1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
5 <title>LLVM Assembly Language Reference Manual</title>
6 <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
7 <meta name="author" content="Chris Lattner">
8 <meta name="description"
9 content="LLVM Assembly Language Reference Manual.">
10 <link rel="stylesheet" href="_static/llvm.css" type="text/css">
15 <h1>LLVM Language Reference Manual</h1>
17 <li><a href="#abstract">Abstract</a></li>
18 <li><a href="#introduction">Introduction</a></li>
19 <li><a href="#identifiers">Identifiers</a></li>
20 <li><a href="#highlevel">High Level Structure</a>
22 <li><a href="#modulestructure">Module Structure</a></li>
23 <li><a href="#linkage">Linkage Types</a>
25 <li><a href="#linkage_private">'<tt>private</tt>' Linkage</a></li>
26 <li><a href="#linkage_linker_private">'<tt>linker_private</tt>' Linkage</a></li>
27 <li><a href="#linkage_linker_private_weak">'<tt>linker_private_weak</tt>' Linkage</a></li>
28 <li><a href="#linkage_linker_private_weak_def_auto">'<tt>linker_private_weak_def_auto</tt>' Linkage</a></li>
29 <li><a href="#linkage_internal">'<tt>internal</tt>' Linkage</a></li>
30 <li><a href="#linkage_available_externally">'<tt>available_externally</tt>' Linkage</a></li>
31 <li><a href="#linkage_linkonce">'<tt>linkonce</tt>' Linkage</a></li>
32 <li><a href="#linkage_common">'<tt>common</tt>' Linkage</a></li>
33 <li><a href="#linkage_weak">'<tt>weak</tt>' Linkage</a></li>
34 <li><a href="#linkage_appending">'<tt>appending</tt>' Linkage</a></li>
35 <li><a href="#linkage_externweak">'<tt>extern_weak</tt>' Linkage</a></li>
36 <li><a href="#linkage_linkonce_odr">'<tt>linkonce_odr</tt>' Linkage</a></li>
37 <li><a href="#linkage_weak">'<tt>weak_odr</tt>' Linkage</a></li>
38 <li><a href="#linkage_external">'<tt>external</tt>' Linkage</a></li>
39 <li><a href="#linkage_dllimport">'<tt>dllimport</tt>' Linkage</a></li>
40 <li><a href="#linkage_dllexport">'<tt>dllexport</tt>' Linkage</a></li>
43 <li><a href="#callingconv">Calling Conventions</a></li>
44 <li><a href="#namedtypes">Named Types</a></li>
45 <li><a href="#globalvars">Global Variables</a></li>
46 <li><a href="#functionstructure">Functions</a></li>
47 <li><a href="#aliasstructure">Aliases</a></li>
48 <li><a href="#namedmetadatastructure">Named Metadata</a></li>
49 <li><a href="#paramattrs">Parameter Attributes</a></li>
50 <li><a href="#fnattrs">Function Attributes</a></li>
51 <li><a href="#gc">Garbage Collector Names</a></li>
52 <li><a href="#moduleasm">Module-Level Inline Assembly</a></li>
53 <li><a href="#datalayout">Data Layout</a></li>
54 <li><a href="#pointeraliasing">Pointer Aliasing Rules</a></li>
55 <li><a href="#volatile">Volatile Memory Accesses</a></li>
56 <li><a href="#memmodel">Memory Model for Concurrent Operations</a></li>
57 <li><a href="#ordering">Atomic Memory Ordering Constraints</a></li>
60 <li><a href="#typesystem">Type System</a>
62 <li><a href="#t_classifications">Type Classifications</a></li>
63 <li><a href="#t_primitive">Primitive Types</a>
65 <li><a href="#t_integer">Integer Type</a></li>
66 <li><a href="#t_floating">Floating Point Types</a></li>
67 <li><a href="#t_x86mmx">X86mmx Type</a></li>
68 <li><a href="#t_void">Void Type</a></li>
69 <li><a href="#t_label">Label Type</a></li>
70 <li><a href="#t_metadata">Metadata Type</a></li>
73 <li><a href="#t_derived">Derived Types</a>
75 <li><a href="#t_aggregate">Aggregate Types</a>
77 <li><a href="#t_array">Array Type</a></li>
78 <li><a href="#t_struct">Structure Type</a></li>
79 <li><a href="#t_opaque">Opaque Structure Types</a></li>
80 <li><a href="#t_vector">Vector Type</a></li>
83 <li><a href="#t_function">Function Type</a></li>
84 <li><a href="#t_pointer">Pointer Type</a></li>
89 <li><a href="#constants">Constants</a>
91 <li><a href="#simpleconstants">Simple Constants</a></li>
92 <li><a href="#complexconstants">Complex Constants</a></li>
93 <li><a href="#globalconstants">Global Variable and Function Addresses</a></li>
94 <li><a href="#undefvalues">Undefined Values</a></li>
95 <li><a href="#poisonvalues">Poison Values</a></li>
96 <li><a href="#blockaddress">Addresses of Basic Blocks</a></li>
97 <li><a href="#constantexprs">Constant Expressions</a></li>
100 <li><a href="#othervalues">Other Values</a>
102 <li><a href="#inlineasm">Inline Assembler Expressions</a></li>
103 <li><a href="#metadata">Metadata Nodes and Metadata Strings</a>
105 <li><a href="#tbaa">'<tt>tbaa</tt>' Metadata</a></li>
106 <li><a href="#fpmath">'<tt>fpmath</tt>' Metadata</a></li>
107 <li><a href="#range">'<tt>range</tt>' Metadata</a></li>
112 <li><a href="#module_flags">Module Flags Metadata</a>
114 <li><a href="#objc_gc_flags">Objective-C Garbage Collection Module Flags Metadata</a></li>
117 <li><a href="#intrinsic_globals">Intrinsic Global Variables</a>
119 <li><a href="#intg_used">The '<tt>llvm.used</tt>' Global Variable</a></li>
120 <li><a href="#intg_compiler_used">The '<tt>llvm.compiler.used</tt>'
121 Global Variable</a></li>
122 <li><a href="#intg_global_ctors">The '<tt>llvm.global_ctors</tt>'
123 Global Variable</a></li>
124 <li><a href="#intg_global_dtors">The '<tt>llvm.global_dtors</tt>'
125 Global Variable</a></li>
128 <li><a href="#instref">Instruction Reference</a>
130 <li><a href="#terminators">Terminator Instructions</a>
132 <li><a href="#i_ret">'<tt>ret</tt>' Instruction</a></li>
133 <li><a href="#i_br">'<tt>br</tt>' Instruction</a></li>
134 <li><a href="#i_switch">'<tt>switch</tt>' Instruction</a></li>
135 <li><a href="#i_indirectbr">'<tt>indirectbr</tt>' Instruction</a></li>
136 <li><a href="#i_invoke">'<tt>invoke</tt>' Instruction</a></li>
137 <li><a href="#i_resume">'<tt>resume</tt>' Instruction</a></li>
138 <li><a href="#i_unreachable">'<tt>unreachable</tt>' Instruction</a></li>
141 <li><a href="#binaryops">Binary Operations</a>
143 <li><a href="#i_add">'<tt>add</tt>' Instruction</a></li>
144 <li><a href="#i_fadd">'<tt>fadd</tt>' Instruction</a></li>
145 <li><a href="#i_sub">'<tt>sub</tt>' Instruction</a></li>
146 <li><a href="#i_fsub">'<tt>fsub</tt>' Instruction</a></li>
147 <li><a href="#i_mul">'<tt>mul</tt>' Instruction</a></li>
148 <li><a href="#i_fmul">'<tt>fmul</tt>' Instruction</a></li>
149 <li><a href="#i_udiv">'<tt>udiv</tt>' Instruction</a></li>
150 <li><a href="#i_sdiv">'<tt>sdiv</tt>' Instruction</a></li>
151 <li><a href="#i_fdiv">'<tt>fdiv</tt>' Instruction</a></li>
152 <li><a href="#i_urem">'<tt>urem</tt>' Instruction</a></li>
153 <li><a href="#i_srem">'<tt>srem</tt>' Instruction</a></li>
154 <li><a href="#i_frem">'<tt>frem</tt>' Instruction</a></li>
157 <li><a href="#bitwiseops">Bitwise Binary Operations</a>
159 <li><a href="#i_shl">'<tt>shl</tt>' Instruction</a></li>
160 <li><a href="#i_lshr">'<tt>lshr</tt>' Instruction</a></li>
161 <li><a href="#i_ashr">'<tt>ashr</tt>' Instruction</a></li>
162 <li><a href="#i_and">'<tt>and</tt>' Instruction</a></li>
163 <li><a href="#i_or">'<tt>or</tt>' Instruction</a></li>
164 <li><a href="#i_xor">'<tt>xor</tt>' Instruction</a></li>
167 <li><a href="#vectorops">Vector Operations</a>
169 <li><a href="#i_extractelement">'<tt>extractelement</tt>' Instruction</a></li>
170 <li><a href="#i_insertelement">'<tt>insertelement</tt>' Instruction</a></li>
171 <li><a href="#i_shufflevector">'<tt>shufflevector</tt>' Instruction</a></li>
174 <li><a href="#aggregateops">Aggregate Operations</a>
176 <li><a href="#i_extractvalue">'<tt>extractvalue</tt>' Instruction</a></li>
177 <li><a href="#i_insertvalue">'<tt>insertvalue</tt>' Instruction</a></li>
180 <li><a href="#memoryops">Memory Access and Addressing Operations</a>
182 <li><a href="#i_alloca">'<tt>alloca</tt>' Instruction</a></li>
183 <li><a href="#i_load">'<tt>load</tt>' Instruction</a></li>
184 <li><a href="#i_store">'<tt>store</tt>' Instruction</a></li>
185 <li><a href="#i_fence">'<tt>fence</tt>' Instruction</a></li>
186 <li><a href="#i_cmpxchg">'<tt>cmpxchg</tt>' Instruction</a></li>
187 <li><a href="#i_atomicrmw">'<tt>atomicrmw</tt>' Instruction</a></li>
188 <li><a href="#i_getelementptr">'<tt>getelementptr</tt>' Instruction</a></li>
191 <li><a href="#convertops">Conversion Operations</a>
193 <li><a href="#i_trunc">'<tt>trunc .. to</tt>' Instruction</a></li>
194 <li><a href="#i_zext">'<tt>zext .. to</tt>' Instruction</a></li>
195 <li><a href="#i_sext">'<tt>sext .. to</tt>' Instruction</a></li>
196 <li><a href="#i_fptrunc">'<tt>fptrunc .. to</tt>' Instruction</a></li>
197 <li><a href="#i_fpext">'<tt>fpext .. to</tt>' Instruction</a></li>
198 <li><a href="#i_fptoui">'<tt>fptoui .. to</tt>' Instruction</a></li>
199 <li><a href="#i_fptosi">'<tt>fptosi .. to</tt>' Instruction</a></li>
200 <li><a href="#i_uitofp">'<tt>uitofp .. to</tt>' Instruction</a></li>
201 <li><a href="#i_sitofp">'<tt>sitofp .. to</tt>' Instruction</a></li>
202 <li><a href="#i_ptrtoint">'<tt>ptrtoint .. to</tt>' Instruction</a></li>
203 <li><a href="#i_inttoptr">'<tt>inttoptr .. to</tt>' Instruction</a></li>
204 <li><a href="#i_bitcast">'<tt>bitcast .. to</tt>' Instruction</a></li>
207 <li><a href="#otherops">Other Operations</a>
209 <li><a href="#i_icmp">'<tt>icmp</tt>' Instruction</a></li>
210 <li><a href="#i_fcmp">'<tt>fcmp</tt>' Instruction</a></li>
211 <li><a href="#i_phi">'<tt>phi</tt>' Instruction</a></li>
212 <li><a href="#i_select">'<tt>select</tt>' Instruction</a></li>
213 <li><a href="#i_call">'<tt>call</tt>' Instruction</a></li>
214 <li><a href="#i_va_arg">'<tt>va_arg</tt>' Instruction</a></li>
215 <li><a href="#i_landingpad">'<tt>landingpad</tt>' Instruction</a></li>
220 <li><a href="#intrinsics">Intrinsic Functions</a>
222 <li><a href="#int_varargs">Variable Argument Handling Intrinsics</a>
224 <li><a href="#int_va_start">'<tt>llvm.va_start</tt>' Intrinsic</a></li>
225 <li><a href="#int_va_end">'<tt>llvm.va_end</tt>' Intrinsic</a></li>
226 <li><a href="#int_va_copy">'<tt>llvm.va_copy</tt>' Intrinsic</a></li>
229 <li><a href="#int_gc">Accurate Garbage Collection Intrinsics</a>
231 <li><a href="#int_gcroot">'<tt>llvm.gcroot</tt>' Intrinsic</a></li>
232 <li><a href="#int_gcread">'<tt>llvm.gcread</tt>' Intrinsic</a></li>
233 <li><a href="#int_gcwrite">'<tt>llvm.gcwrite</tt>' Intrinsic</a></li>
236 <li><a href="#int_codegen">Code Generator Intrinsics</a>
238 <li><a href="#int_returnaddress">'<tt>llvm.returnaddress</tt>' Intrinsic</a></li>
239 <li><a href="#int_frameaddress">'<tt>llvm.frameaddress</tt>' Intrinsic</a></li>
240 <li><a href="#int_stacksave">'<tt>llvm.stacksave</tt>' Intrinsic</a></li>
241 <li><a href="#int_stackrestore">'<tt>llvm.stackrestore</tt>' Intrinsic</a></li>
242 <li><a href="#int_prefetch">'<tt>llvm.prefetch</tt>' Intrinsic</a></li>
243 <li><a href="#int_pcmarker">'<tt>llvm.pcmarker</tt>' Intrinsic</a></li>
244 <li><a href="#int_readcyclecounter">'<tt>llvm.readcyclecounter</tt>' Intrinsic</a></li>
247 <li><a href="#int_libc">Standard C Library Intrinsics</a>
249 <li><a href="#int_memcpy">'<tt>llvm.memcpy.*</tt>' Intrinsic</a></li>
250 <li><a href="#int_memmove">'<tt>llvm.memmove.*</tt>' Intrinsic</a></li>
251 <li><a href="#int_memset">'<tt>llvm.memset.*</tt>' Intrinsic</a></li>
252 <li><a href="#int_sqrt">'<tt>llvm.sqrt.*</tt>' Intrinsic</a></li>
253 <li><a href="#int_powi">'<tt>llvm.powi.*</tt>' Intrinsic</a></li>
254 <li><a href="#int_sin">'<tt>llvm.sin.*</tt>' Intrinsic</a></li>
255 <li><a href="#int_cos">'<tt>llvm.cos.*</tt>' Intrinsic</a></li>
256 <li><a href="#int_pow">'<tt>llvm.pow.*</tt>' Intrinsic</a></li>
257 <li><a href="#int_exp">'<tt>llvm.exp.*</tt>' Intrinsic</a></li>
258 <li><a href="#int_log">'<tt>llvm.log.*</tt>' Intrinsic</a></li>
259 <li><a href="#int_fma">'<tt>llvm.fma.*</tt>' Intrinsic</a></li>
260 <li><a href="#int_fabs">'<tt>llvm.fabs.*</tt>' Intrinsic</a></li>
263 <li><a href="#int_manip">Bit Manipulation Intrinsics</a>
265 <li><a href="#int_bswap">'<tt>llvm.bswap.*</tt>' Intrinsics</a></li>
266 <li><a href="#int_ctpop">'<tt>llvm.ctpop.*</tt>' Intrinsic </a></li>
267 <li><a href="#int_ctlz">'<tt>llvm.ctlz.*</tt>' Intrinsic </a></li>
268 <li><a href="#int_cttz">'<tt>llvm.cttz.*</tt>' Intrinsic </a></li>
271 <li><a href="#int_overflow">Arithmetic with Overflow Intrinsics</a>
273 <li><a href="#int_sadd_overflow">'<tt>llvm.sadd.with.overflow.*</tt> Intrinsics</a></li>
274 <li><a href="#int_uadd_overflow">'<tt>llvm.uadd.with.overflow.*</tt> Intrinsics</a></li>
275 <li><a href="#int_ssub_overflow">'<tt>llvm.ssub.with.overflow.*</tt> Intrinsics</a></li>
276 <li><a href="#int_usub_overflow">'<tt>llvm.usub.with.overflow.*</tt> Intrinsics</a></li>
277 <li><a href="#int_smul_overflow">'<tt>llvm.smul.with.overflow.*</tt> Intrinsics</a></li>
278 <li><a href="#int_umul_overflow">'<tt>llvm.umul.with.overflow.*</tt> Intrinsics</a></li>
281 <li><a href="#spec_arithmetic">Specialised Arithmetic Intrinsics</a>
283 <li><a href="#fmuladd">'<tt>llvm.fmuladd</tt> Intrinsic</a></li>
286 <li><a href="#int_fp16">Half Precision Floating Point Intrinsics</a>
288 <li><a href="#int_convert_to_fp16">'<tt>llvm.convert.to.fp16</tt>' Intrinsic</a></li>
289 <li><a href="#int_convert_from_fp16">'<tt>llvm.convert.from.fp16</tt>' Intrinsic</a></li>
292 <li><a href="#int_debugger">Debugger intrinsics</a></li>
293 <li><a href="#int_eh">Exception Handling intrinsics</a></li>
294 <li><a href="#int_trampoline">Trampoline Intrinsics</a>
296 <li><a href="#int_it">'<tt>llvm.init.trampoline</tt>' Intrinsic</a></li>
297 <li><a href="#int_at">'<tt>llvm.adjust.trampoline</tt>' Intrinsic</a></li>
300 <li><a href="#int_memorymarkers">Memory Use Markers</a>
302 <li><a href="#int_lifetime_start">'<tt>llvm.lifetime.start</tt>' Intrinsic</a></li>
303 <li><a href="#int_lifetime_end">'<tt>llvm.lifetime.end</tt>' Intrinsic</a></li>
304 <li><a href="#int_invariant_start">'<tt>llvm.invariant.start</tt>' Intrinsic</a></li>
305 <li><a href="#int_invariant_end">'<tt>llvm.invariant.end</tt>' Intrinsic</a></li>
308 <li><a href="#int_general">General intrinsics</a>
310 <li><a href="#int_var_annotation">
311 '<tt>llvm.var.annotation</tt>' Intrinsic</a></li>
312 <li><a href="#int_annotation">
313 '<tt>llvm.annotation.*</tt>' Intrinsic</a></li>
314 <li><a href="#int_trap">
315 '<tt>llvm.trap</tt>' Intrinsic</a></li>
316 <li><a href="#int_debugtrap">
317 '<tt>llvm.debugtrap</tt>' Intrinsic</a></li>
318 <li><a href="#int_stackprotector">
319 '<tt>llvm.stackprotector</tt>' Intrinsic</a></li>
320 <li><a href="#int_objectsize">
321 '<tt>llvm.objectsize</tt>' Intrinsic</a></li>
322 <li><a href="#int_expect">
323 '<tt>llvm.expect</tt>' Intrinsic</a></li>
330 <div class="doc_author">
331 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a>
332 and <a href="mailto:vadve@cs.uiuc.edu">Vikram Adve</a></p>
335 <!-- *********************************************************************** -->
336 <h2><a name="abstract">Abstract</a></h2>
337 <!-- *********************************************************************** -->
341 <p>This document is a reference manual for the LLVM assembly language. LLVM is
342 a Static Single Assignment (SSA) based representation that provides type
343 safety, low-level operations, flexibility, and the capability of representing
344 'all' high-level languages cleanly. It is the common code representation
345 used throughout all phases of the LLVM compilation strategy.</p>
349 <!-- *********************************************************************** -->
350 <h2><a name="introduction">Introduction</a></h2>
351 <!-- *********************************************************************** -->
355 <p>The LLVM code representation is designed to be used in three different forms:
356 as an in-memory compiler IR, as an on-disk bitcode representation (suitable
357 for fast loading by a Just-In-Time compiler), and as a human readable
358 assembly language representation. This allows LLVM to provide a powerful
359 intermediate representation for efficient compiler transformations and
360 analysis, while providing a natural means to debug and visualize the
361 transformations. The three different forms of LLVM are all equivalent. This
362 document describes the human readable representation and notation.</p>
364 <p>The LLVM representation aims to be light-weight and low-level while being
365 expressive, typed, and extensible at the same time. It aims to be a
366 "universal IR" of sorts, by being at a low enough level that high-level ideas
367 may be cleanly mapped to it (similar to how microprocessors are "universal
368 IR's", allowing many source languages to be mapped to them). By providing
369 type information, LLVM can be used as the target of optimizations: for
370 example, through pointer analysis, it can be proven that a C automatic
371 variable is never accessed outside of the current function, allowing it to
372 be promoted to a simple SSA value instead of a memory location.</p>
374 <!-- _______________________________________________________________________ -->
376 <a name="wellformed">Well-Formedness</a>
381 <p>It is important to note that this document describes 'well formed' LLVM
382 assembly language. There is a difference between what the parser accepts and
383 what is considered 'well formed'. For example, the following instruction is
384 syntactically okay, but not well formed:</p>
386 <pre class="doc_code">
387 %x = <a href="#i_add">add</a> i32 1, %x
390 <p>because the definition of <tt>%x</tt> does not dominate all of its uses. The
391 LLVM infrastructure provides a verification pass that may be used to verify
392 that an LLVM module is well formed. This pass is automatically run by the
393 parser after parsing input assembly and by the optimizer before it outputs
394 bitcode. The violations pointed out by the verifier pass indicate bugs in
395 transformation passes or input to the parser.</p>
401 <!-- Describe the typesetting conventions here. -->
403 <!-- *********************************************************************** -->
404 <h2><a name="identifiers">Identifiers</a></h2>
405 <!-- *********************************************************************** -->
409 <p>LLVM identifiers come in two basic types: global and local. Global
410 identifiers (functions, global variables) begin with the <tt>'@'</tt>
411 character. Local identifiers (register names, types) begin with
412 the <tt>'%'</tt> character. Additionally, there are three different formats
413 for identifiers, for different purposes:</p>
416 <li>Named values are represented as a string of characters with their prefix.
417 For example, <tt>%foo</tt>, <tt>@DivisionByZero</tt>,
418 <tt>%a.really.long.identifier</tt>. The actual regular expression used is
419 '<tt>[%@][a-zA-Z$._][a-zA-Z$._0-9]*</tt>'. Identifiers which require
420 other characters in their names can be surrounded with quotes. Special
421 characters may be escaped using <tt>"\xx"</tt> where <tt>xx</tt> is the
422 ASCII code for the character in hexadecimal. In this way, any character
423 can be used in a name value, even quotes themselves.</li>
425 <li>Unnamed values are represented as an unsigned numeric value with their
426 prefix. For example, <tt>%12</tt>, <tt>@2</tt>, <tt>%44</tt>.</li>
428 <li>Constants, which are described in a <a href="#constants">section about
429 constants</a>, below.</li>
432 <p>LLVM requires that values start with a prefix for two reasons: Compilers
433 don't need to worry about name clashes with reserved words, and the set of
434 reserved words may be expanded in the future without penalty. Additionally,
435 unnamed identifiers allow a compiler to quickly come up with a temporary
436 variable without having to avoid symbol table conflicts.</p>
438 <p>Reserved words in LLVM are very similar to reserved words in other
439 languages. There are keywords for different opcodes
440 ('<tt><a href="#i_add">add</a></tt>',
441 '<tt><a href="#i_bitcast">bitcast</a></tt>',
442 '<tt><a href="#i_ret">ret</a></tt>', etc...), for primitive type names
443 ('<tt><a href="#t_void">void</a></tt>',
444 '<tt><a href="#t_primitive">i32</a></tt>', etc...), and others. These
445 reserved words cannot conflict with variable names, because none of them
446 start with a prefix character (<tt>'%'</tt> or <tt>'@'</tt>).</p>
448 <p>Here is an example of LLVM code to multiply the integer variable
449 '<tt>%X</tt>' by 8:</p>
453 <pre class="doc_code">
454 %result = <a href="#i_mul">mul</a> i32 %X, 8
457 <p>After strength reduction:</p>
459 <pre class="doc_code">
460 %result = <a href="#i_shl">shl</a> i32 %X, i8 3
463 <p>And the hard way:</p>
465 <pre class="doc_code">
466 %0 = <a href="#i_add">add</a> i32 %X, %X <i>; yields {i32}:%0</i>
467 %1 = <a href="#i_add">add</a> i32 %0, %0 <i>; yields {i32}:%1</i>
468 %result = <a href="#i_add">add</a> i32 %1, %1
471 <p>This last way of multiplying <tt>%X</tt> by 8 illustrates several important
472 lexical features of LLVM:</p>
475 <li>Comments are delimited with a '<tt>;</tt>' and go until the end of
478 <li>Unnamed temporaries are created when the result of a computation is not
479 assigned to a named value.</li>
481 <li>Unnamed temporaries are numbered sequentially</li>
484 <p>It also shows a convention that we follow in this document. When
485 demonstrating instructions, we will follow an instruction with a comment that
486 defines the type and name of value produced. Comments are shown in italic
491 <!-- *********************************************************************** -->
492 <h2><a name="highlevel">High Level Structure</a></h2>
493 <!-- *********************************************************************** -->
495 <!-- ======================================================================= -->
497 <a name="modulestructure">Module Structure</a>
502 <p>LLVM programs are composed of <tt>Module</tt>s, each of which is a
503 translation unit of the input programs. Each module consists of functions,
504 global variables, and symbol table entries. Modules may be combined together
505 with the LLVM linker, which merges function (and global variable)
506 definitions, resolves forward declarations, and merges symbol table
507 entries. Here is an example of the "hello world" module:</p>
509 <pre class="doc_code">
510 <i>; Declare the string constant as a global constant.</i>
511 <a href="#identifiers">@.str</a> = <a href="#linkage_private">private</a> <a href="#globalvars">unnamed_addr</a> <a href="#globalvars">constant</a> <a href="#t_array">[13 x i8]</a> c"hello world\0A\00"
513 <i>; External declaration of the puts function</i>
514 <a href="#functionstructure">declare</a> i32 @puts(i8* <a href="#nocapture">nocapture</a>) <a href="#fnattrs">nounwind</a>
516 <i>; Definition of main function</i>
517 define i32 @main() { <i>; i32()* </i>
518 <i>; Convert [13 x i8]* to i8 *...</i>
519 %cast210 = <a href="#i_getelementptr">getelementptr</a> [13 x i8]* @.str, i64 0, i64 0
521 <i>; Call puts function to write out the string to stdout.</i>
522 <a href="#i_call">call</a> i32 @puts(i8* %cast210)
523 <a href="#i_ret">ret</a> i32 0
526 <i>; Named metadata</i>
527 !1 = metadata !{i32 42}
531 <p>This example is made up of a <a href="#globalvars">global variable</a> named
532 "<tt>.str</tt>", an external declaration of the "<tt>puts</tt>" function,
533 a <a href="#functionstructure">function definition</a> for
534 "<tt>main</tt>" and <a href="#namedmetadatastructure">named metadata</a>
537 <p>In general, a module is made up of a list of global values (where both
538 functions and global variables are global values). Global values are
539 represented by a pointer to a memory location (in this case, a pointer to an
540 array of char, and a pointer to a function), and have one of the
541 following <a href="#linkage">linkage types</a>.</p>
545 <!-- ======================================================================= -->
547 <a name="linkage">Linkage Types</a>
552 <p>All Global Variables and Functions have one of the following types of
556 <dt><tt><b><a name="linkage_private">private</a></b></tt></dt>
557 <dd>Global values with "<tt>private</tt>" linkage are only directly accessible
558 by objects in the current module. In particular, linking code into a
559 module with an private global value may cause the private to be renamed as
560 necessary to avoid collisions. Because the symbol is private to the
561 module, all references can be updated. This doesn't show up in any symbol
562 table in the object file.</dd>
564 <dt><tt><b><a name="linkage_linker_private">linker_private</a></b></tt></dt>
565 <dd>Similar to <tt>private</tt>, but the symbol is passed through the
566 assembler and evaluated by the linker. Unlike normal strong symbols, they
567 are removed by the linker from the final linked image (executable or
568 dynamic library).</dd>
570 <dt><tt><b><a name="linkage_linker_private_weak">linker_private_weak</a></b></tt></dt>
571 <dd>Similar to "<tt>linker_private</tt>", but the symbol is weak. Note that
572 <tt>linker_private_weak</tt> symbols are subject to coalescing by the
573 linker. The symbols are removed by the linker from the final linked image
574 (executable or dynamic library).</dd>
576 <dt><tt><b><a name="linkage_linker_private_weak_def_auto">linker_private_weak_def_auto</a></b></tt></dt>
577 <dd>Similar to "<tt>linker_private_weak</tt>", but it's known that the address
578 of the object is not taken. For instance, functions that had an inline
579 definition, but the compiler decided not to inline it. Note,
580 unlike <tt>linker_private</tt> and <tt>linker_private_weak</tt>,
581 <tt>linker_private_weak_def_auto</tt> may have only <tt>default</tt>
582 visibility. The symbols are removed by the linker from the final linked
583 image (executable or dynamic library).</dd>
585 <dt><tt><b><a name="linkage_internal">internal</a></b></tt></dt>
586 <dd>Similar to private, but the value shows as a local symbol
587 (<tt>STB_LOCAL</tt> in the case of ELF) in the object file. This
588 corresponds to the notion of the '<tt>static</tt>' keyword in C.</dd>
590 <dt><tt><b><a name="linkage_available_externally">available_externally</a></b></tt></dt>
591 <dd>Globals with "<tt>available_externally</tt>" linkage are never emitted
592 into the object file corresponding to the LLVM module. They exist to
593 allow inlining and other optimizations to take place given knowledge of
594 the definition of the global, which is known to be somewhere outside the
595 module. Globals with <tt>available_externally</tt> linkage are allowed to
596 be discarded at will, and are otherwise the same as <tt>linkonce_odr</tt>.
597 This linkage type is only allowed on definitions, not declarations.</dd>
599 <dt><tt><b><a name="linkage_linkonce">linkonce</a></b></tt></dt>
600 <dd>Globals with "<tt>linkonce</tt>" linkage are merged with other globals of
601 the same name when linkage occurs. This can be used to implement
602 some forms of inline functions, templates, or other code which must be
603 generated in each translation unit that uses it, but where the body may
604 be overridden with a more definitive definition later. Unreferenced
605 <tt>linkonce</tt> globals are allowed to be discarded. Note that
606 <tt>linkonce</tt> linkage does not actually allow the optimizer to
607 inline the body of this function into callers because it doesn't know if
608 this definition of the function is the definitive definition within the
609 program or whether it will be overridden by a stronger definition.
610 To enable inlining and other optimizations, use "<tt>linkonce_odr</tt>"
613 <dt><tt><b><a name="linkage_weak">weak</a></b></tt></dt>
614 <dd>"<tt>weak</tt>" linkage has the same merging semantics as
615 <tt>linkonce</tt> linkage, except that unreferenced globals with
616 <tt>weak</tt> linkage may not be discarded. This is used for globals that
617 are declared "weak" in C source code.</dd>
619 <dt><tt><b><a name="linkage_common">common</a></b></tt></dt>
620 <dd>"<tt>common</tt>" linkage is most similar to "<tt>weak</tt>" linkage, but
621 they are used for tentative definitions in C, such as "<tt>int X;</tt>" at
623 Symbols with "<tt>common</tt>" linkage are merged in the same way as
624 <tt>weak symbols</tt>, and they may not be deleted if unreferenced.
625 <tt>common</tt> symbols may not have an explicit section,
626 must have a zero initializer, and may not be marked '<a
627 href="#globalvars"><tt>constant</tt></a>'. Functions and aliases may not
628 have common linkage.</dd>
631 <dt><tt><b><a name="linkage_appending">appending</a></b></tt></dt>
632 <dd>"<tt>appending</tt>" linkage may only be applied to global variables of
633 pointer to array type. When two global variables with appending linkage
634 are linked together, the two global arrays are appended together. This is
635 the LLVM, typesafe, equivalent of having the system linker append together
636 "sections" with identical names when .o files are linked.</dd>
638 <dt><tt><b><a name="linkage_externweak">extern_weak</a></b></tt></dt>
639 <dd>The semantics of this linkage follow the ELF object file model: the symbol
640 is weak until linked, if not linked, the symbol becomes null instead of
641 being an undefined reference.</dd>
643 <dt><tt><b><a name="linkage_linkonce_odr">linkonce_odr</a></b></tt></dt>
644 <dt><tt><b><a name="linkage_weak_odr">weak_odr</a></b></tt></dt>
645 <dd>Some languages allow differing globals to be merged, such as two functions
646 with different semantics. Other languages, such as <tt>C++</tt>, ensure
647 that only equivalent globals are ever merged (the "one definition rule"
648 — "ODR"). Such languages can use the <tt>linkonce_odr</tt>
649 and <tt>weak_odr</tt> linkage types to indicate that the global will only
650 be merged with equivalent globals. These linkage types are otherwise the
651 same as their non-<tt>odr</tt> versions.</dd>
653 <dt><tt><b><a name="linkage_external">external</a></b></tt></dt>
654 <dd>If none of the above identifiers are used, the global is externally
655 visible, meaning that it participates in linkage and can be used to
656 resolve external symbol references.</dd>
659 <p>The next two types of linkage are targeted for Microsoft Windows platform
660 only. They are designed to support importing (exporting) symbols from (to)
661 DLLs (Dynamic Link Libraries).</p>
664 <dt><tt><b><a name="linkage_dllimport">dllimport</a></b></tt></dt>
665 <dd>"<tt>dllimport</tt>" linkage causes the compiler to reference a function
666 or variable via a global pointer to a pointer that is set up by the DLL
667 exporting the symbol. On Microsoft Windows targets, the pointer name is
668 formed by combining <code>__imp_</code> and the function or variable
671 <dt><tt><b><a name="linkage_dllexport">dllexport</a></b></tt></dt>
672 <dd>"<tt>dllexport</tt>" linkage causes the compiler to provide a global
673 pointer to a pointer in a DLL, so that it can be referenced with the
674 <tt>dllimport</tt> attribute. On Microsoft Windows targets, the pointer
675 name is formed by combining <code>__imp_</code> and the function or
679 <p>For example, since the "<tt>.LC0</tt>" variable is defined to be internal, if
680 another module defined a "<tt>.LC0</tt>" variable and was linked with this
681 one, one of the two would be renamed, preventing a collision. Since
682 "<tt>main</tt>" and "<tt>puts</tt>" are external (i.e., lacking any linkage
683 declarations), they are accessible outside of the current module.</p>
685 <p>It is illegal for a function <i>declaration</i> to have any linkage type
686 other than <tt>external</tt>, <tt>dllimport</tt>
687 or <tt>extern_weak</tt>.</p>
689 <p>Aliases can have only <tt>external</tt>, <tt>internal</tt>, <tt>weak</tt>
690 or <tt>weak_odr</tt> linkages.</p>
694 <!-- ======================================================================= -->
696 <a name="callingconv">Calling Conventions</a>
701 <p>LLVM <a href="#functionstructure">functions</a>, <a href="#i_call">calls</a>
702 and <a href="#i_invoke">invokes</a> can all have an optional calling
703 convention specified for the call. The calling convention of any pair of
704 dynamic caller/callee must match, or the behavior of the program is
705 undefined. The following calling conventions are supported by LLVM, and more
706 may be added in the future:</p>
709 <dt><b>"<tt>ccc</tt>" - The C calling convention</b>:</dt>
710 <dd>This calling convention (the default if no other calling convention is
711 specified) matches the target C calling conventions. This calling
712 convention supports varargs function calls and tolerates some mismatch in
713 the declared prototype and implemented declaration of the function (as
716 <dt><b>"<tt>fastcc</tt>" - The fast calling convention</b>:</dt>
717 <dd>This calling convention attempts to make calls as fast as possible
718 (e.g. by passing things in registers). This calling convention allows the
719 target to use whatever tricks it wants to produce fast code for the
720 target, without having to conform to an externally specified ABI
721 (Application Binary Interface).
722 <a href="CodeGenerator.html#tailcallopt">Tail calls can only be optimized
723 when this or the GHC convention is used.</a> This calling convention
724 does not support varargs and requires the prototype of all callees to
725 exactly match the prototype of the function definition.</dd>
727 <dt><b>"<tt>coldcc</tt>" - The cold calling convention</b>:</dt>
728 <dd>This calling convention attempts to make code in the caller as efficient
729 as possible under the assumption that the call is not commonly executed.
730 As such, these calls often preserve all registers so that the call does
731 not break any live ranges in the caller side. This calling convention
732 does not support varargs and requires the prototype of all callees to
733 exactly match the prototype of the function definition.</dd>
735 <dt><b>"<tt>cc <em>10</em></tt>" - GHC convention</b>:</dt>
736 <dd>This calling convention has been implemented specifically for use by the
737 <a href="http://www.haskell.org/ghc">Glasgow Haskell Compiler (GHC)</a>.
738 It passes everything in registers, going to extremes to achieve this by
739 disabling callee save registers. This calling convention should not be
740 used lightly but only for specific situations such as an alternative to
741 the <em>register pinning</em> performance technique often used when
742 implementing functional programming languages.At the moment only X86
743 supports this convention and it has the following limitations:
745 <li>On <em>X86-32</em> only supports up to 4 bit type parameters. No
746 floating point types are supported.</li>
747 <li>On <em>X86-64</em> only supports up to 10 bit type parameters and
748 6 floating point parameters.</li>
750 This calling convention supports
751 <a href="CodeGenerator.html#tailcallopt">tail call optimization</a> but
752 requires both the caller and callee are using it.
755 <dt><b>"<tt>cc <<em>n</em>></tt>" - Numbered convention</b>:</dt>
756 <dd>Any calling convention may be specified by number, allowing
757 target-specific calling conventions to be used. Target specific calling
758 conventions start at 64.</dd>
761 <p>More calling conventions can be added/defined on an as-needed basis, to
762 support Pascal conventions or any other well-known target-independent
767 <!-- ======================================================================= -->
769 <a name="visibility">Visibility Styles</a>
774 <p>All Global Variables and Functions have one of the following visibility
778 <dt><b>"<tt>default</tt>" - Default style</b>:</dt>
779 <dd>On targets that use the ELF object file format, default visibility means
780 that the declaration is visible to other modules and, in shared libraries,
781 means that the declared entity may be overridden. On Darwin, default
782 visibility means that the declaration is visible to other modules. Default
783 visibility corresponds to "external linkage" in the language.</dd>
785 <dt><b>"<tt>hidden</tt>" - Hidden style</b>:</dt>
786 <dd>Two declarations of an object with hidden visibility refer to the same
787 object if they are in the same shared object. Usually, hidden visibility
788 indicates that the symbol will not be placed into the dynamic symbol
789 table, so no other module (executable or shared library) can reference it
792 <dt><b>"<tt>protected</tt>" - Protected style</b>:</dt>
793 <dd>On ELF, protected visibility indicates that the symbol will be placed in
794 the dynamic symbol table, but that references within the defining module
795 will bind to the local symbol. That is, the symbol cannot be overridden by
801 <!-- ======================================================================= -->
803 <a name="namedtypes">Named Types</a>
808 <p>LLVM IR allows you to specify name aliases for certain types. This can make
809 it easier to read the IR and make the IR more condensed (particularly when
810 recursive types are involved). An example of a name specification is:</p>
812 <pre class="doc_code">
813 %mytype = type { %mytype*, i32 }
816 <p>You may give a name to any <a href="#typesystem">type</a> except
817 "<a href="#t_void">void</a>". Type name aliases may be used anywhere a type
818 is expected with the syntax "%mytype".</p>
820 <p>Note that type names are aliases for the structural type that they indicate,
821 and that you can therefore specify multiple names for the same type. This
822 often leads to confusing behavior when dumping out a .ll file. Since LLVM IR
823 uses structural typing, the name is not part of the type. When printing out
824 LLVM IR, the printer will pick <em>one name</em> to render all types of a
825 particular shape. This means that if you have code where two different
826 source types end up having the same LLVM type, that the dumper will sometimes
827 print the "wrong" or unexpected type. This is an important design point and
828 isn't going to change.</p>
832 <!-- ======================================================================= -->
834 <a name="globalvars">Global Variables</a>
839 <p>Global variables define regions of memory allocated at compilation time
840 instead of run-time. Global variables may optionally be initialized, may
841 have an explicit section to be placed in, and may have an optional explicit
842 alignment specified.</p>
844 <p>A variable may be defined as <tt>thread_local</tt>, which
845 means that it will not be shared by threads (each thread will have a
846 separated copy of the variable). Not all targets support thread-local
847 variables. Optionally, a TLS model may be specified:</p>
850 <dt><b><tt>localdynamic</tt></b>:</dt>
851 <dd>For variables that are only used within the current shared library.</dd>
853 <dt><b><tt>initialexec</tt></b>:</dt>
854 <dd>For variables in modules that will not be loaded dynamically.</dd>
856 <dt><b><tt>localexec</tt></b>:</dt>
857 <dd>For variables defined in the executable and only used within it.</dd>
860 <p>The models correspond to the ELF TLS models; see
861 <a href="http://people.redhat.com/drepper/tls.pdf">ELF
862 Handling For Thread-Local Storage</a> for more information on under which
863 circumstances the different models may be used. The target may choose a
864 different TLS model if the specified model is not supported, or if a better
865 choice of model can be made.</p>
867 <p>A variable may be defined as a global
868 "constant," which indicates that the contents of the variable
869 will <b>never</b> be modified (enabling better optimization, allowing the
870 global data to be placed in the read-only section of an executable, etc).
871 Note that variables that need runtime initialization cannot be marked
872 "constant" as there is a store to the variable.</p>
874 <p>LLVM explicitly allows <em>declarations</em> of global variables to be marked
875 constant, even if the final definition of the global is not. This capability
876 can be used to enable slightly better optimization of the program, but
877 requires the language definition to guarantee that optimizations based on the
878 'constantness' are valid for the translation units that do not include the
881 <p>As SSA values, global variables define pointer values that are in scope
882 (i.e. they dominate) all basic blocks in the program. Global variables
883 always define a pointer to their "content" type because they describe a
884 region of memory, and all memory objects in LLVM are accessed through
887 <p>Global variables can be marked with <tt>unnamed_addr</tt> which indicates
888 that the address is not significant, only the content. Constants marked
889 like this can be merged with other constants if they have the same
890 initializer. Note that a constant with significant address <em>can</em>
891 be merged with a <tt>unnamed_addr</tt> constant, the result being a
892 constant whose address is significant.</p>
894 <p>A global variable may be declared to reside in a target-specific numbered
895 address space. For targets that support them, address spaces may affect how
896 optimizations are performed and/or what target instructions are used to
897 access the variable. The default address space is zero. The address space
898 qualifier must precede any other attributes.</p>
900 <p>LLVM allows an explicit section to be specified for globals. If the target
901 supports it, it will emit globals to the section specified.</p>
903 <p>An explicit alignment may be specified for a global, which must be a power
904 of 2. If not present, or if the alignment is set to zero, the alignment of
905 the global is set by the target to whatever it feels convenient. If an
906 explicit alignment is specified, the global is forced to have exactly that
907 alignment. Targets and optimizers are not allowed to over-align the global
908 if the global has an assigned section. In this case, the extra alignment
909 could be observable: for example, code could assume that the globals are
910 densely packed in their section and try to iterate over them as an array,
911 alignment padding would break this iteration.</p>
913 <p>For example, the following defines a global in a numbered address space with
914 an initializer, section, and alignment:</p>
916 <pre class="doc_code">
917 @G = addrspace(5) constant float 1.0, section "foo", align 4
920 <p>The following example defines a thread-local global with
921 the <tt>initialexec</tt> TLS model:</p>
923 <pre class="doc_code">
924 @G = thread_local(initialexec) global i32 0, align 4
930 <!-- ======================================================================= -->
932 <a name="functionstructure">Functions</a>
937 <p>LLVM function definitions consist of the "<tt>define</tt>" keyword, an
938 optional <a href="#linkage">linkage type</a>, an optional
939 <a href="#visibility">visibility style</a>, an optional
940 <a href="#callingconv">calling convention</a>,
941 an optional <tt>unnamed_addr</tt> attribute, a return type, an optional
942 <a href="#paramattrs">parameter attribute</a> for the return type, a function
943 name, a (possibly empty) argument list (each with optional
944 <a href="#paramattrs">parameter attributes</a>), optional
945 <a href="#fnattrs">function attributes</a>, an optional section, an optional
946 alignment, an optional <a href="#gc">garbage collector name</a>, an opening
947 curly brace, a list of basic blocks, and a closing curly brace.</p>
949 <p>LLVM function declarations consist of the "<tt>declare</tt>" keyword, an
950 optional <a href="#linkage">linkage type</a>, an optional
951 <a href="#visibility">visibility style</a>, an optional
952 <a href="#callingconv">calling convention</a>,
953 an optional <tt>unnamed_addr</tt> attribute, a return type, an optional
954 <a href="#paramattrs">parameter attribute</a> for the return type, a function
955 name, a possibly empty list of arguments, an optional alignment, and an
956 optional <a href="#gc">garbage collector name</a>.</p>
958 <p>A function definition contains a list of basic blocks, forming the CFG
959 (Control Flow Graph) for the function. Each basic block may optionally start
960 with a label (giving the basic block a symbol table entry), contains a list
961 of instructions, and ends with a <a href="#terminators">terminator</a>
962 instruction (such as a branch or function return).</p>
964 <p>The first basic block in a function is special in two ways: it is immediately
965 executed on entrance to the function, and it is not allowed to have
966 predecessor basic blocks (i.e. there can not be any branches to the entry
967 block of a function). Because the block can have no predecessors, it also
968 cannot have any <a href="#i_phi">PHI nodes</a>.</p>
970 <p>LLVM allows an explicit section to be specified for functions. If the target
971 supports it, it will emit functions to the section specified.</p>
973 <p>An explicit alignment may be specified for a function. If not present, or if
974 the alignment is set to zero, the alignment of the function is set by the
975 target to whatever it feels convenient. If an explicit alignment is
976 specified, the function is forced to have at least that much alignment. All
977 alignments must be a power of 2.</p>
979 <p>If the <tt>unnamed_addr</tt> attribute is given, the address is know to not
980 be significant and two identical functions can be merged.</p>
983 <pre class="doc_code">
984 define [<a href="#linkage">linkage</a>] [<a href="#visibility">visibility</a>]
985 [<a href="#callingconv">cconv</a>] [<a href="#paramattrs">ret attrs</a>]
986 <ResultType> @<FunctionName> ([argument list])
987 [<a href="#fnattrs">fn Attrs</a>] [section "name"] [align N]
988 [<a href="#gc">gc</a>] { ... }
993 <!-- ======================================================================= -->
995 <a name="aliasstructure">Aliases</a>
1000 <p>Aliases act as "second name" for the aliasee value (which can be either
1001 function, global variable, another alias or bitcast of global value). Aliases
1002 may have an optional <a href="#linkage">linkage type</a>, and an
1003 optional <a href="#visibility">visibility style</a>.</p>
1006 <pre class="doc_code">
1007 @<Name> = alias [Linkage] [Visibility] <AliaseeTy> @<Aliasee>
1012 <!-- ======================================================================= -->
1014 <a name="namedmetadatastructure">Named Metadata</a>
1019 <p>Named metadata is a collection of metadata. <a href="#metadata">Metadata
1020 nodes</a> (but not metadata strings) are the only valid operands for
1021 a named metadata.</p>
1024 <pre class="doc_code">
1025 ; Some unnamed metadata nodes, which are referenced by the named metadata.
1026 !0 = metadata !{metadata !"zero"}
1027 !1 = metadata !{metadata !"one"}
1028 !2 = metadata !{metadata !"two"}
1030 !name = !{!0, !1, !2}
1035 <!-- ======================================================================= -->
1037 <a name="paramattrs">Parameter Attributes</a>
1042 <p>The return type and each parameter of a function type may have a set of
1043 <i>parameter attributes</i> associated with them. Parameter attributes are
1044 used to communicate additional information about the result or parameters of
1045 a function. Parameter attributes are considered to be part of the function,
1046 not of the function type, so functions with different parameter attributes
1047 can have the same function type.</p>
1049 <p>Parameter attributes are simple keywords that follow the type specified. If
1050 multiple parameter attributes are needed, they are space separated. For
1053 <pre class="doc_code">
1054 declare i32 @printf(i8* noalias nocapture, ...)
1055 declare i32 @atoi(i8 zeroext)
1056 declare signext i8 @returns_signed_char()
1059 <p>Note that any attributes for the function result (<tt>nounwind</tt>,
1060 <tt>readonly</tt>) come immediately after the argument list.</p>
1062 <p>Currently, only the following parameter attributes are defined:</p>
1065 <dt><tt><b>zeroext</b></tt></dt>
1066 <dd>This indicates to the code generator that the parameter or return value
1067 should be zero-extended to the extent required by the target's ABI (which
1068 is usually 32-bits, but is 8-bits for a i1 on x86-64) by the caller (for a
1069 parameter) or the callee (for a return value).</dd>
1071 <dt><tt><b>signext</b></tt></dt>
1072 <dd>This indicates to the code generator that the parameter or return value
1073 should be sign-extended to the extent required by the target's ABI (which
1074 is usually 32-bits) by the caller (for a parameter) or the callee (for a
1077 <dt><tt><b>inreg</b></tt></dt>
1078 <dd>This indicates that this parameter or return value should be treated in a
1079 special target-dependent fashion during while emitting code for a function
1080 call or return (usually, by putting it in a register as opposed to memory,
1081 though some targets use it to distinguish between two different kinds of
1082 registers). Use of this attribute is target-specific.</dd>
1084 <dt><tt><b><a name="byval">byval</a></b></tt></dt>
1085 <dd><p>This indicates that the pointer parameter should really be passed by
1086 value to the function. The attribute implies that a hidden copy of the
1088 is made between the caller and the callee, so the callee is unable to
1089 modify the value in the caller. This attribute is only valid on LLVM
1090 pointer arguments. It is generally used to pass structs and arrays by
1091 value, but is also valid on pointers to scalars. The copy is considered
1092 to belong to the caller not the callee (for example,
1093 <tt><a href="#readonly">readonly</a></tt> functions should not write to
1094 <tt>byval</tt> parameters). This is not a valid attribute for return
1097 <p>The byval attribute also supports specifying an alignment with
1098 the align attribute. It indicates the alignment of the stack slot to
1099 form and the known alignment of the pointer specified to the call site. If
1100 the alignment is not specified, then the code generator makes a
1101 target-specific assumption.</p></dd>
1103 <dt><tt><b><a name="sret">sret</a></b></tt></dt>
1104 <dd>This indicates that the pointer parameter specifies the address of a
1105 structure that is the return value of the function in the source program.
1106 This pointer must be guaranteed by the caller to be valid: loads and
1107 stores to the structure may be assumed by the callee to not to trap. This
1108 may only be applied to the first parameter. This is not a valid attribute
1109 for return values. </dd>
1111 <dt><tt><b><a name="noalias">noalias</a></b></tt></dt>
1112 <dd>This indicates that pointer values
1113 <a href="#pointeraliasing"><i>based</i></a> on the argument or return
1114 value do not alias pointer values which are not <i>based</i> on it,
1115 ignoring certain "irrelevant" dependencies.
1116 For a call to the parent function, dependencies between memory
1117 references from before or after the call and from those during the call
1118 are "irrelevant" to the <tt>noalias</tt> keyword for the arguments and
1119 return value used in that call.
1120 The caller shares the responsibility with the callee for ensuring that
1121 these requirements are met.
1122 For further details, please see the discussion of the NoAlias response in
1123 <a href="AliasAnalysis.html#MustMayNo">alias analysis</a>.<br>
1125 Note that this definition of <tt>noalias</tt> is intentionally
1126 similar to the definition of <tt>restrict</tt> in C99 for function
1127 arguments, though it is slightly weaker.
1129 For function return values, C99's <tt>restrict</tt> is not meaningful,
1130 while LLVM's <tt>noalias</tt> is.
1133 <dt><tt><b><a name="nocapture">nocapture</a></b></tt></dt>
1134 <dd>This indicates that the callee does not make any copies of the pointer
1135 that outlive the callee itself. This is not a valid attribute for return
1138 <dt><tt><b><a name="nest">nest</a></b></tt></dt>
1139 <dd>This indicates that the pointer parameter can be excised using the
1140 <a href="#int_trampoline">trampoline intrinsics</a>. This is not a valid
1141 attribute for return values.</dd>
1146 <!-- ======================================================================= -->
1148 <a name="gc">Garbage Collector Names</a>
1153 <p>Each function may specify a garbage collector name, which is simply a
1156 <pre class="doc_code">
1157 define void @f() gc "name" { ... }
1160 <p>The compiler declares the supported values of <i>name</i>. Specifying a
1161 collector which will cause the compiler to alter its output in order to
1162 support the named garbage collection algorithm.</p>
1166 <!-- ======================================================================= -->
1168 <a name="fnattrs">Function Attributes</a>
1173 <p>Function attributes are set to communicate additional information about a
1174 function. Function attributes are considered to be part of the function, not
1175 of the function type, so functions with different parameter attributes can
1176 have the same function type.</p>
1178 <p>Function attributes are simple keywords that follow the type specified. If
1179 multiple attributes are needed, they are space separated. For example:</p>
1181 <pre class="doc_code">
1182 define void @f() noinline { ... }
1183 define void @f() alwaysinline { ... }
1184 define void @f() alwaysinline optsize { ... }
1185 define void @f() optsize { ... }
1189 <dt><tt><b>address_safety</b></tt></dt>
1190 <dd>This attribute indicates that the address safety analysis
1191 is enabled for this function. </dd>
1193 <dt><tt><b>alignstack(<<em>n</em>>)</b></tt></dt>
1194 <dd>This attribute indicates that, when emitting the prologue and epilogue,
1195 the backend should forcibly align the stack pointer. Specify the
1196 desired alignment, which must be a power of two, in parentheses.
1198 <dt><tt><b>alwaysinline</b></tt></dt>
1199 <dd>This attribute indicates that the inliner should attempt to inline this
1200 function into callers whenever possible, ignoring any active inlining size
1201 threshold for this caller.</dd>
1203 <dt><tt><b>nonlazybind</b></tt></dt>
1204 <dd>This attribute suppresses lazy symbol binding for the function. This
1205 may make calls to the function faster, at the cost of extra program
1206 startup time if the function is not called during program startup.</dd>
1208 <dt><tt><b>inlinehint</b></tt></dt>
1209 <dd>This attribute indicates that the source code contained a hint that inlining
1210 this function is desirable (such as the "inline" keyword in C/C++). It
1211 is just a hint; it imposes no requirements on the inliner.</dd>
1213 <dt><tt><b>naked</b></tt></dt>
1214 <dd>This attribute disables prologue / epilogue emission for the function.
1215 This can have very system-specific consequences.</dd>
1217 <dt><tt><b>noimplicitfloat</b></tt></dt>
1218 <dd>This attributes disables implicit floating point instructions.</dd>
1220 <dt><tt><b>noinline</b></tt></dt>
1221 <dd>This attribute indicates that the inliner should never inline this
1222 function in any situation. This attribute may not be used together with
1223 the <tt>alwaysinline</tt> attribute.</dd>
1225 <dt><tt><b>noredzone</b></tt></dt>
1226 <dd>This attribute indicates that the code generator should not use a red
1227 zone, even if the target-specific ABI normally permits it.</dd>
1229 <dt><tt><b>noreturn</b></tt></dt>
1230 <dd>This function attribute indicates that the function never returns
1231 normally. This produces undefined behavior at runtime if the function
1232 ever does dynamically return.</dd>
1234 <dt><tt><b>nounwind</b></tt></dt>
1235 <dd>This function attribute indicates that the function never returns with an
1236 unwind or exceptional control flow. If the function does unwind, its
1237 runtime behavior is undefined.</dd>
1239 <dt><tt><b>optsize</b></tt></dt>
1240 <dd>This attribute suggests that optimization passes and code generator passes
1241 make choices that keep the code size of this function low, and otherwise
1242 do optimizations specifically to reduce code size.</dd>
1244 <dt><tt><b>readnone</b></tt></dt>
1245 <dd>This attribute indicates that the function computes its result (or decides
1246 to unwind an exception) based strictly on its arguments, without
1247 dereferencing any pointer arguments or otherwise accessing any mutable
1248 state (e.g. memory, control registers, etc) visible to caller functions.
1249 It does not write through any pointer arguments
1250 (including <tt><a href="#byval">byval</a></tt> arguments) and never
1251 changes any state visible to callers. This means that it cannot unwind
1252 exceptions by calling the <tt>C++</tt> exception throwing methods.</dd>
1254 <dt><tt><b><a name="readonly">readonly</a></b></tt></dt>
1255 <dd>This attribute indicates that the function does not write through any
1256 pointer arguments (including <tt><a href="#byval">byval</a></tt>
1257 arguments) or otherwise modify any state (e.g. memory, control registers,
1258 etc) visible to caller functions. It may dereference pointer arguments
1259 and read state that may be set in the caller. A readonly function always
1260 returns the same value (or unwinds an exception identically) when called
1261 with the same set of arguments and global state. It cannot unwind an
1262 exception by calling the <tt>C++</tt> exception throwing methods.</dd>
1264 <dt><tt><b><a name="returns_twice">returns_twice</a></b></tt></dt>
1265 <dd>This attribute indicates that this function can return twice. The
1266 C <code>setjmp</code> is an example of such a function. The compiler
1267 disables some optimizations (like tail calls) in the caller of these
1270 <dt><tt><b><a name="ssp">ssp</a></b></tt></dt>
1271 <dd>This attribute indicates that the function should emit a stack smashing
1272 protector. It is in the form of a "canary"—a random value placed on
1273 the stack before the local variables that's checked upon return from the
1274 function to see if it has been overwritten. A heuristic is used to
1275 determine if a function needs stack protectors or not.<br>
1277 If a function that has an <tt>ssp</tt> attribute is inlined into a
1278 function that doesn't have an <tt>ssp</tt> attribute, then the resulting
1279 function will have an <tt>ssp</tt> attribute.</dd>
1281 <dt><tt><b>sspreq</b></tt></dt>
1282 <dd>This attribute indicates that the function should <em>always</em> emit a
1283 stack smashing protector. This overrides
1284 the <tt><a href="#ssp">ssp</a></tt> function attribute.<br>
1286 If a function that has an <tt>sspreq</tt> attribute is inlined into a
1287 function that doesn't have an <tt>sspreq</tt> attribute or which has
1288 an <tt>ssp</tt> attribute, then the resulting function will have
1289 an <tt>sspreq</tt> attribute.</dd>
1291 <dt><tt><b><a name="uwtable">uwtable</a></b></tt></dt>
1292 <dd>This attribute indicates that the ABI being targeted requires that
1293 an unwind table entry be produce for this function even if we can
1294 show that no exceptions passes by it. This is normally the case for
1295 the ELF x86-64 abi, but it can be disabled for some compilation
1301 <!-- ======================================================================= -->
1303 <a name="moduleasm">Module-Level Inline Assembly</a>
1308 <p>Modules may contain "module-level inline asm" blocks, which corresponds to
1309 the GCC "file scope inline asm" blocks. These blocks are internally
1310 concatenated by LLVM and treated as a single unit, but may be separated in
1311 the <tt>.ll</tt> file if desired. The syntax is very simple:</p>
1313 <pre class="doc_code">
1314 module asm "inline asm code goes here"
1315 module asm "more can go here"
1318 <p>The strings can contain any character by escaping non-printable characters.
1319 The escape sequence used is simply "\xx" where "xx" is the two digit hex code
1322 <p>The inline asm code is simply printed to the machine code .s file when
1323 assembly code is generated.</p>
1327 <!-- ======================================================================= -->
1329 <a name="datalayout">Data Layout</a>
1334 <p>A module may specify a target specific data layout string that specifies how
1335 data is to be laid out in memory. The syntax for the data layout is
1338 <pre class="doc_code">
1339 target datalayout = "<i>layout specification</i>"
1342 <p>The <i>layout specification</i> consists of a list of specifications
1343 separated by the minus sign character ('-'). Each specification starts with
1344 a letter and may include other information after the letter to define some
1345 aspect of the data layout. The specifications accepted are as follows:</p>
1349 <dd>Specifies that the target lays out data in big-endian form. That is, the
1350 bits with the most significance have the lowest address location.</dd>
1353 <dd>Specifies that the target lays out data in little-endian form. That is,
1354 the bits with the least significance have the lowest address
1357 <dt><tt>S<i>size</i></tt></dt>
1358 <dd>Specifies the natural alignment of the stack in bits. Alignment promotion
1359 of stack variables is limited to the natural stack alignment to avoid
1360 dynamic stack realignment. The stack alignment must be a multiple of
1361 8-bits. If omitted, the natural stack alignment defaults to "unspecified",
1362 which does not prevent any alignment promotions.</dd>
1364 <dt><tt>p:<i>size</i>:<i>abi</i>:<i>pref</i></tt></dt>
1365 <dd>This specifies the <i>size</i> of a pointer and its <i>abi</i> and
1366 <i>preferred</i> alignments. All sizes are in bits. Specifying
1367 the <i>pref</i> alignment is optional. If omitted, the
1368 preceding <tt>:</tt> should be omitted too.</dd>
1370 <dt><tt>i<i>size</i>:<i>abi</i>:<i>pref</i></tt></dt>
1371 <dd>This specifies the alignment for an integer type of a given bit
1372 <i>size</i>. The value of <i>size</i> must be in the range [1,2^23).</dd>
1374 <dt><tt>v<i>size</i>:<i>abi</i>:<i>pref</i></tt></dt>
1375 <dd>This specifies the alignment for a vector type of a given bit
1378 <dt><tt>f<i>size</i>:<i>abi</i>:<i>pref</i></tt></dt>
1379 <dd>This specifies the alignment for a floating point type of a given bit
1380 <i>size</i>. Only values of <i>size</i> that are supported by the target
1381 will work. 32 (float) and 64 (double) are supported on all targets;
1382 80 or 128 (different flavors of long double) are also supported on some
1385 <dt><tt>a<i>size</i>:<i>abi</i>:<i>pref</i></tt></dt>
1386 <dd>This specifies the alignment for an aggregate type of a given bit
1389 <dt><tt>s<i>size</i>:<i>abi</i>:<i>pref</i></tt></dt>
1390 <dd>This specifies the alignment for a stack object of a given bit
1393 <dt><tt>n<i>size1</i>:<i>size2</i>:<i>size3</i>...</tt></dt>
1394 <dd>This specifies a set of native integer widths for the target CPU
1395 in bits. For example, it might contain "n32" for 32-bit PowerPC,
1396 "n32:64" for PowerPC 64, or "n8:16:32:64" for X86-64. Elements of
1397 this set are considered to support most general arithmetic
1398 operations efficiently.</dd>
1401 <p>When constructing the data layout for a given target, LLVM starts with a
1402 default set of specifications which are then (possibly) overridden by the
1403 specifications in the <tt>datalayout</tt> keyword. The default specifications
1404 are given in this list:</p>
1407 <li><tt>E</tt> - big endian</li>
1408 <li><tt>p:64:64:64</tt> - 64-bit pointers with 64-bit alignment</li>
1409 <li><tt>i1:8:8</tt> - i1 is 8-bit (byte) aligned</li>
1410 <li><tt>i8:8:8</tt> - i8 is 8-bit (byte) aligned</li>
1411 <li><tt>i16:16:16</tt> - i16 is 16-bit aligned</li>
1412 <li><tt>i32:32:32</tt> - i32 is 32-bit aligned</li>
1413 <li><tt>i64:32:64</tt> - i64 has ABI alignment of 32-bits but preferred
1414 alignment of 64-bits</li>
1415 <li><tt>f32:32:32</tt> - float is 32-bit aligned</li>
1416 <li><tt>f64:64:64</tt> - double is 64-bit aligned</li>
1417 <li><tt>v64:64:64</tt> - 64-bit vector is 64-bit aligned</li>
1418 <li><tt>v128:128:128</tt> - 128-bit vector is 128-bit aligned</li>
1419 <li><tt>a0:0:1</tt> - aggregates are 8-bit aligned</li>
1420 <li><tt>s0:64:64</tt> - stack objects are 64-bit aligned</li>
1423 <p>When LLVM is determining the alignment for a given type, it uses the
1424 following rules:</p>
1427 <li>If the type sought is an exact match for one of the specifications, that
1428 specification is used.</li>
1430 <li>If no match is found, and the type sought is an integer type, then the
1431 smallest integer type that is larger than the bitwidth of the sought type
1432 is used. If none of the specifications are larger than the bitwidth then
1433 the the largest integer type is used. For example, given the default
1434 specifications above, the i7 type will use the alignment of i8 (next
1435 largest) while both i65 and i256 will use the alignment of i64 (largest
1438 <li>If no match is found, and the type sought is a vector type, then the
1439 largest vector type that is smaller than the sought vector type will be
1440 used as a fall back. This happens because <128 x double> can be
1441 implemented in terms of 64 <2 x double>, for example.</li>
1444 <p>The function of the data layout string may not be what you expect. Notably,
1445 this is not a specification from the frontend of what alignment the code
1446 generator should use.</p>
1448 <p>Instead, if specified, the target data layout is required to match what the
1449 ultimate <em>code generator</em> expects. This string is used by the
1450 mid-level optimizers to
1451 improve code, and this only works if it matches what the ultimate code
1452 generator uses. If you would like to generate IR that does not embed this
1453 target-specific detail into the IR, then you don't have to specify the
1454 string. This will disable some optimizations that require precise layout
1455 information, but this also prevents those optimizations from introducing
1456 target specificity into the IR.</p>
1462 <!-- ======================================================================= -->
1464 <a name="pointeraliasing">Pointer Aliasing Rules</a>
1469 <p>Any memory access must be done through a pointer value associated
1470 with an address range of the memory access, otherwise the behavior
1471 is undefined. Pointer values are associated with address ranges
1472 according to the following rules:</p>
1475 <li>A pointer value is associated with the addresses associated with
1476 any value it is <i>based</i> on.
1477 <li>An address of a global variable is associated with the address
1478 range of the variable's storage.</li>
1479 <li>The result value of an allocation instruction is associated with
1480 the address range of the allocated storage.</li>
1481 <li>A null pointer in the default address-space is associated with
1483 <li>An integer constant other than zero or a pointer value returned
1484 from a function not defined within LLVM may be associated with address
1485 ranges allocated through mechanisms other than those provided by
1486 LLVM. Such ranges shall not overlap with any ranges of addresses
1487 allocated by mechanisms provided by LLVM.</li>
1490 <p>A pointer value is <i>based</i> on another pointer value according
1491 to the following rules:</p>
1494 <li>A pointer value formed from a
1495 <tt><a href="#i_getelementptr">getelementptr</a></tt> operation
1496 is <i>based</i> on the first operand of the <tt>getelementptr</tt>.</li>
1497 <li>The result value of a
1498 <tt><a href="#i_bitcast">bitcast</a></tt> is <i>based</i> on the operand
1499 of the <tt>bitcast</tt>.</li>
1500 <li>A pointer value formed by an
1501 <tt><a href="#i_inttoptr">inttoptr</a></tt> is <i>based</i> on all
1502 pointer values that contribute (directly or indirectly) to the
1503 computation of the pointer's value.</li>
1504 <li>The "<i>based</i> on" relationship is transitive.</li>
1507 <p>Note that this definition of <i>"based"</i> is intentionally
1508 similar to the definition of <i>"based"</i> in C99, though it is
1509 slightly weaker.</p>
1511 <p>LLVM IR does not associate types with memory. The result type of a
1512 <tt><a href="#i_load">load</a></tt> merely indicates the size and
1513 alignment of the memory from which to load, as well as the
1514 interpretation of the value. The first operand type of a
1515 <tt><a href="#i_store">store</a></tt> similarly only indicates the size
1516 and alignment of the store.</p>
1518 <p>Consequently, type-based alias analysis, aka TBAA, aka
1519 <tt>-fstrict-aliasing</tt>, is not applicable to general unadorned
1520 LLVM IR. <a href="#metadata">Metadata</a> may be used to encode
1521 additional information which specialized optimization passes may use
1522 to implement type-based alias analysis.</p>
1526 <!-- ======================================================================= -->
1528 <a name="volatile">Volatile Memory Accesses</a>
1533 <p>Certain memory accesses, such as <a href="#i_load"><tt>load</tt></a>s, <a
1534 href="#i_store"><tt>store</tt></a>s, and <a
1535 href="#int_memcpy"><tt>llvm.memcpy</tt></a>s may be marked <tt>volatile</tt>.
1536 The optimizers must not change the number of volatile operations or change their
1537 order of execution relative to other volatile operations. The optimizers
1538 <i>may</i> change the order of volatile operations relative to non-volatile
1539 operations. This is not Java's "volatile" and has no cross-thread
1540 synchronization behavior.</p>
1544 <!-- ======================================================================= -->
1546 <a name="memmodel">Memory Model for Concurrent Operations</a>
1551 <p>The LLVM IR does not define any way to start parallel threads of execution
1552 or to register signal handlers. Nonetheless, there are platform-specific
1553 ways to create them, and we define LLVM IR's behavior in their presence. This
1554 model is inspired by the C++0x memory model.</p>
1556 <p>For a more informal introduction to this model, see the
1557 <a href="Atomics.html">LLVM Atomic Instructions and Concurrency Guide</a>.
1559 <p>We define a <i>happens-before</i> partial order as the least partial order
1562 <li>Is a superset of single-thread program order, and</li>
1563 <li>When a <i>synchronizes-with</i> <tt>b</tt>, includes an edge from
1564 <tt>a</tt> to <tt>b</tt>. <i>Synchronizes-with</i> pairs are introduced
1565 by platform-specific techniques, like pthread locks, thread
1566 creation, thread joining, etc., and by atomic instructions.
1567 (See also <a href="#ordering">Atomic Memory Ordering Constraints</a>).
1571 <p>Note that program order does not introduce <i>happens-before</i> edges
1572 between a thread and signals executing inside that thread.</p>
1574 <p>Every (defined) read operation (load instructions, memcpy, atomic
1575 loads/read-modify-writes, etc.) <var>R</var> reads a series of bytes written by
1576 (defined) write operations (store instructions, atomic
1577 stores/read-modify-writes, memcpy, etc.). For the purposes of this section,
1578 initialized globals are considered to have a write of the initializer which is
1579 atomic and happens before any other read or write of the memory in question.
1580 For each byte of a read <var>R</var>, <var>R<sub>byte</sub></var> may see
1581 any write to the same byte, except:</p>
1584 <li>If <var>write<sub>1</sub></var> happens before
1585 <var>write<sub>2</sub></var>, and <var>write<sub>2</sub></var> happens
1586 before <var>R<sub>byte</sub></var>, then <var>R<sub>byte</sub></var>
1587 does not see <var>write<sub>1</sub></var>.
1588 <li>If <var>R<sub>byte</sub></var> happens before
1589 <var>write<sub>3</sub></var>, then <var>R<sub>byte</sub></var> does not
1590 see <var>write<sub>3</sub></var>.
1593 <p>Given that definition, <var>R<sub>byte</sub></var> is defined as follows:
1595 <li>If <var>R</var> is volatile, the result is target-dependent. (Volatile
1596 is supposed to give guarantees which can support
1597 <code>sig_atomic_t</code> in C/C++, and may be used for accesses to
1598 addresses which do not behave like normal memory. It does not generally
1599 provide cross-thread synchronization.)
1600 <li>Otherwise, if there is no write to the same byte that happens before
1601 <var>R<sub>byte</sub></var>, <var>R<sub>byte</sub></var> returns
1602 <tt>undef</tt> for that byte.
1603 <li>Otherwise, if <var>R<sub>byte</sub></var> may see exactly one write,
1604 <var>R<sub>byte</sub></var> returns the value written by that
1606 <li>Otherwise, if <var>R</var> is atomic, and all the writes
1607 <var>R<sub>byte</sub></var> may see are atomic, it chooses one of the
1608 values written. See the <a href="#ordering">Atomic Memory Ordering
1609 Constraints</a> section for additional constraints on how the choice
1611 <li>Otherwise <var>R<sub>byte</sub></var> returns <tt>undef</tt>.</li>
1614 <p><var>R</var> returns the value composed of the series of bytes it read.
1615 This implies that some bytes within the value may be <tt>undef</tt>
1616 <b>without</b> the entire value being <tt>undef</tt>. Note that this only
1617 defines the semantics of the operation; it doesn't mean that targets will
1618 emit more than one instruction to read the series of bytes.</p>
1620 <p>Note that in cases where none of the atomic intrinsics are used, this model
1621 places only one restriction on IR transformations on top of what is required
1622 for single-threaded execution: introducing a store to a byte which might not
1623 otherwise be stored is not allowed in general. (Specifically, in the case
1624 where another thread might write to and read from an address, introducing a
1625 store can change a load that may see exactly one write into a load that may
1626 see multiple writes.)</p>
1628 <!-- FIXME: This model assumes all targets where concurrency is relevant have
1629 a byte-size store which doesn't affect adjacent bytes. As far as I can tell,
1630 none of the backends currently in the tree fall into this category; however,
1631 there might be targets which care. If there are, we want a paragraph
1634 Targets may specify that stores narrower than a certain width are not
1635 available; on such a target, for the purposes of this model, treat any
1636 non-atomic write with an alignment or width less than the minimum width
1637 as if it writes to the relevant surrounding bytes.
1642 <!-- ======================================================================= -->
1644 <a name="ordering">Atomic Memory Ordering Constraints</a>
1649 <p>Atomic instructions (<a href="#i_cmpxchg"><code>cmpxchg</code></a>,
1650 <a href="#i_atomicrmw"><code>atomicrmw</code></a>,
1651 <a href="#i_fence"><code>fence</code></a>,
1652 <a href="#i_load"><code>atomic load</code></a>, and
1653 <a href="#i_store"><code>atomic store</code></a>) take an ordering parameter
1654 that determines which other atomic instructions on the same address they
1655 <i>synchronize with</i>. These semantics are borrowed from Java and C++0x,
1656 but are somewhat more colloquial. If these descriptions aren't precise enough,
1657 check those specs (see spec references in the
1658 <a href="Atomics.html#introduction">atomics guide</a>).
1659 <a href="#i_fence"><code>fence</code></a> instructions
1660 treat these orderings somewhat differently since they don't take an address.
1661 See that instruction's documentation for details.</p>
1663 <p>For a simpler introduction to the ordering constraints, see the
1664 <a href="Atomics.html">LLVM Atomic Instructions and Concurrency Guide</a>.</p>
1667 <dt><code>unordered</code></dt>
1668 <dd>The set of values that can be read is governed by the happens-before
1669 partial order. A value cannot be read unless some operation wrote it.
1670 This is intended to provide a guarantee strong enough to model Java's
1671 non-volatile shared variables. This ordering cannot be specified for
1672 read-modify-write operations; it is not strong enough to make them atomic
1673 in any interesting way.</dd>
1674 <dt><code>monotonic</code></dt>
1675 <dd>In addition to the guarantees of <code>unordered</code>, there is a single
1676 total order for modifications by <code>monotonic</code> operations on each
1677 address. All modification orders must be compatible with the happens-before
1678 order. There is no guarantee that the modification orders can be combined to
1679 a global total order for the whole program (and this often will not be
1680 possible). The read in an atomic read-modify-write operation
1681 (<a href="#i_cmpxchg"><code>cmpxchg</code></a> and
1682 <a href="#i_atomicrmw"><code>atomicrmw</code></a>)
1683 reads the value in the modification order immediately before the value it
1684 writes. If one atomic read happens before another atomic read of the same
1685 address, the later read must see the same value or a later value in the
1686 address's modification order. This disallows reordering of
1687 <code>monotonic</code> (or stronger) operations on the same address. If an
1688 address is written <code>monotonic</code>ally by one thread, and other threads
1689 <code>monotonic</code>ally read that address repeatedly, the other threads must
1690 eventually see the write. This corresponds to the C++0x/C1x
1691 <code>memory_order_relaxed</code>.</dd>
1692 <dt><code>acquire</code></dt>
1693 <dd>In addition to the guarantees of <code>monotonic</code>,
1694 a <i>synchronizes-with</i> edge may be formed with a <code>release</code>
1695 operation. This is intended to model C++'s <code>memory_order_acquire</code>.</dd>
1696 <dt><code>release</code></dt>
1697 <dd>In addition to the guarantees of <code>monotonic</code>, if this operation
1698 writes a value which is subsequently read by an <code>acquire</code> operation,
1699 it <i>synchronizes-with</i> that operation. (This isn't a complete
1700 description; see the C++0x definition of a release sequence.) This corresponds
1701 to the C++0x/C1x <code>memory_order_release</code>.</dd>
1702 <dt><code>acq_rel</code> (acquire+release)</dt><dd>Acts as both an
1703 <code>acquire</code> and <code>release</code> operation on its address.
1704 This corresponds to the C++0x/C1x <code>memory_order_acq_rel</code>.</dd>
1705 <dt><code>seq_cst</code> (sequentially consistent)</dt><dd>
1706 <dd>In addition to the guarantees of <code>acq_rel</code>
1707 (<code>acquire</code> for an operation which only reads, <code>release</code>
1708 for an operation which only writes), there is a global total order on all
1709 sequentially-consistent operations on all addresses, which is consistent with
1710 the <i>happens-before</i> partial order and with the modification orders of
1711 all the affected addresses. Each sequentially-consistent read sees the last
1712 preceding write to the same address in this global order. This corresponds
1713 to the C++0x/C1x <code>memory_order_seq_cst</code> and Java volatile.</dd>
1716 <p id="singlethread">If an atomic operation is marked <code>singlethread</code>,
1717 it only <i>synchronizes with</i> or participates in modification and seq_cst
1718 total orderings with other operations running in the same thread (for example,
1719 in signal handlers).</p>
1725 <!-- *********************************************************************** -->
1726 <h2><a name="typesystem">Type System</a></h2>
1727 <!-- *********************************************************************** -->
1731 <p>The LLVM type system is one of the most important features of the
1732 intermediate representation. Being typed enables a number of optimizations
1733 to be performed on the intermediate representation directly, without having
1734 to do extra analyses on the side before the transformation. A strong type
1735 system makes it easier to read the generated code and enables novel analyses
1736 and transformations that are not feasible to perform on normal three address
1737 code representations.</p>
1739 <!-- ======================================================================= -->
1741 <a name="t_classifications">Type Classifications</a>
1746 <p>The types fall into a few useful classifications:</p>
1748 <table border="1" cellspacing="0" cellpadding="4">
1750 <tr><th>Classification</th><th>Types</th></tr>
1752 <td><a href="#t_integer">integer</a></td>
1753 <td><tt>i1, i2, i3, ... i8, ... i16, ... i32, ... i64, ... </tt></td>
1756 <td><a href="#t_floating">floating point</a></td>
1757 <td><tt>half, float, double, x86_fp80, fp128, ppc_fp128</tt></td>
1760 <td><a name="t_firstclass">first class</a></td>
1761 <td><a href="#t_integer">integer</a>,
1762 <a href="#t_floating">floating point</a>,
1763 <a href="#t_pointer">pointer</a>,
1764 <a href="#t_vector">vector</a>,
1765 <a href="#t_struct">structure</a>,
1766 <a href="#t_array">array</a>,
1767 <a href="#t_label">label</a>,
1768 <a href="#t_metadata">metadata</a>.
1772 <td><a href="#t_primitive">primitive</a></td>
1773 <td><a href="#t_label">label</a>,
1774 <a href="#t_void">void</a>,
1775 <a href="#t_integer">integer</a>,
1776 <a href="#t_floating">floating point</a>,
1777 <a href="#t_x86mmx">x86mmx</a>,
1778 <a href="#t_metadata">metadata</a>.</td>
1781 <td><a href="#t_derived">derived</a></td>
1782 <td><a href="#t_array">array</a>,
1783 <a href="#t_function">function</a>,
1784 <a href="#t_pointer">pointer</a>,
1785 <a href="#t_struct">structure</a>,
1786 <a href="#t_vector">vector</a>,
1787 <a href="#t_opaque">opaque</a>.
1793 <p>The <a href="#t_firstclass">first class</a> types are perhaps the most
1794 important. Values of these types are the only ones which can be produced by
1799 <!-- ======================================================================= -->
1801 <a name="t_primitive">Primitive Types</a>
1806 <p>The primitive types are the fundamental building blocks of the LLVM
1809 <!-- _______________________________________________________________________ -->
1811 <a name="t_integer">Integer Type</a>
1817 <p>The integer type is a very simple type that simply specifies an arbitrary
1818 bit width for the integer type desired. Any bit width from 1 bit to
1819 2<sup>23</sup>-1 (about 8 million) can be specified.</p>
1826 <p>The number of bits the integer will occupy is specified by the <tt>N</tt>
1830 <table class="layout">
1832 <td class="left"><tt>i1</tt></td>
1833 <td class="left">a single-bit integer.</td>
1836 <td class="left"><tt>i32</tt></td>
1837 <td class="left">a 32-bit integer.</td>
1840 <td class="left"><tt>i1942652</tt></td>
1841 <td class="left">a really big integer of over 1 million bits.</td>
1847 <!-- _______________________________________________________________________ -->
1849 <a name="t_floating">Floating Point Types</a>
1856 <tr><th>Type</th><th>Description</th></tr>
1857 <tr><td><tt>half</tt></td><td>16-bit floating point value</td></tr>
1858 <tr><td><tt>float</tt></td><td>32-bit floating point value</td></tr>
1859 <tr><td><tt>double</tt></td><td>64-bit floating point value</td></tr>
1860 <tr><td><tt>fp128</tt></td><td>128-bit floating point value (112-bit mantissa)</td></tr>
1861 <tr><td><tt>x86_fp80</tt></td><td>80-bit floating point value (X87)</td></tr>
1862 <tr><td><tt>ppc_fp128</tt></td><td>128-bit floating point value (two 64-bits)</td></tr>
1868 <!-- _______________________________________________________________________ -->
1870 <a name="t_x86mmx">X86mmx Type</a>
1876 <p>The x86mmx type represents a value held in an MMX register on an x86 machine. The operations allowed on it are quite limited: parameters and return values, load and store, and bitcast. User-specified MMX instructions are represented as intrinsic or asm calls with arguments and/or results of this type. There are no arrays, vectors or constants of this type.</p>
1885 <!-- _______________________________________________________________________ -->
1887 <a name="t_void">Void Type</a>
1893 <p>The void type does not represent any value and has no size.</p>
1902 <!-- _______________________________________________________________________ -->
1904 <a name="t_label">Label Type</a>
1910 <p>The label type represents code labels.</p>
1919 <!-- _______________________________________________________________________ -->
1921 <a name="t_metadata">Metadata Type</a>
1927 <p>The metadata type represents embedded metadata. No derived types may be
1928 created from metadata except for <a href="#t_function">function</a>
1940 <!-- ======================================================================= -->
1942 <a name="t_derived">Derived Types</a>
1947 <p>The real power in LLVM comes from the derived types in the system. This is
1948 what allows a programmer to represent arrays, functions, pointers, and other
1949 useful types. Each of these types contain one or more element types which
1950 may be a primitive type, or another derived type. For example, it is
1951 possible to have a two dimensional array, using an array as the element type
1952 of another array.</p>
1954 <!-- _______________________________________________________________________ -->
1956 <a name="t_aggregate">Aggregate Types</a>
1961 <p>Aggregate Types are a subset of derived types that can contain multiple
1962 member types. <a href="#t_array">Arrays</a> and
1963 <a href="#t_struct">structs</a> are aggregate types.
1964 <a href="#t_vector">Vectors</a> are not considered to be aggregate types.</p>
1968 <!-- _______________________________________________________________________ -->
1970 <a name="t_array">Array Type</a>
1976 <p>The array type is a very simple derived type that arranges elements
1977 sequentially in memory. The array type requires a size (number of elements)
1978 and an underlying data type.</p>
1982 [<# elements> x <elementtype>]
1985 <p>The number of elements is a constant integer value; <tt>elementtype</tt> may
1986 be any type with a size.</p>
1989 <table class="layout">
1991 <td class="left"><tt>[40 x i32]</tt></td>
1992 <td class="left">Array of 40 32-bit integer values.</td>
1995 <td class="left"><tt>[41 x i32]</tt></td>
1996 <td class="left">Array of 41 32-bit integer values.</td>
1999 <td class="left"><tt>[4 x i8]</tt></td>
2000 <td class="left">Array of 4 8-bit integer values.</td>
2003 <p>Here are some examples of multidimensional arrays:</p>
2004 <table class="layout">
2006 <td class="left"><tt>[3 x [4 x i32]]</tt></td>
2007 <td class="left">3x4 array of 32-bit integer values.</td>
2010 <td class="left"><tt>[12 x [10 x float]]</tt></td>
2011 <td class="left">12x10 array of single precision floating point values.</td>
2014 <td class="left"><tt>[2 x [3 x [4 x i16]]]</tt></td>
2015 <td class="left">2x3x4 array of 16-bit integer values.</td>
2019 <p>There is no restriction on indexing beyond the end of the array implied by
2020 a static type (though there are restrictions on indexing beyond the bounds
2021 of an allocated object in some cases). This means that single-dimension
2022 'variable sized array' addressing can be implemented in LLVM with a zero
2023 length array type. An implementation of 'pascal style arrays' in LLVM could
2024 use the type "<tt>{ i32, [0 x float]}</tt>", for example.</p>
2028 <!-- _______________________________________________________________________ -->
2030 <a name="t_function">Function Type</a>
2036 <p>The function type can be thought of as a function signature. It consists of
2037 a return type and a list of formal parameter types. The return type of a
2038 function type is a first class type or a void type.</p>
2042 <returntype> (<parameter list>)
2045 <p>...where '<tt><parameter list></tt>' is a comma-separated list of type
2046 specifiers. Optionally, the parameter list may include a type <tt>...</tt>,
2047 which indicates that the function takes a variable number of arguments.
2048 Variable argument functions can access their arguments with
2049 the <a href="#int_varargs">variable argument handling intrinsic</a>
2050 functions. '<tt><returntype></tt>' is any type except
2051 <a href="#t_label">label</a>.</p>
2054 <table class="layout">
2056 <td class="left"><tt>i32 (i32)</tt></td>
2057 <td class="left">function taking an <tt>i32</tt>, returning an <tt>i32</tt>
2059 </tr><tr class="layout">
2060 <td class="left"><tt>float (i16, i32 *) *
2062 <td class="left"><a href="#t_pointer">Pointer</a> to a function that takes
2063 an <tt>i16</tt> and a <a href="#t_pointer">pointer</a> to <tt>i32</tt>,
2064 returning <tt>float</tt>.
2066 </tr><tr class="layout">
2067 <td class="left"><tt>i32 (i8*, ...)</tt></td>
2068 <td class="left">A vararg function that takes at least one
2069 <a href="#t_pointer">pointer</a> to <tt>i8 </tt> (char in C),
2070 which returns an integer. This is the signature for <tt>printf</tt> in
2073 </tr><tr class="layout">
2074 <td class="left"><tt>{i32, i32} (i32)</tt></td>
2075 <td class="left">A function taking an <tt>i32</tt>, returning a
2076 <a href="#t_struct">structure</a> containing two <tt>i32</tt> values
2083 <!-- _______________________________________________________________________ -->
2085 <a name="t_struct">Structure Type</a>
2091 <p>The structure type is used to represent a collection of data members together
2092 in memory. The elements of a structure may be any type that has a size.</p>
2094 <p>Structures in memory are accessed using '<tt><a href="#i_load">load</a></tt>'
2095 and '<tt><a href="#i_store">store</a></tt>' by getting a pointer to a field
2096 with the '<tt><a href="#i_getelementptr">getelementptr</a></tt>' instruction.
2097 Structures in registers are accessed using the
2098 '<tt><a href="#i_extractvalue">extractvalue</a></tt>' and
2099 '<tt><a href="#i_insertvalue">insertvalue</a></tt>' instructions.</p>
2101 <p>Structures may optionally be "packed" structures, which indicate that the
2102 alignment of the struct is one byte, and that there is no padding between
2103 the elements. In non-packed structs, padding between field types is inserted
2104 as defined by the TargetData string in the module, which is required to match
2105 what the underlying code generator expects.</p>
2107 <p>Structures can either be "literal" or "identified". A literal structure is
2108 defined inline with other types (e.g. <tt>{i32, i32}*</tt>) whereas identified
2109 types are always defined at the top level with a name. Literal types are
2110 uniqued by their contents and can never be recursive or opaque since there is
2111 no way to write one. Identified types can be recursive, can be opaqued, and are
2117 %T1 = type { <type list> } <i>; Identified normal struct type</i>
2118 %T2 = type <{ <type list> }> <i>; Identified packed struct type</i>
2122 <table class="layout">
2124 <td class="left"><tt>{ i32, i32, i32 }</tt></td>
2125 <td class="left">A triple of three <tt>i32</tt> values</td>
2128 <td class="left"><tt>{ float, i32 (i32) * }</tt></td>
2129 <td class="left">A pair, where the first element is a <tt>float</tt> and the
2130 second element is a <a href="#t_pointer">pointer</a> to a
2131 <a href="#t_function">function</a> that takes an <tt>i32</tt>, returning
2132 an <tt>i32</tt>.</td>
2135 <td class="left"><tt><{ i8, i32 }></tt></td>
2136 <td class="left">A packed struct known to be 5 bytes in size.</td>
2142 <!-- _______________________________________________________________________ -->
2144 <a name="t_opaque">Opaque Structure Types</a>
2150 <p>Opaque structure types are used to represent named structure types that do
2151 not have a body specified. This corresponds (for example) to the C notion of
2152 a forward declared structure.</p>
2161 <table class="layout">
2163 <td class="left"><tt>opaque</tt></td>
2164 <td class="left">An opaque type.</td>
2172 <!-- _______________________________________________________________________ -->
2174 <a name="t_pointer">Pointer Type</a>
2180 <p>The pointer type is used to specify memory locations.
2181 Pointers are commonly used to reference objects in memory.</p>
2183 <p>Pointer types may have an optional address space attribute defining the
2184 numbered address space where the pointed-to object resides. The default
2185 address space is number zero. The semantics of non-zero address
2186 spaces are target-specific.</p>
2188 <p>Note that LLVM does not permit pointers to void (<tt>void*</tt>) nor does it
2189 permit pointers to labels (<tt>label*</tt>). Use <tt>i8*</tt> instead.</p>
2197 <table class="layout">
2199 <td class="left"><tt>[4 x i32]*</tt></td>
2200 <td class="left">A <a href="#t_pointer">pointer</a> to <a
2201 href="#t_array">array</a> of four <tt>i32</tt> values.</td>
2204 <td class="left"><tt>i32 (i32*) *</tt></td>
2205 <td class="left"> A <a href="#t_pointer">pointer</a> to a <a
2206 href="#t_function">function</a> that takes an <tt>i32*</tt>, returning an
2210 <td class="left"><tt>i32 addrspace(5)*</tt></td>
2211 <td class="left">A <a href="#t_pointer">pointer</a> to an <tt>i32</tt> value
2212 that resides in address space #5.</td>
2218 <!-- _______________________________________________________________________ -->
2220 <a name="t_vector">Vector Type</a>
2226 <p>A vector type is a simple derived type that represents a vector of elements.
2227 Vector types are used when multiple primitive data are operated in parallel
2228 using a single instruction (SIMD). A vector type requires a size (number of
2229 elements) and an underlying primitive data type. Vector types are considered
2230 <a href="#t_firstclass">first class</a>.</p>
2234 < <# elements> x <elementtype> >
2237 <p>The number of elements is a constant integer value larger than 0; elementtype
2238 may be any integer or floating point type, or a pointer to these types.
2239 Vectors of size zero are not allowed. </p>
2242 <table class="layout">
2244 <td class="left"><tt><4 x i32></tt></td>
2245 <td class="left">Vector of 4 32-bit integer values.</td>
2248 <td class="left"><tt><8 x float></tt></td>
2249 <td class="left">Vector of 8 32-bit floating-point values.</td>
2252 <td class="left"><tt><2 x i64></tt></td>
2253 <td class="left">Vector of 2 64-bit integer values.</td>
2256 <td class="left"><tt><4 x i64*></tt></td>
2257 <td class="left">Vector of 4 pointers to 64-bit integer values.</td>
2267 <!-- *********************************************************************** -->
2268 <h2><a name="constants">Constants</a></h2>
2269 <!-- *********************************************************************** -->
2273 <p>LLVM has several different basic types of constants. This section describes
2274 them all and their syntax.</p>
2276 <!-- ======================================================================= -->
2278 <a name="simpleconstants">Simple Constants</a>
2284 <dt><b>Boolean constants</b></dt>
2285 <dd>The two strings '<tt>true</tt>' and '<tt>false</tt>' are both valid
2286 constants of the <tt><a href="#t_integer">i1</a></tt> type.</dd>
2288 <dt><b>Integer constants</b></dt>
2289 <dd>Standard integers (such as '4') are constants of
2290 the <a href="#t_integer">integer</a> type. Negative numbers may be used
2291 with integer types.</dd>
2293 <dt><b>Floating point constants</b></dt>
2294 <dd>Floating point constants use standard decimal notation (e.g. 123.421),
2295 exponential notation (e.g. 1.23421e+2), or a more precise hexadecimal
2296 notation (see below). The assembler requires the exact decimal value of a
2297 floating-point constant. For example, the assembler accepts 1.25 but
2298 rejects 1.3 because 1.3 is a repeating decimal in binary. Floating point
2299 constants must have a <a href="#t_floating">floating point</a> type. </dd>
2301 <dt><b>Null pointer constants</b></dt>
2302 <dd>The identifier '<tt>null</tt>' is recognized as a null pointer constant
2303 and must be of <a href="#t_pointer">pointer type</a>.</dd>
2306 <p>The one non-intuitive notation for constants is the hexadecimal form of
2307 floating point constants. For example, the form '<tt>double
2308 0x432ff973cafa8000</tt>' is equivalent to (but harder to read than)
2309 '<tt>double 4.5e+15</tt>'. The only time hexadecimal floating point
2310 constants are required (and the only time that they are generated by the
2311 disassembler) is when a floating point constant must be emitted but it cannot
2312 be represented as a decimal floating point number in a reasonable number of
2313 digits. For example, NaN's, infinities, and other special values are
2314 represented in their IEEE hexadecimal format so that assembly and disassembly
2315 do not cause any bits to change in the constants.</p>
2317 <p>When using the hexadecimal form, constants of types half, float, and double are
2318 represented using the 16-digit form shown above (which matches the IEEE754
2319 representation for double); half and float values must, however, be exactly
2320 representable as IEE754 half and single precision, respectively.
2321 Hexadecimal format is always used
2322 for long double, and there are three forms of long double. The 80-bit format
2323 used by x86 is represented as <tt>0xK</tt> followed by 20 hexadecimal digits.
2324 The 128-bit format used by PowerPC (two adjacent doubles) is represented
2325 by <tt>0xM</tt> followed by 32 hexadecimal digits. The IEEE 128-bit format
2326 is represented by <tt>0xL</tt> followed by 32 hexadecimal digits; no
2327 currently supported target uses this format. Long doubles will only work if
2328 they match the long double format on your target. The IEEE 16-bit format
2329 (half precision) is represented by <tt>0xH</tt> followed by 4 hexadecimal
2330 digits. All hexadecimal formats are big-endian (sign bit at the left).</p>
2332 <p>There are no constants of type x86mmx.</p>
2335 <!-- ======================================================================= -->
2337 <a name="aggregateconstants"></a> <!-- old anchor -->
2338 <a name="complexconstants">Complex Constants</a>
2343 <p>Complex constants are a (potentially recursive) combination of simple
2344 constants and smaller complex constants.</p>
2347 <dt><b>Structure constants</b></dt>
2348 <dd>Structure constants are represented with notation similar to structure
2349 type definitions (a comma separated list of elements, surrounded by braces
2350 (<tt>{}</tt>)). For example: "<tt>{ i32 4, float 17.0, i32* @G }</tt>",
2351 where "<tt>@G</tt>" is declared as "<tt>@G = external global i32</tt>".
2352 Structure constants must have <a href="#t_struct">structure type</a>, and
2353 the number and types of elements must match those specified by the
2356 <dt><b>Array constants</b></dt>
2357 <dd>Array constants are represented with notation similar to array type
2358 definitions (a comma separated list of elements, surrounded by square
2359 brackets (<tt>[]</tt>)). For example: "<tt>[ i32 42, i32 11, i32 74
2360 ]</tt>". Array constants must have <a href="#t_array">array type</a>, and
2361 the number and types of elements must match those specified by the
2364 <dt><b>Vector constants</b></dt>
2365 <dd>Vector constants are represented with notation similar to vector type
2366 definitions (a comma separated list of elements, surrounded by
2367 less-than/greater-than's (<tt><></tt>)). For example: "<tt>< i32
2368 42, i32 11, i32 74, i32 100 ></tt>". Vector constants must
2369 have <a href="#t_vector">vector type</a>, and the number and types of
2370 elements must match those specified by the type.</dd>
2372 <dt><b>Zero initialization</b></dt>
2373 <dd>The string '<tt>zeroinitializer</tt>' can be used to zero initialize a
2374 value to zero of <em>any</em> type, including scalar and
2375 <a href="#t_aggregate">aggregate</a> types.
2376 This is often used to avoid having to print large zero initializers
2377 (e.g. for large arrays) and is always exactly equivalent to using explicit
2378 zero initializers.</dd>
2380 <dt><b>Metadata node</b></dt>
2381 <dd>A metadata node is a structure-like constant with
2382 <a href="#t_metadata">metadata type</a>. For example: "<tt>metadata !{
2383 i32 0, metadata !"test" }</tt>". Unlike other constants that are meant to
2384 be interpreted as part of the instruction stream, metadata is a place to
2385 attach additional information such as debug info.</dd>
2390 <!-- ======================================================================= -->
2392 <a name="globalconstants">Global Variable and Function Addresses</a>
2397 <p>The addresses of <a href="#globalvars">global variables</a>
2398 and <a href="#functionstructure">functions</a> are always implicitly valid
2399 (link-time) constants. These constants are explicitly referenced when
2400 the <a href="#identifiers">identifier for the global</a> is used and always
2401 have <a href="#t_pointer">pointer</a> type. For example, the following is a
2402 legal LLVM file:</p>
2404 <pre class="doc_code">
2407 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2412 <!-- ======================================================================= -->
2414 <a name="undefvalues">Undefined Values</a>
2419 <p>The string '<tt>undef</tt>' can be used anywhere a constant is expected, and
2420 indicates that the user of the value may receive an unspecified bit-pattern.
2421 Undefined values may be of any type (other than '<tt>label</tt>'
2422 or '<tt>void</tt>') and be used anywhere a constant is permitted.</p>
2424 <p>Undefined values are useful because they indicate to the compiler that the
2425 program is well defined no matter what value is used. This gives the
2426 compiler more freedom to optimize. Here are some examples of (potentially
2427 surprising) transformations that are valid (in pseudo IR):</p>
2430 <pre class="doc_code">
2440 <p>This is safe because all of the output bits are affected by the undef bits.
2441 Any output bit can have a zero or one depending on the input bits.</p>
2443 <pre class="doc_code">
2454 <p>These logical operations have bits that are not always affected by the input.
2455 For example, if <tt>%X</tt> has a zero bit, then the output of the
2456 '<tt>and</tt>' operation will always be a zero for that bit, no matter what
2457 the corresponding bit from the '<tt>undef</tt>' is. As such, it is unsafe to
2458 optimize or assume that the result of the '<tt>and</tt>' is '<tt>undef</tt>'.
2459 However, it is safe to assume that all bits of the '<tt>undef</tt>' could be
2460 0, and optimize the '<tt>and</tt>' to 0. Likewise, it is safe to assume that
2461 all the bits of the '<tt>undef</tt>' operand to the '<tt>or</tt>' could be
2462 set, allowing the '<tt>or</tt>' to be folded to -1.</p>
2464 <pre class="doc_code">
2465 %A = select undef, %X, %Y
2466 %B = select undef, 42, %Y
2467 %C = select %X, %Y, undef
2478 <p>This set of examples shows that undefined '<tt>select</tt>' (and conditional
2479 branch) conditions can go <em>either way</em>, but they have to come from one
2480 of the two operands. In the <tt>%A</tt> example, if <tt>%X</tt> and
2481 <tt>%Y</tt> were both known to have a clear low bit, then <tt>%A</tt> would
2482 have to have a cleared low bit. However, in the <tt>%C</tt> example, the
2483 optimizer is allowed to assume that the '<tt>undef</tt>' operand could be the
2484 same as <tt>%Y</tt>, allowing the whole '<tt>select</tt>' to be
2487 <pre class="doc_code">
2488 %A = xor undef, undef
2506 <p>This example points out that two '<tt>undef</tt>' operands are not
2507 necessarily the same. This can be surprising to people (and also matches C
2508 semantics) where they assume that "<tt>X^X</tt>" is always zero, even
2509 if <tt>X</tt> is undefined. This isn't true for a number of reasons, but the
2510 short answer is that an '<tt>undef</tt>' "variable" can arbitrarily change
2511 its value over its "live range". This is true because the variable doesn't
2512 actually <em>have a live range</em>. Instead, the value is logically read
2513 from arbitrary registers that happen to be around when needed, so the value
2514 is not necessarily consistent over time. In fact, <tt>%A</tt> and <tt>%C</tt>
2515 need to have the same semantics or the core LLVM "replace all uses with"
2516 concept would not hold.</p>
2518 <pre class="doc_code">
2526 <p>These examples show the crucial difference between an <em>undefined
2527 value</em> and <em>undefined behavior</em>. An undefined value (like
2528 '<tt>undef</tt>') is allowed to have an arbitrary bit-pattern. This means that
2529 the <tt>%A</tt> operation can be constant folded to '<tt>undef</tt>', because
2530 the '<tt>undef</tt>' could be an SNaN, and <tt>fdiv</tt> is not (currently)
2531 defined on SNaN's. However, in the second example, we can make a more
2532 aggressive assumption: because the <tt>undef</tt> is allowed to be an
2533 arbitrary value, we are allowed to assume that it could be zero. Since a
2534 divide by zero has <em>undefined behavior</em>, we are allowed to assume that
2535 the operation does not execute at all. This allows us to delete the divide and
2536 all code after it. Because the undefined operation "can't happen", the
2537 optimizer can assume that it occurs in dead code.</p>
2539 <pre class="doc_code">
2540 a: store undef -> %X
2541 b: store %X -> undef
2547 <p>These examples reiterate the <tt>fdiv</tt> example: a store <em>of</em> an
2548 undefined value can be assumed to not have any effect; we can assume that the
2549 value is overwritten with bits that happen to match what was already there.
2550 However, a store <em>to</em> an undefined location could clobber arbitrary
2551 memory, therefore, it has undefined behavior.</p>
2555 <!-- ======================================================================= -->
2557 <a name="poisonvalues">Poison Values</a>
2562 <p>Poison values are similar to <a href="#undefvalues">undef values</a>, however
2563 they also represent the fact that an instruction or constant expression which
2564 cannot evoke side effects has nevertheless detected a condition which results
2565 in undefined behavior.</p>
2567 <p>There is currently no way of representing a poison value in the IR; they
2568 only exist when produced by operations such as
2569 <a href="#i_add"><tt>add</tt></a> with the <tt>nsw</tt> flag.</p>
2571 <p>Poison value behavior is defined in terms of value <i>dependence</i>:</p>
2574 <li>Values other than <a href="#i_phi"><tt>phi</tt></a> nodes depend on
2575 their operands.</li>
2577 <li><a href="#i_phi"><tt>Phi</tt></a> nodes depend on the operand corresponding
2578 to their dynamic predecessor basic block.</li>
2580 <li>Function arguments depend on the corresponding actual argument values in
2581 the dynamic callers of their functions.</li>
2583 <li><a href="#i_call"><tt>Call</tt></a> instructions depend on the
2584 <a href="#i_ret"><tt>ret</tt></a> instructions that dynamically transfer
2585 control back to them.</li>
2587 <li><a href="#i_invoke"><tt>Invoke</tt></a> instructions depend on the
2588 <a href="#i_ret"><tt>ret</tt></a>, <a href="#i_resume"><tt>resume</tt></a>,
2589 or exception-throwing call instructions that dynamically transfer control
2592 <li>Non-volatile loads and stores depend on the most recent stores to all of the
2593 referenced memory addresses, following the order in the IR
2594 (including loads and stores implied by intrinsics such as
2595 <a href="#int_memcpy"><tt>@llvm.memcpy</tt></a>.)</li>
2597 <!-- TODO: In the case of multiple threads, this only applies if the store
2598 "happens-before" the load or store. -->
2600 <!-- TODO: floating-point exception state -->
2602 <li>An instruction with externally visible side effects depends on the most
2603 recent preceding instruction with externally visible side effects, following
2604 the order in the IR. (This includes
2605 <a href="#volatile">volatile operations</a>.)</li>
2607 <li>An instruction <i>control-depends</i> on a
2608 <a href="#terminators">terminator instruction</a>
2609 if the terminator instruction has multiple successors and the instruction
2610 is always executed when control transfers to one of the successors, and
2611 may not be executed when control is transferred to another.</li>
2613 <li>Additionally, an instruction also <i>control-depends</i> on a terminator
2614 instruction if the set of instructions it otherwise depends on would be
2615 different if the terminator had transferred control to a different
2618 <li>Dependence is transitive.</li>
2622 <p>Poison Values have the same behavior as <a href="#undefvalues">undef values</a>,
2623 with the additional affect that any instruction which has a <i>dependence</i>
2624 on a poison value has undefined behavior.</p>
2626 <p>Here are some examples:</p>
2628 <pre class="doc_code">
2630 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2631 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2632 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2633 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2635 store i32 %poison, i32* @g ; Poison value stored to memory.
2636 %poison2 = load i32* @g ; Poison value loaded back from memory.
2638 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2640 %narrowaddr = bitcast i32* @g to i16*
2641 %wideaddr = bitcast i32* @g to i64*
2642 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2643 %poison4 = load i64* %wideaddr ; Returns a poison value.
2645 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2646 br i1 %cmp, label %true, label %end ; Branch to either destination.
2649 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2650 ; it has undefined behavior.
2654 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2655 ; Both edges into this PHI are
2656 ; control-dependent on %cmp, so this
2657 ; always results in a poison value.
2659 store volatile i32 0, i32* @g ; This would depend on the store in %true
2660 ; if %cmp is true, or the store in %entry
2661 ; otherwise, so this is undefined behavior.
2663 br i1 %cmp, label %second_true, label %second_end
2664 ; The same branch again, but this time the
2665 ; true block doesn't have side effects.
2672 store volatile i32 0, i32* @g ; This time, the instruction always depends
2673 ; on the store in %end. Also, it is
2674 ; control-equivalent to %end, so this is
2675 ; well-defined (ignoring earlier undefined
2676 ; behavior in this example).
2681 <!-- ======================================================================= -->
2683 <a name="blockaddress">Addresses of Basic Blocks</a>
2688 <p><b><tt>blockaddress(@function, %block)</tt></b></p>
2690 <p>The '<tt>blockaddress</tt>' constant computes the address of the specified
2691 basic block in the specified function, and always has an i8* type. Taking
2692 the address of the entry block is illegal.</p>
2694 <p>This value only has defined behavior when used as an operand to the
2695 '<a href="#i_indirectbr"><tt>indirectbr</tt></a>' instruction, or for
2696 comparisons against null. Pointer equality tests between labels addresses
2697 results in undefined behavior — though, again, comparison against null
2698 is ok, and no label is equal to the null pointer. This may be passed around
2699 as an opaque pointer sized value as long as the bits are not inspected. This
2700 allows <tt>ptrtoint</tt> and arithmetic to be performed on these values so
2701 long as the original value is reconstituted before the <tt>indirectbr</tt>
2704 <p>Finally, some targets may provide defined semantics when using the value as
2705 the operand to an inline assembly, but that is target specific.</p>
2710 <!-- ======================================================================= -->
2712 <a name="constantexprs">Constant Expressions</a>
2717 <p>Constant expressions are used to allow expressions involving other constants
2718 to be used as constants. Constant expressions may be of
2719 any <a href="#t_firstclass">first class</a> type and may involve any LLVM
2720 operation that does not have side effects (e.g. load and call are not
2721 supported). The following is the syntax for constant expressions:</p>
2724 <dt><b><tt>trunc (CST to TYPE)</tt></b></dt>
2725 <dd>Truncate a constant to another type. The bit size of CST must be larger
2726 than the bit size of TYPE. Both types must be integers.</dd>
2728 <dt><b><tt>zext (CST to TYPE)</tt></b></dt>
2729 <dd>Zero extend a constant to another type. The bit size of CST must be
2730 smaller than the bit size of TYPE. Both types must be integers.</dd>
2732 <dt><b><tt>sext (CST to TYPE)</tt></b></dt>
2733 <dd>Sign extend a constant to another type. The bit size of CST must be
2734 smaller than the bit size of TYPE. Both types must be integers.</dd>
2736 <dt><b><tt>fptrunc (CST to TYPE)</tt></b></dt>
2737 <dd>Truncate a floating point constant to another floating point type. The
2738 size of CST must be larger than the size of TYPE. Both types must be
2739 floating point.</dd>
2741 <dt><b><tt>fpext (CST to TYPE)</tt></b></dt>
2742 <dd>Floating point extend a constant to another type. The size of CST must be
2743 smaller or equal to the size of TYPE. Both types must be floating
2746 <dt><b><tt>fptoui (CST to TYPE)</tt></b></dt>
2747 <dd>Convert a floating point constant to the corresponding unsigned integer
2748 constant. TYPE must be a scalar or vector integer type. CST must be of
2749 scalar or vector floating point type. Both CST and TYPE must be scalars,
2750 or vectors of the same number of elements. If the value won't fit in the
2751 integer type, the results are undefined.</dd>
2753 <dt><b><tt>fptosi (CST to TYPE)</tt></b></dt>
2754 <dd>Convert a floating point constant to the corresponding signed integer
2755 constant. TYPE must be a scalar or vector integer type. CST must be of
2756 scalar or vector floating point type. Both CST and TYPE must be scalars,
2757 or vectors of the same number of elements. If the value won't fit in the
2758 integer type, the results are undefined.</dd>
2760 <dt><b><tt>uitofp (CST to TYPE)</tt></b></dt>
2761 <dd>Convert an unsigned integer constant to the corresponding floating point
2762 constant. TYPE must be a scalar or vector floating point type. CST must be
2763 of scalar or vector integer type. Both CST and TYPE must be scalars, or
2764 vectors of the same number of elements. If the value won't fit in the
2765 floating point type, the results are undefined.</dd>
2767 <dt><b><tt>sitofp (CST to TYPE)</tt></b></dt>
2768 <dd>Convert a signed integer constant to the corresponding floating point
2769 constant. TYPE must be a scalar or vector floating point type. CST must be
2770 of scalar or vector integer type. Both CST and TYPE must be scalars, or
2771 vectors of the same number of elements. If the value won't fit in the
2772 floating point type, the results are undefined.</dd>
2774 <dt><b><tt>ptrtoint (CST to TYPE)</tt></b></dt>
2775 <dd>Convert a pointer typed constant to the corresponding integer constant
2776 <tt>TYPE</tt> must be an integer type. <tt>CST</tt> must be of pointer
2777 type. The <tt>CST</tt> value is zero extended, truncated, or unchanged to
2778 make it fit in <tt>TYPE</tt>.</dd>
2780 <dt><b><tt>inttoptr (CST to TYPE)</tt></b></dt>
2781 <dd>Convert a integer constant to a pointer constant. TYPE must be a pointer
2782 type. CST must be of integer type. The CST value is zero extended,
2783 truncated, or unchanged to make it fit in a pointer size. This one is
2784 <i>really</i> dangerous!</dd>
2786 <dt><b><tt>bitcast (CST to TYPE)</tt></b></dt>
2787 <dd>Convert a constant, CST, to another TYPE. The constraints of the operands
2788 are the same as those for the <a href="#i_bitcast">bitcast
2789 instruction</a>.</dd>
2791 <dt><b><tt>getelementptr (CSTPTR, IDX0, IDX1, ...)</tt></b></dt>
2792 <dt><b><tt>getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)</tt></b></dt>
2793 <dd>Perform the <a href="#i_getelementptr">getelementptr operation</a> on
2794 constants. As with the <a href="#i_getelementptr">getelementptr</a>
2795 instruction, the index list may have zero or more indexes, which are
2796 required to make sense for the type of "CSTPTR".</dd>
2798 <dt><b><tt>select (COND, VAL1, VAL2)</tt></b></dt>
2799 <dd>Perform the <a href="#i_select">select operation</a> on constants.</dd>
2801 <dt><b><tt>icmp COND (VAL1, VAL2)</tt></b></dt>
2802 <dd>Performs the <a href="#i_icmp">icmp operation</a> on constants.</dd>
2804 <dt><b><tt>fcmp COND (VAL1, VAL2)</tt></b></dt>
2805 <dd>Performs the <a href="#i_fcmp">fcmp operation</a> on constants.</dd>
2807 <dt><b><tt>extractelement (VAL, IDX)</tt></b></dt>
2808 <dd>Perform the <a href="#i_extractelement">extractelement operation</a> on
2811 <dt><b><tt>insertelement (VAL, ELT, IDX)</tt></b></dt>
2812 <dd>Perform the <a href="#i_insertelement">insertelement operation</a> on
2815 <dt><b><tt>shufflevector (VEC1, VEC2, IDXMASK)</tt></b></dt>
2816 <dd>Perform the <a href="#i_shufflevector">shufflevector operation</a> on
2819 <dt><b><tt>extractvalue (VAL, IDX0, IDX1, ...)</tt></b></dt>
2820 <dd>Perform the <a href="#i_extractvalue">extractvalue operation</a> on
2821 constants. The index list is interpreted in a similar manner as indices in
2822 a '<a href="#i_getelementptr">getelementptr</a>' operation. At least one
2823 index value must be specified.</dd>
2825 <dt><b><tt>insertvalue (VAL, ELT, IDX0, IDX1, ...)</tt></b></dt>
2826 <dd>Perform the <a href="#i_insertvalue">insertvalue operation</a> on
2827 constants. The index list is interpreted in a similar manner as indices in
2828 a '<a href="#i_getelementptr">getelementptr</a>' operation. At least one
2829 index value must be specified.</dd>
2831 <dt><b><tt>OPCODE (LHS, RHS)</tt></b></dt>
2832 <dd>Perform the specified operation of the LHS and RHS constants. OPCODE may
2833 be any of the <a href="#binaryops">binary</a>
2834 or <a href="#bitwiseops">bitwise binary</a> operations. The constraints
2835 on operands are the same as those for the corresponding instruction
2836 (e.g. no bitwise operations on floating point values are allowed).</dd>
2843 <!-- *********************************************************************** -->
2844 <h2><a name="othervalues">Other Values</a></h2>
2845 <!-- *********************************************************************** -->
2847 <!-- ======================================================================= -->
2849 <a name="inlineasm">Inline Assembler Expressions</a>
2854 <p>LLVM supports inline assembler expressions (as opposed
2855 to <a href="#moduleasm">Module-Level Inline Assembly</a>) through the use of
2856 a special value. This value represents the inline assembler as a string
2857 (containing the instructions to emit), a list of operand constraints (stored
2858 as a string), a flag that indicates whether or not the inline asm
2859 expression has side effects, and a flag indicating whether the function
2860 containing the asm needs to align its stack conservatively. An example
2861 inline assembler expression is:</p>
2863 <pre class="doc_code">
2864 i32 (i32) asm "bswap $0", "=r,r"
2867 <p>Inline assembler expressions may <b>only</b> be used as the callee operand of
2868 a <a href="#i_call"><tt>call</tt> instruction</a>. Thus, typically we
2871 <pre class="doc_code">
2872 %X = call i32 asm "<a href="#int_bswap">bswap</a> $0", "=r,r"(i32 %Y)
2875 <p>Inline asms with side effects not visible in the constraint list must be
2876 marked as having side effects. This is done through the use of the
2877 '<tt>sideeffect</tt>' keyword, like so:</p>
2879 <pre class="doc_code">
2880 call void asm sideeffect "eieio", ""()
2883 <p>In some cases inline asms will contain code that will not work unless the
2884 stack is aligned in some way, such as calls or SSE instructions on x86,
2885 yet will not contain code that does that alignment within the asm.
2886 The compiler should make conservative assumptions about what the asm might
2887 contain and should generate its usual stack alignment code in the prologue
2888 if the '<tt>alignstack</tt>' keyword is present:</p>
2890 <pre class="doc_code">
2891 call void asm alignstack "eieio", ""()
2894 <p>If both keywords appear the '<tt>sideeffect</tt>' keyword must come
2898 <p>TODO: The format of the asm and constraints string still need to be
2899 documented here. Constraints on what can be done (e.g. duplication, moving,
2900 etc need to be documented). This is probably best done by reference to
2901 another document that covers inline asm from a holistic perspective.</p>
2904 <!-- _______________________________________________________________________ -->
2906 <a name="inlineasm_md">Inline Asm Metadata</a>
2911 <p>The call instructions that wrap inline asm nodes may have a
2912 "<tt>!srcloc</tt>" MDNode attached to it that contains a list of constant
2913 integers. If present, the code generator will use the integer as the
2914 location cookie value when report errors through the <tt>LLVMContext</tt>
2915 error reporting mechanisms. This allows a front-end to correlate backend
2916 errors that occur with inline asm back to the source code that produced it.
2919 <pre class="doc_code">
2920 call void asm sideeffect "something bad", ""()<b>, !srcloc !42</b>
2922 !42 = !{ i32 1234567 }
2925 <p>It is up to the front-end to make sense of the magic numbers it places in the
2926 IR. If the MDNode contains multiple constants, the code generator will use
2927 the one that corresponds to the line of the asm that the error occurs on.</p>
2933 <!-- ======================================================================= -->
2935 <a name="metadata">Metadata Nodes and Metadata Strings</a>
2940 <p>LLVM IR allows metadata to be attached to instructions in the program that
2941 can convey extra information about the code to the optimizers and code
2942 generator. One example application of metadata is source-level debug
2943 information. There are two metadata primitives: strings and nodes. All
2944 metadata has the <tt>metadata</tt> type and is identified in syntax by a
2945 preceding exclamation point ('<tt>!</tt>').</p>
2947 <p>A metadata string is a string surrounded by double quotes. It can contain
2948 any character by escaping non-printable characters with "<tt>\xx</tt>" where
2949 "<tt>xx</tt>" is the two digit hex code. For example:
2950 "<tt>!"test\00"</tt>".</p>
2952 <p>Metadata nodes are represented with notation similar to structure constants
2953 (a comma separated list of elements, surrounded by braces and preceded by an
2954 exclamation point). Metadata nodes can have any values as their operand. For
2957 <div class="doc_code">
2959 !{ metadata !"test\00", i32 10}
2963 <p>A <a href="#namedmetadatastructure">named metadata</a> is a collection of
2964 metadata nodes, which can be looked up in the module symbol table. For
2967 <div class="doc_code">
2969 !foo = metadata !{!4, !3}
2973 <p>Metadata can be used as function arguments. Here <tt>llvm.dbg.value</tt>
2974 function is using two metadata arguments:</p>
2976 <div class="doc_code">
2978 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2982 <p>Metadata can be attached with an instruction. Here metadata <tt>!21</tt> is
2983 attached to the <tt>add</tt> instruction using the <tt>!dbg</tt>
2986 <div class="doc_code">
2988 %indvar.next = add i64 %indvar, 1, !dbg !21
2992 <p>More information about specific metadata nodes recognized by the optimizers
2993 and code generator is found below.</p>
2995 <!-- _______________________________________________________________________ -->
2997 <a name="tbaa">'<tt>tbaa</tt>' Metadata</a>
3002 <p>In LLVM IR, memory does not have types, so LLVM's own type system is not
3003 suitable for doing TBAA. Instead, metadata is added to the IR to describe
3004 a type system of a higher level language. This can be used to implement
3005 typical C/C++ TBAA, but it can also be used to implement custom alias
3006 analysis behavior for other languages.</p>
3008 <p>The current metadata format is very simple. TBAA metadata nodes have up to
3009 three fields, e.g.:</p>
3011 <div class="doc_code">
3013 !0 = metadata !{ metadata !"an example type tree" }
3014 !1 = metadata !{ metadata !"int", metadata !0 }
3015 !2 = metadata !{ metadata !"float", metadata !0 }
3016 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
3020 <p>The first field is an identity field. It can be any value, usually
3021 a metadata string, which uniquely identifies the type. The most important
3022 name in the tree is the name of the root node. Two trees with
3023 different root node names are entirely disjoint, even if they
3024 have leaves with common names.</p>
3026 <p>The second field identifies the type's parent node in the tree, or
3027 is null or omitted for a root node. A type is considered to alias
3028 all of its descendants and all of its ancestors in the tree. Also,
3029 a type is considered to alias all types in other trees, so that
3030 bitcode produced from multiple front-ends is handled conservatively.</p>
3032 <p>If the third field is present, it's an integer which if equal to 1
3033 indicates that the type is "constant" (meaning
3034 <tt>pointsToConstantMemory</tt> should return true; see
3035 <a href="AliasAnalysis.html#OtherItfs">other useful
3036 <tt>AliasAnalysis</tt> methods</a>).</p>
3040 <!-- _______________________________________________________________________ -->
3042 <a name="fpmath">'<tt>fpmath</tt>' Metadata</a>
3047 <p><tt>fpmath</tt> metadata may be attached to any instruction of floating point
3048 type. It can be used to express the maximum acceptable error in the result of
3049 that instruction, in ULPs, thus potentially allowing the compiler to use a
3050 more efficient but less accurate method of computing it. ULP is defined as
3055 <p>If <tt>x</tt> is a real number that lies between two finite consecutive
3056 floating-point numbers <tt>a</tt> and <tt>b</tt>, without being equal to one
3057 of them, then <tt>ulp(x) = |b - a|</tt>, otherwise <tt>ulp(x)</tt> is the
3058 distance between the two non-equal finite floating-point numbers nearest
3059 <tt>x</tt>. Moreover, <tt>ulp(NaN)</tt> is <tt>NaN</tt>.</p>
3063 <p>The metadata node shall consist of a single positive floating point number
3064 representing the maximum relative error, for example:</p>
3066 <div class="doc_code">
3068 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
3074 <!-- _______________________________________________________________________ -->
3076 <a name="range">'<tt>range</tt>' Metadata</a>
3080 <p><tt>range</tt> metadata may be attached only to loads of integer types. It
3081 expresses the possible ranges the loaded value is in. The ranges are
3082 represented with a flattened list of integers. The loaded value is known to
3083 be in the union of the ranges defined by each consecutive pair. Each pair
3084 has the following properties:</p>
3086 <li>The type must match the type loaded by the instruction.</li>
3087 <li>The pair <tt>a,b</tt> represents the range <tt>[a,b)</tt>.</li>
3088 <li>Both <tt>a</tt> and <tt>b</tt> are constants.</li>
3089 <li>The range is allowed to wrap.</li>
3090 <li>The range should not represent the full or empty set. That is,
3091 <tt>a!=b</tt>. </li>
3093 <p> In addition, the pairs must be in signed order of the lower bound and
3094 they must be non-contiguous.</p>
3097 <div class="doc_code">
3099 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
3100 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
3101 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
3102 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
3104 !0 = metadata !{ i8 0, i8 2 }
3105 !1 = metadata !{ i8 255, i8 2 }
3106 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
3107 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
3115 <!-- *********************************************************************** -->
3117 <a name="module_flags">Module Flags Metadata</a>
3119 <!-- *********************************************************************** -->
3123 <p>Information about the module as a whole is difficult to convey to LLVM's
3124 subsystems. The LLVM IR isn't sufficient to transmit this
3125 information. The <tt>llvm.module.flags</tt> named metadata exists in order to
3126 facilitate this. These flags are in the form of key / value pairs —
3127 much like a dictionary — making it easy for any subsystem who cares
3128 about a flag to look it up.</p>
3130 <p>The <tt>llvm.module.flags</tt> metadata contains a list of metadata
3131 triplets. Each triplet has the following form:</p>
3134 <li>The first element is a <i>behavior</i> flag, which specifies the behavior
3135 when two (or more) modules are merged together, and it encounters two (or
3136 more) metadata with the same ID. The supported behaviors are described
3139 <li>The second element is a metadata string that is a unique ID for the
3140 metadata. How each ID is interpreted is documented below.</li>
3142 <li>The third element is the value of the flag.</li>
3145 <p>When two (or more) modules are merged together, the resulting
3146 <tt>llvm.module.flags</tt> metadata is the union of the
3147 modules' <tt>llvm.module.flags</tt> metadata. The only exception being a flag
3148 with the <i>Override</i> behavior, which may override another flag's value
3151 <p>The following behaviors are supported:</p>
3153 <table border="1" cellspacing="0" cellpadding="4">
3163 <dt><b>Error</b></dt>
3164 <dd>Emits an error if two values disagree. It is an error to have an ID
3165 with both an Error and a Warning behavior.</dd>
3173 <dt><b>Warning</b></dt>
3174 <dd>Emits a warning if two values disagree.</dd>
3182 <dt><b>Require</b></dt>
3183 <dd>Emits an error when the specified value is not present or doesn't
3184 have the specified value. It is an error for two (or more)
3185 <tt>llvm.module.flags</tt> with the same ID to have the Require
3186 behavior but different values. There may be multiple Require flags
3195 <dt><b>Override</b></dt>
3196 <dd>Uses the specified value if the two values disagree. It is an
3197 error for two (or more) <tt>llvm.module.flags</tt> with the same
3198 ID to have the Override behavior but different values.</dd>
3205 <p>An example of module flags:</p>
3207 <pre class="doc_code">
3208 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
3209 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
3210 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
3211 !3 = metadata !{ i32 3, metadata !"qux",
3213 metadata !"foo", i32 1
3216 !llvm.module.flags = !{ !0, !1, !2, !3 }
3220 <li><p>Metadata <tt>!0</tt> has the ID <tt>!"foo"</tt> and the value '1'. The
3221 behavior if two or more <tt>!"foo"</tt> flags are seen is to emit an
3222 error if their values are not equal.</p></li>
3224 <li><p>Metadata <tt>!1</tt> has the ID <tt>!"bar"</tt> and the value '37'. The
3225 behavior if two or more <tt>!"bar"</tt> flags are seen is to use the
3226 value '37' if their values are not equal.</p></li>
3228 <li><p>Metadata <tt>!2</tt> has the ID <tt>!"qux"</tt> and the value '42'. The
3229 behavior if two or more <tt>!"qux"</tt> flags are seen is to emit a
3230 warning if their values are not equal.</p></li>
3232 <li><p>Metadata <tt>!3</tt> has the ID <tt>!"qux"</tt> and the value:</p>
3234 <pre class="doc_code">
3235 metadata !{ metadata !"foo", i32 1 }
3238 <p>The behavior is to emit an error if the <tt>llvm.module.flags</tt> does
3239 not contain a flag with the ID <tt>!"foo"</tt> that has the value
3240 '1'. If two or more <tt>!"qux"</tt> flags exist, then they must have
3241 the same value or an error will be issued.</p></li>
3245 <!-- ======================================================================= -->
3247 <a name="objc_gc_flags">Objective-C Garbage Collection Module Flags Metadata</a>
3252 <p>On the Mach-O platform, Objective-C stores metadata about garbage collection
3253 in a special section called "image info". The metadata consists of a version
3254 number and a bitmask specifying what types of garbage collection are
3255 supported (if any) by the file. If two or more modules are linked together
3256 their garbage collection metadata needs to be merged rather than appended
3259 <p>The Objective-C garbage collection module flags metadata consists of the
3260 following key-value pairs:</p>
3262 <table border="1" cellspacing="0" cellpadding="4">
3270 <td><tt>Objective-C Version</tt></td>
3271 <td align="left"><b>[Required]</b> — The Objective-C ABI
3272 version. Valid values are 1 and 2.</td>
3275 <td><tt>Objective-C Image Info Version</tt></td>
3276 <td align="left"><b>[Required]</b> — The version of the image info
3277 section. Currently always 0.</td>
3280 <td><tt>Objective-C Image Info Section</tt></td>
3281 <td align="left"><b>[Required]</b> — The section to place the
3282 metadata. Valid values are <tt>"__OBJC, __image_info, regular"</tt> for
3283 Objective-C ABI version 1, and <tt>"__DATA,__objc_imageinfo, regular,
3284 no_dead_strip"</tt> for Objective-C ABI version 2.</td>
3287 <td><tt>Objective-C Garbage Collection</tt></td>
3288 <td align="left"><b>[Required]</b> — Specifies whether garbage
3289 collection is supported or not. Valid values are 0, for no garbage
3290 collection, and 2, for garbage collection supported.</td>
3293 <td><tt>Objective-C GC Only</tt></td>
3294 <td align="left"><b>[Optional]</b> — Specifies that only garbage
3295 collection is supported. If present, its value must be 6. This flag
3296 requires that the <tt>Objective-C Garbage Collection</tt> flag have the
3302 <p>Some important flag interactions:</p>
3305 <li>If a module with <tt>Objective-C Garbage Collection</tt> set to 0 is
3306 merged with a module with <tt>Objective-C Garbage Collection</tt> set to
3307 2, then the resulting module has the <tt>Objective-C Garbage
3308 Collection</tt> flag set to 0.</li>
3310 <li>A module with <tt>Objective-C Garbage Collection</tt> set to 0 cannot be
3311 merged with a module with <tt>Objective-C GC Only</tt> set to 6.</li>
3318 <!-- *********************************************************************** -->
3320 <a name="intrinsic_globals">Intrinsic Global Variables</a>
3322 <!-- *********************************************************************** -->
3324 <p>LLVM has a number of "magic" global variables that contain data that affect
3325 code generation or other IR semantics. These are documented here. All globals
3326 of this sort should have a section specified as "<tt>llvm.metadata</tt>". This
3327 section and all globals that start with "<tt>llvm.</tt>" are reserved for use
3330 <!-- ======================================================================= -->
3332 <a name="intg_used">The '<tt>llvm.used</tt>' Global Variable</a>
3337 <p>The <tt>@llvm.used</tt> global is an array with i8* element type which has <a
3338 href="#linkage_appending">appending linkage</a>. This array contains a list of
3339 pointers to global variables and functions which may optionally have a pointer
3340 cast formed of bitcast or getelementptr. For example, a legal use of it is:</p>
3342 <div class="doc_code">
3347 @llvm.used = appending global [2 x i8*] [
3349 i8* bitcast (i32* @Y to i8*)
3350 ], section "llvm.metadata"
3354 <p>If a global variable appears in the <tt>@llvm.used</tt> list, then the
3355 compiler, assembler, and linker are required to treat the symbol as if there
3356 is a reference to the global that it cannot see. For example, if a variable
3357 has internal linkage and no references other than that from
3358 the <tt>@llvm.used</tt> list, it cannot be deleted. This is commonly used to
3359 represent references from inline asms and other things the compiler cannot
3360 "see", and corresponds to "<tt>attribute((used))</tt>" in GNU C.</p>
3362 <p>On some targets, the code generator must emit a directive to the assembler or
3363 object file to prevent the assembler and linker from molesting the
3368 <!-- ======================================================================= -->
3370 <a name="intg_compiler_used">
3371 The '<tt>llvm.compiler.used</tt>' Global Variable
3377 <p>The <tt>@llvm.compiler.used</tt> directive is the same as the
3378 <tt>@llvm.used</tt> directive, except that it only prevents the compiler from
3379 touching the symbol. On targets that support it, this allows an intelligent
3380 linker to optimize references to the symbol without being impeded as it would
3381 be by <tt>@llvm.used</tt>.</p>
3383 <p>This is a rare construct that should only be used in rare circumstances, and
3384 should not be exposed to source languages.</p>
3388 <!-- ======================================================================= -->
3390 <a name="intg_global_ctors">The '<tt>llvm.global_ctors</tt>' Global Variable</a>
3395 <div class="doc_code">
3397 %0 = type { i32, void ()* }
3398 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
3402 <p>The <tt>@llvm.global_ctors</tt> array contains a list of constructor
3403 functions and associated priorities. The functions referenced by this array
3404 will be called in ascending order of priority (i.e. lowest first) when the
3405 module is loaded. The order of functions with the same priority is not
3410 <!-- ======================================================================= -->
3412 <a name="intg_global_dtors">The '<tt>llvm.global_dtors</tt>' Global Variable</a>
3417 <div class="doc_code">
3419 %0 = type { i32, void ()* }
3420 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
3424 <p>The <tt>@llvm.global_dtors</tt> array contains a list of destructor functions
3425 and associated priorities. The functions referenced by this array will be
3426 called in descending order of priority (i.e. highest first) when the module
3427 is loaded. The order of functions with the same priority is not defined.</p>
3433 <!-- *********************************************************************** -->
3434 <h2><a name="instref">Instruction Reference</a></h2>
3435 <!-- *********************************************************************** -->
3439 <p>The LLVM instruction set consists of several different classifications of
3440 instructions: <a href="#terminators">terminator
3441 instructions</a>, <a href="#binaryops">binary instructions</a>,
3442 <a href="#bitwiseops">bitwise binary instructions</a>,
3443 <a href="#memoryops">memory instructions</a>, and
3444 <a href="#otherops">other instructions</a>.</p>
3446 <!-- ======================================================================= -->
3448 <a name="terminators">Terminator Instructions</a>
3453 <p>As mentioned <a href="#functionstructure">previously</a>, every basic block
3454 in a program ends with a "Terminator" instruction, which indicates which
3455 block should be executed after the current block is finished. These
3456 terminator instructions typically yield a '<tt>void</tt>' value: they produce
3457 control flow, not values (the one exception being the
3458 '<a href="#i_invoke"><tt>invoke</tt></a>' instruction).</p>
3460 <p>The terminator instructions are:
3461 '<a href="#i_ret"><tt>ret</tt></a>',
3462 '<a href="#i_br"><tt>br</tt></a>',
3463 '<a href="#i_switch"><tt>switch</tt></a>',
3464 '<a href="#i_indirectbr"><tt>indirectbr</tt></a>',
3465 '<a href="#i_invoke"><tt>invoke</tt></a>',
3466 '<a href="#i_resume"><tt>resume</tt></a>', and
3467 '<a href="#i_unreachable"><tt>unreachable</tt></a>'.</p>
3469 <!-- _______________________________________________________________________ -->
3471 <a name="i_ret">'<tt>ret</tt>' Instruction</a>
3478 ret <type> <value> <i>; Return a value from a non-void function</i>
3479 ret void <i>; Return from void function</i>
3483 <p>The '<tt>ret</tt>' instruction is used to return control flow (and optionally
3484 a value) from a function back to the caller.</p>
3486 <p>There are two forms of the '<tt>ret</tt>' instruction: one that returns a
3487 value and then causes control flow, and one that just causes control flow to
3491 <p>The '<tt>ret</tt>' instruction optionally accepts a single argument, the
3492 return value. The type of the return value must be a
3493 '<a href="#t_firstclass">first class</a>' type.</p>
3495 <p>A function is not <a href="#wellformed">well formed</a> if it it has a
3496 non-void return type and contains a '<tt>ret</tt>' instruction with no return
3497 value or a return value with a type that does not match its type, or if it
3498 has a void return type and contains a '<tt>ret</tt>' instruction with a
3502 <p>When the '<tt>ret</tt>' instruction is executed, control flow returns back to
3503 the calling function's context. If the caller is a
3504 "<a href="#i_call"><tt>call</tt></a>" instruction, execution continues at the
3505 instruction after the call. If the caller was an
3506 "<a href="#i_invoke"><tt>invoke</tt></a>" instruction, execution continues at
3507 the beginning of the "normal" destination block. If the instruction returns
3508 a value, that value shall set the call or invoke instruction's return
3513 ret i32 5 <i>; Return an integer value of 5</i>
3514 ret void <i>; Return from a void function</i>
3515 ret { i32, i8 } { i32 4, i8 2 } <i>; Return a struct of values 4 and 2</i>
3519 <!-- _______________________________________________________________________ -->
3521 <a name="i_br">'<tt>br</tt>' Instruction</a>
3528 br i1 <cond>, label <iftrue>, label <iffalse>
3529 br label <dest> <i>; Unconditional branch</i>
3533 <p>The '<tt>br</tt>' instruction is used to cause control flow to transfer to a
3534 different basic block in the current function. There are two forms of this
3535 instruction, corresponding to a conditional branch and an unconditional
3539 <p>The conditional branch form of the '<tt>br</tt>' instruction takes a single
3540 '<tt>i1</tt>' value and two '<tt>label</tt>' values. The unconditional form
3541 of the '<tt>br</tt>' instruction takes a single '<tt>label</tt>' value as a
3545 <p>Upon execution of a conditional '<tt>br</tt>' instruction, the '<tt>i1</tt>'
3546 argument is evaluated. If the value is <tt>true</tt>, control flows to the
3547 '<tt>iftrue</tt>' <tt>label</tt> argument. If "cond" is <tt>false</tt>,
3548 control flows to the '<tt>iffalse</tt>' <tt>label</tt> argument.</p>
3553 %cond = <a href="#i_icmp">icmp</a> eq i32 %a, %b
3554 br i1 %cond, label %IfEqual, label %IfUnequal
3556 <a href="#i_ret">ret</a> i32 1
3558 <a href="#i_ret">ret</a> i32 0
3563 <!-- _______________________________________________________________________ -->
3565 <a name="i_switch">'<tt>switch</tt>' Instruction</a>
3572 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3576 <p>The '<tt>switch</tt>' instruction is used to transfer control flow to one of
3577 several different places. It is a generalization of the '<tt>br</tt>'
3578 instruction, allowing a branch to occur to one of many possible
3582 <p>The '<tt>switch</tt>' instruction uses three parameters: an integer
3583 comparison value '<tt>value</tt>', a default '<tt>label</tt>' destination,
3584 and an array of pairs of comparison value constants and '<tt>label</tt>'s.
3585 The table is not allowed to contain duplicate constant entries.</p>
3588 <p>The <tt>switch</tt> instruction specifies a table of values and
3589 destinations. When the '<tt>switch</tt>' instruction is executed, this table
3590 is searched for the given value. If the value is found, control flow is
3591 transferred to the corresponding destination; otherwise, control flow is
3592 transferred to the default destination.</p>
3594 <h5>Implementation:</h5>
3595 <p>Depending on properties of the target machine and the particular
3596 <tt>switch</tt> instruction, this instruction may be code generated in
3597 different ways. For example, it could be generated as a series of chained
3598 conditional branches or with a lookup table.</p>
3602 <i>; Emulate a conditional br instruction</i>
3603 %Val = <a href="#i_zext">zext</a> i1 %value to i32
3604 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3606 <i>; Emulate an unconditional br instruction</i>
3607 switch i32 0, label %dest [ ]
3609 <i>; Implement a jump table:</i>
3610 switch i32 %val, label %otherwise [ i32 0, label %onzero
3612 i32 2, label %ontwo ]
3618 <!-- _______________________________________________________________________ -->
3620 <a name="i_indirectbr">'<tt>indirectbr</tt>' Instruction</a>
3627 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3632 <p>The '<tt>indirectbr</tt>' instruction implements an indirect branch to a label
3633 within the current function, whose address is specified by
3634 "<tt>address</tt>". Address must be derived from a <a
3635 href="#blockaddress">blockaddress</a> constant.</p>
3639 <p>The '<tt>address</tt>' argument is the address of the label to jump to. The
3640 rest of the arguments indicate the full set of possible destinations that the
3641 address may point to. Blocks are allowed to occur multiple times in the
3642 destination list, though this isn't particularly useful.</p>
3644 <p>This destination list is required so that dataflow analysis has an accurate
3645 understanding of the CFG.</p>
3649 <p>Control transfers to the block specified in the address argument. All
3650 possible destination blocks must be listed in the label list, otherwise this
3651 instruction has undefined behavior. This implies that jumps to labels
3652 defined in other functions have undefined behavior as well.</p>
3654 <h5>Implementation:</h5>
3656 <p>This is typically implemented with a jump through a register.</p>
3660 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3666 <!-- _______________________________________________________________________ -->
3668 <a name="i_invoke">'<tt>invoke</tt>' Instruction</a>
3675 <result> = invoke [<a href="#callingconv">cconv</a>] [<a href="#paramattrs">ret attrs</a>] <ptr to function ty> <function ptr val>(<function args>) [<a href="#fnattrs">fn attrs</a>]
3676 to label <normal label> unwind label <exception label>
3680 <p>The '<tt>invoke</tt>' instruction causes control to transfer to a specified
3681 function, with the possibility of control flow transfer to either the
3682 '<tt>normal</tt>' label or the '<tt>exception</tt>' label. If the callee
3683 function returns with the "<tt><a href="#i_ret">ret</a></tt>" instruction,
3684 control flow will return to the "normal" label. If the callee (or any
3685 indirect callees) returns via the "<a href="#i_resume"><tt>resume</tt></a>"
3686 instruction or other exception handling mechanism, control is interrupted and
3687 continued at the dynamically nearest "exception" label.</p>
3689 <p>The '<tt>exception</tt>' label is a
3690 <i><a href="ExceptionHandling.html#overview">landing pad</a></i> for the
3691 exception. As such, '<tt>exception</tt>' label is required to have the
3692 "<a href="#i_landingpad"><tt>landingpad</tt></a>" instruction, which contains
3693 the information about the behavior of the program after unwinding
3694 happens, as its first non-PHI instruction. The restrictions on the
3695 "<tt>landingpad</tt>" instruction's tightly couples it to the
3696 "<tt>invoke</tt>" instruction, so that the important information contained
3697 within the "<tt>landingpad</tt>" instruction can't be lost through normal
3701 <p>This instruction requires several arguments:</p>
3704 <li>The optional "cconv" marker indicates which <a href="#callingconv">calling
3705 convention</a> the call should use. If none is specified, the call
3706 defaults to using C calling conventions.</li>
3708 <li>The optional <a href="#paramattrs">Parameter Attributes</a> list for
3709 return values. Only '<tt>zeroext</tt>', '<tt>signext</tt>', and
3710 '<tt>inreg</tt>' attributes are valid here.</li>
3712 <li>'<tt>ptr to function ty</tt>': shall be the signature of the pointer to
3713 function value being invoked. In most cases, this is a direct function
3714 invocation, but indirect <tt>invoke</tt>s are just as possible, branching
3715 off an arbitrary pointer to function value.</li>
3717 <li>'<tt>function ptr val</tt>': An LLVM value containing a pointer to a
3718 function to be invoked. </li>
3720 <li>'<tt>function args</tt>': argument list whose types match the function
3721 signature argument types and parameter attributes. All arguments must be
3722 of <a href="#t_firstclass">first class</a> type. If the function
3723 signature indicates the function accepts a variable number of arguments,
3724 the extra arguments can be specified.</li>
3726 <li>'<tt>normal label</tt>': the label reached when the called function
3727 executes a '<tt><a href="#i_ret">ret</a></tt>' instruction. </li>
3729 <li>'<tt>exception label</tt>': the label reached when a callee returns via
3730 the <a href="#i_resume"><tt>resume</tt></a> instruction or other exception
3731 handling mechanism.</li>
3733 <li>The optional <a href="#fnattrs">function attributes</a> list. Only
3734 '<tt>noreturn</tt>', '<tt>nounwind</tt>', '<tt>readonly</tt>' and
3735 '<tt>readnone</tt>' attributes are valid here.</li>
3739 <p>This instruction is designed to operate as a standard
3740 '<tt><a href="#i_call">call</a></tt>' instruction in most regards. The
3741 primary difference is that it establishes an association with a label, which
3742 is used by the runtime library to unwind the stack.</p>
3744 <p>This instruction is used in languages with destructors to ensure that proper
3745 cleanup is performed in the case of either a <tt>longjmp</tt> or a thrown
3746 exception. Additionally, this is important for implementation of
3747 '<tt>catch</tt>' clauses in high-level languages that support them.</p>
3749 <p>For the purposes of the SSA form, the definition of the value returned by the
3750 '<tt>invoke</tt>' instruction is deemed to occur on the edge from the current
3751 block to the "normal" label. If the callee unwinds then no return value is
3756 %retval = invoke i32 @Test(i32 15) to label %Continue
3757 unwind label %TestCleanup <i>; {i32}:retval set</i>
3758 %retval = invoke <a href="#callingconv">coldcc</a> i32 %Testfnptr(i32 15) to label %Continue
3759 unwind label %TestCleanup <i>; {i32}:retval set</i>
3764 <!-- _______________________________________________________________________ -->
3767 <a name="i_resume">'<tt>resume</tt>' Instruction</a>
3774 resume <type> <value>
3778 <p>The '<tt>resume</tt>' instruction is a terminator instruction that has no
3782 <p>The '<tt>resume</tt>' instruction requires one argument, which must have the
3783 same type as the result of any '<tt>landingpad</tt>' instruction in the same
3787 <p>The '<tt>resume</tt>' instruction resumes propagation of an existing
3788 (in-flight) exception whose unwinding was interrupted with
3789 a <a href="#i_landingpad"><tt>landingpad</tt></a> instruction.</p>
3793 resume { i8*, i32 } %exn
3798 <!-- _______________________________________________________________________ -->
3801 <a name="i_unreachable">'<tt>unreachable</tt>' Instruction</a>
3812 <p>The '<tt>unreachable</tt>' instruction has no defined semantics. This
3813 instruction is used to inform the optimizer that a particular portion of the
3814 code is not reachable. This can be used to indicate that the code after a
3815 no-return function cannot be reached, and other facts.</p>
3818 <p>The '<tt>unreachable</tt>' instruction has no defined semantics.</p>
3824 <!-- ======================================================================= -->
3826 <a name="binaryops">Binary Operations</a>
3831 <p>Binary operators are used to do most of the computation in a program. They
3832 require two operands of the same type, execute an operation on them, and
3833 produce a single value. The operands might represent multiple data, as is
3834 the case with the <a href="#t_vector">vector</a> data type. The result value
3835 has the same type as its operands.</p>
3837 <p>There are several different binary operators:</p>
3839 <!-- _______________________________________________________________________ -->
3841 <a name="i_add">'<tt>add</tt>' Instruction</a>
3848 <result> = add <ty> <op1>, <op2> <i>; yields {ty}:result</i>
3849 <result> = add nuw <ty> <op1>, <op2> <i>; yields {ty}:result</i>
3850 <result> = add nsw <ty> <op1>, <op2> <i>; yields {ty}:result</i>
3851 <result> = add nuw nsw <ty> <op1>, <op2> <i>; yields {ty}:result</i>
3855 <p>The '<tt>add</tt>' instruction returns the sum of its two operands.</p>
3858 <p>The two arguments to the '<tt>add</tt>' instruction must
3859 be <a href="#t_integer">integer</a> or <a href="#t_vector">vector</a> of
3860 integer values. Both arguments must have identical types.</p>
3863 <p>The value produced is the integer sum of the two operands.</p>
3865 <p>If the sum has unsigned overflow, the result returned is the mathematical
3866 result modulo 2<sup>n</sup>, where n is the bit width of the result.</p>
3868 <p>Because LLVM integers use a two's complement representation, this instruction
3869 is appropriate for both signed and unsigned integers.</p>
3871 <p><tt>nuw</tt> and <tt>nsw</tt> stand for "No Unsigned Wrap"
3872 and "No Signed Wrap", respectively. If the <tt>nuw</tt> and/or
3873 <tt>nsw</tt> keywords are present, the result value of the <tt>add</tt>
3874 is a <a href="#poisonvalues">poison value</a> if unsigned and/or signed overflow,
3875 respectively, occurs.</p>
3879 <result> = add i32 4, %var <i>; yields {i32}:result = 4 + %var</i>
3884 <!-- _______________________________________________________________________ -->
3886 <a name="i_fadd">'<tt>fadd</tt>' Instruction</a>
3893 <result> = fadd <ty> <op1>, <op2> <i>; yields {ty}:result</i>
3897 <p>The '<tt>fadd</tt>' instruction returns the sum of its two operands.</p>
3900 <p>The two arguments to the '<tt>fadd</tt>' instruction must be
3901 <a href="#t_floating">floating point</a> or <a href="#t_vector">vector</a> of
3902 floating point values. Both arguments must have identical types.</p>
3905 <p>The value produced is the floating point sum of the two operands.</p>
3909 <result> = fadd float 4.0, %var <i>; yields {float}:result = 4.0 + %var</i>
3914 <!-- _______________________________________________________________________ -->
3916 <a name="i_sub">'<tt>sub</tt>' Instruction</a>
3923 <result> = sub <ty> <op1>, <op2> <i>; yields {ty}:result</i>
3924 <result> = sub nuw <ty> <op1>, <op2> <i>; yields {ty}:result</i>
3925 <result> = sub nsw <ty> <op1>, <op2> <i>; yields {ty}:result</i>
3926 <result> = sub nuw nsw <ty> <op1>, <op2> <i>; yields {ty}:result</i>
3930 <p>The '<tt>sub</tt>' instruction returns the difference of its two
3933 <p>Note that the '<tt>sub</tt>' instruction is used to represent the
3934 '<tt>neg</tt>' instruction present in most other intermediate
3935 representations.</p>
3938 <p>The two arguments to the '<tt>sub</tt>' instruction must
3939 be <a href="#t_integer">integer</a> or <a href="#t_vector">vector</a> of
3940 integer values. Both arguments must have identical types.</p>
3943 <p>The value produced is the integer difference of the two operands.</p>
3945 <p>If the difference has unsigned overflow, the result returned is the
3946 mathematical result modulo 2<sup>n</sup>, where n is the bit width of the
3949 <p>Because LLVM integers use a two's complement representation, this instruction
3950 is appropriate for both signed and unsigned integers.</p>
3952 <p><tt>nuw</tt> and <tt>nsw</tt> stand for "No Unsigned Wrap"
3953 and "No Signed Wrap", respectively. If the <tt>nuw</tt> and/or
3954 <tt>nsw</tt> keywords are present, the result value of the <tt>sub</tt>
3955 is a <a href="#poisonvalues">poison value</a> if unsigned and/or signed overflow,
3956 respectively, occurs.</p>
3960 <result> = sub i32 4, %var <i>; yields {i32}:result = 4 - %var</i>
3961 <result> = sub i32 0, %val <i>; yields {i32}:result = -%var</i>
3966 <!-- _______________________________________________________________________ -->
3968 <a name="i_fsub">'<tt>fsub</tt>' Instruction</a>
3975 <result> = fsub <ty> <op1>, <op2> <i>; yields {ty}:result</i>
3979 <p>The '<tt>fsub</tt>' instruction returns the difference of its two
3982 <p>Note that the '<tt>fsub</tt>' instruction is used to represent the
3983 '<tt>fneg</tt>' instruction present in most other intermediate
3984 representations.</p>
3987 <p>The two arguments to the '<tt>fsub</tt>' instruction must be
3988 <a href="#t_floating">floating point</a> or <a href="#t_vector">vector</a> of
3989 floating point values. Both arguments must have identical types.</p>
3992 <p>The value produced is the floating point difference of the two operands.</p>
3996 <result> = fsub float 4.0, %var <i>; yields {float}:result = 4.0 - %var</i>
3997 <result> = fsub float -0.0, %val <i>; yields {float}:result = -%var</i>
4002 <!-- _______________________________________________________________________ -->
4004 <a name="i_mul">'<tt>mul</tt>' Instruction</a>
4011 <result> = mul <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4012 <result> = mul nuw <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4013 <result> = mul nsw <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4014 <result> = mul nuw nsw <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4018 <p>The '<tt>mul</tt>' instruction returns the product of its two operands.</p>
4021 <p>The two arguments to the '<tt>mul</tt>' instruction must
4022 be <a href="#t_integer">integer</a> or <a href="#t_vector">vector</a> of
4023 integer values. Both arguments must have identical types.</p>
4026 <p>The value produced is the integer product of the two operands.</p>
4028 <p>If the result of the multiplication has unsigned overflow, the result
4029 returned is the mathematical result modulo 2<sup>n</sup>, where n is the bit
4030 width of the result.</p>
4032 <p>Because LLVM integers use a two's complement representation, and the result
4033 is the same width as the operands, this instruction returns the correct
4034 result for both signed and unsigned integers. If a full product
4035 (e.g. <tt>i32</tt>x<tt>i32</tt>-><tt>i64</tt>) is needed, the operands should
4036 be sign-extended or zero-extended as appropriate to the width of the full
4039 <p><tt>nuw</tt> and <tt>nsw</tt> stand for "No Unsigned Wrap"
4040 and "No Signed Wrap", respectively. If the <tt>nuw</tt> and/or
4041 <tt>nsw</tt> keywords are present, the result value of the <tt>mul</tt>
4042 is a <a href="#poisonvalues">poison value</a> if unsigned and/or signed overflow,
4043 respectively, occurs.</p>
4047 <result> = mul i32 4, %var <i>; yields {i32}:result = 4 * %var</i>
4052 <!-- _______________________________________________________________________ -->
4054 <a name="i_fmul">'<tt>fmul</tt>' Instruction</a>
4061 <result> = fmul <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4065 <p>The '<tt>fmul</tt>' instruction returns the product of its two operands.</p>
4068 <p>The two arguments to the '<tt>fmul</tt>' instruction must be
4069 <a href="#t_floating">floating point</a> or <a href="#t_vector">vector</a> of
4070 floating point values. Both arguments must have identical types.</p>
4073 <p>The value produced is the floating point product of the two operands.</p>
4077 <result> = fmul float 4.0, %var <i>; yields {float}:result = 4.0 * %var</i>
4082 <!-- _______________________________________________________________________ -->
4084 <a name="i_udiv">'<tt>udiv</tt>' Instruction</a>
4091 <result> = udiv <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4092 <result> = udiv exact <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4096 <p>The '<tt>udiv</tt>' instruction returns the quotient of its two operands.</p>
4099 <p>The two arguments to the '<tt>udiv</tt>' instruction must be
4100 <a href="#t_integer">integer</a> or <a href="#t_vector">vector</a> of integer
4101 values. Both arguments must have identical types.</p>
4104 <p>The value produced is the unsigned integer quotient of the two operands.</p>
4106 <p>Note that unsigned integer division and signed integer division are distinct
4107 operations; for signed integer division, use '<tt>sdiv</tt>'.</p>
4109 <p>Division by zero leads to undefined behavior.</p>
4111 <p>If the <tt>exact</tt> keyword is present, the result value of the
4112 <tt>udiv</tt> is a <a href="#poisonvalues">poison value</a> if %op1 is not a
4113 multiple of %op2 (as such, "((a udiv exact b) mul b) == a").</p>
4118 <result> = udiv i32 4, %var <i>; yields {i32}:result = 4 / %var</i>
4123 <!-- _______________________________________________________________________ -->
4125 <a name="i_sdiv">'<tt>sdiv</tt>' Instruction</a>
4132 <result> = sdiv <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4133 <result> = sdiv exact <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4137 <p>The '<tt>sdiv</tt>' instruction returns the quotient of its two operands.</p>
4140 <p>The two arguments to the '<tt>sdiv</tt>' instruction must be
4141 <a href="#t_integer">integer</a> or <a href="#t_vector">vector</a> of integer
4142 values. Both arguments must have identical types.</p>
4145 <p>The value produced is the signed integer quotient of the two operands rounded
4148 <p>Note that signed integer division and unsigned integer division are distinct
4149 operations; for unsigned integer division, use '<tt>udiv</tt>'.</p>
4151 <p>Division by zero leads to undefined behavior. Overflow also leads to
4152 undefined behavior; this is a rare case, but can occur, for example, by doing
4153 a 32-bit division of -2147483648 by -1.</p>
4155 <p>If the <tt>exact</tt> keyword is present, the result value of the
4156 <tt>sdiv</tt> is a <a href="#poisonvalues">poison value</a> if the result would
4161 <result> = sdiv i32 4, %var <i>; yields {i32}:result = 4 / %var</i>
4166 <!-- _______________________________________________________________________ -->
4168 <a name="i_fdiv">'<tt>fdiv</tt>' Instruction</a>
4175 <result> = fdiv <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4179 <p>The '<tt>fdiv</tt>' instruction returns the quotient of its two operands.</p>
4182 <p>The two arguments to the '<tt>fdiv</tt>' instruction must be
4183 <a href="#t_floating">floating point</a> or <a href="#t_vector">vector</a> of
4184 floating point values. Both arguments must have identical types.</p>
4187 <p>The value produced is the floating point quotient of the two operands.</p>
4191 <result> = fdiv float 4.0, %var <i>; yields {float}:result = 4.0 / %var</i>
4196 <!-- _______________________________________________________________________ -->
4198 <a name="i_urem">'<tt>urem</tt>' Instruction</a>
4205 <result> = urem <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4209 <p>The '<tt>urem</tt>' instruction returns the remainder from the unsigned
4210 division of its two arguments.</p>
4213 <p>The two arguments to the '<tt>urem</tt>' instruction must be
4214 <a href="#t_integer">integer</a> or <a href="#t_vector">vector</a> of integer
4215 values. Both arguments must have identical types.</p>
4218 <p>This instruction returns the unsigned integer <i>remainder</i> of a division.
4219 This instruction always performs an unsigned division to get the
4222 <p>Note that unsigned integer remainder and signed integer remainder are
4223 distinct operations; for signed integer remainder, use '<tt>srem</tt>'.</p>
4225 <p>Taking the remainder of a division by zero leads to undefined behavior.</p>
4229 <result> = urem i32 4, %var <i>; yields {i32}:result = 4 % %var</i>
4234 <!-- _______________________________________________________________________ -->
4236 <a name="i_srem">'<tt>srem</tt>' Instruction</a>
4243 <result> = srem <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4247 <p>The '<tt>srem</tt>' instruction returns the remainder from the signed
4248 division of its two operands. This instruction can also take
4249 <a href="#t_vector">vector</a> versions of the values in which case the
4250 elements must be integers.</p>
4253 <p>The two arguments to the '<tt>srem</tt>' instruction must be
4254 <a href="#t_integer">integer</a> or <a href="#t_vector">vector</a> of integer
4255 values. Both arguments must have identical types.</p>
4258 <p>This instruction returns the <i>remainder</i> of a division (where the result
4259 is either zero or has the same sign as the dividend, <tt>op1</tt>), not the
4260 <i>modulo</i> operator (where the result is either zero or has the same sign
4261 as the divisor, <tt>op2</tt>) of a value.
4262 For more information about the difference,
4263 see <a href="http://mathforum.org/dr.math/problems/anne.4.28.99.html">The
4264 Math Forum</a>. For a table of how this is implemented in various languages,
4265 please see <a href="http://en.wikipedia.org/wiki/Modulo_operation">
4266 Wikipedia: modulo operation</a>.</p>
4268 <p>Note that signed integer remainder and unsigned integer remainder are
4269 distinct operations; for unsigned integer remainder, use '<tt>urem</tt>'.</p>
4271 <p>Taking the remainder of a division by zero leads to undefined behavior.
4272 Overflow also leads to undefined behavior; this is a rare case, but can
4273 occur, for example, by taking the remainder of a 32-bit division of
4274 -2147483648 by -1. (The remainder doesn't actually overflow, but this rule
4275 lets srem be implemented using instructions that return both the result of
4276 the division and the remainder.)</p>
4280 <result> = srem i32 4, %var <i>; yields {i32}:result = 4 % %var</i>
4285 <!-- _______________________________________________________________________ -->
4287 <a name="i_frem">'<tt>frem</tt>' Instruction</a>
4294 <result> = frem <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4298 <p>The '<tt>frem</tt>' instruction returns the remainder from the division of
4299 its two operands.</p>
4302 <p>The two arguments to the '<tt>frem</tt>' instruction must be
4303 <a href="#t_floating">floating point</a> or <a href="#t_vector">vector</a> of
4304 floating point values. Both arguments must have identical types.</p>
4307 <p>This instruction returns the <i>remainder</i> of a division. The remainder
4308 has the same sign as the dividend.</p>
4312 <result> = frem float 4.0, %var <i>; yields {float}:result = 4.0 % %var</i>
4319 <!-- ======================================================================= -->
4321 <a name="bitwiseops">Bitwise Binary Operations</a>
4326 <p>Bitwise binary operators are used to do various forms of bit-twiddling in a
4327 program. They are generally very efficient instructions and can commonly be
4328 strength reduced from other instructions. They require two operands of the
4329 same type, execute an operation on them, and produce a single value. The
4330 resulting value is the same type as its operands.</p>
4332 <!-- _______________________________________________________________________ -->
4334 <a name="i_shl">'<tt>shl</tt>' Instruction</a>
4341 <result> = shl <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4342 <result> = shl nuw <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4343 <result> = shl nsw <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4344 <result> = shl nuw nsw <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4348 <p>The '<tt>shl</tt>' instruction returns the first operand shifted to the left
4349 a specified number of bits.</p>
4352 <p>Both arguments to the '<tt>shl</tt>' instruction must be the
4353 same <a href="#t_integer">integer</a> or <a href="#t_vector">vector</a> of
4354 integer type. '<tt>op2</tt>' is treated as an unsigned value.</p>
4357 <p>The value produced is <tt>op1</tt> * 2<sup><tt>op2</tt></sup> mod
4358 2<sup>n</sup>, where <tt>n</tt> is the width of the result. If <tt>op2</tt>
4359 is (statically or dynamically) negative or equal to or larger than the number
4360 of bits in <tt>op1</tt>, the result is undefined. If the arguments are
4361 vectors, each vector element of <tt>op1</tt> is shifted by the corresponding
4362 shift amount in <tt>op2</tt>.</p>
4364 <p>If the <tt>nuw</tt> keyword is present, then the shift produces a
4365 <a href="#poisonvalues">poison value</a> if it shifts out any non-zero bits. If
4366 the <tt>nsw</tt> keyword is present, then the shift produces a
4367 <a href="#poisonvalues">poison value</a> if it shifts out any bits that disagree
4368 with the resultant sign bit. As such, NUW/NSW have the same semantics as
4369 they would if the shift were expressed as a mul instruction with the same
4370 nsw/nuw bits in (mul %op1, (shl 1, %op2)).</p>
4374 <result> = shl i32 4, %var <i>; yields {i32}: 4 << %var</i>
4375 <result> = shl i32 4, 2 <i>; yields {i32}: 16</i>
4376 <result> = shl i32 1, 10 <i>; yields {i32}: 1024</i>
4377 <result> = shl i32 1, 32 <i>; undefined</i>
4378 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> <i>; yields: result=<2 x i32> < i32 2, i32 4></i>
4383 <!-- _______________________________________________________________________ -->
4385 <a name="i_lshr">'<tt>lshr</tt>' Instruction</a>
4392 <result> = lshr <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4393 <result> = lshr exact <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4397 <p>The '<tt>lshr</tt>' instruction (logical shift right) returns the first
4398 operand shifted to the right a specified number of bits with zero fill.</p>
4401 <p>Both arguments to the '<tt>lshr</tt>' instruction must be the same
4402 <a href="#t_integer">integer</a> or <a href="#t_vector">vector</a> of integer
4403 type. '<tt>op2</tt>' is treated as an unsigned value.</p>
4406 <p>This instruction always performs a logical shift right operation. The most
4407 significant bits of the result will be filled with zero bits after the shift.
4408 If <tt>op2</tt> is (statically or dynamically) equal to or larger than the
4409 number of bits in <tt>op1</tt>, the result is undefined. If the arguments are
4410 vectors, each vector element of <tt>op1</tt> is shifted by the corresponding
4411 shift amount in <tt>op2</tt>.</p>
4413 <p>If the <tt>exact</tt> keyword is present, the result value of the
4414 <tt>lshr</tt> is a <a href="#poisonvalues">poison value</a> if any of the bits
4415 shifted out are non-zero.</p>
4420 <result> = lshr i32 4, 1 <i>; yields {i32}:result = 2</i>
4421 <result> = lshr i32 4, 2 <i>; yields {i32}:result = 1</i>
4422 <result> = lshr i8 4, 3 <i>; yields {i8}:result = 0</i>
4423 <result> = lshr i8 -2, 1 <i>; yields {i8}:result = 0x7FFFFFFF </i>
4424 <result> = lshr i32 1, 32 <i>; undefined</i>
4425 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> <i>; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1></i>
4430 <!-- _______________________________________________________________________ -->
4432 <a name="i_ashr">'<tt>ashr</tt>' Instruction</a>
4439 <result> = ashr <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4440 <result> = ashr exact <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4444 <p>The '<tt>ashr</tt>' instruction (arithmetic shift right) returns the first
4445 operand shifted to the right a specified number of bits with sign
4449 <p>Both arguments to the '<tt>ashr</tt>' instruction must be the same
4450 <a href="#t_integer">integer</a> or <a href="#t_vector">vector</a> of integer
4451 type. '<tt>op2</tt>' is treated as an unsigned value.</p>
4454 <p>This instruction always performs an arithmetic shift right operation, The
4455 most significant bits of the result will be filled with the sign bit
4456 of <tt>op1</tt>. If <tt>op2</tt> is (statically or dynamically) equal to or
4457 larger than the number of bits in <tt>op1</tt>, the result is undefined. If
4458 the arguments are vectors, each vector element of <tt>op1</tt> is shifted by
4459 the corresponding shift amount in <tt>op2</tt>.</p>
4461 <p>If the <tt>exact</tt> keyword is present, the result value of the
4462 <tt>ashr</tt> is a <a href="#poisonvalues">poison value</a> if any of the bits
4463 shifted out are non-zero.</p>
4467 <result> = ashr i32 4, 1 <i>; yields {i32}:result = 2</i>
4468 <result> = ashr i32 4, 2 <i>; yields {i32}:result = 1</i>
4469 <result> = ashr i8 4, 3 <i>; yields {i8}:result = 0</i>
4470 <result> = ashr i8 -2, 1 <i>; yields {i8}:result = -1</i>
4471 <result> = ashr i32 1, 32 <i>; undefined</i>
4472 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> <i>; yields: result=<2 x i32> < i32 -1, i32 0></i>
4477 <!-- _______________________________________________________________________ -->
4479 <a name="i_and">'<tt>and</tt>' Instruction</a>
4486 <result> = and <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4490 <p>The '<tt>and</tt>' instruction returns the bitwise logical and of its two
4494 <p>The two arguments to the '<tt>and</tt>' instruction must be
4495 <a href="#t_integer">integer</a> or <a href="#t_vector">vector</a> of integer
4496 values. Both arguments must have identical types.</p>
4499 <p>The truth table used for the '<tt>and</tt>' instruction is:</p>
4501 <table border="1" cellspacing="0" cellpadding="4">
4533 <result> = and i32 4, %var <i>; yields {i32}:result = 4 & %var</i>
4534 <result> = and i32 15, 40 <i>; yields {i32}:result = 8</i>
4535 <result> = and i32 4, 8 <i>; yields {i32}:result = 0</i>
4538 <!-- _______________________________________________________________________ -->
4540 <a name="i_or">'<tt>or</tt>' Instruction</a>
4547 <result> = or <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4551 <p>The '<tt>or</tt>' instruction returns the bitwise logical inclusive or of its
4555 <p>The two arguments to the '<tt>or</tt>' instruction must be
4556 <a href="#t_integer">integer</a> or <a href="#t_vector">vector</a> of integer
4557 values. Both arguments must have identical types.</p>
4560 <p>The truth table used for the '<tt>or</tt>' instruction is:</p>
4562 <table border="1" cellspacing="0" cellpadding="4">
4594 <result> = or i32 4, %var <i>; yields {i32}:result = 4 | %var</i>
4595 <result> = or i32 15, 40 <i>; yields {i32}:result = 47</i>
4596 <result> = or i32 4, 8 <i>; yields {i32}:result = 12</i>
4601 <!-- _______________________________________________________________________ -->
4603 <a name="i_xor">'<tt>xor</tt>' Instruction</a>
4610 <result> = xor <ty> <op1>, <op2> <i>; yields {ty}:result</i>
4614 <p>The '<tt>xor</tt>' instruction returns the bitwise logical exclusive or of
4615 its two operands. The <tt>xor</tt> is used to implement the "one's
4616 complement" operation, which is the "~" operator in C.</p>
4619 <p>The two arguments to the '<tt>xor</tt>' instruction must be
4620 <a href="#t_integer">integer</a> or <a href="#t_vector">vector</a> of integer
4621 values. Both arguments must have identical types.</p>
4624 <p>The truth table used for the '<tt>xor</tt>' instruction is:</p>
4626 <table border="1" cellspacing="0" cellpadding="4">
4658 <result> = xor i32 4, %var <i>; yields {i32}:result = 4 ^ %var</i>
4659 <result> = xor i32 15, 40 <i>; yields {i32}:result = 39</i>
4660 <result> = xor i32 4, 8 <i>; yields {i32}:result = 12</i>
4661 <result> = xor i32 %V, -1 <i>; yields {i32}:result = ~%V</i>
4668 <!-- ======================================================================= -->
4670 <a name="vectorops">Vector Operations</a>
4675 <p>LLVM supports several instructions to represent vector operations in a
4676 target-independent manner. These instructions cover the element-access and
4677 vector-specific operations needed to process vectors effectively. While LLVM
4678 does directly support these vector operations, many sophisticated algorithms
4679 will want to use target-specific intrinsics to take full advantage of a
4680 specific target.</p>
4682 <!-- _______________________________________________________________________ -->
4684 <a name="i_extractelement">'<tt>extractelement</tt>' Instruction</a>
4691 <result> = extractelement <n x <ty>> <val>, i32 <idx> <i>; yields <ty></i>
4695 <p>The '<tt>extractelement</tt>' instruction extracts a single scalar element
4696 from a vector at a specified index.</p>
4700 <p>The first operand of an '<tt>extractelement</tt>' instruction is a value
4701 of <a href="#t_vector">vector</a> type. The second operand is an index
4702 indicating the position from which to extract the element. The index may be
4706 <p>The result is a scalar of the same type as the element type of
4707 <tt>val</tt>. Its value is the value at position <tt>idx</tt> of
4708 <tt>val</tt>. If <tt>idx</tt> exceeds the length of <tt>val</tt>, the
4709 results are undefined.</p>
4713 <result> = extractelement <4 x i32> %vec, i32 0 <i>; yields i32</i>
4718 <!-- _______________________________________________________________________ -->
4720 <a name="i_insertelement">'<tt>insertelement</tt>' Instruction</a>
4727 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> <i>; yields <n x <ty>></i>
4731 <p>The '<tt>insertelement</tt>' instruction inserts a scalar element into a
4732 vector at a specified index.</p>
4735 <p>The first operand of an '<tt>insertelement</tt>' instruction is a value
4736 of <a href="#t_vector">vector</a> type. The second operand is a scalar value
4737 whose type must equal the element type of the first operand. The third
4738 operand is an index indicating the position at which to insert the value.
4739 The index may be a variable.</p>
4742 <p>The result is a vector of the same type as <tt>val</tt>. Its element values
4743 are those of <tt>val</tt> except at position <tt>idx</tt>, where it gets the
4744 value <tt>elt</tt>. If <tt>idx</tt> exceeds the length of <tt>val</tt>, the
4745 results are undefined.</p>
4749 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 <i>; yields <4 x i32></i>
4754 <!-- _______________________________________________________________________ -->
4756 <a name="i_shufflevector">'<tt>shufflevector</tt>' Instruction</a>
4763 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> <i>; yields <m x <ty>></i>
4767 <p>The '<tt>shufflevector</tt>' instruction constructs a permutation of elements
4768 from two input vectors, returning a vector with the same element type as the
4769 input and length that is the same as the shuffle mask.</p>
4772 <p>The first two operands of a '<tt>shufflevector</tt>' instruction are vectors
4773 with the same type. The third argument is a shuffle mask whose
4774 element type is always 'i32'. The result of the instruction is a vector
4775 whose length is the same as the shuffle mask and whose element type is the
4776 same as the element type of the first two operands.</p>
4778 <p>The shuffle mask operand is required to be a constant vector with either
4779 constant integer or undef values.</p>
4782 <p>The elements of the two input vectors are numbered from left to right across
4783 both of the vectors. The shuffle mask operand specifies, for each element of
4784 the result vector, which element of the two input vectors the result element
4785 gets. The element selector may be undef (meaning "don't care") and the
4786 second operand may be undef if performing a shuffle from only one vector.</p>
4790 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4791 <4 x i32> <i32 0, i32 4, i32 1, i32 5> <i>; yields <4 x i32></i>
4792 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4793 <4 x i32> <i32 0, i32 1, i32 2, i32 3> <i>; yields <4 x i32></i> - Identity shuffle.
4794 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4795 <4 x i32> <i32 0, i32 1, i32 2, i32 3> <i>; yields <4 x i32></i>
4796 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4797 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > <i>; yields <8 x i32></i>
4804 <!-- ======================================================================= -->
4806 <a name="aggregateops">Aggregate Operations</a>
4811 <p>LLVM supports several instructions for working with
4812 <a href="#t_aggregate">aggregate</a> values.</p>
4814 <!-- _______________________________________________________________________ -->
4816 <a name="i_extractvalue">'<tt>extractvalue</tt>' Instruction</a>
4823 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4827 <p>The '<tt>extractvalue</tt>' instruction extracts the value of a member field
4828 from an <a href="#t_aggregate">aggregate</a> value.</p>
4831 <p>The first operand of an '<tt>extractvalue</tt>' instruction is a value
4832 of <a href="#t_struct">struct</a> or
4833 <a href="#t_array">array</a> type. The operands are constant indices to
4834 specify which value to extract in a similar manner as indices in a
4835 '<tt><a href="#i_getelementptr">getelementptr</a></tt>' instruction.</p>
4836 <p>The major differences to <tt>getelementptr</tt> indexing are:</p>
4838 <li>Since the value being indexed is not a pointer, the first index is
4839 omitted and assumed to be zero.</li>
4840 <li>At least one index must be specified.</li>
4841 <li>Not only struct indices but also array indices must be in
4846 <p>The result is the value at the position in the aggregate specified by the
4851 <result> = extractvalue {i32, float} %agg, 0 <i>; yields i32</i>
4856 <!-- _______________________________________________________________________ -->
4858 <a name="i_insertvalue">'<tt>insertvalue</tt>' Instruction</a>
4865 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* <i>; yields <aggregate type></i>
4869 <p>The '<tt>insertvalue</tt>' instruction inserts a value into a member field
4870 in an <a href="#t_aggregate">aggregate</a> value.</p>
4873 <p>The first operand of an '<tt>insertvalue</tt>' instruction is a value
4874 of <a href="#t_struct">struct</a> or
4875 <a href="#t_array">array</a> type. The second operand is a first-class
4876 value to insert. The following operands are constant indices indicating
4877 the position at which to insert the value in a similar manner as indices in a
4878 '<tt><a href="#i_extractvalue">extractvalue</a></tt>' instruction. The
4879 value to insert must have the same type as the value identified by the
4883 <p>The result is an aggregate of the same type as <tt>val</tt>. Its value is
4884 that of <tt>val</tt> except that the value at the position specified by the
4885 indices is that of <tt>elt</tt>.</p>
4889 %agg1 = insertvalue {i32, float} undef, i32 1, 0 <i>; yields {i32 1, float undef}</i>
4890 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 <i>; yields {i32 1, float %val}</i>
4891 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 <i>; yields {i32 1, float %val}</i>
4898 <!-- ======================================================================= -->
4900 <a name="memoryops">Memory Access and Addressing Operations</a>
4905 <p>A key design point of an SSA-based representation is how it represents
4906 memory. In LLVM, no memory locations are in SSA form, which makes things
4907 very simple. This section describes how to read, write, and allocate
4910 <!-- _______________________________________________________________________ -->
4912 <a name="i_alloca">'<tt>alloca</tt>' Instruction</a>
4919 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] <i>; yields {type*}:result</i>
4923 <p>The '<tt>alloca</tt>' instruction allocates memory on the stack frame of the
4924 currently executing function, to be automatically released when this function
4925 returns to its caller. The object is always allocated in the generic address
4926 space (address space zero).</p>
4929 <p>The '<tt>alloca</tt>' instruction
4930 allocates <tt>sizeof(<type>)*NumElements</tt> bytes of memory on the
4931 runtime stack, returning a pointer of the appropriate type to the program.
4932 If "NumElements" is specified, it is the number of elements allocated,
4933 otherwise "NumElements" is defaulted to be one. If a constant alignment is
4934 specified, the value result of the allocation is guaranteed to be aligned to
4935 at least that boundary. If not specified, or if zero, the target can choose
4936 to align the allocation on any convenient boundary compatible with the
4939 <p>'<tt>type</tt>' may be any sized type.</p>
4942 <p>Memory is allocated; a pointer is returned. The operation is undefined if
4943 there is insufficient stack space for the allocation. '<tt>alloca</tt>'d
4944 memory is automatically released when the function returns. The
4945 '<tt>alloca</tt>' instruction is commonly used to represent automatic
4946 variables that must have an address available. When the function returns
4947 (either with the <tt><a href="#i_ret">ret</a></tt>
4948 or <tt><a href="#i_resume">resume</a></tt> instructions), the memory is
4949 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4950 The order in which memory is allocated (ie., which way the stack grows) is
4957 %ptr = alloca i32 <i>; yields {i32*}:ptr</i>
4958 %ptr = alloca i32, i32 4 <i>; yields {i32*}:ptr</i>
4959 %ptr = alloca i32, i32 4, align 1024 <i>; yields {i32*}:ptr</i>
4960 %ptr = alloca i32, align 1024 <i>; yields {i32*}:ptr</i>
4965 <!-- _______________________________________________________________________ -->
4967 <a name="i_load">'<tt>load</tt>' Instruction</a>
4974 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4975 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4976 !<index> = !{ i32 1 }
4980 <p>The '<tt>load</tt>' instruction is used to read from memory.</p>
4983 <p>The argument to the '<tt>load</tt>' instruction specifies the memory address
4984 from which to load. The pointer must point to
4985 a <a href="#t_firstclass">first class</a> type. If the <tt>load</tt> is
4986 marked as <tt>volatile</tt>, then the optimizer is not allowed to modify the
4987 number or order of execution of this <tt>load</tt> with other <a
4988 href="#volatile">volatile operations</a>.</p>
4990 <p>If the <code>load</code> is marked as <code>atomic</code>, it takes an extra
4991 <a href="#ordering">ordering</a> and optional <code>singlethread</code>
4992 argument. The <code>release</code> and <code>acq_rel</code> orderings are
4993 not valid on <code>load</code> instructions. Atomic loads produce <a
4994 href="#memorymodel">defined</a> results when they may see multiple atomic
4995 stores. The type of the pointee must be an integer type whose bit width
4996 is a power of two greater than or equal to eight and less than or equal
4997 to a target-specific size limit. <code>align</code> must be explicitly
4998 specified on atomic loads, and the load has undefined behavior if the
4999 alignment is not set to a value which is at least the size in bytes of
5000 the pointee. <code>!nontemporal</code> does not have any defined semantics
5001 for atomic loads.</p>
5003 <p>The optional constant <tt>align</tt> argument specifies the alignment of the
5004 operation (that is, the alignment of the memory address). A value of 0 or an
5005 omitted <tt>align</tt> argument means that the operation has the preferential
5006 alignment for the target. It is the responsibility of the code emitter to
5007 ensure that the alignment information is correct. Overestimating the
5008 alignment results in undefined behavior. Underestimating the alignment may
5009 produce less efficient code. An alignment of 1 is always safe.</p>
5011 <p>The optional <tt>!nontemporal</tt> metadata must reference a single
5012 metatadata name <index> corresponding to a metadata node with
5013 one <tt>i32</tt> entry of value 1. The existence of
5014 the <tt>!nontemporal</tt> metatadata on the instruction tells the optimizer
5015 and code generator that this load is not expected to be reused in the cache.
5016 The code generator may select special instructions to save cache bandwidth,
5017 such as the <tt>MOVNT</tt> instruction on x86.</p>
5019 <p>The optional <tt>!invariant.load</tt> metadata must reference a single
5020 metatadata name <index> corresponding to a metadata node with no
5021 entries. The existence of the <tt>!invariant.load</tt> metatadata on the
5022 instruction tells the optimizer and code generator that this load address
5023 points to memory which does not change value during program execution.
5024 The optimizer may then move this load around, for example, by hoisting it
5025 out of loops using loop invariant code motion.</p>
5028 <p>The location of memory pointed to is loaded. If the value being loaded is of
5029 scalar type then the number of bytes read does not exceed the minimum number
5030 of bytes needed to hold all bits of the type. For example, loading an
5031 <tt>i24</tt> reads at most three bytes. When loading a value of a type like
5032 <tt>i20</tt> with a size that is not an integral number of bytes, the result
5033 is undefined if the value was not originally written using a store of the
5038 %ptr = <a href="#i_alloca">alloca</a> i32 <i>; yields {i32*}:ptr</i>
5039 <a href="#i_store">store</a> i32 3, i32* %ptr <i>; yields {void}</i>
5040 %val = load i32* %ptr <i>; yields {i32}:val = i32 3</i>
5045 <!-- _______________________________________________________________________ -->
5047 <a name="i_store">'<tt>store</tt>' Instruction</a>
5054 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] <i>; yields {void}</i>
5055 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> <i>; yields {void}</i>
5059 <p>The '<tt>store</tt>' instruction is used to write to memory.</p>
5062 <p>There are two arguments to the '<tt>store</tt>' instruction: a value to store
5063 and an address at which to store it. The type of the
5064 '<tt><pointer></tt>' operand must be a pointer to
5065 the <a href="#t_firstclass">first class</a> type of the
5066 '<tt><value></tt>' operand. If the <tt>store</tt> is marked as
5067 <tt>volatile</tt>, then the optimizer is not allowed to modify the number or
5068 order of execution of this <tt>store</tt> with other <a
5069 href="#volatile">volatile operations</a>.</p>
5071 <p>If the <code>store</code> is marked as <code>atomic</code>, it takes an extra
5072 <a href="#ordering">ordering</a> and optional <code>singlethread</code>
5073 argument. The <code>acquire</code> and <code>acq_rel</code> orderings aren't
5074 valid on <code>store</code> instructions. Atomic loads produce <a
5075 href="#memorymodel">defined</a> results when they may see multiple atomic
5076 stores. The type of the pointee must be an integer type whose bit width
5077 is a power of two greater than or equal to eight and less than or equal
5078 to a target-specific size limit. <code>align</code> must be explicitly
5079 specified on atomic stores, and the store has undefined behavior if the
5080 alignment is not set to a value which is at least the size in bytes of
5081 the pointee. <code>!nontemporal</code> does not have any defined semantics
5082 for atomic stores.</p>
5084 <p>The optional constant "align" argument specifies the alignment of the
5085 operation (that is, the alignment of the memory address). A value of 0 or an
5086 omitted "align" argument means that the operation has the preferential
5087 alignment for the target. It is the responsibility of the code emitter to
5088 ensure that the alignment information is correct. Overestimating the
5089 alignment results in an undefined behavior. Underestimating the alignment may
5090 produce less efficient code. An alignment of 1 is always safe.</p>
5092 <p>The optional !nontemporal metadata must reference a single metatadata
5093 name <index> corresponding to a metadata node with one i32 entry of
5094 value 1. The existence of the !nontemporal metatadata on the
5095 instruction tells the optimizer and code generator that this load is
5096 not expected to be reused in the cache. The code generator may
5097 select special instructions to save cache bandwidth, such as the
5098 MOVNT instruction on x86.</p>
5102 <p>The contents of memory are updated to contain '<tt><value></tt>' at the
5103 location specified by the '<tt><pointer></tt>' operand. If
5104 '<tt><value></tt>' is of scalar type then the number of bytes written
5105 does not exceed the minimum number of bytes needed to hold all bits of the
5106 type. For example, storing an <tt>i24</tt> writes at most three bytes. When
5107 writing a value of a type like <tt>i20</tt> with a size that is not an
5108 integral number of bytes, it is unspecified what happens to the extra bits
5109 that do not belong to the type, but they will typically be overwritten.</p>
5113 %ptr = <a href="#i_alloca">alloca</a> i32 <i>; yields {i32*}:ptr</i>
5114 store i32 3, i32* %ptr <i>; yields {void}</i>
5115 %val = <a href="#i_load">load</a> i32* %ptr <i>; yields {i32}:val = i32 3</i>
5120 <!-- _______________________________________________________________________ -->
5122 <a name="i_fence">'<tt>fence</tt>' Instruction</a>
5129 fence [singlethread] <ordering> <i>; yields {void}</i>
5133 <p>The '<tt>fence</tt>' instruction is used to introduce happens-before edges
5134 between operations.</p>
5136 <h5>Arguments:</h5> <p>'<code>fence</code>' instructions take an <a
5137 href="#ordering">ordering</a> argument which defines what
5138 <i>synchronizes-with</i> edges they add. They can only be given
5139 <code>acquire</code>, <code>release</code>, <code>acq_rel</code>, and
5140 <code>seq_cst</code> orderings.</p>
5143 <p>A fence <var>A</var> which has (at least) <code>release</code> ordering
5144 semantics <i>synchronizes with</i> a fence <var>B</var> with (at least)
5145 <code>acquire</code> ordering semantics if and only if there exist atomic
5146 operations <var>X</var> and <var>Y</var>, both operating on some atomic object
5147 <var>M</var>, such that <var>A</var> is sequenced before <var>X</var>,
5148 <var>X</var> modifies <var>M</var> (either directly or through some side effect
5149 of a sequence headed by <var>X</var>), <var>Y</var> is sequenced before
5150 <var>B</var>, and <var>Y</var> observes <var>M</var>. This provides a
5151 <i>happens-before</i> dependency between <var>A</var> and <var>B</var>. Rather
5152 than an explicit <code>fence</code>, one (but not both) of the atomic operations
5153 <var>X</var> or <var>Y</var> might provide a <code>release</code> or
5154 <code>acquire</code> (resp.) ordering constraint and still
5155 <i>synchronize-with</i> the explicit <code>fence</code> and establish the
5156 <i>happens-before</i> edge.</p>
5158 <p>A <code>fence</code> which has <code>seq_cst</code> ordering, in addition to
5159 having both <code>acquire</code> and <code>release</code> semantics specified
5160 above, participates in the global program order of other <code>seq_cst</code>
5161 operations and/or fences.</p>
5163 <p>The optional "<a href="#singlethread"><code>singlethread</code></a>" argument
5164 specifies that the fence only synchronizes with other fences in the same
5165 thread. (This is useful for interacting with signal handlers.)</p>
5169 fence acquire <i>; yields {void}</i>
5170 fence singlethread seq_cst <i>; yields {void}</i>
5175 <!-- _______________________________________________________________________ -->
5177 <a name="i_cmpxchg">'<tt>cmpxchg</tt>' Instruction</a>
5184 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> <i>; yields {ty}</i>
5188 <p>The '<tt>cmpxchg</tt>' instruction is used to atomically modify memory.
5189 It loads a value in memory and compares it to a given value. If they are
5190 equal, it stores a new value into the memory.</p>
5193 <p>There are three arguments to the '<code>cmpxchg</code>' instruction: an
5194 address to operate on, a value to compare to the value currently be at that
5195 address, and a new value to place at that address if the compared values are
5196 equal. The type of '<var><cmp></var>' must be an integer type whose
5197 bit width is a power of two greater than or equal to eight and less than
5198 or equal to a target-specific size limit. '<var><cmp></var>' and
5199 '<var><new></var>' must have the same type, and the type of
5200 '<var><pointer></var>' must be a pointer to that type. If the
5201 <code>cmpxchg</code> is marked as <code>volatile</code>, then the
5202 optimizer is not allowed to modify the number or order of execution
5203 of this <code>cmpxchg</code> with other <a href="#volatile">volatile
5206 <!-- FIXME: Extend allowed types. -->
5208 <p>The <a href="#ordering"><var>ordering</var></a> argument specifies how this
5209 <code>cmpxchg</code> synchronizes with other atomic operations.</p>
5211 <p>The optional "<code>singlethread</code>" argument declares that the
5212 <code>cmpxchg</code> is only atomic with respect to code (usually signal
5213 handlers) running in the same thread as the <code>cmpxchg</code>. Otherwise the
5214 cmpxchg is atomic with respect to all other code in the system.</p>
5216 <p>The pointer passed into cmpxchg must have alignment greater than or equal to
5217 the size in memory of the operand.
5220 <p>The contents of memory at the location specified by the
5221 '<tt><pointer></tt>' operand is read and compared to
5222 '<tt><cmp></tt>'; if the read value is the equal,
5223 '<tt><new></tt>' is written. The original value at the location
5226 <p>A successful <code>cmpxchg</code> is a read-modify-write instruction for the
5227 purpose of identifying <a href="#release_sequence">release sequences</a>. A
5228 failed <code>cmpxchg</code> is equivalent to an atomic load with an ordering
5229 parameter determined by dropping any <code>release</code> part of the
5230 <code>cmpxchg</code>'s ordering.</p>
5233 FIXME: Is compare_exchange_weak() necessary? (Consider after we've done
5234 optimization work on ARM.)
5236 FIXME: Is a weaker ordering constraint on failure helpful in practice?
5242 %orig = atomic <a href="#i_load">load</a> i32* %ptr unordered <i>; yields {i32}</i>
5243 <a href="#i_br">br</a> label %loop
5246 %cmp = <a href="#i_phi">phi</a> i32 [ %orig, %entry ], [%old, %loop]
5247 %squared = <a href="#i_mul">mul</a> i32 %cmp, %cmp
5248 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared <i>; yields {i32}</i>
5249 %success = <a href="#i_icmp">icmp</a> eq i32 %cmp, %old
5250 <a href="#i_br">br</a> i1 %success, label %done, label %loop
5258 <!-- _______________________________________________________________________ -->
5260 <a name="i_atomicrmw">'<tt>atomicrmw</tt>' Instruction</a>
5267 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> <i>; yields {ty}</i>
5271 <p>The '<tt>atomicrmw</tt>' instruction is used to atomically modify memory.</p>
5274 <p>There are three arguments to the '<code>atomicrmw</code>' instruction: an
5275 operation to apply, an address whose value to modify, an argument to the
5276 operation. The operation must be one of the following keywords:</p>
5291 <p>The type of '<var><value></var>' must be an integer type whose
5292 bit width is a power of two greater than or equal to eight and less than
5293 or equal to a target-specific size limit. The type of the
5294 '<code><pointer></code>' operand must be a pointer to that type.
5295 If the <code>atomicrmw</code> is marked as <code>volatile</code>, then the
5296 optimizer is not allowed to modify the number or order of execution of this
5297 <code>atomicrmw</code> with other <a href="#volatile">volatile
5300 <!-- FIXME: Extend allowed types. -->
5303 <p>The contents of memory at the location specified by the
5304 '<tt><pointer></tt>' operand are atomically read, modified, and written
5305 back. The original value at the location is returned. The modification is
5306 specified by the <var>operation</var> argument:</p>
5309 <li>xchg: <code>*ptr = val</code></li>
5310 <li>add: <code>*ptr = *ptr + val</code></li>
5311 <li>sub: <code>*ptr = *ptr - val</code></li>
5312 <li>and: <code>*ptr = *ptr & val</code></li>
5313 <li>nand: <code>*ptr = ~(*ptr & val)</code></li>
5314 <li>or: <code>*ptr = *ptr | val</code></li>
5315 <li>xor: <code>*ptr = *ptr ^ val</code></li>
5316 <li>max: <code>*ptr = *ptr > val ? *ptr : val</code> (using a signed comparison)</li>
5317 <li>min: <code>*ptr = *ptr < val ? *ptr : val</code> (using a signed comparison)</li>
5318 <li>umax: <code>*ptr = *ptr > val ? *ptr : val</code> (using an unsigned comparison)</li>
5319 <li>umin: <code>*ptr = *ptr < val ? *ptr : val</code> (using an unsigned comparison)</li>
5324 %old = atomicrmw add i32* %ptr, i32 1 acquire <i>; yields {i32}</i>
5329 <!-- _______________________________________________________________________ -->
5331 <a name="i_getelementptr">'<tt>getelementptr</tt>' Instruction</a>
5338 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5339 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5340 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5344 <p>The '<tt>getelementptr</tt>' instruction is used to get the address of a
5345 subelement of an <a href="#t_aggregate">aggregate</a> data structure.
5346 It performs address calculation only and does not access memory.</p>
5349 <p>The first argument is always a pointer or a vector of pointers,
5350 and forms the basis of the
5351 calculation. The remaining arguments are indices that indicate which of the
5352 elements of the aggregate object are indexed. The interpretation of each
5353 index is dependent on the type being indexed into. The first index always
5354 indexes the pointer value given as the first argument, the second index
5355 indexes a value of the type pointed to (not necessarily the value directly
5356 pointed to, since the first index can be non-zero), etc. The first type
5357 indexed into must be a pointer value, subsequent types can be arrays,
5358 vectors, and structs. Note that subsequent types being indexed into
5359 can never be pointers, since that would require loading the pointer before
5360 continuing calculation.</p>
5362 <p>The type of each index argument depends on the type it is indexing into.
5363 When indexing into a (optionally packed) structure, only <tt>i32</tt>
5364 integer <b>constants</b> are allowed. When indexing into an array, pointer
5365 or vector, integers of any width are allowed, and they are not required to be
5366 constant. These integers are treated as signed values where relevant.</p>
5368 <p>For example, let's consider a C code fragment and how it gets compiled to
5371 <pre class="doc_code">
5383 int *foo(struct ST *s) {
5384 return &s[1].Z.B[5][13];
5388 <p>The LLVM code generated by Clang is:</p>
5390 <pre class="doc_code">
5391 %struct.RT = <a href="#namedtypes">type</a> { i8, [10 x [20 x i32]], i8 }
5392 %struct.ST = <a href="#namedtypes">type</a> { i32, double, %struct.RT }
5394 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5396 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5402 <p>In the example above, the first index is indexing into the
5403 '<tt>%struct.ST*</tt>' type, which is a pointer, yielding a
5404 '<tt>%struct.ST</tt>' = '<tt>{ i32, double, %struct.RT }</tt>' type, a
5405 structure. The second index indexes into the third element of the structure,
5406 yielding a '<tt>%struct.RT</tt>' = '<tt>{ i8 , [10 x [20 x i32]], i8 }</tt>'
5407 type, another structure. The third index indexes into the second element of
5408 the structure, yielding a '<tt>[10 x [20 x i32]]</tt>' type, an array. The
5409 two dimensions of the array are subscripted into, yielding an '<tt>i32</tt>'
5410 type. The '<tt>getelementptr</tt>' instruction returns a pointer to this
5411 element, thus computing a value of '<tt>i32*</tt>' type.</p>
5413 <p>Note that it is perfectly legal to index partially through a structure,
5414 returning a pointer to an inner element. Because of this, the LLVM code for
5415 the given testcase is equivalent to:</p>
5417 <pre class="doc_code">
5418 define i32* @foo(%struct.ST* %s) {
5419 %t1 = getelementptr %struct.ST* %s, i32 1 <i>; yields %struct.ST*:%t1</i>
5420 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 <i>; yields %struct.RT*:%t2</i>
5421 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 <i>; yields [10 x [20 x i32]]*:%t3</i>
5422 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 <i>; yields [20 x i32]*:%t4</i>
5423 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 <i>; yields i32*:%t5</i>
5428 <p>If the <tt>inbounds</tt> keyword is present, the result value of the
5429 <tt>getelementptr</tt> is a <a href="#poisonvalues">poison value</a> if the
5430 base pointer is not an <i>in bounds</i> address of an allocated object,
5431 or if any of the addresses that would be formed by successive addition of
5432 the offsets implied by the indices to the base address with infinitely
5433 precise signed arithmetic are not an <i>in bounds</i> address of that
5434 allocated object. The <i>in bounds</i> addresses for an allocated object
5435 are all the addresses that point into the object, plus the address one
5437 In cases where the base is a vector of pointers the <tt>inbounds</tt> keyword
5438 applies to each of the computations element-wise. </p>
5440 <p>If the <tt>inbounds</tt> keyword is not present, the offsets are added to
5441 the base address with silently-wrapping two's complement arithmetic. If the
5442 offsets have a different width from the pointer, they are sign-extended or
5443 truncated to the width of the pointer. The result value of the
5444 <tt>getelementptr</tt> may be outside the object pointed to by the base
5445 pointer. The result value may not necessarily be used to access memory
5446 though, even if it happens to point into allocated storage. See the
5447 <a href="#pointeraliasing">Pointer Aliasing Rules</a> section for more
5450 <p>The getelementptr instruction is often confusing. For some more insight into
5451 how it works, see <a href="GetElementPtr.html">the getelementptr FAQ</a>.</p>
5455 <i>; yields [12 x i8]*:aptr</i>
5456 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5457 <i>; yields i8*:vptr</i>
5458 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5459 <i>; yields i8*:eptr</i>
5460 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5461 <i>; yields i32*:iptr</i>
5462 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5465 <p>In cases where the pointer argument is a vector of pointers, only a
5466 single index may be used, and the number of vector elements has to be
5467 the same. For example: </p>
5468 <pre class="doc_code">
5469 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5476 <!-- ======================================================================= -->
5478 <a name="convertops">Conversion Operations</a>
5483 <p>The instructions in this category are the conversion instructions (casting)
5484 which all take a single operand and a type. They perform various bit
5485 conversions on the operand.</p>
5487 <!-- _______________________________________________________________________ -->
5489 <a name="i_trunc">'<tt>trunc .. to</tt>' Instruction</a>
5496 <result> = trunc <ty> <value> to <ty2> <i>; yields ty2</i>
5500 <p>The '<tt>trunc</tt>' instruction truncates its operand to the
5501 type <tt>ty2</tt>.</p>
5504 <p>The '<tt>trunc</tt>' instruction takes a value to trunc, and a type to trunc it to.
5505 Both types must be of <a href="#t_integer">integer</a> types, or vectors
5506 of the same number of integers.
5507 The bit size of the <tt>value</tt> must be larger than
5508 the bit size of the destination type, <tt>ty2</tt>.
5509 Equal sized types are not allowed.</p>
5512 <p>The '<tt>trunc</tt>' instruction truncates the high order bits
5513 in <tt>value</tt> and converts the remaining bits to <tt>ty2</tt>. Since the
5514 source size must be larger than the destination size, <tt>trunc</tt> cannot
5515 be a <i>no-op cast</i>. It will always truncate bits.</p>
5519 %X = trunc i32 257 to i8 <i>; yields i8:1</i>
5520 %Y = trunc i32 123 to i1 <i>; yields i1:true</i>
5521 %Z = trunc i32 122 to i1 <i>; yields i1:false</i>
5522 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> <i>; yields <i8 8, i8 7></i>
5527 <!-- _______________________________________________________________________ -->
5529 <a name="i_zext">'<tt>zext .. to</tt>' Instruction</a>
5536 <result> = zext <ty> <value> to <ty2> <i>; yields ty2</i>
5540 <p>The '<tt>zext</tt>' instruction zero extends its operand to type
5545 <p>The '<tt>zext</tt>' instruction takes a value to cast, and a type to cast it to.
5546 Both types must be of <a href="#t_integer">integer</a> types, or vectors
5547 of the same number of integers.
5548 The bit size of the <tt>value</tt> must be smaller than
5549 the bit size of the destination type,
5553 <p>The <tt>zext</tt> fills the high order bits of the <tt>value</tt> with zero
5554 bits until it reaches the size of the destination type, <tt>ty2</tt>.</p>
5556 <p>When zero extending from i1, the result will always be either 0 or 1.</p>
5560 %X = zext i32 257 to i64 <i>; yields i64:257</i>
5561 %Y = zext i1 true to i32 <i>; yields i32:1</i>
5562 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> <i>; yields <i32 8, i32 7></i>
5567 <!-- _______________________________________________________________________ -->
5569 <a name="i_sext">'<tt>sext .. to</tt>' Instruction</a>
5576 <result> = sext <ty> <value> to <ty2> <i>; yields ty2</i>
5580 <p>The '<tt>sext</tt>' sign extends <tt>value</tt> to the type <tt>ty2</tt>.</p>
5583 <p>The '<tt>sext</tt>' instruction takes a value to cast, and a type to cast it to.
5584 Both types must be of <a href="#t_integer">integer</a> types, or vectors
5585 of the same number of integers.
5586 The bit size of the <tt>value</tt> must be smaller than
5587 the bit size of the destination type,
5591 <p>The '<tt>sext</tt>' instruction performs a sign extension by copying the sign
5592 bit (highest order bit) of the <tt>value</tt> until it reaches the bit size
5593 of the type <tt>ty2</tt>.</p>
5595 <p>When sign extending from i1, the extension always results in -1 or 0.</p>
5599 %X = sext i8 -1 to i16 <i>; yields i16 :65535</i>
5600 %Y = sext i1 true to i32 <i>; yields i32:-1</i>
5601 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> <i>; yields <i32 8, i32 7></i>
5606 <!-- _______________________________________________________________________ -->
5608 <a name="i_fptrunc">'<tt>fptrunc .. to</tt>' Instruction</a>
5615 <result> = fptrunc <ty> <value> to <ty2> <i>; yields ty2</i>
5619 <p>The '<tt>fptrunc</tt>' instruction truncates <tt>value</tt> to type
5623 <p>The '<tt>fptrunc</tt>' instruction takes a <a href="#t_floating">floating
5624 point</a> value to cast and a <a href="#t_floating">floating point</a> type
5625 to cast it to. The size of <tt>value</tt> must be larger than the size of
5626 <tt>ty2</tt>. This implies that <tt>fptrunc</tt> cannot be used to make a
5627 <i>no-op cast</i>.</p>
5630 <p>The '<tt>fptrunc</tt>' instruction truncates a <tt>value</tt> from a larger
5631 <a href="#t_floating">floating point</a> type to a smaller
5632 <a href="#t_floating">floating point</a> type. If the value cannot fit
5633 within the destination type, <tt>ty2</tt>, then the results are
5638 %X = fptrunc double 123.0 to float <i>; yields float:123.0</i>
5639 %Y = fptrunc double 1.0E+300 to float <i>; yields undefined</i>
5644 <!-- _______________________________________________________________________ -->
5646 <a name="i_fpext">'<tt>fpext .. to</tt>' Instruction</a>
5653 <result> = fpext <ty> <value> to <ty2> <i>; yields ty2</i>
5657 <p>The '<tt>fpext</tt>' extends a floating point <tt>value</tt> to a larger
5658 floating point value.</p>
5661 <p>The '<tt>fpext</tt>' instruction takes a
5662 <a href="#t_floating">floating point</a> <tt>value</tt> to cast, and
5663 a <a href="#t_floating">floating point</a> type to cast it to. The source
5664 type must be smaller than the destination type.</p>
5667 <p>The '<tt>fpext</tt>' instruction extends the <tt>value</tt> from a smaller
5668 <a href="#t_floating">floating point</a> type to a larger
5669 <a href="#t_floating">floating point</a> type. The <tt>fpext</tt> cannot be
5670 used to make a <i>no-op cast</i> because it always changes bits. Use
5671 <tt>bitcast</tt> to make a <i>no-op cast</i> for a floating point cast.</p>
5675 %X = fpext float 3.125 to double <i>; yields double:3.125000e+00</i>
5676 %Y = fpext double %X to fp128 <i>; yields fp128:0xL00000000000000004000900000000000</i>
5681 <!-- _______________________________________________________________________ -->
5683 <a name="i_fptoui">'<tt>fptoui .. to</tt>' Instruction</a>
5690 <result> = fptoui <ty> <value> to <ty2> <i>; yields ty2</i>
5694 <p>The '<tt>fptoui</tt>' converts a floating point <tt>value</tt> to its
5695 unsigned integer equivalent of type <tt>ty2</tt>.</p>
5698 <p>The '<tt>fptoui</tt>' instruction takes a value to cast, which must be a
5699 scalar or vector <a href="#t_floating">floating point</a> value, and a type
5700 to cast it to <tt>ty2</tt>, which must be an <a href="#t_integer">integer</a>
5701 type. If <tt>ty</tt> is a vector floating point type, <tt>ty2</tt> must be a
5702 vector integer type with the same number of elements as <tt>ty</tt></p>
5705 <p>The '<tt>fptoui</tt>' instruction converts its
5706 <a href="#t_floating">floating point</a> operand into the nearest (rounding
5707 towards zero) unsigned integer value. If the value cannot fit
5708 in <tt>ty2</tt>, the results are undefined.</p>
5712 %X = fptoui double 123.0 to i32 <i>; yields i32:123</i>
5713 %Y = fptoui float 1.0E+300 to i1 <i>; yields undefined:1</i>
5714 %Z = fptoui float 1.04E+17 to i8 <i>; yields undefined:1</i>
5719 <!-- _______________________________________________________________________ -->
5721 <a name="i_fptosi">'<tt>fptosi .. to</tt>' Instruction</a>
5728 <result> = fptosi <ty> <value> to <ty2> <i>; yields ty2</i>
5732 <p>The '<tt>fptosi</tt>' instruction converts
5733 <a href="#t_floating">floating point</a> <tt>value</tt> to
5734 type <tt>ty2</tt>.</p>
5737 <p>The '<tt>fptosi</tt>' instruction takes a value to cast, which must be a
5738 scalar or vector <a href="#t_floating">floating point</a> value, and a type
5739 to cast it to <tt>ty2</tt>, which must be an <a href="#t_integer">integer</a>
5740 type. If <tt>ty</tt> is a vector floating point type, <tt>ty2</tt> must be a
5741 vector integer type with the same number of elements as <tt>ty</tt></p>
5744 <p>The '<tt>fptosi</tt>' instruction converts its
5745 <a href="#t_floating">floating point</a> operand into the nearest (rounding
5746 towards zero) signed integer value. If the value cannot fit in <tt>ty2</tt>,
5747 the results are undefined.</p>
5751 %X = fptosi double -123.0 to i32 <i>; yields i32:-123</i>
5752 %Y = fptosi float 1.0E-247 to i1 <i>; yields undefined:1</i>
5753 %Z = fptosi float 1.04E+17 to i8 <i>; yields undefined:1</i>
5758 <!-- _______________________________________________________________________ -->
5760 <a name="i_uitofp">'<tt>uitofp .. to</tt>' Instruction</a>
5767 <result> = uitofp <ty> <value> to <ty2> <i>; yields ty2</i>
5771 <p>The '<tt>uitofp</tt>' instruction regards <tt>value</tt> as an unsigned
5772 integer and converts that value to the <tt>ty2</tt> type.</p>
5775 <p>The '<tt>uitofp</tt>' instruction takes a value to cast, which must be a
5776 scalar or vector <a href="#t_integer">integer</a> value, and a type to cast
5777 it to <tt>ty2</tt>, which must be an <a href="#t_floating">floating point</a>
5778 type. If <tt>ty</tt> is a vector integer type, <tt>ty2</tt> must be a vector
5779 floating point type with the same number of elements as <tt>ty</tt></p>
5782 <p>The '<tt>uitofp</tt>' instruction interprets its operand as an unsigned
5783 integer quantity and converts it to the corresponding floating point
5784 value. If the value cannot fit in the floating point value, the results are
5789 %X = uitofp i32 257 to float <i>; yields float:257.0</i>
5790 %Y = uitofp i8 -1 to double <i>; yields double:255.0</i>
5795 <!-- _______________________________________________________________________ -->
5797 <a name="i_sitofp">'<tt>sitofp .. to</tt>' Instruction</a>
5804 <result> = sitofp <ty> <value> to <ty2> <i>; yields ty2</i>
5808 <p>The '<tt>sitofp</tt>' instruction regards <tt>value</tt> as a signed integer
5809 and converts that value to the <tt>ty2</tt> type.</p>
5812 <p>The '<tt>sitofp</tt>' instruction takes a value to cast, which must be a
5813 scalar or vector <a href="#t_integer">integer</a> value, and a type to cast
5814 it to <tt>ty2</tt>, which must be an <a href="#t_floating">floating point</a>
5815 type. If <tt>ty</tt> is a vector integer type, <tt>ty2</tt> must be a vector
5816 floating point type with the same number of elements as <tt>ty</tt></p>
5819 <p>The '<tt>sitofp</tt>' instruction interprets its operand as a signed integer
5820 quantity and converts it to the corresponding floating point value. If the
5821 value cannot fit in the floating point value, the results are undefined.</p>
5825 %X = sitofp i32 257 to float <i>; yields float:257.0</i>
5826 %Y = sitofp i8 -1 to double <i>; yields double:-1.0</i>
5831 <!-- _______________________________________________________________________ -->
5833 <a name="i_ptrtoint">'<tt>ptrtoint .. to</tt>' Instruction</a>
5840 <result> = ptrtoint <ty> <value> to <ty2> <i>; yields ty2</i>
5844 <p>The '<tt>ptrtoint</tt>' instruction converts the pointer or a vector of
5845 pointers <tt>value</tt> to
5846 the integer (or vector of integers) type <tt>ty2</tt>.</p>
5849 <p>The '<tt>ptrtoint</tt>' instruction takes a <tt>value</tt> to cast, which
5850 must be a a value of type <a href="#t_pointer">pointer</a> or a vector of
5851 pointers, and a type to cast it to
5852 <tt>ty2</tt>, which must be an <a href="#t_integer">integer</a> or a vector
5853 of integers type.</p>
5856 <p>The '<tt>ptrtoint</tt>' instruction converts <tt>value</tt> to integer type
5857 <tt>ty2</tt> by interpreting the pointer value as an integer and either
5858 truncating or zero extending that value to the size of the integer type. If
5859 <tt>value</tt> is smaller than <tt>ty2</tt> then a zero extension is done. If
5860 <tt>value</tt> is larger than <tt>ty2</tt> then a truncation is done. If they
5861 are the same size, then nothing is done (<i>no-op cast</i>) other than a type
5866 %X = ptrtoint i32* %P to i8 <i>; yields truncation on 32-bit architecture</i>
5867 %Y = ptrtoint i32* %P to i64 <i>; yields zero extension on 32-bit architecture</i>
5868 %Z = ptrtoint <4 x i32*> %P to <4 x i64><i>; yields vector zero extension for a vector of addresses on 32-bit architecture</i>
5873 <!-- _______________________________________________________________________ -->
5875 <a name="i_inttoptr">'<tt>inttoptr .. to</tt>' Instruction</a>
5882 <result> = inttoptr <ty> <value> to <ty2> <i>; yields ty2</i>
5886 <p>The '<tt>inttoptr</tt>' instruction converts an integer <tt>value</tt> to a
5887 pointer type, <tt>ty2</tt>.</p>
5890 <p>The '<tt>inttoptr</tt>' instruction takes an <a href="#t_integer">integer</a>
5891 value to cast, and a type to cast it to, which must be a
5892 <a href="#t_pointer">pointer</a> type.</p>
5895 <p>The '<tt>inttoptr</tt>' instruction converts <tt>value</tt> to type
5896 <tt>ty2</tt> by applying either a zero extension or a truncation depending on
5897 the size of the integer <tt>value</tt>. If <tt>value</tt> is larger than the
5898 size of a pointer then a truncation is done. If <tt>value</tt> is smaller
5899 than the size of a pointer then a zero extension is done. If they are the
5900 same size, nothing is done (<i>no-op cast</i>).</p>
5904 %X = inttoptr i32 255 to i32* <i>; yields zero extension on 64-bit architecture</i>
5905 %Y = inttoptr i32 255 to i32* <i>; yields no-op on 32-bit architecture</i>
5906 %Z = inttoptr i64 0 to i32* <i>; yields truncation on 32-bit architecture</i>
5907 %Z = inttoptr <4 x i32> %G to <4 x i8*><i>; yields truncation of vector G to four pointers</i>
5912 <!-- _______________________________________________________________________ -->
5914 <a name="i_bitcast">'<tt>bitcast .. to</tt>' Instruction</a>
5921 <result> = bitcast <ty> <value> to <ty2> <i>; yields ty2</i>
5925 <p>The '<tt>bitcast</tt>' instruction converts <tt>value</tt> to type
5926 <tt>ty2</tt> without changing any bits.</p>
5929 <p>The '<tt>bitcast</tt>' instruction takes a value to cast, which must be a
5930 non-aggregate first class value, and a type to cast it to, which must also be
5931 a non-aggregate <a href="#t_firstclass">first class</a> type. The bit sizes
5932 of <tt>value</tt> and the destination type, <tt>ty2</tt>, must be
5933 identical. If the source type is a pointer, the destination type must also be
5934 a pointer. This instruction supports bitwise conversion of vectors to
5935 integers and to vectors of other types (as long as they have the same
5939 <p>The '<tt>bitcast</tt>' instruction converts <tt>value</tt> to type
5940 <tt>ty2</tt>. It is always a <i>no-op cast</i> because no bits change with
5941 this conversion. The conversion is done as if the <tt>value</tt> had been
5942 stored to memory and read back as type <tt>ty2</tt>.
5943 Pointer (or vector of pointers) types may only be converted to other pointer
5944 (or vector of pointers) types with this instruction. To convert
5945 pointers to other types, use the <a href="#i_inttoptr">inttoptr</a> or
5946 <a href="#i_ptrtoint">ptrtoint</a> instructions first.</p>
5950 %X = bitcast i8 255 to i8 <i>; yields i8 :-1</i>
5951 %Y = bitcast i32* %x to sint* <i>; yields sint*:%x</i>
5952 %Z = bitcast <2 x int> %V to i64; <i>; yields i64: %V</i>
5953 %Z = bitcast <2 x i32*> %V to <2 x i64*> <i>; yields <2 x i64*></i>
5960 <!-- ======================================================================= -->
5962 <a name="otherops">Other Operations</a>
5967 <p>The instructions in this category are the "miscellaneous" instructions, which
5968 defy better classification.</p>
5970 <!-- _______________________________________________________________________ -->
5972 <a name="i_icmp">'<tt>icmp</tt>' Instruction</a>
5979 <result> = icmp <cond> <ty> <op1>, <op2> <i>; yields {i1} or {<N x i1>}:result</i>
5983 <p>The '<tt>icmp</tt>' instruction returns a boolean value or a vector of
5984 boolean values based on comparison of its two integer, integer vector,
5985 pointer, or pointer vector operands.</p>
5988 <p>The '<tt>icmp</tt>' instruction takes three operands. The first operand is
5989 the condition code indicating the kind of comparison to perform. It is not a
5990 value, just a keyword. The possible condition code are:</p>
5993 <li><tt>eq</tt>: equal</li>
5994 <li><tt>ne</tt>: not equal </li>
5995 <li><tt>ugt</tt>: unsigned greater than</li>
5996 <li><tt>uge</tt>: unsigned greater or equal</li>
5997 <li><tt>ult</tt>: unsigned less than</li>
5998 <li><tt>ule</tt>: unsigned less or equal</li>
5999 <li><tt>sgt</tt>: signed greater than</li>
6000 <li><tt>sge</tt>: signed greater or equal</li>
6001 <li><tt>slt</tt>: signed less than</li>
6002 <li><tt>sle</tt>: signed less or equal</li>
6005 <p>The remaining two arguments must be <a href="#t_integer">integer</a> or
6006 <a href="#t_pointer">pointer</a> or integer <a href="#t_vector">vector</a>
6007 typed. They must also be identical types.</p>
6010 <p>The '<tt>icmp</tt>' compares <tt>op1</tt> and <tt>op2</tt> according to the
6011 condition code given as <tt>cond</tt>. The comparison performed always yields
6012 either an <a href="#t_integer"><tt>i1</tt></a> or vector of <tt>i1</tt>
6013 result, as follows:</p>
6016 <li><tt>eq</tt>: yields <tt>true</tt> if the operands are equal,
6017 <tt>false</tt> otherwise. No sign interpretation is necessary or
6020 <li><tt>ne</tt>: yields <tt>true</tt> if the operands are unequal,
6021 <tt>false</tt> otherwise. No sign interpretation is necessary or
6024 <li><tt>ugt</tt>: interprets the operands as unsigned values and yields
6025 <tt>true</tt> if <tt>op1</tt> is greater than <tt>op2</tt>.</li>
6027 <li><tt>uge</tt>: interprets the operands as unsigned values and yields
6028 <tt>true</tt> if <tt>op1</tt> is greater than or equal
6029 to <tt>op2</tt>.</li>
6031 <li><tt>ult</tt>: interprets the operands as unsigned values and yields
6032 <tt>true</tt> if <tt>op1</tt> is less than <tt>op2</tt>.</li>
6034 <li><tt>ule</tt>: interprets the operands as unsigned values and yields
6035 <tt>true</tt> if <tt>op1</tt> is less than or equal to <tt>op2</tt>.</li>
6037 <li><tt>sgt</tt>: interprets the operands as signed values and yields
6038 <tt>true</tt> if <tt>op1</tt> is greater than <tt>op2</tt>.</li>
6040 <li><tt>sge</tt>: interprets the operands as signed values and yields
6041 <tt>true</tt> if <tt>op1</tt> is greater than or equal
6042 to <tt>op2</tt>.</li>
6044 <li><tt>slt</tt>: interprets the operands as signed values and yields
6045 <tt>true</tt> if <tt>op1</tt> is less than <tt>op2</tt>.</li>
6047 <li><tt>sle</tt>: interprets the operands as signed values and yields
6048 <tt>true</tt> if <tt>op1</tt> is less than or equal to <tt>op2</tt>.</li>
6051 <p>If the operands are <a href="#t_pointer">pointer</a> typed, the pointer
6052 values are compared as if they were integers.</p>
6054 <p>If the operands are integer vectors, then they are compared element by
6055 element. The result is an <tt>i1</tt> vector with the same number of elements
6056 as the values being compared. Otherwise, the result is an <tt>i1</tt>.</p>
6060 <result> = icmp eq i32 4, 5 <i>; yields: result=false</i>
6061 <result> = icmp ne float* %X, %X <i>; yields: result=false</i>
6062 <result> = icmp ult i16 4, 5 <i>; yields: result=true</i>
6063 <result> = icmp sgt i16 4, 5 <i>; yields: result=false</i>
6064 <result> = icmp ule i16 -4, 5 <i>; yields: result=false</i>
6065 <result> = icmp sge i16 4, 5 <i>; yields: result=false</i>
6068 <p>Note that the code generator does not yet support vector types with
6069 the <tt>icmp</tt> instruction.</p>
6073 <!-- _______________________________________________________________________ -->
6075 <a name="i_fcmp">'<tt>fcmp</tt>' Instruction</a>
6082 <result> = fcmp <cond> <ty> <op1>, <op2> <i>; yields {i1} or {<N x i1>}:result</i>
6086 <p>The '<tt>fcmp</tt>' instruction returns a boolean value or vector of boolean
6087 values based on comparison of its operands.</p>
6089 <p>If the operands are floating point scalars, then the result type is a boolean
6090 (<a href="#t_integer"><tt>i1</tt></a>).</p>
6092 <p>If the operands are floating point vectors, then the result type is a vector
6093 of boolean with the same number of elements as the operands being
6097 <p>The '<tt>fcmp</tt>' instruction takes three operands. The first operand is
6098 the condition code indicating the kind of comparison to perform. It is not a
6099 value, just a keyword. The possible condition code are:</p>
6102 <li><tt>false</tt>: no comparison, always returns false</li>
6103 <li><tt>oeq</tt>: ordered and equal</li>
6104 <li><tt>ogt</tt>: ordered and greater than </li>
6105 <li><tt>oge</tt>: ordered and greater than or equal</li>
6106 <li><tt>olt</tt>: ordered and less than </li>
6107 <li><tt>ole</tt>: ordered and less than or equal</li>
6108 <li><tt>one</tt>: ordered and not equal</li>
6109 <li><tt>ord</tt>: ordered (no nans)</li>
6110 <li><tt>ueq</tt>: unordered or equal</li>
6111 <li><tt>ugt</tt>: unordered or greater than </li>
6112 <li><tt>uge</tt>: unordered or greater than or equal</li>
6113 <li><tt>ult</tt>: unordered or less than </li>
6114 <li><tt>ule</tt>: unordered or less than or equal</li>
6115 <li><tt>une</tt>: unordered or not equal</li>
6116 <li><tt>uno</tt>: unordered (either nans)</li>
6117 <li><tt>true</tt>: no comparison, always returns true</li>
6120 <p><i>Ordered</i> means that neither operand is a QNAN while
6121 <i>unordered</i> means that either operand may be a QNAN.</p>
6123 <p>Each of <tt>val1</tt> and <tt>val2</tt> arguments must be either
6124 a <a href="#t_floating">floating point</a> type or
6125 a <a href="#t_vector">vector</a> of floating point type. They must have
6126 identical types.</p>
6129 <p>The '<tt>fcmp</tt>' instruction compares <tt>op1</tt> and <tt>op2</tt>
6130 according to the condition code given as <tt>cond</tt>. If the operands are
6131 vectors, then the vectors are compared element by element. Each comparison
6132 performed always yields an <a href="#t_integer">i1</a> result, as
6136 <li><tt>false</tt>: always yields <tt>false</tt>, regardless of operands.</li>
6138 <li><tt>oeq</tt>: yields <tt>true</tt> if both operands are not a QNAN and
6139 <tt>op1</tt> is equal to <tt>op2</tt>.</li>
6141 <li><tt>ogt</tt>: yields <tt>true</tt> if both operands are not a QNAN and
6142 <tt>op1</tt> is greater than <tt>op2</tt>.</li>
6144 <li><tt>oge</tt>: yields <tt>true</tt> if both operands are not a QNAN and
6145 <tt>op1</tt> is greater than or equal to <tt>op2</tt>.</li>
6147 <li><tt>olt</tt>: yields <tt>true</tt> if both operands are not a QNAN and
6148 <tt>op1</tt> is less than <tt>op2</tt>.</li>
6150 <li><tt>ole</tt>: yields <tt>true</tt> if both operands are not a QNAN and
6151 <tt>op1</tt> is less than or equal to <tt>op2</tt>.</li>
6153 <li><tt>one</tt>: yields <tt>true</tt> if both operands are not a QNAN and
6154 <tt>op1</tt> is not equal to <tt>op2</tt>.</li>
6156 <li><tt>ord</tt>: yields <tt>true</tt> if both operands are not a QNAN.</li>
6158 <li><tt>ueq</tt>: yields <tt>true</tt> if either operand is a QNAN or
6159 <tt>op1</tt> is equal to <tt>op2</tt>.</li>
6161 <li><tt>ugt</tt>: yields <tt>true</tt> if either operand is a QNAN or
6162 <tt>op1</tt> is greater than <tt>op2</tt>.</li>
6164 <li><tt>uge</tt>: yields <tt>true</tt> if either operand is a QNAN or
6165 <tt>op1</tt> is greater than or equal to <tt>op2</tt>.</li>
6167 <li><tt>ult</tt>: yields <tt>true</tt> if either operand is a QNAN or
6168 <tt>op1</tt> is less than <tt>op2</tt>.</li>
6170 <li><tt>ule</tt>: yields <tt>true</tt> if either operand is a QNAN or
6171 <tt>op1</tt> is less than or equal to <tt>op2</tt>.</li>
6173 <li><tt>une</tt>: yields <tt>true</tt> if either operand is a QNAN or
6174 <tt>op1</tt> is not equal to <tt>op2</tt>.</li>
6176 <li><tt>uno</tt>: yields <tt>true</tt> if either operand is a QNAN.</li>
6178 <li><tt>true</tt>: always yields <tt>true</tt>, regardless of operands.</li>
6183 <result> = fcmp oeq float 4.0, 5.0 <i>; yields: result=false</i>
6184 <result> = fcmp one float 4.0, 5.0 <i>; yields: result=true</i>
6185 <result> = fcmp olt float 4.0, 5.0 <i>; yields: result=true</i>
6186 <result> = fcmp ueq double 1.0, 2.0 <i>; yields: result=false</i>
6189 <p>Note that the code generator does not yet support vector types with
6190 the <tt>fcmp</tt> instruction.</p>
6194 <!-- _______________________________________________________________________ -->
6196 <a name="i_phi">'<tt>phi</tt>' Instruction</a>
6203 <result> = phi <ty> [ <val0>, <label0>], ...
6207 <p>The '<tt>phi</tt>' instruction is used to implement the φ node in the
6208 SSA graph representing the function.</p>
6211 <p>The type of the incoming values is specified with the first type field. After
6212 this, the '<tt>phi</tt>' instruction takes a list of pairs as arguments, with
6213 one pair for each predecessor basic block of the current block. Only values
6214 of <a href="#t_firstclass">first class</a> type may be used as the value
6215 arguments to the PHI node. Only labels may be used as the label
6218 <p>There must be no non-phi instructions between the start of a basic block and
6219 the PHI instructions: i.e. PHI instructions must be first in a basic
6222 <p>For the purposes of the SSA form, the use of each incoming value is deemed to
6223 occur on the edge from the corresponding predecessor block to the current
6224 block (but after any definition of an '<tt>invoke</tt>' instruction's return
6225 value on the same edge).</p>
6228 <p>At runtime, the '<tt>phi</tt>' instruction logically takes on the value
6229 specified by the pair corresponding to the predecessor basic block that
6230 executed just prior to the current block.</p>
6234 Loop: ; Infinite loop that counts from 0 on up...
6235 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6236 %nextindvar = add i32 %indvar, 1
6242 <!-- _______________________________________________________________________ -->
6244 <a name="i_select">'<tt>select</tt>' Instruction</a>
6251 <result> = select <i>selty</i> <cond>, <ty> <val1>, <ty> <val2> <i>; yields ty</i>
6253 <i>selty</i> is either i1 or {<N x i1>}
6257 <p>The '<tt>select</tt>' instruction is used to choose one value based on a
6258 condition, without branching.</p>
6262 <p>The '<tt>select</tt>' instruction requires an 'i1' value or a vector of 'i1'
6263 values indicating the condition, and two values of the
6264 same <a href="#t_firstclass">first class</a> type. If the val1/val2 are
6265 vectors and the condition is a scalar, then entire vectors are selected, not
6266 individual elements.</p>
6269 <p>If the condition is an i1 and it evaluates to 1, the instruction returns the
6270 first value argument; otherwise, it returns the second value argument.</p>
6272 <p>If the condition is a vector of i1, then the value arguments must be vectors
6273 of the same size, and the selection is done element by element.</p>
6277 %X = select i1 true, i8 17, i8 42 <i>; yields i8:17</i>
6282 <!-- _______________________________________________________________________ -->
6284 <a name="i_call">'<tt>call</tt>' Instruction</a>
6291 <result> = [tail] call [<a href="#callingconv">cconv</a>] [<a href="#paramattrs">ret attrs</a>] <ty> [<fnty>*] <fnptrval>(<function args>) [<a href="#fnattrs">fn attrs</a>]
6295 <p>The '<tt>call</tt>' instruction represents a simple function call.</p>
6298 <p>This instruction requires several arguments:</p>
6301 <li>The optional "tail" marker indicates that the callee function does not
6302 access any allocas or varargs in the caller. Note that calls may be
6303 marked "tail" even if they do not occur before
6304 a <a href="#i_ret"><tt>ret</tt></a> instruction. If the "tail" marker is
6305 present, the function call is eligible for tail call optimization,
6306 but <a href="CodeGenerator.html#tailcallopt">might not in fact be
6307 optimized into a jump</a>. The code generator may optimize calls marked
6308 "tail" with either 1) automatic <a href="CodeGenerator.html#sibcallopt">
6309 sibling call optimization</a> when the caller and callee have
6310 matching signatures, or 2) forced tail call optimization when the
6311 following extra requirements are met:
6313 <li>Caller and callee both have the calling
6314 convention <tt>fastcc</tt>.</li>
6315 <li>The call is in tail position (ret immediately follows call and ret
6316 uses value of call or is void).</li>
6317 <li>Option <tt>-tailcallopt</tt> is enabled,
6318 or <code>llvm::GuaranteedTailCallOpt</code> is <code>true</code>.</li>
6319 <li><a href="CodeGenerator.html#tailcallopt">Platform specific
6320 constraints are met.</a></li>
6324 <li>The optional "cconv" marker indicates which <a href="#callingconv">calling
6325 convention</a> the call should use. If none is specified, the call
6326 defaults to using C calling conventions. The calling convention of the
6327 call must match the calling convention of the target function, or else the
6328 behavior is undefined.</li>
6330 <li>The optional <a href="#paramattrs">Parameter Attributes</a> list for
6331 return values. Only '<tt>zeroext</tt>', '<tt>signext</tt>', and
6332 '<tt>inreg</tt>' attributes are valid here.</li>
6334 <li>'<tt>ty</tt>': the type of the call instruction itself which is also the
6335 type of the return value. Functions that return no value are marked
6336 <tt><a href="#t_void">void</a></tt>.</li>
6338 <li>'<tt>fnty</tt>': shall be the signature of the pointer to function value
6339 being invoked. The argument types must match the types implied by this
6340 signature. This type can be omitted if the function is not varargs and if
6341 the function type does not return a pointer to a function.</li>
6343 <li>'<tt>fnptrval</tt>': An LLVM value containing a pointer to a function to
6344 be invoked. In most cases, this is a direct function invocation, but
6345 indirect <tt>call</tt>s are just as possible, calling an arbitrary pointer
6346 to function value.</li>
6348 <li>'<tt>function args</tt>': argument list whose types match the function
6349 signature argument types and parameter attributes. All arguments must be
6350 of <a href="#t_firstclass">first class</a> type. If the function
6351 signature indicates the function accepts a variable number of arguments,
6352 the extra arguments can be specified.</li>
6354 <li>The optional <a href="#fnattrs">function attributes</a> list. Only
6355 '<tt>noreturn</tt>', '<tt>nounwind</tt>', '<tt>readonly</tt>' and
6356 '<tt>readnone</tt>' attributes are valid here.</li>
6360 <p>The '<tt>call</tt>' instruction is used to cause control flow to transfer to
6361 a specified function, with its incoming arguments bound to the specified
6362 values. Upon a '<tt><a href="#i_ret">ret</a></tt>' instruction in the called
6363 function, control flow continues with the instruction after the function
6364 call, and the return value of the function is bound to the result
6369 %retval = call i32 @test(i32 %argc)
6370 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) <i>; yields i32</i>
6371 %X = tail call i32 @foo() <i>; yields i32</i>
6372 %Y = tail call <a href="#callingconv">fastcc</a> i32 @foo() <i>; yields i32</i>
6373 call void %foo(i8 97 signext)
6375 %struct.A = type { i32, i8 }
6376 %r = call %struct.A @foo() <i>; yields { 32, i8 }</i>
6377 %gr = extractvalue %struct.A %r, 0 <i>; yields i32</i>
6378 %gr1 = extractvalue %struct.A %r, 1 <i>; yields i8</i>
6379 %Z = call void @foo() noreturn <i>; indicates that %foo never returns normally</i>
6380 %ZZ = call zeroext i32 @bar() <i>; Return value is %zero extended</i>
6383 <p>llvm treats calls to some functions with names and arguments that match the
6384 standard C99 library as being the C99 library functions, and may perform
6385 optimizations or generate code for them under that assumption. This is
6386 something we'd like to change in the future to provide better support for
6387 freestanding environments and non-C-based languages.</p>
6391 <!-- _______________________________________________________________________ -->
6393 <a name="i_va_arg">'<tt>va_arg</tt>' Instruction</a>
6400 <resultval> = va_arg <va_list*> <arglist>, <argty>
6404 <p>The '<tt>va_arg</tt>' instruction is used to access arguments passed through
6405 the "variable argument" area of a function call. It is used to implement the
6406 <tt>va_arg</tt> macro in C.</p>
6409 <p>This instruction takes a <tt>va_list*</tt> value and the type of the
6410 argument. It returns a value of the specified argument type and increments
6411 the <tt>va_list</tt> to point to the next argument. The actual type
6412 of <tt>va_list</tt> is target specific.</p>
6415 <p>The '<tt>va_arg</tt>' instruction loads an argument of the specified type
6416 from the specified <tt>va_list</tt> and causes the <tt>va_list</tt> to point
6417 to the next argument. For more information, see the variable argument
6418 handling <a href="#int_varargs">Intrinsic Functions</a>.</p>
6420 <p>It is legal for this instruction to be called in a function which does not
6421 take a variable number of arguments, for example, the <tt>vfprintf</tt>
6424 <p><tt>va_arg</tt> is an LLVM instruction instead of
6425 an <a href="#intrinsics">intrinsic function</a> because it takes a type as an
6429 <p>See the <a href="#int_varargs">variable argument processing</a> section.</p>
6431 <p>Note that the code generator does not yet fully support va_arg on many
6432 targets. Also, it does not currently support va_arg with aggregate types on
6437 <!-- _______________________________________________________________________ -->
6439 <a name="i_landingpad">'<tt>landingpad</tt>' Instruction</a>
6446 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6447 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6449 <clause> := catch <type> <value>
6450 <clause> := filter <array constant type> <array constant>
6454 <p>The '<tt>landingpad</tt>' instruction is used by
6455 <a href="ExceptionHandling.html#overview">LLVM's exception handling
6456 system</a> to specify that a basic block is a landing pad — one where
6457 the exception lands, and corresponds to the code found in the
6458 <i><tt>catch</tt></i> portion of a <i><tt>try/catch</tt></i> sequence. It
6459 defines values supplied by the personality function (<tt>pers_fn</tt>) upon
6460 re-entry to the function. The <tt>resultval</tt> has the
6461 type <tt>resultty</tt>.</p>
6464 <p>This instruction takes a <tt>pers_fn</tt> value. This is the personality
6465 function associated with the unwinding mechanism. The optional
6466 <tt>cleanup</tt> flag indicates that the landing pad block is a cleanup.</p>
6468 <p>A <tt>clause</tt> begins with the clause type — <tt>catch</tt>
6469 or <tt>filter</tt> — and contains the global variable representing the
6470 "type" that may be caught or filtered respectively. Unlike the
6471 <tt>catch</tt> clause, the <tt>filter</tt> clause takes an array constant as
6472 its argument. Use "<tt>[0 x i8**] undef</tt>" for a filter which cannot
6473 throw. The '<tt>landingpad</tt>' instruction must contain <em>at least</em>
6474 one <tt>clause</tt> or the <tt>cleanup</tt> flag.</p>
6477 <p>The '<tt>landingpad</tt>' instruction defines the values which are set by the
6478 personality function (<tt>pers_fn</tt>) upon re-entry to the function, and
6479 therefore the "result type" of the <tt>landingpad</tt> instruction. As with
6480 calling conventions, how the personality function results are represented in
6481 LLVM IR is target specific.</p>
6483 <p>The clauses are applied in order from top to bottom. If two
6484 <tt>landingpad</tt> instructions are merged together through inlining, the
6485 clauses from the calling function are appended to the list of clauses.
6486 When the call stack is being unwound due to an exception being thrown, the
6487 exception is compared against each <tt>clause</tt> in turn. If it doesn't
6488 match any of the clauses, and the <tt>cleanup</tt> flag is not set, then
6489 unwinding continues further up the call stack.</p>
6491 <p>The <tt>landingpad</tt> instruction has several restrictions:</p>
6494 <li>A landing pad block is a basic block which is the unwind destination of an
6495 '<tt>invoke</tt>' instruction.</li>
6496 <li>A landing pad block must have a '<tt>landingpad</tt>' instruction as its
6497 first non-PHI instruction.</li>
6498 <li>There can be only one '<tt>landingpad</tt>' instruction within the landing
6500 <li>A basic block that is not a landing pad block may not include a
6501 '<tt>landingpad</tt>' instruction.</li>
6502 <li>All '<tt>landingpad</tt>' instructions in a function must have the same
6503 personality function.</li>
6508 ;; A landing pad which can catch an integer.
6509 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6511 ;; A landing pad that is a cleanup.
6512 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6514 ;; A landing pad which can catch an integer and can only throw a double.
6515 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6517 filter [1 x i8**] [@_ZTId]
6526 <!-- *********************************************************************** -->
6527 <h2><a name="intrinsics">Intrinsic Functions</a></h2>
6528 <!-- *********************************************************************** -->
6532 <p>LLVM supports the notion of an "intrinsic function". These functions have
6533 well known names and semantics and are required to follow certain
6534 restrictions. Overall, these intrinsics represent an extension mechanism for
6535 the LLVM language that does not require changing all of the transformations
6536 in LLVM when adding to the language (or the bitcode reader/writer, the
6537 parser, etc...).</p>
6539 <p>Intrinsic function names must all start with an "<tt>llvm.</tt>" prefix. This
6540 prefix is reserved in LLVM for intrinsic names; thus, function names may not
6541 begin with this prefix. Intrinsic functions must always be external
6542 functions: you cannot define the body of intrinsic functions. Intrinsic
6543 functions may only be used in call or invoke instructions: it is illegal to
6544 take the address of an intrinsic function. Additionally, because intrinsic
6545 functions are part of the LLVM language, it is required if any are added that
6546 they be documented here.</p>
6548 <p>Some intrinsic functions can be overloaded, i.e., the intrinsic represents a
6549 family of functions that perform the same operation but on different data
6550 types. Because LLVM can represent over 8 million different integer types,
6551 overloading is used commonly to allow an intrinsic function to operate on any
6552 integer type. One or more of the argument types or the result type can be
6553 overloaded to accept any integer type. Argument types may also be defined as
6554 exactly matching a previous argument's type or the result type. This allows
6555 an intrinsic function which accepts multiple arguments, but needs all of them
6556 to be of the same type, to only be overloaded with respect to a single
6557 argument or the result.</p>
6559 <p>Overloaded intrinsics will have the names of its overloaded argument types
6560 encoded into its function name, each preceded by a period. Only those types
6561 which are overloaded result in a name suffix. Arguments whose type is matched
6562 against another type do not. For example, the <tt>llvm.ctpop</tt> function
6563 can take an integer of any width and returns an integer of exactly the same
6564 integer width. This leads to a family of functions such as
6565 <tt>i8 @llvm.ctpop.i8(i8 %val)</tt> and <tt>i29 @llvm.ctpop.i29(i29
6566 %val)</tt>. Only one type, the return type, is overloaded, and only one type
6567 suffix is required. Because the argument's type is matched against the return
6568 type, it does not require its own name suffix.</p>
6570 <p>To learn how to add an intrinsic function, please see the
6571 <a href="ExtendingLLVM.html">Extending LLVM Guide</a>.</p>
6573 <!-- ======================================================================= -->
6575 <a name="int_varargs">Variable Argument Handling Intrinsics</a>
6580 <p>Variable argument support is defined in LLVM with
6581 the <a href="#i_va_arg"><tt>va_arg</tt></a> instruction and these three
6582 intrinsic functions. These functions are related to the similarly named
6583 macros defined in the <tt><stdarg.h></tt> header file.</p>
6585 <p>All of these functions operate on arguments that use a target-specific value
6586 type "<tt>va_list</tt>". The LLVM assembly language reference manual does
6587 not define what this type is, so all transformations should be prepared to
6588 handle these functions regardless of the type used.</p>
6590 <p>This example shows how the <a href="#i_va_arg"><tt>va_arg</tt></a>
6591 instruction and the variable argument handling intrinsic functions are
6594 <pre class="doc_code">
6595 define i32 @test(i32 %X, ...) {
6596 ; Initialize variable argument processing
6598 %ap2 = bitcast i8** %ap to i8*
6599 call void @llvm.va_start(i8* %ap2)
6601 ; Read a single integer argument
6602 %tmp = va_arg i8** %ap, i32
6604 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6606 %aq2 = bitcast i8** %aq to i8*
6607 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6608 call void @llvm.va_end(i8* %aq2)
6610 ; Stop processing of arguments.
6611 call void @llvm.va_end(i8* %ap2)
6615 declare void @llvm.va_start(i8*)
6616 declare void @llvm.va_copy(i8*, i8*)
6617 declare void @llvm.va_end(i8*)
6620 <!-- _______________________________________________________________________ -->
6622 <a name="int_va_start">'<tt>llvm.va_start</tt>' Intrinsic</a>
6630 declare void %llvm.va_start(i8* <arglist>)
6634 <p>The '<tt>llvm.va_start</tt>' intrinsic initializes <tt>*<arglist></tt>
6635 for subsequent use by <tt><a href="#i_va_arg">va_arg</a></tt>.</p>
6638 <p>The argument is a pointer to a <tt>va_list</tt> element to initialize.</p>
6641 <p>The '<tt>llvm.va_start</tt>' intrinsic works just like the <tt>va_start</tt>
6642 macro available in C. In a target-dependent way, it initializes
6643 the <tt>va_list</tt> element to which the argument points, so that the next
6644 call to <tt>va_arg</tt> will produce the first variable argument passed to
6645 the function. Unlike the C <tt>va_start</tt> macro, this intrinsic does not
6646 need to know the last argument of the function as the compiler can figure
6651 <!-- _______________________________________________________________________ -->
6653 <a name="int_va_end">'<tt>llvm.va_end</tt>' Intrinsic</a>
6660 declare void @llvm.va_end(i8* <arglist>)
6664 <p>The '<tt>llvm.va_end</tt>' intrinsic destroys <tt>*<arglist></tt>,
6665 which has been initialized previously
6666 with <tt><a href="#int_va_start">llvm.va_start</a></tt>
6667 or <tt><a href="#i_va_copy">llvm.va_copy</a></tt>.</p>
6670 <p>The argument is a pointer to a <tt>va_list</tt> to destroy.</p>
6673 <p>The '<tt>llvm.va_end</tt>' intrinsic works just like the <tt>va_end</tt>
6674 macro available in C. In a target-dependent way, it destroys
6675 the <tt>va_list</tt> element to which the argument points. Calls
6676 to <a href="#int_va_start"><tt>llvm.va_start</tt></a>
6677 and <a href="#int_va_copy"> <tt>llvm.va_copy</tt></a> must be matched exactly
6678 with calls to <tt>llvm.va_end</tt>.</p>
6682 <!-- _______________________________________________________________________ -->
6684 <a name="int_va_copy">'<tt>llvm.va_copy</tt>' Intrinsic</a>
6691 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6695 <p>The '<tt>llvm.va_copy</tt>' intrinsic copies the current argument position
6696 from the source argument list to the destination argument list.</p>
6699 <p>The first argument is a pointer to a <tt>va_list</tt> element to initialize.
6700 The second argument is a pointer to a <tt>va_list</tt> element to copy
6704 <p>The '<tt>llvm.va_copy</tt>' intrinsic works just like the <tt>va_copy</tt>
6705 macro available in C. In a target-dependent way, it copies the
6706 source <tt>va_list</tt> element into the destination <tt>va_list</tt>
6707 element. This intrinsic is necessary because
6708 the <tt><a href="#int_va_start"> llvm.va_start</a></tt> intrinsic may be
6709 arbitrarily complex and require, for example, memory allocation.</p>
6715 <!-- ======================================================================= -->
6717 <a name="int_gc">Accurate Garbage Collection Intrinsics</a>
6722 <p>LLVM support for <a href="GarbageCollection.html">Accurate Garbage
6723 Collection</a> (GC) requires the implementation and generation of these
6724 intrinsics. These intrinsics allow identification of <a href="#int_gcroot">GC
6725 roots on the stack</a>, as well as garbage collector implementations that
6726 require <a href="#int_gcread">read</a> and <a href="#int_gcwrite">write</a>
6727 barriers. Front-ends for type-safe garbage collected languages should generate
6728 these intrinsics to make use of the LLVM garbage collectors. For more details,
6729 see <a href="GarbageCollection.html">Accurate Garbage Collection with
6732 <p>The garbage collection intrinsics only operate on objects in the generic
6733 address space (address space zero).</p>
6735 <!-- _______________________________________________________________________ -->
6737 <a name="int_gcroot">'<tt>llvm.gcroot</tt>' Intrinsic</a>
6744 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6748 <p>The '<tt>llvm.gcroot</tt>' intrinsic declares the existence of a GC root to
6749 the code generator, and allows some metadata to be associated with it.</p>
6752 <p>The first argument specifies the address of a stack object that contains the
6753 root pointer. The second pointer (which must be either a constant or a
6754 global value address) contains the meta-data to be associated with the
6758 <p>At runtime, a call to this intrinsic stores a null pointer into the "ptrloc"
6759 location. At compile-time, the code generator generates information to allow
6760 the runtime to find the pointer at GC safe points. The '<tt>llvm.gcroot</tt>'
6761 intrinsic may only be used in a function which <a href="#gc">specifies a GC
6766 <!-- _______________________________________________________________________ -->
6768 <a name="int_gcread">'<tt>llvm.gcread</tt>' Intrinsic</a>
6775 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6779 <p>The '<tt>llvm.gcread</tt>' intrinsic identifies reads of references from heap
6780 locations, allowing garbage collector implementations that require read
6784 <p>The second argument is the address to read from, which should be an address
6785 allocated from the garbage collector. The first object is a pointer to the
6786 start of the referenced object, if needed by the language runtime (otherwise
6790 <p>The '<tt>llvm.gcread</tt>' intrinsic has the same semantics as a load
6791 instruction, but may be replaced with substantially more complex code by the
6792 garbage collector runtime, as needed. The '<tt>llvm.gcread</tt>' intrinsic
6793 may only be used in a function which <a href="#gc">specifies a GC
6798 <!-- _______________________________________________________________________ -->
6800 <a name="int_gcwrite">'<tt>llvm.gcwrite</tt>' Intrinsic</a>
6807 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6811 <p>The '<tt>llvm.gcwrite</tt>' intrinsic identifies writes of references to heap
6812 locations, allowing garbage collector implementations that require write
6813 barriers (such as generational or reference counting collectors).</p>
6816 <p>The first argument is the reference to store, the second is the start of the
6817 object to store it to, and the third is the address of the field of Obj to
6818 store to. If the runtime does not require a pointer to the object, Obj may
6822 <p>The '<tt>llvm.gcwrite</tt>' intrinsic has the same semantics as a store
6823 instruction, but may be replaced with substantially more complex code by the
6824 garbage collector runtime, as needed. The '<tt>llvm.gcwrite</tt>' intrinsic
6825 may only be used in a function which <a href="#gc">specifies a GC
6832 <!-- ======================================================================= -->
6834 <a name="int_codegen">Code Generator Intrinsics</a>
6839 <p>These intrinsics are provided by LLVM to expose special features that may
6840 only be implemented with code generator support.</p>
6842 <!-- _______________________________________________________________________ -->
6844 <a name="int_returnaddress">'<tt>llvm.returnaddress</tt>' Intrinsic</a>
6851 declare i8 *@llvm.returnaddress(i32 <level>)
6855 <p>The '<tt>llvm.returnaddress</tt>' intrinsic attempts to compute a
6856 target-specific value indicating the return address of the current function
6857 or one of its callers.</p>
6860 <p>The argument to this intrinsic indicates which function to return the address
6861 for. Zero indicates the calling function, one indicates its caller, etc.
6862 The argument is <b>required</b> to be a constant integer value.</p>
6865 <p>The '<tt>llvm.returnaddress</tt>' intrinsic either returns a pointer
6866 indicating the return address of the specified call frame, or zero if it
6867 cannot be identified. The value returned by this intrinsic is likely to be
6868 incorrect or 0 for arguments other than zero, so it should only be used for
6869 debugging purposes.</p>
6871 <p>Note that calling this intrinsic does not prevent function inlining or other
6872 aggressive transformations, so the value returned may not be that of the
6873 obvious source-language caller.</p>
6877 <!-- _______________________________________________________________________ -->
6879 <a name="int_frameaddress">'<tt>llvm.frameaddress</tt>' Intrinsic</a>
6886 declare i8* @llvm.frameaddress(i32 <level>)
6890 <p>The '<tt>llvm.frameaddress</tt>' intrinsic attempts to return the
6891 target-specific frame pointer value for the specified stack frame.</p>
6894 <p>The argument to this intrinsic indicates which function to return the frame
6895 pointer for. Zero indicates the calling function, one indicates its caller,
6896 etc. The argument is <b>required</b> to be a constant integer value.</p>
6899 <p>The '<tt>llvm.frameaddress</tt>' intrinsic either returns a pointer
6900 indicating the frame address of the specified call frame, or zero if it
6901 cannot be identified. The value returned by this intrinsic is likely to be
6902 incorrect or 0 for arguments other than zero, so it should only be used for
6903 debugging purposes.</p>
6905 <p>Note that calling this intrinsic does not prevent function inlining or other
6906 aggressive transformations, so the value returned may not be that of the
6907 obvious source-language caller.</p>
6911 <!-- _______________________________________________________________________ -->
6913 <a name="int_stacksave">'<tt>llvm.stacksave</tt>' Intrinsic</a>
6920 declare i8* @llvm.stacksave()
6924 <p>The '<tt>llvm.stacksave</tt>' intrinsic is used to remember the current state
6925 of the function stack, for use
6926 with <a href="#int_stackrestore"> <tt>llvm.stackrestore</tt></a>. This is
6927 useful for implementing language features like scoped automatic variable
6928 sized arrays in C99.</p>
6931 <p>This intrinsic returns a opaque pointer value that can be passed
6932 to <a href="#int_stackrestore"><tt>llvm.stackrestore</tt></a>. When
6933 an <tt>llvm.stackrestore</tt> intrinsic is executed with a value saved
6934 from <tt>llvm.stacksave</tt>, it effectively restores the state of the stack
6935 to the state it was in when the <tt>llvm.stacksave</tt> intrinsic executed.
6936 In practice, this pops any <a href="#i_alloca">alloca</a> blocks from the
6937 stack that were allocated after the <tt>llvm.stacksave</tt> was executed.</p>
6941 <!-- _______________________________________________________________________ -->
6943 <a name="int_stackrestore">'<tt>llvm.stackrestore</tt>' Intrinsic</a>
6950 declare void @llvm.stackrestore(i8* %ptr)
6954 <p>The '<tt>llvm.stackrestore</tt>' intrinsic is used to restore the state of
6955 the function stack to the state it was in when the
6956 corresponding <a href="#int_stacksave"><tt>llvm.stacksave</tt></a> intrinsic
6957 executed. This is useful for implementing language features like scoped
6958 automatic variable sized arrays in C99.</p>
6961 <p>See the description
6962 for <a href="#int_stacksave"><tt>llvm.stacksave</tt></a>.</p>
6966 <!-- _______________________________________________________________________ -->
6968 <a name="int_prefetch">'<tt>llvm.prefetch</tt>' Intrinsic</a>
6975 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6979 <p>The '<tt>llvm.prefetch</tt>' intrinsic is a hint to the code generator to
6980 insert a prefetch instruction if supported; otherwise, it is a noop.
6981 Prefetches have no effect on the behavior of the program but can change its
6982 performance characteristics.</p>
6985 <p><tt>address</tt> is the address to be prefetched, <tt>rw</tt> is the
6986 specifier determining if the fetch should be for a read (0) or write (1),
6987 and <tt>locality</tt> is a temporal locality specifier ranging from (0) - no
6988 locality, to (3) - extremely local keep in cache. The <tt>cache type</tt>
6989 specifies whether the prefetch is performed on the data (1) or instruction (0)
6990 cache. The <tt>rw</tt>, <tt>locality</tt> and <tt>cache type</tt> arguments
6991 must be constant integers.</p>
6994 <p>This intrinsic does not modify the behavior of the program. In particular,
6995 prefetches cannot trap and do not produce a value. On targets that support
6996 this intrinsic, the prefetch can provide hints to the processor cache for
6997 better performance.</p>
7001 <!-- _______________________________________________________________________ -->
7003 <a name="int_pcmarker">'<tt>llvm.pcmarker</tt>' Intrinsic</a>
7010 declare void @llvm.pcmarker(i32 <id>)
7014 <p>The '<tt>llvm.pcmarker</tt>' intrinsic is a method to export a Program
7015 Counter (PC) in a region of code to simulators and other tools. The method
7016 is target specific, but it is expected that the marker will use exported
7017 symbols to transmit the PC of the marker. The marker makes no guarantees
7018 that it will remain with any specific instruction after optimizations. It is
7019 possible that the presence of a marker will inhibit optimizations. The
7020 intended use is to be inserted after optimizations to allow correlations of
7021 simulation runs.</p>
7024 <p><tt>id</tt> is a numerical id identifying the marker.</p>
7027 <p>This intrinsic does not modify the behavior of the program. Backends that do
7028 not support this intrinsic may ignore it.</p>
7032 <!-- _______________________________________________________________________ -->
7034 <a name="int_readcyclecounter">'<tt>llvm.readcyclecounter</tt>' Intrinsic</a>
7041 declare i64 @llvm.readcyclecounter()
7045 <p>The '<tt>llvm.readcyclecounter</tt>' intrinsic provides access to the cycle
7046 counter register (or similar low latency, high accuracy clocks) on those
7047 targets that support it. On X86, it should map to RDTSC. On Alpha, it
7048 should map to RPCC. As the backing counters overflow quickly (on the order
7049 of 9 seconds on alpha), this should only be used for small timings.</p>
7052 <p>When directly supported, reading the cycle counter should not modify any
7053 memory. Implementations are allowed to either return a application specific
7054 value or a system wide value. On backends without support, this is lowered
7055 to a constant 0.</p>
7061 <!-- ======================================================================= -->
7063 <a name="int_libc">Standard C Library Intrinsics</a>
7068 <p>LLVM provides intrinsics for a few important standard C library functions.
7069 These intrinsics allow source-language front-ends to pass information about
7070 the alignment of the pointer arguments to the code generator, providing
7071 opportunity for more efficient code generation.</p>
7073 <!-- _______________________________________________________________________ -->
7075 <a name="int_memcpy">'<tt>llvm.memcpy</tt>' Intrinsic</a>
7081 <p>This is an overloaded intrinsic. You can use <tt>llvm.memcpy</tt> on any
7082 integer bit width and for different address spaces. Not all targets support
7083 all bit widths however.</p>
7086 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7087 i32 <len>, i32 <align>, i1 <isvolatile>)
7088 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7089 i64 <len>, i32 <align>, i1 <isvolatile>)
7093 <p>The '<tt>llvm.memcpy.*</tt>' intrinsics copy a block of memory from the
7094 source location to the destination location.</p>
7096 <p>Note that, unlike the standard libc function, the <tt>llvm.memcpy.*</tt>
7097 intrinsics do not return a value, takes extra alignment/isvolatile arguments
7098 and the pointers can be in specified address spaces.</p>
7102 <p>The first argument is a pointer to the destination, the second is a pointer
7103 to the source. The third argument is an integer argument specifying the
7104 number of bytes to copy, the fourth argument is the alignment of the
7105 source and destination locations, and the fifth is a boolean indicating a
7106 volatile access.</p>
7108 <p>If the call to this intrinsic has an alignment value that is not 0 or 1,
7109 then the caller guarantees that both the source and destination pointers are
7110 aligned to that boundary.</p>
7112 <p>If the <tt>isvolatile</tt> parameter is <tt>true</tt>, the
7113 <tt>llvm.memcpy</tt> call is a <a href="#volatile">volatile operation</a>.
7114 The detailed access behavior is not very cleanly specified and it is unwise
7115 to depend on it.</p>
7119 <p>The '<tt>llvm.memcpy.*</tt>' intrinsics copy a block of memory from the
7120 source location to the destination location, which are not allowed to
7121 overlap. It copies "len" bytes of memory over. If the argument is known to
7122 be aligned to some boundary, this can be specified as the fourth argument,
7123 otherwise it should be set to 0 or 1.</p>
7127 <!-- _______________________________________________________________________ -->
7129 <a name="int_memmove">'<tt>llvm.memmove</tt>' Intrinsic</a>
7135 <p>This is an overloaded intrinsic. You can use llvm.memmove on any integer bit
7136 width and for different address space. Not all targets support all bit
7140 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7141 i32 <len>, i32 <align>, i1 <isvolatile>)
7142 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7143 i64 <len>, i32 <align>, i1 <isvolatile>)
7147 <p>The '<tt>llvm.memmove.*</tt>' intrinsics move a block of memory from the
7148 source location to the destination location. It is similar to the
7149 '<tt>llvm.memcpy</tt>' intrinsic but allows the two memory locations to
7152 <p>Note that, unlike the standard libc function, the <tt>llvm.memmove.*</tt>
7153 intrinsics do not return a value, takes extra alignment/isvolatile arguments
7154 and the pointers can be in specified address spaces.</p>
7158 <p>The first argument is a pointer to the destination, the second is a pointer
7159 to the source. The third argument is an integer argument specifying the
7160 number of bytes to copy, the fourth argument is the alignment of the
7161 source and destination locations, and the fifth is a boolean indicating a
7162 volatile access.</p>
7164 <p>If the call to this intrinsic has an alignment value that is not 0 or 1,
7165 then the caller guarantees that the source and destination pointers are
7166 aligned to that boundary.</p>
7168 <p>If the <tt>isvolatile</tt> parameter is <tt>true</tt>, the
7169 <tt>llvm.memmove</tt> call is a <a href="#volatile">volatile operation</a>.
7170 The detailed access behavior is not very cleanly specified and it is unwise
7171 to depend on it.</p>
7175 <p>The '<tt>llvm.memmove.*</tt>' intrinsics copy a block of memory from the
7176 source location to the destination location, which may overlap. It copies
7177 "len" bytes of memory over. If the argument is known to be aligned to some
7178 boundary, this can be specified as the fourth argument, otherwise it should
7179 be set to 0 or 1.</p>
7183 <!-- _______________________________________________________________________ -->
7185 <a name="int_memset">'<tt>llvm.memset.*</tt>' Intrinsics</a>
7191 <p>This is an overloaded intrinsic. You can use llvm.memset on any integer bit
7192 width and for different address spaces. However, not all targets support all
7196 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7197 i32 <len>, i32 <align>, i1 <isvolatile>)
7198 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7199 i64 <len>, i32 <align>, i1 <isvolatile>)
7203 <p>The '<tt>llvm.memset.*</tt>' intrinsics fill a block of memory with a
7204 particular byte value.</p>
7206 <p>Note that, unlike the standard libc function, the <tt>llvm.memset</tt>
7207 intrinsic does not return a value and takes extra alignment/volatile
7208 arguments. Also, the destination can be in an arbitrary address space.</p>
7211 <p>The first argument is a pointer to the destination to fill, the second is the
7212 byte value with which to fill it, the third argument is an integer argument
7213 specifying the number of bytes to fill, and the fourth argument is the known
7214 alignment of the destination location.</p>
7216 <p>If the call to this intrinsic has an alignment value that is not 0 or 1,
7217 then the caller guarantees that the destination pointer is aligned to that
7220 <p>If the <tt>isvolatile</tt> parameter is <tt>true</tt>, the
7221 <tt>llvm.memset</tt> call is a <a href="#volatile">volatile operation</a>.
7222 The detailed access behavior is not very cleanly specified and it is unwise
7223 to depend on it.</p>
7226 <p>The '<tt>llvm.memset.*</tt>' intrinsics fill "len" bytes of memory starting
7227 at the destination location. If the argument is known to be aligned to some
7228 boundary, this can be specified as the fourth argument, otherwise it should
7229 be set to 0 or 1.</p>
7233 <!-- _______________________________________________________________________ -->
7235 <a name="int_sqrt">'<tt>llvm.sqrt.*</tt>' Intrinsic</a>
7241 <p>This is an overloaded intrinsic. You can use <tt>llvm.sqrt</tt> on any
7242 floating point or vector of floating point type. Not all targets support all
7246 declare float @llvm.sqrt.f32(float %Val)
7247 declare double @llvm.sqrt.f64(double %Val)
7248 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7249 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7250 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7254 <p>The '<tt>llvm.sqrt</tt>' intrinsics return the sqrt of the specified operand,
7255 returning the same value as the libm '<tt>sqrt</tt>' functions would.
7256 Unlike <tt>sqrt</tt> in libm, however, <tt>llvm.sqrt</tt> has undefined
7257 behavior for negative numbers other than -0.0 (which allows for better
7258 optimization, because there is no need to worry about errno being
7259 set). <tt>llvm.sqrt(-0.0)</tt> is defined to return -0.0 like IEEE sqrt.</p>
7262 <p>The argument and return value are floating point numbers of the same
7266 <p>This function returns the sqrt of the specified operand if it is a
7267 nonnegative floating point number.</p>
7271 <!-- _______________________________________________________________________ -->
7273 <a name="int_powi">'<tt>llvm.powi.*</tt>' Intrinsic</a>
7279 <p>This is an overloaded intrinsic. You can use <tt>llvm.powi</tt> on any
7280 floating point or vector of floating point type. Not all targets support all
7284 declare float @llvm.powi.f32(float %Val, i32 %power)
7285 declare double @llvm.powi.f64(double %Val, i32 %power)
7286 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7287 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7288 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7292 <p>The '<tt>llvm.powi.*</tt>' intrinsics return the first operand raised to the
7293 specified (positive or negative) power. The order of evaluation of
7294 multiplications is not defined. When a vector of floating point type is
7295 used, the second argument remains a scalar integer value.</p>
7298 <p>The second argument is an integer power, and the first is a value to raise to
7302 <p>This function returns the first value raised to the second power with an
7303 unspecified sequence of rounding operations.</p>
7307 <!-- _______________________________________________________________________ -->
7309 <a name="int_sin">'<tt>llvm.sin.*</tt>' Intrinsic</a>
7315 <p>This is an overloaded intrinsic. You can use <tt>llvm.sin</tt> on any
7316 floating point or vector of floating point type. Not all targets support all
7320 declare float @llvm.sin.f32(float %Val)
7321 declare double @llvm.sin.f64(double %Val)
7322 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7323 declare fp128 @llvm.sin.f128(fp128 %Val)
7324 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7328 <p>The '<tt>llvm.sin.*</tt>' intrinsics return the sine of the operand.</p>
7331 <p>The argument and return value are floating point numbers of the same
7335 <p>This function returns the sine of the specified operand, returning the same
7336 values as the libm <tt>sin</tt> functions would, and handles error conditions
7337 in the same way.</p>
7341 <!-- _______________________________________________________________________ -->
7343 <a name="int_cos">'<tt>llvm.cos.*</tt>' Intrinsic</a>
7349 <p>This is an overloaded intrinsic. You can use <tt>llvm.cos</tt> on any
7350 floating point or vector of floating point type. Not all targets support all
7354 declare float @llvm.cos.f32(float %Val)
7355 declare double @llvm.cos.f64(double %Val)
7356 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7357 declare fp128 @llvm.cos.f128(fp128 %Val)
7358 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7362 <p>The '<tt>llvm.cos.*</tt>' intrinsics return the cosine of the operand.</p>
7365 <p>The argument and return value are floating point numbers of the same
7369 <p>This function returns the cosine of the specified operand, returning the same
7370 values as the libm <tt>cos</tt> functions would, and handles error conditions
7371 in the same way.</p>
7375 <!-- _______________________________________________________________________ -->
7377 <a name="int_pow">'<tt>llvm.pow.*</tt>' Intrinsic</a>
7383 <p>This is an overloaded intrinsic. You can use <tt>llvm.pow</tt> on any
7384 floating point or vector of floating point type. Not all targets support all
7388 declare float @llvm.pow.f32(float %Val, float %Power)
7389 declare double @llvm.pow.f64(double %Val, double %Power)
7390 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7391 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7392 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7396 <p>The '<tt>llvm.pow.*</tt>' intrinsics return the first operand raised to the
7397 specified (positive or negative) power.</p>
7400 <p>The second argument is a floating point power, and the first is a value to
7401 raise to that power.</p>
7404 <p>This function returns the first value raised to the second power, returning
7405 the same values as the libm <tt>pow</tt> functions would, and handles error
7406 conditions in the same way.</p>
7410 <!-- _______________________________________________________________________ -->
7412 <a name="int_exp">'<tt>llvm.exp.*</tt>' Intrinsic</a>
7418 <p>This is an overloaded intrinsic. You can use <tt>llvm.exp</tt> on any
7419 floating point or vector of floating point type. Not all targets support all
7423 declare float @llvm.exp.f32(float %Val)
7424 declare double @llvm.exp.f64(double %Val)
7425 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7426 declare fp128 @llvm.exp.f128(fp128 %Val)
7427 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7431 <p>The '<tt>llvm.exp.*</tt>' intrinsics perform the exp function.</p>
7434 <p>The argument and return value are floating point numbers of the same
7438 <p>This function returns the same values as the libm <tt>exp</tt> functions
7439 would, and handles error conditions in the same way.</p>
7443 <!-- _______________________________________________________________________ -->
7445 <a name="int_log">'<tt>llvm.log.*</tt>' Intrinsic</a>
7451 <p>This is an overloaded intrinsic. You can use <tt>llvm.log</tt> on any
7452 floating point or vector of floating point type. Not all targets support all
7456 declare float @llvm.log.f32(float %Val)
7457 declare double @llvm.log.f64(double %Val)
7458 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7459 declare fp128 @llvm.log.f128(fp128 %Val)
7460 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7464 <p>The '<tt>llvm.log.*</tt>' intrinsics perform the log function.</p>
7467 <p>The argument and return value are floating point numbers of the same
7471 <p>This function returns the same values as the libm <tt>log</tt> functions
7472 would, and handles error conditions in the same way.</p>
7476 <!-- _______________________________________________________________________ -->
7478 <a name="int_fma">'<tt>llvm.fma.*</tt>' Intrinsic</a>
7484 <p>This is an overloaded intrinsic. You can use <tt>llvm.fma</tt> on any
7485 floating point or vector of floating point type. Not all targets support all
7489 declare float @llvm.fma.f32(float %a, float %b, float %c)
7490 declare double @llvm.fma.f64(double %a, double %b, double %c)
7491 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7492 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7493 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7497 <p>The '<tt>llvm.fma.*</tt>' intrinsics perform the fused multiply-add
7501 <p>The argument and return value are floating point numbers of the same
7505 <p>This function returns the same values as the libm <tt>fma</tt> functions
7510 <!-- _______________________________________________________________________ -->
7512 <a name="int_fabs">'<tt>llvm.fabs.*</tt>' Intrinsic</a>
7518 <p>This is an overloaded intrinsic. You can use <tt>llvm.fabs</tt> on any
7519 floating point or vector of floating point type. Not all targets support all
7523 declare float @llvm.fabs.f32(float %Val)
7524 declare double @llvm.fabs.f64(double %Val)
7525 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7526 declare fp128 @llvm.fabs.f128(fp128 %Val)
7527 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7531 <p>The '<tt>llvm.fabs.*</tt>' intrinsics return the absolute value of
7535 <p>The argument and return value are floating point numbers of the same
7539 <p>This function returns the same values as the libm <tt>fabs</tt> functions
7540 would, and handles error conditions in the same way.</p>
7546 <!-- ======================================================================= -->
7548 <a name="int_manip">Bit Manipulation Intrinsics</a>
7553 <p>LLVM provides intrinsics for a few important bit manipulation operations.
7554 These allow efficient code generation for some algorithms.</p>
7556 <!-- _______________________________________________________________________ -->
7558 <a name="int_bswap">'<tt>llvm.bswap.*</tt>' Intrinsics</a>
7564 <p>This is an overloaded intrinsic function. You can use bswap on any integer
7565 type that is an even number of bytes (i.e. BitWidth % 16 == 0).</p>
7568 declare i16 @llvm.bswap.i16(i16 <id>)
7569 declare i32 @llvm.bswap.i32(i32 <id>)
7570 declare i64 @llvm.bswap.i64(i64 <id>)
7574 <p>The '<tt>llvm.bswap</tt>' family of intrinsics is used to byte swap integer
7575 values with an even number of bytes (positive multiple of 16 bits). These
7576 are useful for performing operations on data that is not in the target's
7577 native byte order.</p>
7580 <p>The <tt>llvm.bswap.i16</tt> intrinsic returns an i16 value that has the high
7581 and low byte of the input i16 swapped. Similarly,
7582 the <tt>llvm.bswap.i32</tt> intrinsic returns an i32 value that has the four
7583 bytes of the input i32 swapped, so that if the input bytes are numbered 0, 1,
7584 2, 3 then the returned i32 will have its bytes in 3, 2, 1, 0 order.
7585 The <tt>llvm.bswap.i48</tt>, <tt>llvm.bswap.i64</tt> and other intrinsics
7586 extend this concept to additional even-byte lengths (6 bytes, 8 bytes and
7587 more, respectively).</p>
7591 <!-- _______________________________________________________________________ -->
7593 <a name="int_ctpop">'<tt>llvm.ctpop.*</tt>' Intrinsic</a>
7599 <p>This is an overloaded intrinsic. You can use llvm.ctpop on any integer bit
7600 width, or on any vector with integer elements. Not all targets support all
7601 bit widths or vector types, however.</p>
7604 declare i8 @llvm.ctpop.i8(i8 <src>)
7605 declare i16 @llvm.ctpop.i16(i16 <src>)
7606 declare i32 @llvm.ctpop.i32(i32 <src>)
7607 declare i64 @llvm.ctpop.i64(i64 <src>)
7608 declare i256 @llvm.ctpop.i256(i256 <src>)
7609 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7613 <p>The '<tt>llvm.ctpop</tt>' family of intrinsics counts the number of bits set
7617 <p>The only argument is the value to be counted. The argument may be of any
7618 integer type, or a vector with integer elements.
7619 The return type must match the argument type.</p>
7622 <p>The '<tt>llvm.ctpop</tt>' intrinsic counts the 1's in a variable, or within each
7623 element of a vector.</p>
7627 <!-- _______________________________________________________________________ -->
7629 <a name="int_ctlz">'<tt>llvm.ctlz.*</tt>' Intrinsic</a>
7635 <p>This is an overloaded intrinsic. You can use <tt>llvm.ctlz</tt> on any
7636 integer bit width, or any vector whose elements are integers. Not all
7637 targets support all bit widths or vector types, however.</p>
7640 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7641 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7642 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7643 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7644 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7645 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7649 <p>The '<tt>llvm.ctlz</tt>' family of intrinsic functions counts the number of
7650 leading zeros in a variable.</p>
7653 <p>The first argument is the value to be counted. This argument may be of any
7654 integer type, or a vectory with integer element type. The return type
7655 must match the first argument type.</p>
7657 <p>The second argument must be a constant and is a flag to indicate whether the
7658 intrinsic should ensure that a zero as the first argument produces a defined
7659 result. Historically some architectures did not provide a defined result for
7660 zero values as efficiently, and many algorithms are now predicated on
7661 avoiding zero-value inputs.</p>
7664 <p>The '<tt>llvm.ctlz</tt>' intrinsic counts the leading (most significant)
7665 zeros in a variable, or within each element of the vector.
7666 If <tt>src == 0</tt> then the result is the size in bits of the type of
7667 <tt>src</tt> if <tt>is_zero_undef == 0</tt> and <tt>undef</tt> otherwise.
7668 For example, <tt>llvm.ctlz(i32 2) = 30</tt>.</p>
7672 <!-- _______________________________________________________________________ -->
7674 <a name="int_cttz">'<tt>llvm.cttz.*</tt>' Intrinsic</a>
7680 <p>This is an overloaded intrinsic. You can use <tt>llvm.cttz</tt> on any
7681 integer bit width, or any vector of integer elements. Not all targets
7682 support all bit widths or vector types, however.</p>
7685 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7686 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7687 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7688 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7689 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7690 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7694 <p>The '<tt>llvm.cttz</tt>' family of intrinsic functions counts the number of
7698 <p>The first argument is the value to be counted. This argument may be of any
7699 integer type, or a vectory with integer element type. The return type
7700 must match the first argument type.</p>
7702 <p>The second argument must be a constant and is a flag to indicate whether the
7703 intrinsic should ensure that a zero as the first argument produces a defined
7704 result. Historically some architectures did not provide a defined result for
7705 zero values as efficiently, and many algorithms are now predicated on
7706 avoiding zero-value inputs.</p>
7709 <p>The '<tt>llvm.cttz</tt>' intrinsic counts the trailing (least significant)
7710 zeros in a variable, or within each element of a vector.
7711 If <tt>src == 0</tt> then the result is the size in bits of the type of
7712 <tt>src</tt> if <tt>is_zero_undef == 0</tt> and <tt>undef</tt> otherwise.
7713 For example, <tt>llvm.cttz(2) = 1</tt>.</p>
7719 <!-- ======================================================================= -->
7721 <a name="int_overflow">Arithmetic with Overflow Intrinsics</a>
7726 <p>LLVM provides intrinsics for some arithmetic with overflow operations.</p>
7728 <!-- _______________________________________________________________________ -->
7730 <a name="int_sadd_overflow">
7731 '<tt>llvm.sadd.with.overflow.*</tt>' Intrinsics
7738 <p>This is an overloaded intrinsic. You can use <tt>llvm.sadd.with.overflow</tt>
7739 on any integer bit width.</p>
7742 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7743 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7744 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7748 <p>The '<tt>llvm.sadd.with.overflow</tt>' family of intrinsic functions perform
7749 a signed addition of the two arguments, and indicate whether an overflow
7750 occurred during the signed summation.</p>
7753 <p>The arguments (%a and %b) and the first element of the result structure may
7754 be of integer types of any bit width, but they must have the same bit
7755 width. The second element of the result structure must be of
7756 type <tt>i1</tt>. <tt>%a</tt> and <tt>%b</tt> are the two values that will
7757 undergo signed addition.</p>
7760 <p>The '<tt>llvm.sadd.with.overflow</tt>' family of intrinsic functions perform
7761 a signed addition of the two variables. They return a structure — the
7762 first element of which is the signed summation, and the second element of
7763 which is a bit specifying if the signed summation resulted in an
7768 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7769 %sum = extractvalue {i32, i1} %res, 0
7770 %obit = extractvalue {i32, i1} %res, 1
7771 br i1 %obit, label %overflow, label %normal
7776 <!-- _______________________________________________________________________ -->
7778 <a name="int_uadd_overflow">
7779 '<tt>llvm.uadd.with.overflow.*</tt>' Intrinsics
7786 <p>This is an overloaded intrinsic. You can use <tt>llvm.uadd.with.overflow</tt>
7787 on any integer bit width.</p>
7790 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
7791 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7792 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
7796 <p>The '<tt>llvm.uadd.with.overflow</tt>' family of intrinsic functions perform
7797 an unsigned addition of the two arguments, and indicate whether a carry
7798 occurred during the unsigned summation.</p>
7801 <p>The arguments (%a and %b) and the first element of the result structure may
7802 be of integer types of any bit width, but they must have the same bit
7803 width. The second element of the result structure must be of
7804 type <tt>i1</tt>. <tt>%a</tt> and <tt>%b</tt> are the two values that will
7805 undergo unsigned addition.</p>
7808 <p>The '<tt>llvm.uadd.with.overflow</tt>' family of intrinsic functions perform
7809 an unsigned addition of the two arguments. They return a structure —
7810 the first element of which is the sum, and the second element of which is a
7811 bit specifying if the unsigned summation resulted in a carry.</p>
7815 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7816 %sum = extractvalue {i32, i1} %res, 0
7817 %obit = extractvalue {i32, i1} %res, 1
7818 br i1 %obit, label %carry, label %normal
7823 <!-- _______________________________________________________________________ -->
7825 <a name="int_ssub_overflow">
7826 '<tt>llvm.ssub.with.overflow.*</tt>' Intrinsics
7833 <p>This is an overloaded intrinsic. You can use <tt>llvm.ssub.with.overflow</tt>
7834 on any integer bit width.</p>
7837 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
7838 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7839 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
7843 <p>The '<tt>llvm.ssub.with.overflow</tt>' family of intrinsic functions perform
7844 a signed subtraction of the two arguments, and indicate whether an overflow
7845 occurred during the signed subtraction.</p>
7848 <p>The arguments (%a and %b) and the first element of the result structure may
7849 be of integer types of any bit width, but they must have the same bit
7850 width. The second element of the result structure must be of
7851 type <tt>i1</tt>. <tt>%a</tt> and <tt>%b</tt> are the two values that will
7852 undergo signed subtraction.</p>
7855 <p>The '<tt>llvm.ssub.with.overflow</tt>' family of intrinsic functions perform
7856 a signed subtraction of the two arguments. They return a structure —
7857 the first element of which is the subtraction, and the second element of
7858 which is a bit specifying if the signed subtraction resulted in an
7863 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7864 %sum = extractvalue {i32, i1} %res, 0
7865 %obit = extractvalue {i32, i1} %res, 1
7866 br i1 %obit, label %overflow, label %normal
7871 <!-- _______________________________________________________________________ -->
7873 <a name="int_usub_overflow">
7874 '<tt>llvm.usub.with.overflow.*</tt>' Intrinsics
7881 <p>This is an overloaded intrinsic. You can use <tt>llvm.usub.with.overflow</tt>
7882 on any integer bit width.</p>
7885 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
7886 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7887 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
7891 <p>The '<tt>llvm.usub.with.overflow</tt>' family of intrinsic functions perform
7892 an unsigned subtraction of the two arguments, and indicate whether an
7893 overflow occurred during the unsigned subtraction.</p>
7896 <p>The arguments (%a and %b) and the first element of the result structure may
7897 be of integer types of any bit width, but they must have the same bit
7898 width. The second element of the result structure must be of
7899 type <tt>i1</tt>. <tt>%a</tt> and <tt>%b</tt> are the two values that will
7900 undergo unsigned subtraction.</p>
7903 <p>The '<tt>llvm.usub.with.overflow</tt>' family of intrinsic functions perform
7904 an unsigned subtraction of the two arguments. They return a structure —
7905 the first element of which is the subtraction, and the second element of
7906 which is a bit specifying if the unsigned subtraction resulted in an
7911 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7912 %sum = extractvalue {i32, i1} %res, 0
7913 %obit = extractvalue {i32, i1} %res, 1
7914 br i1 %obit, label %overflow, label %normal
7919 <!-- _______________________________________________________________________ -->
7921 <a name="int_smul_overflow">
7922 '<tt>llvm.smul.with.overflow.*</tt>' Intrinsics
7929 <p>This is an overloaded intrinsic. You can use <tt>llvm.smul.with.overflow</tt>
7930 on any integer bit width.</p>
7933 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
7934 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7935 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
7940 <p>The '<tt>llvm.smul.with.overflow</tt>' family of intrinsic functions perform
7941 a signed multiplication of the two arguments, and indicate whether an
7942 overflow occurred during the signed multiplication.</p>
7945 <p>The arguments (%a and %b) and the first element of the result structure may
7946 be of integer types of any bit width, but they must have the same bit
7947 width. The second element of the result structure must be of
7948 type <tt>i1</tt>. <tt>%a</tt> and <tt>%b</tt> are the two values that will
7949 undergo signed multiplication.</p>
7952 <p>The '<tt>llvm.smul.with.overflow</tt>' family of intrinsic functions perform
7953 a signed multiplication of the two arguments. They return a structure —
7954 the first element of which is the multiplication, and the second element of
7955 which is a bit specifying if the signed multiplication resulted in an
7960 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7961 %sum = extractvalue {i32, i1} %res, 0
7962 %obit = extractvalue {i32, i1} %res, 1
7963 br i1 %obit, label %overflow, label %normal
7968 <!-- _______________________________________________________________________ -->
7970 <a name="int_umul_overflow">
7971 '<tt>llvm.umul.with.overflow.*</tt>' Intrinsics
7978 <p>This is an overloaded intrinsic. You can use <tt>llvm.umul.with.overflow</tt>
7979 on any integer bit width.</p>
7982 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
7983 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
7984 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
7988 <p>The '<tt>llvm.umul.with.overflow</tt>' family of intrinsic functions perform
7989 a unsigned multiplication of the two arguments, and indicate whether an
7990 overflow occurred during the unsigned multiplication.</p>
7993 <p>The arguments (%a and %b) and the first element of the result structure may
7994 be of integer types of any bit width, but they must have the same bit
7995 width. The second element of the result structure must be of
7996 type <tt>i1</tt>. <tt>%a</tt> and <tt>%b</tt> are the two values that will
7997 undergo unsigned multiplication.</p>
8000 <p>The '<tt>llvm.umul.with.overflow</tt>' family of intrinsic functions perform
8001 an unsigned multiplication of the two arguments. They return a structure
8002 — the first element of which is the multiplication, and the second
8003 element of which is a bit specifying if the unsigned multiplication resulted
8008 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8009 %sum = extractvalue {i32, i1} %res, 0
8010 %obit = extractvalue {i32, i1} %res, 1
8011 br i1 %obit, label %overflow, label %normal
8018 <!-- ======================================================================= -->
8020 <a name="spec_arithmetic">Specialised Arithmetic Intrinsics</a>
8023 <!-- _______________________________________________________________________ -->
8026 <a name="fmuladd">'<tt>llvm.fmuladd.*</tt>' Intrinsic</a>
8033 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8034 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8038 <p>The '<tt>llvm.fmuladd.*</tt>' intrinsic functions represent multiply-add
8039 expressions that can be fused if the code generator determines that the fused
8040 expression would be legal and efficient.</p>
8043 <p>The '<tt>llvm.fmuladd.*</tt>' intrinsics each take three arguments: two
8044 multiplicands, a and b, and an addend c.</p>
8047 <p>The expression:</p>
8049 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8051 <p>is equivalent to the expression a * b + c, except that rounding will not be
8052 performed between the multiplication and addition steps if the code generator
8053 fuses the operations. Fusion is not guaranteed, even if the target platform
8054 supports it. If a fused multiply-add is required the corresponding llvm.fma.*
8055 intrinsic function should be used instead.</p>
8059 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8064 <!-- ======================================================================= -->
8066 <a name="int_fp16">Half Precision Floating Point Intrinsics</a>
8071 <p>For most target platforms, half precision floating point is a storage-only
8072 format. This means that it is
8073 a dense encoding (in memory) but does not support computation in the
8076 <p>This means that code must first load the half-precision floating point
8077 value as an i16, then convert it to float with <a
8078 href="#int_convert_from_fp16"><tt>llvm.convert.from.fp16</tt></a>.
8079 Computation can then be performed on the float value (including extending to
8080 double etc). To store the value back to memory, it is first converted to
8081 float if needed, then converted to i16 with
8082 <a href="#int_convert_to_fp16"><tt>llvm.convert.to.fp16</tt></a>, then
8083 storing as an i16 value.</p>
8085 <!-- _______________________________________________________________________ -->
8087 <a name="int_convert_to_fp16">
8088 '<tt>llvm.convert.to.fp16</tt>' Intrinsic
8096 declare i16 @llvm.convert.to.fp16(f32 %a)
8100 <p>The '<tt>llvm.convert.to.fp16</tt>' intrinsic function performs
8101 a conversion from single precision floating point format to half precision
8102 floating point format.</p>
8105 <p>The intrinsic function contains single argument - the value to be
8109 <p>The '<tt>llvm.convert.to.fp16</tt>' intrinsic function performs
8110 a conversion from single precision floating point format to half precision
8111 floating point format. The return value is an <tt>i16</tt> which
8112 contains the converted number.</p>
8116 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8117 store i16 %res, i16* @x, align 2
8122 <!-- _______________________________________________________________________ -->
8124 <a name="int_convert_from_fp16">
8125 '<tt>llvm.convert.from.fp16</tt>' Intrinsic
8133 declare f32 @llvm.convert.from.fp16(i16 %a)
8137 <p>The '<tt>llvm.convert.from.fp16</tt>' intrinsic function performs
8138 a conversion from half precision floating point format to single precision
8139 floating point format.</p>
8142 <p>The intrinsic function contains single argument - the value to be
8146 <p>The '<tt>llvm.convert.from.fp16</tt>' intrinsic function performs a
8147 conversion from half single precision floating point format to single
8148 precision floating point format. The input half-float value is represented by
8149 an <tt>i16</tt> value.</p>
8153 %a = load i16* @x, align 2
8154 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8161 <!-- ======================================================================= -->
8163 <a name="int_debugger">Debugger Intrinsics</a>
8168 <p>The LLVM debugger intrinsics (which all start with <tt>llvm.dbg.</tt>
8169 prefix), are described in
8170 the <a href="SourceLevelDebugging.html#format_common_intrinsics">LLVM Source
8171 Level Debugging</a> document.</p>
8175 <!-- ======================================================================= -->
8177 <a name="int_eh">Exception Handling Intrinsics</a>
8182 <p>The LLVM exception handling intrinsics (which all start with
8183 <tt>llvm.eh.</tt> prefix), are described in
8184 the <a href="ExceptionHandling.html#format_common_intrinsics">LLVM Exception
8185 Handling</a> document.</p>
8189 <!-- ======================================================================= -->
8191 <a name="int_trampoline">Trampoline Intrinsics</a>
8196 <p>These intrinsics make it possible to excise one parameter, marked with
8197 the <a href="#nest"><tt>nest</tt></a> attribute, from a function.
8198 The result is a callable
8199 function pointer lacking the nest parameter - the caller does not need to
8200 provide a value for it. Instead, the value to use is stored in advance in a
8201 "trampoline", a block of memory usually allocated on the stack, which also
8202 contains code to splice the nest value into the argument list. This is used
8203 to implement the GCC nested function address extension.</p>
8205 <p>For example, if the function is
8206 <tt>i32 f(i8* nest %c, i32 %x, i32 %y)</tt> then the resulting function
8207 pointer has signature <tt>i32 (i32, i32)*</tt>. It can be created as
8210 <pre class="doc_code">
8211 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8212 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8213 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8214 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8215 %fp = bitcast i8* %p to i32 (i32, i32)*
8218 <p>The call <tt>%val = call i32 %fp(i32 %x, i32 %y)</tt> is then equivalent
8219 to <tt>%val = call i32 %f(i8* %nval, i32 %x, i32 %y)</tt>.</p>
8221 <!-- _______________________________________________________________________ -->
8224 '<tt>llvm.init.trampoline</tt>' Intrinsic
8232 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8236 <p>This fills the memory pointed to by <tt>tramp</tt> with executable code,
8237 turning it into a trampoline.</p>
8240 <p>The <tt>llvm.init.trampoline</tt> intrinsic takes three arguments, all
8241 pointers. The <tt>tramp</tt> argument must point to a sufficiently large and
8242 sufficiently aligned block of memory; this memory is written to by the
8243 intrinsic. Note that the size and the alignment are target-specific - LLVM
8244 currently provides no portable way of determining them, so a front-end that
8245 generates this intrinsic needs to have some target-specific knowledge.
8246 The <tt>func</tt> argument must hold a function bitcast to
8247 an <tt>i8*</tt>.</p>
8250 <p>The block of memory pointed to by <tt>tramp</tt> is filled with target
8251 dependent code, turning it into a function. Then <tt>tramp</tt> needs to be
8252 passed to <a href="#int_at">llvm.adjust.trampoline</a> to get a pointer
8253 which can be <a href="#int_trampoline">bitcast (to a new function) and
8254 called</a>. The new function's signature is the same as that of
8255 <tt>func</tt> with any arguments marked with the <tt>nest</tt> attribute
8256 removed. At most one such <tt>nest</tt> argument is allowed, and it must be of
8257 pointer type. Calling the new function is equivalent to calling <tt>func</tt>
8258 with the same argument list, but with <tt>nval</tt> used for the missing
8259 <tt>nest</tt> argument. If, after calling <tt>llvm.init.trampoline</tt>, the
8260 memory pointed to by <tt>tramp</tt> is modified, then the effect of any later call
8261 to the returned function pointer is undefined.</p>
8264 <!-- _______________________________________________________________________ -->
8267 '<tt>llvm.adjust.trampoline</tt>' Intrinsic
8275 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8279 <p>This performs any required machine-specific adjustment to the address of a
8280 trampoline (passed as <tt>tramp</tt>).</p>
8283 <p><tt>tramp</tt> must point to a block of memory which already has trampoline code
8284 filled in by a previous call to <a href="#int_it"><tt>llvm.init.trampoline</tt>
8288 <p>On some architectures the address of the code to be executed needs to be
8289 different to the address where the trampoline is actually stored. This
8290 intrinsic returns the executable address corresponding to <tt>tramp</tt>
8291 after performing the required machine specific adjustments.
8292 The pointer returned can then be <a href="#int_trampoline"> bitcast and
8300 <!-- ======================================================================= -->
8302 <a name="int_memorymarkers">Memory Use Markers</a>
8307 <p>This class of intrinsics exists to information about the lifetime of memory
8308 objects and ranges where variables are immutable.</p>
8310 <!-- _______________________________________________________________________ -->
8312 <a name="int_lifetime_start">'<tt>llvm.lifetime.start</tt>' Intrinsic</a>
8319 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8323 <p>The '<tt>llvm.lifetime.start</tt>' intrinsic specifies the start of a memory
8324 object's lifetime.</p>
8327 <p>The first argument is a constant integer representing the size of the
8328 object, or -1 if it is variable sized. The second argument is a pointer to
8332 <p>This intrinsic indicates that before this point in the code, the value of the
8333 memory pointed to by <tt>ptr</tt> is dead. This means that it is known to
8334 never be used and has an undefined value. A load from the pointer that
8335 precedes this intrinsic can be replaced with
8336 <tt>'<a href="#undefvalues">undef</a>'</tt>.</p>
8340 <!-- _______________________________________________________________________ -->
8342 <a name="int_lifetime_end">'<tt>llvm.lifetime.end</tt>' Intrinsic</a>
8349 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8353 <p>The '<tt>llvm.lifetime.end</tt>' intrinsic specifies the end of a memory
8354 object's lifetime.</p>
8357 <p>The first argument is a constant integer representing the size of the
8358 object, or -1 if it is variable sized. The second argument is a pointer to
8362 <p>This intrinsic indicates that after this point in the code, the value of the
8363 memory pointed to by <tt>ptr</tt> is dead. This means that it is known to
8364 never be used and has an undefined value. Any stores into the memory object
8365 following this intrinsic may be removed as dead.
8369 <!-- _______________________________________________________________________ -->
8371 <a name="int_invariant_start">'<tt>llvm.invariant.start</tt>' Intrinsic</a>
8378 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8382 <p>The '<tt>llvm.invariant.start</tt>' intrinsic specifies that the contents of
8383 a memory object will not change.</p>
8386 <p>The first argument is a constant integer representing the size of the
8387 object, or -1 if it is variable sized. The second argument is a pointer to
8391 <p>This intrinsic indicates that until an <tt>llvm.invariant.end</tt> that uses
8392 the return value, the referenced memory location is constant and
8397 <!-- _______________________________________________________________________ -->
8399 <a name="int_invariant_end">'<tt>llvm.invariant.end</tt>' Intrinsic</a>
8406 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8410 <p>The '<tt>llvm.invariant.end</tt>' intrinsic specifies that the contents of
8411 a memory object are mutable.</p>
8414 <p>The first argument is the matching <tt>llvm.invariant.start</tt> intrinsic.
8415 The second argument is a constant integer representing the size of the
8416 object, or -1 if it is variable sized and the third argument is a pointer
8420 <p>This intrinsic indicates that the memory is mutable again.</p>
8426 <!-- ======================================================================= -->
8428 <a name="int_general">General Intrinsics</a>
8433 <p>This class of intrinsics is designed to be generic and has no specific
8436 <!-- _______________________________________________________________________ -->
8438 <a name="int_var_annotation">'<tt>llvm.var.annotation</tt>' Intrinsic</a>
8445 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8449 <p>The '<tt>llvm.var.annotation</tt>' intrinsic.</p>
8452 <p>The first argument is a pointer to a value, the second is a pointer to a
8453 global string, the third is a pointer to a global string which is the source
8454 file name, and the last argument is the line number.</p>
8457 <p>This intrinsic allows annotation of local variables with arbitrary strings.
8458 This can be useful for special purpose optimizations that want to look for
8459 these annotations. These have no other defined use; they are ignored by code
8460 generation and optimization.</p>
8464 <!-- _______________________________________________________________________ -->
8466 <a name="int_annotation">'<tt>llvm.annotation.*</tt>' Intrinsic</a>
8472 <p>This is an overloaded intrinsic. You can use '<tt>llvm.annotation</tt>' on
8473 any integer bit width.</p>
8476 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8477 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8478 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8479 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8480 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8484 <p>The '<tt>llvm.annotation</tt>' intrinsic.</p>
8487 <p>The first argument is an integer value (result of some expression), the
8488 second is a pointer to a global string, the third is a pointer to a global
8489 string which is the source file name, and the last argument is the line
8490 number. It returns the value of the first argument.</p>
8493 <p>This intrinsic allows annotations to be put on arbitrary expressions with
8494 arbitrary strings. This can be useful for special purpose optimizations that
8495 want to look for these annotations. These have no other defined use; they
8496 are ignored by code generation and optimization.</p>
8500 <!-- _______________________________________________________________________ -->
8502 <a name="int_trap">'<tt>llvm.trap</tt>' Intrinsic</a>
8509 declare void @llvm.trap() noreturn nounwind
8513 <p>The '<tt>llvm.trap</tt>' intrinsic.</p>
8519 <p>This intrinsic is lowered to the target dependent trap instruction. If the
8520 target does not have a trap instruction, this intrinsic will be lowered to
8521 a call of the <tt>abort()</tt> function.</p>
8525 <!-- _______________________________________________________________________ -->
8527 <a name="int_debugtrap">'<tt>llvm.debugtrap</tt>' Intrinsic</a>
8534 declare void @llvm.debugtrap() nounwind
8538 <p>The '<tt>llvm.debugtrap</tt>' intrinsic.</p>
8544 <p>This intrinsic is lowered to code which is intended to cause an execution
8545 trap with the intention of requesting the attention of a debugger.</p>
8549 <!-- _______________________________________________________________________ -->
8551 <a name="int_stackprotector">'<tt>llvm.stackprotector</tt>' Intrinsic</a>
8558 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8562 <p>The <tt>llvm.stackprotector</tt> intrinsic takes the <tt>guard</tt> and
8563 stores it onto the stack at <tt>slot</tt>. The stack slot is adjusted to
8564 ensure that it is placed on the stack before local variables.</p>
8567 <p>The <tt>llvm.stackprotector</tt> intrinsic requires two pointer
8568 arguments. The first argument is the value loaded from the stack
8569 guard <tt>@__stack_chk_guard</tt>. The second variable is an <tt>alloca</tt>
8570 that has enough space to hold the value of the guard.</p>
8573 <p>This intrinsic causes the prologue/epilogue inserter to force the position of
8574 the <tt>AllocaInst</tt> stack slot to be before local variables on the
8575 stack. This is to ensure that if a local variable on the stack is
8576 overwritten, it will destroy the value of the guard. When the function exits,
8577 the guard on the stack is checked against the original guard. If they are
8578 different, then the program aborts by calling the <tt>__stack_chk_fail()</tt>
8583 <!-- _______________________________________________________________________ -->
8585 <a name="int_objectsize">'<tt>llvm.objectsize</tt>' Intrinsic</a>
8592 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8593 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8597 <p>The <tt>llvm.objectsize</tt> intrinsic is designed to provide information to
8598 the optimizers to determine at compile time whether a) an operation (like
8599 memcpy) will overflow a buffer that corresponds to an object, or b) that a
8600 runtime check for overflow isn't necessary. An object in this context means
8601 an allocation of a specific class, structure, array, or other object.</p>
8604 <p>The <tt>llvm.objectsize</tt> intrinsic takes two arguments. The first
8605 argument is a pointer to or into the <tt>object</tt>. The second argument
8606 is a boolean and determines whether <tt>llvm.objectsize</tt> returns 0 (if
8607 true) or -1 (if false) when the object size is unknown.
8608 The second argument only accepts constants.</p>
8611 <p>The <tt>llvm.objectsize</tt> intrinsic is lowered to a constant representing
8612 the size of the object concerned. If the size cannot be determined at compile
8613 time, <tt>llvm.objectsize</tt> returns <tt>i32/i64 -1 or 0</tt>
8614 (depending on the <tt>min</tt> argument).</p>
8617 <!-- _______________________________________________________________________ -->
8619 <a name="int_expect">'<tt>llvm.expect</tt>' Intrinsic</a>
8626 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8627 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8631 <p>The <tt>llvm.expect</tt> intrinsic provides information about expected (the
8632 most probable) value of <tt>val</tt>, which can be used by optimizers.</p>
8635 <p>The <tt>llvm.expect</tt> intrinsic takes two arguments. The first
8636 argument is a value. The second argument is an expected value, this needs to
8637 be a constant value, variables are not allowed.</p>
8640 <p>This intrinsic is lowered to the <tt>val</tt>.</p>
8646 <!-- *********************************************************************** -->
8649 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
8650 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
8651 <a href="http://validator.w3.org/check/referer"><img
8652 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
8654 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
8655 <a href="http://llvm.org/">The LLVM Compiler Infrastructure</a><br>
8656 Last modified: $Date$