre>
3438 </pre>
3440 </div>
3442 <!-- _______________________________________________________________________ -->
3443 <h4>
3445 </h4>
3447 <div>
3450 <pre>
3455 </pre>
3463 integer values. Both arguments must have identical types.</p>
3468 <p>If the result of the multiplication has unsigned overflow, the result
3470 width of the result.</p>
3472 <p>Because LLVM integers use a two's complement representation, and the result
3473 is the same width as the operands, this instruction returns the correct
3474 result for both signed and unsigned integers. If a full product
3476 be sign-extended or zero-extended as appropriate to the width of the full
3477 product.</p>
3483 respectively, occurs.</p>
3486 <pre>
3488 </pre>
3490 </div>
3492 <!-- _______________________________________________________________________ -->
3493 <h4>
3495 </h4>
3497 <div>
3500 <pre>
3502 </pre>
3510 floating point values. Both arguments must have identical types.</p>
3516 <pre>
3518 </pre>
3520 </div>
3522 <!-- _______________________________________________________________________ -->
3523 <h4>
3525 </h4>
3527 <div>
3530 <pre>
3533 </pre>
3541 values. Both arguments must have identical types.</p>
3546 <p>Note that unsigned integer division and signed integer division are distinct
3557 <pre>
3559 </pre>
3561 </div>
3563 <!-- _______________________________________________________________________ -->
3564 <h4>
3566 </h4>
3568 <div>
3571 <pre>
3574 </pre>
3582 values. Both arguments must have identical types.</p>
3585 <p>The value produced is the signed integer quotient of the two operands rounded
3586 towards zero.</p>
3588 <p>Note that signed integer division and unsigned integer division are distinct
3591 <p>Division by zero leads to undefined behavior. Overflow also leads to
3592 undefined behavior; this is a rare case, but can occur, for example, by doing
3597 be rounded.</p>
3600 <pre>
3602 </pre>
3604 </div>
3606 <!-- _______________________________________________________________________ -->
3607 <h4>
3609 </h4>
3611 <div>
3614 <pre>
3616 </pre>
3624 floating point values. Both arguments must have identical types.</p>
3630 <pre>
3632 </pre>
3634 </div>
3636 <!-- _______________________________________________________________________ -->
3637 <h4>
3639 </h4>
3641 <div>
3644 <pre>
3646 </pre>
3650 division of its two arguments.</p>
3655 values. Both arguments must have identical types.</p>
3659 This instruction always performs an unsigned division to get the
3660 remainder.</p>
3662 <p>Note that unsigned integer remainder and signed integer remainder are
3668 <pre>
3670 </pre>
3672 </div>
3674 <!-- _______________________________________________________________________ -->
3675 <h4>
3677 </h4>
3679 <div>
3682 <pre>
3684 </pre>
3688 division of its two operands. This instruction can also take
3690 elements must be integers.</p>
3695 values. Both arguments must have identical types.</p>
3702 For more information about the difference,
3704 Math Forum</a>. For a table of how this is implemented in various languages,
3708 <p>Note that signed integer remainder and unsigned integer remainder are
3711 <p>Taking the remainder of a division by zero leads to undefined behavior.
3712 Overflow also leads to undefined behavior; this is a rare case, but can
3713 occur, for example, by taking the remainder of a 32-bit division of
3715 lets srem be implemented using instructions that return both the result of
3716 the division and the remainder.)</p>
3719 <pre>
3721 </pre>
3723 </div>
3725 <!-- _______________________________________________________________________ -->
3726 <h4>
3728 </h4>
3730 <div>
3733 <pre>
3735 </pre>
3739 its two operands.</p>
3744 floating point values. Both arguments must have identical types.</p>
3748 has the same sign as the dividend.</p>
3751 <pre>
3753 </pre>
3755 </div>
3757 </div>
3759 <!-- ======================================================================= -->
3760 <h3>
3762 </h3>
3764 <div>
3766 <p>Bitwise binary operators are used to do various forms of bit-twiddling in a
3767 program. They are generally very efficient instructions and can commonly be
3768 strength reduced from other instructions. They require two operands of the
3769 same type, execute an operation on them, and produce a single value. The
3770 resulting value is the same type as its operands.</p>
3772 <!-- _______________________________________________________________________ -->
3773 <h4>
3775 </h4>
3777 <div>
3780 <pre>
3785 </pre>
3789 a specified number of bits.</p>
3799 is (statically or dynamically) negative or equal to or larger than the number
3808 with the resultant sign bit. As such, NUW/NSW have the same semantics as
3809 they would if the shift were expressed as a mul instruction with the same
3813 <pre>
3818 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> <i>; yields: result=<2 x i32> < i32 2, i32 4></i>
3819 </pre>
3821 </div>
3823 <!-- _______________________________________________________________________ -->
3824 <h4>
3826 </h4>
3828 <div>
3831 <pre>
3834 </pre>
3838 operand shifted to the right a specified number of bits with zero fill.</p>
3846 <p>This instruction always performs a logical shift right operation. The most
3847 significant bits of the result will be filled with zero bits after the shift.
3855 shifted out are non-zero.</p>
3859 <pre>
3865 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> <i>; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1></i>
3866 </pre>
3868 </div>
3870 <!-- _______________________________________________________________________ -->
3871 <h4>
3873 </h4>
3875 <div>
3878 <pre>
3881 </pre>
3885 operand shifted to the right a specified number of bits with sign
3886 extension.</p>
3894 <p>This instruction always performs an arithmetic shift right operation, The
3895 most significant bits of the result will be filled with the sign bit
3903 shifted out are non-zero.</p>
3906 <pre>
3912 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> <i>; yields: result=<2 x i32> < i32 -1, i32 0></i>
3913 </pre>
3915 </div>
3917 <!-- _______________________________________________________________________ -->
3918 <h4>
3920 </h4>
3922 <div>
3925 <pre>
3927 </pre>
3931 operands.</p>
3936 values. Both arguments must have identical types.</p>
3942 <tbody>
3943 <tr>
3947 </tr>
3948 <tr>
3952 </tr>
3953 <tr>
3957 </tr>
3958 <tr>
3962 </tr>
3963 <tr>
3967 </tr>
3968 </tbody>
3969 </table>
3972 <pre>
3976 </pre>
3977 </div>
3978 <!-- _______________________________________________________________________ -->
3979 <h4>
3981 </h4>
3983 <div>
3986 <pre>
3988 </pre>
3992 two operands.</p>
3997 values. Both arguments must have identical types.</p>
4003 <tbody>
4004 <tr>
4008 </tr>
4009 <tr>
4013 </tr>
4014 <tr>
4018 </tr>
4019 <tr>
4023 </tr>
4024 <tr>
4028 </tr>
4029 </tbody>
4030 </table>
4033 <pre>
4037 </pre>
4039 </div>
4041 <!-- _______________________________________________________________________ -->
4042 <h4>
4044 </h4>
4046 <div>
4049 <pre>
4051 </pre>
4061 values. Both arguments must have identical types.</p>
4067 <tbody>
4068 <tr>
4072 </tr>
4073 <tr>
4077 </tr>
4078 <tr>
4082 </tr>
4083 <tr>
4087 </tr>
4088 <tr>
4092 </tr>
4093 </tbody>
4094 </table>
4097 <pre>
4102 </pre>
4104 </div>
4106 </div>
4108 <!-- ======================================================================= -->
4109 <h3>
4111 </h3>
4113 <div>
4115 <p>LLVM supports several instructions to represent vector operations in a
4116 target-independent manner. These instructions cover the element-access and
4117 vector-specific operations needed to process vectors effectively. While LLVM
4118 does directly support these vector operations, many sophisticated algorithms
4119 will want to use target-specific intrinsics to take full advantage of a
4120 specific target.</p>
4122 <!-- _______________________________________________________________________ -->
4123 <h4>
4125 </h4>
4127 <div>
4130 <pre>
4131 <result> = extractelement <n x <ty>> <val>, i32 <idx> <i>; yields <ty></i>
4132 </pre>
4136 from a vector at a specified index.</p>
4142 indicating the position from which to extract the element. The index may be
4143 a variable.</p>
4146 <p>The result is a scalar of the same type as the element type of
4149 results are undefined.</p>
4152 <pre>
4154 </pre>
4156 </div>
4158 <!-- _______________________________________________________________________ -->
4159 <h4>
4161 </h4>
4163 <div>
4166 <pre>
4167 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> <i>; yields <n x <ty>></i>
4168 </pre>
4172 vector at a specified index.</p>
4177 whose type must equal the element type of the first operand. The third
4178 operand is an index indicating the position at which to insert the value.
4179 The index may be a variable.</p>
4185 results are undefined.</p>
4188 <pre>
4189 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 <i>; yields <4 x i32></i>
4190 </pre>
4192 </div>
4194 <!-- _______________________________________________________________________ -->
4195 <h4>
4197 </h4>
4199 <div>
4202 <pre>
4203 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> <i>; yields <m x <ty>></i>
4204 </pre>
4208 from two input vectors, returning a vector with the same element type as the
4209 input and length that is the same as the shuffle mask.</p>
4213 with types that match each other. The third argument is a shuffle mask whose
4214 element type is always 'i32'. The result of the instruction is a vector
4215 whose length is the same as the shuffle mask and whose element type is the
4216 same as the element type of the first two operands.</p>
4218 <p>The shuffle mask operand is required to be a constant vector with either
4219 constant integer or undef values.</p>
4222 <p>The elements of the two input vectors are numbered from left to right across
4223 both of the vectors. The shuffle mask operand specifies, for each element of
4224 the result vector, which element of the two input vectors the result element
4225 gets. The element selector may be undef (meaning "don't care") and the
4226 second operand may be undef if performing a shuffle from only one vector.</p>
4229 <pre>
4233 <4 x i32> <i32 0, i32 1, i32 2, i32 3> <i>; yields <4 x i32></i> - Identity shuffle.
4237 <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>
4238 </pre>
4240 </div>
4242 </div>
4244 <!-- ======================================================================= -->
4245 <h3>
4247 </h3>
4249 <div>
4251 <p>LLVM supports several instructions for working with
4254 <!-- _______________________________________________________________________ -->
4255 <h4>
4257 </h4>
4259 <div>
4262 <pre>
4264 </pre>
4274 specify which value to extract in a similar manner as indices in a
4277 <ul>
4278 <li>Since the value being indexed is not a pointer, the first index is
4279 omitted and assumed to be zero.</li>
4281 <li>Not only struct indices but also array indices must be in
4282 bounds.</li>
4283 </ul>
4286 <p>The result is the value at the position in the aggregate specified by the
4287 index operands.</p>
4290 <pre>
4292 </pre>
4294 </div>
4296 <!-- _______________________________________________________________________ -->
4297 <h4>
4299 </h4>
4301 <div>
4304 <pre>
4305 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx> <i>; yields <aggregate type></i>
4306 </pre>
4316 value to insert. The following operands are constant indices indicating
4317 the position at which to insert the value in a similar manner as indices in a
4319 value to insert must have the same type as the value identified by the
4320 indices.</p>
4328 <pre>
4331 </pre>
4333 </div>
4335 </div>
4337 <!-- ======================================================================= -->
4338 <h3>
4340 </h3>
4342 <div>
4344 <p>A key design point of an SSA-based representation is how it represents
4345 memory. In LLVM, no memory locations are in SSA form, which makes things
4346 very simple. This section describes how to read, write, and allocate
4347 memory in LLVM.</p>
4349 <!-- _______________________________________________________________________ -->
4350 <h4>
4352 </h4>
4354 <div>
4357 <pre>
4358 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] <i>; yields {type*}:result</i>
4359 </pre>
4363 currently executing function, to be automatically released when this function
4364 returns to its caller. The object is always allocated in the generic address
4365 space (address space zero).</p>
4370 runtime stack, returning a pointer of the appropriate type to the program.
4371 If "NumElements" is specified, it is the number of elements allocated,
4372 otherwise "NumElements" is defaulted to be one. If a constant alignment is
4373 specified, the value result of the allocation is guaranteed to be aligned to
4374 at least that boundary. If not specified, or if zero, the target can choose
4375 to align the allocation on any convenient boundary compatible with the
4376 type.</p>
4381 <p>Memory is allocated; a pointer is returned. The operation is undefined if
4383 memory is automatically released when the function returns. The
4385 variables that must have an address available. When the function returns
4388 reclaimed. Allocating zero bytes is legal, but the result is undefined.</p>
4391 <pre>
4396 </pre>
4398 </div>
4400 <!-- _______________________________________________________________________ -->
4401 <h4>
4403 </h4>
4405 <div>
4408 <pre>
4409 <result> = load <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>]
4410 <result> = volatile load <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>]
4412 </pre>
4419 from which to load. The pointer must point to
4426 operation (that is, the alignment of the memory address). A value of 0 or an
4428 alignment for the target. It is the responsibility of the code emitter to
4429 ensure that the alignment information is correct. Overestimating the
4430 alignment results in undefined behavior. Underestimating the alignment may
4437 and code generator that this load is not expected to be reused in the cache.
4438 The code generator may select special instructions to save cache bandwidth,
4442 <p>The location of memory pointed to is loaded. If the value being loaded is of
4443 scalar type then the number of bytes read does not exceed the minimum number
4444 of bytes needed to hold all bits of the type. For example, loading an
4447 is undefined if the value was not originally written using a store of the
4448 same type.</p>
4451 <pre>
4455 </pre>
4457 </div>
4459 <!-- _______________________________________________________________________ -->
4460 <h4>
4462 </h4>
4464 <div>
4467 <pre>
4468 store <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] <i>; yields {void}</i>
4469 volatile store <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] <i>; yields {void}</i>
4470 </pre>
4477 and an address at which to store it. The type of the
4486 operation (that is, the alignment of the memory address). A value of 0 or an
4487 omitted "align" argument means that the operation has the preferential
4488 alignment for the target. It is the responsibility of the code emitter to
4489 ensure that the alignment information is correct. Overestimating the
4490 alignment results in an undefined behavior. Underestimating the alignment may
4493 <p>The optional !nontemporal metadata must reference a single metatadata
4495 value 1. The existence of the !nontemporal metatadata on the
4496 instruction tells the optimizer and code generator that this load is
4497 not expected to be reused in the cache. The code generator may
4498 select special instructions to save cache bandwidth, such as the
4499 MOVNT instruction on x86.</p>
4506 does not exceed the minimum number of bytes needed to hold all bits of the
4509 integral number of bytes, it is unspecified what happens to the extra bits
4510 that do not belong to the type, but they will typically be overwritten.</p>
4513 <pre>
4517 </pre>
4519 </div>
4521 <!-- _______________________________________________________________________ -->
4522 <h4>
4524 </h4>
4526 <div>
4529 <pre>
4532 </pre>
4537 It performs address calculation only and does not access memory.</p>
4540 <p>The first argument is always a pointer, and forms the basis of the
4541 calculation. The remaining arguments are indices that indicate which of the
4542 elements of the aggregate object are indexed. The interpretation of each
4543 index is dependent on the type being indexed into. The first index always
4544 indexes the pointer value given as the first argument, the second index
4545 indexes a value of the type pointed to (not necessarily the value directly
4546 pointed to, since the first index can be non-zero), etc. The first type
4547 indexed into must be a pointer value, subsequent types can be arrays,
4548 vectors, and structs. Note that subsequent types being indexed into
4549 can never be pointers, since that would require loading the pointer before
4550 continuing calculation.</p>
4552 <p>The type of each index argument depends on the type it is indexing into.
4555 or vector, integers of any width are allowed, and they are not required to be
4556 constant.</p>
4558 <p>For example, let's consider a C code fragment and how it gets compiled to
4559 LLVM:</p>
4562 struct RT {
4563 char A;
4565 char C;
4566 };
4567 struct ST {
4568 int X;
4569 double Y;
4570 struct RT Z;
4571 };
4573 int *foo(struct ST *s) {
4575 }
4576 </pre>
4584 define i32* @foo(%ST* %s) {
4585 entry:
4587 ret i32* %reg
4588 }
4589 </pre>
4594 }</tt>' type, a structure. The second index indexes into the third element
4596 i8 }</tt>' type, another structure. The third index indexes into the second
4598 array. The two dimensions of the array are subscripted into, yielding an
4602 <p>Note that it is perfectly legal to index partially through a structure,
4603 returning a pointer to an inner element. Because of this, the LLVM code for
4604 the given testcase is equivalent to:</p>
4606 <pre>
4607 define i32* @foo(%ST* %s) {
4613 ret i32* %t5
4614 }
4615 </pre>
4620 or if any of the addresses that would be formed by successive addition of
4621 the offsets implied by the indices to the base address with infinitely
4624 the addresses that point into the object, plus the address one byte past
4625 the end.</p>
4628 the base address with silently-wrapping two's complement arithmetic, and
4630 pointed to by the base pointer. The result value may not necessarily be
4631 used to access memory though, even if it happens to point into allocated
4633 section for more information.</p>
4635 <p>The getelementptr instruction is often confusing. For some more insight into
4639 <pre>
4648 </pre>
4650 </div>
4652 </div>
4654 <!-- ======================================================================= -->
4655 <h3>
4657 </h3>
4659 <div>
4661 <p>The instructions in this category are the conversion instructions (casting)
4662 which all take a single operand and a type. They perform various bit
4663 conversions on the operand.</p>
4665 <!-- _______________________________________________________________________ -->
4666 <h4>
4668 </h4>
4670 <div>
4673 <pre>
4675 </pre>
4684 of the same number of integers.
4687 Equal sized types are not allowed.</p>
4696 <pre>
4700 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> <i>; yields <i8 8, i8 7></i>
4701 </pre>
4703 </div>
4705 <!-- _______________________________________________________________________ -->
4706 <h4>
4708 </h4>
4710 <div>
4713 <pre>
4715 </pre>
4725 of the same number of integers.
4727 the bit size of the destination type,
4737 <pre>
4740 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> <i>; yields <i32 8, i32 7></i>
4741 </pre>
4743 </div>
4745 <!-- _______________________________________________________________________ -->
4746 <h4>
4748 </h4>
4750 <div>
4753 <pre>
4755 </pre>
4763 of the same number of integers.
4765 the bit size of the destination type,
4776 <pre>
4779 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> <i>; yields <i32 8, i32 7></i>
4780 </pre>
4782 </div>
4784 <!-- _______________________________________________________________________ -->
4785 <h4>
4787 </h4>
4789 <div>
4792 <pre>
4794 </pre>
4812 undefined.</p>
4815 <pre>
4818 </pre>
4820 </div>
4822 <!-- _______________________________________________________________________ -->
4823 <h4>
4825 </h4>
4827 <div>
4830 <pre>
4832 </pre>
4836 floating point value.</p>
4842 type must be smaller than the destination type.</p>
4852 <pre>
4855 </pre>
4857 </div>
4859 <!-- _______________________________________________________________________ -->
4860 <h4>
4862 </h4>
4864 <div>
4867 <pre>
4869 </pre>
4885 towards zero) unsigned integer value. If the value cannot fit
4889 <pre>
4893 </pre>
4895 </div>
4897 <!-- _______________________________________________________________________ -->
4898 <h4>
4900 </h4>
4902 <div>
4905 <pre>
4907 </pre>
4925 the results are undefined.</p>
4928 <pre>
4932 </pre>
4934 </div>
4936 <!-- _______________________________________________________________________ -->
4937 <h4>
4939 </h4>
4941 <div>
4944 <pre>
4946 </pre>
4961 integer quantity and converts it to the corresponding floating point
4962 value. If the value cannot fit in the floating point value, the results are
4963 undefined.</p>
4966 <pre>
4969 </pre>
4971 </div>
4973 <!-- _______________________________________________________________________ -->
4974 <h4>
4976 </h4>
4978 <div>
4981 <pre>
4983 </pre>
4998 quantity and converts it to the corresponding floating point value. If the
4999 value cannot fit in the floating point value, the results are undefined.</p>
5002 <pre>
5005 </pre>
5007 </div>
5009 <!-- _______________________________________________________________________ -->
5010 <h4>
5012 </h4>
5014 <div>
5017 <pre>
5019 </pre>
5033 truncating or zero extending that value to the size of the integer type. If
5037 change.</p>
5040 <pre>
5043 </pre>
5045 </div>
5047 <!-- _______________________________________________________________________ -->
5048 <h4>
5050 </h4>
5052 <div>
5055 <pre>
5057 </pre>
5065 value to cast, and a type to cast it to, which must be a
5073 than the size of a pointer then a zero extension is done. If they are the
5077 <pre>
5081 </pre>
5083 </div>
5085 <!-- _______________________________________________________________________ -->
5086 <h4>
5088 </h4>
5090 <div>
5093 <pre>
5095 </pre>
5103 non-aggregate first class value, and a type to cast it to, which must also be
5106 identical. If the source type is a pointer, the destination type must also be
5107 a pointer. This instruction supports bitwise conversion of vectors to
5108 integers and to vectors of other types (as long as they have the same
5109 size).</p>
5116 be converted to other pointer types with this instruction. To convert
5121 <pre>
5125 </pre>
5127 </div>
5129 </div>
5131 <!-- ======================================================================= -->
5132 <h3>
5134 </h3>
5136 <div>
5139 defy better classification.</p>
5141 <!-- _______________________________________________________________________ -->
5142 <h4>
5144 </h4>
5146 <div>
5149 <pre>
5150 <result> = icmp <cond> <ty> <op1>, <op2> <i>; yields {i1} or {<N x i1>}:result</i>
5151 </pre>
5155 boolean values based on comparison of its two integer, integer vector, or
5156 pointer operands.</p>
5160 the condition code indicating the kind of comparison to perform. It is not a
5161 value, just a keyword. The possible condition code are:</p>
5163 <ol>
5174 </ol>
5178 typed. They must also be identical types.</p>
5184 result, as follows:</p>
5186 <ol>
5189 performed.</li>
5193 performed.</li>
5220 </ol>
5223 values are compared as if they were integers.</p>
5225 <p>If the operands are integer vectors, then they are compared element by
5230 <pre>
5237 </pre>
5239 <p>Note that the code generator does not yet support vector types with
5242 </div>
5244 <!-- _______________________________________________________________________ -->
5245 <h4>
5247 </h4>
5249 <div>
5252 <pre>
5253 <result> = fcmp <cond> <ty> <op1>, <op2> <i>; yields {i1} or {<N x i1>}:result</i>
5254 </pre>
5258 values based on comparison of its operands.</p>
5260 <p>If the operands are floating point scalars, then the result type is a boolean
5263 <p>If the operands are floating point vectors, then the result type is a vector
5264 of boolean with the same number of elements as the operands being
5265 compared.</p>
5269 the condition code indicating the kind of comparison to perform. It is not a
5270 value, just a keyword. The possible condition code are:</p>
5272 <ol>
5289 </ol>
5297 identical types.</p>
5302 vectors, then the vectors are compared element by element. Each comparison
5304 follows:</p>
5306 <ol>
5350 </ol>
5353 <pre>
5358 </pre>
5360 <p>Note that the code generator does not yet support vector types with
5363 </div>
5365 <!-- _______________________________________________________________________ -->
5366 <h4>
5368 </h4>
5370 <div>
5373 <pre>
5375 </pre>
5379 SSA graph representing the function.</p>
5382 <p>The type of the incoming values is specified with the first type field. After
5384 one pair for each predecessor basic block of the current block. Only values
5386 arguments to the PHI node. Only labels may be used as the label
5387 arguments.</p>
5389 <p>There must be no non-phi instructions between the start of a basic block and
5390 the PHI instructions: i.e. PHI instructions must be first in a basic
5391 block.</p>
5393 <p>For the purposes of the SSA form, the use of each incoming value is deemed to
5394 occur on the edge from the corresponding predecessor block to the current
5396 value on the same edge).</p>
5400 specified by the pair corresponding to the predecessor basic block that
5401 executed just prior to the current block.</p>
5404 <pre>
5405 Loop: ; Infinite loop that counts from 0 on up...
5406 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
5407 %nextindvar = add i32 %indvar, 1
5408 br label %Loop
5409 </pre>
5411 </div>
5413 <!-- _______________________________________________________________________ -->
5414 <h4>
5416 </h4>
5418 <div>
5421 <pre>
5422 <result> = select <i>selty</i> <cond>, <ty> <val1>, <ty> <val2> <i>; yields ty</i>
5425 </pre>
5429 condition, without branching.</p>
5434 values indicating the condition, and two values of the
5436 vectors and the condition is a scalar, then entire vectors are selected, not
5437 individual elements.</p>
5441 first value argument; otherwise, it returns the second value argument.</p>
5443 <p>If the condition is a vector of i1, then the value arguments must be vectors
5444 of the same size, and the selection is done element by element.</p>
5447 <pre>
5449 </pre>
5451 <p>Note that the code generator does not yet support conditions
5452 with vector type.</p>
5454 </div>
5456 <!-- _______________________________________________________________________ -->
5457 <h4>
5459 </h4>
5461 <div>
5464 <pre>
5465 <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>]
5466 </pre>
5474 <ol>
5476 access any allocas or varargs in the caller. Note that calls may be
5477 marked "tail" even if they do not occur before
5479 present, the function call is eligible for tail call optimization,
5481 optimized into a jump</a>. The code generator may optimize calls marked
5483 sibling call optimization</a> when the caller and callee have
5484 matching signatures, or 2) forced tail call optimization when the
5485 following extra requirements are met:
5486 <ul>
5487 <li>Caller and callee both have the calling
5489 <li>The call is in tail position (ret immediately follows call and ret
5490 uses value of call or is void).</li>
5494 constraints are met.</a></li>
5495 </ul>
5496 </li>
5499 convention</a> the call should use. If none is specified, the call
5500 defaults to using C calling conventions. The calling convention of the
5501 call must match the calling convention of the target function, or else the
5502 behavior is undefined.</li>
5509 type of the return value. Functions that return no value are marked
5513 being invoked. The argument types must match the types implied by this
5514 signature. This type can be omitted if the function is not varargs and if
5515 the function type does not return a pointer to a function.</li>
5518 be invoked. In most cases, this is a direct function invocation, but
5520 to function value.</li>
5523 signature argument types and parameter attributes. All arguments must be
5525 signature indicates the function accepts a variable number of arguments,
5526 the extra arguments can be specified.</li>
5531 </ol>
5535 a specified function, with its incoming arguments bound to the specified
5537 function, control flow continues with the instruction after the function
5538 call, and the return value of the function is bound to the result
5539 argument.</p>
5542 <pre>
5543 %retval = call i32 @test(i32 %argc)
5547 call void %foo(i8 97 signext)
5549 %struct.A = type { i32, i8 }
5555 </pre>
5557 <p>llvm treats calls to some functions with names and arguments that match the
5558 standard C99 library as being the C99 library functions, and may perform
5559 optimizations or generate code for them under that assumption. This is
5560 something we'd like to change in the future to provide better support for
5561 freestanding environments and non-C-based languages.</p>
5563 </div>
5565 <!-- _______________________________________________________________________ -->
5566 <h4>
5568 </h4>
5570 <div>
5573 <pre>
5575 </pre>
5579 the "variable argument" area of a function call. It is used to implement the
5584 argument. It returns a value of the specified argument type and increments
5591 to the next argument. For more information, see the variable argument
5594 <p>It is legal for this instruction to be called in a function which does not
5596 function.</p>
5600 argument.</p>
5605 <p>Note that the code generator does not yet fully support va_arg on many
5606 targets. Also, it does not currently support va_arg with aggregate types on
5607 any target.</p>
5609 </div>
5611 </div>
5613 </div>
5615 <!-- *********************************************************************** -->
5617 <!-- *********************************************************************** -->
5619 <div>
5622 well known names and semantics and are required to follow certain
5623 restrictions. Overall, these intrinsics represent an extension mechanism for
5624 the LLVM language that does not require changing all of the transformations
5625 in LLVM when adding to the language (or the bitcode reader/writer, the
5626 parser, etc...).</p>
5629 prefix is reserved in LLVM for intrinsic names; thus, function names may not
5630 begin with this prefix. Intrinsic functions must always be external
5631 functions: you cannot define the body of intrinsic functions. Intrinsic
5632 functions may only be used in call or invoke instructions: it is illegal to
5633 take the address of an intrinsic function. Additionally, because intrinsic
5634 functions are part of the LLVM language, it is required if any are added that
5635 they be documented here.</p>
5637 <p>Some intrinsic functions can be overloaded, i.e., the intrinsic represents a
5638 family of functions that perform the same operation but on different data
5639 types. Because LLVM can represent over 8 million different integer types,
5640 overloading is used commonly to allow an intrinsic function to operate on any
5641 integer type. One or more of the argument types or the result type can be
5642 overloaded to accept any integer type. Argument types may also be defined as
5643 exactly matching a previous argument's type or the result type. This allows
5644 an intrinsic function which accepts multiple arguments, but needs all of them
5645 to be of the same type, to only be overloaded with respect to a single
5646 argument or the result.</p>
5648 <p>Overloaded intrinsics will have the names of its overloaded argument types
5649 encoded into its function name, each preceded by a period. Only those types
5650 which are overloaded result in a name suffix. Arguments whose type is matched
5652 can take an integer of any width and returns an integer of exactly the same
5653 integer width. This leads to a family of functions such as
5655 %val)</tt>. Only one type, the return type, is overloaded, and only one type
5656 suffix is required. Because the argument's type is matched against the return
5657 type, it does not require its own name suffix.</p>
5659 <p>To learn how to add an intrinsic function, please see the
5662 <!-- ======================================================================= -->
5663 <h3>
5665 </h3>
5667 <div>
5669 <p>Variable argument support is defined in LLVM with
5671 intrinsic functions. These functions are related to the similarly named
5674 <p>All of these functions operate on arguments that use a target-specific value
5675 type "<tt>va_list</tt>". The LLVM assembly language reference manual does
5676 not define what this type is, so all transformations should be prepared to
5677 handle these functions regardless of the type used.</p>
5680 instruction and the variable argument handling intrinsic functions are
5681 used.</p>
5684 define i32 @test(i32 %X, ...) {
5685 ; Initialize variable argument processing
5686 %ap = alloca i8*
5687 %ap2 = bitcast i8** %ap to i8*
5688 call void @llvm.va_start(i8* %ap2)
5690 ; Read a single integer argument
5691 %tmp = va_arg i8** %ap, i32
5693 ; Demonstrate usage of llvm.va_copy and llvm.va_end
5694 %aq = alloca i8*
5695 %aq2 = bitcast i8** %aq to i8*
5696 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
5697 call void @llvm.va_end(i8* %aq2)
5699 ; Stop processing of arguments.
5700 call void @llvm.va_end(i8* %ap2)
5701 ret i32 %tmp
5702 }
5704 declare void @llvm.va_start(i8*)
5705 declare void @llvm.va_copy(i8*, i8*)
5706 declare void @llvm.va_end(i8*)
5707 </pre>
5709 <!-- _______________________________________________________________________ -->
5710 <h4>
5712 </h4>
5715 <div>
5718 <pre>
5720 </pre>
5731 macro available in C. In a target-dependent way, it initializes
5735 need to know the last argument of the function as the compiler can figure
5736 that out.</p>
5738 </div>
5740 <!-- _______________________________________________________________________ -->
5741 <h4>
5743 </h4>
5745 <div>
5748 <pre>
5750 </pre>
5754 which has been initialized previously
5763 macro available in C. In a target-dependent way, it destroys
5769 </div>
5771 <!-- _______________________________________________________________________ -->
5772 <h4>
5774 </h4>
5776 <div>
5779 <pre>
5781 </pre>
5785 from the source argument list to the destination argument list.</p>
5790 from.</p>
5794 macro available in C. In a target-dependent way, it copies the
5796 element. This intrinsic is necessary because
5798 arbitrarily complex and require, for example, memory allocation.</p>
5800 </div>
5802 </div>
5804 <!-- ======================================================================= -->
5805 <h3>
5807 </h3>
5809 <div>
5812 Collection</a> (GC) requires the implementation and generation of these
5814 roots on the stack</a>, as well as garbage collector implementations that
5816 barriers. Front-ends for type-safe garbage collected languages should generate
5817 these intrinsics to make use of the LLVM garbage collectors. For more details,
5821 <p>The garbage collection intrinsics only operate on objects in the generic
5822 address space (address space zero).</p>
5824 <!-- _______________________________________________________________________ -->
5825 <h4>
5827 </h4>
5829 <div>
5832 <pre>
5833 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
5834 </pre>
5838 the code generator, and allows some metadata to be associated with it.</p>
5841 <p>The first argument specifies the address of a stack object that contains the
5842 root pointer. The second pointer (which must be either a constant or a
5843 global value address) contains the meta-data to be associated with the
5844 root.</p>
5848 location. At compile-time, the code generator generates information to allow
5853 </div>
5855 <!-- _______________________________________________________________________ -->
5856 <h4>
5858 </h4>
5860 <div>
5863 <pre>
5864 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
5865 </pre>
5869 locations, allowing garbage collector implementations that require read
5870 barriers.</p>
5873 <p>The second argument is the address to read from, which should be an address
5874 allocated from the garbage collector. The first object is a pointer to the
5875 start of the referenced object, if needed by the language runtime (otherwise
5876 null).</p>
5880 instruction, but may be replaced with substantially more complex code by the
5885 </div>
5887 <!-- _______________________________________________________________________ -->
5888 <h4>
5890 </h4>
5892 <div>
5895 <pre>
5896 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
5897 </pre>
5901 locations, allowing garbage collector implementations that require write
5902 barriers (such as generational or reference counting collectors).</p>
5905 <p>The first argument is the reference to store, the second is the start of the
5906 object to store it to, and the third is the address of the field of Obj to
5907 store to. If the runtime does not require a pointer to the object, Obj may
5908 be null.</p>
5912 instruction, but may be replaced with substantially more complex code by the
5917 </div>
5919 </div>
5921 <!-- ======================================================================= -->
5922 <h3>
5924 </h3>
5926 <div>
5928 <p>These intrinsics are provided by LLVM to expose special features that may
5929 only be implemented with code generator support.</p>
5931 <!-- _______________________________________________________________________ -->
5932 <h4>
5934 </h4>
5936 <div>
5939 <pre>
5941 </pre>
5945 target-specific value indicating the return address of the current function
5946 or one of its callers.</p>
5949 <p>The argument to this intrinsic indicates which function to return the address
5950 for. Zero indicates the calling function, one indicates its caller, etc.
5955 indicating the return address of the specified call frame, or zero if it
5956 cannot be identified. The value returned by this intrinsic is likely to be
5957 incorrect or 0 for arguments other than zero, so it should only be used for
5958 debugging purposes.</p>
5960 <p>Note that calling this intrinsic does not prevent function inlining or other
5961 aggressive transformations, so the value returned may not be that of the
5962 obvious source-language caller.</p>
5964 </div>
5966 <!-- _______________________________________________________________________ -->
5967 <h4>
5969 </h4>
5971 <div>
5974 <pre>
5976 </pre>
5980 target-specific frame pointer value for the specified stack frame.</p>
5983 <p>The argument to this intrinsic indicates which function to return the frame
5984 pointer for. Zero indicates the calling function, one indicates its caller,
5989 indicating the frame address of the specified call frame, or zero if it
5990 cannot be identified. The value returned by this intrinsic is likely to be
5991 incorrect or 0 for arguments other than zero, so it should only be used for
5992 debugging purposes.</p>
5994 <p>Note that calling this intrinsic does not prevent function inlining or other
5995 aggressive transformations, so the value returned may not be that of the
5996 obvious source-language caller.</p>
5998 </div>
6000 <!-- _______________________________________________________________________ -->
6001 <h4>
6003 </h4>
6005 <div>
6008 <pre>
6009 declare i8* @llvm.stacksave()
6010 </pre>
6014 of the function stack, for use
6016 useful for implementing language features like scoped automatic variable
6017 sized arrays in C99.</p>
6020 <p>This intrinsic returns a opaque pointer value that can be passed
6028 </div>
6030 <!-- _______________________________________________________________________ -->
6031 <h4>
6033 </h4>
6035 <div>
6038 <pre>
6039 declare void @llvm.stackrestore(i8* %ptr)
6040 </pre>
6044 the function stack to the state it was in when the
6046 executed. This is useful for implementing language features like scoped
6047 automatic variable sized arrays in C99.</p>
6050 <p>See the description
6053 </div>
6055 <!-- _______________________________________________________________________ -->
6056 <h4>
6058 </h4>
6060 <div>
6063 <pre>
6065 </pre>
6069 insert a prefetch instruction if supported; otherwise, it is a noop.
6070 Prefetches have no effect on the behavior of the program but can change its
6071 performance characteristics.</p>
6081 <p>This intrinsic does not modify the behavior of the program. In particular,
6082 prefetches cannot trap and do not produce a value. On targets that support
6083 this intrinsic, the prefetch can provide hints to the processor cache for
6084 better performance.</p>
6086 </div>
6088 <!-- _______________________________________________________________________ -->
6089 <h4>
6091 </h4>
6093 <div>
6096 <pre>
6098 </pre>
6102 Counter (PC) in a region of code to simulators and other tools. The method
6103 is target specific, but it is expected that the marker will use exported
6104 symbols to transmit the PC of the marker. The marker makes no guarantees
6105 that it will remain with any specific instruction after optimizations. It is
6106 possible that the presence of a marker will inhibit optimizations. The
6107 intended use is to be inserted after optimizations to allow correlations of
6108 simulation runs.</p>
6114 <p>This intrinsic does not modify the behavior of the program. Backends that do
6115 not support this intrinsic may ignore it.</p>
6117 </div>
6119 <!-- _______________________________________________________________________ -->
6120 <h4>
6122 </h4>
6124 <div>
6127 <pre>
6128 declare i64 @llvm.readcyclecounter()
6129 </pre>
6133 counter register (or similar low latency, high accuracy clocks) on those
6134 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6135 should map to RPCC. As the backing counters overflow quickly (on the order
6139 <p>When directly supported, reading the cycle counter should not modify any
6140 memory. Implementations are allowed to either return a application specific
6141 value or a system wide value. On backends without support, this is lowered
6144 </div>
6146 </div>
6148 <!-- ======================================================================= -->
6149 <h3>
6151 </h3>
6153 <div>
6155 <p>LLVM provides intrinsics for a few important standard C library functions.
6156 These intrinsics allow source-language front-ends to pass information about
6157 the alignment of the pointer arguments to the code generator, providing
6158 opportunity for more efficient code generation.</p>
6160 <!-- _______________________________________________________________________ -->
6161 <h4>
6163 </h4>
6165 <div>
6169 integer bit width and for different address spaces. Not all targets support
6170 all bit widths however.</p>
6172 <pre>
6177 </pre>
6181 source location to the destination location.</p>
6184 intrinsics do not return a value, takes extra alignment/isvolatile arguments
6185 and the pointers can be in specified address spaces.</p>
6189 <p>The first argument is a pointer to the destination, the second is a pointer
6190 to the source. The third argument is an integer argument specifying the
6191 number of bytes to copy, the fourth argument is the alignment of the
6192 source and destination locations, and the fifth is a boolean indicating a
6193 volatile access.</p>
6196 then the caller guarantees that both the source and destination pointers are
6197 aligned to that boundary.</p>
6201 The detailed access behavior is not very cleanly specified and it is unwise
6202 to depend on it.</p>
6207 source location to the destination location, which are not allowed to
6208 overlap. It copies "len" bytes of memory over. If the argument is known to
6209 be aligned to some boundary, this can be specified as the fourth argument,
6212 </div>
6214 <!-- _______________________________________________________________________ -->
6215 <h4>
6217 </h4>
6219 <div>
6222 <p>This is an overloaded intrinsic. You can use llvm.memmove on any integer bit
6223 width and for different address space. Not all targets support all bit
6224 widths however.</p>
6226 <pre>
6231 </pre>
6235 source location to the destination location. It is similar to the
6237 overlap.</p>
6240 intrinsics do not return a value, takes extra alignment/isvolatile arguments
6241 and the pointers can be in specified address spaces.</p>
6245 <p>The first argument is a pointer to the destination, the second is a pointer
6246 to the source. The third argument is an integer argument specifying the
6247 number of bytes to copy, the fourth argument is the alignment of the
6248 source and destination locations, and the fifth is a boolean indicating a
6249 volatile access.</p>
6252 then the caller guarantees that the source and destination pointers are
6253 aligned to that boundary.</p>
6257 The detailed access behavior is not very cleanly specified and it is unwise
6258 to depend on it.</p>
6263 source location to the destination location, which may overlap. It copies
6264 "len" bytes of memory over. If the argument is known to be aligned to some
6265 boundary, this can be specified as the fourth argument, otherwise it should
6268 </div>
6270 <!-- _______________________________________________________________________ -->
6271 <h4>
6273 </h4>
6275 <div>
6278 <p>This is an overloaded intrinsic. You can use llvm.memset on any integer bit
6279 width and for different address spaces. However, not all targets support all
6280 bit widths.</p>
6282 <pre>
6287 </pre>
6291 particular byte value.</p>
6294 intrinsic does not return a value and takes extra alignment/volatile
6295 arguments. Also, the destination can be in an arbitrary address space.</p>
6298 <p>The first argument is a pointer to the destination to fill, the second is the
6299 byte value with which to fill it, the third argument is an integer argument
6300 specifying the number of bytes to fill, and the fourth argument is the known
6301 alignment of the destination location.</p>
6304 then the caller guarantees that the destination pointer is aligned to that
6305 boundary.</p>
6309 The detailed access behavior is not very cleanly specified and it is unwise
6310 to depend on it.</p>
6314 at the destination location. If the argument is known to be aligned to some
6315 boundary, this can be specified as the fourth argument, otherwise it should
6318 </div>
6320 <!-- _______________________________________________________________________ -->
6321 <h4>
6323 </h4>
6325 <div>
6329 floating point or vector of floating point type. Not all targets support all
6330 types however.</p>
6332 <pre>
6333 declare float @llvm.sqrt.f32(float %Val)
6334 declare double @llvm.sqrt.f64(double %Val)
6335 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
6336 declare fp128 @llvm.sqrt.f128(fp128 %Val)
6337 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
6338 </pre>
6344 behavior for negative numbers other than -0.0 (which allows for better
6345 optimization, because there is no need to worry about errno being
6349 <p>The argument and return value are floating point numbers of the same
6350 type.</p>
6353 <p>This function returns the sqrt of the specified operand if it is a
6354 nonnegative floating point number.</p>
6356 </div>
6358 <!-- _______________________________________________________________________ -->
6359 <h4>
6361 </h4>
6363 <div>
6367 floating point or vector of floating point type. Not all targets support all
6368 types however.</p>
6370 <pre>
6371 declare float @llvm.powi.f32(float %Val, i32 %power)
6372 declare double @llvm.powi.f64(double %Val, i32 %power)
6373 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
6374 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
6375 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
6376 </pre>
6380 specified (positive or negative) power. The order of evaluation of
6381 multiplications is not defined. When a vector of floating point type is
6382 used, the second argument remains a scalar integer value.</p>
6385 <p>The second argument is an integer power, and the first is a value to raise to
6386 that power.</p>
6389 <p>This function returns the first value raised to the second power with an
6390 unspecified sequence of rounding operations.</p>
6392 </div>
6394 <!-- _______________________________________________________________________ -->
6395 <h4>
6397 </h4>
6399 <div>
6403 floating point or vector of floating point type. Not all targets support all
6404 types however.</p>
6406 <pre>
6407 declare float @llvm.sin.f32(float %Val)
6408 declare double @llvm.sin.f64(double %Val)
6409 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
6410 declare fp128 @llvm.sin.f128(fp128 %Val)
6411 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
6412 </pre>
6418 <p>The argument and return value are floating point numbers of the same
6419 type.</p>
6422 <p>This function returns the sine of the specified operand, returning the same
6424 in the same way.</p>
6426 </div>
6428 <!-- _______________________________________________________________________ -->
6429 <h4>
6431 </h4>
6433 <div>
6437 floating point or vector of floating point type. Not all targets support all
6438 types however.</p>
6440 <pre>
6441 declare float @llvm.cos.f32(float %Val)
6442 declare double @llvm.cos.f64(double %Val)
6443 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
6444 declare fp128 @llvm.cos.f128(fp128 %Val)
6445 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
6446 </pre>
6452 <p>The argument and return value are floating point numbers of the same
6453 type.</p>
6456 <p>This function returns the cosine of the specified operand, returning the same
6458 in the same way.</p>
6460 </div>
6462 <!-- _______________________________________________________________________ -->
6463 <h4>
6465 </h4>
6467 <div>
6471 floating point or vector of floating point type. Not all targets support all
6472 types however.</p>
6474 <pre>
6475 declare float @llvm.pow.f32(float %Val, float %Power)
6476 declare double @llvm.pow.f64(double %Val, double %Power)
6477 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
6478 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
6479 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
6480 </pre>
6484 specified (positive or negative) power.</p>
6487 <p>The second argument is a floating point power, and the first is a value to
6488 raise to that power.</p>
6491 <p>This function returns the first value raised to the second power, returning
6493 conditions in the same way.</p>
6495 </div>
6497 </div>
6499 <!-- ======================================================================= -->
6500 <h3>
6502 </h3>
6504 <div>
6506 <p>LLVM provides intrinsics for a few important bit manipulation operations.
6507 These allow efficient code generation for some algorithms.</p>
6509 <!-- _______________________________________________________________________ -->
6510 <h4>
6512 </h4>
6514 <div>
6517 <p>This is an overloaded intrinsic function. You can use bswap on any integer
6520 <pre>
6524 </pre>
6528 values with an even number of bytes (positive multiple of 16 bits). These
6529 are useful for performing operations on data that is not in the target's
6530 native byte order.</p>
6534 and low byte of the input i16 swapped. Similarly,
6540 more, respectively).</p>
6542 </div>
6544 <!-- _______________________________________________________________________ -->
6545 <h4>
6547 </h4>
6549 <div>
6552 <p>This is an overloaded intrinsic. You can use llvm.ctpop on any integer bit
6553 width. Not all targets support all bit widths however.</p>
6555 <pre>
6561 </pre>
6565 in a value.</p>
6568 <p>The only argument is the value to be counted. The argument may be of any
6569 integer type. The return type must match the argument type.</p>
6574 </div>
6576 <!-- _______________________________________________________________________ -->
6577 <h4>
6579 </h4>
6581 <div>
6585 integer bit width. Not all targets support all bit widths however.</p>
6587 <pre>
6593 </pre>
6597 leading zeros in a variable.</p>
6600 <p>The only argument is the value to be counted. The argument may be of any
6601 integer type. The return type must match the argument type.</p>
6605 zeros in a variable. If the src == 0 then the result is the size in bits of
6608 </div>
6610 <!-- _______________________________________________________________________ -->
6611 <h4>
6613 </h4>
6615 <div>
6619 integer bit width. Not all targets support all bit widths however.</p>
6621 <pre>
6627 </pre>
6631 trailing zeros.</p>
6634 <p>The only argument is the value to be counted. The argument may be of any
6635 integer type. The return type must match the argument type.</p>
6639 zeros in a variable. If the src == 0 then the result is the size in bits of
6642 </div>
6644 </div>
6646 <!-- ======================================================================= -->
6647 <h3>
6649 </h3>
6651 <div>
6655 <!-- _______________________________________________________________________ -->
6656 <h4>
6659 </a>
6660 </h4>
6662 <div>
6666 on any integer bit width.</p>
6668 <pre>
6669 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
6670 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
6671 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
6672 </pre>
6676 a signed addition of the two arguments, and indicate whether an overflow
6677 occurred during the signed summation.</p>
6680 <p>The arguments (%a and %b) and the first element of the result structure may
6681 be of integer types of any bit width, but they must have the same bit
6682 width. The second element of the result structure must be of
6684 undergo signed addition.</p>
6688 a signed addition of the two variables. They return a structure — the
6689 first element of which is the signed summation, and the second element of
6690 which is a bit specifying if the signed summation resulted in an
6691 overflow.</p>
6694 <pre>
6695 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
6696 %sum = extractvalue {i32, i1} %res, 0
6697 %obit = extractvalue {i32, i1} %res, 1
6698 br i1 %obit, label %overflow, label %normal
6699 </pre>
6701 </div>
6703 <!-- _______________________________________________________________________ -->
6704 <h4>
6707 </a>
6708 </h4>
6710 <div>
6714 on any integer bit width.</p>
6716 <pre>
6717 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
6718 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
6719 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
6720 </pre>
6724 an unsigned addition of the two arguments, and indicate whether a carry
6725 occurred during the unsigned summation.</p>
6728 <p>The arguments (%a and %b) and the first element of the result structure may
6729 be of integer types of any bit width, but they must have the same bit
6730 width. The second element of the result structure must be of
6732 undergo unsigned addition.</p>
6736 an unsigned addition of the two arguments. They return a structure —
6737 the first element of which is the sum, and the second element of which is a
6738 bit specifying if the unsigned summation resulted in a carry.</p>
6741 <pre>
6742 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
6743 %sum = extractvalue {i32, i1} %res, 0
6744 %obit = extractvalue {i32, i1} %res, 1
6745 br i1 %obit, label %carry, label %normal
6746 </pre>
6748 </div>
6750 <!-- _______________________________________________________________________ -->
6751 <h4>
6754 </a>
6755 </h4>
6757 <div>
6761 on any integer bit width.</p>
6763 <pre>
6764 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
6765 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
6766 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
6767 </pre>
6771 a signed subtraction of the two arguments, and indicate whether an overflow
6772 occurred during the signed subtraction.</p>
6775 <p>The arguments (%a and %b) and the first element of the result structure may
6776 be of integer types of any bit width, but they must have the same bit
6777 width. The second element of the result structure must be of
6779 undergo signed subtraction.</p>
6783 a signed subtraction of the two arguments. They return a structure —
6784 the first element of which is the subtraction, and the second element of
6785 which is a bit specifying if the signed subtraction resulted in an
6786 overflow.</p>
6789 <pre>
6790 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
6791 %sum = extractvalue {i32, i1} %res, 0
6792 %obit = extractvalue {i32, i1} %res, 1
6793 br i1 %obit, label %overflow, label %normal
6794 </pre>
6796 </div>
6798 <!-- _______________________________________________________________________ -->
6799 <h4>
6802 </a>
6803 </h4>
6805 <div>
6809 on any integer bit width.</p>
6811 <pre>
6812 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
6813 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
6814 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
6815 </pre>
6819 an unsigned subtraction of the two arguments, and indicate whether an
6820 overflow occurred during the unsigned subtraction.</p>
6823 <p>The arguments (%a and %b) and the first element of the result structure may
6824 be of integer types of any bit width, but they must have the same bit
6825 width. The second element of the result structure must be of
6827 undergo unsigned subtraction.</p>
6831 an unsigned subtraction of the two arguments. They return a structure —
6832 the first element of which is the subtraction, and the second element of
6833 which is a bit specifying if the unsigned subtraction resulted in an
6834 overflow.</p>
6837 <pre>
6838 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
6839 %sum = extractvalue {i32, i1} %res, 0
6840 %obit = extractvalue {i32, i1} %res, 1
6841 br i1 %obit, label %overflow, label %normal
6842 </pre>
6844 </div>
6846 <!-- _______________________________________________________________________ -->
6847 <h4>
6850 </a>
6851 </h4>
6853 <div>
6857 on any integer bit width.</p>
6859 <pre>
6860 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
6861 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
6862 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
6863 </pre>
6868 a signed multiplication of the two arguments, and indicate whether an
6869 overflow occurred during the signed multiplication.</p>
6872 <p>The arguments (%a and %b) and the first element of the result structure may
6873 be of integer types of any bit width, but they must have the same bit
6874 width. The second element of the result structure must be of
6876 undergo signed multiplication.</p>
6880 a signed multiplication of the two arguments. They return a structure —
6881 the first element of which is the multiplication, and the second element of
6882 which is a bit specifying if the signed multiplication resulted in an
6883 overflow.</p>
6886 <pre>
6887 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
6888 %sum = extractvalue {i32, i1} %res, 0
6889 %obit = extractvalue {i32, i1} %res, 1
6890 br i1 %obit, label %overflow, label %normal
6891 </pre>
6893 </div>
6895 <!-- _______________________________________________________________________ -->
6896 <h4>
6899 </a>
6900 </h4>
6902 <div>
6906 on any integer bit width.</p>
6908 <pre>
6909 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
6910 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
6911 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
6912 </pre>
6916 a unsigned multiplication of the two arguments, and indicate whether an
6917 overflow occurred during the unsigned multiplication.</p>
6920 <p>The arguments (%a and %b) and the first element of the result structure may
6921 be of integer types of any bit width, but they must have the same bit
6922 width. The second element of the result structure must be of
6924 undergo unsigned multiplication.</p>
6928 an unsigned multiplication of the two arguments. They return a structure
6929 — the first element of which is the multiplication, and the second
6930 element of which is a bit specifying if the unsigned multiplication resulted
6931 in an overflow.</p>
6934 <pre>
6935 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
6936 %sum = extractvalue {i32, i1} %res, 0
6937 %obit = extractvalue {i32, i1} %res, 1
6938 br i1 %obit, label %overflow, label %normal
6939 </pre>
6941 </div>
6943 </div>
6945 <!-- ======================================================================= -->
6946 <h3>
6948 </h3>
6950 <div>
6952 <p>Half precision floating point is a storage-only format. This means that it is
6953 a dense encoding (in memory) but does not support computation in the
6954 format.</p>
6956 <p>This means that code must first load the half-precision floating point
6957 value as an i16, then convert it to float with <a
6959 Computation can then be performed on the float value (including extending to
6960 double etc). To store the value back to memory, it is first converted to
6961 float if needed, then converted to i16 with
6963 storing as an i16 value.</p>
6965 <!-- _______________________________________________________________________ -->
6966 <h4>
6969 </a>
6970 </h4>
6972 <div>
6975 <pre>
6976 declare i16 @llvm.convert.to.fp16(f32 %a)
6977 </pre>
6981 a conversion from single precision floating point format to half precision
6982 floating point format.</p>
6985 <p>The intrinsic function contains single argument - the value to be
6986 converted.</p>
6990 a conversion from single precision floating point format to half precision
6992 contains the converted number.</p>
6995 <pre>
6996 %res = call i16 @llvm.convert.to.fp16(f32 %a)
6997 store i16 %res, i16* @x, align 2
6998 </pre>
7000 </div>
7002 <!-- _______________________________________________________________________ -->
7003 <h4>
7006 </a>
7007 </h4>
7009 <div>
7012 <pre>
7013 declare f32 @llvm.convert.from.fp16(i16 %a)
7014 </pre>
7018 a conversion from half precision floating point format to single precision
7019 floating point format.</p>
7022 <p>The intrinsic function contains single argument - the value to be
7023 converted.</p>
7027 conversion from half single precision floating point format to single
7028 precision floating point format. The input half-float value is represented by
7032 <pre>
7033 %a = load i16* @x, align 2
7034 %res = call f32 @llvm.convert.from.fp16(i16 %a)
7035 </pre>
7037 </div>
7039 </div>
7041 <!-- ======================================================================= -->
7042 <h3>
7044 </h3>
7046 <div>
7049 prefix), are described in
7053 </div>
7055 <!-- ======================================================================= -->
7056 <h3>
7058 </h3>
7060 <div>
7062 <p>The LLVM exception handling intrinsics (which all start with
7067 </div>
7069 <!-- ======================================================================= -->
7070 <h3>
7072 </h3>
7074 <div>
7076 <p>This intrinsic makes it possible to excise one parameter, marked with
7078 The result is a callable
7079 function pointer lacking the nest parameter - the caller does not need to
7080 provide a value for it. Instead, the value to use is stored in advance in a
7081 "trampoline", a block of memory usually allocated on the stack, which also
7082 contains code to splice the nest value into the argument list. This is used
7083 to implement the GCC nested function address extension.</p>
7085 <p>For example, if the function is
7088 follows:</p>
7093 %p = call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8* nest , i32, i32)* @f to i8*), i8* %nval)
7094 %fp = bitcast i8* %p to i32 (i32, i32)*
7095 </pre>
7100 <!-- _______________________________________________________________________ -->
7101 <h4>
7104 </a>
7105 </h4>
7107 <div>
7110 <pre>
7112 </pre>
7116 function pointer suitable for executing it.</p>
7121 sufficiently aligned block of memory; this memory is written to by the
7122 intrinsic. Note that the size and the alignment are target-specific - LLVM
7123 currently provides no portable way of determining them, so a front-end that
7124 generates this intrinsic needs to have some target-specific knowledge.
7130 dependent code, turning it into a function. A pointer to this function is
7132 function pointer type</a> before being called. The new function's signature
7135 is allowed, and it must be of pointer type. Calling the new function is
7140 returned function pointer is undefined.</p>
7142 </div>
7144 </div>
7146 <!-- ======================================================================= -->
7147 <h3>
7149 </h3>
7151 <div>
7154 hardware constructs for atomic operations and memory synchronization. This
7155 provides an interface to the hardware, not an interface to the programmer. It
7156 is aimed at a low enough level to allow any programming models or APIs
7157 (Application Programming Interfaces) which need atomic behaviors to map
7158 cleanly onto it. It is also modeled primarily on hardware behavior. Just as
7159 hardware provides a "universal IR" for source languages, it also provides a
7160 starting point for developing a "universal" atomic operation and
7161 synchronization IR.</p>
7164 software transaction memory systems, atomic primitives, and intrinsic
7165 functions as found in BSD, GNU libc, atomic_ops, APR, and other system and
7166 application libraries. The hardware interface provided by LLVM should allow
7167 a clean implementation of all of these APIs and parallel programming models.
7168 No one model or paradigm should be selected above others unless the hardware
7169 itself ubiquitously does so.</p>
7171 <!-- _______________________________________________________________________ -->
7172 <h4>
7174 </h4>
7176 <div>
7178 <pre>
7179 declare void @llvm.memory.barrier(i1 <ll>, i1 <ls>, i1 <sl>, i1 <ss>, i1 <device>)
7180 </pre>
7184 specific pairs of memory access types.</p>
7188 The first four arguments enables a specific barrier as listed below. The
7189 fifth argument specifies that the barrier applies to io or device or uncached
7190 memory.</p>
7192 <ul>
7198 </ul>
7201 <p>This intrinsic causes the system to enforce some ordering constraints upon
7202 the loads and stores of the program. This barrier does not
7205 and store operations (f.ex. load-load, or store-load), all of the first
7206 operations preceding the barrier will complete before any of the second
7207 operations succeeding the barrier begin. Specifically the semantics for each
7208 pairing is as follows:</p>
7210 <ul>
7212 after the barrier begins.</li>
7214 store after the barrier begins.</li>
7216 store after the barrier begins.</li>
7218 load after the barrier begins.</li>
7219 </ul>
7222 is enabled in a single memory barrier intrinsic.</p>
7224 <p>Backends may implement stronger barriers than those requested when they do
7225 not support as fine grained a barrier as requested. Some architectures do
7226 not need all types of barriers and on such architectures, these become
7227 noops.</p>
7230 <pre>
7231 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7232 %ptr = bitcast i8* %mallocP to i32*
7233 store i32 4, %ptr
7236 call void @llvm.memory.barrier(i1 false, i1 true, i1 false, i1 false)
7239 </pre>
7241 </div>
7243 <!-- _______________________________________________________________________ -->
7244 <h4>
7246 </h4>
7248 <div>
7252 any integer bit width and for different address spaces. Not all targets
7253 support all bit widths however.</p>
7255 <pre>
7257 declare i16 @llvm.atomic.cmp.swap.i16.p0i16(i16* <ptr>, i16 <cmp>, i16 <val>)
7258 declare i32 @llvm.atomic.cmp.swap.i32.p0i32(i32* <ptr>, i32 <cmp>, i32 <val>)
7259 declare i64 @llvm.atomic.cmp.swap.i64.p0i64(i64* <ptr>, i64 <cmp>, i64 <val>)
7260 </pre>
7263 <p>This loads a value in memory and compares it to a given value. If they are
7264 equal, it stores a new value into the memory.</p>
7270 this integer type. While any bit width integer may be used, targets may only
7271 lower representations they support in hardware.</p>
7274 <p>This entire intrinsic must be executed atomically. It first loads the value
7277 memory. The loaded value is yielded in all cases. This provides the
7278 equivalent of an atomic compare-and-swap operation within the SSA
7279 framework.</p>
7282 <pre>
7283 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7284 %ptr = bitcast i8* %mallocP to i32*
7285 store i32 4, %ptr
7288 %result1 = call i32 @llvm.atomic.cmp.swap.i32.p0i32(i32* %ptr, i32 4, %val1)
7294 %result2 = call i32 @llvm.atomic.cmp.swap.i32.p0i32(i32* %ptr, i32 5, %val2)
7299 </pre>
7301 </div>
7303 <!-- _______________________________________________________________________ -->
7304 <h4>
7306 </h4>
7308 <div>
7312 integer bit width. Not all targets support all bit widths however.</p>
7314 <pre>
7319 </pre>
7330 integer type. The targets may only lower integer representations they
7331 support.</p>
7336 equivalent of an atomic swap operation within the SSA framework.</p>
7339 <pre>
7340 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7341 %ptr = bitcast i8* %mallocP to i32*
7342 store i32 4, %ptr
7345 %result1 = call i32 @llvm.atomic.swap.i32.p0i32(i32* %ptr, i32 %val1)
7351 %result2 = call i32 @llvm.atomic.swap.i32.p0i32(i32* %ptr, i32 %val2)
7356 </pre>
7358 </div>
7360 <!-- _______________________________________________________________________ -->
7361 <h4>
7363 </h4>
7365 <div>
7369 any integer bit width. Not all targets support all bit widths however.</p>
7371 <pre>
7376 </pre>
7383 <p>The intrinsic takes two arguments, the first a pointer to an integer value
7384 and the second an integer value. The result is also an integer value. These
7385 integer types can have any bit width, but they must all have the same bit
7386 width. The targets may only lower integer representations they support.</p>
7389 <p>This intrinsic does a series of operations atomically. It first loads the
7394 <pre>
7395 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7396 %ptr = bitcast i8* %mallocP to i32*
7397 store i32 4, %ptr
7398 %result1 = call i32 @llvm.atomic.load.add.i32.p0i32(i32* %ptr, i32 4)
7400 %result2 = call i32 @llvm.atomic.load.add.i32.p0i32(i32* %ptr, i32 2)
7402 %result3 = call i32 @llvm.atomic.load.add.i32.p0i32(i32* %ptr, i32 5)
7405 </pre>
7407 </div>
7409 <!-- _______________________________________________________________________ -->
7410 <h4>
7412 </h4>
7414 <div>
7418 any integer bit width and for different address spaces. Not all targets
7419 support all bit widths however.</p>
7421 <pre>
7426 </pre>
7433 <p>The intrinsic takes two arguments, the first a pointer to an integer value
7434 and the second an integer value. The result is also an integer value. These
7435 integer types can have any bit width, but they must all have the same bit
7436 width. The targets may only lower integer representations they support.</p>
7439 <p>This intrinsic does a series of operations atomically. It first loads the
7445 <pre>
7446 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7447 %ptr = bitcast i8* %mallocP to i32*
7448 store i32 8, %ptr
7449 %result1 = call i32 @llvm.atomic.load.sub.i32.p0i32(i32* %ptr, i32 4)
7451 %result2 = call i32 @llvm.atomic.load.sub.i32.p0i32(i32* %ptr, i32 2)
7453 %result3 = call i32 @llvm.atomic.load.sub.i32.p0i32(i32* %ptr, i32 5)
7456 </pre>
7458 </div>
7460 <!-- _______________________________________________________________________ -->
7461 <h4>
7464 </a>
7465 <br>
7468 </a>
7469 <br>
7472 </a>
7473 <br>
7476 </a>
7477 </h4>
7479 <div>
7482 <p>These are overloaded intrinsics. You can
7485 bit width and for different address spaces. Not all targets support all bit
7486 widths however.</p>
7488 <pre>
7493 </pre>
7495 <pre>
7500 </pre>
7502 <pre>
7507 </pre>
7509 <pre>
7514 </pre>
7522 <p>These intrinsics take two arguments, the first a pointer to an integer value
7523 and the second an integer value. The result is also an integer value. These
7524 integer types can have any bit width, but they must all have the same bit
7525 width. The targets may only lower integer representations they support.</p>
7528 <p>These intrinsics does a series of operations atomically. They first load the
7534 <pre>
7535 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7536 %ptr = bitcast i8* %mallocP to i32*
7537 store i32 0x0F0F, %ptr
7538 %result0 = call i32 @llvm.atomic.load.nand.i32.p0i32(i32* %ptr, i32 0xFF)
7540 %result1 = call i32 @llvm.atomic.load.and.i32.p0i32(i32* %ptr, i32 0xFF)
7542 %result2 = call i32 @llvm.atomic.load.or.i32.p0i32(i32* %ptr, i32 0F)
7544 %result3 = call i32 @llvm.atomic.load.xor.i32.p0i32(i32* %ptr, i32 0F)
7547 </pre>
7549 </div>
7551 <!-- _______________________________________________________________________ -->
7552 <h4>
7555 </a>
7556 <br>
7559 </a>
7560 <br>
7563 </a>
7564 <br>
7567 </a>
7568 </h4>
7570 <div>
7576 address spaces. Not all targets support all bit widths however.</p>
7578 <pre>
7583 </pre>
7585 <pre>
7590 </pre>
7592 <pre>
7597 </pre>
7599 <pre>
7604 </pre>
7607 <p>These intrinsics takes the signed or unsigned minimum or maximum of
7612 <p>These intrinsics take two arguments, the first a pointer to an integer value
7613 and the second an integer value. The result is also an integer value. These
7614 integer types can have any bit width, but they must all have the same bit
7615 width. The targets may only lower integer representations they support.</p>
7618 <p>These intrinsics does a series of operations atomically. They first load the
7624 <pre>
7625 %mallocP = tail call i8* @malloc(i32 ptrtoint (i32* getelementptr (i32* null, i32 1) to i32))
7626 %ptr = bitcast i8* %mallocP to i32*
7627 store i32 7, %ptr
7628 %result0 = call i32 @llvm.atomic.load.min.i32.p0i32(i32* %ptr, i32 -2)
7630 %result1 = call i32 @llvm.atomic.load.max.i32.p0i32(i32* %ptr, i32 8)
7632 %result2 = call i32 @llvm.atomic.load.umin.i32.p0i32(i32* %ptr, i32 10)
7634 %result3 = call i32 @llvm.atomic.load.umax.i32.p0i32(i32* %ptr, i32 30)
7637 </pre>
7639 </div>
7641 </div>
7643 <!-- ======================================================================= -->
7644 <h3>
7646 </h3>
7648 <div>
7650 <p>This class of intrinsics exists to information about the lifetime of memory
7651 objects and ranges where variables are immutable.</p>
7653 <!-- _______________________________________________________________________ -->
7654 <h4>
7656 </h4>
7658 <div>
7661 <pre>
7663 </pre>
7667 object's lifetime.</p>
7670 <p>The first argument is a constant integer representing the size of the
7671 object, or -1 if it is variable sized. The second argument is a pointer to
7672 the object.</p>
7675 <p>This intrinsic indicates that before this point in the code, the value of the
7677 never be used and has an undefined value. A load from the pointer that
7678 precedes this intrinsic can be replaced with
7681 </div>
7683 <!-- _______________________________________________________________________ -->
7684 <h4>
7686 </h4>
7688 <div>
7691 <pre>
7693 </pre>
7697 object's lifetime.</p>
7700 <p>The first argument is a constant integer representing the size of the
7701 object, or -1 if it is variable sized. The second argument is a pointer to
7702 the object.</p>
7705 <p>This intrinsic indicates that after this point in the code, the value of the
7707 never be used and has an undefined value. Any stores into the memory object
7708 following this intrinsic may be removed as dead.
7710 </div>
7712 <!-- _______________________________________________________________________ -->
7713 <h4>
7715 </h4>
7717 <div>
7720 <pre>
7722 </pre>
7726 a memory object will not change.</p>
7729 <p>The first argument is a constant integer representing the size of the
7730 object, or -1 if it is variable sized. The second argument is a pointer to
7731 the object.</p>
7735 the return value, the referenced memory location is constant and
7736 unchanging.</p>
7738 </div>
7740 <!-- _______________________________________________________________________ -->
7741 <h4>
7743 </h4>
7745 <div>
7748 <pre>
7749 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
7750 </pre>
7754 a memory object are mutable.</p>
7758 The second argument is a constant integer representing the size of the
7759 object, or -1 if it is variable sized and the third argument is a pointer
7760 to the object.</p>
7765 </div>
7767 </div>
7769 <!-- ======================================================================= -->
7770 <h3>
7772 </h3>
7774 <div>
7776 <p>This class of intrinsics is designed to be generic and has no specific
7777 purpose.</p>
7779 <!-- _______________________________________________________________________ -->
7780 <h4>
7782 </h4>
7784 <div>
7787 <pre>
7788 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
7789 </pre>
7795 <p>The first argument is a pointer to a value, the second is a pointer to a
7796 global string, the third is a pointer to a global string which is the source
7797 file name, and the last argument is the line number.</p>
7800 <p>This intrinsic allows annotation of local variables with arbitrary strings.
7801 This can be useful for special purpose optimizations that want to look for
7802 these annotations. These have no other defined use, they are ignored by code
7803 generation and optimization.</p>
7805 </div>
7807 <!-- _______________________________________________________________________ -->
7808 <h4>
7810 </h4>
7812 <div>
7816 any integer bit width.</p>
7818 <pre>
7819 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
7820 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
7821 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
7822 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
7823 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
7824 </pre>
7830 <p>The first argument is an integer value (result of some expression), the
7831 second is a pointer to a global string, the third is a pointer to a global
7832 string which is the source file name, and the last argument is the line
7833 number. It returns the value of the first argument.</p>
7836 <p>This intrinsic allows annotations to be put on arbitrary expressions with
7837 arbitrary strings. This can be useful for special purpose optimizations that
7838 want to look for these annotations. These have no other defined use, they
7839 are ignored by code generation and optimization.</p>
7841 </div>
7843 <!-- _______________________________________________________________________ -->
7844 <h4>
7846 </h4>
7848 <div>
7851 <pre>
7852 declare void @llvm.trap()
7853 </pre>
7862 <p>This intrinsics is lowered to the target dependent trap instruction. If the
7863 target does not have a trap instruction, this intrinsic will be lowered to
7866 </div>
7868 <!-- _______________________________________________________________________ -->
7869 <h4>
7871 </h4>
7873 <div>
7876 <pre>
7878 </pre>
7883 ensure that it is placed on the stack before local variables.</p>
7887 arguments. The first argument is the value loaded from the stack
7889 that has enough space to hold the value of the guard.</p>
7892 <p>This intrinsic causes the prologue/epilogue inserter to force the position of
7894 stack. This is to ensure that if a local variable on the stack is
7895 overwritten, it will destroy the value of the guard. When the function exits,
7896 the guard on the stack is checked against the original guard. If they are
7898 function.</p>
7900 </div>
7902 <!-- _______________________________________________________________________ -->
7903 <h4>
7905 </h4>
7907 <div>
7910 <pre>
7913 </pre>
7917 the optimizers to determine at compile time whether a) an operation (like
7918 memcpy) will overflow a buffer that corresponds to an object, or b) that a
7919 runtime check for overflow isn't necessary. An object in this context means
7920 an allocation of a specific class, structure, array, or other object.</p>
7933 compile time.</p>
7935 </div>
7937 </div>
7939 </div>
7941 <!-- *********************************************************************** -->
7942 <hr>
7943 <address>
7951 Last modified: $Date$
7952 </address>
7954 </body>
7955 </html>