1=============== 2LLVM Extensions 3=============== 4 5.. contents:: 6 :local: 7 8.. toctree:: 9 :hidden: 10 11Introduction 12============ 13 14This document describes extensions to tools and formats LLVM seeks compatibility 15with. 16 17General Assembly Syntax 18=========================== 19 20C99-style Hexadecimal Floating-point Constants 21---------------------------------------------- 22 23LLVM's assemblers allow floating-point constants to be written in C99's 24hexadecimal format instead of decimal if desired. 25 26.. code-block:: gas 27 28 .section .data 29 .float 0x1c2.2ap3 30 31Machine-specific Assembly Syntax 32================================ 33 34X86/COFF-Dependent 35------------------ 36 37Relocations 38^^^^^^^^^^^ 39 40The following additional relocation types are supported: 41 42**@IMGREL** (AT&T syntax only) generates an image-relative relocation that 43corresponds to the COFF relocation types ``IMAGE_REL_I386_DIR32NB`` (32-bit) or 44``IMAGE_REL_AMD64_ADDR32NB`` (64-bit). 45 46.. code-block:: text 47 48 .text 49 fun: 50 mov foo@IMGREL(%ebx, %ecx, 4), %eax 51 52 .section .pdata 53 .long fun@IMGREL 54 .long (fun@imgrel + 0x3F) 55 .long $unwind$fun@imgrel 56 57**.secrel32** generates a relocation that corresponds to the COFF relocation 58types ``IMAGE_REL_I386_SECREL`` (32-bit) or ``IMAGE_REL_AMD64_SECREL`` (64-bit). 59 60**.secidx** relocation generates an index of the section that contains 61the target. It corresponds to the COFF relocation types 62``IMAGE_REL_I386_SECTION`` (32-bit) or ``IMAGE_REL_AMD64_SECTION`` (64-bit). 63 64.. code-block:: none 65 66 .section .debug$S,"rn" 67 .long 4 68 .long 242 69 .long 40 70 .secrel32 _function_name + 0 71 .secidx _function_name 72 ... 73 74``.linkonce`` Directive 75^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 76 77Syntax: 78 79 ``.linkonce [ comdat type ]`` 80 81Supported COMDAT types: 82 83``discard`` 84 Discards duplicate sections with the same COMDAT symbol. This is the default 85 if no type is specified. 86 87``one_only`` 88 If the symbol is defined multiple times, the linker issues an error. 89 90``same_size`` 91 Duplicates are discarded, but the linker issues an error if any have 92 different sizes. 93 94``same_contents`` 95 Duplicates are discarded, but the linker issues an error if any duplicates 96 do not have exactly the same content. 97 98``largest`` 99 Links the largest section from among the duplicates. 100 101``newest`` 102 Links the newest section from among the duplicates. 103 104 105.. code-block:: gas 106 107 .section .text$foo 108 .linkonce 109 ... 110 111``.section`` Directive 112^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 113 114MC supports passing the information in ``.linkonce`` at the end of 115``.section``. For example, these two codes are equivalent 116 117.. code-block:: gas 118 119 .section secName, "dr", discard, "Symbol1" 120 .globl Symbol1 121 Symbol1: 122 .long 1 123 124.. code-block:: gas 125 126 .section secName, "dr" 127 .linkonce discard 128 .globl Symbol1 129 Symbol1: 130 .long 1 131 132Note that in the combined form the COMDAT symbol is explicit. This 133extension exists to support multiple sections with the same name in 134different COMDATs: 135 136 137.. code-block:: gas 138 139 .section secName, "dr", discard, "Symbol1" 140 .globl Symbol1 141 Symbol1: 142 .long 1 143 144 .section secName, "dr", discard, "Symbol2" 145 .globl Symbol2 146 Symbol2: 147 .long 1 148 149In addition to the types allowed with ``.linkonce``, ``.section`` also accepts 150``associative``. The meaning is that the section is linked if a certain other 151COMDAT section is linked. This other section is indicated by the comdat symbol 152in this directive. It can be any symbol defined in the associated section, but 153is usually the associated section's comdat. 154 155 The following restrictions apply to the associated section: 156 157 1. It must be a COMDAT section. 158 2. It cannot be another associative COMDAT section. 159 160In the following example the symbol ``sym`` is the comdat symbol of ``.foo`` 161and ``.bar`` is associated to ``.foo``. 162 163.. code-block:: gas 164 165 .section .foo,"bw",discard, "sym" 166 .section .bar,"rd",associative, "sym" 167 168MC supports these flags in the COFF ``.section`` directive: 169 170 - ``b``: BSS section (``IMAGE_SCN_CNT_INITIALIZED_DATA``) 171 - ``d``: Data section (``IMAGE_SCN_CNT_UNINITIALIZED_DATA``) 172 - ``n``: Section is not loaded (``IMAGE_SCN_LNK_REMOVE``) 173 - ``r``: Read-only 174 - ``s``: Shared section 175 - ``w``: Writable 176 - ``x``: Executable section 177 - ``y``: Not readable 178 - ``D``: Discardable (``IMAGE_SCN_MEM_DISCARDABLE``) 179 180These flags are all compatible with gas, with the exception of the ``D`` flag, 181which gnu as does not support. For gas compatibility, sections with a name 182starting with ".debug" are implicitly discardable. 183 184 185ARM64/COFF-Dependent 186-------------------- 187 188Relocations 189^^^^^^^^^^^ 190 191The following additional symbol variants are supported: 192 193**:secrel_lo12:** generates a relocation that corresponds to the COFF relocation 194types ``IMAGE_REL_ARM64_SECREL_LOW12A`` or ``IMAGE_REL_ARM64_SECREL_LOW12L``. 195 196**:secrel_hi12:** generates a relocation that corresponds to the COFF relocation 197type ``IMAGE_REL_ARM64_SECREL_HIGH12A``. 198 199.. code-block:: gas 200 201 add x0, x0, :secrel_hi12:symbol 202 ldr x0, [x0, :secrel_lo12:symbol] 203 204 add x1, x1, :secrel_hi12:symbol 205 add x1, x1, :secrel_lo12:symbol 206 ... 207 208 209ELF-Dependent 210------------- 211 212``.section`` Directive 213^^^^^^^^^^^^^^^^^^^^^^ 214 215In order to support creating multiple sections with the same name and comdat, 216it is possible to add an unique number at the end of the ``.section`` directive. 217For example, the following code creates two sections named ``.text``. 218 219.. code-block:: gas 220 221 .section .text,"ax",@progbits,unique,1 222 nop 223 224 .section .text,"ax",@progbits,unique,2 225 nop 226 227 228The unique number is not present in the resulting object at all. It is just used 229in the assembler to differentiate the sections. 230 231The 'o' flag is mapped to SHF_LINK_ORDER. If it is present, a symbol 232must be given that identifies the section to be placed is the 233.sh_link. 234 235.. code-block:: gas 236 237 .section .foo,"a",@progbits 238 .Ltmp: 239 .section .bar,"ao",@progbits,.Ltmp 240 241which is equivalent to just 242 243.. code-block:: gas 244 245 .section .foo,"a",@progbits 246 .section .bar,"ao",@progbits,.foo 247 248``.linker-options`` Section (linker options) 249^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 250 251In order to support passing linker options from the frontend to the linker, a 252special section of type ``SHT_LLVM_LINKER_OPTIONS`` (usually named 253``.linker-options`` though the name is not significant as it is identified by 254the type). The contents of this section is a simple pair-wise encoding of 255directives for consideration by the linker. The strings are encoded as standard 256null-terminated UTF-8 strings. They are emitted inline to avoid having the 257linker traverse the object file for retrieving the value. The linker is 258permitted to not honour the option and instead provide a warning/error to the 259user that the requested option was not honoured. 260 261The section has type ``SHT_LLVM_LINKER_OPTIONS`` and has the ``SHF_EXCLUDE`` 262flag to ensure that the section is treated as opaque by linkers which do not 263support the feature and will not be emitted into the final linked binary. 264 265This would be equivalent to the follow raw assembly: 266 267.. code-block:: gas 268 269 .section ".linker-options","e",@llvm_linker_options 270 .asciz "option 1" 271 .asciz "value 1" 272 .asciz "option 2" 273 .asciz "value 2" 274 275The following directives are specified: 276 277 - lib 278 279 The parameter identifies a library to be linked against. The library will 280 be looked up in the default and any specified library search paths 281 (specified to this point). 282 283 - libpath 284 285 The parameter identifies an additional library search path to be considered 286 when looking up libraries after the inclusion of this option. 287 288``SHT_LLVM_DEPENDENT_LIBRARIES`` Section (Dependent Libraries) 289^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 290 291This section contains strings specifying libraries to be added to the link by 292the linker. 293 294The section should be consumed by the linker and not written to the output. 295 296The strings are encoded as standard null-terminated UTF-8 strings. 297 298For example: 299 300.. code-block:: gas 301 302 .section ".deplibs","MS",@llvm_dependent_libraries,1 303 .asciz "library specifier 1" 304 .asciz "library specifier 2" 305 306The interpretation of the library specifiers is defined by the consuming linker. 307 308``SHT_LLVM_CALL_GRAPH_PROFILE`` Section (Call Graph Profile) 309^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 310 311This section is used to pass a call graph profile to the linker which can be 312used to optimize the placement of sections. It contains a sequence of 313(from symbol, to symbol, weight) tuples. 314 315It shall have a type of ``SHT_LLVM_CALL_GRAPH_PROFILE`` (0x6fff4c02), shall 316have the ``SHF_EXCLUDE`` flag set, the ``sh_link`` member shall hold the section 317header index of the associated symbol table, and shall have a ``sh_entsize`` of 31816. It should be named ``.llvm.call-graph-profile``. 319 320The contents of the section shall be a sequence of ``Elf_CGProfile`` entries. 321 322.. code-block:: c 323 324 typedef struct { 325 Elf_Word cgp_from; 326 Elf_Word cgp_to; 327 Elf_Xword cgp_weight; 328 } Elf_CGProfile; 329 330cgp_from 331 The symbol index of the source of the edge. 332 333cgp_to 334 The symbol index of the destination of the edge. 335 336cgp_weight 337 The weight of the edge. 338 339This is represented in assembly as: 340 341.. code-block:: gas 342 343 .cg_profile from, to, 42 344 345``.cg_profile`` directives are processed at the end of the file. It is an error 346if either ``from`` or ``to`` are undefined temporary symbols. If either symbol 347is a temporary symbol, then the section symbol is used instead. If either 348symbol is undefined, then that symbol is defined as if ``.weak symbol`` has been 349written at the end of the file. This forces the symbol to show up in the symbol 350table. 351 352``SHT_LLVM_ADDRSIG`` Section (address-significance table) 353^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 354 355This section is used to mark symbols as address-significant, i.e. the address 356of the symbol is used in a comparison or leaks outside the translation unit. It 357has the same meaning as the absence of the LLVM attributes ``unnamed_addr`` 358and ``local_unnamed_addr``. 359 360Any sections referred to by symbols that are not marked as address-significant 361in any object file may be safely merged by a linker without breaking the 362address uniqueness guarantee provided by the C and C++ language standards. 363 364The contents of the section are a sequence of ULEB128-encoded integers 365referring to the symbol table indexes of the address-significant symbols. 366 367There are two associated assembly directives: 368 369.. code-block:: gas 370 371 .addrsig 372 373This instructs the assembler to emit an address-significance table. Without 374this directive, all symbols are considered address-significant. 375 376.. code-block:: gas 377 378 .addrsig_sym sym 379 380This marks ``sym`` as address-significant. 381 382``SHT_LLVM_SYMPART`` Section (symbol partition specification) 383^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 384 385This section is used to mark symbols with the `partition`_ that they 386belong to. An ``.llvm_sympart`` section consists of a null-terminated string 387specifying the name of the partition followed by a relocation referring to 388the symbol that belongs to the partition. It may be constructed as follows: 389 390.. code-block:: gas 391 392 .section ".llvm_sympart","",@llvm_sympart 393 .asciz "libpartition.so" 394 .word symbol_in_partition 395 396.. _partition: https://lld.llvm.org/Partitions.html 397 398``SHT_LLVM_BB_ADDR_MAP`` Section (basic block address map) 399^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 400This section stores the binary address of basic blocks along with other related 401metadata. This information can be used to map binary profiles (like perf 402profiles) directly to machine basic blocks. 403This section is emitted with ``-basic-block-sections=labels`` and will contain 404a BB address map table for every function. 405 406The ``SHT_LLVM_BB_ADDR_MAP`` type provides backward compatibility to allow 407reading older versions of the BB address map generated by older compilers. Each 408function entry starts with a version byte which specifies the encoding version 409to use. The following versioning schemes are currently supported. 410 411Version 1 (newest): basic block address offsets are computed relative to the end 412of previous blocks. 413 414Example: 415 416.. code-block:: gas 417 418 .section ".llvm_bb_addr_map","",@llvm_bb_addr_map 419 .byte 1 # version number 420 .byte 0 # feature byte (reserved for future use) 421 .quad .Lfunc_begin0 # address of the function 422 .byte 2 # number of basic blocks 423 # BB record for BB_0 424 .uleb128 .Lfunc_beign0-.Lfunc_begin0 # BB_0 offset relative to function entry (always zero) 425 .uleb128 .LBB_END0_0-.Lfunc_begin0 # BB_0 size 426 .byte x # BB_0 metadata 427 # BB record for BB_1 428 .uleb128 .LBB0_1-.LBB_END0_0 # BB_1 offset relative to the end of last block (BB_0). 429 .uleb128 .LBB_END0_1-.LBB0_1 # BB_1 size 430 .byte y # BB_1 metadata 431 432Version 0: basic block address offsets are computed relative to the function 433address. This uses the unversioned ``SHT_LLVM_BB_ADDR_MAP_V0`` section type and 434is semantically equivalent to using ``SHT_LLVM_BB_ADDR_MAP`` with a zero 435version field. 436 437Example: 438 439.. code-block:: gas 440 441 .section ".llvm_bb_addr_map","",@llvm_bb_addr_map_v0 442 .quad .Lfunc_begin0 # address of the function 443 .byte 2 # number of basic blocks 444 # BB record for BB_0 445 .uleb128 .Lfunc_beign0-.Lfunc_begin0 # BB_0 offset relative to the function entry (always zero) 446 .uleb128 .LBB_END0_0-.Lfunc_begin0 # BB_0 size 447 .byte x # BB_0 metadata 448 # BB record for BB_1 449 .uleb128 .LBB0_1-.Lfunc_begin0 # BB_1 offset relative to the function entry 450 .uleb128 .LBB_END0_1-.LBB0_1 # BB_1 size 451 .byte y # BB_1 metadata 452 453``SHT_LLVM_OFFLOADING`` Section (offloading data) 454^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 455This section stores the binary data used to perform offloading device linking 456and execution, creating a fat binary. This section is emitted during compilation 457of offloading languages such as OpenMP or CUDA. If the data is intended to be 458used by the device linker only, it should use the ``SHF_EXCLUDE`` flag so it is 459automatically stripped from the final executable or shared library. 460 461The binary data stored in this section conforms to a custom binary format used 462for storing offloading metadata. This format is effectively a string table 463containing metadata accompanied by a device image. 464 465CodeView-Dependent 466------------------ 467 468``.cv_file`` Directive 469^^^^^^^^^^^^^^^^^^^^^^ 470Syntax: 471 ``.cv_file`` *FileNumber FileName* [ *checksum* ] [ *checksumkind* ] 472 473``.cv_func_id`` Directive 474^^^^^^^^^^^^^^^^^^^^^^^^^ 475Introduces a function ID that can be used with ``.cv_loc``. 476 477Syntax: 478 ``.cv_func_id`` *FunctionId* 479 480``.cv_inline_site_id`` Directive 481^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 482Introduces a function ID that can be used with ``.cv_loc``. Includes 483``inlined at`` source location information for use in the line table of the 484caller, whether the caller is a real function or another inlined call site. 485 486Syntax: 487 ``.cv_inline_site_id`` *FunctionId* ``within`` *Function* ``inlined_at`` *FileNumber Line* [ *Column* ] 488 489``.cv_loc`` Directive 490^^^^^^^^^^^^^^^^^^^^^ 491The first number is a file number, must have been previously assigned with a 492``.file`` directive, the second number is the line number and optionally the 493third number is a column position (zero if not specified). The remaining 494optional items are ``.loc`` sub-directives. 495 496Syntax: 497 ``.cv_loc`` *FunctionId FileNumber* [ *Line* ] [ *Column* ] [ *prologue_end* ] [ ``is_stmt`` *value* ] 498 499``.cv_linetable`` Directive 500^^^^^^^^^^^^^^^^^^^^^^^^^^^ 501Syntax: 502 ``.cv_linetable`` *FunctionId* ``,`` *FunctionStart* ``,`` *FunctionEnd* 503 504``.cv_inline_linetable`` Directive 505^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 506Syntax: 507 ``.cv_inline_linetable`` *PrimaryFunctionId* ``,`` *FileNumber Line FunctionStart FunctionEnd* 508 509``.cv_def_range`` Directive 510^^^^^^^^^^^^^^^^^^^^^^^^^^^ 511The *GapStart* and *GapEnd* options may be repeated as needed. 512 513Syntax: 514 ``.cv_def_range`` *RangeStart RangeEnd* [ *GapStart GapEnd* ] ``,`` *bytes* 515 516``.cv_stringtable`` Directive 517^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 518 519``.cv_filechecksums`` Directive 520^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 521 522``.cv_filechecksumoffset`` Directive 523^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 524Syntax: 525 ``.cv_filechecksumoffset`` *FileNumber* 526 527``.cv_fpo_data`` Directive 528^^^^^^^^^^^^^^^^^^^^^^^^^^ 529Syntax: 530 ``.cv_fpo_data`` *procsym* 531 532Target Specific Behaviour 533========================= 534 535X86 536--- 537 538Relocations 539^^^^^^^^^^^ 540 541**@ABS8** can be applied to symbols which appear as immediate operands to 542instructions that have an 8-bit immediate form for that operand. It causes 543the assembler to use the 8-bit form and an 8-bit relocation (e.g. ``R_386_8`` 544or ``R_X86_64_8``) for the symbol. 545 546For example: 547 548.. code-block:: gas 549 550 cmpq $foo@ABS8, %rdi 551 552This causes the assembler to select the form of the 64-bit ``cmpq`` instruction 553that takes an 8-bit immediate operand that is sign extended to 64 bits, as 554opposed to ``cmpq $foo, %rdi`` which takes a 32-bit immediate operand. This 555is also not the same as ``cmpb $foo, %dil``, which is an 8-bit comparison. 556 557 558**@GOTPCREL_NORELAX** can be used in place of ``@GOTPCREL`` to guarantee that 559the assembler emits an ``R_X86_64_GOTPCREL`` relocation instead of a relaxable 560``R_X86_64[_REX]_GOTPCRELX`` relocation. 561 562Windows on ARM 563-------------- 564 565Stack Probe Emission 566^^^^^^^^^^^^^^^^^^^^ 567 568The reference implementation (Microsoft Visual Studio 2012) emits stack probes 569in the following fashion: 570 571.. code-block:: gas 572 573 movw r4, #constant 574 bl __chkstk 575 sub.w sp, sp, r4 576 577However, this has the limitation of 32 MiB (±16MiB). In order to accommodate 578larger binaries, LLVM supports the use of ``-mcmodel=large`` to allow a 4GiB 579range via a slight deviation. It will generate an indirect jump as follows: 580 581.. code-block:: gas 582 583 movw r4, #constant 584 movw r12, :lower16:__chkstk 585 movt r12, :upper16:__chkstk 586 blx r12 587 sub.w sp, sp, r4 588 589Variable Length Arrays 590^^^^^^^^^^^^^^^^^^^^^^ 591 592The reference implementation (Microsoft Visual Studio 2012) does not permit the 593emission of Variable Length Arrays (VLAs). 594 595The Windows ARM Itanium ABI extends the base ABI by adding support for emitting 596a dynamic stack allocation. When emitting a variable stack allocation, a call 597to ``__chkstk`` is emitted unconditionally to ensure that guard pages are setup 598properly. The emission of this stack probe emission is handled similar to the 599standard stack probe emission. 600 601The MSVC environment does not emit code for VLAs currently. 602 603Windows on ARM64 604---------------- 605 606Stack Probe Emission 607^^^^^^^^^^^^^^^^^^^^ 608 609The reference implementation (Microsoft Visual Studio 2017) emits stack probes 610in the following fashion: 611 612.. code-block:: gas 613 614 mov x15, #constant 615 bl __chkstk 616 sub sp, sp, x15, lsl #4 617 618However, this has the limitation of 256 MiB (±128MiB). In order to accommodate 619larger binaries, LLVM supports the use of ``-mcmodel=large`` to allow a 8GiB 620(±4GiB) range via a slight deviation. It will generate an indirect jump as 621follows: 622 623.. code-block:: gas 624 625 mov x15, #constant 626 adrp x16, __chkstk 627 add x16, x16, :lo12:__chkstk 628 blr x16 629 sub sp, sp, x15, lsl #4 630 631