BC file format: metadata and summaries

Metadata carries debug information and other data attached to a module. LLVM’s Bitcode File Format document leaves the metadata blocks as placeholders; the codes below are those in LLVM’s LLVMBitCodes.h with the operand lists from its comments, and the layouts of the strings, attachment and kind records from LLVM’s writer (see sources). Operand lists that end in “…” continue with fields that only the reader and writer source define; they are not listed here.

METADATA block (15)

Code	Record	Operands (header comment)
`1`	`STRING_OLD`	[values]
`2`	`VALUE`	[type num, value num]
`3`	`NODE`	[n × md num]
`4`	`NAME`	[values]
`5`	`DISTINCT_NODE`	[n × md num]
`6`	`KIND`	[n × [id, name]]
`7`	`LOCATION`	[distinct, line, col, scope, inlined-at?]
`8`	`OLD_NODE`	[n × (type num, value num)]
`9`	`OLD_FN_NODE`	[n × (type num, value num)]
`10`	`NAMED_NODE`	[n × mdnodes]
`11`	`ATTACHMENT`	[m × [value, [n × [id, mdnode]]]]
`12`	`GENERIC_DEBUG`	[distinct, tag, vers, header, n × md num]
`13`	`SUBRANGE`	[distinct, count, lo]
`14`	`ENUMERATOR`	[isUnsigned\|distinct, value, name]
`15`	`BASIC_TYPE`	[distinct, tag, name, size, align, enc]
`16`	`FILE`	[distinct, filename, directory, checksumkind, checksum]
`17`	`DERIVED_TYPE`	[distinct, …]
`18`	`COMPOSITE_TYPE`	[distinct, …]
`19`	`SUBROUTINE_TYPE`	[distinct, flags, types, cc]
`20`	`COMPILE_UNIT`	[distinct, …]
`21`	`SUBPROGRAM`	[distinct, …]
`22`	`LEXICAL_BLOCK`	[distinct, scope, file, line, column]
`23`	`LEXICAL_BLOCK_FILE`	[distinct, scope, file, discriminator]
`24`	`NAMESPACE`	[distinct, scope, file, name, line, exportSymbols]
`25`	`TEMPLATE_TYPE`	[distinct, scope, name, type, …]
`26`	`TEMPLATE_VALUE`	[distinct, scope, name, type, value, …]
`27`	`GLOBAL_VAR`	[distinct, …]
`28`	`LOCAL_VAR`	[distinct, …]
`29`	`EXPRESSION`	[distinct, n × element]
`30`	`OBJC_PROPERTY`	[distinct, name, file, line, …]
`31`	`IMPORTED_ENTITY`	[distinct, tag, scope, entity, line, name]
`32`	`MODULE`	[distinct, scope, name, …]
`33`	`MACRO`	[distinct, macinfo, line, name, value]
`34`	`MACRO_FILE`	[distinct, macinfo, line, file, …]
`35`	`STRINGS`	[count, offset] blob([lengths][chars]); below
`36`	`GLOBAL_DECL_ATTACHMENT`	[valueid, n × [id, mdnode]]
`37`	`GLOBAL_VAR_EXPR`	[distinct, var, expr]
`38`	`INDEX_OFFSET`	[offset]
`39`	`INDEX`	[bitpos]
`40`	`LABEL`	[distinct, scope, name, file, line]
`41`	`STRING_TYPE`	[distinct, name, size, align, …]
`42`, `43`	—	reserved for Fortran array debug information
`44`	`COMMON_BLOCK`	[distinct, scope, name, variable, …]
`45`	`GENERIC_SUBRANGE`	[distinct, count, lo, up, stride]
`46`	`ARG_LIST`	[n × [type num, value num]]
`47`	`ASSIGN_ID`	[distinct, …]
`48`	`SUBRANGE_TYPE`	[distinct, …]
`49`	`FIXED_POINT_TYPE`	[distinct, …]
`50`	`PROPERTY`	[distinct, name, file, line, type, backing_storage]
`51`	`LAYERLOC`	[distinct, line, column, file, kind]; recent, not documented
`52`	`LAYERLOCLIST`	[distinct, n × layerloc]; recent, not documented

The STRINGS record

As written by LLVM: [count, offset] and a Blob. The blob starts with count string lengths, each written as a VBR6 in bitstream form, padded to a 32-bit boundary; offset is the position in the blob where the characters start, and the characters of all the strings follow one after another.

METADATA_ATTACHMENT block (16)

Inside a FUNCTION block. As written by LLVM, the function’s own attachments come first in an ATTACHMENT record (code 11) of (kind, metadata) pairs with no instruction number; then one ATTACHMENT record per instruction that has attachments, starting with the instruction’s number and followed by its pairs. Debug locations are not written here (they use DEBUG_LOC records in the function block).

METADATA_KIND block (22)

One KIND record (code 6) per metadata kind name: the kind ID, then the name’s characters. The writer leaves the block out when there are no names.

Summary blocks (19, 20, 24)

GLOBALVAL_SUMMARY (20), FULL_LTO_GLOBALVAL_SUMMARY (24) and MODULE_STRTAB (19) carry the summaries used for link-time optimisation. Record codes of the summary blocks:

Code	Record	Code	Record
`1`	`FS_PERMODULE`	`18`	`FS_CFI_FUNCTION_DECLS`
`2`	`FS_PERMODULE_PROFILE`	`19`	`FS_PERMODULE_RELBF`
`3`	`FS_PERMODULE_GLOBALVAR_INIT_REFS`	`20`	`FS_FLAGS`
`4`	`FS_COMBINED`	`21`	`FS_TYPE_ID`
`5`	`FS_COMBINED_PROFILE`	`22`	`FS_TYPE_ID_METADATA`
`6`	`FS_COMBINED_GLOBALVAR_INIT_REFS`	`23`	`FS_PERMODULE_VTABLE_GLOBALVAR_INIT_REFS`
`7`	`FS_ALIAS`	`24`	`FS_BLOCK_COUNT`
`8`	`FS_COMBINED_ALIAS`	`25`	`FS_PARAM_ACCESS`
`9`	`FS_COMBINED_ORIGINAL_NAME`	`26`	`FS_PERMODULE_CALLSITE_INFO`
`10`	`FS_VERSION`	`27`	`FS_PERMODULE_ALLOC_INFO`
`11`	`FS_TYPE_TESTS`	`28`	`FS_COMBINED_CALLSITE_INFO`
`12`	`FS_TYPE_TEST_ASSUME_VCALLS`	`29`	`FS_COMBINED_ALLOC_INFO`
`13`	`FS_TYPE_CHECKED_LOAD_VCALLS`	`30`	`FS_STACK_IDS`
`14`	`FS_TYPE_TEST_ASSUME_CONST_VCALL`	`31`	`FS_ALLOC_CONTEXT_IDS`
`15`	`FS_TYPE_CHECKED_LOAD_CONST_VCALL`	`32`	`FS_CONTEXT_RADIX_TREE_ARRAY`
`16`	`FS_VALUE_GUID`	`33`	`FS_COMBINED_ALLOC_INFO_NO_CONTEXT`
`17`	`FS_CFI_FUNCTION_DEFS`

The operands of these records are not described in the sources read and are not listed.

MODULE_STRTAB block (19)

Code	Record	Operands
`1`	`MST_ENTRY`	[modid, namechar…]
`2`	`MST_HASH`	[5 × i32]

Sources

LLVMBitCodes.h – MetadataCodes, GlobalValueSummarySymtabCodes, ModulePathSymtabCodes, with their comments (main branch, November 2026)
BitcodeWriter.cpp – the metadata strings blob, the attachment records and the kind records
LLVM Bitcode File Format – METADATA_BLOCK and METADATA_ATTACHMENT (placeholders)

.BC File Format – Metadata and summaries