1 ==============================
2 LLVM Language Reference Manual
3 ==============================
12 This document is a reference manual for the LLVM assembly language. LLVM
13 is a Static Single Assignment (SSA) based representation that provides
14 type safety, low-level operations, flexibility, and the capability of
15 representing 'all' high-level languages cleanly. It is the common code
16 representation used throughout all phases of the LLVM compilation
22 The LLVM code representation is designed to be used in three different
23 forms: as an in-memory compiler IR, as an on-disk bitcode representation
24 (suitable for fast loading by a Just-In-Time compiler), and as a human
25 readable assembly language representation. This allows LLVM to provide a
26 powerful intermediate representation for efficient compiler
27 transformations and analysis, while providing a natural means to debug
28 and visualize the transformations. The three different forms of LLVM are
29 all equivalent. This document describes the human readable
30 representation and notation.
32 The LLVM representation aims to be light-weight and low-level while
33 being expressive, typed, and extensible at the same time. It aims to be
34 a "universal IR" of sorts, by being at a low enough level that
35 high-level ideas may be cleanly mapped to it (similar to how
36 microprocessors are "universal IR's", allowing many source languages to
37 be mapped to them). By providing type information, LLVM can be used as
38 the target of optimizations: for example, through pointer analysis, it
39 can be proven that a C automatic variable is never accessed outside of
40 the current function, allowing it to be promoted to a simple SSA value
41 instead of a memory location.
48 It is important to note that this document describes 'well formed' LLVM
49 assembly language. There is a difference between what the parser accepts
50 and what is considered 'well formed'. For example, the following
51 instruction is syntactically okay, but not well formed:
57 because the definition of ``%x`` does not dominate all of its uses. The
58 LLVM infrastructure provides a verification pass that may be used to
59 verify that an LLVM module is well formed. This pass is automatically
60 run by the parser after parsing input assembly and by the optimizer
61 before it outputs bitcode. The violations pointed out by the verifier
62 pass indicate bugs in transformation passes or input to the parser.
69 LLVM identifiers come in two basic types: global and local. Global
70 identifiers (functions, global variables) begin with the ``'@'``
71 character. Local identifiers (register names, types) begin with the
72 ``'%'`` character. Additionally, there are three different formats for
73 identifiers, for different purposes:
75 #. Named values are represented as a string of characters with their
76 prefix. For example, ``%foo``, ``@DivisionByZero``,
77 ``%a.really.long.identifier``. The actual regular expression used is
78 '``[%@][-a-zA-Z$._][-a-zA-Z$._0-9]*``'. Identifiers that require other
79 characters in their names can be surrounded with quotes. Special
80 characters may be escaped using ``"\xx"`` where ``xx`` is the ASCII
81 code for the character in hexadecimal. In this way, any character can
82 be used in a name value, even quotes themselves. The ``"\01"`` prefix
83 can be used on global variables to suppress mangling.
84 #. Unnamed values are represented as an unsigned numeric value with
85 their prefix. For example, ``%12``, ``@2``, ``%44``.
86 #. Constants, which are described in the section Constants_ below.
88 LLVM requires that values start with a prefix for two reasons: Compilers
89 don't need to worry about name clashes with reserved words, and the set
90 of reserved words may be expanded in the future without penalty.
91 Additionally, unnamed identifiers allow a compiler to quickly come up
92 with a temporary variable without having to avoid symbol table
95 Reserved words in LLVM are very similar to reserved words in other
96 languages. There are keywords for different opcodes ('``add``',
97 '``bitcast``', '``ret``', etc...), for primitive type names ('``void``',
98 '``i32``', etc...), and others. These reserved words cannot conflict
99 with variable names, because none of them start with a prefix character
100 (``'%'`` or ``'@'``).
102 Here is an example of LLVM code to multiply the integer variable
109 %result = mul i32 %X, 8
111 After strength reduction:
115 %result = shl i32 %X, 3
121 %0 = add i32 %X, %X ; yields i32:%0
122 %1 = add i32 %0, %0 ; yields i32:%1
123 %result = add i32 %1, %1
125 This last way of multiplying ``%X`` by 8 illustrates several important
126 lexical features of LLVM:
128 #. Comments are delimited with a '``;``' and go until the end of line.
129 #. Unnamed temporaries are created when the result of a computation is
130 not assigned to a named value.
131 #. Unnamed temporaries are numbered sequentially (using a per-function
132 incrementing counter, starting with 0). Note that basic blocks and unnamed
133 function parameters are included in this numbering. For example, if the
134 entry basic block is not given a label name and all function parameters are
135 named, then it will get number 0.
137 It also shows a convention that we follow in this document. When
138 demonstrating instructions, we will follow an instruction with a comment
139 that defines the type and name of value produced.
147 LLVM programs are composed of ``Module``'s, each of which is a
148 translation unit of the input programs. Each module consists of
149 functions, global variables, and symbol table entries. Modules may be
150 combined together with the LLVM linker, which merges function (and
151 global variable) definitions, resolves forward declarations, and merges
152 symbol table entries. Here is an example of the "hello world" module:
156 ; Declare the string constant as a global constant.
157 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
159 ; External declaration of the puts function
160 declare i32 @puts(i8* nocapture) nounwind
162 ; Definition of main function
163 define i32 @main() { ; i32()*
164 ; Convert [13 x i8]* to i8 *...
165 %cast210 = getelementptr [13 x i8], [13 x i8]* @.str, i64 0, i64 0
167 ; Call puts function to write out the string to stdout.
168 call i32 @puts(i8* %cast210)
173 !0 = !{i32 42, null, !"string"}
176 This example is made up of a :ref:`global variable <globalvars>` named
177 "``.str``", an external declaration of the "``puts``" function, a
178 :ref:`function definition <functionstructure>` for "``main``" and
179 :ref:`named metadata <namedmetadatastructure>` "``foo``".
181 In general, a module is made up of a list of global values (where both
182 functions and global variables are global values). Global values are
183 represented by a pointer to a memory location (in this case, a pointer
184 to an array of char, and a pointer to a function), and have one of the
185 following :ref:`linkage types <linkage>`.
192 All Global Variables and Functions have one of the following types of
196 Global values with "``private``" linkage are only directly
197 accessible by objects in the current module. In particular, linking
198 code into a module with an private global value may cause the
199 private to be renamed as necessary to avoid collisions. Because the
200 symbol is private to the module, all references can be updated. This
201 doesn't show up in any symbol table in the object file.
203 Similar to private, but the value shows as a local symbol
204 (``STB_LOCAL`` in the case of ELF) in the object file. This
205 corresponds to the notion of the '``static``' keyword in C.
206 ``available_externally``
207 Globals with "``available_externally``" linkage are never emitted
208 into the object file corresponding to the LLVM module. They exist to
209 allow inlining and other optimizations to take place given knowledge
210 of the definition of the global, which is known to be somewhere
211 outside the module. Globals with ``available_externally`` linkage
212 are allowed to be discarded at will, and are otherwise the same as
213 ``linkonce_odr``. This linkage type is only allowed on definitions,
216 Globals with "``linkonce``" linkage are merged with other globals of
217 the same name when linkage occurs. This can be used to implement
218 some forms of inline functions, templates, or other code which must
219 be generated in each translation unit that uses it, but where the
220 body may be overridden with a more definitive definition later.
221 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
222 that ``linkonce`` linkage does not actually allow the optimizer to
223 inline the body of this function into callers because it doesn't
224 know if this definition of the function is the definitive definition
225 within the program or whether it will be overridden by a stronger
226 definition. To enable inlining and other optimizations, use
227 "``linkonce_odr``" linkage.
229 "``weak``" linkage has the same merging semantics as ``linkonce``
230 linkage, except that unreferenced globals with ``weak`` linkage may
231 not be discarded. This is used for globals that are declared "weak"
234 "``common``" linkage is most similar to "``weak``" linkage, but they
235 are used for tentative definitions in C, such as "``int X;``" at
236 global scope. Symbols with "``common``" linkage are merged in the
237 same way as ``weak symbols``, and they may not be deleted if
238 unreferenced. ``common`` symbols may not have an explicit section,
239 must have a zero initializer, and may not be marked
240 ':ref:`constant <globalvars>`'. Functions and aliases may not have
243 .. _linkage_appending:
246 "``appending``" linkage may only be applied to global variables of
247 pointer to array type. When two global variables with appending
248 linkage are linked together, the two global arrays are appended
249 together. This is the LLVM, typesafe, equivalent of having the
250 system linker append together "sections" with identical names when
253 The semantics of this linkage follow the ELF object file model: the
254 symbol is weak until linked, if not linked, the symbol becomes null
255 instead of being an undefined reference.
256 ``linkonce_odr``, ``weak_odr``
257 Some languages allow differing globals to be merged, such as two
258 functions with different semantics. Other languages, such as
259 ``C++``, ensure that only equivalent globals are ever merged (the
260 "one definition rule" --- "ODR"). Such languages can use the
261 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
262 global will only be merged with equivalent globals. These linkage
263 types are otherwise the same as their non-``odr`` versions.
265 If none of the above identifiers are used, the global is externally
266 visible, meaning that it participates in linkage and can be used to
267 resolve external symbol references.
269 It is illegal for a function *declaration* to have any linkage type
270 other than ``external`` or ``extern_weak``.
277 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
278 :ref:`invokes <i_invoke>` can all have an optional calling convention
279 specified for the call. The calling convention of any pair of dynamic
280 caller/callee must match, or the behavior of the program is undefined.
281 The following calling conventions are supported by LLVM, and more may be
284 "``ccc``" - The C calling convention
285 This calling convention (the default if no other calling convention
286 is specified) matches the target C calling conventions. This calling
287 convention supports varargs function calls and tolerates some
288 mismatch in the declared prototype and implemented declaration of
289 the function (as does normal C).
290 "``fastcc``" - The fast calling convention
291 This calling convention attempts to make calls as fast as possible
292 (e.g. by passing things in registers). This calling convention
293 allows the target to use whatever tricks it wants to produce fast
294 code for the target, without having to conform to an externally
295 specified ABI (Application Binary Interface). `Tail calls can only
296 be optimized when this, the GHC or the HiPE convention is
297 used. <CodeGenerator.html#id80>`_ This calling convention does not
298 support varargs and requires the prototype of all callees to exactly
299 match the prototype of the function definition.
300 "``coldcc``" - The cold calling convention
301 This calling convention attempts to make code in the caller as
302 efficient as possible under the assumption that the call is not
303 commonly executed. As such, these calls often preserve all registers
304 so that the call does not break any live ranges in the caller side.
305 This calling convention does not support varargs and requires the
306 prototype of all callees to exactly match the prototype of the
307 function definition. Furthermore the inliner doesn't consider such function
309 "``cc 10``" - GHC convention
310 This calling convention has been implemented specifically for use by
311 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
312 It passes everything in registers, going to extremes to achieve this
313 by disabling callee save registers. This calling convention should
314 not be used lightly but only for specific situations such as an
315 alternative to the *register pinning* performance technique often
316 used when implementing functional programming languages. At the
317 moment only X86 supports this convention and it has the following
320 - On *X86-32* only supports up to 4 bit type parameters. No
321 floating point types are supported.
322 - On *X86-64* only supports up to 10 bit type parameters and 6
323 floating point parameters.
325 This calling convention supports `tail call
326 optimization <CodeGenerator.html#id80>`_ but requires both the
327 caller and callee are using it.
328 "``cc 11``" - The HiPE calling convention
329 This calling convention has been implemented specifically for use by
330 the `High-Performance Erlang
331 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
332 native code compiler of the `Ericsson's Open Source Erlang/OTP
333 system <http://www.erlang.org/download.shtml>`_. It uses more
334 registers for argument passing than the ordinary C calling
335 convention and defines no callee-saved registers. The calling
336 convention properly supports `tail call
337 optimization <CodeGenerator.html#id80>`_ but requires that both the
338 caller and the callee use it. It uses a *register pinning*
339 mechanism, similar to GHC's convention, for keeping frequently
340 accessed runtime components pinned to specific hardware registers.
341 At the moment only X86 supports this convention (both 32 and 64
343 "``webkit_jscc``" - WebKit's JavaScript calling convention
344 This calling convention has been implemented for `WebKit FTL JIT
345 <https://trac.webkit.org/wiki/FTLJIT>`_. It passes arguments on the
346 stack right to left (as cdecl does), and returns a value in the
347 platform's customary return register.
348 "``anyregcc``" - Dynamic calling convention for code patching
349 This is a special convention that supports patching an arbitrary code
350 sequence in place of a call site. This convention forces the call
351 arguments into registers but allows them to be dynamically
352 allocated. This can currently only be used with calls to
353 llvm.experimental.patchpoint because only this intrinsic records
354 the location of its arguments in a side table. See :doc:`StackMaps`.
355 "``preserve_mostcc``" - The `PreserveMost` calling convention
356 This calling convention attempts to make the code in the caller as
357 unintrusive as possible. This convention behaves identically to the `C`
358 calling convention on how arguments and return values are passed, but it
359 uses a different set of caller/callee-saved registers. This alleviates the
360 burden of saving and recovering a large register set before and after the
361 call in the caller. If the arguments are passed in callee-saved registers,
362 then they will be preserved by the callee across the call. This doesn't
363 apply for values returned in callee-saved registers.
365 - On X86-64 the callee preserves all general purpose registers, except for
366 R11. R11 can be used as a scratch register. Floating-point registers
367 (XMMs/YMMs) are not preserved and need to be saved by the caller.
369 The idea behind this convention is to support calls to runtime functions
370 that have a hot path and a cold path. The hot path is usually a small piece
371 of code that doesn't use many registers. The cold path might need to call out to
372 another function and therefore only needs to preserve the caller-saved
373 registers, which haven't already been saved by the caller. The
374 `PreserveMost` calling convention is very similar to the `cold` calling
375 convention in terms of caller/callee-saved registers, but they are used for
376 different types of function calls. `coldcc` is for function calls that are
377 rarely executed, whereas `preserve_mostcc` function calls are intended to be
378 on the hot path and definitely executed a lot. Furthermore `preserve_mostcc`
379 doesn't prevent the inliner from inlining the function call.
381 This calling convention will be used by a future version of the ObjectiveC
382 runtime and should therefore still be considered experimental at this time.
383 Although this convention was created to optimize certain runtime calls to
384 the ObjectiveC runtime, it is not limited to this runtime and might be used
385 by other runtimes in the future too. The current implementation only
386 supports X86-64, but the intention is to support more architectures in the
388 "``preserve_allcc``" - The `PreserveAll` calling convention
389 This calling convention attempts to make the code in the caller even less
390 intrusive than the `PreserveMost` calling convention. This calling
391 convention also behaves identical to the `C` calling convention on how
392 arguments and return values are passed, but it uses a different set of
393 caller/callee-saved registers. This removes the burden of saving and
394 recovering a large register set before and after the call in the caller. If
395 the arguments are passed in callee-saved registers, then they will be
396 preserved by the callee across the call. This doesn't apply for values
397 returned in callee-saved registers.
399 - On X86-64 the callee preserves all general purpose registers, except for
400 R11. R11 can be used as a scratch register. Furthermore it also preserves
401 all floating-point registers (XMMs/YMMs).
403 The idea behind this convention is to support calls to runtime functions
404 that don't need to call out to any other functions.
406 This calling convention, like the `PreserveMost` calling convention, will be
407 used by a future version of the ObjectiveC runtime and should be considered
408 experimental at this time.
409 "``cxx_fast_tlscc``" - The `CXX_FAST_TLS` calling convention for access functions
410 This calling convention aims to minimize overhead in the caller by
411 preserving as many registers as possible. This calling convention behaves
412 identical to the `C` calling convention on how arguments and return values
413 are passed, but it uses a different set of caller/callee-saved registers.
414 Given that C-style TLS on Darwin has its own special CSRs, we can't use the
415 existing `PreserveMost`.
417 - On X86-64 the callee preserves all general purpose registers, except for
419 "``cc <n>``" - Numbered convention
420 Any calling convention may be specified by number, allowing
421 target-specific calling conventions to be used. Target specific
422 calling conventions start at 64.
424 More calling conventions can be added/defined on an as-needed basis, to
425 support Pascal conventions or any other well-known target-independent
428 .. _visibilitystyles:
433 All Global Variables and Functions have one of the following visibility
436 "``default``" - Default style
437 On targets that use the ELF object file format, default visibility
438 means that the declaration is visible to other modules and, in
439 shared libraries, means that the declared entity may be overridden.
440 On Darwin, default visibility means that the declaration is visible
441 to other modules. Default visibility corresponds to "external
442 linkage" in the language.
443 "``hidden``" - Hidden style
444 Two declarations of an object with hidden visibility refer to the
445 same object if they are in the same shared object. Usually, hidden
446 visibility indicates that the symbol will not be placed into the
447 dynamic symbol table, so no other module (executable or shared
448 library) can reference it directly.
449 "``protected``" - Protected style
450 On ELF, protected visibility indicates that the symbol will be
451 placed in the dynamic symbol table, but that references within the
452 defining module will bind to the local symbol. That is, the symbol
453 cannot be overridden by another module.
455 A symbol with ``internal`` or ``private`` linkage must have ``default``
463 All Global Variables, Functions and Aliases can have one of the following
467 "``dllimport``" causes the compiler to reference a function or variable via
468 a global pointer to a pointer that is set up by the DLL exporting the
469 symbol. On Microsoft Windows targets, the pointer name is formed by
470 combining ``__imp_`` and the function or variable name.
472 "``dllexport``" causes the compiler to provide a global pointer to a pointer
473 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
474 Microsoft Windows targets, the pointer name is formed by combining
475 ``__imp_`` and the function or variable name. Since this storage class
476 exists for defining a dll interface, the compiler, assembler and linker know
477 it is externally referenced and must refrain from deleting the symbol.
481 Thread Local Storage Models
482 ---------------------------
484 A variable may be defined as ``thread_local``, which means that it will
485 not be shared by threads (each thread will have a separated copy of the
486 variable). Not all targets support thread-local variables. Optionally, a
487 TLS model may be specified:
490 For variables that are only used within the current shared library.
492 For variables in modules that will not be loaded dynamically.
494 For variables defined in the executable and only used within it.
496 If no explicit model is given, the "general dynamic" model is used.
498 The models correspond to the ELF TLS models; see `ELF Handling For
499 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
500 more information on under which circumstances the different models may
501 be used. The target may choose a different TLS model if the specified
502 model is not supported, or if a better choice of model can be made.
504 A model can also be specified in an alias, but then it only governs how
505 the alias is accessed. It will not have any effect in the aliasee.
507 For platforms without linker support of ELF TLS model, the -femulated-tls
508 flag can be used to generate GCC compatible emulated TLS code.
515 LLVM IR allows you to specify both "identified" and "literal" :ref:`structure
516 types <t_struct>`. Literal types are uniqued structurally, but identified types
517 are never uniqued. An :ref:`opaque structural type <t_opaque>` can also be used
518 to forward declare a type that is not yet available.
520 An example of an identified structure specification is:
524 %mytype = type { %mytype*, i32 }
526 Prior to the LLVM 3.0 release, identified types were structurally uniqued. Only
527 literal types are uniqued in recent versions of LLVM.
534 Global variables define regions of memory allocated at compilation time
537 Global variable definitions must be initialized.
539 Global variables in other translation units can also be declared, in which
540 case they don't have an initializer.
542 Either global variable definitions or declarations may have an explicit section
543 to be placed in and may have an optional explicit alignment specified.
545 A variable may be defined as a global ``constant``, which indicates that
546 the contents of the variable will **never** be modified (enabling better
547 optimization, allowing the global data to be placed in the read-only
548 section of an executable, etc). Note that variables that need runtime
549 initialization cannot be marked ``constant`` as there is a store to the
552 LLVM explicitly allows *declarations* of global variables to be marked
553 constant, even if the final definition of the global is not. This
554 capability can be used to enable slightly better optimization of the
555 program, but requires the language definition to guarantee that
556 optimizations based on the 'constantness' are valid for the translation
557 units that do not include the definition.
559 As SSA values, global variables define pointer values that are in scope
560 (i.e. they dominate) all basic blocks in the program. Global variables
561 always define a pointer to their "content" type because they describe a
562 region of memory, and all memory objects in LLVM are accessed through
565 Global variables can be marked with ``unnamed_addr`` which indicates
566 that the address is not significant, only the content. Constants marked
567 like this can be merged with other constants if they have the same
568 initializer. Note that a constant with significant address *can* be
569 merged with a ``unnamed_addr`` constant, the result being a constant
570 whose address is significant.
572 A global variable may be declared to reside in a target-specific
573 numbered address space. For targets that support them, address spaces
574 may affect how optimizations are performed and/or what target
575 instructions are used to access the variable. The default address space
576 is zero. The address space qualifier must precede any other attributes.
578 LLVM allows an explicit section to be specified for globals. If the
579 target supports it, it will emit globals to the section specified.
580 Additionally, the global can placed in a comdat if the target has the necessary
583 By default, global initializers are optimized by assuming that global
584 variables defined within the module are not modified from their
585 initial values before the start of the global initializer. This is
586 true even for variables potentially accessible from outside the
587 module, including those with external linkage or appearing in
588 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
589 by marking the variable with ``externally_initialized``.
591 An explicit alignment may be specified for a global, which must be a
592 power of 2. If not present, or if the alignment is set to zero, the
593 alignment of the global is set by the target to whatever it feels
594 convenient. If an explicit alignment is specified, the global is forced
595 to have exactly that alignment. Targets and optimizers are not allowed
596 to over-align the global if the global has an assigned section. In this
597 case, the extra alignment could be observable: for example, code could
598 assume that the globals are densely packed in their section and try to
599 iterate over them as an array, alignment padding would break this
600 iteration. The maximum alignment is ``1 << 29``.
602 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
604 Variables and aliases can have a
605 :ref:`Thread Local Storage Model <tls_model>`.
609 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
610 [unnamed_addr] [AddrSpace] [ExternallyInitialized]
611 <global | constant> <Type> [<InitializerConstant>]
612 [, section "name"] [, comdat [($name)]]
613 [, align <Alignment>]
615 For example, the following defines a global in a numbered address space
616 with an initializer, section, and alignment:
620 @G = addrspace(5) constant float 1.0, section "foo", align 4
622 The following example just declares a global variable
626 @G = external global i32
628 The following example defines a thread-local global with the
629 ``initialexec`` TLS model:
633 @G = thread_local(initialexec) global i32 0, align 4
635 .. _functionstructure:
640 LLVM function definitions consist of the "``define``" keyword, an
641 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
642 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
643 an optional :ref:`calling convention <callingconv>`,
644 an optional ``unnamed_addr`` attribute, a return type, an optional
645 :ref:`parameter attribute <paramattrs>` for the return type, a function
646 name, a (possibly empty) argument list (each with optional :ref:`parameter
647 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
648 an optional section, an optional alignment,
649 an optional :ref:`comdat <langref_comdats>`,
650 an optional :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
651 an optional :ref:`prologue <prologuedata>`,
652 an optional :ref:`personality <personalityfn>`,
653 an optional list of attached :ref:`metadata <metadata>`,
654 an opening curly brace, a list of basic blocks, and a closing curly brace.
656 LLVM function declarations consist of the "``declare``" keyword, an
657 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
658 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
659 an optional :ref:`calling convention <callingconv>`,
660 an optional ``unnamed_addr`` attribute, a return type, an optional
661 :ref:`parameter attribute <paramattrs>` for the return type, a function
662 name, a possibly empty list of arguments, an optional alignment, an optional
663 :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
664 and an optional :ref:`prologue <prologuedata>`.
666 A function definition contains a list of basic blocks, forming the CFG (Control
667 Flow Graph) for the function. Each basic block may optionally start with a label
668 (giving the basic block a symbol table entry), contains a list of instructions,
669 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
670 function return). If an explicit label is not provided, a block is assigned an
671 implicit numbered label, using the next value from the same counter as used for
672 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
673 entry block does not have an explicit label, it will be assigned label "%0",
674 then the first unnamed temporary in that block will be "%1", etc.
676 The first basic block in a function is special in two ways: it is
677 immediately executed on entrance to the function, and it is not allowed
678 to have predecessor basic blocks (i.e. there can not be any branches to
679 the entry block of a function). Because the block can have no
680 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
682 LLVM allows an explicit section to be specified for functions. If the
683 target supports it, it will emit functions to the section specified.
684 Additionally, the function can be placed in a COMDAT.
686 An explicit alignment may be specified for a function. If not present,
687 or if the alignment is set to zero, the alignment of the function is set
688 by the target to whatever it feels convenient. If an explicit alignment
689 is specified, the function is forced to have at least that much
690 alignment. All alignments must be a power of 2.
692 If the ``unnamed_addr`` attribute is given, the address is known to not
693 be significant and two identical functions can be merged.
697 define [linkage] [visibility] [DLLStorageClass]
699 <ResultType> @<FunctionName> ([argument list])
700 [unnamed_addr] [fn Attrs] [section "name"] [comdat [($name)]]
701 [align N] [gc] [prefix Constant] [prologue Constant]
702 [personality Constant] (!name !N)* { ... }
704 The argument list is a comma separated sequence of arguments where each
705 argument is of the following form:
709 <type> [parameter Attrs] [name]
717 Aliases, unlike function or variables, don't create any new data. They
718 are just a new symbol and metadata for an existing position.
720 Aliases have a name and an aliasee that is either a global value or a
723 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
724 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
725 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
729 @<Name> = [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal] [unnamed_addr] alias <AliaseeTy>, <AliaseeTy>* @<Aliasee>
731 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
732 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
733 might not correctly handle dropping a weak symbol that is aliased.
735 Aliases that are not ``unnamed_addr`` are guaranteed to have the same address as
736 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
739 Since aliases are only a second name, some restrictions apply, of which
740 some can only be checked when producing an object file:
742 * The expression defining the aliasee must be computable at assembly
743 time. Since it is just a name, no relocations can be used.
745 * No alias in the expression can be weak as the possibility of the
746 intermediate alias being overridden cannot be represented in an
749 * No global value in the expression can be a declaration, since that
750 would require a relocation, which is not possible.
757 Comdat IR provides access to COFF and ELF object file COMDAT functionality.
759 Comdats have a name which represents the COMDAT key. All global objects that
760 specify this key will only end up in the final object file if the linker chooses
761 that key over some other key. Aliases are placed in the same COMDAT that their
762 aliasee computes to, if any.
764 Comdats have a selection kind to provide input on how the linker should
765 choose between keys in two different object files.
769 $<Name> = comdat SelectionKind
771 The selection kind must be one of the following:
774 The linker may choose any COMDAT key, the choice is arbitrary.
776 The linker may choose any COMDAT key but the sections must contain the
779 The linker will choose the section containing the largest COMDAT key.
781 The linker requires that only section with this COMDAT key exist.
783 The linker may choose any COMDAT key but the sections must contain the
786 Note that the Mach-O platform doesn't support COMDATs and ELF only supports
787 ``any`` as a selection kind.
789 Here is an example of a COMDAT group where a function will only be selected if
790 the COMDAT key's section is the largest:
794 $foo = comdat largest
795 @foo = global i32 2, comdat($foo)
797 define void @bar() comdat($foo) {
801 As a syntactic sugar the ``$name`` can be omitted if the name is the same as
807 @foo = global i32 2, comdat
810 In a COFF object file, this will create a COMDAT section with selection kind
811 ``IMAGE_COMDAT_SELECT_LARGEST`` containing the contents of the ``@foo`` symbol
812 and another COMDAT section with selection kind
813 ``IMAGE_COMDAT_SELECT_ASSOCIATIVE`` which is associated with the first COMDAT
814 section and contains the contents of the ``@bar`` symbol.
816 There are some restrictions on the properties of the global object.
817 It, or an alias to it, must have the same name as the COMDAT group when
819 The contents and size of this object may be used during link-time to determine
820 which COMDAT groups get selected depending on the selection kind.
821 Because the name of the object must match the name of the COMDAT group, the
822 linkage of the global object must not be local; local symbols can get renamed
823 if a collision occurs in the symbol table.
825 The combined use of COMDATS and section attributes may yield surprising results.
832 @g1 = global i32 42, section "sec", comdat($foo)
833 @g2 = global i32 42, section "sec", comdat($bar)
835 From the object file perspective, this requires the creation of two sections
836 with the same name. This is necessary because both globals belong to different
837 COMDAT groups and COMDATs, at the object file level, are represented by
840 Note that certain IR constructs like global variables and functions may
841 create COMDATs in the object file in addition to any which are specified using
842 COMDAT IR. This arises when the code generator is configured to emit globals
843 in individual sections (e.g. when `-data-sections` or `-function-sections`
844 is supplied to `llc`).
846 .. _namedmetadatastructure:
851 Named metadata is a collection of metadata. :ref:`Metadata
852 nodes <metadata>` (but not metadata strings) are the only valid
853 operands for a named metadata.
855 #. Named metadata are represented as a string of characters with the
856 metadata prefix. The rules for metadata names are the same as for
857 identifiers, but quoted names are not allowed. ``"\xx"`` type escapes
858 are still valid, which allows any character to be part of a name.
862 ; Some unnamed metadata nodes, which are referenced by the named metadata.
867 !name = !{!0, !1, !2}
874 The return type and each parameter of a function type may have a set of
875 *parameter attributes* associated with them. Parameter attributes are
876 used to communicate additional information about the result or
877 parameters of a function. Parameter attributes are considered to be part
878 of the function, not of the function type, so functions with different
879 parameter attributes can have the same function type.
881 Parameter attributes are simple keywords that follow the type specified.
882 If multiple parameter attributes are needed, they are space separated.
887 declare i32 @printf(i8* noalias nocapture, ...)
888 declare i32 @atoi(i8 zeroext)
889 declare signext i8 @returns_signed_char()
891 Note that any attributes for the function result (``nounwind``,
892 ``readonly``) come immediately after the argument list.
894 Currently, only the following parameter attributes are defined:
897 This indicates to the code generator that the parameter or return
898 value should be zero-extended to the extent required by the target's
899 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
900 the caller (for a parameter) or the callee (for a return value).
902 This indicates to the code generator that the parameter or return
903 value should be sign-extended to the extent required by the target's
904 ABI (which is usually 32-bits) by the caller (for a parameter) or
905 the callee (for a return value).
907 This indicates that this parameter or return value should be treated
908 in a special target-dependent fashion while emitting code for
909 a function call or return (usually, by putting it in a register as
910 opposed to memory, though some targets use it to distinguish between
911 two different kinds of registers). Use of this attribute is
914 This indicates that the pointer parameter should really be passed by
915 value to the function. The attribute implies that a hidden copy of
916 the pointee is made between the caller and the callee, so the callee
917 is unable to modify the value in the caller. This attribute is only
918 valid on LLVM pointer arguments. It is generally used to pass
919 structs and arrays by value, but is also valid on pointers to
920 scalars. The copy is considered to belong to the caller not the
921 callee (for example, ``readonly`` functions should not write to
922 ``byval`` parameters). This is not a valid attribute for return
925 The byval attribute also supports specifying an alignment with the
926 align attribute. It indicates the alignment of the stack slot to
927 form and the known alignment of the pointer specified to the call
928 site. If the alignment is not specified, then the code generator
929 makes a target-specific assumption.
935 The ``inalloca`` argument attribute allows the caller to take the
936 address of outgoing stack arguments. An ``inalloca`` argument must
937 be a pointer to stack memory produced by an ``alloca`` instruction.
938 The alloca, or argument allocation, must also be tagged with the
939 inalloca keyword. Only the last argument may have the ``inalloca``
940 attribute, and that argument is guaranteed to be passed in memory.
942 An argument allocation may be used by a call at most once because
943 the call may deallocate it. The ``inalloca`` attribute cannot be
944 used in conjunction with other attributes that affect argument
945 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
946 ``inalloca`` attribute also disables LLVM's implicit lowering of
947 large aggregate return values, which means that frontend authors
948 must lower them with ``sret`` pointers.
950 When the call site is reached, the argument allocation must have
951 been the most recent stack allocation that is still live, or the
952 results are undefined. It is possible to allocate additional stack
953 space after an argument allocation and before its call site, but it
954 must be cleared off with :ref:`llvm.stackrestore
957 See :doc:`InAlloca` for more information on how to use this
961 This indicates that the pointer parameter specifies the address of a
962 structure that is the return value of the function in the source
963 program. This pointer must be guaranteed by the caller to be valid:
964 loads and stores to the structure may be assumed by the callee
965 not to trap and to be properly aligned. This may only be applied to
966 the first parameter. This is not a valid attribute for return
970 This indicates that the pointer value may be assumed by the optimizer to
971 have the specified alignment.
973 Note that this attribute has additional semantics when combined with the
979 This indicates that objects accessed via pointer values
980 :ref:`based <pointeraliasing>` on the argument or return value are not also
981 accessed, during the execution of the function, via pointer values not
982 *based* on the argument or return value. The attribute on a return value
983 also has additional semantics described below. The caller shares the
984 responsibility with the callee for ensuring that these requirements are met.
985 For further details, please see the discussion of the NoAlias response in
986 :ref:`alias analysis <Must, May, or No>`.
988 Note that this definition of ``noalias`` is intentionally similar
989 to the definition of ``restrict`` in C99 for function arguments.
991 For function return values, C99's ``restrict`` is not meaningful,
992 while LLVM's ``noalias`` is. Furthermore, the semantics of the ``noalias``
993 attribute on return values are stronger than the semantics of the attribute
994 when used on function arguments. On function return values, the ``noalias``
995 attribute indicates that the function acts like a system memory allocation
996 function, returning a pointer to allocated storage disjoint from the
997 storage for any other object accessible to the caller.
1000 This indicates that the callee does not make any copies of the
1001 pointer that outlive the callee itself. This is not a valid
1002 attribute for return values.
1007 This indicates that the pointer parameter can be excised using the
1008 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
1009 attribute for return values and can only be applied to one parameter.
1012 This indicates that the function always returns the argument as its return
1013 value. This is an optimization hint to the code generator when generating
1014 the caller, allowing tail call optimization and omission of register saves
1015 and restores in some cases; it is not checked or enforced when generating
1016 the callee. The parameter and the function return type must be valid
1017 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
1018 valid attribute for return values and can only be applied to one parameter.
1021 This indicates that the parameter or return pointer is not null. This
1022 attribute may only be applied to pointer typed parameters. This is not
1023 checked or enforced by LLVM, the caller must ensure that the pointer
1024 passed in is non-null, or the callee must ensure that the returned pointer
1027 ``dereferenceable(<n>)``
1028 This indicates that the parameter or return pointer is dereferenceable. This
1029 attribute may only be applied to pointer typed parameters. A pointer that
1030 is dereferenceable can be loaded from speculatively without a risk of
1031 trapping. The number of bytes known to be dereferenceable must be provided
1032 in parentheses. It is legal for the number of bytes to be less than the
1033 size of the pointee type. The ``nonnull`` attribute does not imply
1034 dereferenceability (consider a pointer to one element past the end of an
1035 array), however ``dereferenceable(<n>)`` does imply ``nonnull`` in
1036 ``addrspace(0)`` (which is the default address space).
1038 ``dereferenceable_or_null(<n>)``
1039 This indicates that the parameter or return value isn't both
1040 non-null and non-dereferenceable (up to ``<n>`` bytes) at the same
1041 time. All non-null pointers tagged with
1042 ``dereferenceable_or_null(<n>)`` are ``dereferenceable(<n>)``.
1043 For address space 0 ``dereferenceable_or_null(<n>)`` implies that
1044 a pointer is exactly one of ``dereferenceable(<n>)`` or ``null``,
1045 and in other address spaces ``dereferenceable_or_null(<n>)``
1046 implies that a pointer is at least one of ``dereferenceable(<n>)``
1047 or ``null`` (i.e. it may be both ``null`` and
1048 ``dereferenceable(<n>)``). This attribute may only be applied to
1049 pointer typed parameters.
1053 Garbage Collector Strategy Names
1054 --------------------------------
1056 Each function may specify a garbage collector strategy name, which is simply a
1059 .. code-block:: llvm
1061 define void @f() gc "name" { ... }
1063 The supported values of *name* includes those :ref:`built in to LLVM
1064 <builtin-gc-strategies>` and any provided by loaded plugins. Specifying a GC
1065 strategy will cause the compiler to alter its output in order to support the
1066 named garbage collection algorithm. Note that LLVM itself does not contain a
1067 garbage collector, this functionality is restricted to generating machine code
1068 which can interoperate with a collector provided externally.
1075 Prefix data is data associated with a function which the code
1076 generator will emit immediately before the function's entrypoint.
1077 The purpose of this feature is to allow frontends to associate
1078 language-specific runtime metadata with specific functions and make it
1079 available through the function pointer while still allowing the
1080 function pointer to be called.
1082 To access the data for a given function, a program may bitcast the
1083 function pointer to a pointer to the constant's type and dereference
1084 index -1. This implies that the IR symbol points just past the end of
1085 the prefix data. For instance, take the example of a function annotated
1086 with a single ``i32``,
1088 .. code-block:: llvm
1090 define void @f() prefix i32 123 { ... }
1092 The prefix data can be referenced as,
1094 .. code-block:: llvm
1096 %0 = bitcast void* () @f to i32*
1097 %a = getelementptr inbounds i32, i32* %0, i32 -1
1098 %b = load i32, i32* %a
1100 Prefix data is laid out as if it were an initializer for a global variable
1101 of the prefix data's type. The function will be placed such that the
1102 beginning of the prefix data is aligned. This means that if the size
1103 of the prefix data is not a multiple of the alignment size, the
1104 function's entrypoint will not be aligned. If alignment of the
1105 function's entrypoint is desired, padding must be added to the prefix
1108 A function may have prefix data but no body. This has similar semantics
1109 to the ``available_externally`` linkage in that the data may be used by the
1110 optimizers but will not be emitted in the object file.
1117 The ``prologue`` attribute allows arbitrary code (encoded as bytes) to
1118 be inserted prior to the function body. This can be used for enabling
1119 function hot-patching and instrumentation.
1121 To maintain the semantics of ordinary function calls, the prologue data must
1122 have a particular format. Specifically, it must begin with a sequence of
1123 bytes which decode to a sequence of machine instructions, valid for the
1124 module's target, which transfer control to the point immediately succeeding
1125 the prologue data, without performing any other visible action. This allows
1126 the inliner and other passes to reason about the semantics of the function
1127 definition without needing to reason about the prologue data. Obviously this
1128 makes the format of the prologue data highly target dependent.
1130 A trivial example of valid prologue data for the x86 architecture is ``i8 144``,
1131 which encodes the ``nop`` instruction:
1133 .. code-block:: llvm
1135 define void @f() prologue i8 144 { ... }
1137 Generally prologue data can be formed by encoding a relative branch instruction
1138 which skips the metadata, as in this example of valid prologue data for the
1139 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
1141 .. code-block:: llvm
1143 %0 = type <{ i8, i8, i8* }>
1145 define void @f() prologue %0 <{ i8 235, i8 8, i8* @md}> { ... }
1147 A function may have prologue data but no body. This has similar semantics
1148 to the ``available_externally`` linkage in that the data may be used by the
1149 optimizers but will not be emitted in the object file.
1153 Personality Function
1154 --------------------
1156 The ``personality`` attribute permits functions to specify what function
1157 to use for exception handling.
1164 Attribute groups are groups of attributes that are referenced by objects within
1165 the IR. They are important for keeping ``.ll`` files readable, because a lot of
1166 functions will use the same set of attributes. In the degenerative case of a
1167 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
1168 group will capture the important command line flags used to build that file.
1170 An attribute group is a module-level object. To use an attribute group, an
1171 object references the attribute group's ID (e.g. ``#37``). An object may refer
1172 to more than one attribute group. In that situation, the attributes from the
1173 different groups are merged.
1175 Here is an example of attribute groups for a function that should always be
1176 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
1178 .. code-block:: llvm
1180 ; Target-independent attributes:
1181 attributes #0 = { alwaysinline alignstack=4 }
1183 ; Target-dependent attributes:
1184 attributes #1 = { "no-sse" }
1186 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
1187 define void @f() #0 #1 { ... }
1194 Function attributes are set to communicate additional information about
1195 a function. Function attributes are considered to be part of the
1196 function, not of the function type, so functions with different function
1197 attributes can have the same function type.
1199 Function attributes are simple keywords that follow the type specified.
1200 If multiple attributes are needed, they are space separated. For
1203 .. code-block:: llvm
1205 define void @f() noinline { ... }
1206 define void @f() alwaysinline { ... }
1207 define void @f() alwaysinline optsize { ... }
1208 define void @f() optsize { ... }
1211 This attribute indicates that, when emitting the prologue and
1212 epilogue, the backend should forcibly align the stack pointer.
1213 Specify the desired alignment, which must be a power of two, in
1216 This attribute indicates that the inliner should attempt to inline
1217 this function into callers whenever possible, ignoring any active
1218 inlining size threshold for this caller.
1220 This indicates that the callee function at a call site should be
1221 recognized as a built-in function, even though the function's declaration
1222 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1223 direct calls to functions that are declared with the ``nobuiltin``
1226 This attribute indicates that this function is rarely called. When
1227 computing edge weights, basic blocks post-dominated by a cold
1228 function call are also considered to be cold; and, thus, given low
1231 This attribute indicates that the callee is dependent on a convergent
1232 thread execution pattern under certain parallel execution models.
1233 Transformations that are execution model agnostic may not make the execution
1234 of a convergent operation control dependent on any additional values.
1236 This attribute indicates that the source code contained a hint that
1237 inlining this function is desirable (such as the "inline" keyword in
1238 C/C++). It is just a hint; it imposes no requirements on the
1241 This attribute indicates that the function should be added to a
1242 jump-instruction table at code-generation time, and that all address-taken
1243 references to this function should be replaced with a reference to the
1244 appropriate jump-instruction-table function pointer. Note that this creates
1245 a new pointer for the original function, which means that code that depends
1246 on function-pointer identity can break. So, any function annotated with
1247 ``jumptable`` must also be ``unnamed_addr``.
1249 This attribute suggests that optimization passes and code generator
1250 passes make choices that keep the code size of this function as small
1251 as possible and perform optimizations that may sacrifice runtime
1252 performance in order to minimize the size of the generated code.
1254 This attribute disables prologue / epilogue emission for the
1255 function. This can have very system-specific consequences.
1257 This indicates that the callee function at a call site is not recognized as
1258 a built-in function. LLVM will retain the original call and not replace it
1259 with equivalent code based on the semantics of the built-in function, unless
1260 the call site uses the ``builtin`` attribute. This is valid at call sites
1261 and on function declarations and definitions.
1263 This attribute indicates that calls to the function cannot be
1264 duplicated. A call to a ``noduplicate`` function may be moved
1265 within its parent function, but may not be duplicated within
1266 its parent function.
1268 A function containing a ``noduplicate`` call may still
1269 be an inlining candidate, provided that the call is not
1270 duplicated by inlining. That implies that the function has
1271 internal linkage and only has one call site, so the original
1272 call is dead after inlining.
1274 This attributes disables implicit floating point instructions.
1276 This attribute indicates that the inliner should never inline this
1277 function in any situation. This attribute may not be used together
1278 with the ``alwaysinline`` attribute.
1280 This attribute suppresses lazy symbol binding for the function. This
1281 may make calls to the function faster, at the cost of extra program
1282 startup time if the function is not called during program startup.
1284 This attribute indicates that the code generator should not use a
1285 red zone, even if the target-specific ABI normally permits it.
1287 This function attribute indicates that the function never returns
1288 normally. This produces undefined behavior at runtime if the
1289 function ever does dynamically return.
1291 This function attribute indicates that the function does not call itself
1292 either directly or indirectly down any possible call path. This produces
1293 undefined behavior at runtime if the function ever does recurse.
1295 This function attribute indicates that the function never raises an
1296 exception. If the function does raise an exception, its runtime
1297 behavior is undefined. However, functions marked nounwind may still
1298 trap or generate asynchronous exceptions. Exception handling schemes
1299 that are recognized by LLVM to handle asynchronous exceptions, such
1300 as SEH, will still provide their implementation defined semantics.
1302 This function attribute indicates that most optimization passes will skip
1303 this function, with the exception of interprocedural optimization passes.
1304 Code generation defaults to the "fast" instruction selector.
1305 This attribute cannot be used together with the ``alwaysinline``
1306 attribute; this attribute is also incompatible
1307 with the ``minsize`` attribute and the ``optsize`` attribute.
1309 This attribute requires the ``noinline`` attribute to be specified on
1310 the function as well, so the function is never inlined into any caller.
1311 Only functions with the ``alwaysinline`` attribute are valid
1312 candidates for inlining into the body of this function.
1314 This attribute suggests that optimization passes and code generator
1315 passes make choices that keep the code size of this function low,
1316 and otherwise do optimizations specifically to reduce code size as
1317 long as they do not significantly impact runtime performance.
1319 On a function, this attribute indicates that the function computes its
1320 result (or decides to unwind an exception) based strictly on its arguments,
1321 without dereferencing any pointer arguments or otherwise accessing
1322 any mutable state (e.g. memory, control registers, etc) visible to
1323 caller functions. It does not write through any pointer arguments
1324 (including ``byval`` arguments) and never changes any state visible
1325 to callers. This means that it cannot unwind exceptions by calling
1326 the ``C++`` exception throwing methods.
1328 On an argument, this attribute indicates that the function does not
1329 dereference that pointer argument, even though it may read or write the
1330 memory that the pointer points to if accessed through other pointers.
1332 On a function, this attribute indicates that the function does not write
1333 through any pointer arguments (including ``byval`` arguments) or otherwise
1334 modify any state (e.g. memory, control registers, etc) visible to
1335 caller functions. It may dereference pointer arguments and read
1336 state that may be set in the caller. A readonly function always
1337 returns the same value (or unwinds an exception identically) when
1338 called with the same set of arguments and global state. It cannot
1339 unwind an exception by calling the ``C++`` exception throwing
1342 On an argument, this attribute indicates that the function does not write
1343 through this pointer argument, even though it may write to the memory that
1344 the pointer points to.
1346 This attribute indicates that the only memory accesses inside function are
1347 loads and stores from objects pointed to by its pointer-typed arguments,
1348 with arbitrary offsets. Or in other words, all memory operations in the
1349 function can refer to memory only using pointers based on its function
1351 Note that ``argmemonly`` can be used together with ``readonly`` attribute
1352 in order to specify that function reads only from its arguments.
1354 This attribute indicates that this function can return twice. The C
1355 ``setjmp`` is an example of such a function. The compiler disables
1356 some optimizations (like tail calls) in the caller of these
1359 This attribute indicates that
1360 `SafeStack <http://clang.llvm.org/docs/SafeStack.html>`_
1361 protection is enabled for this function.
1363 If a function that has a ``safestack`` attribute is inlined into a
1364 function that doesn't have a ``safestack`` attribute or which has an
1365 ``ssp``, ``sspstrong`` or ``sspreq`` attribute, then the resulting
1366 function will have a ``safestack`` attribute.
1367 ``sanitize_address``
1368 This attribute indicates that AddressSanitizer checks
1369 (dynamic address safety analysis) are enabled for this function.
1371 This attribute indicates that MemorySanitizer checks (dynamic detection
1372 of accesses to uninitialized memory) are enabled for this function.
1374 This attribute indicates that ThreadSanitizer checks
1375 (dynamic thread safety analysis) are enabled for this function.
1377 This attribute indicates that the function should emit a stack
1378 smashing protector. It is in the form of a "canary" --- a random value
1379 placed on the stack before the local variables that's checked upon
1380 return from the function to see if it has been overwritten. A
1381 heuristic is used to determine if a function needs stack protectors
1382 or not. The heuristic used will enable protectors for functions with:
1384 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1385 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1386 - Calls to alloca() with variable sizes or constant sizes greater than
1387 ``ssp-buffer-size``.
1389 Variables that are identified as requiring a protector will be arranged
1390 on the stack such that they are adjacent to the stack protector guard.
1392 If a function that has an ``ssp`` attribute is inlined into a
1393 function that doesn't have an ``ssp`` attribute, then the resulting
1394 function will have an ``ssp`` attribute.
1396 This attribute indicates that the function should *always* emit a
1397 stack smashing protector. This overrides the ``ssp`` function
1400 Variables that are identified as requiring a protector will be arranged
1401 on the stack such that they are adjacent to the stack protector guard.
1402 The specific layout rules are:
1404 #. Large arrays and structures containing large arrays
1405 (``>= ssp-buffer-size``) are closest to the stack protector.
1406 #. Small arrays and structures containing small arrays
1407 (``< ssp-buffer-size``) are 2nd closest to the protector.
1408 #. Variables that have had their address taken are 3rd closest to the
1411 If a function that has an ``sspreq`` attribute is inlined into a
1412 function that doesn't have an ``sspreq`` attribute or which has an
1413 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1414 an ``sspreq`` attribute.
1416 This attribute indicates that the function should emit a stack smashing
1417 protector. This attribute causes a strong heuristic to be used when
1418 determining if a function needs stack protectors. The strong heuristic
1419 will enable protectors for functions with:
1421 - Arrays of any size and type
1422 - Aggregates containing an array of any size and type.
1423 - Calls to alloca().
1424 - Local variables that have had their address taken.
1426 Variables that are identified as requiring a protector will be arranged
1427 on the stack such that they are adjacent to the stack protector guard.
1428 The specific layout rules are:
1430 #. Large arrays and structures containing large arrays
1431 (``>= ssp-buffer-size``) are closest to the stack protector.
1432 #. Small arrays and structures containing small arrays
1433 (``< ssp-buffer-size``) are 2nd closest to the protector.
1434 #. Variables that have had their address taken are 3rd closest to the
1437 This overrides the ``ssp`` function attribute.
1439 If a function that has an ``sspstrong`` attribute is inlined into a
1440 function that doesn't have an ``sspstrong`` attribute, then the
1441 resulting function will have an ``sspstrong`` attribute.
1443 This attribute indicates that the function will delegate to some other
1444 function with a tail call. The prototype of a thunk should not be used for
1445 optimization purposes. The caller is expected to cast the thunk prototype to
1446 match the thunk target prototype.
1448 This attribute indicates that the ABI being targeted requires that
1449 an unwind table entry be produced for this function even if we can
1450 show that no exceptions passes by it. This is normally the case for
1451 the ELF x86-64 abi, but it can be disabled for some compilation
1460 Note: operand bundles are a work in progress, and they should be
1461 considered experimental at this time.
1463 Operand bundles are tagged sets of SSA values that can be associated
1464 with certain LLVM instructions (currently only ``call`` s and
1465 ``invoke`` s). In a way they are like metadata, but dropping them is
1466 incorrect and will change program semantics.
1470 operand bundle set ::= '[' operand bundle (, operand bundle )* ']'
1471 operand bundle ::= tag '(' [ bundle operand ] (, bundle operand )* ')'
1472 bundle operand ::= SSA value
1473 tag ::= string constant
1475 Operand bundles are **not** part of a function's signature, and a
1476 given function may be called from multiple places with different kinds
1477 of operand bundles. This reflects the fact that the operand bundles
1478 are conceptually a part of the ``call`` (or ``invoke``), not the
1479 callee being dispatched to.
1481 Operand bundles are a generic mechanism intended to support
1482 runtime-introspection-like functionality for managed languages. While
1483 the exact semantics of an operand bundle depend on the bundle tag,
1484 there are certain limitations to how much the presence of an operand
1485 bundle can influence the semantics of a program. These restrictions
1486 are described as the semantics of an "unknown" operand bundle. As
1487 long as the behavior of an operand bundle is describable within these
1488 restrictions, LLVM does not need to have special knowledge of the
1489 operand bundle to not miscompile programs containing it.
1491 - The bundle operands for an unknown operand bundle escape in unknown
1492 ways before control is transferred to the callee or invokee.
1493 - Calls and invokes with operand bundles have unknown read / write
1494 effect on the heap on entry and exit (even if the call target is
1495 ``readnone`` or ``readonly``), unless they're overriden with
1496 callsite specific attributes.
1497 - An operand bundle at a call site cannot change the implementation
1498 of the called function. Inter-procedural optimizations work as
1499 usual as long as they take into account the first two properties.
1501 More specific types of operand bundles are described below.
1503 Deoptimization Operand Bundles
1504 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1506 Deoptimization operand bundles are characterized by the ``"deopt"``
1507 operand bundle tag. These operand bundles represent an alternate
1508 "safe" continuation for the call site they're attached to, and can be
1509 used by a suitable runtime to deoptimize the compiled frame at the
1510 specified call site. There can be at most one ``"deopt"`` operand
1511 bundle attached to a call site. Exact details of deoptimization is
1512 out of scope for the language reference, but it usually involves
1513 rewriting a compiled frame into a set of interpreted frames.
1515 From the compiler's perspective, deoptimization operand bundles make
1516 the call sites they're attached to at least ``readonly``. They read
1517 through all of their pointer typed operands (even if they're not
1518 otherwise escaped) and the entire visible heap. Deoptimization
1519 operand bundles do not capture their operands except during
1520 deoptimization, in which case control will not be returned to the
1523 The inliner knows how to inline through calls that have deoptimization
1524 operand bundles. Just like inlining through a normal call site
1525 involves composing the normal and exceptional continuations, inlining
1526 through a call site with a deoptimization operand bundle needs to
1527 appropriately compose the "safe" deoptimization continuation. The
1528 inliner does this by prepending the parent's deoptimization
1529 continuation to every deoptimization continuation in the inlined body.
1530 E.g. inlining ``@f`` into ``@g`` in the following example
1532 .. code-block:: llvm
1535 call void @x() ;; no deopt state
1536 call void @y() [ "deopt"(i32 10) ]
1537 call void @y() [ "deopt"(i32 10), "unknown"(i8* null) ]
1542 call void @f() [ "deopt"(i32 20) ]
1548 .. code-block:: llvm
1551 call void @x() ;; still no deopt state
1552 call void @y() [ "deopt"(i32 20, i32 10) ]
1553 call void @y() [ "deopt"(i32 20, i32 10), "unknown"(i8* null) ]
1557 It is the frontend's responsibility to structure or encode the
1558 deoptimization state in a way that syntactically prepending the
1559 caller's deoptimization state to the callee's deoptimization state is
1560 semantically equivalent to composing the caller's deoptimization
1561 continuation after the callee's deoptimization continuation.
1565 Module-Level Inline Assembly
1566 ----------------------------
1568 Modules may contain "module-level inline asm" blocks, which corresponds
1569 to the GCC "file scope inline asm" blocks. These blocks are internally
1570 concatenated by LLVM and treated as a single unit, but may be separated
1571 in the ``.ll`` file if desired. The syntax is very simple:
1573 .. code-block:: llvm
1575 module asm "inline asm code goes here"
1576 module asm "more can go here"
1578 The strings can contain any character by escaping non-printable
1579 characters. The escape sequence used is simply "\\xx" where "xx" is the
1580 two digit hex code for the number.
1582 Note that the assembly string *must* be parseable by LLVM's integrated assembler
1583 (unless it is disabled), even when emitting a ``.s`` file.
1585 .. _langref_datalayout:
1590 A module may specify a target specific data layout string that specifies
1591 how data is to be laid out in memory. The syntax for the data layout is
1594 .. code-block:: llvm
1596 target datalayout = "layout specification"
1598 The *layout specification* consists of a list of specifications
1599 separated by the minus sign character ('-'). Each specification starts
1600 with a letter and may include other information after the letter to
1601 define some aspect of the data layout. The specifications accepted are
1605 Specifies that the target lays out data in big-endian form. That is,
1606 the bits with the most significance have the lowest address
1609 Specifies that the target lays out data in little-endian form. That
1610 is, the bits with the least significance have the lowest address
1613 Specifies the natural alignment of the stack in bits. Alignment
1614 promotion of stack variables is limited to the natural stack
1615 alignment to avoid dynamic stack realignment. The stack alignment
1616 must be a multiple of 8-bits. If omitted, the natural stack
1617 alignment defaults to "unspecified", which does not prevent any
1618 alignment promotions.
1619 ``p[n]:<size>:<abi>:<pref>``
1620 This specifies the *size* of a pointer and its ``<abi>`` and
1621 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1622 bits. The address space, ``n``, is optional, and if not specified,
1623 denotes the default address space 0. The value of ``n`` must be
1624 in the range [1,2^23).
1625 ``i<size>:<abi>:<pref>``
1626 This specifies the alignment for an integer type of a given bit
1627 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1628 ``v<size>:<abi>:<pref>``
1629 This specifies the alignment for a vector type of a given bit
1631 ``f<size>:<abi>:<pref>``
1632 This specifies the alignment for a floating point type of a given bit
1633 ``<size>``. Only values of ``<size>`` that are supported by the target
1634 will work. 32 (float) and 64 (double) are supported on all targets; 80
1635 or 128 (different flavors of long double) are also supported on some
1638 This specifies the alignment for an object of aggregate type.
1640 If present, specifies that llvm names are mangled in the output. The
1643 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1644 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1645 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1646 symbols get a ``_`` prefix.
1647 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1648 functions also get a suffix based on the frame size.
1649 * ``x``: Windows x86 COFF prefix: Similar to Windows COFF, but use a ``_``
1650 prefix for ``__cdecl`` functions.
1651 ``n<size1>:<size2>:<size3>...``
1652 This specifies a set of native integer widths for the target CPU in
1653 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1654 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1655 this set are considered to support most general arithmetic operations
1658 On every specification that takes a ``<abi>:<pref>``, specifying the
1659 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1660 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1662 When constructing the data layout for a given target, LLVM starts with a
1663 default set of specifications which are then (possibly) overridden by
1664 the specifications in the ``datalayout`` keyword. The default
1665 specifications are given in this list:
1667 - ``E`` - big endian
1668 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1669 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1670 same as the default address space.
1671 - ``S0`` - natural stack alignment is unspecified
1672 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1673 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1674 - ``i16:16:16`` - i16 is 16-bit aligned
1675 - ``i32:32:32`` - i32 is 32-bit aligned
1676 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1677 alignment of 64-bits
1678 - ``f16:16:16`` - half is 16-bit aligned
1679 - ``f32:32:32`` - float is 32-bit aligned
1680 - ``f64:64:64`` - double is 64-bit aligned
1681 - ``f128:128:128`` - quad is 128-bit aligned
1682 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1683 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1684 - ``a:0:64`` - aggregates are 64-bit aligned
1686 When LLVM is determining the alignment for a given type, it uses the
1689 #. If the type sought is an exact match for one of the specifications,
1690 that specification is used.
1691 #. If no match is found, and the type sought is an integer type, then
1692 the smallest integer type that is larger than the bitwidth of the
1693 sought type is used. If none of the specifications are larger than
1694 the bitwidth then the largest integer type is used. For example,
1695 given the default specifications above, the i7 type will use the
1696 alignment of i8 (next largest) while both i65 and i256 will use the
1697 alignment of i64 (largest specified).
1698 #. If no match is found, and the type sought is a vector type, then the
1699 largest vector type that is smaller than the sought vector type will
1700 be used as a fall back. This happens because <128 x double> can be
1701 implemented in terms of 64 <2 x double>, for example.
1703 The function of the data layout string may not be what you expect.
1704 Notably, this is not a specification from the frontend of what alignment
1705 the code generator should use.
1707 Instead, if specified, the target data layout is required to match what
1708 the ultimate *code generator* expects. This string is used by the
1709 mid-level optimizers to improve code, and this only works if it matches
1710 what the ultimate code generator uses. There is no way to generate IR
1711 that does not embed this target-specific detail into the IR. If you
1712 don't specify the string, the default specifications will be used to
1713 generate a Data Layout and the optimization phases will operate
1714 accordingly and introduce target specificity into the IR with respect to
1715 these default specifications.
1722 A module may specify a target triple string that describes the target
1723 host. The syntax for the target triple is simply:
1725 .. code-block:: llvm
1727 target triple = "x86_64-apple-macosx10.7.0"
1729 The *target triple* string consists of a series of identifiers delimited
1730 by the minus sign character ('-'). The canonical forms are:
1734 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1735 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1737 This information is passed along to the backend so that it generates
1738 code for the proper architecture. It's possible to override this on the
1739 command line with the ``-mtriple`` command line option.
1741 .. _pointeraliasing:
1743 Pointer Aliasing Rules
1744 ----------------------
1746 Any memory access must be done through a pointer value associated with
1747 an address range of the memory access, otherwise the behavior is
1748 undefined. Pointer values are associated with address ranges according
1749 to the following rules:
1751 - A pointer value is associated with the addresses associated with any
1752 value it is *based* on.
1753 - An address of a global variable is associated with the address range
1754 of the variable's storage.
1755 - The result value of an allocation instruction is associated with the
1756 address range of the allocated storage.
1757 - A null pointer in the default address-space is associated with no
1759 - An integer constant other than zero or a pointer value returned from
1760 a function not defined within LLVM may be associated with address
1761 ranges allocated through mechanisms other than those provided by
1762 LLVM. Such ranges shall not overlap with any ranges of addresses
1763 allocated by mechanisms provided by LLVM.
1765 A pointer value is *based* on another pointer value according to the
1768 - A pointer value formed from a ``getelementptr`` operation is *based*
1769 on the first value operand of the ``getelementptr``.
1770 - The result value of a ``bitcast`` is *based* on the operand of the
1772 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1773 values that contribute (directly or indirectly) to the computation of
1774 the pointer's value.
1775 - The "*based* on" relationship is transitive.
1777 Note that this definition of *"based"* is intentionally similar to the
1778 definition of *"based"* in C99, though it is slightly weaker.
1780 LLVM IR does not associate types with memory. The result type of a
1781 ``load`` merely indicates the size and alignment of the memory from
1782 which to load, as well as the interpretation of the value. The first
1783 operand type of a ``store`` similarly only indicates the size and
1784 alignment of the store.
1786 Consequently, type-based alias analysis, aka TBAA, aka
1787 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1788 :ref:`Metadata <metadata>` may be used to encode additional information
1789 which specialized optimization passes may use to implement type-based
1794 Volatile Memory Accesses
1795 ------------------------
1797 Certain memory accesses, such as :ref:`load <i_load>`'s,
1798 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1799 marked ``volatile``. The optimizers must not change the number of
1800 volatile operations or change their order of execution relative to other
1801 volatile operations. The optimizers *may* change the order of volatile
1802 operations relative to non-volatile operations. This is not Java's
1803 "volatile" and has no cross-thread synchronization behavior.
1805 IR-level volatile loads and stores cannot safely be optimized into
1806 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1807 flagged volatile. Likewise, the backend should never split or merge
1808 target-legal volatile load/store instructions.
1810 .. admonition:: Rationale
1812 Platforms may rely on volatile loads and stores of natively supported
1813 data width to be executed as single instruction. For example, in C
1814 this holds for an l-value of volatile primitive type with native
1815 hardware support, but not necessarily for aggregate types. The
1816 frontend upholds these expectations, which are intentionally
1817 unspecified in the IR. The rules above ensure that IR transformations
1818 do not violate the frontend's contract with the language.
1822 Memory Model for Concurrent Operations
1823 --------------------------------------
1825 The LLVM IR does not define any way to start parallel threads of
1826 execution or to register signal handlers. Nonetheless, there are
1827 platform-specific ways to create them, and we define LLVM IR's behavior
1828 in their presence. This model is inspired by the C++0x memory model.
1830 For a more informal introduction to this model, see the :doc:`Atomics`.
1832 We define a *happens-before* partial order as the least partial order
1835 - Is a superset of single-thread program order, and
1836 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1837 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1838 techniques, like pthread locks, thread creation, thread joining,
1839 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1840 Constraints <ordering>`).
1842 Note that program order does not introduce *happens-before* edges
1843 between a thread and signals executing inside that thread.
1845 Every (defined) read operation (load instructions, memcpy, atomic
1846 loads/read-modify-writes, etc.) R reads a series of bytes written by
1847 (defined) write operations (store instructions, atomic
1848 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1849 section, initialized globals are considered to have a write of the
1850 initializer which is atomic and happens before any other read or write
1851 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1852 may see any write to the same byte, except:
1854 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1855 write\ :sub:`2` happens before R\ :sub:`byte`, then
1856 R\ :sub:`byte` does not see write\ :sub:`1`.
1857 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1858 R\ :sub:`byte` does not see write\ :sub:`3`.
1860 Given that definition, R\ :sub:`byte` is defined as follows:
1862 - If R is volatile, the result is target-dependent. (Volatile is
1863 supposed to give guarantees which can support ``sig_atomic_t`` in
1864 C/C++, and may be used for accesses to addresses that do not behave
1865 like normal memory. It does not generally provide cross-thread
1867 - Otherwise, if there is no write to the same byte that happens before
1868 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1869 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1870 R\ :sub:`byte` returns the value written by that write.
1871 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1872 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1873 Memory Ordering Constraints <ordering>` section for additional
1874 constraints on how the choice is made.
1875 - Otherwise R\ :sub:`byte` returns ``undef``.
1877 R returns the value composed of the series of bytes it read. This
1878 implies that some bytes within the value may be ``undef`` **without**
1879 the entire value being ``undef``. Note that this only defines the
1880 semantics of the operation; it doesn't mean that targets will emit more
1881 than one instruction to read the series of bytes.
1883 Note that in cases where none of the atomic intrinsics are used, this
1884 model places only one restriction on IR transformations on top of what
1885 is required for single-threaded execution: introducing a store to a byte
1886 which might not otherwise be stored is not allowed in general.
1887 (Specifically, in the case where another thread might write to and read
1888 from an address, introducing a store can change a load that may see
1889 exactly one write into a load that may see multiple writes.)
1893 Atomic Memory Ordering Constraints
1894 ----------------------------------
1896 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1897 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1898 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1899 ordering parameters that determine which other atomic instructions on
1900 the same address they *synchronize with*. These semantics are borrowed
1901 from Java and C++0x, but are somewhat more colloquial. If these
1902 descriptions aren't precise enough, check those specs (see spec
1903 references in the :doc:`atomics guide <Atomics>`).
1904 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1905 differently since they don't take an address. See that instruction's
1906 documentation for details.
1908 For a simpler introduction to the ordering constraints, see the
1912 The set of values that can be read is governed by the happens-before
1913 partial order. A value cannot be read unless some operation wrote
1914 it. This is intended to provide a guarantee strong enough to model
1915 Java's non-volatile shared variables. This ordering cannot be
1916 specified for read-modify-write operations; it is not strong enough
1917 to make them atomic in any interesting way.
1919 In addition to the guarantees of ``unordered``, there is a single
1920 total order for modifications by ``monotonic`` operations on each
1921 address. All modification orders must be compatible with the
1922 happens-before order. There is no guarantee that the modification
1923 orders can be combined to a global total order for the whole program
1924 (and this often will not be possible). The read in an atomic
1925 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1926 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1927 order immediately before the value it writes. If one atomic read
1928 happens before another atomic read of the same address, the later
1929 read must see the same value or a later value in the address's
1930 modification order. This disallows reordering of ``monotonic`` (or
1931 stronger) operations on the same address. If an address is written
1932 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1933 read that address repeatedly, the other threads must eventually see
1934 the write. This corresponds to the C++0x/C1x
1935 ``memory_order_relaxed``.
1937 In addition to the guarantees of ``monotonic``, a
1938 *synchronizes-with* edge may be formed with a ``release`` operation.
1939 This is intended to model C++'s ``memory_order_acquire``.
1941 In addition to the guarantees of ``monotonic``, if this operation
1942 writes a value which is subsequently read by an ``acquire``
1943 operation, it *synchronizes-with* that operation. (This isn't a
1944 complete description; see the C++0x definition of a release
1945 sequence.) This corresponds to the C++0x/C1x
1946 ``memory_order_release``.
1947 ``acq_rel`` (acquire+release)
1948 Acts as both an ``acquire`` and ``release`` operation on its
1949 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1950 ``seq_cst`` (sequentially consistent)
1951 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1952 operation that only reads, ``release`` for an operation that only
1953 writes), there is a global total order on all
1954 sequentially-consistent operations on all addresses, which is
1955 consistent with the *happens-before* partial order and with the
1956 modification orders of all the affected addresses. Each
1957 sequentially-consistent read sees the last preceding write to the
1958 same address in this global order. This corresponds to the C++0x/C1x
1959 ``memory_order_seq_cst`` and Java volatile.
1963 If an atomic operation is marked ``singlethread``, it only *synchronizes
1964 with* or participates in modification and seq\_cst total orderings with
1965 other operations running in the same thread (for example, in signal
1973 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1974 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1975 :ref:`frem <i_frem>`, :ref:`fcmp <i_fcmp>`) have the following flags that can
1976 be set to enable otherwise unsafe floating point operations
1979 No NaNs - Allow optimizations to assume the arguments and result are not
1980 NaN. Such optimizations are required to retain defined behavior over
1981 NaNs, but the value of the result is undefined.
1984 No Infs - Allow optimizations to assume the arguments and result are not
1985 +/-Inf. Such optimizations are required to retain defined behavior over
1986 +/-Inf, but the value of the result is undefined.
1989 No Signed Zeros - Allow optimizations to treat the sign of a zero
1990 argument or result as insignificant.
1993 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1994 argument rather than perform division.
1997 Fast - Allow algebraically equivalent transformations that may
1998 dramatically change results in floating point (e.g. reassociate). This
1999 flag implies all the others.
2003 Use-list Order Directives
2004 -------------------------
2006 Use-list directives encode the in-memory order of each use-list, allowing the
2007 order to be recreated. ``<order-indexes>`` is a comma-separated list of
2008 indexes that are assigned to the referenced value's uses. The referenced
2009 value's use-list is immediately sorted by these indexes.
2011 Use-list directives may appear at function scope or global scope. They are not
2012 instructions, and have no effect on the semantics of the IR. When they're at
2013 function scope, they must appear after the terminator of the final basic block.
2015 If basic blocks have their address taken via ``blockaddress()`` expressions,
2016 ``uselistorder_bb`` can be used to reorder their use-lists from outside their
2023 uselistorder <ty> <value>, { <order-indexes> }
2024 uselistorder_bb @function, %block { <order-indexes> }
2030 define void @foo(i32 %arg1, i32 %arg2) {
2032 ; ... instructions ...
2034 ; ... instructions ...
2036 ; At function scope.
2037 uselistorder i32 %arg1, { 1, 0, 2 }
2038 uselistorder label %bb, { 1, 0 }
2042 uselistorder i32* @global, { 1, 2, 0 }
2043 uselistorder i32 7, { 1, 0 }
2044 uselistorder i32 (i32) @bar, { 1, 0 }
2045 uselistorder_bb @foo, %bb, { 5, 1, 3, 2, 0, 4 }
2052 The LLVM type system is one of the most important features of the
2053 intermediate representation. Being typed enables a number of
2054 optimizations to be performed on the intermediate representation
2055 directly, without having to do extra analyses on the side before the
2056 transformation. A strong type system makes it easier to read the
2057 generated code and enables novel analyses and transformations that are
2058 not feasible to perform on normal three address code representations.
2068 The void type does not represent any value and has no size.
2086 The function type can be thought of as a function signature. It consists of a
2087 return type and a list of formal parameter types. The return type of a function
2088 type is a void type or first class type --- except for :ref:`label <t_label>`
2089 and :ref:`metadata <t_metadata>` types.
2095 <returntype> (<parameter list>)
2097 ...where '``<parameter list>``' is a comma-separated list of type
2098 specifiers. Optionally, the parameter list may include a type ``...``, which
2099 indicates that the function takes a variable number of arguments. Variable
2100 argument functions can access their arguments with the :ref:`variable argument
2101 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
2102 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
2106 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2107 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
2108 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2109 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
2110 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2111 | ``i32 (i8*, ...)`` | A vararg function that takes at least one :ref:`pointer <t_pointer>` to ``i8`` (char in C), which returns an integer. This is the signature for ``printf`` in LLVM. |
2112 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2113 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
2114 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2121 The :ref:`first class <t_firstclass>` types are perhaps the most important.
2122 Values of these types are the only ones which can be produced by
2130 These are the types that are valid in registers from CodeGen's perspective.
2139 The integer type is a very simple type that simply specifies an
2140 arbitrary bit width for the integer type desired. Any bit width from 1
2141 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
2149 The number of bits the integer will occupy is specified by the ``N``
2155 +----------------+------------------------------------------------+
2156 | ``i1`` | a single-bit integer. |
2157 +----------------+------------------------------------------------+
2158 | ``i32`` | a 32-bit integer. |
2159 +----------------+------------------------------------------------+
2160 | ``i1942652`` | a really big integer of over 1 million bits. |
2161 +----------------+------------------------------------------------+
2165 Floating Point Types
2166 """"""""""""""""""""
2175 - 16-bit floating point value
2178 - 32-bit floating point value
2181 - 64-bit floating point value
2184 - 128-bit floating point value (112-bit mantissa)
2187 - 80-bit floating point value (X87)
2190 - 128-bit floating point value (two 64-bits)
2197 The x86_mmx type represents a value held in an MMX register on an x86
2198 machine. The operations allowed on it are quite limited: parameters and
2199 return values, load and store, and bitcast. User-specified MMX
2200 instructions are represented as intrinsic or asm calls with arguments
2201 and/or results of this type. There are no arrays, vectors or constants
2218 The pointer type is used to specify memory locations. Pointers are
2219 commonly used to reference objects in memory.
2221 Pointer types may have an optional address space attribute defining the
2222 numbered address space where the pointed-to object resides. The default
2223 address space is number zero. The semantics of non-zero address spaces
2224 are target-specific.
2226 Note that LLVM does not permit pointers to void (``void*``) nor does it
2227 permit pointers to labels (``label*``). Use ``i8*`` instead.
2237 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2238 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
2239 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2240 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
2241 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2242 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
2243 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2252 A vector type is a simple derived type that represents a vector of
2253 elements. Vector types are used when multiple primitive data are
2254 operated in parallel using a single instruction (SIMD). A vector type
2255 requires a size (number of elements) and an underlying primitive data
2256 type. Vector types are considered :ref:`first class <t_firstclass>`.
2262 < <# elements> x <elementtype> >
2264 The number of elements is a constant integer value larger than 0;
2265 elementtype may be any integer, floating point or pointer type. Vectors
2266 of size zero are not allowed.
2270 +-------------------+--------------------------------------------------+
2271 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
2272 +-------------------+--------------------------------------------------+
2273 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
2274 +-------------------+--------------------------------------------------+
2275 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
2276 +-------------------+--------------------------------------------------+
2277 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
2278 +-------------------+--------------------------------------------------+
2287 The label type represents code labels.
2302 The token type is used when a value is associated with an instruction
2303 but all uses of the value must not attempt to introspect or obscure it.
2304 As such, it is not appropriate to have a :ref:`phi <i_phi>` or
2305 :ref:`select <i_select>` of type token.
2322 The metadata type represents embedded metadata. No derived types may be
2323 created from metadata except for :ref:`function <t_function>` arguments.
2336 Aggregate Types are a subset of derived types that can contain multiple
2337 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2338 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2348 The array type is a very simple derived type that arranges elements
2349 sequentially in memory. The array type requires a size (number of
2350 elements) and an underlying data type.
2356 [<# elements> x <elementtype>]
2358 The number of elements is a constant integer value; ``elementtype`` may
2359 be any type with a size.
2363 +------------------+--------------------------------------+
2364 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
2365 +------------------+--------------------------------------+
2366 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
2367 +------------------+--------------------------------------+
2368 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
2369 +------------------+--------------------------------------+
2371 Here are some examples of multidimensional arrays:
2373 +-----------------------------+----------------------------------------------------------+
2374 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
2375 +-----------------------------+----------------------------------------------------------+
2376 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
2377 +-----------------------------+----------------------------------------------------------+
2378 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
2379 +-----------------------------+----------------------------------------------------------+
2381 There is no restriction on indexing beyond the end of the array implied
2382 by a static type (though there are restrictions on indexing beyond the
2383 bounds of an allocated object in some cases). This means that
2384 single-dimension 'variable sized array' addressing can be implemented in
2385 LLVM with a zero length array type. An implementation of 'pascal style
2386 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2396 The structure type is used to represent a collection of data members
2397 together in memory. The elements of a structure may be any type that has
2400 Structures in memory are accessed using '``load``' and '``store``' by
2401 getting a pointer to a field with the '``getelementptr``' instruction.
2402 Structures in registers are accessed using the '``extractvalue``' and
2403 '``insertvalue``' instructions.
2405 Structures may optionally be "packed" structures, which indicate that
2406 the alignment of the struct is one byte, and that there is no padding
2407 between the elements. In non-packed structs, padding between field types
2408 is inserted as defined by the DataLayout string in the module, which is
2409 required to match what the underlying code generator expects.
2411 Structures can either be "literal" or "identified". A literal structure
2412 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2413 identified types are always defined at the top level with a name.
2414 Literal types are uniqued by their contents and can never be recursive
2415 or opaque since there is no way to write one. Identified types can be
2416 recursive, can be opaqued, and are never uniqued.
2422 %T1 = type { <type list> } ; Identified normal struct type
2423 %T2 = type <{ <type list> }> ; Identified packed struct type
2427 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2428 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2429 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2430 | ``{ float, i32 (i32) * }`` | A pair, where the first element is a ``float`` and the second element is a :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32``, returning an ``i32``. |
2431 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2432 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2433 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2437 Opaque Structure Types
2438 """"""""""""""""""""""
2442 Opaque structure types are used to represent named structure types that
2443 do not have a body specified. This corresponds (for example) to the C
2444 notion of a forward declared structure.
2455 +--------------+-------------------+
2456 | ``opaque`` | An opaque type. |
2457 +--------------+-------------------+
2464 LLVM has several different basic types of constants. This section
2465 describes them all and their syntax.
2470 **Boolean constants**
2471 The two strings '``true``' and '``false``' are both valid constants
2473 **Integer constants**
2474 Standard integers (such as '4') are constants of the
2475 :ref:`integer <t_integer>` type. Negative numbers may be used with
2477 **Floating point constants**
2478 Floating point constants use standard decimal notation (e.g.
2479 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2480 hexadecimal notation (see below). The assembler requires the exact
2481 decimal value of a floating-point constant. For example, the
2482 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2483 decimal in binary. Floating point constants must have a :ref:`floating
2484 point <t_floating>` type.
2485 **Null pointer constants**
2486 The identifier '``null``' is recognized as a null pointer constant
2487 and must be of :ref:`pointer type <t_pointer>`.
2489 The identifier '``none``' is recognized as an empty token constant
2490 and must be of :ref:`token type <t_token>`.
2492 The one non-intuitive notation for constants is the hexadecimal form of
2493 floating point constants. For example, the form
2494 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2495 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2496 constants are required (and the only time that they are generated by the
2497 disassembler) is when a floating point constant must be emitted but it
2498 cannot be represented as a decimal floating point number in a reasonable
2499 number of digits. For example, NaN's, infinities, and other special
2500 values are represented in their IEEE hexadecimal format so that assembly
2501 and disassembly do not cause any bits to change in the constants.
2503 When using the hexadecimal form, constants of types half, float, and
2504 double are represented using the 16-digit form shown above (which
2505 matches the IEEE754 representation for double); half and float values
2506 must, however, be exactly representable as IEEE 754 half and single
2507 precision, respectively. Hexadecimal format is always used for long
2508 double, and there are three forms of long double. The 80-bit format used
2509 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2510 128-bit format used by PowerPC (two adjacent doubles) is represented by
2511 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2512 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2513 will only work if they match the long double format on your target.
2514 The IEEE 16-bit format (half precision) is represented by ``0xH``
2515 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2516 (sign bit at the left).
2518 There are no constants of type x86_mmx.
2520 .. _complexconstants:
2525 Complex constants are a (potentially recursive) combination of simple
2526 constants and smaller complex constants.
2528 **Structure constants**
2529 Structure constants are represented with notation similar to
2530 structure type definitions (a comma separated list of elements,
2531 surrounded by braces (``{}``)). For example:
2532 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2533 "``@G = external global i32``". Structure constants must have
2534 :ref:`structure type <t_struct>`, and the number and types of elements
2535 must match those specified by the type.
2537 Array constants are represented with notation similar to array type
2538 definitions (a comma separated list of elements, surrounded by
2539 square brackets (``[]``)). For example:
2540 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2541 :ref:`array type <t_array>`, and the number and types of elements must
2542 match those specified by the type. As a special case, character array
2543 constants may also be represented as a double-quoted string using the ``c``
2544 prefix. For example: "``c"Hello World\0A\00"``".
2545 **Vector constants**
2546 Vector constants are represented with notation similar to vector
2547 type definitions (a comma separated list of elements, surrounded by
2548 less-than/greater-than's (``<>``)). For example:
2549 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2550 must have :ref:`vector type <t_vector>`, and the number and types of
2551 elements must match those specified by the type.
2552 **Zero initialization**
2553 The string '``zeroinitializer``' can be used to zero initialize a
2554 value to zero of *any* type, including scalar and
2555 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2556 having to print large zero initializers (e.g. for large arrays) and
2557 is always exactly equivalent to using explicit zero initializers.
2559 A metadata node is a constant tuple without types. For example:
2560 "``!{!0, !{!2, !0}, !"test"}``". Metadata can reference constant values,
2561 for example: "``!{!0, i32 0, i8* @global, i64 (i64)* @function, !"str"}``".
2562 Unlike other typed constants that are meant to be interpreted as part of
2563 the instruction stream, metadata is a place to attach additional
2564 information such as debug info.
2566 Global Variable and Function Addresses
2567 --------------------------------------
2569 The addresses of :ref:`global variables <globalvars>` and
2570 :ref:`functions <functionstructure>` are always implicitly valid
2571 (link-time) constants. These constants are explicitly referenced when
2572 the :ref:`identifier for the global <identifiers>` is used and always have
2573 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2576 .. code-block:: llvm
2580 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2587 The string '``undef``' can be used anywhere a constant is expected, and
2588 indicates that the user of the value may receive an unspecified
2589 bit-pattern. Undefined values may be of any type (other than '``label``'
2590 or '``void``') and be used anywhere a constant is permitted.
2592 Undefined values are useful because they indicate to the compiler that
2593 the program is well defined no matter what value is used. This gives the
2594 compiler more freedom to optimize. Here are some examples of
2595 (potentially surprising) transformations that are valid (in pseudo IR):
2597 .. code-block:: llvm
2607 This is safe because all of the output bits are affected by the undef
2608 bits. Any output bit can have a zero or one depending on the input bits.
2610 .. code-block:: llvm
2621 These logical operations have bits that are not always affected by the
2622 input. For example, if ``%X`` has a zero bit, then the output of the
2623 '``and``' operation will always be a zero for that bit, no matter what
2624 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2625 optimize or assume that the result of the '``and``' is '``undef``'.
2626 However, it is safe to assume that all bits of the '``undef``' could be
2627 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2628 all the bits of the '``undef``' operand to the '``or``' could be set,
2629 allowing the '``or``' to be folded to -1.
2631 .. code-block:: llvm
2633 %A = select undef, %X, %Y
2634 %B = select undef, 42, %Y
2635 %C = select %X, %Y, undef
2645 This set of examples shows that undefined '``select``' (and conditional
2646 branch) conditions can go *either way*, but they have to come from one
2647 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2648 both known to have a clear low bit, then ``%A`` would have to have a
2649 cleared low bit. However, in the ``%C`` example, the optimizer is
2650 allowed to assume that the '``undef``' operand could be the same as
2651 ``%Y``, allowing the whole '``select``' to be eliminated.
2653 .. code-block:: llvm
2655 %A = xor undef, undef
2672 This example points out that two '``undef``' operands are not
2673 necessarily the same. This can be surprising to people (and also matches
2674 C semantics) where they assume that "``X^X``" is always zero, even if
2675 ``X`` is undefined. This isn't true for a number of reasons, but the
2676 short answer is that an '``undef``' "variable" can arbitrarily change
2677 its value over its "live range". This is true because the variable
2678 doesn't actually *have a live range*. Instead, the value is logically
2679 read from arbitrary registers that happen to be around when needed, so
2680 the value is not necessarily consistent over time. In fact, ``%A`` and
2681 ``%C`` need to have the same semantics or the core LLVM "replace all
2682 uses with" concept would not hold.
2684 .. code-block:: llvm
2692 These examples show the crucial difference between an *undefined value*
2693 and *undefined behavior*. An undefined value (like '``undef``') is
2694 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2695 operation can be constant folded to '``undef``', because the '``undef``'
2696 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2697 However, in the second example, we can make a more aggressive
2698 assumption: because the ``undef`` is allowed to be an arbitrary value,
2699 we are allowed to assume that it could be zero. Since a divide by zero
2700 has *undefined behavior*, we are allowed to assume that the operation
2701 does not execute at all. This allows us to delete the divide and all
2702 code after it. Because the undefined operation "can't happen", the
2703 optimizer can assume that it occurs in dead code.
2705 .. code-block:: llvm
2707 a: store undef -> %X
2708 b: store %X -> undef
2713 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2714 value can be assumed to not have any effect; we can assume that the
2715 value is overwritten with bits that happen to match what was already
2716 there. However, a store *to* an undefined location could clobber
2717 arbitrary memory, therefore, it has undefined behavior.
2724 Poison values are similar to :ref:`undef values <undefvalues>`, however
2725 they also represent the fact that an instruction or constant expression
2726 that cannot evoke side effects has nevertheless detected a condition
2727 that results in undefined behavior.
2729 There is currently no way of representing a poison value in the IR; they
2730 only exist when produced by operations such as :ref:`add <i_add>` with
2733 Poison value behavior is defined in terms of value *dependence*:
2735 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2736 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2737 their dynamic predecessor basic block.
2738 - Function arguments depend on the corresponding actual argument values
2739 in the dynamic callers of their functions.
2740 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2741 instructions that dynamically transfer control back to them.
2742 - :ref:`Invoke <i_invoke>` instructions depend on the
2743 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2744 call instructions that dynamically transfer control back to them.
2745 - Non-volatile loads and stores depend on the most recent stores to all
2746 of the referenced memory addresses, following the order in the IR
2747 (including loads and stores implied by intrinsics such as
2748 :ref:`@llvm.memcpy <int_memcpy>`.)
2749 - An instruction with externally visible side effects depends on the
2750 most recent preceding instruction with externally visible side
2751 effects, following the order in the IR. (This includes :ref:`volatile
2752 operations <volatile>`.)
2753 - An instruction *control-depends* on a :ref:`terminator
2754 instruction <terminators>` if the terminator instruction has
2755 multiple successors and the instruction is always executed when
2756 control transfers to one of the successors, and may not be executed
2757 when control is transferred to another.
2758 - Additionally, an instruction also *control-depends* on a terminator
2759 instruction if the set of instructions it otherwise depends on would
2760 be different if the terminator had transferred control to a different
2762 - Dependence is transitive.
2764 Poison values have the same behavior as :ref:`undef values <undefvalues>`,
2765 with the additional effect that any instruction that has a *dependence*
2766 on a poison value has undefined behavior.
2768 Here are some examples:
2770 .. code-block:: llvm
2773 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2774 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2775 %poison_yet_again = getelementptr i32, i32* @h, i32 %still_poison
2776 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2778 store i32 %poison, i32* @g ; Poison value stored to memory.
2779 %poison2 = load i32, i32* @g ; Poison value loaded back from memory.
2781 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2783 %narrowaddr = bitcast i32* @g to i16*
2784 %wideaddr = bitcast i32* @g to i64*
2785 %poison3 = load i16, i16* %narrowaddr ; Returns a poison value.
2786 %poison4 = load i64, i64* %wideaddr ; Returns a poison value.
2788 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2789 br i1 %cmp, label %true, label %end ; Branch to either destination.
2792 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2793 ; it has undefined behavior.
2797 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2798 ; Both edges into this PHI are
2799 ; control-dependent on %cmp, so this
2800 ; always results in a poison value.
2802 store volatile i32 0, i32* @g ; This would depend on the store in %true
2803 ; if %cmp is true, or the store in %entry
2804 ; otherwise, so this is undefined behavior.
2806 br i1 %cmp, label %second_true, label %second_end
2807 ; The same branch again, but this time the
2808 ; true block doesn't have side effects.
2815 store volatile i32 0, i32* @g ; This time, the instruction always depends
2816 ; on the store in %end. Also, it is
2817 ; control-equivalent to %end, so this is
2818 ; well-defined (ignoring earlier undefined
2819 ; behavior in this example).
2823 Addresses of Basic Blocks
2824 -------------------------
2826 ``blockaddress(@function, %block)``
2828 The '``blockaddress``' constant computes the address of the specified
2829 basic block in the specified function, and always has an ``i8*`` type.
2830 Taking the address of the entry block is illegal.
2832 This value only has defined behavior when used as an operand to the
2833 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2834 against null. Pointer equality tests between labels addresses results in
2835 undefined behavior --- though, again, comparison against null is ok, and
2836 no label is equal to the null pointer. This may be passed around as an
2837 opaque pointer sized value as long as the bits are not inspected. This
2838 allows ``ptrtoint`` and arithmetic to be performed on these values so
2839 long as the original value is reconstituted before the ``indirectbr``
2842 Finally, some targets may provide defined semantics when using the value
2843 as the operand to an inline assembly, but that is target specific.
2847 Constant Expressions
2848 --------------------
2850 Constant expressions are used to allow expressions involving other
2851 constants to be used as constants. Constant expressions may be of any
2852 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2853 that does not have side effects (e.g. load and call are not supported).
2854 The following is the syntax for constant expressions:
2856 ``trunc (CST to TYPE)``
2857 Truncate a constant to another type. The bit size of CST must be
2858 larger than the bit size of TYPE. Both types must be integers.
2859 ``zext (CST to TYPE)``
2860 Zero extend a constant to another type. The bit size of CST must be
2861 smaller than the bit size of TYPE. Both types must be integers.
2862 ``sext (CST to TYPE)``
2863 Sign extend a constant to another type. The bit size of CST must be
2864 smaller than the bit size of TYPE. Both types must be integers.
2865 ``fptrunc (CST to TYPE)``
2866 Truncate a floating point constant to another floating point type.
2867 The size of CST must be larger than the size of TYPE. Both types
2868 must be floating point.
2869 ``fpext (CST to TYPE)``
2870 Floating point extend a constant to another type. The size of CST
2871 must be smaller or equal to the size of TYPE. Both types must be
2873 ``fptoui (CST to TYPE)``
2874 Convert a floating point constant to the corresponding unsigned
2875 integer constant. TYPE must be a scalar or vector integer type. CST
2876 must be of scalar or vector floating point type. Both CST and TYPE
2877 must be scalars, or vectors of the same number of elements. If the
2878 value won't fit in the integer type, the results are undefined.
2879 ``fptosi (CST to TYPE)``
2880 Convert a floating point constant to the corresponding signed
2881 integer constant. TYPE must be a scalar or vector integer type. CST
2882 must be of scalar or vector floating point type. Both CST and TYPE
2883 must be scalars, or vectors of the same number of elements. If the
2884 value won't fit in the integer type, the results are undefined.
2885 ``uitofp (CST to TYPE)``
2886 Convert an unsigned integer constant to the corresponding floating
2887 point constant. TYPE must be a scalar or vector floating point type.
2888 CST must be of scalar or vector integer type. Both CST and TYPE must
2889 be scalars, or vectors of the same number of elements. If the value
2890 won't fit in the floating point type, the results are undefined.
2891 ``sitofp (CST to TYPE)``
2892 Convert a signed integer constant to the corresponding floating
2893 point constant. TYPE must be a scalar or vector floating point type.
2894 CST must be of scalar or vector integer type. Both CST and TYPE must
2895 be scalars, or vectors of the same number of elements. If the value
2896 won't fit in the floating point type, the results are undefined.
2897 ``ptrtoint (CST to TYPE)``
2898 Convert a pointer typed constant to the corresponding integer
2899 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2900 pointer type. The ``CST`` value is zero extended, truncated, or
2901 unchanged to make it fit in ``TYPE``.
2902 ``inttoptr (CST to TYPE)``
2903 Convert an integer constant to a pointer constant. TYPE must be a
2904 pointer type. CST must be of integer type. The CST value is zero
2905 extended, truncated, or unchanged to make it fit in a pointer size.
2906 This one is *really* dangerous!
2907 ``bitcast (CST to TYPE)``
2908 Convert a constant, CST, to another TYPE. The constraints of the
2909 operands are the same as those for the :ref:`bitcast
2910 instruction <i_bitcast>`.
2911 ``addrspacecast (CST to TYPE)``
2912 Convert a constant pointer or constant vector of pointer, CST, to another
2913 TYPE in a different address space. The constraints of the operands are the
2914 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2915 ``getelementptr (TY, CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (TY, CSTPTR, IDX0, IDX1, ...)``
2916 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2917 constants. As with the :ref:`getelementptr <i_getelementptr>`
2918 instruction, the index list may have zero or more indexes, which are
2919 required to make sense for the type of "pointer to TY".
2920 ``select (COND, VAL1, VAL2)``
2921 Perform the :ref:`select operation <i_select>` on constants.
2922 ``icmp COND (VAL1, VAL2)``
2923 Performs the :ref:`icmp operation <i_icmp>` on constants.
2924 ``fcmp COND (VAL1, VAL2)``
2925 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2926 ``extractelement (VAL, IDX)``
2927 Perform the :ref:`extractelement operation <i_extractelement>` on
2929 ``insertelement (VAL, ELT, IDX)``
2930 Perform the :ref:`insertelement operation <i_insertelement>` on
2932 ``shufflevector (VEC1, VEC2, IDXMASK)``
2933 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2935 ``extractvalue (VAL, IDX0, IDX1, ...)``
2936 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2937 constants. The index list is interpreted in a similar manner as
2938 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2939 least one index value must be specified.
2940 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2941 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2942 The index list is interpreted in a similar manner as indices in a
2943 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2944 value must be specified.
2945 ``OPCODE (LHS, RHS)``
2946 Perform the specified operation of the LHS and RHS constants. OPCODE
2947 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2948 binary <bitwiseops>` operations. The constraints on operands are
2949 the same as those for the corresponding instruction (e.g. no bitwise
2950 operations on floating point values are allowed).
2957 Inline Assembler Expressions
2958 ----------------------------
2960 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2961 Inline Assembly <moduleasm>`) through the use of a special value. This value
2962 represents the inline assembler as a template string (containing the
2963 instructions to emit), a list of operand constraints (stored as a string), a
2964 flag that indicates whether or not the inline asm expression has side effects,
2965 and a flag indicating whether the function containing the asm needs to align its
2966 stack conservatively.
2968 The template string supports argument substitution of the operands using "``$``"
2969 followed by a number, to indicate substitution of the given register/memory
2970 location, as specified by the constraint string. "``${NUM:MODIFIER}``" may also
2971 be used, where ``MODIFIER`` is a target-specific annotation for how to print the
2972 operand (See :ref:`inline-asm-modifiers`).
2974 A literal "``$``" may be included by using "``$$``" in the template. To include
2975 other special characters into the output, the usual "``\XX``" escapes may be
2976 used, just as in other strings. Note that after template substitution, the
2977 resulting assembly string is parsed by LLVM's integrated assembler unless it is
2978 disabled -- even when emitting a ``.s`` file -- and thus must contain assembly
2979 syntax known to LLVM.
2981 LLVM's support for inline asm is modeled closely on the requirements of Clang's
2982 GCC-compatible inline-asm support. Thus, the feature-set and the constraint and
2983 modifier codes listed here are similar or identical to those in GCC's inline asm
2984 support. However, to be clear, the syntax of the template and constraint strings
2985 described here is *not* the same as the syntax accepted by GCC and Clang, and,
2986 while most constraint letters are passed through as-is by Clang, some get
2987 translated to other codes when converting from the C source to the LLVM
2990 An example inline assembler expression is:
2992 .. code-block:: llvm
2994 i32 (i32) asm "bswap $0", "=r,r"
2996 Inline assembler expressions may **only** be used as the callee operand
2997 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2998 Thus, typically we have:
3000 .. code-block:: llvm
3002 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
3004 Inline asms with side effects not visible in the constraint list must be
3005 marked as having side effects. This is done through the use of the
3006 '``sideeffect``' keyword, like so:
3008 .. code-block:: llvm
3010 call void asm sideeffect "eieio", ""()
3012 In some cases inline asms will contain code that will not work unless
3013 the stack is aligned in some way, such as calls or SSE instructions on
3014 x86, yet will not contain code that does that alignment within the asm.
3015 The compiler should make conservative assumptions about what the asm
3016 might contain and should generate its usual stack alignment code in the
3017 prologue if the '``alignstack``' keyword is present:
3019 .. code-block:: llvm
3021 call void asm alignstack "eieio", ""()
3023 Inline asms also support using non-standard assembly dialects. The
3024 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
3025 the inline asm is using the Intel dialect. Currently, ATT and Intel are
3026 the only supported dialects. An example is:
3028 .. code-block:: llvm
3030 call void asm inteldialect "eieio", ""()
3032 If multiple keywords appear the '``sideeffect``' keyword must come
3033 first, the '``alignstack``' keyword second and the '``inteldialect``'
3036 Inline Asm Constraint String
3037 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3039 The constraint list is a comma-separated string, each element containing one or
3040 more constraint codes.
3042 For each element in the constraint list an appropriate register or memory
3043 operand will be chosen, and it will be made available to assembly template
3044 string expansion as ``$0`` for the first constraint in the list, ``$1`` for the
3047 There are three different types of constraints, which are distinguished by a
3048 prefix symbol in front of the constraint code: Output, Input, and Clobber. The
3049 constraints must always be given in that order: outputs first, then inputs, then
3050 clobbers. They cannot be intermingled.
3052 There are also three different categories of constraint codes:
3054 - Register constraint. This is either a register class, or a fixed physical
3055 register. This kind of constraint will allocate a register, and if necessary,
3056 bitcast the argument or result to the appropriate type.
3057 - Memory constraint. This kind of constraint is for use with an instruction
3058 taking a memory operand. Different constraints allow for different addressing
3059 modes used by the target.
3060 - Immediate value constraint. This kind of constraint is for an integer or other
3061 immediate value which can be rendered directly into an instruction. The
3062 various target-specific constraints allow the selection of a value in the
3063 proper range for the instruction you wish to use it with.
3068 Output constraints are specified by an "``=``" prefix (e.g. "``=r``"). This
3069 indicates that the assembly will write to this operand, and the operand will
3070 then be made available as a return value of the ``asm`` expression. Output
3071 constraints do not consume an argument from the call instruction. (Except, see
3072 below about indirect outputs).
3074 Normally, it is expected that no output locations are written to by the assembly
3075 expression until *all* of the inputs have been read. As such, LLVM may assign
3076 the same register to an output and an input. If this is not safe (e.g. if the
3077 assembly contains two instructions, where the first writes to one output, and
3078 the second reads an input and writes to a second output), then the "``&``"
3079 modifier must be used (e.g. "``=&r``") to specify that the output is an
3080 "early-clobber" output. Marking an ouput as "early-clobber" ensures that LLVM
3081 will not use the same register for any inputs (other than an input tied to this
3087 Input constraints do not have a prefix -- just the constraint codes. Each input
3088 constraint will consume one argument from the call instruction. It is not
3089 permitted for the asm to write to any input register or memory location (unless
3090 that input is tied to an output). Note also that multiple inputs may all be
3091 assigned to the same register, if LLVM can determine that they necessarily all
3092 contain the same value.
3094 Instead of providing a Constraint Code, input constraints may also "tie"
3095 themselves to an output constraint, by providing an integer as the constraint
3096 string. Tied inputs still consume an argument from the call instruction, and
3097 take up a position in the asm template numbering as is usual -- they will simply
3098 be constrained to always use the same register as the output they've been tied
3099 to. For example, a constraint string of "``=r,0``" says to assign a register for
3100 output, and use that register as an input as well (it being the 0'th
3103 It is permitted to tie an input to an "early-clobber" output. In that case, no
3104 *other* input may share the same register as the input tied to the early-clobber
3105 (even when the other input has the same value).
3107 You may only tie an input to an output which has a register constraint, not a
3108 memory constraint. Only a single input may be tied to an output.
3110 There is also an "interesting" feature which deserves a bit of explanation: if a
3111 register class constraint allocates a register which is too small for the value
3112 type operand provided as input, the input value will be split into multiple
3113 registers, and all of them passed to the inline asm.
3115 However, this feature is often not as useful as you might think.
3117 Firstly, the registers are *not* guaranteed to be consecutive. So, on those
3118 architectures that have instructions which operate on multiple consecutive
3119 instructions, this is not an appropriate way to support them. (e.g. the 32-bit
3120 SparcV8 has a 64-bit load, which instruction takes a single 32-bit register. The
3121 hardware then loads into both the named register, and the next register. This
3122 feature of inline asm would not be useful to support that.)
3124 A few of the targets provide a template string modifier allowing explicit access
3125 to the second register of a two-register operand (e.g. MIPS ``L``, ``M``, and
3126 ``D``). On such an architecture, you can actually access the second allocated
3127 register (yet, still, not any subsequent ones). But, in that case, you're still
3128 probably better off simply splitting the value into two separate operands, for
3129 clarity. (e.g. see the description of the ``A`` constraint on X86, which,
3130 despite existing only for use with this feature, is not really a good idea to
3133 Indirect inputs and outputs
3134 """""""""""""""""""""""""""
3136 Indirect output or input constraints can be specified by the "``*``" modifier
3137 (which goes after the "``=``" in case of an output). This indicates that the asm
3138 will write to or read from the contents of an *address* provided as an input
3139 argument. (Note that in this way, indirect outputs act more like an *input* than
3140 an output: just like an input, they consume an argument of the call expression,
3141 rather than producing a return value. An indirect output constraint is an
3142 "output" only in that the asm is expected to write to the contents of the input
3143 memory location, instead of just read from it).
3145 This is most typically used for memory constraint, e.g. "``=*m``", to pass the
3146 address of a variable as a value.
3148 It is also possible to use an indirect *register* constraint, but only on output
3149 (e.g. "``=*r``"). This will cause LLVM to allocate a register for an output
3150 value normally, and then, separately emit a store to the address provided as
3151 input, after the provided inline asm. (It's not clear what value this
3152 functionality provides, compared to writing the store explicitly after the asm
3153 statement, and it can only produce worse code, since it bypasses many
3154 optimization passes. I would recommend not using it.)
3160 A clobber constraint is indicated by a "``~``" prefix. A clobber does not
3161 consume an input operand, nor generate an output. Clobbers cannot use any of the
3162 general constraint code letters -- they may use only explicit register
3163 constraints, e.g. "``~{eax}``". The one exception is that a clobber string of
3164 "``~{memory}``" indicates that the assembly writes to arbitrary undeclared
3165 memory locations -- not only the memory pointed to by a declared indirect
3171 After a potential prefix comes constraint code, or codes.
3173 A Constraint Code is either a single letter (e.g. "``r``"), a "``^``" character
3174 followed by two letters (e.g. "``^wc``"), or "``{``" register-name "``}``"
3177 The one and two letter constraint codes are typically chosen to be the same as
3178 GCC's constraint codes.
3180 A single constraint may include one or more than constraint code in it, leaving
3181 it up to LLVM to choose which one to use. This is included mainly for
3182 compatibility with the translation of GCC inline asm coming from clang.
3184 There are two ways to specify alternatives, and either or both may be used in an
3185 inline asm constraint list:
3187 1) Append the codes to each other, making a constraint code set. E.g. "``im``"
3188 or "``{eax}m``". This means "choose any of the options in the set". The
3189 choice of constraint is made independently for each constraint in the
3192 2) Use "``|``" between constraint code sets, creating alternatives. Every
3193 constraint in the constraint list must have the same number of alternative
3194 sets. With this syntax, the same alternative in *all* of the items in the
3195 constraint list will be chosen together.
3197 Putting those together, you might have a two operand constraint string like
3198 ``"rm|r,ri|rm"``. This indicates that if operand 0 is ``r`` or ``m``, then
3199 operand 1 may be one of ``r`` or ``i``. If operand 0 is ``r``, then operand 1
3200 may be one of ``r`` or ``m``. But, operand 0 and 1 cannot both be of type m.
3202 However, the use of either of the alternatives features is *NOT* recommended, as
3203 LLVM is not able to make an intelligent choice about which one to use. (At the
3204 point it currently needs to choose, not enough information is available to do so
3205 in a smart way.) Thus, it simply tries to make a choice that's most likely to
3206 compile, not one that will be optimal performance. (e.g., given "``rm``", it'll
3207 always choose to use memory, not registers). And, if given multiple registers,
3208 or multiple register classes, it will simply choose the first one. (In fact, it
3209 doesn't currently even ensure explicitly specified physical registers are
3210 unique, so specifying multiple physical registers as alternatives, like
3211 ``{r11}{r12},{r11}{r12}``, will assign r11 to both operands, not at all what was
3214 Supported Constraint Code List
3215 """"""""""""""""""""""""""""""
3217 The constraint codes are, in general, expected to behave the same way they do in
3218 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3219 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3220 and GCC likely indicates a bug in LLVM.
3222 Some constraint codes are typically supported by all targets:
3224 - ``r``: A register in the target's general purpose register class.
3225 - ``m``: A memory address operand. It is target-specific what addressing modes
3226 are supported, typical examples are register, or register + register offset,
3227 or register + immediate offset (of some target-specific size).
3228 - ``i``: An integer constant (of target-specific width). Allows either a simple
3229 immediate, or a relocatable value.
3230 - ``n``: An integer constant -- *not* including relocatable values.
3231 - ``s``: An integer constant, but allowing *only* relocatable values.
3232 - ``X``: Allows an operand of any kind, no constraint whatsoever. Typically
3233 useful to pass a label for an asm branch or call.
3235 .. FIXME: but that surely isn't actually okay to jump out of an asm
3236 block without telling llvm about the control transfer???)
3238 - ``{register-name}``: Requires exactly the named physical register.
3240 Other constraints are target-specific:
3244 - ``z``: An immediate integer 0. Outputs ``WZR`` or ``XZR``, as appropriate.
3245 - ``I``: An immediate integer valid for an ``ADD`` or ``SUB`` instruction,
3246 i.e. 0 to 4095 with optional shift by 12.
3247 - ``J``: An immediate integer that, when negated, is valid for an ``ADD`` or
3248 ``SUB`` instruction, i.e. -1 to -4095 with optional left shift by 12.
3249 - ``K``: An immediate integer that is valid for the 'bitmask immediate 32' of a
3250 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 32-bit register.
3251 - ``L``: An immediate integer that is valid for the 'bitmask immediate 64' of a
3252 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 64-bit register.
3253 - ``M``: An immediate integer for use with the ``MOV`` assembly alias on a
3254 32-bit register. This is a superset of ``K``: in addition to the bitmask
3255 immediate, also allows immediate integers which can be loaded with a single
3256 ``MOVZ`` or ``MOVL`` instruction.
3257 - ``N``: An immediate integer for use with the ``MOV`` assembly alias on a
3258 64-bit register. This is a superset of ``L``.
3259 - ``Q``: Memory address operand must be in a single register (no
3260 offsets). (However, LLVM currently does this for the ``m`` constraint as
3262 - ``r``: A 32 or 64-bit integer register (W* or X*).
3263 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register.
3264 - ``x``: A lower 128-bit floating-point/SIMD register (``V0`` to ``V15``).
3268 - ``r``: A 32 or 64-bit integer register.
3269 - ``[0-9]v``: The 32-bit VGPR register, number 0-9.
3270 - ``[0-9]s``: The 32-bit SGPR register, number 0-9.
3275 - ``Q``, ``Um``, ``Un``, ``Uq``, ``Us``, ``Ut``, ``Uv``, ``Uy``: Memory address
3276 operand. Treated the same as operand ``m``, at the moment.
3278 ARM and ARM's Thumb2 mode:
3280 - ``j``: An immediate integer between 0 and 65535 (valid for ``MOVW``)
3281 - ``I``: An immediate integer valid for a data-processing instruction.
3282 - ``J``: An immediate integer between -4095 and 4095.
3283 - ``K``: An immediate integer whose bitwise inverse is valid for a
3284 data-processing instruction. (Can be used with template modifier "``B``" to
3285 print the inverted value).
3286 - ``L``: An immediate integer whose negation is valid for a data-processing
3287 instruction. (Can be used with template modifier "``n``" to print the negated
3289 - ``M``: A power of two or a integer between 0 and 32.
3290 - ``N``: Invalid immediate constraint.
3291 - ``O``: Invalid immediate constraint.
3292 - ``r``: A general-purpose 32-bit integer register (``r0-r15``).
3293 - ``l``: In Thumb2 mode, low 32-bit GPR registers (``r0-r7``). In ARM mode, same
3295 - ``h``: In Thumb2 mode, a high 32-bit GPR register (``r8-r15``). In ARM mode,
3297 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3298 ``d0-d31``, or ``q0-q15``.
3299 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3300 ``d0-d7``, or ``q0-q3``.
3301 - ``t``: A floating-point/SIMD register, only supports 32-bit values:
3306 - ``I``: An immediate integer between 0 and 255.
3307 - ``J``: An immediate integer between -255 and -1.
3308 - ``K``: An immediate integer between 0 and 255, with optional left-shift by
3310 - ``L``: An immediate integer between -7 and 7.
3311 - ``M``: An immediate integer which is a multiple of 4 between 0 and 1020.
3312 - ``N``: An immediate integer between 0 and 31.
3313 - ``O``: An immediate integer which is a multiple of 4 between -508 and 508.
3314 - ``r``: A low 32-bit GPR register (``r0-r7``).
3315 - ``l``: A low 32-bit GPR register (``r0-r7``).
3316 - ``h``: A high GPR register (``r0-r7``).
3317 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3318 ``d0-d31``, or ``q0-q15``.
3319 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3320 ``d0-d7``, or ``q0-q3``.
3321 - ``t``: A floating-point/SIMD register, only supports 32-bit values:
3327 - ``o``, ``v``: A memory address operand, treated the same as constraint ``m``,
3329 - ``r``: A 32 or 64-bit register.
3333 - ``r``: An 8 or 16-bit register.
3337 - ``I``: An immediate signed 16-bit integer.
3338 - ``J``: An immediate integer zero.
3339 - ``K``: An immediate unsigned 16-bit integer.
3340 - ``L``: An immediate 32-bit integer, where the lower 16 bits are 0.
3341 - ``N``: An immediate integer between -65535 and -1.
3342 - ``O``: An immediate signed 15-bit integer.
3343 - ``P``: An immediate integer between 1 and 65535.
3344 - ``m``: A memory address operand. In MIPS-SE mode, allows a base address
3345 register plus 16-bit immediate offset. In MIPS mode, just a base register.
3346 - ``R``: A memory address operand. In MIPS-SE mode, allows a base address
3347 register plus a 9-bit signed offset. In MIPS mode, the same as constraint
3349 - ``ZC``: A memory address operand, suitable for use in a ``pref``, ``ll``, or
3350 ``sc`` instruction on the given subtarget (details vary).
3351 - ``r``, ``d``, ``y``: A 32 or 64-bit GPR register.
3352 - ``f``: A 32 or 64-bit FPU register (``F0-F31``), or a 128-bit MSA register
3353 (``W0-W31``). In the case of MSA registers, it is recommended to use the ``w``
3354 argument modifier for compatibility with GCC.
3355 - ``c``: A 32-bit or 64-bit GPR register suitable for indirect jump (always
3357 - ``l``: The ``lo`` register, 32 or 64-bit.
3362 - ``b``: A 1-bit integer register.
3363 - ``c`` or ``h``: A 16-bit integer register.
3364 - ``r``: A 32-bit integer register.
3365 - ``l`` or ``N``: A 64-bit integer register.
3366 - ``f``: A 32-bit float register.
3367 - ``d``: A 64-bit float register.
3372 - ``I``: An immediate signed 16-bit integer.
3373 - ``J``: An immediate unsigned 16-bit integer, shifted left 16 bits.
3374 - ``K``: An immediate unsigned 16-bit integer.
3375 - ``L``: An immediate signed 16-bit integer, shifted left 16 bits.
3376 - ``M``: An immediate integer greater than 31.
3377 - ``N``: An immediate integer that is an exact power of 2.
3378 - ``O``: The immediate integer constant 0.
3379 - ``P``: An immediate integer constant whose negation is a signed 16-bit
3381 - ``es``, ``o``, ``Q``, ``Z``, ``Zy``: A memory address operand, currently
3382 treated the same as ``m``.
3383 - ``r``: A 32 or 64-bit integer register.
3384 - ``b``: A 32 or 64-bit integer register, excluding ``R0`` (that is:
3386 - ``f``: A 32 or 64-bit float register (``F0-F31``), or when QPX is enabled, a
3387 128 or 256-bit QPX register (``Q0-Q31``; aliases the ``F`` registers).
3388 - ``v``: For ``4 x f32`` or ``4 x f64`` types, when QPX is enabled, a
3389 128 or 256-bit QPX register (``Q0-Q31``), otherwise a 128-bit
3390 altivec vector register (``V0-V31``).
3392 .. FIXME: is this a bug that v accepts QPX registers? I think this
3393 is supposed to only use the altivec vector registers?
3395 - ``y``: Condition register (``CR0-CR7``).
3396 - ``wc``: An individual CR bit in a CR register.
3397 - ``wa``, ``wd``, ``wf``: Any 128-bit VSX vector register, from the full VSX
3398 register set (overlapping both the floating-point and vector register files).
3399 - ``ws``: A 32 or 64-bit floating point register, from the full VSX register
3404 - ``I``: An immediate 13-bit signed integer.
3405 - ``r``: A 32-bit integer register.
3409 - ``I``: An immediate unsigned 8-bit integer.
3410 - ``J``: An immediate unsigned 12-bit integer.
3411 - ``K``: An immediate signed 16-bit integer.
3412 - ``L``: An immediate signed 20-bit integer.
3413 - ``M``: An immediate integer 0x7fffffff.
3414 - ``Q``, ``R``, ``S``, ``T``: A memory address operand, treated the same as
3415 ``m``, at the moment.
3416 - ``r`` or ``d``: A 32, 64, or 128-bit integer register.
3417 - ``a``: A 32, 64, or 128-bit integer address register (excludes R0, which in an
3418 address context evaluates as zero).
3419 - ``h``: A 32-bit value in the high part of a 64bit data register
3421 - ``f``: A 32, 64, or 128-bit floating point register.
3425 - ``I``: An immediate integer between 0 and 31.
3426 - ``J``: An immediate integer between 0 and 64.
3427 - ``K``: An immediate signed 8-bit integer.
3428 - ``L``: An immediate integer, 0xff or 0xffff or (in 64-bit mode only)
3430 - ``M``: An immediate integer between 0 and 3.
3431 - ``N``: An immediate unsigned 8-bit integer.
3432 - ``O``: An immediate integer between 0 and 127.
3433 - ``e``: An immediate 32-bit signed integer.
3434 - ``Z``: An immediate 32-bit unsigned integer.
3435 - ``o``, ``v``: Treated the same as ``m``, at the moment.
3436 - ``q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3437 ``l`` integer register. On X86-32, this is the ``a``, ``b``, ``c``, and ``d``
3438 registers, and on X86-64, it is all of the integer registers.
3439 - ``Q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3440 ``h`` integer register. This is the ``a``, ``b``, ``c``, and ``d`` registers.
3441 - ``r`` or ``l``: An 8, 16, 32, or 64-bit integer register.
3442 - ``R``: An 8, 16, 32, or 64-bit "legacy" integer register -- one which has
3443 existed since i386, and can be accessed without the REX prefix.
3444 - ``f``: A 32, 64, or 80-bit '387 FPU stack pseudo-register.
3445 - ``y``: A 64-bit MMX register, if MMX is enabled.
3446 - ``x``: If SSE is enabled: a 32 or 64-bit scalar operand, or 128-bit vector
3447 operand in a SSE register. If AVX is also enabled, can also be a 256-bit
3448 vector operand in an AVX register. If AVX-512 is also enabled, can also be a
3449 512-bit vector operand in an AVX512 register, Otherwise, an error.
3450 - ``Y``: The same as ``x``, if *SSE2* is enabled, otherwise an error.
3451 - ``A``: Special case: allocates EAX first, then EDX, for a single operand (in
3452 32-bit mode, a 64-bit integer operand will get split into two registers). It
3453 is not recommended to use this constraint, as in 64-bit mode, the 64-bit
3454 operand will get allocated only to RAX -- if two 32-bit operands are needed,
3455 you're better off splitting it yourself, before passing it to the asm
3460 - ``r``: A 32-bit integer register.
3463 .. _inline-asm-modifiers:
3465 Asm template argument modifiers
3466 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3468 In the asm template string, modifiers can be used on the operand reference, like
3471 The modifiers are, in general, expected to behave the same way they do in
3472 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3473 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3474 and GCC likely indicates a bug in LLVM.
3478 - ``c``: Print an immediate integer constant unadorned, without
3479 the target-specific immediate punctuation (e.g. no ``$`` prefix).
3480 - ``n``: Negate and print immediate integer constant unadorned, without the
3481 target-specific immediate punctuation (e.g. no ``$`` prefix).
3482 - ``l``: Print as an unadorned label, without the target-specific label
3483 punctuation (e.g. no ``$`` prefix).
3487 - ``w``: Print a GPR register with a ``w*`` name instead of ``x*`` name. E.g.,
3488 instead of ``x30``, print ``w30``.
3489 - ``x``: Print a GPR register with a ``x*`` name. (this is the default, anyhow).
3490 - ``b``, ``h``, ``s``, ``d``, ``q``: Print a floating-point/SIMD register with a
3491 ``b*``, ``h*``, ``s*``, ``d*``, or ``q*`` name, rather than the default of
3500 - ``a``: Print an operand as an address (with ``[`` and ``]`` surrounding a
3504 - ``y``: Print a VFP single-precision register as an indexed double (e.g. print
3505 as ``d4[1]`` instead of ``s9``)
3506 - ``B``: Bitwise invert and print an immediate integer constant without ``#``
3508 - ``L``: Print the low 16-bits of an immediate integer constant.
3509 - ``M``: Print as a register set suitable for ldm/stm. Also prints *all*
3510 register operands subsequent to the specified one (!), so use carefully.
3511 - ``Q``: Print the low-order register of a register-pair, or the low-order
3512 register of a two-register operand.
3513 - ``R``: Print the high-order register of a register-pair, or the high-order
3514 register of a two-register operand.
3515 - ``H``: Print the second register of a register-pair. (On a big-endian system,
3516 ``H`` is equivalent to ``Q``, and on little-endian system, ``H`` is equivalent
3519 .. FIXME: H doesn't currently support printing the second register
3520 of a two-register operand.
3522 - ``e``: Print the low doubleword register of a NEON quad register.
3523 - ``f``: Print the high doubleword register of a NEON quad register.
3524 - ``m``: Print the base register of a memory operand without the ``[`` and ``]``
3529 - ``L``: Print the second register of a two-register operand. Requires that it
3530 has been allocated consecutively to the first.
3532 .. FIXME: why is it restricted to consecutive ones? And there's
3533 nothing that ensures that happens, is there?
3535 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3536 nothing. Used to print 'addi' vs 'add' instructions.
3540 No additional modifiers.
3544 - ``X``: Print an immediate integer as hexadecimal
3545 - ``x``: Print the low 16 bits of an immediate integer as hexadecimal.
3546 - ``d``: Print an immediate integer as decimal.
3547 - ``m``: Subtract one and print an immediate integer as decimal.
3548 - ``z``: Print $0 if an immediate zero, otherwise print normally.
3549 - ``L``: Print the low-order register of a two-register operand, or prints the
3550 address of the low-order word of a double-word memory operand.
3552 .. FIXME: L seems to be missing memory operand support.
3554 - ``M``: Print the high-order register of a two-register operand, or prints the
3555 address of the high-order word of a double-word memory operand.
3557 .. FIXME: M seems to be missing memory operand support.
3559 - ``D``: Print the second register of a two-register operand, or prints the
3560 second word of a double-word memory operand. (On a big-endian system, ``D`` is
3561 equivalent to ``L``, and on little-endian system, ``D`` is equivalent to
3563 - ``w``: No effect. Provided for compatibility with GCC which requires this
3564 modifier in order to print MSA registers (``W0-W31``) with the ``f``
3573 - ``L``: Print the second register of a two-register operand. Requires that it
3574 has been allocated consecutively to the first.
3576 .. FIXME: why is it restricted to consecutive ones? And there's
3577 nothing that ensures that happens, is there?
3579 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3580 nothing. Used to print 'addi' vs 'add' instructions.
3581 - ``y``: For a memory operand, prints formatter for a two-register X-form
3582 instruction. (Currently always prints ``r0,OPERAND``).
3583 - ``U``: Prints 'u' if the memory operand is an update form, and nothing
3584 otherwise. (NOTE: LLVM does not support update form, so this will currently
3585 always print nothing)
3586 - ``X``: Prints 'x' if the memory operand is an indexed form. (NOTE: LLVM does
3587 not support indexed form, so this will currently always print nothing)
3595 SystemZ implements only ``n``, and does *not* support any of the other
3596 target-independent modifiers.
3600 - ``c``: Print an unadorned integer or symbol name. (The latter is
3601 target-specific behavior for this typically target-independent modifier).
3602 - ``A``: Print a register name with a '``*``' before it.
3603 - ``b``: Print an 8-bit register name (e.g. ``al``); do nothing on a memory
3605 - ``h``: Print the upper 8-bit register name (e.g. ``ah``); do nothing on a
3607 - ``w``: Print the 16-bit register name (e.g. ``ax``); do nothing on a memory
3609 - ``k``: Print the 32-bit register name (e.g. ``eax``); do nothing on a memory
3611 - ``q``: Print the 64-bit register name (e.g. ``rax``), if 64-bit registers are
3612 available, otherwise the 32-bit register name; do nothing on a memory operand.
3613 - ``n``: Negate and print an unadorned integer, or, for operands other than an
3614 immediate integer (e.g. a relocatable symbol expression), print a '-' before
3615 the operand. (The behavior for relocatable symbol expressions is a
3616 target-specific behavior for this typically target-independent modifier)
3617 - ``H``: Print a memory reference with additional offset +8.
3618 - ``P``: Print a memory reference or operand for use as the argument of a call
3619 instruction. (E.g. omit ``(rip)``, even though it's PC-relative.)
3623 No additional modifiers.
3629 The call instructions that wrap inline asm nodes may have a
3630 "``!srcloc``" MDNode attached to it that contains a list of constant
3631 integers. If present, the code generator will use the integer as the
3632 location cookie value when report errors through the ``LLVMContext``
3633 error reporting mechanisms. This allows a front-end to correlate backend
3634 errors that occur with inline asm back to the source code that produced
3637 .. code-block:: llvm
3639 call void asm sideeffect "something bad", ""(), !srcloc !42
3641 !42 = !{ i32 1234567 }
3643 It is up to the front-end to make sense of the magic numbers it places
3644 in the IR. If the MDNode contains multiple constants, the code generator
3645 will use the one that corresponds to the line of the asm that the error
3653 LLVM IR allows metadata to be attached to instructions in the program
3654 that can convey extra information about the code to the optimizers and
3655 code generator. One example application of metadata is source-level
3656 debug information. There are two metadata primitives: strings and nodes.
3658 Metadata does not have a type, and is not a value. If referenced from a
3659 ``call`` instruction, it uses the ``metadata`` type.
3661 All metadata are identified in syntax by a exclamation point ('``!``').
3663 .. _metadata-string:
3665 Metadata Nodes and Metadata Strings
3666 -----------------------------------
3668 A metadata string is a string surrounded by double quotes. It can
3669 contain any character by escaping non-printable characters with
3670 "``\xx``" where "``xx``" is the two digit hex code. For example:
3673 Metadata nodes are represented with notation similar to structure
3674 constants (a comma separated list of elements, surrounded by braces and
3675 preceded by an exclamation point). Metadata nodes can have any values as
3676 their operand. For example:
3678 .. code-block:: llvm
3680 !{ !"test\00", i32 10}
3682 Metadata nodes that aren't uniqued use the ``distinct`` keyword. For example:
3684 .. code-block:: llvm
3686 !0 = distinct !{!"test\00", i32 10}
3688 ``distinct`` nodes are useful when nodes shouldn't be merged based on their
3689 content. They can also occur when transformations cause uniquing collisions
3690 when metadata operands change.
3692 A :ref:`named metadata <namedmetadatastructure>` is a collection of
3693 metadata nodes, which can be looked up in the module symbol table. For
3696 .. code-block:: llvm
3700 Metadata can be used as function arguments. Here ``llvm.dbg.value``
3701 function is using two metadata arguments:
3703 .. code-block:: llvm
3705 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
3707 Metadata can be attached to an instruction. Here metadata ``!21`` is attached
3708 to the ``add`` instruction using the ``!dbg`` identifier:
3710 .. code-block:: llvm
3712 %indvar.next = add i64 %indvar, 1, !dbg !21
3714 Metadata can also be attached to a function definition. Here metadata ``!22``
3715 is attached to the ``foo`` function using the ``!dbg`` identifier:
3717 .. code-block:: llvm
3719 define void @foo() !dbg !22 {
3723 More information about specific metadata nodes recognized by the
3724 optimizers and code generator is found below.
3726 .. _specialized-metadata:
3728 Specialized Metadata Nodes
3729 ^^^^^^^^^^^^^^^^^^^^^^^^^^
3731 Specialized metadata nodes are custom data structures in metadata (as opposed
3732 to generic tuples). Their fields are labelled, and can be specified in any
3735 These aren't inherently debug info centric, but currently all the specialized
3736 metadata nodes are related to debug info.
3743 ``DICompileUnit`` nodes represent a compile unit. The ``enums:``,
3744 ``retainedTypes:``, ``subprograms:``, ``globals:`` and ``imports:`` fields are
3745 tuples containing the debug info to be emitted along with the compile unit,
3746 regardless of code optimizations (some nodes are only emitted if there are
3747 references to them from instructions).
3749 .. code-block:: llvm
3751 !0 = !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang",
3752 isOptimized: true, flags: "-O2", runtimeVersion: 2,
3753 splitDebugFilename: "abc.debug", emissionKind: 1,
3754 enums: !2, retainedTypes: !3, subprograms: !4,
3755 globals: !5, imports: !6)
3757 Compile unit descriptors provide the root scope for objects declared in a
3758 specific compilation unit. File descriptors are defined using this scope.
3759 These descriptors are collected by a named metadata ``!llvm.dbg.cu``. They
3760 keep track of subprograms, global variables, type information, and imported
3761 entities (declarations and namespaces).
3768 ``DIFile`` nodes represent files. The ``filename:`` can include slashes.
3770 .. code-block:: llvm
3772 !0 = !DIFile(filename: "path/to/file", directory: "/path/to/dir")
3774 Files are sometimes used in ``scope:`` fields, and are the only valid target
3775 for ``file:`` fields.
3782 ``DIBasicType`` nodes represent primitive types, such as ``int``, ``bool`` and
3783 ``float``. ``tag:`` defaults to ``DW_TAG_base_type``.
3785 .. code-block:: llvm
3787 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
3788 encoding: DW_ATE_unsigned_char)
3789 !1 = !DIBasicType(tag: DW_TAG_unspecified_type, name: "decltype(nullptr)")
3791 The ``encoding:`` describes the details of the type. Usually it's one of the
3794 .. code-block:: llvm
3800 DW_ATE_signed_char = 6
3802 DW_ATE_unsigned_char = 8
3804 .. _DISubroutineType:
3809 ``DISubroutineType`` nodes represent subroutine types. Their ``types:`` field
3810 refers to a tuple; the first operand is the return type, while the rest are the
3811 types of the formal arguments in order. If the first operand is ``null``, that
3812 represents a function with no return value (such as ``void foo() {}`` in C++).
3814 .. code-block:: llvm
3816 !0 = !BasicType(name: "int", size: 32, align: 32, DW_ATE_signed)
3817 !1 = !BasicType(name: "char", size: 8, align: 8, DW_ATE_signed_char)
3818 !2 = !DISubroutineType(types: !{null, !0, !1}) ; void (int, char)
3825 ``DIDerivedType`` nodes represent types derived from other types, such as
3828 .. code-block:: llvm
3830 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
3831 encoding: DW_ATE_unsigned_char)
3832 !1 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !0, size: 32,
3835 The following ``tag:`` values are valid:
3837 .. code-block:: llvm
3839 DW_TAG_formal_parameter = 5
3841 DW_TAG_pointer_type = 15
3842 DW_TAG_reference_type = 16
3844 DW_TAG_ptr_to_member_type = 31
3845 DW_TAG_const_type = 38
3846 DW_TAG_volatile_type = 53
3847 DW_TAG_restrict_type = 55
3849 ``DW_TAG_member`` is used to define a member of a :ref:`composite type
3850 <DICompositeType>` or :ref:`subprogram <DISubprogram>`. The type of the member
3851 is the ``baseType:``. The ``offset:`` is the member's bit offset.
3852 ``DW_TAG_formal_parameter`` is used to define a member which is a formal
3853 argument of a subprogram.
3855 ``DW_TAG_typedef`` is used to provide a name for the ``baseType:``.
3857 ``DW_TAG_pointer_type``, ``DW_TAG_reference_type``, ``DW_TAG_const_type``,
3858 ``DW_TAG_volatile_type`` and ``DW_TAG_restrict_type`` are used to qualify the
3861 Note that the ``void *`` type is expressed as a type derived from NULL.
3863 .. _DICompositeType:
3868 ``DICompositeType`` nodes represent types composed of other types, like
3869 structures and unions. ``elements:`` points to a tuple of the composed types.
3871 If the source language supports ODR, the ``identifier:`` field gives the unique
3872 identifier used for type merging between modules. When specified, other types
3873 can refer to composite types indirectly via a :ref:`metadata string
3874 <metadata-string>` that matches their identifier.
3876 .. code-block:: llvm
3878 !0 = !DIEnumerator(name: "SixKind", value: 7)
3879 !1 = !DIEnumerator(name: "SevenKind", value: 7)
3880 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
3881 !3 = !DICompositeType(tag: DW_TAG_enumeration_type, name: "Enum", file: !12,
3882 line: 2, size: 32, align: 32, identifier: "_M4Enum",
3883 elements: !{!0, !1, !2})
3885 The following ``tag:`` values are valid:
3887 .. code-block:: llvm
3889 DW_TAG_array_type = 1
3890 DW_TAG_class_type = 2
3891 DW_TAG_enumeration_type = 4
3892 DW_TAG_structure_type = 19
3893 DW_TAG_union_type = 23
3894 DW_TAG_subroutine_type = 21
3895 DW_TAG_inheritance = 28
3898 For ``DW_TAG_array_type``, the ``elements:`` should be :ref:`subrange
3899 descriptors <DISubrange>`, each representing the range of subscripts at that
3900 level of indexing. The ``DIFlagVector`` flag to ``flags:`` indicates that an
3901 array type is a native packed vector.
3903 For ``DW_TAG_enumeration_type``, the ``elements:`` should be :ref:`enumerator
3904 descriptors <DIEnumerator>`, each representing the definition of an enumeration
3905 value for the set. All enumeration type descriptors are collected in the
3906 ``enums:`` field of the :ref:`compile unit <DICompileUnit>`.
3908 For ``DW_TAG_structure_type``, ``DW_TAG_class_type``, and
3909 ``DW_TAG_union_type``, the ``elements:`` should be :ref:`derived types
3910 <DIDerivedType>` with ``tag: DW_TAG_member`` or ``tag: DW_TAG_inheritance``.
3917 ``DISubrange`` nodes are the elements for ``DW_TAG_array_type`` variants of
3918 :ref:`DICompositeType`. ``count: -1`` indicates an empty array.
3920 .. code-block:: llvm
3922 !0 = !DISubrange(count: 5, lowerBound: 0) ; array counting from 0
3923 !1 = !DISubrange(count: 5, lowerBound: 1) ; array counting from 1
3924 !2 = !DISubrange(count: -1) ; empty array.
3931 ``DIEnumerator`` nodes are the elements for ``DW_TAG_enumeration_type``
3932 variants of :ref:`DICompositeType`.
3934 .. code-block:: llvm
3936 !0 = !DIEnumerator(name: "SixKind", value: 7)
3937 !1 = !DIEnumerator(name: "SevenKind", value: 7)
3938 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
3940 DITemplateTypeParameter
3941 """""""""""""""""""""""
3943 ``DITemplateTypeParameter`` nodes represent type parameters to generic source
3944 language constructs. They are used (optionally) in :ref:`DICompositeType` and
3945 :ref:`DISubprogram` ``templateParams:`` fields.
3947 .. code-block:: llvm
3949 !0 = !DITemplateTypeParameter(name: "Ty", type: !1)
3951 DITemplateValueParameter
3952 """"""""""""""""""""""""
3954 ``DITemplateValueParameter`` nodes represent value parameters to generic source
3955 language constructs. ``tag:`` defaults to ``DW_TAG_template_value_parameter``,
3956 but if specified can also be set to ``DW_TAG_GNU_template_template_param`` or
3957 ``DW_TAG_GNU_template_param_pack``. They are used (optionally) in
3958 :ref:`DICompositeType` and :ref:`DISubprogram` ``templateParams:`` fields.
3960 .. code-block:: llvm
3962 !0 = !DITemplateValueParameter(name: "Ty", type: !1, value: i32 7)
3967 ``DINamespace`` nodes represent namespaces in the source language.
3969 .. code-block:: llvm
3971 !0 = !DINamespace(name: "myawesomeproject", scope: !1, file: !2, line: 7)
3976 ``DIGlobalVariable`` nodes represent global variables in the source language.
3978 .. code-block:: llvm
3980 !0 = !DIGlobalVariable(name: "foo", linkageName: "foo", scope: !1,
3981 file: !2, line: 7, type: !3, isLocal: true,
3982 isDefinition: false, variable: i32* @foo,
3985 All global variables should be referenced by the `globals:` field of a
3986 :ref:`compile unit <DICompileUnit>`.
3993 ``DISubprogram`` nodes represent functions from the source language. A
3994 ``DISubprogram`` may be attached to a function definition using ``!dbg``
3995 metadata. The ``variables:`` field points at :ref:`variables <DILocalVariable>`
3996 that must be retained, even if their IR counterparts are optimized out of
3997 the IR. The ``type:`` field must point at an :ref:`DISubroutineType`.
3999 .. code-block:: llvm
4001 define void @_Z3foov() !dbg !0 {
4005 !0 = distinct !DISubprogram(name: "foo", linkageName: "_Zfoov", scope: !1,
4006 file: !2, line: 7, type: !3, isLocal: true,
4007 isDefinition: false, scopeLine: 8,
4009 virtuality: DW_VIRTUALITY_pure_virtual,
4010 virtualIndex: 10, flags: DIFlagPrototyped,
4011 isOptimized: true, templateParams: !5,
4012 declaration: !6, variables: !7)
4019 ``DILexicalBlock`` nodes describe nested blocks within a :ref:`subprogram
4020 <DISubprogram>`. The line number and column numbers are used to distinguish
4021 two lexical blocks at same depth. They are valid targets for ``scope:``
4024 .. code-block:: llvm
4026 !0 = distinct !DILexicalBlock(scope: !1, file: !2, line: 7, column: 35)
4028 Usually lexical blocks are ``distinct`` to prevent node merging based on
4031 .. _DILexicalBlockFile:
4036 ``DILexicalBlockFile`` nodes are used to discriminate between sections of a
4037 :ref:`lexical block <DILexicalBlock>`. The ``file:`` field can be changed to
4038 indicate textual inclusion, or the ``discriminator:`` field can be used to
4039 discriminate between control flow within a single block in the source language.
4041 .. code-block:: llvm
4043 !0 = !DILexicalBlock(scope: !3, file: !4, line: 7, column: 35)
4044 !1 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 0)
4045 !2 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 1)
4052 ``DILocation`` nodes represent source debug locations. The ``scope:`` field is
4053 mandatory, and points at an :ref:`DILexicalBlockFile`, an
4054 :ref:`DILexicalBlock`, or an :ref:`DISubprogram`.
4056 .. code-block:: llvm
4058 !0 = !DILocation(line: 2900, column: 42, scope: !1, inlinedAt: !2)
4060 .. _DILocalVariable:
4065 ``DILocalVariable`` nodes represent local variables in the source language. If
4066 the ``arg:`` field is set to non-zero, then this variable is a subprogram
4067 parameter, and it will be included in the ``variables:`` field of its
4068 :ref:`DISubprogram`.
4070 .. code-block:: llvm
4072 !0 = !DILocalVariable(name: "this", arg: 1, scope: !3, file: !2, line: 7,
4073 type: !3, flags: DIFlagArtificial)
4074 !1 = !DILocalVariable(name: "x", arg: 2, scope: !4, file: !2, line: 7,
4076 !2 = !DILocalVariable(name: "y", scope: !5, file: !2, line: 7, type: !3)
4081 ``DIExpression`` nodes represent DWARF expression sequences. They are used in
4082 :ref:`debug intrinsics<dbg_intrinsics>` (such as ``llvm.dbg.declare``) to
4083 describe how the referenced LLVM variable relates to the source language
4086 The current supported vocabulary is limited:
4088 - ``DW_OP_deref`` dereferences the working expression.
4089 - ``DW_OP_plus, 93`` adds ``93`` to the working expression.
4090 - ``DW_OP_bit_piece, 16, 8`` specifies the offset and size (``16`` and ``8``
4091 here, respectively) of the variable piece from the working expression.
4093 .. code-block:: llvm
4095 !0 = !DIExpression(DW_OP_deref)
4096 !1 = !DIExpression(DW_OP_plus, 3)
4097 !2 = !DIExpression(DW_OP_bit_piece, 3, 7)
4098 !3 = !DIExpression(DW_OP_deref, DW_OP_plus, 3, DW_OP_bit_piece, 3, 7)
4103 ``DIObjCProperty`` nodes represent Objective-C property nodes.
4105 .. code-block:: llvm
4107 !3 = !DIObjCProperty(name: "foo", file: !1, line: 7, setter: "setFoo",
4108 getter: "getFoo", attributes: 7, type: !2)
4113 ``DIImportedEntity`` nodes represent entities (such as modules) imported into a
4116 .. code-block:: llvm
4118 !2 = !DIImportedEntity(tag: DW_TAG_imported_module, name: "foo", scope: !0,
4119 entity: !1, line: 7)
4124 In LLVM IR, memory does not have types, so LLVM's own type system is not
4125 suitable for doing TBAA. Instead, metadata is added to the IR to
4126 describe a type system of a higher level language. This can be used to
4127 implement typical C/C++ TBAA, but it can also be used to implement
4128 custom alias analysis behavior for other languages.
4130 The current metadata format is very simple. TBAA metadata nodes have up
4131 to three fields, e.g.:
4133 .. code-block:: llvm
4135 !0 = !{ !"an example type tree" }
4136 !1 = !{ !"int", !0 }
4137 !2 = !{ !"float", !0 }
4138 !3 = !{ !"const float", !2, i64 1 }
4140 The first field is an identity field. It can be any value, usually a
4141 metadata string, which uniquely identifies the type. The most important
4142 name in the tree is the name of the root node. Two trees with different
4143 root node names are entirely disjoint, even if they have leaves with
4146 The second field identifies the type's parent node in the tree, or is
4147 null or omitted for a root node. A type is considered to alias all of
4148 its descendants and all of its ancestors in the tree. Also, a type is
4149 considered to alias all types in other trees, so that bitcode produced
4150 from multiple front-ends is handled conservatively.
4152 If the third field is present, it's an integer which if equal to 1
4153 indicates that the type is "constant" (meaning
4154 ``pointsToConstantMemory`` should return true; see `other useful
4155 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
4157 '``tbaa.struct``' Metadata
4158 ^^^^^^^^^^^^^^^^^^^^^^^^^^
4160 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
4161 aggregate assignment operations in C and similar languages, however it
4162 is defined to copy a contiguous region of memory, which is more than
4163 strictly necessary for aggregate types which contain holes due to
4164 padding. Also, it doesn't contain any TBAA information about the fields
4167 ``!tbaa.struct`` metadata can describe which memory subregions in a
4168 memcpy are padding and what the TBAA tags of the struct are.
4170 The current metadata format is very simple. ``!tbaa.struct`` metadata
4171 nodes are a list of operands which are in conceptual groups of three.
4172 For each group of three, the first operand gives the byte offset of a
4173 field in bytes, the second gives its size in bytes, and the third gives
4176 .. code-block:: llvm
4178 !4 = !{ i64 0, i64 4, !1, i64 8, i64 4, !2 }
4180 This describes a struct with two fields. The first is at offset 0 bytes
4181 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
4182 and has size 4 bytes and has tbaa tag !2.
4184 Note that the fields need not be contiguous. In this example, there is a
4185 4 byte gap between the two fields. This gap represents padding which
4186 does not carry useful data and need not be preserved.
4188 '``noalias``' and '``alias.scope``' Metadata
4189 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4191 ``noalias`` and ``alias.scope`` metadata provide the ability to specify generic
4192 noalias memory-access sets. This means that some collection of memory access
4193 instructions (loads, stores, memory-accessing calls, etc.) that carry
4194 ``noalias`` metadata can specifically be specified not to alias with some other
4195 collection of memory access instructions that carry ``alias.scope`` metadata.
4196 Each type of metadata specifies a list of scopes where each scope has an id and
4197 a domain. When evaluating an aliasing query, if for some domain, the set
4198 of scopes with that domain in one instruction's ``alias.scope`` list is a
4199 subset of (or equal to) the set of scopes for that domain in another
4200 instruction's ``noalias`` list, then the two memory accesses are assumed not to
4203 The metadata identifying each domain is itself a list containing one or two
4204 entries. The first entry is the name of the domain. Note that if the name is a
4205 string then it can be combined across functions and translation units. A
4206 self-reference can be used to create globally unique domain names. A
4207 descriptive string may optionally be provided as a second list entry.
4209 The metadata identifying each scope is also itself a list containing two or
4210 three entries. The first entry is the name of the scope. Note that if the name
4211 is a string then it can be combined across functions and translation units. A
4212 self-reference can be used to create globally unique scope names. A metadata
4213 reference to the scope's domain is the second entry. A descriptive string may
4214 optionally be provided as a third list entry.
4218 .. code-block:: llvm
4220 ; Two scope domains:
4224 ; Some scopes in these domains:
4230 !5 = !{!4} ; A list containing only scope !4
4234 ; These two instructions don't alias:
4235 %0 = load float, float* %c, align 4, !alias.scope !5
4236 store float %0, float* %arrayidx.i, align 4, !noalias !5
4238 ; These two instructions also don't alias (for domain !1, the set of scopes
4239 ; in the !alias.scope equals that in the !noalias list):
4240 %2 = load float, float* %c, align 4, !alias.scope !5
4241 store float %2, float* %arrayidx.i2, align 4, !noalias !6
4243 ; These two instructions may alias (for domain !0, the set of scopes in
4244 ; the !noalias list is not a superset of, or equal to, the scopes in the
4245 ; !alias.scope list):
4246 %2 = load float, float* %c, align 4, !alias.scope !6
4247 store float %0, float* %arrayidx.i, align 4, !noalias !7
4249 '``fpmath``' Metadata
4250 ^^^^^^^^^^^^^^^^^^^^^
4252 ``fpmath`` metadata may be attached to any instruction of floating point
4253 type. It can be used to express the maximum acceptable error in the
4254 result of that instruction, in ULPs, thus potentially allowing the
4255 compiler to use a more efficient but less accurate method of computing
4256 it. ULP is defined as follows:
4258 If ``x`` is a real number that lies between two finite consecutive
4259 floating-point numbers ``a`` and ``b``, without being equal to one
4260 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
4261 distance between the two non-equal finite floating-point numbers
4262 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
4264 The metadata node shall consist of a single positive floating point
4265 number representing the maximum relative error, for example:
4267 .. code-block:: llvm
4269 !0 = !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
4273 '``range``' Metadata
4274 ^^^^^^^^^^^^^^^^^^^^
4276 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
4277 integer types. It expresses the possible ranges the loaded value or the value
4278 returned by the called function at this call site is in. The ranges are
4279 represented with a flattened list of integers. The loaded value or the value
4280 returned is known to be in the union of the ranges defined by each consecutive
4281 pair. Each pair has the following properties:
4283 - The type must match the type loaded by the instruction.
4284 - The pair ``a,b`` represents the range ``[a,b)``.
4285 - Both ``a`` and ``b`` are constants.
4286 - The range is allowed to wrap.
4287 - The range should not represent the full or empty set. That is,
4290 In addition, the pairs must be in signed order of the lower bound and
4291 they must be non-contiguous.
4295 .. code-block:: llvm
4297 %a = load i8, i8* %x, align 1, !range !0 ; Can only be 0 or 1
4298 %b = load i8, i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
4299 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
4300 %d = invoke i8 @bar() to label %cont
4301 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
4303 !0 = !{ i8 0, i8 2 }
4304 !1 = !{ i8 255, i8 2 }
4305 !2 = !{ i8 0, i8 2, i8 3, i8 6 }
4306 !3 = !{ i8 -2, i8 0, i8 3, i8 6 }
4308 '``unpredictable``' Metadata
4309 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4311 ``unpredictable`` metadata may be attached to any branch or switch
4312 instruction. It can be used to express the unpredictability of control
4313 flow. Similar to the llvm.expect intrinsic, it may be used to alter
4314 optimizations related to compare and branch instructions. The metadata
4315 is treated as a boolean value; if it exists, it signals that the branch
4316 or switch that it is attached to is completely unpredictable.
4321 It is sometimes useful to attach information to loop constructs. Currently,
4322 loop metadata is implemented as metadata attached to the branch instruction
4323 in the loop latch block. This type of metadata refer to a metadata node that is
4324 guaranteed to be separate for each loop. The loop identifier metadata is
4325 specified with the name ``llvm.loop``.
4327 The loop identifier metadata is implemented using a metadata that refers to
4328 itself to avoid merging it with any other identifier metadata, e.g.,
4329 during module linkage or function inlining. That is, each loop should refer
4330 to their own identification metadata even if they reside in separate functions.
4331 The following example contains loop identifier metadata for two separate loop
4334 .. code-block:: llvm
4339 The loop identifier metadata can be used to specify additional
4340 per-loop metadata. Any operands after the first operand can be treated
4341 as user-defined metadata. For example the ``llvm.loop.unroll.count``
4342 suggests an unroll factor to the loop unroller:
4344 .. code-block:: llvm
4346 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
4349 !1 = !{!"llvm.loop.unroll.count", i32 4}
4351 '``llvm.loop.vectorize``' and '``llvm.loop.interleave``'
4352 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4354 Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
4355 used to control per-loop vectorization and interleaving parameters such as
4356 vectorization width and interleave count. These metadata should be used in
4357 conjunction with ``llvm.loop`` loop identification metadata. The
4358 ``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
4359 optimization hints and the optimizer will only interleave and vectorize loops if
4360 it believes it is safe to do so. The ``llvm.mem.parallel_loop_access`` metadata
4361 which contains information about loop-carried memory dependencies can be helpful
4362 in determining the safety of these transformations.
4364 '``llvm.loop.interleave.count``' Metadata
4365 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4367 This metadata suggests an interleave count to the loop interleaver.
4368 The first operand is the string ``llvm.loop.interleave.count`` and the
4369 second operand is an integer specifying the interleave count. For
4372 .. code-block:: llvm
4374 !0 = !{!"llvm.loop.interleave.count", i32 4}
4376 Note that setting ``llvm.loop.interleave.count`` to 1 disables interleaving
4377 multiple iterations of the loop. If ``llvm.loop.interleave.count`` is set to 0
4378 then the interleave count will be determined automatically.
4380 '``llvm.loop.vectorize.enable``' Metadata
4381 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4383 This metadata selectively enables or disables vectorization for the loop. The
4384 first operand is the string ``llvm.loop.vectorize.enable`` and the second operand
4385 is a bit. If the bit operand value is 1 vectorization is enabled. A value of
4386 0 disables vectorization:
4388 .. code-block:: llvm
4390 !0 = !{!"llvm.loop.vectorize.enable", i1 0}
4391 !1 = !{!"llvm.loop.vectorize.enable", i1 1}
4393 '``llvm.loop.vectorize.width``' Metadata
4394 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4396 This metadata sets the target width of the vectorizer. The first
4397 operand is the string ``llvm.loop.vectorize.width`` and the second
4398 operand is an integer specifying the width. For example:
4400 .. code-block:: llvm
4402 !0 = !{!"llvm.loop.vectorize.width", i32 4}
4404 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
4405 vectorization of the loop. If ``llvm.loop.vectorize.width`` is set to
4406 0 or if the loop does not have this metadata the width will be
4407 determined automatically.
4409 '``llvm.loop.unroll``'
4410 ^^^^^^^^^^^^^^^^^^^^^^
4412 Metadata prefixed with ``llvm.loop.unroll`` are loop unrolling
4413 optimization hints such as the unroll factor. ``llvm.loop.unroll``
4414 metadata should be used in conjunction with ``llvm.loop`` loop
4415 identification metadata. The ``llvm.loop.unroll`` metadata are only
4416 optimization hints and the unrolling will only be performed if the
4417 optimizer believes it is safe to do so.
4419 '``llvm.loop.unroll.count``' Metadata
4420 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4422 This metadata suggests an unroll factor to the loop unroller. The
4423 first operand is the string ``llvm.loop.unroll.count`` and the second
4424 operand is a positive integer specifying the unroll factor. For
4427 .. code-block:: llvm
4429 !0 = !{!"llvm.loop.unroll.count", i32 4}
4431 If the trip count of the loop is less than the unroll count the loop
4432 will be partially unrolled.
4434 '``llvm.loop.unroll.disable``' Metadata
4435 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4437 This metadata disables loop unrolling. The metadata has a single operand
4438 which is the string ``llvm.loop.unroll.disable``. For example:
4440 .. code-block:: llvm
4442 !0 = !{!"llvm.loop.unroll.disable"}
4444 '``llvm.loop.unroll.runtime.disable``' Metadata
4445 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4447 This metadata disables runtime loop unrolling. The metadata has a single
4448 operand which is the string ``llvm.loop.unroll.runtime.disable``. For example:
4450 .. code-block:: llvm
4452 !0 = !{!"llvm.loop.unroll.runtime.disable"}
4454 '``llvm.loop.unroll.enable``' Metadata
4455 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4457 This metadata suggests that the loop should be fully unrolled if the trip count
4458 is known at compile time and partially unrolled if the trip count is not known
4459 at compile time. The metadata has a single operand which is the string
4460 ``llvm.loop.unroll.enable``. For example:
4462 .. code-block:: llvm
4464 !0 = !{!"llvm.loop.unroll.enable"}
4466 '``llvm.loop.unroll.full``' Metadata
4467 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4469 This metadata suggests that the loop should be unrolled fully. The
4470 metadata has a single operand which is the string ``llvm.loop.unroll.full``.
4473 .. code-block:: llvm
4475 !0 = !{!"llvm.loop.unroll.full"}
4480 Metadata types used to annotate memory accesses with information helpful
4481 for optimizations are prefixed with ``llvm.mem``.
4483 '``llvm.mem.parallel_loop_access``' Metadata
4484 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4486 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
4487 or metadata containing a list of loop identifiers for nested loops.
4488 The metadata is attached to memory accessing instructions and denotes that
4489 no loop carried memory dependence exist between it and other instructions denoted
4490 with the same loop identifier.
4492 Precisely, given two instructions ``m1`` and ``m2`` that both have the
4493 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
4494 set of loops associated with that metadata, respectively, then there is no loop
4495 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
4498 As a special case, if all memory accessing instructions in a loop have
4499 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
4500 loop has no loop carried memory dependences and is considered to be a parallel
4503 Note that if not all memory access instructions have such metadata referring to
4504 the loop, then the loop is considered not being trivially parallel. Additional
4505 memory dependence analysis is required to make that determination. As a fail
4506 safe mechanism, this causes loops that were originally parallel to be considered
4507 sequential (if optimization passes that are unaware of the parallel semantics
4508 insert new memory instructions into the loop body).
4510 Example of a loop that is considered parallel due to its correct use of
4511 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
4512 metadata types that refer to the same loop identifier metadata.
4514 .. code-block:: llvm
4518 %val0 = load i32, i32* %arrayidx, !llvm.mem.parallel_loop_access !0
4520 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
4522 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
4528 It is also possible to have nested parallel loops. In that case the
4529 memory accesses refer to a list of loop identifier metadata nodes instead of
4530 the loop identifier metadata node directly:
4532 .. code-block:: llvm
4536 %val1 = load i32, i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
4538 br label %inner.for.body
4542 %val0 = load i32, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
4544 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
4546 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
4550 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
4552 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
4554 outer.for.end: ; preds = %for.body
4556 !0 = !{!1, !2} ; a list of loop identifiers
4557 !1 = !{!1} ; an identifier for the inner loop
4558 !2 = !{!2} ; an identifier for the outer loop
4563 The ``llvm.bitsets`` global metadata is used to implement
4564 :doc:`bitsets <BitSets>`.
4566 '``invariant.group``' Metadata
4567 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4569 The ``invariant.group`` metadata may be attached to ``load``/``store`` instructions.
4570 The existence of the ``invariant.group`` metadata on the instruction tells
4571 the optimizer that every ``load`` and ``store`` to the same pointer operand
4572 within the same invariant group can be assumed to load or store the same
4573 value (but see the ``llvm.invariant.group.barrier`` intrinsic which affects
4574 when two pointers are considered the same).
4578 .. code-block:: llvm
4580 @unknownPtr = external global i8
4583 store i8 42, i8* %ptr, !invariant.group !0
4584 call void @foo(i8* %ptr)
4586 %a = load i8, i8* %ptr, !invariant.group !0 ; Can assume that value under %ptr didn't change
4587 call void @foo(i8* %ptr)
4588 %b = load i8, i8* %ptr, !invariant.group !1 ; Can't assume anything, because group changed
4590 %newPtr = call i8* @getPointer(i8* %ptr)
4591 %c = load i8, i8* %newPtr, !invariant.group !0 ; Can't assume anything, because we only have information about %ptr
4593 %unknownValue = load i8, i8* @unknownPtr
4594 store i8 %unknownValue, i8* %ptr, !invariant.group !0 ; Can assume that %unknownValue == 42
4596 call void @foo(i8* %ptr)
4597 %newPtr2 = call i8* @llvm.invariant.group.barrier(i8* %ptr)
4598 %d = load i8, i8* %newPtr2, !invariant.group !0 ; Can't step through invariant.group.barrier to get value of %ptr
4601 declare void @foo(i8*)
4602 declare i8* @getPointer(i8*)
4603 declare i8* @llvm.invariant.group.barrier(i8*)
4605 !0 = !{!"magic ptr"}
4606 !1 = !{!"other ptr"}
4610 Module Flags Metadata
4611 =====================
4613 Information about the module as a whole is difficult to convey to LLVM's
4614 subsystems. The LLVM IR isn't sufficient to transmit this information.
4615 The ``llvm.module.flags`` named metadata exists in order to facilitate
4616 this. These flags are in the form of key / value pairs --- much like a
4617 dictionary --- making it easy for any subsystem who cares about a flag to
4620 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
4621 Each triplet has the following form:
4623 - The first element is a *behavior* flag, which specifies the behavior
4624 when two (or more) modules are merged together, and it encounters two
4625 (or more) metadata with the same ID. The supported behaviors are
4627 - The second element is a metadata string that is a unique ID for the
4628 metadata. Each module may only have one flag entry for each unique ID (not
4629 including entries with the **Require** behavior).
4630 - The third element is the value of the flag.
4632 When two (or more) modules are merged together, the resulting
4633 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
4634 each unique metadata ID string, there will be exactly one entry in the merged
4635 modules ``llvm.module.flags`` metadata table, and the value for that entry will
4636 be determined by the merge behavior flag, as described below. The only exception
4637 is that entries with the *Require* behavior are always preserved.
4639 The following behaviors are supported:
4650 Emits an error if two values disagree, otherwise the resulting value
4651 is that of the operands.
4655 Emits a warning if two values disagree. The result value will be the
4656 operand for the flag from the first module being linked.
4660 Adds a requirement that another module flag be present and have a
4661 specified value after linking is performed. The value must be a
4662 metadata pair, where the first element of the pair is the ID of the
4663 module flag to be restricted, and the second element of the pair is
4664 the value the module flag should be restricted to. This behavior can
4665 be used to restrict the allowable results (via triggering of an
4666 error) of linking IDs with the **Override** behavior.
4670 Uses the specified value, regardless of the behavior or value of the
4671 other module. If both modules specify **Override**, but the values
4672 differ, an error will be emitted.
4676 Appends the two values, which are required to be metadata nodes.
4680 Appends the two values, which are required to be metadata
4681 nodes. However, duplicate entries in the second list are dropped
4682 during the append operation.
4684 It is an error for a particular unique flag ID to have multiple behaviors,
4685 except in the case of **Require** (which adds restrictions on another metadata
4686 value) or **Override**.
4688 An example of module flags:
4690 .. code-block:: llvm
4692 !0 = !{ i32 1, !"foo", i32 1 }
4693 !1 = !{ i32 4, !"bar", i32 37 }
4694 !2 = !{ i32 2, !"qux", i32 42 }
4695 !3 = !{ i32 3, !"qux",
4700 !llvm.module.flags = !{ !0, !1, !2, !3 }
4702 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
4703 if two or more ``!"foo"`` flags are seen is to emit an error if their
4704 values are not equal.
4706 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
4707 behavior if two or more ``!"bar"`` flags are seen is to use the value
4710 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
4711 behavior if two or more ``!"qux"`` flags are seen is to emit a
4712 warning if their values are not equal.
4714 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
4720 The behavior is to emit an error if the ``llvm.module.flags`` does not
4721 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
4724 Objective-C Garbage Collection Module Flags Metadata
4725 ----------------------------------------------------
4727 On the Mach-O platform, Objective-C stores metadata about garbage
4728 collection in a special section called "image info". The metadata
4729 consists of a version number and a bitmask specifying what types of
4730 garbage collection are supported (if any) by the file. If two or more
4731 modules are linked together their garbage collection metadata needs to
4732 be merged rather than appended together.
4734 The Objective-C garbage collection module flags metadata consists of the
4735 following key-value pairs:
4744 * - ``Objective-C Version``
4745 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
4747 * - ``Objective-C Image Info Version``
4748 - **[Required]** --- The version of the image info section. Currently
4751 * - ``Objective-C Image Info Section``
4752 - **[Required]** --- The section to place the metadata. Valid values are
4753 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
4754 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
4755 Objective-C ABI version 2.
4757 * - ``Objective-C Garbage Collection``
4758 - **[Required]** --- Specifies whether garbage collection is supported or
4759 not. Valid values are 0, for no garbage collection, and 2, for garbage
4760 collection supported.
4762 * - ``Objective-C GC Only``
4763 - **[Optional]** --- Specifies that only garbage collection is supported.
4764 If present, its value must be 6. This flag requires that the
4765 ``Objective-C Garbage Collection`` flag have the value 2.
4767 Some important flag interactions:
4769 - If a module with ``Objective-C Garbage Collection`` set to 0 is
4770 merged with a module with ``Objective-C Garbage Collection`` set to
4771 2, then the resulting module has the
4772 ``Objective-C Garbage Collection`` flag set to 0.
4773 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
4774 merged with a module with ``Objective-C GC Only`` set to 6.
4776 Automatic Linker Flags Module Flags Metadata
4777 --------------------------------------------
4779 Some targets support embedding flags to the linker inside individual object
4780 files. Typically this is used in conjunction with language extensions which
4781 allow source files to explicitly declare the libraries they depend on, and have
4782 these automatically be transmitted to the linker via object files.
4784 These flags are encoded in the IR using metadata in the module flags section,
4785 using the ``Linker Options`` key. The merge behavior for this flag is required
4786 to be ``AppendUnique``, and the value for the key is expected to be a metadata
4787 node which should be a list of other metadata nodes, each of which should be a
4788 list of metadata strings defining linker options.
4790 For example, the following metadata section specifies two separate sets of
4791 linker options, presumably to link against ``libz`` and the ``Cocoa``
4794 !0 = !{ i32 6, !"Linker Options",
4797 !{ !"-framework", !"Cocoa" } } }
4798 !llvm.module.flags = !{ !0 }
4800 The metadata encoding as lists of lists of options, as opposed to a collapsed
4801 list of options, is chosen so that the IR encoding can use multiple option
4802 strings to specify e.g., a single library, while still having that specifier be
4803 preserved as an atomic element that can be recognized by a target specific
4804 assembly writer or object file emitter.
4806 Each individual option is required to be either a valid option for the target's
4807 linker, or an option that is reserved by the target specific assembly writer or
4808 object file emitter. No other aspect of these options is defined by the IR.
4810 C type width Module Flags Metadata
4811 ----------------------------------
4813 The ARM backend emits a section into each generated object file describing the
4814 options that it was compiled with (in a compiler-independent way) to prevent
4815 linking incompatible objects, and to allow automatic library selection. Some
4816 of these options are not visible at the IR level, namely wchar_t width and enum
4819 To pass this information to the backend, these options are encoded in module
4820 flags metadata, using the following key-value pairs:
4830 - * 0 --- sizeof(wchar_t) == 4
4831 * 1 --- sizeof(wchar_t) == 2
4834 - * 0 --- Enums are at least as large as an ``int``.
4835 * 1 --- Enums are stored in the smallest integer type which can
4836 represent all of its values.
4838 For example, the following metadata section specifies that the module was
4839 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
4840 enum is the smallest type which can represent all of its values::
4842 !llvm.module.flags = !{!0, !1}
4843 !0 = !{i32 1, !"short_wchar", i32 1}
4844 !1 = !{i32 1, !"short_enum", i32 0}
4846 .. _intrinsicglobalvariables:
4848 Intrinsic Global Variables
4849 ==========================
4851 LLVM has a number of "magic" global variables that contain data that
4852 affect code generation or other IR semantics. These are documented here.
4853 All globals of this sort should have a section specified as
4854 "``llvm.metadata``". This section and all globals that start with
4855 "``llvm.``" are reserved for use by LLVM.
4859 The '``llvm.used``' Global Variable
4860 -----------------------------------
4862 The ``@llvm.used`` global is an array which has
4863 :ref:`appending linkage <linkage_appending>`. This array contains a list of
4864 pointers to named global variables, functions and aliases which may optionally
4865 have a pointer cast formed of bitcast or getelementptr. For example, a legal
4868 .. code-block:: llvm
4873 @llvm.used = appending global [2 x i8*] [
4875 i8* bitcast (i32* @Y to i8*)
4876 ], section "llvm.metadata"
4878 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
4879 and linker are required to treat the symbol as if there is a reference to the
4880 symbol that it cannot see (which is why they have to be named). For example, if
4881 a variable has internal linkage and no references other than that from the
4882 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
4883 references from inline asms and other things the compiler cannot "see", and
4884 corresponds to "``attribute((used))``" in GNU C.
4886 On some targets, the code generator must emit a directive to the
4887 assembler or object file to prevent the assembler and linker from
4888 molesting the symbol.
4890 .. _gv_llvmcompilerused:
4892 The '``llvm.compiler.used``' Global Variable
4893 --------------------------------------------
4895 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
4896 directive, except that it only prevents the compiler from touching the
4897 symbol. On targets that support it, this allows an intelligent linker to
4898 optimize references to the symbol without being impeded as it would be
4901 This is a rare construct that should only be used in rare circumstances,
4902 and should not be exposed to source languages.
4904 .. _gv_llvmglobalctors:
4906 The '``llvm.global_ctors``' Global Variable
4907 -------------------------------------------
4909 .. code-block:: llvm
4911 %0 = type { i32, void ()*, i8* }
4912 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
4914 The ``@llvm.global_ctors`` array contains a list of constructor
4915 functions, priorities, and an optional associated global or function.
4916 The functions referenced by this array will be called in ascending order
4917 of priority (i.e. lowest first) when the module is loaded. The order of
4918 functions with the same priority is not defined.
4920 If the third field is present, non-null, and points to a global variable
4921 or function, the initializer function will only run if the associated
4922 data from the current module is not discarded.
4924 .. _llvmglobaldtors:
4926 The '``llvm.global_dtors``' Global Variable
4927 -------------------------------------------
4929 .. code-block:: llvm
4931 %0 = type { i32, void ()*, i8* }
4932 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
4934 The ``@llvm.global_dtors`` array contains a list of destructor
4935 functions, priorities, and an optional associated global or function.
4936 The functions referenced by this array will be called in descending
4937 order of priority (i.e. highest first) when the module is unloaded. The
4938 order of functions with the same priority is not defined.
4940 If the third field is present, non-null, and points to a global variable
4941 or function, the destructor function will only run if the associated
4942 data from the current module is not discarded.
4944 Instruction Reference
4945 =====================
4947 The LLVM instruction set consists of several different classifications
4948 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
4949 instructions <binaryops>`, :ref:`bitwise binary
4950 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
4951 :ref:`other instructions <otherops>`.
4955 Terminator Instructions
4956 -----------------------
4958 As mentioned :ref:`previously <functionstructure>`, every basic block in a
4959 program ends with a "Terminator" instruction, which indicates which
4960 block should be executed after the current block is finished. These
4961 terminator instructions typically yield a '``void``' value: they produce
4962 control flow, not values (the one exception being the
4963 ':ref:`invoke <i_invoke>`' instruction).
4965 The terminator instructions are: ':ref:`ret <i_ret>`',
4966 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
4967 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
4968 ':ref:`resume <i_resume>`', ':ref:`catchpad <i_catchpad>`',
4969 ':ref:`catchendpad <i_catchendpad>`',
4970 ':ref:`catchret <i_catchret>`',
4971 ':ref:`cleanupendpad <i_cleanupendpad>`',
4972 ':ref:`cleanupret <i_cleanupret>`',
4973 ':ref:`terminatepad <i_terminatepad>`',
4974 and ':ref:`unreachable <i_unreachable>`'.
4978 '``ret``' Instruction
4979 ^^^^^^^^^^^^^^^^^^^^^
4986 ret <type> <value> ; Return a value from a non-void function
4987 ret void ; Return from void function
4992 The '``ret``' instruction is used to return control flow (and optionally
4993 a value) from a function back to the caller.
4995 There are two forms of the '``ret``' instruction: one that returns a
4996 value and then causes control flow, and one that just causes control
5002 The '``ret``' instruction optionally accepts a single argument, the
5003 return value. The type of the return value must be a ':ref:`first
5004 class <t_firstclass>`' type.
5006 A function is not :ref:`well formed <wellformed>` if it it has a non-void
5007 return type and contains a '``ret``' instruction with no return value or
5008 a return value with a type that does not match its type, or if it has a
5009 void return type and contains a '``ret``' instruction with a return
5015 When the '``ret``' instruction is executed, control flow returns back to
5016 the calling function's context. If the caller is a
5017 ":ref:`call <i_call>`" instruction, execution continues at the
5018 instruction after the call. If the caller was an
5019 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
5020 beginning of the "normal" destination block. If the instruction returns
5021 a value, that value shall set the call or invoke instruction's return
5027 .. code-block:: llvm
5029 ret i32 5 ; Return an integer value of 5
5030 ret void ; Return from a void function
5031 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
5035 '``br``' Instruction
5036 ^^^^^^^^^^^^^^^^^^^^
5043 br i1 <cond>, label <iftrue>, label <iffalse>
5044 br label <dest> ; Unconditional branch
5049 The '``br``' instruction is used to cause control flow to transfer to a
5050 different basic block in the current function. There are two forms of
5051 this instruction, corresponding to a conditional branch and an
5052 unconditional branch.
5057 The conditional branch form of the '``br``' instruction takes a single
5058 '``i1``' value and two '``label``' values. The unconditional form of the
5059 '``br``' instruction takes a single '``label``' value as a target.
5064 Upon execution of a conditional '``br``' instruction, the '``i1``'
5065 argument is evaluated. If the value is ``true``, control flows to the
5066 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
5067 to the '``iffalse``' ``label`` argument.
5072 .. code-block:: llvm
5075 %cond = icmp eq i32 %a, %b
5076 br i1 %cond, label %IfEqual, label %IfUnequal
5084 '``switch``' Instruction
5085 ^^^^^^^^^^^^^^^^^^^^^^^^
5092 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
5097 The '``switch``' instruction is used to transfer control flow to one of
5098 several different places. It is a generalization of the '``br``'
5099 instruction, allowing a branch to occur to one of many possible
5105 The '``switch``' instruction uses three parameters: an integer
5106 comparison value '``value``', a default '``label``' destination, and an
5107 array of pairs of comparison value constants and '``label``'s. The table
5108 is not allowed to contain duplicate constant entries.
5113 The ``switch`` instruction specifies a table of values and destinations.
5114 When the '``switch``' instruction is executed, this table is searched
5115 for the given value. If the value is found, control flow is transferred
5116 to the corresponding destination; otherwise, control flow is transferred
5117 to the default destination.
5122 Depending on properties of the target machine and the particular
5123 ``switch`` instruction, this instruction may be code generated in
5124 different ways. For example, it could be generated as a series of
5125 chained conditional branches or with a lookup table.
5130 .. code-block:: llvm
5132 ; Emulate a conditional br instruction
5133 %Val = zext i1 %value to i32
5134 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
5136 ; Emulate an unconditional br instruction
5137 switch i32 0, label %dest [ ]
5139 ; Implement a jump table:
5140 switch i32 %val, label %otherwise [ i32 0, label %onzero
5142 i32 2, label %ontwo ]
5146 '``indirectbr``' Instruction
5147 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5154 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
5159 The '``indirectbr``' instruction implements an indirect branch to a
5160 label within the current function, whose address is specified by
5161 "``address``". Address must be derived from a
5162 :ref:`blockaddress <blockaddress>` constant.
5167 The '``address``' argument is the address of the label to jump to. The
5168 rest of the arguments indicate the full set of possible destinations
5169 that the address may point to. Blocks are allowed to occur multiple
5170 times in the destination list, though this isn't particularly useful.
5172 This destination list is required so that dataflow analysis has an
5173 accurate understanding of the CFG.
5178 Control transfers to the block specified in the address argument. All
5179 possible destination blocks must be listed in the label list, otherwise
5180 this instruction has undefined behavior. This implies that jumps to
5181 labels defined in other functions have undefined behavior as well.
5186 This is typically implemented with a jump through a register.
5191 .. code-block:: llvm
5193 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
5197 '``invoke``' Instruction
5198 ^^^^^^^^^^^^^^^^^^^^^^^^
5205 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
5206 [operand bundles] to label <normal label> unwind label <exception label>
5211 The '``invoke``' instruction causes control to transfer to a specified
5212 function, with the possibility of control flow transfer to either the
5213 '``normal``' label or the '``exception``' label. If the callee function
5214 returns with the "``ret``" instruction, control flow will return to the
5215 "normal" label. If the callee (or any indirect callees) returns via the
5216 ":ref:`resume <i_resume>`" instruction or other exception handling
5217 mechanism, control is interrupted and continued at the dynamically
5218 nearest "exception" label.
5220 The '``exception``' label is a `landing
5221 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
5222 '``exception``' label is required to have the
5223 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
5224 information about the behavior of the program after unwinding happens,
5225 as its first non-PHI instruction. The restrictions on the
5226 "``landingpad``" instruction's tightly couples it to the "``invoke``"
5227 instruction, so that the important information contained within the
5228 "``landingpad``" instruction can't be lost through normal code motion.
5233 This instruction requires several arguments:
5235 #. The optional "cconv" marker indicates which :ref:`calling
5236 convention <callingconv>` the call should use. If none is
5237 specified, the call defaults to using C calling conventions.
5238 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
5239 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
5241 #. '``ptr to function ty``': shall be the signature of the pointer to
5242 function value being invoked. In most cases, this is a direct
5243 function invocation, but indirect ``invoke``'s are just as possible,
5244 branching off an arbitrary pointer to function value.
5245 #. '``function ptr val``': An LLVM value containing a pointer to a
5246 function to be invoked.
5247 #. '``function args``': argument list whose types match the function
5248 signature argument types and parameter attributes. All arguments must
5249 be of :ref:`first class <t_firstclass>` type. If the function signature
5250 indicates the function accepts a variable number of arguments, the
5251 extra arguments can be specified.
5252 #. '``normal label``': the label reached when the called function
5253 executes a '``ret``' instruction.
5254 #. '``exception label``': the label reached when a callee returns via
5255 the :ref:`resume <i_resume>` instruction or other exception handling
5257 #. The optional :ref:`function attributes <fnattrs>` list. Only
5258 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
5259 attributes are valid here.
5260 #. The optional :ref:`operand bundles <opbundles>` list.
5265 This instruction is designed to operate as a standard '``call``'
5266 instruction in most regards. The primary difference is that it
5267 establishes an association with a label, which is used by the runtime
5268 library to unwind the stack.
5270 This instruction is used in languages with destructors to ensure that
5271 proper cleanup is performed in the case of either a ``longjmp`` or a
5272 thrown exception. Additionally, this is important for implementation of
5273 '``catch``' clauses in high-level languages that support them.
5275 For the purposes of the SSA form, the definition of the value returned
5276 by the '``invoke``' instruction is deemed to occur on the edge from the
5277 current block to the "normal" label. If the callee unwinds then no
5278 return value is available.
5283 .. code-block:: llvm
5285 %retval = invoke i32 @Test(i32 15) to label %Continue
5286 unwind label %TestCleanup ; i32:retval set
5287 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
5288 unwind label %TestCleanup ; i32:retval set
5292 '``resume``' Instruction
5293 ^^^^^^^^^^^^^^^^^^^^^^^^
5300 resume <type> <value>
5305 The '``resume``' instruction is a terminator instruction that has no
5311 The '``resume``' instruction requires one argument, which must have the
5312 same type as the result of any '``landingpad``' instruction in the same
5318 The '``resume``' instruction resumes propagation of an existing
5319 (in-flight) exception whose unwinding was interrupted with a
5320 :ref:`landingpad <i_landingpad>` instruction.
5325 .. code-block:: llvm
5327 resume { i8*, i32 } %exn
5331 '``catchpad``' Instruction
5332 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5339 <resultval> = catchpad [<args>*]
5340 to label <normal label> unwind label <exception label>
5345 The '``catchpad``' instruction is used by `LLVM's exception handling
5346 system <ExceptionHandling.html#overview>`_ to specify that a basic block
5347 is a catch block --- one where a personality routine attempts to transfer
5348 control to catch an exception.
5349 The ``args`` correspond to whatever information the personality
5350 routine requires to know if this is an appropriate place to catch the
5351 exception. Control is transfered to the ``exception`` label if the
5352 ``catchpad`` is not an appropriate handler for the in-flight exception.
5353 The ``normal`` label should contain the code found in the ``catch``
5354 portion of a ``try``/``catch`` sequence. The ``resultval`` has the type
5355 :ref:`token <t_token>` and is used to match the ``catchpad`` to
5356 corresponding :ref:`catchrets <i_catchret>`.
5361 The instruction takes a list of arbitrary values which are interpreted
5362 by the :ref:`personality function <personalityfn>`.
5364 The ``catchpad`` must be provided a ``normal`` label to transfer control
5365 to if the ``catchpad`` matches the exception and an ``exception``
5366 label to transfer control to if it doesn't.
5371 When the call stack is being unwound due to an exception being thrown,
5372 the exception is compared against the ``args``. If it doesn't match,
5373 then control is transfered to the ``exception`` basic block.
5374 As with calling conventions, how the personality function results are
5375 represented in LLVM IR is target specific.
5377 The ``catchpad`` instruction has several restrictions:
5379 - A catch block is a basic block which is the unwind destination of
5380 an exceptional instruction.
5381 - A catch block must have a '``catchpad``' instruction as its
5382 first non-PHI instruction.
5383 - A catch block's ``exception`` edge must refer to a catch block or a
5385 - There can be only one '``catchpad``' instruction within the
5387 - A basic block that is not a catch block may not include a
5388 '``catchpad``' instruction.
5389 - A catch block which has another catch block as a predecessor may not have
5390 any other predecessors.
5391 - It is undefined behavior for control to transfer from a ``catchpad`` to a
5392 ``ret`` without first executing a ``catchret`` that consumes the
5393 ``catchpad`` or unwinding through its ``catchendpad``.
5394 - It is undefined behavior for control to transfer from a ``catchpad`` to
5395 itself without first executing a ``catchret`` that consumes the
5396 ``catchpad`` or unwinding through its ``catchendpad``.
5401 .. code-block:: llvm
5403 ;; A catch block which can catch an integer.
5404 %tok = catchpad [i8** @_ZTIi]
5405 to label %int.handler unwind label %terminate
5409 '``catchendpad``' Instruction
5410 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5417 catchendpad unwind label <nextaction>
5418 catchendpad unwind to caller
5423 The '``catchendpad``' instruction is used by `LLVM's exception handling
5424 system <ExceptionHandling.html#overview>`_ to communicate to the
5425 :ref:`personality function <personalityfn>` which invokes are associated
5426 with a chain of :ref:`catchpad <i_catchpad>` instructions; propagating an
5427 exception out of a catch handler is represented by unwinding through its
5428 ``catchendpad``. Unwinding to the outer scope when a chain of catch handlers
5429 do not handle an exception is also represented by unwinding through their
5432 The ``nextaction`` label indicates where control should transfer to if
5433 none of the ``catchpad`` instructions are suitable for catching the
5434 in-flight exception.
5436 If a ``nextaction`` label is not present, the instruction unwinds out of
5437 its parent function. The
5438 :ref:`personality function <personalityfn>` will continue processing
5439 exception handling actions in the caller.
5444 The instruction optionally takes a label, ``nextaction``, indicating
5445 where control should transfer to if none of the preceding
5446 ``catchpad`` instructions are suitable for the in-flight exception.
5451 When the call stack is being unwound due to an exception being thrown
5452 and none of the constituent ``catchpad`` instructions match, then
5453 control is transfered to ``nextaction`` if it is present. If it is not
5454 present, control is transfered to the caller.
5456 The ``catchendpad`` instruction has several restrictions:
5458 - A catch-end block is a basic block which is the unwind destination of
5459 an exceptional instruction.
5460 - A catch-end block must have a '``catchendpad``' instruction as its
5461 first non-PHI instruction.
5462 - There can be only one '``catchendpad``' instruction within the
5464 - A basic block that is not a catch-end block may not include a
5465 '``catchendpad``' instruction.
5466 - Exactly one catch block may unwind to a ``catchendpad``.
5467 - It is undefined behavior to execute a ``catchendpad`` if none of the
5468 '``catchpad``'s chained to it have been executed.
5469 - It is undefined behavior to execute a ``catchendpad`` twice without an
5470 intervening execution of one or more of the '``catchpad``'s chained to it.
5471 - It is undefined behavior to execute a ``catchendpad`` if, after the most
5472 recent execution of the normal successor edge of any ``catchpad`` chained
5473 to it, some ``catchret`` consuming that ``catchpad`` has already been
5475 - It is undefined behavior to execute a ``catchendpad`` if, after the most
5476 recent execution of the normal successor edge of any ``catchpad`` chained
5477 to it, any other ``catchpad`` or ``cleanuppad`` has been executed but has
5478 not had a corresponding
5479 ``catchret``/``cleanupret``/``catchendpad``/``cleanupendpad`` executed.
5484 .. code-block:: llvm
5486 catchendpad unwind label %terminate
5487 catchendpad unwind to caller
5491 '``catchret``' Instruction
5492 ^^^^^^^^^^^^^^^^^^^^^^^^^^
5499 catchret <value> to label <normal>
5504 The '``catchret``' instruction is a terminator instruction that has a
5511 The first argument to a '``catchret``' indicates which ``catchpad`` it
5512 exits. It must be a :ref:`catchpad <i_catchpad>`.
5513 The second argument to a '``catchret``' specifies where control will
5519 The '``catchret``' instruction ends the existing (in-flight) exception
5520 whose unwinding was interrupted with a
5521 :ref:`catchpad <i_catchpad>` instruction.
5522 The :ref:`personality function <personalityfn>` gets a chance to execute
5523 arbitrary code to, for example, run a C++ destructor.
5524 Control then transfers to ``normal``.
5525 It may be passed an optional, personality specific, value.
5527 It is undefined behavior to execute a ``catchret`` whose ``catchpad`` has
5530 It is undefined behavior to execute a ``catchret`` if, after the most recent
5531 execution of its ``catchpad``, some ``catchret`` or ``catchendpad`` linked
5532 to the same ``catchpad`` has already been executed.
5534 It is undefined behavior to execute a ``catchret`` if, after the most recent
5535 execution of its ``catchpad``, any other ``catchpad`` or ``cleanuppad`` has
5536 been executed but has not had a corresponding
5537 ``catchret``/``cleanupret``/``catchendpad``/``cleanupendpad`` executed.
5542 .. code-block:: llvm
5544 catchret %catch label %continue
5546 .. _i_cleanupendpad:
5548 '``cleanupendpad``' Instruction
5549 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5556 cleanupendpad <value> unwind label <nextaction>
5557 cleanupendpad <value> unwind to caller
5562 The '``cleanupendpad``' instruction is used by `LLVM's exception handling
5563 system <ExceptionHandling.html#overview>`_ to communicate to the
5564 :ref:`personality function <personalityfn>` which invokes are associated
5565 with a :ref:`cleanuppad <i_cleanuppad>` instructions; propagating an exception
5566 out of a cleanup is represented by unwinding through its ``cleanupendpad``.
5568 The ``nextaction`` label indicates where control should unwind to next, in the
5569 event that a cleanup is exited by means of an(other) exception being raised.
5571 If a ``nextaction`` label is not present, the instruction unwinds out of
5572 its parent function. The
5573 :ref:`personality function <personalityfn>` will continue processing
5574 exception handling actions in the caller.
5579 The '``cleanupendpad``' instruction requires one argument, which indicates
5580 which ``cleanuppad`` it exits, and must be a :ref:`cleanuppad <i_cleanuppad>`.
5581 It also has an optional successor, ``nextaction``, indicating where control
5587 When and exception propagates to a ``cleanupendpad``, control is transfered to
5588 ``nextaction`` if it is present. If it is not present, control is transfered to
5591 The ``cleanupendpad`` instruction has several restrictions:
5593 - A cleanup-end block is a basic block which is the unwind destination of
5594 an exceptional instruction.
5595 - A cleanup-end block must have a '``cleanupendpad``' instruction as its
5596 first non-PHI instruction.
5597 - There can be only one '``cleanupendpad``' instruction within the
5599 - A basic block that is not a cleanup-end block may not include a
5600 '``cleanupendpad``' instruction.
5601 - It is undefined behavior to execute a ``cleanupendpad`` whose ``cleanuppad``
5602 has not been executed.
5603 - It is undefined behavior to execute a ``cleanupendpad`` if, after the most
5604 recent execution of its ``cleanuppad``, some ``cleanupret`` or ``cleanupendpad``
5605 consuming the same ``cleanuppad`` has already been executed.
5606 - It is undefined behavior to execute a ``cleanupendpad`` if, after the most
5607 recent execution of its ``cleanuppad``, any other ``cleanuppad`` or
5608 ``catchpad`` has been executed but has not had a corresponding
5609 ``cleanupret``/``catchret``/``cleanupendpad``/``catchendpad`` executed.
5614 .. code-block:: llvm
5616 cleanupendpad %cleanup unwind label %terminate
5617 cleanupendpad %cleanup unwind to caller
5621 '``cleanupret``' Instruction
5622 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5629 cleanupret <value> unwind label <continue>
5630 cleanupret <value> unwind to caller
5635 The '``cleanupret``' instruction is a terminator instruction that has
5636 an optional successor.
5642 The '``cleanupret``' instruction requires one argument, which indicates
5643 which ``cleanuppad`` it exits, and must be a :ref:`cleanuppad <i_cleanuppad>`.
5644 It also has an optional successor, ``continue``.
5649 The '``cleanupret``' instruction indicates to the
5650 :ref:`personality function <personalityfn>` that one
5651 :ref:`cleanuppad <i_cleanuppad>` it transferred control to has ended.
5652 It transfers control to ``continue`` or unwinds out of the function.
5654 It is undefined behavior to execute a ``cleanupret`` whose ``cleanuppad`` has
5657 It is undefined behavior to execute a ``cleanupret`` if, after the most recent
5658 execution of its ``cleanuppad``, some ``cleanupret`` or ``cleanupendpad``
5659 consuming the same ``cleanuppad`` has already been executed.
5661 It is undefined behavior to execute a ``cleanupret`` if, after the most recent
5662 execution of its ``cleanuppad``, any other ``cleanuppad`` or ``catchpad`` has
5663 been executed but has not had a corresponding
5664 ``cleanupret``/``catchret``/``cleanupendpad``/``catchendpad`` executed.
5669 .. code-block:: llvm
5671 cleanupret %cleanup unwind to caller
5672 cleanupret %cleanup unwind label %continue
5676 '``terminatepad``' Instruction
5677 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5684 terminatepad [<args>*] unwind label <exception label>
5685 terminatepad [<args>*] unwind to caller
5690 The '``terminatepad``' instruction is used by `LLVM's exception handling
5691 system <ExceptionHandling.html#overview>`_ to specify that a basic block
5692 is a terminate block --- one where a personality routine may decide to
5693 terminate the program.
5694 The ``args`` correspond to whatever information the personality
5695 routine requires to know if this is an appropriate place to terminate the
5696 program. Control is transferred to the ``exception`` label if the
5697 personality routine decides not to terminate the program for the
5698 in-flight exception.
5703 The instruction takes a list of arbitrary values which are interpreted
5704 by the :ref:`personality function <personalityfn>`.
5706 The ``terminatepad`` may be given an ``exception`` label to
5707 transfer control to if the in-flight exception matches the ``args``.
5712 When the call stack is being unwound due to an exception being thrown,
5713 the exception is compared against the ``args``. If it matches,
5714 then control is transfered to the ``exception`` basic block. Otherwise,
5715 the program is terminated via personality-specific means. Typically,
5716 the first argument to ``terminatepad`` specifies what function the
5717 personality should defer to in order to terminate the program.
5719 The ``terminatepad`` instruction has several restrictions:
5721 - A terminate block is a basic block which is the unwind destination of
5722 an exceptional instruction.
5723 - A terminate block must have a '``terminatepad``' instruction as its
5724 first non-PHI instruction.
5725 - There can be only one '``terminatepad``' instruction within the
5727 - A basic block that is not a terminate block may not include a
5728 '``terminatepad``' instruction.
5733 .. code-block:: llvm
5735 ;; A terminate block which only permits integers.
5736 terminatepad [i8** @_ZTIi] unwind label %continue
5740 '``unreachable``' Instruction
5741 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5753 The '``unreachable``' instruction has no defined semantics. This
5754 instruction is used to inform the optimizer that a particular portion of
5755 the code is not reachable. This can be used to indicate that the code
5756 after a no-return function cannot be reached, and other facts.
5761 The '``unreachable``' instruction has no defined semantics.
5768 Binary operators are used to do most of the computation in a program.
5769 They require two operands of the same type, execute an operation on
5770 them, and produce a single value. The operands might represent multiple
5771 data, as is the case with the :ref:`vector <t_vector>` data type. The
5772 result value has the same type as its operands.
5774 There are several different binary operators:
5778 '``add``' Instruction
5779 ^^^^^^^^^^^^^^^^^^^^^
5786 <result> = add <ty> <op1>, <op2> ; yields ty:result
5787 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
5788 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
5789 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
5794 The '``add``' instruction returns the sum of its two operands.
5799 The two arguments to the '``add``' instruction must be
5800 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5801 arguments must have identical types.
5806 The value produced is the integer sum of the two operands.
5808 If the sum has unsigned overflow, the result returned is the
5809 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
5812 Because LLVM integers use a two's complement representation, this
5813 instruction is appropriate for both signed and unsigned integers.
5815 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
5816 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
5817 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
5818 unsigned and/or signed overflow, respectively, occurs.
5823 .. code-block:: llvm
5825 <result> = add i32 4, %var ; yields i32:result = 4 + %var
5829 '``fadd``' Instruction
5830 ^^^^^^^^^^^^^^^^^^^^^^
5837 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5842 The '``fadd``' instruction returns the sum of its two operands.
5847 The two arguments to the '``fadd``' instruction must be :ref:`floating
5848 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5849 Both arguments must have identical types.
5854 The value produced is the floating point sum of the two operands. This
5855 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
5856 which are optimization hints to enable otherwise unsafe floating point
5862 .. code-block:: llvm
5864 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
5866 '``sub``' Instruction
5867 ^^^^^^^^^^^^^^^^^^^^^
5874 <result> = sub <ty> <op1>, <op2> ; yields ty:result
5875 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
5876 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
5877 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
5882 The '``sub``' instruction returns the difference of its two operands.
5884 Note that the '``sub``' instruction is used to represent the '``neg``'
5885 instruction present in most other intermediate representations.
5890 The two arguments to the '``sub``' instruction must be
5891 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5892 arguments must have identical types.
5897 The value produced is the integer difference of the two operands.
5899 If the difference has unsigned overflow, the result returned is the
5900 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
5903 Because LLVM integers use a two's complement representation, this
5904 instruction is appropriate for both signed and unsigned integers.
5906 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
5907 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
5908 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
5909 unsigned and/or signed overflow, respectively, occurs.
5914 .. code-block:: llvm
5916 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
5917 <result> = sub i32 0, %val ; yields i32:result = -%var
5921 '``fsub``' Instruction
5922 ^^^^^^^^^^^^^^^^^^^^^^
5929 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5934 The '``fsub``' instruction returns the difference of its two operands.
5936 Note that the '``fsub``' instruction is used to represent the '``fneg``'
5937 instruction present in most other intermediate representations.
5942 The two arguments to the '``fsub``' instruction must be :ref:`floating
5943 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5944 Both arguments must have identical types.
5949 The value produced is the floating point difference of the two operands.
5950 This instruction can also take any number of :ref:`fast-math
5951 flags <fastmath>`, which are optimization hints to enable otherwise
5952 unsafe floating point optimizations:
5957 .. code-block:: llvm
5959 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
5960 <result> = fsub float -0.0, %val ; yields float:result = -%var
5962 '``mul``' Instruction
5963 ^^^^^^^^^^^^^^^^^^^^^
5970 <result> = mul <ty> <op1>, <op2> ; yields ty:result
5971 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
5972 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
5973 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
5978 The '``mul``' instruction returns the product of its two operands.
5983 The two arguments to the '``mul``' instruction must be
5984 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5985 arguments must have identical types.
5990 The value produced is the integer product of the two operands.
5992 If the result of the multiplication has unsigned overflow, the result
5993 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
5994 bit width of the result.
5996 Because LLVM integers use a two's complement representation, and the
5997 result is the same width as the operands, this instruction returns the
5998 correct result for both signed and unsigned integers. If a full product
5999 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
6000 sign-extended or zero-extended as appropriate to the width of the full
6003 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
6004 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
6005 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
6006 unsigned and/or signed overflow, respectively, occurs.
6011 .. code-block:: llvm
6013 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
6017 '``fmul``' Instruction
6018 ^^^^^^^^^^^^^^^^^^^^^^
6025 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6030 The '``fmul``' instruction returns the product of its two operands.
6035 The two arguments to the '``fmul``' instruction must be :ref:`floating
6036 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6037 Both arguments must have identical types.
6042 The value produced is the floating point product of the two operands.
6043 This instruction can also take any number of :ref:`fast-math
6044 flags <fastmath>`, which are optimization hints to enable otherwise
6045 unsafe floating point optimizations:
6050 .. code-block:: llvm
6052 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
6054 '``udiv``' Instruction
6055 ^^^^^^^^^^^^^^^^^^^^^^
6062 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
6063 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
6068 The '``udiv``' instruction returns the quotient of its two operands.
6073 The two arguments to the '``udiv``' instruction must be
6074 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6075 arguments must have identical types.
6080 The value produced is the unsigned integer quotient of the two operands.
6082 Note that unsigned integer division and signed integer division are
6083 distinct operations; for signed integer division, use '``sdiv``'.
6085 Division by zero leads to undefined behavior.
6087 If the ``exact`` keyword is present, the result value of the ``udiv`` is
6088 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
6089 such, "((a udiv exact b) mul b) == a").
6094 .. code-block:: llvm
6096 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
6098 '``sdiv``' Instruction
6099 ^^^^^^^^^^^^^^^^^^^^^^
6106 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
6107 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
6112 The '``sdiv``' instruction returns the quotient of its two operands.
6117 The two arguments to the '``sdiv``' instruction must be
6118 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6119 arguments must have identical types.
6124 The value produced is the signed integer quotient of the two operands
6125 rounded towards zero.
6127 Note that signed integer division and unsigned integer division are
6128 distinct operations; for unsigned integer division, use '``udiv``'.
6130 Division by zero leads to undefined behavior. Overflow also leads to
6131 undefined behavior; this is a rare case, but can occur, for example, by
6132 doing a 32-bit division of -2147483648 by -1.
6134 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
6135 a :ref:`poison value <poisonvalues>` if the result would be rounded.
6140 .. code-block:: llvm
6142 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
6146 '``fdiv``' Instruction
6147 ^^^^^^^^^^^^^^^^^^^^^^
6154 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6159 The '``fdiv``' instruction returns the quotient of its two operands.
6164 The two arguments to the '``fdiv``' instruction must be :ref:`floating
6165 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6166 Both arguments must have identical types.
6171 The value produced is the floating point quotient of the two operands.
6172 This instruction can also take any number of :ref:`fast-math
6173 flags <fastmath>`, which are optimization hints to enable otherwise
6174 unsafe floating point optimizations:
6179 .. code-block:: llvm
6181 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
6183 '``urem``' Instruction
6184 ^^^^^^^^^^^^^^^^^^^^^^
6191 <result> = urem <ty> <op1>, <op2> ; yields ty:result
6196 The '``urem``' instruction returns the remainder from the unsigned
6197 division of its two arguments.
6202 The two arguments to the '``urem``' instruction must be
6203 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6204 arguments must have identical types.
6209 This instruction returns the unsigned integer *remainder* of a division.
6210 This instruction always performs an unsigned division to get the
6213 Note that unsigned integer remainder and signed integer remainder are
6214 distinct operations; for signed integer remainder, use '``srem``'.
6216 Taking the remainder of a division by zero leads to undefined behavior.
6221 .. code-block:: llvm
6223 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
6225 '``srem``' Instruction
6226 ^^^^^^^^^^^^^^^^^^^^^^
6233 <result> = srem <ty> <op1>, <op2> ; yields ty:result
6238 The '``srem``' instruction returns the remainder from the signed
6239 division of its two operands. This instruction can also take
6240 :ref:`vector <t_vector>` versions of the values in which case the elements
6246 The two arguments to the '``srem``' instruction must be
6247 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6248 arguments must have identical types.
6253 This instruction returns the *remainder* of a division (where the result
6254 is either zero or has the same sign as the dividend, ``op1``), not the
6255 *modulo* operator (where the result is either zero or has the same sign
6256 as the divisor, ``op2``) of a value. For more information about the
6257 difference, see `The Math
6258 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
6259 table of how this is implemented in various languages, please see
6261 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
6263 Note that signed integer remainder and unsigned integer remainder are
6264 distinct operations; for unsigned integer remainder, use '``urem``'.
6266 Taking the remainder of a division by zero leads to undefined behavior.
6267 Overflow also leads to undefined behavior; this is a rare case, but can
6268 occur, for example, by taking the remainder of a 32-bit division of
6269 -2147483648 by -1. (The remainder doesn't actually overflow, but this
6270 rule lets srem be implemented using instructions that return both the
6271 result of the division and the remainder.)
6276 .. code-block:: llvm
6278 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
6282 '``frem``' Instruction
6283 ^^^^^^^^^^^^^^^^^^^^^^
6290 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6295 The '``frem``' instruction returns the remainder from the division of
6301 The two arguments to the '``frem``' instruction must be :ref:`floating
6302 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6303 Both arguments must have identical types.
6308 This instruction returns the *remainder* of a division. The remainder
6309 has the same sign as the dividend. This instruction can also take any
6310 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
6311 to enable otherwise unsafe floating point optimizations:
6316 .. code-block:: llvm
6318 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
6322 Bitwise Binary Operations
6323 -------------------------
6325 Bitwise binary operators are used to do various forms of bit-twiddling
6326 in a program. They are generally very efficient instructions and can
6327 commonly be strength reduced from other instructions. They require two
6328 operands of the same type, execute an operation on them, and produce a
6329 single value. The resulting value is the same type as its operands.
6331 '``shl``' Instruction
6332 ^^^^^^^^^^^^^^^^^^^^^
6339 <result> = shl <ty> <op1>, <op2> ; yields ty:result
6340 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
6341 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
6342 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
6347 The '``shl``' instruction returns the first operand shifted to the left
6348 a specified number of bits.
6353 Both arguments to the '``shl``' instruction must be the same
6354 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6355 '``op2``' is treated as an unsigned value.
6360 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
6361 where ``n`` is the width of the result. If ``op2`` is (statically or
6362 dynamically) equal to or larger than the number of bits in
6363 ``op1``, the result is undefined. If the arguments are vectors, each
6364 vector element of ``op1`` is shifted by the corresponding shift amount
6367 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
6368 value <poisonvalues>` if it shifts out any non-zero bits. If the
6369 ``nsw`` keyword is present, then the shift produces a :ref:`poison
6370 value <poisonvalues>` if it shifts out any bits that disagree with the
6371 resultant sign bit. As such, NUW/NSW have the same semantics as they
6372 would if the shift were expressed as a mul instruction with the same
6373 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
6378 .. code-block:: llvm
6380 <result> = shl i32 4, %var ; yields i32: 4 << %var
6381 <result> = shl i32 4, 2 ; yields i32: 16
6382 <result> = shl i32 1, 10 ; yields i32: 1024
6383 <result> = shl i32 1, 32 ; undefined
6384 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
6386 '``lshr``' Instruction
6387 ^^^^^^^^^^^^^^^^^^^^^^
6394 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
6395 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
6400 The '``lshr``' instruction (logical shift right) returns the first
6401 operand shifted to the right a specified number of bits with zero fill.
6406 Both arguments to the '``lshr``' instruction must be the same
6407 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6408 '``op2``' is treated as an unsigned value.
6413 This instruction always performs a logical shift right operation. The
6414 most significant bits of the result will be filled with zero bits after
6415 the shift. If ``op2`` is (statically or dynamically) equal to or larger
6416 than the number of bits in ``op1``, the result is undefined. If the
6417 arguments are vectors, each vector element of ``op1`` is shifted by the
6418 corresponding shift amount in ``op2``.
6420 If the ``exact`` keyword is present, the result value of the ``lshr`` is
6421 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
6427 .. code-block:: llvm
6429 <result> = lshr i32 4, 1 ; yields i32:result = 2
6430 <result> = lshr i32 4, 2 ; yields i32:result = 1
6431 <result> = lshr i8 4, 3 ; yields i8:result = 0
6432 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
6433 <result> = lshr i32 1, 32 ; undefined
6434 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
6436 '``ashr``' Instruction
6437 ^^^^^^^^^^^^^^^^^^^^^^
6444 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
6445 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
6450 The '``ashr``' instruction (arithmetic shift right) returns the first
6451 operand shifted to the right a specified number of bits with sign
6457 Both arguments to the '``ashr``' instruction must be the same
6458 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6459 '``op2``' is treated as an unsigned value.
6464 This instruction always performs an arithmetic shift right operation,
6465 The most significant bits of the result will be filled with the sign bit
6466 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
6467 than the number of bits in ``op1``, the result is undefined. If the
6468 arguments are vectors, each vector element of ``op1`` is shifted by the
6469 corresponding shift amount in ``op2``.
6471 If the ``exact`` keyword is present, the result value of the ``ashr`` is
6472 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
6478 .. code-block:: llvm
6480 <result> = ashr i32 4, 1 ; yields i32:result = 2
6481 <result> = ashr i32 4, 2 ; yields i32:result = 1
6482 <result> = ashr i8 4, 3 ; yields i8:result = 0
6483 <result> = ashr i8 -2, 1 ; yields i8:result = -1
6484 <result> = ashr i32 1, 32 ; undefined
6485 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
6487 '``and``' Instruction
6488 ^^^^^^^^^^^^^^^^^^^^^
6495 <result> = and <ty> <op1>, <op2> ; yields ty:result
6500 The '``and``' instruction returns the bitwise logical and of its two
6506 The two arguments to the '``and``' instruction must be
6507 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6508 arguments must have identical types.
6513 The truth table used for the '``and``' instruction is:
6530 .. code-block:: llvm
6532 <result> = and i32 4, %var ; yields i32:result = 4 & %var
6533 <result> = and i32 15, 40 ; yields i32:result = 8
6534 <result> = and i32 4, 8 ; yields i32:result = 0
6536 '``or``' Instruction
6537 ^^^^^^^^^^^^^^^^^^^^
6544 <result> = or <ty> <op1>, <op2> ; yields ty:result
6549 The '``or``' instruction returns the bitwise logical inclusive or of its
6555 The two arguments to the '``or``' instruction must be
6556 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6557 arguments must have identical types.
6562 The truth table used for the '``or``' instruction is:
6581 <result> = or i32 4, %var ; yields i32:result = 4 | %var
6582 <result> = or i32 15, 40 ; yields i32:result = 47
6583 <result> = or i32 4, 8 ; yields i32:result = 12
6585 '``xor``' Instruction
6586 ^^^^^^^^^^^^^^^^^^^^^
6593 <result> = xor <ty> <op1>, <op2> ; yields ty:result
6598 The '``xor``' instruction returns the bitwise logical exclusive or of
6599 its two operands. The ``xor`` is used to implement the "one's
6600 complement" operation, which is the "~" operator in C.
6605 The two arguments to the '``xor``' instruction must be
6606 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6607 arguments must have identical types.
6612 The truth table used for the '``xor``' instruction is:
6629 .. code-block:: llvm
6631 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
6632 <result> = xor i32 15, 40 ; yields i32:result = 39
6633 <result> = xor i32 4, 8 ; yields i32:result = 12
6634 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
6639 LLVM supports several instructions to represent vector operations in a
6640 target-independent manner. These instructions cover the element-access
6641 and vector-specific operations needed to process vectors effectively.
6642 While LLVM does directly support these vector operations, many
6643 sophisticated algorithms will want to use target-specific intrinsics to
6644 take full advantage of a specific target.
6646 .. _i_extractelement:
6648 '``extractelement``' Instruction
6649 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6656 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
6661 The '``extractelement``' instruction extracts a single scalar element
6662 from a vector at a specified index.
6667 The first operand of an '``extractelement``' instruction is a value of
6668 :ref:`vector <t_vector>` type. The second operand is an index indicating
6669 the position from which to extract the element. The index may be a
6670 variable of any integer type.
6675 The result is a scalar of the same type as the element type of ``val``.
6676 Its value is the value at position ``idx`` of ``val``. If ``idx``
6677 exceeds the length of ``val``, the results are undefined.
6682 .. code-block:: llvm
6684 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
6686 .. _i_insertelement:
6688 '``insertelement``' Instruction
6689 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6696 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
6701 The '``insertelement``' instruction inserts a scalar element into a
6702 vector at a specified index.
6707 The first operand of an '``insertelement``' instruction is a value of
6708 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
6709 type must equal the element type of the first operand. The third operand
6710 is an index indicating the position at which to insert the value. The
6711 index may be a variable of any integer type.
6716 The result is a vector of the same type as ``val``. Its element values
6717 are those of ``val`` except at position ``idx``, where it gets the value
6718 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
6724 .. code-block:: llvm
6726 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
6728 .. _i_shufflevector:
6730 '``shufflevector``' Instruction
6731 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6738 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
6743 The '``shufflevector``' instruction constructs a permutation of elements
6744 from two input vectors, returning a vector with the same element type as
6745 the input and length that is the same as the shuffle mask.
6750 The first two operands of a '``shufflevector``' instruction are vectors
6751 with the same type. The third argument is a shuffle mask whose element
6752 type is always 'i32'. The result of the instruction is a vector whose
6753 length is the same as the shuffle mask and whose element type is the
6754 same as the element type of the first two operands.
6756 The shuffle mask operand is required to be a constant vector with either
6757 constant integer or undef values.
6762 The elements of the two input vectors are numbered from left to right
6763 across both of the vectors. The shuffle mask operand specifies, for each
6764 element of the result vector, which element of the two input vectors the
6765 result element gets. The element selector may be undef (meaning "don't
6766 care") and the second operand may be undef if performing a shuffle from
6772 .. code-block:: llvm
6774 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
6775 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
6776 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
6777 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
6778 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
6779 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
6780 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
6781 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
6783 Aggregate Operations
6784 --------------------
6786 LLVM supports several instructions for working with
6787 :ref:`aggregate <t_aggregate>` values.
6791 '``extractvalue``' Instruction
6792 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6799 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
6804 The '``extractvalue``' instruction extracts the value of a member field
6805 from an :ref:`aggregate <t_aggregate>` value.
6810 The first operand of an '``extractvalue``' instruction is a value of
6811 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The other operands are
6812 constant indices to specify which value to extract in a similar manner
6813 as indices in a '``getelementptr``' instruction.
6815 The major differences to ``getelementptr`` indexing are:
6817 - Since the value being indexed is not a pointer, the first index is
6818 omitted and assumed to be zero.
6819 - At least one index must be specified.
6820 - Not only struct indices but also array indices must be in bounds.
6825 The result is the value at the position in the aggregate specified by
6831 .. code-block:: llvm
6833 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
6837 '``insertvalue``' Instruction
6838 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6845 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
6850 The '``insertvalue``' instruction inserts a value into a member field in
6851 an :ref:`aggregate <t_aggregate>` value.
6856 The first operand of an '``insertvalue``' instruction is a value of
6857 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
6858 a first-class value to insert. The following operands are constant
6859 indices indicating the position at which to insert the value in a
6860 similar manner as indices in a '``extractvalue``' instruction. The value
6861 to insert must have the same type as the value identified by the
6867 The result is an aggregate of the same type as ``val``. Its value is
6868 that of ``val`` except that the value at the position specified by the
6869 indices is that of ``elt``.
6874 .. code-block:: llvm
6876 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
6877 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
6878 %agg3 = insertvalue {i32, {float}} undef, float %val, 1, 0 ; yields {i32 undef, {float %val}}
6882 Memory Access and Addressing Operations
6883 ---------------------------------------
6885 A key design point of an SSA-based representation is how it represents
6886 memory. In LLVM, no memory locations are in SSA form, which makes things
6887 very simple. This section describes how to read, write, and allocate
6892 '``alloca``' Instruction
6893 ^^^^^^^^^^^^^^^^^^^^^^^^
6900 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields type*:result
6905 The '``alloca``' instruction allocates memory on the stack frame of the
6906 currently executing function, to be automatically released when this
6907 function returns to its caller. The object is always allocated in the
6908 generic address space (address space zero).
6913 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
6914 bytes of memory on the runtime stack, returning a pointer of the
6915 appropriate type to the program. If "NumElements" is specified, it is
6916 the number of elements allocated, otherwise "NumElements" is defaulted
6917 to be one. If a constant alignment is specified, the value result of the
6918 allocation is guaranteed to be aligned to at least that boundary. The
6919 alignment may not be greater than ``1 << 29``. If not specified, or if
6920 zero, the target can choose to align the allocation on any convenient
6921 boundary compatible with the type.
6923 '``type``' may be any sized type.
6928 Memory is allocated; a pointer is returned. The operation is undefined
6929 if there is insufficient stack space for the allocation. '``alloca``'d
6930 memory is automatically released when the function returns. The
6931 '``alloca``' instruction is commonly used to represent automatic
6932 variables that must have an address available. When the function returns
6933 (either with the ``ret`` or ``resume`` instructions), the memory is
6934 reclaimed. Allocating zero bytes is legal, but the result is undefined.
6935 The order in which memory is allocated (ie., which way the stack grows)
6941 .. code-block:: llvm
6943 %ptr = alloca i32 ; yields i32*:ptr
6944 %ptr = alloca i32, i32 4 ; yields i32*:ptr
6945 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
6946 %ptr = alloca i32, align 1024 ; yields i32*:ptr
6950 '``load``' Instruction
6951 ^^^^^^^^^^^^^^^^^^^^^^
6958 <result> = load [volatile] <ty>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>][, !invariant.group !<index>][, !nonnull !<index>][, !dereferenceable !<deref_bytes_node>][, !dereferenceable_or_null !<deref_bytes_node>][, !align !<align_node>]
6959 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment> [, !invariant.group !<index>]
6960 !<index> = !{ i32 1 }
6961 !<deref_bytes_node> = !{i64 <dereferenceable_bytes>}
6962 !<align_node> = !{ i64 <value_alignment> }
6967 The '``load``' instruction is used to read from memory.
6972 The argument to the ``load`` instruction specifies the memory address
6973 from which to load. The type specified must be a :ref:`first
6974 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
6975 then the optimizer is not allowed to modify the number or order of
6976 execution of this ``load`` with other :ref:`volatile
6977 operations <volatile>`.
6979 If the ``load`` is marked as ``atomic``, it takes an extra
6980 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
6981 ``release`` and ``acq_rel`` orderings are not valid on ``load``
6982 instructions. Atomic loads produce :ref:`defined <memmodel>` results
6983 when they may see multiple atomic stores. The type of the pointee must
6984 be an integer type whose bit width is a power of two greater than or
6985 equal to eight and less than or equal to a target-specific size limit.
6986 ``align`` must be explicitly specified on atomic loads, and the load has
6987 undefined behavior if the alignment is not set to a value which is at
6988 least the size in bytes of the pointee. ``!nontemporal`` does not have
6989 any defined semantics for atomic loads.
6991 The optional constant ``align`` argument specifies the alignment of the
6992 operation (that is, the alignment of the memory address). A value of 0
6993 or an omitted ``align`` argument means that the operation has the ABI
6994 alignment for the target. It is the responsibility of the code emitter
6995 to ensure that the alignment information is correct. Overestimating the
6996 alignment results in undefined behavior. Underestimating the alignment
6997 may produce less efficient code. An alignment of 1 is always safe. The
6998 maximum possible alignment is ``1 << 29``.
7000 The optional ``!nontemporal`` metadata must reference a single
7001 metadata name ``<index>`` corresponding to a metadata node with one
7002 ``i32`` entry of value 1. The existence of the ``!nontemporal``
7003 metadata on the instruction tells the optimizer and code generator
7004 that this load is not expected to be reused in the cache. The code
7005 generator may select special instructions to save cache bandwidth, such
7006 as the ``MOVNT`` instruction on x86.
7008 The optional ``!invariant.load`` metadata must reference a single
7009 metadata name ``<index>`` corresponding to a metadata node with no
7010 entries. The existence of the ``!invariant.load`` metadata on the
7011 instruction tells the optimizer and code generator that the address
7012 operand to this load points to memory which can be assumed unchanged.
7013 Being invariant does not imply that a location is dereferenceable,
7014 but it does imply that once the location is known dereferenceable
7015 its value is henceforth unchanging.
7017 The optional ``!invariant.group`` metadata must reference a single metadata name
7018 ``<index>`` corresponding to a metadata node. See ``invariant.group`` metadata.
7020 The optional ``!nonnull`` metadata must reference a single
7021 metadata name ``<index>`` corresponding to a metadata node with no
7022 entries. The existence of the ``!nonnull`` metadata on the
7023 instruction tells the optimizer that the value loaded is known to
7024 never be null. This is analogous to the ``nonnull`` attribute
7025 on parameters and return values. This metadata can only be applied
7026 to loads of a pointer type.
7028 The optional ``!dereferenceable`` metadata must reference a single metadata
7029 name ``<deref_bytes_node>`` corresponding to a metadata node with one ``i64``
7030 entry. The existence of the ``!dereferenceable`` metadata on the instruction
7031 tells the optimizer that the value loaded is known to be dereferenceable.
7032 The number of bytes known to be dereferenceable is specified by the integer
7033 value in the metadata node. This is analogous to the ''dereferenceable''
7034 attribute on parameters and return values. This metadata can only be applied
7035 to loads of a pointer type.
7037 The optional ``!dereferenceable_or_null`` metadata must reference a single
7038 metadata name ``<deref_bytes_node>`` corresponding to a metadata node with one
7039 ``i64`` entry. The existence of the ``!dereferenceable_or_null`` metadata on the
7040 instruction tells the optimizer that the value loaded is known to be either
7041 dereferenceable or null.
7042 The number of bytes known to be dereferenceable is specified by the integer
7043 value in the metadata node. This is analogous to the ''dereferenceable_or_null''
7044 attribute on parameters and return values. This metadata can only be applied
7045 to loads of a pointer type.
7047 The optional ``!align`` metadata must reference a single metadata name
7048 ``<align_node>`` corresponding to a metadata node with one ``i64`` entry.
7049 The existence of the ``!align`` metadata on the instruction tells the
7050 optimizer that the value loaded is known to be aligned to a boundary specified
7051 by the integer value in the metadata node. The alignment must be a power of 2.
7052 This is analogous to the ''align'' attribute on parameters and return values.
7053 This metadata can only be applied to loads of a pointer type.
7058 The location of memory pointed to is loaded. If the value being loaded
7059 is of scalar type then the number of bytes read does not exceed the
7060 minimum number of bytes needed to hold all bits of the type. For
7061 example, loading an ``i24`` reads at most three bytes. When loading a
7062 value of a type like ``i20`` with a size that is not an integral number
7063 of bytes, the result is undefined if the value was not originally
7064 written using a store of the same type.
7069 .. code-block:: llvm
7071 %ptr = alloca i32 ; yields i32*:ptr
7072 store i32 3, i32* %ptr ; yields void
7073 %val = load i32, i32* %ptr ; yields i32:val = i32 3
7077 '``store``' Instruction
7078 ^^^^^^^^^^^^^^^^^^^^^^^
7085 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.group !<index>] ; yields void
7086 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> [, !invariant.group !<index>] ; yields void
7091 The '``store``' instruction is used to write to memory.
7096 There are two arguments to the ``store`` instruction: a value to store
7097 and an address at which to store it. The type of the ``<pointer>``
7098 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
7099 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
7100 then the optimizer is not allowed to modify the number or order of
7101 execution of this ``store`` with other :ref:`volatile
7102 operations <volatile>`.
7104 If the ``store`` is marked as ``atomic``, it takes an extra
7105 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
7106 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
7107 instructions. Atomic loads produce :ref:`defined <memmodel>` results
7108 when they may see multiple atomic stores. The type of the pointee must
7109 be an integer type whose bit width is a power of two greater than or
7110 equal to eight and less than or equal to a target-specific size limit.
7111 ``align`` must be explicitly specified on atomic stores, and the store
7112 has undefined behavior if the alignment is not set to a value which is
7113 at least the size in bytes of the pointee. ``!nontemporal`` does not
7114 have any defined semantics for atomic stores.
7116 The optional constant ``align`` argument specifies the alignment of the
7117 operation (that is, the alignment of the memory address). A value of 0
7118 or an omitted ``align`` argument means that the operation has the ABI
7119 alignment for the target. It is the responsibility of the code emitter
7120 to ensure that the alignment information is correct. Overestimating the
7121 alignment results in undefined behavior. Underestimating the
7122 alignment may produce less efficient code. An alignment of 1 is always
7123 safe. The maximum possible alignment is ``1 << 29``.
7125 The optional ``!nontemporal`` metadata must reference a single metadata
7126 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
7127 value 1. The existence of the ``!nontemporal`` metadata on the instruction
7128 tells the optimizer and code generator that this load is not expected to
7129 be reused in the cache. The code generator may select special
7130 instructions to save cache bandwidth, such as the MOVNT instruction on
7133 The optional ``!invariant.group`` metadata must reference a
7134 single metadata name ``<index>``. See ``invariant.group`` metadata.
7139 The contents of memory are updated to contain ``<value>`` at the
7140 location specified by the ``<pointer>`` operand. If ``<value>`` is
7141 of scalar type then the number of bytes written does not exceed the
7142 minimum number of bytes needed to hold all bits of the type. For
7143 example, storing an ``i24`` writes at most three bytes. When writing a
7144 value of a type like ``i20`` with a size that is not an integral number
7145 of bytes, it is unspecified what happens to the extra bits that do not
7146 belong to the type, but they will typically be overwritten.
7151 .. code-block:: llvm
7153 %ptr = alloca i32 ; yields i32*:ptr
7154 store i32 3, i32* %ptr ; yields void
7155 %val = load i32, i32* %ptr ; yields i32:val = i32 3
7159 '``fence``' Instruction
7160 ^^^^^^^^^^^^^^^^^^^^^^^
7167 fence [singlethread] <ordering> ; yields void
7172 The '``fence``' instruction is used to introduce happens-before edges
7178 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
7179 defines what *synchronizes-with* edges they add. They can only be given
7180 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
7185 A fence A which has (at least) ``release`` ordering semantics
7186 *synchronizes with* a fence B with (at least) ``acquire`` ordering
7187 semantics if and only if there exist atomic operations X and Y, both
7188 operating on some atomic object M, such that A is sequenced before X, X
7189 modifies M (either directly or through some side effect of a sequence
7190 headed by X), Y is sequenced before B, and Y observes M. This provides a
7191 *happens-before* dependency between A and B. Rather than an explicit
7192 ``fence``, one (but not both) of the atomic operations X or Y might
7193 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
7194 still *synchronize-with* the explicit ``fence`` and establish the
7195 *happens-before* edge.
7197 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
7198 ``acquire`` and ``release`` semantics specified above, participates in
7199 the global program order of other ``seq_cst`` operations and/or fences.
7201 The optional ":ref:`singlethread <singlethread>`" argument specifies
7202 that the fence only synchronizes with other fences in the same thread.
7203 (This is useful for interacting with signal handlers.)
7208 .. code-block:: llvm
7210 fence acquire ; yields void
7211 fence singlethread seq_cst ; yields void
7215 '``cmpxchg``' Instruction
7216 ^^^^^^^^^^^^^^^^^^^^^^^^^
7223 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields { ty, i1 }
7228 The '``cmpxchg``' instruction is used to atomically modify memory. It
7229 loads a value in memory and compares it to a given value. If they are
7230 equal, it tries to store a new value into the memory.
7235 There are three arguments to the '``cmpxchg``' instruction: an address
7236 to operate on, a value to compare to the value currently be at that
7237 address, and a new value to place at that address if the compared values
7238 are equal. The type of '<cmp>' must be an integer type whose bit width
7239 is a power of two greater than or equal to eight and less than or equal
7240 to a target-specific size limit. '<cmp>' and '<new>' must have the same
7241 type, and the type of '<pointer>' must be a pointer to that type. If the
7242 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
7243 to modify the number or order of execution of this ``cmpxchg`` with
7244 other :ref:`volatile operations <volatile>`.
7246 The success and failure :ref:`ordering <ordering>` arguments specify how this
7247 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
7248 must be at least ``monotonic``, the ordering constraint on failure must be no
7249 stronger than that on success, and the failure ordering cannot be either
7250 ``release`` or ``acq_rel``.
7252 The optional "``singlethread``" argument declares that the ``cmpxchg``
7253 is only atomic with respect to code (usually signal handlers) running in
7254 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
7255 respect to all other code in the system.
7257 The pointer passed into cmpxchg must have alignment greater than or
7258 equal to the size in memory of the operand.
7263 The contents of memory at the location specified by the '``<pointer>``' operand
7264 is read and compared to '``<cmp>``'; if the read value is the equal, the
7265 '``<new>``' is written. The original value at the location is returned, together
7266 with a flag indicating success (true) or failure (false).
7268 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
7269 permitted: the operation may not write ``<new>`` even if the comparison
7272 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
7273 if the value loaded equals ``cmp``.
7275 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
7276 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
7277 load with an ordering parameter determined the second ordering parameter.
7282 .. code-block:: llvm
7285 %orig = atomic load i32, i32* %ptr unordered ; yields i32
7289 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
7290 %squared = mul i32 %cmp, %cmp
7291 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
7292 %value_loaded = extractvalue { i32, i1 } %val_success, 0
7293 %success = extractvalue { i32, i1 } %val_success, 1
7294 br i1 %success, label %done, label %loop
7301 '``atomicrmw``' Instruction
7302 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7309 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields ty
7314 The '``atomicrmw``' instruction is used to atomically modify memory.
7319 There are three arguments to the '``atomicrmw``' instruction: an
7320 operation to apply, an address whose value to modify, an argument to the
7321 operation. The operation must be one of the following keywords:
7335 The type of '<value>' must be an integer type whose bit width is a power
7336 of two greater than or equal to eight and less than or equal to a
7337 target-specific size limit. The type of the '``<pointer>``' operand must
7338 be a pointer to that type. If the ``atomicrmw`` is marked as
7339 ``volatile``, then the optimizer is not allowed to modify the number or
7340 order of execution of this ``atomicrmw`` with other :ref:`volatile
7341 operations <volatile>`.
7346 The contents of memory at the location specified by the '``<pointer>``'
7347 operand are atomically read, modified, and written back. The original
7348 value at the location is returned. The modification is specified by the
7351 - xchg: ``*ptr = val``
7352 - add: ``*ptr = *ptr + val``
7353 - sub: ``*ptr = *ptr - val``
7354 - and: ``*ptr = *ptr & val``
7355 - nand: ``*ptr = ~(*ptr & val)``
7356 - or: ``*ptr = *ptr | val``
7357 - xor: ``*ptr = *ptr ^ val``
7358 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
7359 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
7360 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
7362 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
7368 .. code-block:: llvm
7370 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
7372 .. _i_getelementptr:
7374 '``getelementptr``' Instruction
7375 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7382 <result> = getelementptr <ty>, <ty>* <ptrval>{, <ty> <idx>}*
7383 <result> = getelementptr inbounds <ty>, <ty>* <ptrval>{, <ty> <idx>}*
7384 <result> = getelementptr <ty>, <ptr vector> <ptrval>, <vector index type> <idx>
7389 The '``getelementptr``' instruction is used to get the address of a
7390 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
7391 address calculation only and does not access memory. The instruction can also
7392 be used to calculate a vector of such addresses.
7397 The first argument is always a type used as the basis for the calculations.
7398 The second argument is always a pointer or a vector of pointers, and is the
7399 base address to start from. The remaining arguments are indices
7400 that indicate which of the elements of the aggregate object are indexed.
7401 The interpretation of each index is dependent on the type being indexed
7402 into. The first index always indexes the pointer value given as the
7403 first argument, the second index indexes a value of the type pointed to
7404 (not necessarily the value directly pointed to, since the first index
7405 can be non-zero), etc. The first type indexed into must be a pointer
7406 value, subsequent types can be arrays, vectors, and structs. Note that
7407 subsequent types being indexed into can never be pointers, since that
7408 would require loading the pointer before continuing calculation.
7410 The type of each index argument depends on the type it is indexing into.
7411 When indexing into a (optionally packed) structure, only ``i32`` integer
7412 **constants** are allowed (when using a vector of indices they must all
7413 be the **same** ``i32`` integer constant). When indexing into an array,
7414 pointer or vector, integers of any width are allowed, and they are not
7415 required to be constant. These integers are treated as signed values
7418 For example, let's consider a C code fragment and how it gets compiled
7434 int *foo(struct ST *s) {
7435 return &s[1].Z.B[5][13];
7438 The LLVM code generated by Clang is:
7440 .. code-block:: llvm
7442 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
7443 %struct.ST = type { i32, double, %struct.RT }
7445 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
7447 %arrayidx = getelementptr inbounds %struct.ST, %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
7454 In the example above, the first index is indexing into the
7455 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
7456 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
7457 indexes into the third element of the structure, yielding a
7458 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
7459 structure. The third index indexes into the second element of the
7460 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
7461 dimensions of the array are subscripted into, yielding an '``i32``'
7462 type. The '``getelementptr``' instruction returns a pointer to this
7463 element, thus computing a value of '``i32*``' type.
7465 Note that it is perfectly legal to index partially through a structure,
7466 returning a pointer to an inner element. Because of this, the LLVM code
7467 for the given testcase is equivalent to:
7469 .. code-block:: llvm
7471 define i32* @foo(%struct.ST* %s) {
7472 %t1 = getelementptr %struct.ST, %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
7473 %t2 = getelementptr %struct.ST, %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
7474 %t3 = getelementptr %struct.RT, %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
7475 %t4 = getelementptr [10 x [20 x i32]], [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
7476 %t5 = getelementptr [20 x i32], [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
7480 If the ``inbounds`` keyword is present, the result value of the
7481 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
7482 pointer is not an *in bounds* address of an allocated object, or if any
7483 of the addresses that would be formed by successive addition of the
7484 offsets implied by the indices to the base address with infinitely
7485 precise signed arithmetic are not an *in bounds* address of that
7486 allocated object. The *in bounds* addresses for an allocated object are
7487 all the addresses that point into the object, plus the address one byte
7488 past the end. In cases where the base is a vector of pointers the
7489 ``inbounds`` keyword applies to each of the computations element-wise.
7491 If the ``inbounds`` keyword is not present, the offsets are added to the
7492 base address with silently-wrapping two's complement arithmetic. If the
7493 offsets have a different width from the pointer, they are sign-extended
7494 or truncated to the width of the pointer. The result value of the
7495 ``getelementptr`` may be outside the object pointed to by the base
7496 pointer. The result value may not necessarily be used to access memory
7497 though, even if it happens to point into allocated storage. See the
7498 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
7501 The getelementptr instruction is often confusing. For some more insight
7502 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
7507 .. code-block:: llvm
7509 ; yields [12 x i8]*:aptr
7510 %aptr = getelementptr {i32, [12 x i8]}, {i32, [12 x i8]}* %saptr, i64 0, i32 1
7512 %vptr = getelementptr {i32, <2 x i8>}, {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
7514 %eptr = getelementptr [12 x i8], [12 x i8]* %aptr, i64 0, i32 1
7516 %iptr = getelementptr [10 x i32], [10 x i32]* @arr, i16 0, i16 0
7521 The ``getelementptr`` returns a vector of pointers, instead of a single address,
7522 when one or more of its arguments is a vector. In such cases, all vector
7523 arguments should have the same number of elements, and every scalar argument
7524 will be effectively broadcast into a vector during address calculation.
7526 .. code-block:: llvm
7528 ; All arguments are vectors:
7529 ; A[i] = ptrs[i] + offsets[i]*sizeof(i8)
7530 %A = getelementptr i8, <4 x i8*> %ptrs, <4 x i64> %offsets
7532 ; Add the same scalar offset to each pointer of a vector:
7533 ; A[i] = ptrs[i] + offset*sizeof(i8)
7534 %A = getelementptr i8, <4 x i8*> %ptrs, i64 %offset
7536 ; Add distinct offsets to the same pointer:
7537 ; A[i] = ptr + offsets[i]*sizeof(i8)
7538 %A = getelementptr i8, i8* %ptr, <4 x i64> %offsets
7540 ; In all cases described above the type of the result is <4 x i8*>
7542 The two following instructions are equivalent:
7544 .. code-block:: llvm
7546 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
7547 <4 x i32> <i32 2, i32 2, i32 2, i32 2>,
7548 <4 x i32> <i32 1, i32 1, i32 1, i32 1>,
7550 <4 x i64> <i64 13, i64 13, i64 13, i64 13>
7552 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
7553 i32 2, i32 1, <4 x i32> %ind4, i64 13
7555 Let's look at the C code, where the vector version of ``getelementptr``
7560 // Let's assume that we vectorize the following loop:
7561 double *A, B; int *C;
7562 for (int i = 0; i < size; ++i) {
7566 .. code-block:: llvm
7568 ; get pointers for 8 elements from array B
7569 %ptrs = getelementptr double, double* %B, <8 x i32> %C
7570 ; load 8 elements from array B into A
7571 %A = call <8 x double> @llvm.masked.gather.v8f64(<8 x double*> %ptrs,
7572 i32 8, <8 x i1> %mask, <8 x double> %passthru)
7574 Conversion Operations
7575 ---------------------
7577 The instructions in this category are the conversion instructions
7578 (casting) which all take a single operand and a type. They perform
7579 various bit conversions on the operand.
7581 '``trunc .. to``' Instruction
7582 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7589 <result> = trunc <ty> <value> to <ty2> ; yields ty2
7594 The '``trunc``' instruction truncates its operand to the type ``ty2``.
7599 The '``trunc``' instruction takes a value to trunc, and a type to trunc
7600 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
7601 of the same number of integers. The bit size of the ``value`` must be
7602 larger than the bit size of the destination type, ``ty2``. Equal sized
7603 types are not allowed.
7608 The '``trunc``' instruction truncates the high order bits in ``value``
7609 and converts the remaining bits to ``ty2``. Since the source size must
7610 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
7611 It will always truncate bits.
7616 .. code-block:: llvm
7618 %X = trunc i32 257 to i8 ; yields i8:1
7619 %Y = trunc i32 123 to i1 ; yields i1:true
7620 %Z = trunc i32 122 to i1 ; yields i1:false
7621 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
7623 '``zext .. to``' Instruction
7624 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7631 <result> = zext <ty> <value> to <ty2> ; yields ty2
7636 The '``zext``' instruction zero extends its operand to type ``ty2``.
7641 The '``zext``' instruction takes a value to cast, and a type to cast it
7642 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
7643 the same number of integers. The bit size of the ``value`` must be
7644 smaller than the bit size of the destination type, ``ty2``.
7649 The ``zext`` fills the high order bits of the ``value`` with zero bits
7650 until it reaches the size of the destination type, ``ty2``.
7652 When zero extending from i1, the result will always be either 0 or 1.
7657 .. code-block:: llvm
7659 %X = zext i32 257 to i64 ; yields i64:257
7660 %Y = zext i1 true to i32 ; yields i32:1
7661 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
7663 '``sext .. to``' Instruction
7664 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7671 <result> = sext <ty> <value> to <ty2> ; yields ty2
7676 The '``sext``' sign extends ``value`` to the type ``ty2``.
7681 The '``sext``' instruction takes a value to cast, and a type to cast it
7682 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
7683 the same number of integers. The bit size of the ``value`` must be
7684 smaller than the bit size of the destination type, ``ty2``.
7689 The '``sext``' instruction performs a sign extension by copying the sign
7690 bit (highest order bit) of the ``value`` until it reaches the bit size
7691 of the type ``ty2``.
7693 When sign extending from i1, the extension always results in -1 or 0.
7698 .. code-block:: llvm
7700 %X = sext i8 -1 to i16 ; yields i16 :65535
7701 %Y = sext i1 true to i32 ; yields i32:-1
7702 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
7704 '``fptrunc .. to``' Instruction
7705 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7712 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
7717 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
7722 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
7723 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
7724 The size of ``value`` must be larger than the size of ``ty2``. This
7725 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
7730 The '``fptrunc``' instruction casts a ``value`` from a larger
7731 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
7732 point <t_floating>` type. If the value cannot fit (i.e. overflows) within the
7733 destination type, ``ty2``, then the results are undefined. If the cast produces
7734 an inexact result, how rounding is performed (e.g. truncation, also known as
7735 round to zero) is undefined.
7740 .. code-block:: llvm
7742 %X = fptrunc double 123.0 to float ; yields float:123.0
7743 %Y = fptrunc double 1.0E+300 to float ; yields undefined
7745 '``fpext .. to``' Instruction
7746 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7753 <result> = fpext <ty> <value> to <ty2> ; yields ty2
7758 The '``fpext``' extends a floating point ``value`` to a larger floating
7764 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
7765 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
7766 to. The source type must be smaller than the destination type.
7771 The '``fpext``' instruction extends the ``value`` from a smaller
7772 :ref:`floating point <t_floating>` type to a larger :ref:`floating
7773 point <t_floating>` type. The ``fpext`` cannot be used to make a
7774 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
7775 *no-op cast* for a floating point cast.
7780 .. code-block:: llvm
7782 %X = fpext float 3.125 to double ; yields double:3.125000e+00
7783 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
7785 '``fptoui .. to``' Instruction
7786 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7793 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
7798 The '``fptoui``' converts a floating point ``value`` to its unsigned
7799 integer equivalent of type ``ty2``.
7804 The '``fptoui``' instruction takes a value to cast, which must be a
7805 scalar or vector :ref:`floating point <t_floating>` value, and a type to
7806 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
7807 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
7808 type with the same number of elements as ``ty``
7813 The '``fptoui``' instruction converts its :ref:`floating
7814 point <t_floating>` operand into the nearest (rounding towards zero)
7815 unsigned integer value. If the value cannot fit in ``ty2``, the results
7821 .. code-block:: llvm
7823 %X = fptoui double 123.0 to i32 ; yields i32:123
7824 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
7825 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
7827 '``fptosi .. to``' Instruction
7828 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7835 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
7840 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
7841 ``value`` to type ``ty2``.
7846 The '``fptosi``' instruction takes a value to cast, which must be a
7847 scalar or vector :ref:`floating point <t_floating>` value, and a type to
7848 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
7849 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
7850 type with the same number of elements as ``ty``
7855 The '``fptosi``' instruction converts its :ref:`floating
7856 point <t_floating>` operand into the nearest (rounding towards zero)
7857 signed integer value. If the value cannot fit in ``ty2``, the results
7863 .. code-block:: llvm
7865 %X = fptosi double -123.0 to i32 ; yields i32:-123
7866 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
7867 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
7869 '``uitofp .. to``' Instruction
7870 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7877 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
7882 The '``uitofp``' instruction regards ``value`` as an unsigned integer
7883 and converts that value to the ``ty2`` type.
7888 The '``uitofp``' instruction takes a value to cast, which must be a
7889 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
7890 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
7891 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
7892 type with the same number of elements as ``ty``
7897 The '``uitofp``' instruction interprets its operand as an unsigned
7898 integer quantity and converts it to the corresponding floating point
7899 value. If the value cannot fit in the floating point value, the results
7905 .. code-block:: llvm
7907 %X = uitofp i32 257 to float ; yields float:257.0
7908 %Y = uitofp i8 -1 to double ; yields double:255.0
7910 '``sitofp .. to``' Instruction
7911 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7918 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
7923 The '``sitofp``' instruction regards ``value`` as a signed integer and
7924 converts that value to the ``ty2`` type.
7929 The '``sitofp``' instruction takes a value to cast, which must be a
7930 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
7931 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
7932 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
7933 type with the same number of elements as ``ty``
7938 The '``sitofp``' instruction interprets its operand as a signed integer
7939 quantity and converts it to the corresponding floating point value. If
7940 the value cannot fit in the floating point value, the results are
7946 .. code-block:: llvm
7948 %X = sitofp i32 257 to float ; yields float:257.0
7949 %Y = sitofp i8 -1 to double ; yields double:-1.0
7953 '``ptrtoint .. to``' Instruction
7954 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7961 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
7966 The '``ptrtoint``' instruction converts the pointer or a vector of
7967 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
7972 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
7973 a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
7974 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
7975 a vector of integers type.
7980 The '``ptrtoint``' instruction converts ``value`` to integer type
7981 ``ty2`` by interpreting the pointer value as an integer and either
7982 truncating or zero extending that value to the size of the integer type.
7983 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
7984 ``value`` is larger than ``ty2`` then a truncation is done. If they are
7985 the same size, then nothing is done (*no-op cast*) other than a type
7991 .. code-block:: llvm
7993 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
7994 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
7995 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
7999 '``inttoptr .. to``' Instruction
8000 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8007 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
8012 The '``inttoptr``' instruction converts an integer ``value`` to a
8013 pointer type, ``ty2``.
8018 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
8019 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
8025 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
8026 applying either a zero extension or a truncation depending on the size
8027 of the integer ``value``. If ``value`` is larger than the size of a
8028 pointer then a truncation is done. If ``value`` is smaller than the size
8029 of a pointer then a zero extension is done. If they are the same size,
8030 nothing is done (*no-op cast*).
8035 .. code-block:: llvm
8037 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
8038 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
8039 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
8040 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
8044 '``bitcast .. to``' Instruction
8045 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8052 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
8057 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
8063 The '``bitcast``' instruction takes a value to cast, which must be a
8064 non-aggregate first class value, and a type to cast it to, which must
8065 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
8066 bit sizes of ``value`` and the destination type, ``ty2``, must be
8067 identical. If the source type is a pointer, the destination type must
8068 also be a pointer of the same size. This instruction supports bitwise
8069 conversion of vectors to integers and to vectors of other types (as
8070 long as they have the same size).
8075 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
8076 is always a *no-op cast* because no bits change with this
8077 conversion. The conversion is done as if the ``value`` had been stored
8078 to memory and read back as type ``ty2``. Pointer (or vector of
8079 pointers) types may only be converted to other pointer (or vector of
8080 pointers) types with the same address space through this instruction.
8081 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
8082 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
8087 .. code-block:: llvm
8089 %X = bitcast i8 255 to i8 ; yields i8 :-1
8090 %Y = bitcast i32* %x to sint* ; yields sint*:%x
8091 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
8092 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
8094 .. _i_addrspacecast:
8096 '``addrspacecast .. to``' Instruction
8097 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8104 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
8109 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
8110 address space ``n`` to type ``pty2`` in address space ``m``.
8115 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
8116 to cast and a pointer type to cast it to, which must have a different
8122 The '``addrspacecast``' instruction converts the pointer value
8123 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
8124 value modification, depending on the target and the address space
8125 pair. Pointer conversions within the same address space must be
8126 performed with the ``bitcast`` instruction. Note that if the address space
8127 conversion is legal then both result and operand refer to the same memory
8133 .. code-block:: llvm
8135 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
8136 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
8137 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
8144 The instructions in this category are the "miscellaneous" instructions,
8145 which defy better classification.
8149 '``icmp``' Instruction
8150 ^^^^^^^^^^^^^^^^^^^^^^
8157 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
8162 The '``icmp``' instruction returns a boolean value or a vector of
8163 boolean values based on comparison of its two integer, integer vector,
8164 pointer, or pointer vector operands.
8169 The '``icmp``' instruction takes three operands. The first operand is
8170 the condition code indicating the kind of comparison to perform. It is
8171 not a value, just a keyword. The possible condition code are:
8174 #. ``ne``: not equal
8175 #. ``ugt``: unsigned greater than
8176 #. ``uge``: unsigned greater or equal
8177 #. ``ult``: unsigned less than
8178 #. ``ule``: unsigned less or equal
8179 #. ``sgt``: signed greater than
8180 #. ``sge``: signed greater or equal
8181 #. ``slt``: signed less than
8182 #. ``sle``: signed less or equal
8184 The remaining two arguments must be :ref:`integer <t_integer>` or
8185 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
8186 must also be identical types.
8191 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
8192 code given as ``cond``. The comparison performed always yields either an
8193 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
8195 #. ``eq``: yields ``true`` if the operands are equal, ``false``
8196 otherwise. No sign interpretation is necessary or performed.
8197 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
8198 otherwise. No sign interpretation is necessary or performed.
8199 #. ``ugt``: interprets the operands as unsigned values and yields
8200 ``true`` if ``op1`` is greater than ``op2``.
8201 #. ``uge``: interprets the operands as unsigned values and yields
8202 ``true`` if ``op1`` is greater than or equal to ``op2``.
8203 #. ``ult``: interprets the operands as unsigned values and yields
8204 ``true`` if ``op1`` is less than ``op2``.
8205 #. ``ule``: interprets the operands as unsigned values and yields
8206 ``true`` if ``op1`` is less than or equal to ``op2``.
8207 #. ``sgt``: interprets the operands as signed values and yields ``true``
8208 if ``op1`` is greater than ``op2``.
8209 #. ``sge``: interprets the operands as signed values and yields ``true``
8210 if ``op1`` is greater than or equal to ``op2``.
8211 #. ``slt``: interprets the operands as signed values and yields ``true``
8212 if ``op1`` is less than ``op2``.
8213 #. ``sle``: interprets the operands as signed values and yields ``true``
8214 if ``op1`` is less than or equal to ``op2``.
8216 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
8217 are compared as if they were integers.
8219 If the operands are integer vectors, then they are compared element by
8220 element. The result is an ``i1`` vector with the same number of elements
8221 as the values being compared. Otherwise, the result is an ``i1``.
8226 .. code-block:: llvm
8228 <result> = icmp eq i32 4, 5 ; yields: result=false
8229 <result> = icmp ne float* %X, %X ; yields: result=false
8230 <result> = icmp ult i16 4, 5 ; yields: result=true
8231 <result> = icmp sgt i16 4, 5 ; yields: result=false
8232 <result> = icmp ule i16 -4, 5 ; yields: result=false
8233 <result> = icmp sge i16 4, 5 ; yields: result=false
8235 Note that the code generator does not yet support vector types with the
8236 ``icmp`` instruction.
8240 '``fcmp``' Instruction
8241 ^^^^^^^^^^^^^^^^^^^^^^
8248 <result> = fcmp [fast-math flags]* <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
8253 The '``fcmp``' instruction returns a boolean value or vector of boolean
8254 values based on comparison of its operands.
8256 If the operands are floating point scalars, then the result type is a
8257 boolean (:ref:`i1 <t_integer>`).
8259 If the operands are floating point vectors, then the result type is a
8260 vector of boolean with the same number of elements as the operands being
8266 The '``fcmp``' instruction takes three operands. The first operand is
8267 the condition code indicating the kind of comparison to perform. It is
8268 not a value, just a keyword. The possible condition code are:
8270 #. ``false``: no comparison, always returns false
8271 #. ``oeq``: ordered and equal
8272 #. ``ogt``: ordered and greater than
8273 #. ``oge``: ordered and greater than or equal
8274 #. ``olt``: ordered and less than
8275 #. ``ole``: ordered and less than or equal
8276 #. ``one``: ordered and not equal
8277 #. ``ord``: ordered (no nans)
8278 #. ``ueq``: unordered or equal
8279 #. ``ugt``: unordered or greater than
8280 #. ``uge``: unordered or greater than or equal
8281 #. ``ult``: unordered or less than
8282 #. ``ule``: unordered or less than or equal
8283 #. ``une``: unordered or not equal
8284 #. ``uno``: unordered (either nans)
8285 #. ``true``: no comparison, always returns true
8287 *Ordered* means that neither operand is a QNAN while *unordered* means
8288 that either operand may be a QNAN.
8290 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
8291 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
8292 type. They must have identical types.
8297 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
8298 condition code given as ``cond``. If the operands are vectors, then the
8299 vectors are compared element by element. Each comparison performed
8300 always yields an :ref:`i1 <t_integer>` result, as follows:
8302 #. ``false``: always yields ``false``, regardless of operands.
8303 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
8304 is equal to ``op2``.
8305 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
8306 is greater than ``op2``.
8307 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
8308 is greater than or equal to ``op2``.
8309 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
8310 is less than ``op2``.
8311 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
8312 is less than or equal to ``op2``.
8313 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
8314 is not equal to ``op2``.
8315 #. ``ord``: yields ``true`` if both operands are not a QNAN.
8316 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
8318 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
8319 greater than ``op2``.
8320 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
8321 greater than or equal to ``op2``.
8322 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
8324 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
8325 less than or equal to ``op2``.
8326 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
8327 not equal to ``op2``.
8328 #. ``uno``: yields ``true`` if either operand is a QNAN.
8329 #. ``true``: always yields ``true``, regardless of operands.
8331 The ``fcmp`` instruction can also optionally take any number of
8332 :ref:`fast-math flags <fastmath>`, which are optimization hints to enable
8333 otherwise unsafe floating point optimizations.
8335 Any set of fast-math flags are legal on an ``fcmp`` instruction, but the
8336 only flags that have any effect on its semantics are those that allow
8337 assumptions to be made about the values of input arguments; namely
8338 ``nnan``, ``ninf``, and ``nsz``. See :ref:`fastmath` for more information.
8343 .. code-block:: llvm
8345 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
8346 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
8347 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
8348 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
8350 Note that the code generator does not yet support vector types with the
8351 ``fcmp`` instruction.
8355 '``phi``' Instruction
8356 ^^^^^^^^^^^^^^^^^^^^^
8363 <result> = phi <ty> [ <val0>, <label0>], ...
8368 The '``phi``' instruction is used to implement the φ node in the SSA
8369 graph representing the function.
8374 The type of the incoming values is specified with the first type field.
8375 After this, the '``phi``' instruction takes a list of pairs as
8376 arguments, with one pair for each predecessor basic block of the current
8377 block. Only values of :ref:`first class <t_firstclass>` type may be used as
8378 the value arguments to the PHI node. Only labels may be used as the
8381 There must be no non-phi instructions between the start of a basic block
8382 and the PHI instructions: i.e. PHI instructions must be first in a basic
8385 For the purposes of the SSA form, the use of each incoming value is
8386 deemed to occur on the edge from the corresponding predecessor block to
8387 the current block (but after any definition of an '``invoke``'
8388 instruction's return value on the same edge).
8393 At runtime, the '``phi``' instruction logically takes on the value
8394 specified by the pair corresponding to the predecessor basic block that
8395 executed just prior to the current block.
8400 .. code-block:: llvm
8402 Loop: ; Infinite loop that counts from 0 on up...
8403 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
8404 %nextindvar = add i32 %indvar, 1
8409 '``select``' Instruction
8410 ^^^^^^^^^^^^^^^^^^^^^^^^
8417 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
8419 selty is either i1 or {<N x i1>}
8424 The '``select``' instruction is used to choose one value based on a
8425 condition, without IR-level branching.
8430 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
8431 values indicating the condition, and two values of the same :ref:`first
8432 class <t_firstclass>` type.
8437 If the condition is an i1 and it evaluates to 1, the instruction returns
8438 the first value argument; otherwise, it returns the second value
8441 If the condition is a vector of i1, then the value arguments must be
8442 vectors of the same size, and the selection is done element by element.
8444 If the condition is an i1 and the value arguments are vectors of the
8445 same size, then an entire vector is selected.
8450 .. code-block:: llvm
8452 %X = select i1 true, i8 17, i8 42 ; yields i8:17
8456 '``call``' Instruction
8457 ^^^^^^^^^^^^^^^^^^^^^^
8464 <result> = [tail | musttail | notail ] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
8470 The '``call``' instruction represents a simple function call.
8475 This instruction requires several arguments:
8477 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
8478 should perform tail call optimization. The ``tail`` marker is a hint that
8479 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
8480 means that the call must be tail call optimized in order for the program to
8481 be correct. The ``musttail`` marker provides these guarantees:
8483 #. The call will not cause unbounded stack growth if it is part of a
8484 recursive cycle in the call graph.
8485 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
8488 Both markers imply that the callee does not access allocas or varargs from
8489 the caller. Calls marked ``musttail`` must obey the following additional
8492 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
8493 or a pointer bitcast followed by a ret instruction.
8494 - The ret instruction must return the (possibly bitcasted) value
8495 produced by the call or void.
8496 - The caller and callee prototypes must match. Pointer types of
8497 parameters or return types may differ in pointee type, but not
8499 - The calling conventions of the caller and callee must match.
8500 - All ABI-impacting function attributes, such as sret, byval, inreg,
8501 returned, and inalloca, must match.
8502 - The callee must be varargs iff the caller is varargs. Bitcasting a
8503 non-varargs function to the appropriate varargs type is legal so
8504 long as the non-varargs prefixes obey the other rules.
8506 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
8507 the following conditions are met:
8509 - Caller and callee both have the calling convention ``fastcc``.
8510 - The call is in tail position (ret immediately follows call and ret
8511 uses value of call or is void).
8512 - Option ``-tailcallopt`` is enabled, or
8513 ``llvm::GuaranteedTailCallOpt`` is ``true``.
8514 - `Platform-specific constraints are
8515 met. <CodeGenerator.html#tailcallopt>`_
8517 #. The optional ``notail`` marker indicates that the optimizers should not add
8518 ``tail`` or ``musttail`` markers to the call. It is used to prevent tail
8519 call optimization from being performed on the call.
8521 #. The optional "cconv" marker indicates which :ref:`calling
8522 convention <callingconv>` the call should use. If none is
8523 specified, the call defaults to using C calling conventions. The
8524 calling convention of the call must match the calling convention of
8525 the target function, or else the behavior is undefined.
8526 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
8527 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
8529 #. '``ty``': the type of the call instruction itself which is also the
8530 type of the return value. Functions that return no value are marked
8532 #. '``fnty``': shall be the signature of the pointer to function value
8533 being invoked. The argument types must match the types implied by
8534 this signature. This type can be omitted if the function is not
8535 varargs and if the function type does not return a pointer to a
8537 #. '``fnptrval``': An LLVM value containing a pointer to a function to
8538 be invoked. In most cases, this is a direct function invocation, but
8539 indirect ``call``'s are just as possible, calling an arbitrary pointer
8541 #. '``function args``': argument list whose types match the function
8542 signature argument types and parameter attributes. All arguments must
8543 be of :ref:`first class <t_firstclass>` type. If the function signature
8544 indicates the function accepts a variable number of arguments, the
8545 extra arguments can be specified.
8546 #. The optional :ref:`function attributes <fnattrs>` list. Only
8547 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
8548 attributes are valid here.
8549 #. The optional :ref:`operand bundles <opbundles>` list.
8554 The '``call``' instruction is used to cause control flow to transfer to
8555 a specified function, with its incoming arguments bound to the specified
8556 values. Upon a '``ret``' instruction in the called function, control
8557 flow continues with the instruction after the function call, and the
8558 return value of the function is bound to the result argument.
8563 .. code-block:: llvm
8565 %retval = call i32 @test(i32 %argc)
8566 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
8567 %X = tail call i32 @foo() ; yields i32
8568 %Y = tail call fastcc i32 @foo() ; yields i32
8569 call void %foo(i8 97 signext)
8571 %struct.A = type { i32, i8 }
8572 %r = call %struct.A @foo() ; yields { i32, i8 }
8573 %gr = extractvalue %struct.A %r, 0 ; yields i32
8574 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
8575 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
8576 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
8578 llvm treats calls to some functions with names and arguments that match
8579 the standard C99 library as being the C99 library functions, and may
8580 perform optimizations or generate code for them under that assumption.
8581 This is something we'd like to change in the future to provide better
8582 support for freestanding environments and non-C-based languages.
8586 '``va_arg``' Instruction
8587 ^^^^^^^^^^^^^^^^^^^^^^^^
8594 <resultval> = va_arg <va_list*> <arglist>, <argty>
8599 The '``va_arg``' instruction is used to access arguments passed through
8600 the "variable argument" area of a function call. It is used to implement
8601 the ``va_arg`` macro in C.
8606 This instruction takes a ``va_list*`` value and the type of the
8607 argument. It returns a value of the specified argument type and
8608 increments the ``va_list`` to point to the next argument. The actual
8609 type of ``va_list`` is target specific.
8614 The '``va_arg``' instruction loads an argument of the specified type
8615 from the specified ``va_list`` and causes the ``va_list`` to point to
8616 the next argument. For more information, see the variable argument
8617 handling :ref:`Intrinsic Functions <int_varargs>`.
8619 It is legal for this instruction to be called in a function which does
8620 not take a variable number of arguments, for example, the ``vfprintf``
8623 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
8624 function <intrinsics>` because it takes a type as an argument.
8629 See the :ref:`variable argument processing <int_varargs>` section.
8631 Note that the code generator does not yet fully support va\_arg on many
8632 targets. Also, it does not currently support va\_arg with aggregate
8633 types on any target.
8637 '``landingpad``' Instruction
8638 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8645 <resultval> = landingpad <resultty> <clause>+
8646 <resultval> = landingpad <resultty> cleanup <clause>*
8648 <clause> := catch <type> <value>
8649 <clause> := filter <array constant type> <array constant>
8654 The '``landingpad``' instruction is used by `LLVM's exception handling
8655 system <ExceptionHandling.html#overview>`_ to specify that a basic block
8656 is a landing pad --- one where the exception lands, and corresponds to the
8657 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
8658 defines values supplied by the :ref:`personality function <personalityfn>` upon
8659 re-entry to the function. The ``resultval`` has the type ``resultty``.
8665 ``cleanup`` flag indicates that the landing pad block is a cleanup.
8667 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
8668 contains the global variable representing the "type" that may be caught
8669 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
8670 clause takes an array constant as its argument. Use
8671 "``[0 x i8**] undef``" for a filter which cannot throw. The
8672 '``landingpad``' instruction must contain *at least* one ``clause`` or
8673 the ``cleanup`` flag.
8678 The '``landingpad``' instruction defines the values which are set by the
8679 :ref:`personality function <personalityfn>` upon re-entry to the function, and
8680 therefore the "result type" of the ``landingpad`` instruction. As with
8681 calling conventions, how the personality function results are
8682 represented in LLVM IR is target specific.
8684 The clauses are applied in order from top to bottom. If two
8685 ``landingpad`` instructions are merged together through inlining, the
8686 clauses from the calling function are appended to the list of clauses.
8687 When the call stack is being unwound due to an exception being thrown,
8688 the exception is compared against each ``clause`` in turn. If it doesn't
8689 match any of the clauses, and the ``cleanup`` flag is not set, then
8690 unwinding continues further up the call stack.
8692 The ``landingpad`` instruction has several restrictions:
8694 - A landing pad block is a basic block which is the unwind destination
8695 of an '``invoke``' instruction.
8696 - A landing pad block must have a '``landingpad``' instruction as its
8697 first non-PHI instruction.
8698 - There can be only one '``landingpad``' instruction within the landing
8700 - A basic block that is not a landing pad block may not include a
8701 '``landingpad``' instruction.
8706 .. code-block:: llvm
8708 ;; A landing pad which can catch an integer.
8709 %res = landingpad { i8*, i32 }
8711 ;; A landing pad that is a cleanup.
8712 %res = landingpad { i8*, i32 }
8714 ;; A landing pad which can catch an integer and can only throw a double.
8715 %res = landingpad { i8*, i32 }
8717 filter [1 x i8**] [@_ZTId]
8721 '``cleanuppad``' Instruction
8722 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8729 <resultval> = cleanuppad [<args>*]
8734 The '``cleanuppad``' instruction is used by `LLVM's exception handling
8735 system <ExceptionHandling.html#overview>`_ to specify that a basic block
8736 is a cleanup block --- one where a personality routine attempts to
8737 transfer control to run cleanup actions.
8738 The ``args`` correspond to whatever additional
8739 information the :ref:`personality function <personalityfn>` requires to
8740 execute the cleanup.
8741 The ``resultval`` has the type :ref:`token <t_token>` and is used to
8742 match the ``cleanuppad`` to corresponding :ref:`cleanuprets <i_cleanupret>`
8743 and :ref:`cleanupendpads <i_cleanupendpad>`.
8748 The instruction takes a list of arbitrary values which are interpreted
8749 by the :ref:`personality function <personalityfn>`.
8754 When the call stack is being unwound due to an exception being thrown,
8755 the :ref:`personality function <personalityfn>` transfers control to the
8756 ``cleanuppad`` with the aid of the personality-specific arguments.
8757 As with calling conventions, how the personality function results are
8758 represented in LLVM IR is target specific.
8760 The ``cleanuppad`` instruction has several restrictions:
8762 - A cleanup block is a basic block which is the unwind destination of
8763 an exceptional instruction.
8764 - A cleanup block must have a '``cleanuppad``' instruction as its
8765 first non-PHI instruction.
8766 - There can be only one '``cleanuppad``' instruction within the
8768 - A basic block that is not a cleanup block may not include a
8769 '``cleanuppad``' instruction.
8770 - All '``cleanupret``'s and '``cleanupendpad``'s which consume a ``cleanuppad``
8771 must have the same exceptional successor.
8772 - It is undefined behavior for control to transfer from a ``cleanuppad`` to a
8773 ``ret`` without first executing a ``cleanupret`` or ``cleanupendpad`` that
8774 consumes the ``cleanuppad``.
8775 - It is undefined behavior for control to transfer from a ``cleanuppad`` to
8776 itself without first executing a ``cleanupret`` or ``cleanupendpad`` that
8777 consumes the ``cleanuppad``.
8782 .. code-block:: llvm
8784 %tok = cleanuppad []
8791 LLVM supports the notion of an "intrinsic function". These functions
8792 have well known names and semantics and are required to follow certain
8793 restrictions. Overall, these intrinsics represent an extension mechanism
8794 for the LLVM language that does not require changing all of the
8795 transformations in LLVM when adding to the language (or the bitcode
8796 reader/writer, the parser, etc...).
8798 Intrinsic function names must all start with an "``llvm.``" prefix. This
8799 prefix is reserved in LLVM for intrinsic names; thus, function names may
8800 not begin with this prefix. Intrinsic functions must always be external
8801 functions: you cannot define the body of intrinsic functions. Intrinsic
8802 functions may only be used in call or invoke instructions: it is illegal
8803 to take the address of an intrinsic function. Additionally, because
8804 intrinsic functions are part of the LLVM language, it is required if any
8805 are added that they be documented here.
8807 Some intrinsic functions can be overloaded, i.e., the intrinsic
8808 represents a family of functions that perform the same operation but on
8809 different data types. Because LLVM can represent over 8 million
8810 different integer types, overloading is used commonly to allow an
8811 intrinsic function to operate on any integer type. One or more of the
8812 argument types or the result type can be overloaded to accept any
8813 integer type. Argument types may also be defined as exactly matching a
8814 previous argument's type or the result type. This allows an intrinsic
8815 function which accepts multiple arguments, but needs all of them to be
8816 of the same type, to only be overloaded with respect to a single
8817 argument or the result.
8819 Overloaded intrinsics will have the names of its overloaded argument
8820 types encoded into its function name, each preceded by a period. Only
8821 those types which are overloaded result in a name suffix. Arguments
8822 whose type is matched against another type do not. For example, the
8823 ``llvm.ctpop`` function can take an integer of any width and returns an
8824 integer of exactly the same integer width. This leads to a family of
8825 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
8826 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
8827 overloaded, and only one type suffix is required. Because the argument's
8828 type is matched against the return type, it does not require its own
8831 To learn how to add an intrinsic function, please see the `Extending
8832 LLVM Guide <ExtendingLLVM.html>`_.
8836 Variable Argument Handling Intrinsics
8837 -------------------------------------
8839 Variable argument support is defined in LLVM with the
8840 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
8841 functions. These functions are related to the similarly named macros
8842 defined in the ``<stdarg.h>`` header file.
8844 All of these functions operate on arguments that use a target-specific
8845 value type "``va_list``". The LLVM assembly language reference manual
8846 does not define what this type is, so all transformations should be
8847 prepared to handle these functions regardless of the type used.
8849 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
8850 variable argument handling intrinsic functions are used.
8852 .. code-block:: llvm
8854 ; This struct is different for every platform. For most platforms,
8855 ; it is merely an i8*.
8856 %struct.va_list = type { i8* }
8858 ; For Unix x86_64 platforms, va_list is the following struct:
8859 ; %struct.va_list = type { i32, i32, i8*, i8* }
8861 define i32 @test(i32 %X, ...) {
8862 ; Initialize variable argument processing
8863 %ap = alloca %struct.va_list
8864 %ap2 = bitcast %struct.va_list* %ap to i8*
8865 call void @llvm.va_start(i8* %ap2)
8867 ; Read a single integer argument
8868 %tmp = va_arg i8* %ap2, i32
8870 ; Demonstrate usage of llvm.va_copy and llvm.va_end
8872 %aq2 = bitcast i8** %aq to i8*
8873 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
8874 call void @llvm.va_end(i8* %aq2)
8876 ; Stop processing of arguments.
8877 call void @llvm.va_end(i8* %ap2)
8881 declare void @llvm.va_start(i8*)
8882 declare void @llvm.va_copy(i8*, i8*)
8883 declare void @llvm.va_end(i8*)
8887 '``llvm.va_start``' Intrinsic
8888 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8895 declare void @llvm.va_start(i8* <arglist>)
8900 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
8901 subsequent use by ``va_arg``.
8906 The argument is a pointer to a ``va_list`` element to initialize.
8911 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
8912 available in C. In a target-dependent way, it initializes the
8913 ``va_list`` element to which the argument points, so that the next call
8914 to ``va_arg`` will produce the first variable argument passed to the
8915 function. Unlike the C ``va_start`` macro, this intrinsic does not need
8916 to know the last argument of the function as the compiler can figure
8919 '``llvm.va_end``' Intrinsic
8920 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8927 declare void @llvm.va_end(i8* <arglist>)
8932 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
8933 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
8938 The argument is a pointer to a ``va_list`` to destroy.
8943 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
8944 available in C. In a target-dependent way, it destroys the ``va_list``
8945 element to which the argument points. Calls to
8946 :ref:`llvm.va_start <int_va_start>` and
8947 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
8952 '``llvm.va_copy``' Intrinsic
8953 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8960 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
8965 The '``llvm.va_copy``' intrinsic copies the current argument position
8966 from the source argument list to the destination argument list.
8971 The first argument is a pointer to a ``va_list`` element to initialize.
8972 The second argument is a pointer to a ``va_list`` element to copy from.
8977 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
8978 available in C. In a target-dependent way, it copies the source
8979 ``va_list`` element into the destination ``va_list`` element. This
8980 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
8981 arbitrarily complex and require, for example, memory allocation.
8983 Accurate Garbage Collection Intrinsics
8984 --------------------------------------
8986 LLVM's support for `Accurate Garbage Collection <GarbageCollection.html>`_
8987 (GC) requires the frontend to generate code containing appropriate intrinsic
8988 calls and select an appropriate GC strategy which knows how to lower these
8989 intrinsics in a manner which is appropriate for the target collector.
8991 These intrinsics allow identification of :ref:`GC roots on the
8992 stack <int_gcroot>`, as well as garbage collector implementations that
8993 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
8994 Frontends for type-safe garbage collected languages should generate
8995 these intrinsics to make use of the LLVM garbage collectors. For more
8996 details, see `Garbage Collection with LLVM <GarbageCollection.html>`_.
8998 Experimental Statepoint Intrinsics
8999 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9001 LLVM provides an second experimental set of intrinsics for describing garbage
9002 collection safepoints in compiled code. These intrinsics are an alternative
9003 to the ``llvm.gcroot`` intrinsics, but are compatible with the ones for
9004 :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers. The
9005 differences in approach are covered in the `Garbage Collection with LLVM
9006 <GarbageCollection.html>`_ documentation. The intrinsics themselves are
9007 described in :doc:`Statepoints`.
9011 '``llvm.gcroot``' Intrinsic
9012 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9019 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
9024 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
9025 the code generator, and allows some metadata to be associated with it.
9030 The first argument specifies the address of a stack object that contains
9031 the root pointer. The second pointer (which must be either a constant or
9032 a global value address) contains the meta-data to be associated with the
9038 At runtime, a call to this intrinsic stores a null pointer into the
9039 "ptrloc" location. At compile-time, the code generator generates
9040 information to allow the runtime to find the pointer at GC safe points.
9041 The '``llvm.gcroot``' intrinsic may only be used in a function which
9042 :ref:`specifies a GC algorithm <gc>`.
9046 '``llvm.gcread``' Intrinsic
9047 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9054 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
9059 The '``llvm.gcread``' intrinsic identifies reads of references from heap
9060 locations, allowing garbage collector implementations that require read
9066 The second argument is the address to read from, which should be an
9067 address allocated from the garbage collector. The first object is a
9068 pointer to the start of the referenced object, if needed by the language
9069 runtime (otherwise null).
9074 The '``llvm.gcread``' intrinsic has the same semantics as a load
9075 instruction, but may be replaced with substantially more complex code by
9076 the garbage collector runtime, as needed. The '``llvm.gcread``'
9077 intrinsic may only be used in a function which :ref:`specifies a GC
9082 '``llvm.gcwrite``' Intrinsic
9083 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9090 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
9095 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
9096 locations, allowing garbage collector implementations that require write
9097 barriers (such as generational or reference counting collectors).
9102 The first argument is the reference to store, the second is the start of
9103 the object to store it to, and the third is the address of the field of
9104 Obj to store to. If the runtime does not require a pointer to the
9105 object, Obj may be null.
9110 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
9111 instruction, but may be replaced with substantially more complex code by
9112 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
9113 intrinsic may only be used in a function which :ref:`specifies a GC
9116 Code Generator Intrinsics
9117 -------------------------
9119 These intrinsics are provided by LLVM to expose special features that
9120 may only be implemented with code generator support.
9122 '``llvm.returnaddress``' Intrinsic
9123 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9130 declare i8 *@llvm.returnaddress(i32 <level>)
9135 The '``llvm.returnaddress``' intrinsic attempts to compute a
9136 target-specific value indicating the return address of the current
9137 function or one of its callers.
9142 The argument to this intrinsic indicates which function to return the
9143 address for. Zero indicates the calling function, one indicates its
9144 caller, etc. The argument is **required** to be a constant integer
9150 The '``llvm.returnaddress``' intrinsic either returns a pointer
9151 indicating the return address of the specified call frame, or zero if it
9152 cannot be identified. The value returned by this intrinsic is likely to
9153 be incorrect or 0 for arguments other than zero, so it should only be
9154 used for debugging purposes.
9156 Note that calling this intrinsic does not prevent function inlining or
9157 other aggressive transformations, so the value returned may not be that
9158 of the obvious source-language caller.
9160 '``llvm.frameaddress``' Intrinsic
9161 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9168 declare i8* @llvm.frameaddress(i32 <level>)
9173 The '``llvm.frameaddress``' intrinsic attempts to return the
9174 target-specific frame pointer value for the specified stack frame.
9179 The argument to this intrinsic indicates which function to return the
9180 frame pointer for. Zero indicates the calling function, one indicates
9181 its caller, etc. The argument is **required** to be a constant integer
9187 The '``llvm.frameaddress``' intrinsic either returns a pointer
9188 indicating the frame address of the specified call frame, or zero if it
9189 cannot be identified. The value returned by this intrinsic is likely to
9190 be incorrect or 0 for arguments other than zero, so it should only be
9191 used for debugging purposes.
9193 Note that calling this intrinsic does not prevent function inlining or
9194 other aggressive transformations, so the value returned may not be that
9195 of the obvious source-language caller.
9197 '``llvm.localescape``' and '``llvm.localrecover``' Intrinsics
9198 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9205 declare void @llvm.localescape(...)
9206 declare i8* @llvm.localrecover(i8* %func, i8* %fp, i32 %idx)
9211 The '``llvm.localescape``' intrinsic escapes offsets of a collection of static
9212 allocas, and the '``llvm.localrecover``' intrinsic applies those offsets to a
9213 live frame pointer to recover the address of the allocation. The offset is
9214 computed during frame layout of the caller of ``llvm.localescape``.
9219 All arguments to '``llvm.localescape``' must be pointers to static allocas or
9220 casts of static allocas. Each function can only call '``llvm.localescape``'
9221 once, and it can only do so from the entry block.
9223 The ``func`` argument to '``llvm.localrecover``' must be a constant
9224 bitcasted pointer to a function defined in the current module. The code
9225 generator cannot determine the frame allocation offset of functions defined in
9228 The ``fp`` argument to '``llvm.localrecover``' must be a frame pointer of a
9229 call frame that is currently live. The return value of '``llvm.localaddress``'
9230 is one way to produce such a value, but various runtimes also expose a suitable
9231 pointer in platform-specific ways.
9233 The ``idx`` argument to '``llvm.localrecover``' indicates which alloca passed to
9234 '``llvm.localescape``' to recover. It is zero-indexed.
9239 These intrinsics allow a group of functions to share access to a set of local
9240 stack allocations of a one parent function. The parent function may call the
9241 '``llvm.localescape``' intrinsic once from the function entry block, and the
9242 child functions can use '``llvm.localrecover``' to access the escaped allocas.
9243 The '``llvm.localescape``' intrinsic blocks inlining, as inlining changes where
9244 the escaped allocas are allocated, which would break attempts to use
9245 '``llvm.localrecover``'.
9247 .. _int_read_register:
9248 .. _int_write_register:
9250 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
9251 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9258 declare i32 @llvm.read_register.i32(metadata)
9259 declare i64 @llvm.read_register.i64(metadata)
9260 declare void @llvm.write_register.i32(metadata, i32 @value)
9261 declare void @llvm.write_register.i64(metadata, i64 @value)
9267 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
9268 provides access to the named register. The register must be valid on
9269 the architecture being compiled to. The type needs to be compatible
9270 with the register being read.
9275 The '``llvm.read_register``' intrinsic returns the current value of the
9276 register, where possible. The '``llvm.write_register``' intrinsic sets
9277 the current value of the register, where possible.
9279 This is useful to implement named register global variables that need
9280 to always be mapped to a specific register, as is common practice on
9281 bare-metal programs including OS kernels.
9283 The compiler doesn't check for register availability or use of the used
9284 register in surrounding code, including inline assembly. Because of that,
9285 allocatable registers are not supported.
9287 Warning: So far it only works with the stack pointer on selected
9288 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
9289 work is needed to support other registers and even more so, allocatable
9294 '``llvm.stacksave``' Intrinsic
9295 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9302 declare i8* @llvm.stacksave()
9307 The '``llvm.stacksave``' intrinsic is used to remember the current state
9308 of the function stack, for use with
9309 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
9310 implementing language features like scoped automatic variable sized
9316 This intrinsic returns a opaque pointer value that can be passed to
9317 :ref:`llvm.stackrestore <int_stackrestore>`. When an
9318 ``llvm.stackrestore`` intrinsic is executed with a value saved from
9319 ``llvm.stacksave``, it effectively restores the state of the stack to
9320 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
9321 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
9322 were allocated after the ``llvm.stacksave`` was executed.
9324 .. _int_stackrestore:
9326 '``llvm.stackrestore``' Intrinsic
9327 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9334 declare void @llvm.stackrestore(i8* %ptr)
9339 The '``llvm.stackrestore``' intrinsic is used to restore the state of
9340 the function stack to the state it was in when the corresponding
9341 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
9342 useful for implementing language features like scoped automatic variable
9343 sized arrays in C99.
9348 See the description for :ref:`llvm.stacksave <int_stacksave>`.
9350 .. _int_get_dynamic_area_offset:
9352 '``llvm.get.dynamic.area.offset``' Intrinsic
9353 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9360 declare i32 @llvm.get.dynamic.area.offset.i32()
9361 declare i64 @llvm.get.dynamic.area.offset.i64()
9366 The '``llvm.get.dynamic.area.offset.*``' intrinsic family is used to
9367 get the offset from native stack pointer to the address of the most
9368 recent dynamic alloca on the caller's stack. These intrinsics are
9369 intendend for use in combination with
9370 :ref:`llvm.stacksave <int_stacksave>` to get a
9371 pointer to the most recent dynamic alloca. This is useful, for example,
9372 for AddressSanitizer's stack unpoisoning routines.
9377 These intrinsics return a non-negative integer value that can be used to
9378 get the address of the most recent dynamic alloca, allocated by :ref:`alloca <i_alloca>`
9379 on the caller's stack. In particular, for targets where stack grows downwards,
9380 adding this offset to the native stack pointer would get the address of the most
9381 recent dynamic alloca. For targets where stack grows upwards, the situation is a bit more
9382 complicated, because substracting this value from stack pointer would get the address
9383 one past the end of the most recent dynamic alloca.
9385 Although for most targets `llvm.get.dynamic.area.offset <int_get_dynamic_area_offset>`
9386 returns just a zero, for others, such as PowerPC and PowerPC64, it returns a
9387 compile-time-known constant value.
9389 The return value type of :ref:`llvm.get.dynamic.area.offset <int_get_dynamic_area_offset>`
9390 must match the target's generic address space's (address space 0) pointer type.
9392 '``llvm.prefetch``' Intrinsic
9393 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9400 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
9405 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
9406 insert a prefetch instruction if supported; otherwise, it is a noop.
9407 Prefetches have no effect on the behavior of the program but can change
9408 its performance characteristics.
9413 ``address`` is the address to be prefetched, ``rw`` is the specifier
9414 determining if the fetch should be for a read (0) or write (1), and
9415 ``locality`` is a temporal locality specifier ranging from (0) - no
9416 locality, to (3) - extremely local keep in cache. The ``cache type``
9417 specifies whether the prefetch is performed on the data (1) or
9418 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
9419 arguments must be constant integers.
9424 This intrinsic does not modify the behavior of the program. In
9425 particular, prefetches cannot trap and do not produce a value. On
9426 targets that support this intrinsic, the prefetch can provide hints to
9427 the processor cache for better performance.
9429 '``llvm.pcmarker``' Intrinsic
9430 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9437 declare void @llvm.pcmarker(i32 <id>)
9442 The '``llvm.pcmarker``' intrinsic is a method to export a Program
9443 Counter (PC) in a region of code to simulators and other tools. The
9444 method is target specific, but it is expected that the marker will use
9445 exported symbols to transmit the PC of the marker. The marker makes no
9446 guarantees that it will remain with any specific instruction after
9447 optimizations. It is possible that the presence of a marker will inhibit
9448 optimizations. The intended use is to be inserted after optimizations to
9449 allow correlations of simulation runs.
9454 ``id`` is a numerical id identifying the marker.
9459 This intrinsic does not modify the behavior of the program. Backends
9460 that do not support this intrinsic may ignore it.
9462 '``llvm.readcyclecounter``' Intrinsic
9463 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9470 declare i64 @llvm.readcyclecounter()
9475 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
9476 counter register (or similar low latency, high accuracy clocks) on those
9477 targets that support it. On X86, it should map to RDTSC. On Alpha, it
9478 should map to RPCC. As the backing counters overflow quickly (on the
9479 order of 9 seconds on alpha), this should only be used for small
9485 When directly supported, reading the cycle counter should not modify any
9486 memory. Implementations are allowed to either return a application
9487 specific value or a system wide value. On backends without support, this
9488 is lowered to a constant 0.
9490 Note that runtime support may be conditional on the privilege-level code is
9491 running at and the host platform.
9493 '``llvm.clear_cache``' Intrinsic
9494 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9501 declare void @llvm.clear_cache(i8*, i8*)
9506 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
9507 in the specified range to the execution unit of the processor. On
9508 targets with non-unified instruction and data cache, the implementation
9509 flushes the instruction cache.
9514 On platforms with coherent instruction and data caches (e.g. x86), this
9515 intrinsic is a nop. On platforms with non-coherent instruction and data
9516 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
9517 instructions or a system call, if cache flushing requires special
9520 The default behavior is to emit a call to ``__clear_cache`` from the run
9523 This instrinsic does *not* empty the instruction pipeline. Modifications
9524 of the current function are outside the scope of the intrinsic.
9526 '``llvm.instrprof_increment``' Intrinsic
9527 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9534 declare void @llvm.instrprof_increment(i8* <name>, i64 <hash>,
9535 i32 <num-counters>, i32 <index>)
9540 The '``llvm.instrprof_increment``' intrinsic can be emitted by a
9541 frontend for use with instrumentation based profiling. These will be
9542 lowered by the ``-instrprof`` pass to generate execution counts of a
9548 The first argument is a pointer to a global variable containing the
9549 name of the entity being instrumented. This should generally be the
9550 (mangled) function name for a set of counters.
9552 The second argument is a hash value that can be used by the consumer
9553 of the profile data to detect changes to the instrumented source, and
9554 the third is the number of counters associated with ``name``. It is an
9555 error if ``hash`` or ``num-counters`` differ between two instances of
9556 ``instrprof_increment`` that refer to the same name.
9558 The last argument refers to which of the counters for ``name`` should
9559 be incremented. It should be a value between 0 and ``num-counters``.
9564 This intrinsic represents an increment of a profiling counter. It will
9565 cause the ``-instrprof`` pass to generate the appropriate data
9566 structures and the code to increment the appropriate value, in a
9567 format that can be written out by a compiler runtime and consumed via
9568 the ``llvm-profdata`` tool.
9570 '``llvm.instrprof_value_profile``' Intrinsic
9571 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9578 declare void @llvm.instrprof_value_profile(i8* <name>, i64 <hash>,
9579 i64 <value>, i32 <value_kind>,
9585 The '``llvm.instrprof_value_profile``' intrinsic can be emitted by a
9586 frontend for use with instrumentation based profiling. This will be
9587 lowered by the ``-instrprof`` pass to find out the target values,
9588 instrumented expressions take in a program at runtime.
9593 The first argument is a pointer to a global variable containing the
9594 name of the entity being instrumented. ``name`` should generally be the
9595 (mangled) function name for a set of counters.
9597 The second argument is a hash value that can be used by the consumer
9598 of the profile data to detect changes to the instrumented source. It
9599 is an error if ``hash`` differs between two instances of
9600 ``llvm.instrprof_*`` that refer to the same name.
9602 The third argument is the value of the expression being profiled. The profiled
9603 expression's value should be representable as an unsigned 64-bit value. The
9604 fourth argument represents the kind of value profiling that is being done. The
9605 supported value profiling kinds are enumerated through the
9606 ``InstrProfValueKind`` type declared in the
9607 ``<include/llvm/ProfileData/InstrProf.h>`` header file. The last argument is the
9608 index of the instrumented expression within ``name``. It should be >= 0.
9613 This intrinsic represents the point where a call to a runtime routine
9614 should be inserted for value profiling of target expressions. ``-instrprof``
9615 pass will generate the appropriate data structures and replace the
9616 ``llvm.instrprof_value_profile`` intrinsic with the call to the profile
9617 runtime library with proper arguments.
9619 Standard C Library Intrinsics
9620 -----------------------------
9622 LLVM provides intrinsics for a few important standard C library
9623 functions. These intrinsics allow source-language front-ends to pass
9624 information about the alignment of the pointer arguments to the code
9625 generator, providing opportunity for more efficient code generation.
9629 '``llvm.memcpy``' Intrinsic
9630 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9635 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
9636 integer bit width and for different address spaces. Not all targets
9637 support all bit widths however.
9641 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
9642 i32 <len>, i32 <align>, i1 <isvolatile>)
9643 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
9644 i64 <len>, i32 <align>, i1 <isvolatile>)
9649 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
9650 source location to the destination location.
9652 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
9653 intrinsics do not return a value, takes extra alignment/isvolatile
9654 arguments and the pointers can be in specified address spaces.
9659 The first argument is a pointer to the destination, the second is a
9660 pointer to the source. The third argument is an integer argument
9661 specifying the number of bytes to copy, the fourth argument is the
9662 alignment of the source and destination locations, and the fifth is a
9663 boolean indicating a volatile access.
9665 If the call to this intrinsic has an alignment value that is not 0 or 1,
9666 then the caller guarantees that both the source and destination pointers
9667 are aligned to that boundary.
9669 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
9670 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
9671 very cleanly specified and it is unwise to depend on it.
9676 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
9677 source location to the destination location, which are not allowed to
9678 overlap. It copies "len" bytes of memory over. If the argument is known
9679 to be aligned to some boundary, this can be specified as the fourth
9680 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
9682 '``llvm.memmove``' Intrinsic
9683 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9688 This is an overloaded intrinsic. You can use llvm.memmove on any integer
9689 bit width and for different address space. Not all targets support all
9694 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
9695 i32 <len>, i32 <align>, i1 <isvolatile>)
9696 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
9697 i64 <len>, i32 <align>, i1 <isvolatile>)
9702 The '``llvm.memmove.*``' intrinsics move a block of memory from the
9703 source location to the destination location. It is similar to the
9704 '``llvm.memcpy``' intrinsic but allows the two memory locations to
9707 Note that, unlike the standard libc function, the ``llvm.memmove.*``
9708 intrinsics do not return a value, takes extra alignment/isvolatile
9709 arguments and the pointers can be in specified address spaces.
9714 The first argument is a pointer to the destination, the second is a
9715 pointer to the source. The third argument is an integer argument
9716 specifying the number of bytes to copy, the fourth argument is the
9717 alignment of the source and destination locations, and the fifth is a
9718 boolean indicating a volatile access.
9720 If the call to this intrinsic has an alignment value that is not 0 or 1,
9721 then the caller guarantees that the source and destination pointers are
9722 aligned to that boundary.
9724 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
9725 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
9726 not very cleanly specified and it is unwise to depend on it.
9731 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
9732 source location to the destination location, which may overlap. It
9733 copies "len" bytes of memory over. If the argument is known to be
9734 aligned to some boundary, this can be specified as the fourth argument,
9735 otherwise it should be set to 0 or 1 (both meaning no alignment).
9737 '``llvm.memset.*``' Intrinsics
9738 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9743 This is an overloaded intrinsic. You can use llvm.memset on any integer
9744 bit width and for different address spaces. However, not all targets
9745 support all bit widths.
9749 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
9750 i32 <len>, i32 <align>, i1 <isvolatile>)
9751 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
9752 i64 <len>, i32 <align>, i1 <isvolatile>)
9757 The '``llvm.memset.*``' intrinsics fill a block of memory with a
9758 particular byte value.
9760 Note that, unlike the standard libc function, the ``llvm.memset``
9761 intrinsic does not return a value and takes extra alignment/volatile
9762 arguments. Also, the destination can be in an arbitrary address space.
9767 The first argument is a pointer to the destination to fill, the second
9768 is the byte value with which to fill it, the third argument is an
9769 integer argument specifying the number of bytes to fill, and the fourth
9770 argument is the known alignment of the destination location.
9772 If the call to this intrinsic has an alignment value that is not 0 or 1,
9773 then the caller guarantees that the destination pointer is aligned to
9776 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
9777 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
9778 very cleanly specified and it is unwise to depend on it.
9783 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
9784 at the destination location. If the argument is known to be aligned to
9785 some boundary, this can be specified as the fourth argument, otherwise
9786 it should be set to 0 or 1 (both meaning no alignment).
9788 '``llvm.sqrt.*``' Intrinsic
9789 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9794 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
9795 floating point or vector of floating point type. Not all targets support
9800 declare float @llvm.sqrt.f32(float %Val)
9801 declare double @llvm.sqrt.f64(double %Val)
9802 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
9803 declare fp128 @llvm.sqrt.f128(fp128 %Val)
9804 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
9809 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
9810 returning the same value as the libm '``sqrt``' functions would. Unlike
9811 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
9812 negative numbers other than -0.0 (which allows for better optimization,
9813 because there is no need to worry about errno being set).
9814 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
9819 The argument and return value are floating point numbers of the same
9825 This function returns the sqrt of the specified operand if it is a
9826 nonnegative floating point number.
9828 '``llvm.powi.*``' Intrinsic
9829 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9834 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
9835 floating point or vector of floating point type. Not all targets support
9840 declare float @llvm.powi.f32(float %Val, i32 %power)
9841 declare double @llvm.powi.f64(double %Val, i32 %power)
9842 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
9843 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
9844 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
9849 The '``llvm.powi.*``' intrinsics return the first operand raised to the
9850 specified (positive or negative) power. The order of evaluation of
9851 multiplications is not defined. When a vector of floating point type is
9852 used, the second argument remains a scalar integer value.
9857 The second argument is an integer power, and the first is a value to
9858 raise to that power.
9863 This function returns the first value raised to the second power with an
9864 unspecified sequence of rounding operations.
9866 '``llvm.sin.*``' Intrinsic
9867 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9872 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
9873 floating point or vector of floating point type. Not all targets support
9878 declare float @llvm.sin.f32(float %Val)
9879 declare double @llvm.sin.f64(double %Val)
9880 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
9881 declare fp128 @llvm.sin.f128(fp128 %Val)
9882 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
9887 The '``llvm.sin.*``' intrinsics return the sine of the operand.
9892 The argument and return value are floating point numbers of the same
9898 This function returns the sine of the specified operand, returning the
9899 same values as the libm ``sin`` functions would, and handles error
9900 conditions in the same way.
9902 '``llvm.cos.*``' Intrinsic
9903 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9908 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
9909 floating point or vector of floating point type. Not all targets support
9914 declare float @llvm.cos.f32(float %Val)
9915 declare double @llvm.cos.f64(double %Val)
9916 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
9917 declare fp128 @llvm.cos.f128(fp128 %Val)
9918 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
9923 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
9928 The argument and return value are floating point numbers of the same
9934 This function returns the cosine of the specified operand, returning the
9935 same values as the libm ``cos`` functions would, and handles error
9936 conditions in the same way.
9938 '``llvm.pow.*``' Intrinsic
9939 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9944 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
9945 floating point or vector of floating point type. Not all targets support
9950 declare float @llvm.pow.f32(float %Val, float %Power)
9951 declare double @llvm.pow.f64(double %Val, double %Power)
9952 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
9953 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
9954 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
9959 The '``llvm.pow.*``' intrinsics return the first operand raised to the
9960 specified (positive or negative) power.
9965 The second argument is a floating point power, and the first is a value
9966 to raise to that power.
9971 This function returns the first value raised to the second power,
9972 returning the same values as the libm ``pow`` functions would, and
9973 handles error conditions in the same way.
9975 '``llvm.exp.*``' Intrinsic
9976 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9981 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
9982 floating point or vector of floating point type. Not all targets support
9987 declare float @llvm.exp.f32(float %Val)
9988 declare double @llvm.exp.f64(double %Val)
9989 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
9990 declare fp128 @llvm.exp.f128(fp128 %Val)
9991 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
9996 The '``llvm.exp.*``' intrinsics perform the exp function.
10001 The argument and return value are floating point numbers of the same
10007 This function returns the same values as the libm ``exp`` functions
10008 would, and handles error conditions in the same way.
10010 '``llvm.exp2.*``' Intrinsic
10011 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10016 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
10017 floating point or vector of floating point type. Not all targets support
10022 declare float @llvm.exp2.f32(float %Val)
10023 declare double @llvm.exp2.f64(double %Val)
10024 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
10025 declare fp128 @llvm.exp2.f128(fp128 %Val)
10026 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
10031 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
10036 The argument and return value are floating point numbers of the same
10042 This function returns the same values as the libm ``exp2`` functions
10043 would, and handles error conditions in the same way.
10045 '``llvm.log.*``' Intrinsic
10046 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10051 This is an overloaded intrinsic. You can use ``llvm.log`` on any
10052 floating point or vector of floating point type. Not all targets support
10057 declare float @llvm.log.f32(float %Val)
10058 declare double @llvm.log.f64(double %Val)
10059 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
10060 declare fp128 @llvm.log.f128(fp128 %Val)
10061 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
10066 The '``llvm.log.*``' intrinsics perform the log function.
10071 The argument and return value are floating point numbers of the same
10077 This function returns the same values as the libm ``log`` functions
10078 would, and handles error conditions in the same way.
10080 '``llvm.log10.*``' Intrinsic
10081 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10086 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
10087 floating point or vector of floating point type. Not all targets support
10092 declare float @llvm.log10.f32(float %Val)
10093 declare double @llvm.log10.f64(double %Val)
10094 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
10095 declare fp128 @llvm.log10.f128(fp128 %Val)
10096 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
10101 The '``llvm.log10.*``' intrinsics perform the log10 function.
10106 The argument and return value are floating point numbers of the same
10112 This function returns the same values as the libm ``log10`` functions
10113 would, and handles error conditions in the same way.
10115 '``llvm.log2.*``' Intrinsic
10116 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10121 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
10122 floating point or vector of floating point type. Not all targets support
10127 declare float @llvm.log2.f32(float %Val)
10128 declare double @llvm.log2.f64(double %Val)
10129 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
10130 declare fp128 @llvm.log2.f128(fp128 %Val)
10131 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
10136 The '``llvm.log2.*``' intrinsics perform the log2 function.
10141 The argument and return value are floating point numbers of the same
10147 This function returns the same values as the libm ``log2`` functions
10148 would, and handles error conditions in the same way.
10150 '``llvm.fma.*``' Intrinsic
10151 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10156 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
10157 floating point or vector of floating point type. Not all targets support
10162 declare float @llvm.fma.f32(float %a, float %b, float %c)
10163 declare double @llvm.fma.f64(double %a, double %b, double %c)
10164 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
10165 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
10166 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
10171 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
10177 The argument and return value are floating point numbers of the same
10183 This function returns the same values as the libm ``fma`` functions
10184 would, and does not set errno.
10186 '``llvm.fabs.*``' Intrinsic
10187 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10192 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
10193 floating point or vector of floating point type. Not all targets support
10198 declare float @llvm.fabs.f32(float %Val)
10199 declare double @llvm.fabs.f64(double %Val)
10200 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
10201 declare fp128 @llvm.fabs.f128(fp128 %Val)
10202 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
10207 The '``llvm.fabs.*``' intrinsics return the absolute value of the
10213 The argument and return value are floating point numbers of the same
10219 This function returns the same values as the libm ``fabs`` functions
10220 would, and handles error conditions in the same way.
10222 '``llvm.minnum.*``' Intrinsic
10223 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10228 This is an overloaded intrinsic. You can use ``llvm.minnum`` on any
10229 floating point or vector of floating point type. Not all targets support
10234 declare float @llvm.minnum.f32(float %Val0, float %Val1)
10235 declare double @llvm.minnum.f64(double %Val0, double %Val1)
10236 declare x86_fp80 @llvm.minnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
10237 declare fp128 @llvm.minnum.f128(fp128 %Val0, fp128 %Val1)
10238 declare ppc_fp128 @llvm.minnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
10243 The '``llvm.minnum.*``' intrinsics return the minimum of the two
10250 The arguments and return value are floating point numbers of the same
10256 Follows the IEEE-754 semantics for minNum, which also match for libm's
10259 If either operand is a NaN, returns the other non-NaN operand. Returns
10260 NaN only if both operands are NaN. If the operands compare equal,
10261 returns a value that compares equal to both operands. This means that
10262 fmin(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
10264 '``llvm.maxnum.*``' Intrinsic
10265 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10270 This is an overloaded intrinsic. You can use ``llvm.maxnum`` on any
10271 floating point or vector of floating point type. Not all targets support
10276 declare float @llvm.maxnum.f32(float %Val0, float %Val1l)
10277 declare double @llvm.maxnum.f64(double %Val0, double %Val1)
10278 declare x86_fp80 @llvm.maxnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
10279 declare fp128 @llvm.maxnum.f128(fp128 %Val0, fp128 %Val1)
10280 declare ppc_fp128 @llvm.maxnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
10285 The '``llvm.maxnum.*``' intrinsics return the maximum of the two
10292 The arguments and return value are floating point numbers of the same
10297 Follows the IEEE-754 semantics for maxNum, which also match for libm's
10300 If either operand is a NaN, returns the other non-NaN operand. Returns
10301 NaN only if both operands are NaN. If the operands compare equal,
10302 returns a value that compares equal to both operands. This means that
10303 fmax(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
10305 '``llvm.copysign.*``' Intrinsic
10306 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10311 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
10312 floating point or vector of floating point type. Not all targets support
10317 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
10318 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
10319 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
10320 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
10321 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
10326 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
10327 first operand and the sign of the second operand.
10332 The arguments and return value are floating point numbers of the same
10338 This function returns the same values as the libm ``copysign``
10339 functions would, and handles error conditions in the same way.
10341 '``llvm.floor.*``' Intrinsic
10342 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10347 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
10348 floating point or vector of floating point type. Not all targets support
10353 declare float @llvm.floor.f32(float %Val)
10354 declare double @llvm.floor.f64(double %Val)
10355 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
10356 declare fp128 @llvm.floor.f128(fp128 %Val)
10357 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
10362 The '``llvm.floor.*``' intrinsics return the floor of the operand.
10367 The argument and return value are floating point numbers of the same
10373 This function returns the same values as the libm ``floor`` functions
10374 would, and handles error conditions in the same way.
10376 '``llvm.ceil.*``' Intrinsic
10377 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10382 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
10383 floating point or vector of floating point type. Not all targets support
10388 declare float @llvm.ceil.f32(float %Val)
10389 declare double @llvm.ceil.f64(double %Val)
10390 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
10391 declare fp128 @llvm.ceil.f128(fp128 %Val)
10392 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
10397 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
10402 The argument and return value are floating point numbers of the same
10408 This function returns the same values as the libm ``ceil`` functions
10409 would, and handles error conditions in the same way.
10411 '``llvm.trunc.*``' Intrinsic
10412 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10417 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
10418 floating point or vector of floating point type. Not all targets support
10423 declare float @llvm.trunc.f32(float %Val)
10424 declare double @llvm.trunc.f64(double %Val)
10425 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
10426 declare fp128 @llvm.trunc.f128(fp128 %Val)
10427 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
10432 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
10433 nearest integer not larger in magnitude than the operand.
10438 The argument and return value are floating point numbers of the same
10444 This function returns the same values as the libm ``trunc`` functions
10445 would, and handles error conditions in the same way.
10447 '``llvm.rint.*``' Intrinsic
10448 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10453 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
10454 floating point or vector of floating point type. Not all targets support
10459 declare float @llvm.rint.f32(float %Val)
10460 declare double @llvm.rint.f64(double %Val)
10461 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
10462 declare fp128 @llvm.rint.f128(fp128 %Val)
10463 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
10468 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
10469 nearest integer. It may raise an inexact floating-point exception if the
10470 operand isn't an integer.
10475 The argument and return value are floating point numbers of the same
10481 This function returns the same values as the libm ``rint`` functions
10482 would, and handles error conditions in the same way.
10484 '``llvm.nearbyint.*``' Intrinsic
10485 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10490 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
10491 floating point or vector of floating point type. Not all targets support
10496 declare float @llvm.nearbyint.f32(float %Val)
10497 declare double @llvm.nearbyint.f64(double %Val)
10498 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
10499 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
10500 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
10505 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
10511 The argument and return value are floating point numbers of the same
10517 This function returns the same values as the libm ``nearbyint``
10518 functions would, and handles error conditions in the same way.
10520 '``llvm.round.*``' Intrinsic
10521 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10526 This is an overloaded intrinsic. You can use ``llvm.round`` on any
10527 floating point or vector of floating point type. Not all targets support
10532 declare float @llvm.round.f32(float %Val)
10533 declare double @llvm.round.f64(double %Val)
10534 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
10535 declare fp128 @llvm.round.f128(fp128 %Val)
10536 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
10541 The '``llvm.round.*``' intrinsics returns the operand rounded to the
10547 The argument and return value are floating point numbers of the same
10553 This function returns the same values as the libm ``round``
10554 functions would, and handles error conditions in the same way.
10556 Bit Manipulation Intrinsics
10557 ---------------------------
10559 LLVM provides intrinsics for a few important bit manipulation
10560 operations. These allow efficient code generation for some algorithms.
10562 '``llvm.bitreverse.*``' Intrinsics
10563 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10568 This is an overloaded intrinsic function. You can use bitreverse on any
10573 declare i16 @llvm.bitreverse.i16(i16 <id>)
10574 declare i32 @llvm.bitreverse.i32(i32 <id>)
10575 declare i64 @llvm.bitreverse.i64(i64 <id>)
10580 The '``llvm.bitreverse``' family of intrinsics is used to reverse the
10581 bitpattern of an integer value; for example ``0b1234567`` becomes
10587 The ``llvm.bitreverse.iN`` intrinsic returns an i16 value that has bit
10588 ``M`` in the input moved to bit ``N-M`` in the output.
10590 '``llvm.bswap.*``' Intrinsics
10591 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10596 This is an overloaded intrinsic function. You can use bswap on any
10597 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
10601 declare i16 @llvm.bswap.i16(i16 <id>)
10602 declare i32 @llvm.bswap.i32(i32 <id>)
10603 declare i64 @llvm.bswap.i64(i64 <id>)
10608 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
10609 values with an even number of bytes (positive multiple of 16 bits).
10610 These are useful for performing operations on data that is not in the
10611 target's native byte order.
10616 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
10617 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
10618 intrinsic returns an i32 value that has the four bytes of the input i32
10619 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
10620 returned i32 will have its bytes in 3, 2, 1, 0 order. The
10621 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
10622 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
10625 '``llvm.ctpop.*``' Intrinsic
10626 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10631 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
10632 bit width, or on any vector with integer elements. Not all targets
10633 support all bit widths or vector types, however.
10637 declare i8 @llvm.ctpop.i8(i8 <src>)
10638 declare i16 @llvm.ctpop.i16(i16 <src>)
10639 declare i32 @llvm.ctpop.i32(i32 <src>)
10640 declare i64 @llvm.ctpop.i64(i64 <src>)
10641 declare i256 @llvm.ctpop.i256(i256 <src>)
10642 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
10647 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
10653 The only argument is the value to be counted. The argument may be of any
10654 integer type, or a vector with integer elements. The return type must
10655 match the argument type.
10660 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
10661 each element of a vector.
10663 '``llvm.ctlz.*``' Intrinsic
10664 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10669 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
10670 integer bit width, or any vector whose elements are integers. Not all
10671 targets support all bit widths or vector types, however.
10675 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
10676 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
10677 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
10678 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
10679 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
10680 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
10685 The '``llvm.ctlz``' family of intrinsic functions counts the number of
10686 leading zeros in a variable.
10691 The first argument is the value to be counted. This argument may be of
10692 any integer type, or a vector with integer element type. The return
10693 type must match the first argument type.
10695 The second argument must be a constant and is a flag to indicate whether
10696 the intrinsic should ensure that a zero as the first argument produces a
10697 defined result. Historically some architectures did not provide a
10698 defined result for zero values as efficiently, and many algorithms are
10699 now predicated on avoiding zero-value inputs.
10704 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
10705 zeros in a variable, or within each element of the vector. If
10706 ``src == 0`` then the result is the size in bits of the type of ``src``
10707 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
10708 ``llvm.ctlz(i32 2) = 30``.
10710 '``llvm.cttz.*``' Intrinsic
10711 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10716 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
10717 integer bit width, or any vector of integer elements. Not all targets
10718 support all bit widths or vector types, however.
10722 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
10723 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
10724 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
10725 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
10726 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
10727 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
10732 The '``llvm.cttz``' family of intrinsic functions counts the number of
10738 The first argument is the value to be counted. This argument may be of
10739 any integer type, or a vector with integer element type. The return
10740 type must match the first argument type.
10742 The second argument must be a constant and is a flag to indicate whether
10743 the intrinsic should ensure that a zero as the first argument produces a
10744 defined result. Historically some architectures did not provide a
10745 defined result for zero values as efficiently, and many algorithms are
10746 now predicated on avoiding zero-value inputs.
10751 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
10752 zeros in a variable, or within each element of a vector. If ``src == 0``
10753 then the result is the size in bits of the type of ``src`` if
10754 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
10755 ``llvm.cttz(2) = 1``.
10759 Arithmetic with Overflow Intrinsics
10760 -----------------------------------
10762 LLVM provides intrinsics for some arithmetic with overflow operations.
10764 '``llvm.sadd.with.overflow.*``' Intrinsics
10765 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10770 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
10771 on any integer bit width.
10775 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
10776 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
10777 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
10782 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
10783 a signed addition of the two arguments, and indicate whether an overflow
10784 occurred during the signed summation.
10789 The arguments (%a and %b) and the first element of the result structure
10790 may be of integer types of any bit width, but they must have the same
10791 bit width. The second element of the result structure must be of type
10792 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
10798 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
10799 a signed addition of the two variables. They return a structure --- the
10800 first element of which is the signed summation, and the second element
10801 of which is a bit specifying if the signed summation resulted in an
10807 .. code-block:: llvm
10809 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
10810 %sum = extractvalue {i32, i1} %res, 0
10811 %obit = extractvalue {i32, i1} %res, 1
10812 br i1 %obit, label %overflow, label %normal
10814 '``llvm.uadd.with.overflow.*``' Intrinsics
10815 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10820 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
10821 on any integer bit width.
10825 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
10826 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
10827 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
10832 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
10833 an unsigned addition of the two arguments, and indicate whether a carry
10834 occurred during the unsigned summation.
10839 The arguments (%a and %b) and the first element of the result structure
10840 may be of integer types of any bit width, but they must have the same
10841 bit width. The second element of the result structure must be of type
10842 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
10848 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
10849 an unsigned addition of the two arguments. They return a structure --- the
10850 first element of which is the sum, and the second element of which is a
10851 bit specifying if the unsigned summation resulted in a carry.
10856 .. code-block:: llvm
10858 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
10859 %sum = extractvalue {i32, i1} %res, 0
10860 %obit = extractvalue {i32, i1} %res, 1
10861 br i1 %obit, label %carry, label %normal
10863 '``llvm.ssub.with.overflow.*``' Intrinsics
10864 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10869 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
10870 on any integer bit width.
10874 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
10875 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
10876 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
10881 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
10882 a signed subtraction of the two arguments, and indicate whether an
10883 overflow occurred during the signed subtraction.
10888 The arguments (%a and %b) and the first element of the result structure
10889 may be of integer types of any bit width, but they must have the same
10890 bit width. The second element of the result structure must be of type
10891 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
10897 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
10898 a signed subtraction of the two arguments. They return a structure --- the
10899 first element of which is the subtraction, and the second element of
10900 which is a bit specifying if the signed subtraction resulted in an
10906 .. code-block:: llvm
10908 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
10909 %sum = extractvalue {i32, i1} %res, 0
10910 %obit = extractvalue {i32, i1} %res, 1
10911 br i1 %obit, label %overflow, label %normal
10913 '``llvm.usub.with.overflow.*``' Intrinsics
10914 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10919 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
10920 on any integer bit width.
10924 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
10925 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
10926 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
10931 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
10932 an unsigned subtraction of the two arguments, and indicate whether an
10933 overflow occurred during the unsigned subtraction.
10938 The arguments (%a and %b) and the first element of the result structure
10939 may be of integer types of any bit width, but they must have the same
10940 bit width. The second element of the result structure must be of type
10941 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
10947 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
10948 an unsigned subtraction of the two arguments. They return a structure ---
10949 the first element of which is the subtraction, and the second element of
10950 which is a bit specifying if the unsigned subtraction resulted in an
10956 .. code-block:: llvm
10958 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
10959 %sum = extractvalue {i32, i1} %res, 0
10960 %obit = extractvalue {i32, i1} %res, 1
10961 br i1 %obit, label %overflow, label %normal
10963 '``llvm.smul.with.overflow.*``' Intrinsics
10964 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10969 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
10970 on any integer bit width.
10974 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
10975 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
10976 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
10981 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
10982 a signed multiplication of the two arguments, and indicate whether an
10983 overflow occurred during the signed multiplication.
10988 The arguments (%a and %b) and the first element of the result structure
10989 may be of integer types of any bit width, but they must have the same
10990 bit width. The second element of the result structure must be of type
10991 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
10997 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
10998 a signed multiplication of the two arguments. They return a structure ---
10999 the first element of which is the multiplication, and the second element
11000 of which is a bit specifying if the signed multiplication resulted in an
11006 .. code-block:: llvm
11008 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
11009 %sum = extractvalue {i32, i1} %res, 0
11010 %obit = extractvalue {i32, i1} %res, 1
11011 br i1 %obit, label %overflow, label %normal
11013 '``llvm.umul.with.overflow.*``' Intrinsics
11014 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11019 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
11020 on any integer bit width.
11024 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
11025 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
11026 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
11031 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
11032 a unsigned multiplication of the two arguments, and indicate whether an
11033 overflow occurred during the unsigned multiplication.
11038 The arguments (%a and %b) and the first element of the result structure
11039 may be of integer types of any bit width, but they must have the same
11040 bit width. The second element of the result structure must be of type
11041 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
11047 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
11048 an unsigned multiplication of the two arguments. They return a structure ---
11049 the first element of which is the multiplication, and the second
11050 element of which is a bit specifying if the unsigned multiplication
11051 resulted in an overflow.
11056 .. code-block:: llvm
11058 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
11059 %sum = extractvalue {i32, i1} %res, 0
11060 %obit = extractvalue {i32, i1} %res, 1
11061 br i1 %obit, label %overflow, label %normal
11063 Specialised Arithmetic Intrinsics
11064 ---------------------------------
11066 '``llvm.canonicalize.*``' Intrinsic
11067 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11074 declare float @llvm.canonicalize.f32(float %a)
11075 declare double @llvm.canonicalize.f64(double %b)
11080 The '``llvm.canonicalize.*``' intrinsic returns the platform specific canonical
11081 encoding of a floating point number. This canonicalization is useful for
11082 implementing certain numeric primitives such as frexp. The canonical encoding is
11083 defined by IEEE-754-2008 to be:
11087 2.1.8 canonical encoding: The preferred encoding of a floating-point
11088 representation in a format. Applied to declets, significands of finite
11089 numbers, infinities, and NaNs, especially in decimal formats.
11091 This operation can also be considered equivalent to the IEEE-754-2008
11092 conversion of a floating-point value to the same format. NaNs are handled
11093 according to section 6.2.
11095 Examples of non-canonical encodings:
11097 - x87 pseudo denormals, pseudo NaNs, pseudo Infinity, Unnormals. These are
11098 converted to a canonical representation per hardware-specific protocol.
11099 - Many normal decimal floating point numbers have non-canonical alternative
11101 - Some machines, like GPUs or ARMv7 NEON, do not support subnormal values.
11102 These are treated as non-canonical encodings of zero and with be flushed to
11103 a zero of the same sign by this operation.
11105 Note that per IEEE-754-2008 6.2, systems that support signaling NaNs with
11106 default exception handling must signal an invalid exception, and produce a
11109 This function should always be implementable as multiplication by 1.0, provided
11110 that the compiler does not constant fold the operation. Likewise, division by
11111 1.0 and ``llvm.minnum(x, x)`` are possible implementations. Addition with
11112 -0.0 is also sufficient provided that the rounding mode is not -Infinity.
11114 ``@llvm.canonicalize`` must preserve the equality relation. That is:
11116 - ``(@llvm.canonicalize(x) == x)`` is equivalent to ``(x == x)``
11117 - ``(@llvm.canonicalize(x) == @llvm.canonicalize(y))`` is equivalent to
11120 Additionally, the sign of zero must be conserved:
11121 ``@llvm.canonicalize(-0.0) = -0.0`` and ``@llvm.canonicalize(+0.0) = +0.0``
11123 The payload bits of a NaN must be conserved, with two exceptions.
11124 First, environments which use only a single canonical representation of NaN
11125 must perform said canonicalization. Second, SNaNs must be quieted per the
11128 The canonicalization operation may be optimized away if:
11130 - The input is known to be canonical. For example, it was produced by a
11131 floating-point operation that is required by the standard to be canonical.
11132 - The result is consumed only by (or fused with) other floating-point
11133 operations. That is, the bits of the floating point value are not examined.
11135 '``llvm.fmuladd.*``' Intrinsic
11136 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11143 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
11144 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
11149 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
11150 expressions that can be fused if the code generator determines that (a) the
11151 target instruction set has support for a fused operation, and (b) that the
11152 fused operation is more efficient than the equivalent, separate pair of mul
11153 and add instructions.
11158 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
11159 multiplicands, a and b, and an addend c.
11168 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
11170 is equivalent to the expression a \* b + c, except that rounding will
11171 not be performed between the multiplication and addition steps if the
11172 code generator fuses the operations. Fusion is not guaranteed, even if
11173 the target platform supports it. If a fused multiply-add is required the
11174 corresponding llvm.fma.\* intrinsic function should be used
11175 instead. This never sets errno, just as '``llvm.fma.*``'.
11180 .. code-block:: llvm
11182 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
11185 '``llvm.uabsdiff.*``' and '``llvm.sabsdiff.*``' Intrinsics
11186 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11190 This is an overloaded intrinsic. The loaded data is a vector of any integer bit width.
11192 .. code-block:: llvm
11194 declare <4 x integer> @llvm.uabsdiff.v4i32(<4 x integer> %a, <4 x integer> %b)
11200 The ``llvm.uabsdiff`` intrinsic returns a vector result of the absolute difference
11201 of the two operands, treating them both as unsigned integers. The intermediate
11202 calculations are computed using infinitely precise unsigned arithmetic. The final
11203 result will be truncated to the given type.
11205 The ``llvm.sabsdiff`` intrinsic returns a vector result of the absolute difference of
11206 the two operands, treating them both as signed integers. If the result overflows, the
11207 behavior is undefined.
11211 These intrinsics are primarily used during the code generation stage of compilation.
11212 They are generated by compiler passes such as the Loop and SLP vectorizers. It is not
11213 recommended for users to create them manually.
11218 Both intrinsics take two integer of the same bitwidth.
11225 call <4 x i32> @llvm.uabsdiff.v4i32(<4 x i32> %a, <4 x i32> %b)
11229 %1 = zext <4 x i32> %a to <4 x i64>
11230 %2 = zext <4 x i32> %b to <4 x i64>
11231 %sub = sub <4 x i64> %1, %2
11232 %trunc = trunc <4 x i64> to <4 x i32>
11234 and the expression::
11236 call <4 x i32> @llvm.sabsdiff.v4i32(<4 x i32> %a, <4 x i32> %b)
11240 %sub = sub nsw <4 x i32> %a, %b
11241 %ispos = icmp sge <4 x i32> %sub, zeroinitializer
11242 %neg = sub nsw <4 x i32> zeroinitializer, %sub
11243 %1 = select <4 x i1> %ispos, <4 x i32> %sub, <4 x i32> %neg
11246 Half Precision Floating Point Intrinsics
11247 ----------------------------------------
11249 For most target platforms, half precision floating point is a
11250 storage-only format. This means that it is a dense encoding (in memory)
11251 but does not support computation in the format.
11253 This means that code must first load the half-precision floating point
11254 value as an i16, then convert it to float with
11255 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
11256 then be performed on the float value (including extending to double
11257 etc). To store the value back to memory, it is first converted to float
11258 if needed, then converted to i16 with
11259 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
11262 .. _int_convert_to_fp16:
11264 '``llvm.convert.to.fp16``' Intrinsic
11265 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11272 declare i16 @llvm.convert.to.fp16.f32(float %a)
11273 declare i16 @llvm.convert.to.fp16.f64(double %a)
11278 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
11279 conventional floating point type to half precision floating point format.
11284 The intrinsic function contains single argument - the value to be
11290 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
11291 conventional floating point format to half precision floating point format. The
11292 return value is an ``i16`` which contains the converted number.
11297 .. code-block:: llvm
11299 %res = call i16 @llvm.convert.to.fp16.f32(float %a)
11300 store i16 %res, i16* @x, align 2
11302 .. _int_convert_from_fp16:
11304 '``llvm.convert.from.fp16``' Intrinsic
11305 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11312 declare float @llvm.convert.from.fp16.f32(i16 %a)
11313 declare double @llvm.convert.from.fp16.f64(i16 %a)
11318 The '``llvm.convert.from.fp16``' intrinsic function performs a
11319 conversion from half precision floating point format to single precision
11320 floating point format.
11325 The intrinsic function contains single argument - the value to be
11331 The '``llvm.convert.from.fp16``' intrinsic function performs a
11332 conversion from half single precision floating point format to single
11333 precision floating point format. The input half-float value is
11334 represented by an ``i16`` value.
11339 .. code-block:: llvm
11341 %a = load i16, i16* @x, align 2
11342 %res = call float @llvm.convert.from.fp16(i16 %a)
11344 .. _dbg_intrinsics:
11346 Debugger Intrinsics
11347 -------------------
11349 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
11350 prefix), are described in the `LLVM Source Level
11351 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
11354 Exception Handling Intrinsics
11355 -----------------------------
11357 The LLVM exception handling intrinsics (which all start with
11358 ``llvm.eh.`` prefix), are described in the `LLVM Exception
11359 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
11361 .. _int_trampoline:
11363 Trampoline Intrinsics
11364 ---------------------
11366 These intrinsics make it possible to excise one parameter, marked with
11367 the :ref:`nest <nest>` attribute, from a function. The result is a
11368 callable function pointer lacking the nest parameter - the caller does
11369 not need to provide a value for it. Instead, the value to use is stored
11370 in advance in a "trampoline", a block of memory usually allocated on the
11371 stack, which also contains code to splice the nest value into the
11372 argument list. This is used to implement the GCC nested function address
11375 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
11376 then the resulting function pointer has signature ``i32 (i32, i32)*``.
11377 It can be created as follows:
11379 .. code-block:: llvm
11381 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
11382 %tramp1 = getelementptr [10 x i8], [10 x i8]* %tramp, i32 0, i32 0
11383 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
11384 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
11385 %fp = bitcast i8* %p to i32 (i32, i32)*
11387 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
11388 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
11392 '``llvm.init.trampoline``' Intrinsic
11393 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11400 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
11405 This fills the memory pointed to by ``tramp`` with executable code,
11406 turning it into a trampoline.
11411 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
11412 pointers. The ``tramp`` argument must point to a sufficiently large and
11413 sufficiently aligned block of memory; this memory is written to by the
11414 intrinsic. Note that the size and the alignment are target-specific -
11415 LLVM currently provides no portable way of determining them, so a
11416 front-end that generates this intrinsic needs to have some
11417 target-specific knowledge. The ``func`` argument must hold a function
11418 bitcast to an ``i8*``.
11423 The block of memory pointed to by ``tramp`` is filled with target
11424 dependent code, turning it into a function. Then ``tramp`` needs to be
11425 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
11426 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
11427 function's signature is the same as that of ``func`` with any arguments
11428 marked with the ``nest`` attribute removed. At most one such ``nest``
11429 argument is allowed, and it must be of pointer type. Calling the new
11430 function is equivalent to calling ``func`` with the same argument list,
11431 but with ``nval`` used for the missing ``nest`` argument. If, after
11432 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
11433 modified, then the effect of any later call to the returned function
11434 pointer is undefined.
11438 '``llvm.adjust.trampoline``' Intrinsic
11439 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11446 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
11451 This performs any required machine-specific adjustment to the address of
11452 a trampoline (passed as ``tramp``).
11457 ``tramp`` must point to a block of memory which already has trampoline
11458 code filled in by a previous call to
11459 :ref:`llvm.init.trampoline <int_it>`.
11464 On some architectures the address of the code to be executed needs to be
11465 different than the address where the trampoline is actually stored. This
11466 intrinsic returns the executable address corresponding to ``tramp``
11467 after performing the required machine specific adjustments. The pointer
11468 returned can then be :ref:`bitcast and executed <int_trampoline>`.
11470 .. _int_mload_mstore:
11472 Masked Vector Load and Store Intrinsics
11473 ---------------------------------------
11475 LLVM provides intrinsics for predicated vector load and store operations. The predicate is specified by a mask operand, which holds one bit per vector element, switching the associated vector lane on or off. The memory addresses corresponding to the "off" lanes are not accessed. When all bits of the mask are on, the intrinsic is identical to a regular vector load or store. When all bits are off, no memory is accessed.
11479 '``llvm.masked.load.*``' Intrinsics
11480 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11484 This is an overloaded intrinsic. The loaded data is a vector of any integer, floating point or pointer data type.
11488 declare <16 x float> @llvm.masked.load.v16f32 (<16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
11489 declare <2 x double> @llvm.masked.load.v2f64 (<2 x double>* <ptr>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
11490 ;; The data is a vector of pointers to double
11491 declare <8 x double*> @llvm.masked.load.v8p0f64 (<8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>, <8 x double*> <passthru>)
11492 ;; The data is a vector of function pointers
11493 declare <8 x i32 ()*> @llvm.masked.load.v8p0f_i32f (<8 x i32 ()*>* <ptr>, i32 <alignment>, <8 x i1> <mask>, <8 x i32 ()*> <passthru>)
11498 Reads a vector from memory according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes. The masked-off lanes in the result vector are taken from the corresponding lanes of the '``passthru``' operand.
11504 The first operand is the base pointer for the load. The second operand is the alignment of the source location. It must be a constant integer value. The third operand, mask, is a vector of boolean values with the same number of elements as the return type. The fourth is a pass-through value that is used to fill the masked-off lanes of the result. The return type, underlying type of the base pointer and the type of the '``passthru``' operand are the same vector types.
11510 The '``llvm.masked.load``' intrinsic is designed for conditional reading of selected vector elements in a single IR operation. It is useful for targets that support vector masked loads and allows vectorizing predicated basic blocks on these targets. Other targets may support this intrinsic differently, for example by lowering it into a sequence of branches that guard scalar load operations.
11511 The result of this operation is equivalent to a regular vector load instruction followed by a 'select' between the loaded and the passthru values, predicated on the same mask. However, using this intrinsic prevents exceptions on memory access to masked-off lanes.
11516 %res = call <16 x float> @llvm.masked.load.v16f32 (<16 x float>* %ptr, i32 4, <16 x i1>%mask, <16 x float> %passthru)
11518 ;; The result of the two following instructions is identical aside from potential memory access exception
11519 %loadlal = load <16 x float>, <16 x float>* %ptr, align 4
11520 %res = select <16 x i1> %mask, <16 x float> %loadlal, <16 x float> %passthru
11524 '``llvm.masked.store.*``' Intrinsics
11525 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11529 This is an overloaded intrinsic. The data stored in memory is a vector of any integer, floating point or pointer data type.
11533 declare void @llvm.masked.store.v8i32 (<8 x i32> <value>, <8 x i32>* <ptr>, i32 <alignment>, <8 x i1> <mask>)
11534 declare void @llvm.masked.store.v16f32 (<16 x float> <value>, <16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>)
11535 ;; The data is a vector of pointers to double
11536 declare void @llvm.masked.store.v8p0f64 (<8 x double*> <value>, <8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>)
11537 ;; The data is a vector of function pointers
11538 declare void @llvm.masked.store.v4p0f_i32f (<4 x i32 ()*> <value>, <4 x i32 ()*>* <ptr>, i32 <alignment>, <4 x i1> <mask>)
11543 Writes a vector to memory according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes.
11548 The first operand is the vector value to be written to memory. The second operand is the base pointer for the store, it has the same underlying type as the value operand. The third operand is the alignment of the destination location. The fourth operand, mask, is a vector of boolean values. The types of the mask and the value operand must have the same number of vector elements.
11554 The '``llvm.masked.store``' intrinsics is designed for conditional writing of selected vector elements in a single IR operation. It is useful for targets that support vector masked store and allows vectorizing predicated basic blocks on these targets. Other targets may support this intrinsic differently, for example by lowering it into a sequence of branches that guard scalar store operations.
11555 The result of this operation is equivalent to a load-modify-store sequence. However, using this intrinsic prevents exceptions and data races on memory access to masked-off lanes.
11559 call void @llvm.masked.store.v16f32(<16 x float> %value, <16 x float>* %ptr, i32 4, <16 x i1> %mask)
11561 ;; The result of the following instructions is identical aside from potential data races and memory access exceptions
11562 %oldval = load <16 x float>, <16 x float>* %ptr, align 4
11563 %res = select <16 x i1> %mask, <16 x float> %value, <16 x float> %oldval
11564 store <16 x float> %res, <16 x float>* %ptr, align 4
11567 Masked Vector Gather and Scatter Intrinsics
11568 -------------------------------------------
11570 LLVM provides intrinsics for vector gather and scatter operations. They are similar to :ref:`Masked Vector Load and Store <int_mload_mstore>`, except they are designed for arbitrary memory accesses, rather than sequential memory accesses. Gather and scatter also employ a mask operand, which holds one bit per vector element, switching the associated vector lane on or off. The memory addresses corresponding to the "off" lanes are not accessed. When all bits are off, no memory is accessed.
11574 '``llvm.masked.gather.*``' Intrinsics
11575 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11579 This is an overloaded intrinsic. The loaded data are multiple scalar values of any integer, floating point or pointer data type gathered together into one vector.
11583 declare <16 x float> @llvm.masked.gather.v16f32 (<16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
11584 declare <2 x double> @llvm.masked.gather.v2f64 (<2 x double*> <ptrs>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
11585 declare <8 x float*> @llvm.masked.gather.v8p0f32 (<8 x float**> <ptrs>, i32 <alignment>, <8 x i1> <mask>, <8 x float*> <passthru>)
11590 Reads scalar values from arbitrary memory locations and gathers them into one vector. The memory locations are provided in the vector of pointers '``ptrs``'. The memory is accessed according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes. The masked-off lanes in the result vector are taken from the corresponding lanes of the '``passthru``' operand.
11596 The first operand is a vector of pointers which holds all memory addresses to read. The second operand is an alignment of the source addresses. It must be a constant integer value. The third operand, mask, is a vector of boolean values with the same number of elements as the return type. The fourth is a pass-through value that is used to fill the masked-off lanes of the result. The return type, underlying type of the vector of pointers and the type of the '``passthru``' operand are the same vector types.
11602 The '``llvm.masked.gather``' intrinsic is designed for conditional reading of multiple scalar values from arbitrary memory locations in a single IR operation. It is useful for targets that support vector masked gathers and allows vectorizing basic blocks with data and control divergence. Other targets may support this intrinsic differently, for example by lowering it into a sequence of scalar load operations.
11603 The semantics of this operation are equivalent to a sequence of conditional scalar loads with subsequent gathering all loaded values into a single vector. The mask restricts memory access to certain lanes and facilitates vectorization of predicated basic blocks.
11608 %res = call <4 x double> @llvm.masked.gather.v4f64 (<4 x double*> %ptrs, i32 8, <4 x i1>%mask, <4 x double> <true, true, true, true>)
11610 ;; The gather with all-true mask is equivalent to the following instruction sequence
11611 %ptr0 = extractelement <4 x double*> %ptrs, i32 0
11612 %ptr1 = extractelement <4 x double*> %ptrs, i32 1
11613 %ptr2 = extractelement <4 x double*> %ptrs, i32 2
11614 %ptr3 = extractelement <4 x double*> %ptrs, i32 3
11616 %val0 = load double, double* %ptr0, align 8
11617 %val1 = load double, double* %ptr1, align 8
11618 %val2 = load double, double* %ptr2, align 8
11619 %val3 = load double, double* %ptr3, align 8
11621 %vec0 = insertelement <4 x double>undef, %val0, 0
11622 %vec01 = insertelement <4 x double>%vec0, %val1, 1
11623 %vec012 = insertelement <4 x double>%vec01, %val2, 2
11624 %vec0123 = insertelement <4 x double>%vec012, %val3, 3
11628 '``llvm.masked.scatter.*``' Intrinsics
11629 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11633 This is an overloaded intrinsic. The data stored in memory is a vector of any integer, floating point or pointer data type. Each vector element is stored in an arbitrary memory address. Scatter with overlapping addresses is guaranteed to be ordered from least-significant to most-significant element.
11637 declare void @llvm.masked.scatter.v8i32 (<8 x i32> <value>, <8 x i32*> <ptrs>, i32 <alignment>, <8 x i1> <mask>)
11638 declare void @llvm.masked.scatter.v16f32 (<16 x float> <value>, <16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>)
11639 declare void @llvm.masked.scatter.v4p0f64 (<4 x double*> <value>, <4 x double**> <ptrs>, i32 <alignment>, <4 x i1> <mask>)
11644 Writes each element from the value vector to the corresponding memory address. The memory addresses are represented as a vector of pointers. Writing is done according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes.
11649 The first operand is a vector value to be written to memory. The second operand is a vector of pointers, pointing to where the value elements should be stored. It has the same underlying type as the value operand. The third operand is an alignment of the destination addresses. The fourth operand, mask, is a vector of boolean values. The types of the mask and the value operand must have the same number of vector elements.
11655 The '``llvm.masked.scatter``' intrinsics is designed for writing selected vector elements to arbitrary memory addresses in a single IR operation. The operation may be conditional, when not all bits in the mask are switched on. It is useful for targets that support vector masked scatter and allows vectorizing basic blocks with data and control divergence. Other targets may support this intrinsic differently, for example by lowering it into a sequence of branches that guard scalar store operations.
11659 ;; This instruction unconditionaly stores data vector in multiple addresses
11660 call @llvm.masked.scatter.v8i32 (<8 x i32> %value, <8 x i32*> %ptrs, i32 4, <8 x i1> <true, true, .. true>)
11662 ;; It is equivalent to a list of scalar stores
11663 %val0 = extractelement <8 x i32> %value, i32 0
11664 %val1 = extractelement <8 x i32> %value, i32 1
11666 %val7 = extractelement <8 x i32> %value, i32 7
11667 %ptr0 = extractelement <8 x i32*> %ptrs, i32 0
11668 %ptr1 = extractelement <8 x i32*> %ptrs, i32 1
11670 %ptr7 = extractelement <8 x i32*> %ptrs, i32 7
11671 ;; Note: the order of the following stores is important when they overlap:
11672 store i32 %val0, i32* %ptr0, align 4
11673 store i32 %val1, i32* %ptr1, align 4
11675 store i32 %val7, i32* %ptr7, align 4
11681 This class of intrinsics provides information about the lifetime of
11682 memory objects and ranges where variables are immutable.
11686 '``llvm.lifetime.start``' Intrinsic
11687 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11694 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
11699 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
11705 The first argument is a constant integer representing the size of the
11706 object, or -1 if it is variable sized. The second argument is a pointer
11712 This intrinsic indicates that before this point in the code, the value
11713 of the memory pointed to by ``ptr`` is dead. This means that it is known
11714 to never be used and has an undefined value. A load from the pointer
11715 that precedes this intrinsic can be replaced with ``'undef'``.
11719 '``llvm.lifetime.end``' Intrinsic
11720 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11727 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
11732 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
11738 The first argument is a constant integer representing the size of the
11739 object, or -1 if it is variable sized. The second argument is a pointer
11745 This intrinsic indicates that after this point in the code, the value of
11746 the memory pointed to by ``ptr`` is dead. This means that it is known to
11747 never be used and has an undefined value. Any stores into the memory
11748 object following this intrinsic may be removed as dead.
11750 '``llvm.invariant.start``' Intrinsic
11751 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11758 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
11763 The '``llvm.invariant.start``' intrinsic specifies that the contents of
11764 a memory object will not change.
11769 The first argument is a constant integer representing the size of the
11770 object, or -1 if it is variable sized. The second argument is a pointer
11776 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
11777 the return value, the referenced memory location is constant and
11780 '``llvm.invariant.end``' Intrinsic
11781 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11788 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
11793 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
11794 memory object are mutable.
11799 The first argument is the matching ``llvm.invariant.start`` intrinsic.
11800 The second argument is a constant integer representing the size of the
11801 object, or -1 if it is variable sized and the third argument is a
11802 pointer to the object.
11807 This intrinsic indicates that the memory is mutable again.
11809 '``llvm.invariant.group.barrier``' Intrinsic
11810 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11817 declare i8* @llvm.invariant.group.barrier(i8* <ptr>)
11822 The '``llvm.invariant.group.barrier``' intrinsic can be used when an invariant
11823 established by invariant.group metadata no longer holds, to obtain a new pointer
11824 value that does not carry the invariant information.
11830 The ``llvm.invariant.group.barrier`` takes only one argument, which is
11831 the pointer to the memory for which the ``invariant.group`` no longer holds.
11836 Returns another pointer that aliases its argument but which is considered different
11837 for the purposes of ``load``/``store`` ``invariant.group`` metadata.
11842 This class of intrinsics is designed to be generic and has no specific
11845 '``llvm.var.annotation``' Intrinsic
11846 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11853 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
11858 The '``llvm.var.annotation``' intrinsic.
11863 The first argument is a pointer to a value, the second is a pointer to a
11864 global string, the third is a pointer to a global string which is the
11865 source file name, and the last argument is the line number.
11870 This intrinsic allows annotation of local variables with arbitrary
11871 strings. This can be useful for special purpose optimizations that want
11872 to look for these annotations. These have no other defined use; they are
11873 ignored by code generation and optimization.
11875 '``llvm.ptr.annotation.*``' Intrinsic
11876 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11881 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
11882 pointer to an integer of any width. *NOTE* you must specify an address space for
11883 the pointer. The identifier for the default address space is the integer
11888 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
11889 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
11890 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
11891 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
11892 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
11897 The '``llvm.ptr.annotation``' intrinsic.
11902 The first argument is a pointer to an integer value of arbitrary bitwidth
11903 (result of some expression), the second is a pointer to a global string, the
11904 third is a pointer to a global string which is the source file name, and the
11905 last argument is the line number. It returns the value of the first argument.
11910 This intrinsic allows annotation of a pointer to an integer with arbitrary
11911 strings. This can be useful for special purpose optimizations that want to look
11912 for these annotations. These have no other defined use; they are ignored by code
11913 generation and optimization.
11915 '``llvm.annotation.*``' Intrinsic
11916 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11921 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
11922 any integer bit width.
11926 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
11927 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
11928 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
11929 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
11930 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
11935 The '``llvm.annotation``' intrinsic.
11940 The first argument is an integer value (result of some expression), the
11941 second is a pointer to a global string, the third is a pointer to a
11942 global string which is the source file name, and the last argument is
11943 the line number. It returns the value of the first argument.
11948 This intrinsic allows annotations to be put on arbitrary expressions
11949 with arbitrary strings. This can be useful for special purpose
11950 optimizations that want to look for these annotations. These have no
11951 other defined use; they are ignored by code generation and optimization.
11953 '``llvm.trap``' Intrinsic
11954 ^^^^^^^^^^^^^^^^^^^^^^^^^
11961 declare void @llvm.trap() noreturn nounwind
11966 The '``llvm.trap``' intrinsic.
11976 This intrinsic is lowered to the target dependent trap instruction. If
11977 the target does not have a trap instruction, this intrinsic will be
11978 lowered to a call of the ``abort()`` function.
11980 '``llvm.debugtrap``' Intrinsic
11981 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11988 declare void @llvm.debugtrap() nounwind
11993 The '``llvm.debugtrap``' intrinsic.
12003 This intrinsic is lowered to code which is intended to cause an
12004 execution trap with the intention of requesting the attention of a
12007 '``llvm.stackprotector``' Intrinsic
12008 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12015 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
12020 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
12021 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
12022 is placed on the stack before local variables.
12027 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
12028 The first argument is the value loaded from the stack guard
12029 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
12030 enough space to hold the value of the guard.
12035 This intrinsic causes the prologue/epilogue inserter to force the position of
12036 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
12037 to ensure that if a local variable on the stack is overwritten, it will destroy
12038 the value of the guard. When the function exits, the guard on the stack is
12039 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
12040 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
12041 calling the ``__stack_chk_fail()`` function.
12043 '``llvm.stackprotectorcheck``' Intrinsic
12044 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12051 declare void @llvm.stackprotectorcheck(i8** <guard>)
12056 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
12057 created stack protector and if they are not equal calls the
12058 ``__stack_chk_fail()`` function.
12063 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
12064 the variable ``@__stack_chk_guard``.
12069 This intrinsic is provided to perform the stack protector check by comparing
12070 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
12071 values do not match call the ``__stack_chk_fail()`` function.
12073 The reason to provide this as an IR level intrinsic instead of implementing it
12074 via other IR operations is that in order to perform this operation at the IR
12075 level without an intrinsic, one would need to create additional basic blocks to
12076 handle the success/failure cases. This makes it difficult to stop the stack
12077 protector check from disrupting sibling tail calls in Codegen. With this
12078 intrinsic, we are able to generate the stack protector basic blocks late in
12079 codegen after the tail call decision has occurred.
12081 '``llvm.objectsize``' Intrinsic
12082 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12089 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
12090 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
12095 The ``llvm.objectsize`` intrinsic is designed to provide information to
12096 the optimizers to determine at compile time whether a) an operation
12097 (like memcpy) will overflow a buffer that corresponds to an object, or
12098 b) that a runtime check for overflow isn't necessary. An object in this
12099 context means an allocation of a specific class, structure, array, or
12105 The ``llvm.objectsize`` intrinsic takes two arguments. The first
12106 argument is a pointer to or into the ``object``. The second argument is
12107 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
12108 or -1 (if false) when the object size is unknown. The second argument
12109 only accepts constants.
12114 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
12115 the size of the object concerned. If the size cannot be determined at
12116 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
12117 on the ``min`` argument).
12119 '``llvm.expect``' Intrinsic
12120 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
12125 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
12130 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
12131 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
12132 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
12137 The ``llvm.expect`` intrinsic provides information about expected (the
12138 most probable) value of ``val``, which can be used by optimizers.
12143 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
12144 a value. The second argument is an expected value, this needs to be a
12145 constant value, variables are not allowed.
12150 This intrinsic is lowered to the ``val``.
12154 '``llvm.assume``' Intrinsic
12155 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12162 declare void @llvm.assume(i1 %cond)
12167 The ``llvm.assume`` allows the optimizer to assume that the provided
12168 condition is true. This information can then be used in simplifying other parts
12174 The condition which the optimizer may assume is always true.
12179 The intrinsic allows the optimizer to assume that the provided condition is
12180 always true whenever the control flow reaches the intrinsic call. No code is
12181 generated for this intrinsic, and instructions that contribute only to the
12182 provided condition are not used for code generation. If the condition is
12183 violated during execution, the behavior is undefined.
12185 Note that the optimizer might limit the transformations performed on values
12186 used by the ``llvm.assume`` intrinsic in order to preserve the instructions
12187 only used to form the intrinsic's input argument. This might prove undesirable
12188 if the extra information provided by the ``llvm.assume`` intrinsic does not cause
12189 sufficient overall improvement in code quality. For this reason,
12190 ``llvm.assume`` should not be used to document basic mathematical invariants
12191 that the optimizer can otherwise deduce or facts that are of little use to the
12196 '``llvm.bitset.test``' Intrinsic
12197 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12204 declare i1 @llvm.bitset.test(i8* %ptr, metadata %bitset) nounwind readnone
12210 The first argument is a pointer to be tested. The second argument is a
12211 metadata object representing an identifier for a :doc:`bitset <BitSets>`.
12216 The ``llvm.bitset.test`` intrinsic tests whether the given pointer is a
12217 member of the given bitset.
12219 '``llvm.donothing``' Intrinsic
12220 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12227 declare void @llvm.donothing() nounwind readnone
12232 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's one of only
12233 two intrinsics (besides ``llvm.experimental.patchpoint``) that can be called
12234 with an invoke instruction.
12244 This intrinsic does nothing, and it's removed by optimizers and ignored
12247 Stack Map Intrinsics
12248 --------------------
12250 LLVM provides experimental intrinsics to support runtime patching
12251 mechanisms commonly desired in dynamic language JITs. These intrinsics
12252 are described in :doc:`StackMaps`.