cmd/link: rearrange compiler/linker generated data for clarity

[ianlancetaylor]

A Go executable contains information generated by the compiler and linker that is not program code or data. This data currently appears in various different locations in the executable. I think we should try to clarify it to make it easier for users to understand.

This information I am discussing is at least the following:

• buildinfo, which includes module dependencies and build settings. In ELF this is found in the .go.buildinfo section. It can be found without a section table by searching for a magic string, which is "\xff Go buildinf:". It contains a 32-byte header followed by a pair of strings. It's written out by (*Link).buildinfo in cmd/link/internal/ld/data.go. It's parsed by readRawBuildInfo in debug/buildinfo/buildinfo.go.
• The go:buildinfo.ref symbol is a pointer-sized symbol that refers to the buildinfo data, to keep the C linker from removing the .go.buildinfo section. This symbol appears in the .rodata section. It's written out by (*Link).buildinfo in cmd/link/internal/ld/data.go.
• moduledata, which has pointers to most of the other generated information. In ELF this is found at the symbol runtime.firstmoduledata (for an executable). There is no way to find it if the symbol table is missing. The format is simply the struct runtime.moduledata. It appears in the .noptrdata section. It is written out by (*Link).symtab in cmd/link/internal/ld/symtab.go.
• pcheader, which points to more of the generated information. The moduledata points to the pcheader. The format is the struct runtime.pcHeader. In ELF it appears at the start of the .gopclntab section. It starts with a four byte magic number 0xfffffff1. It is written out by (*pclntab).generatePCHeader in cmd/link/internal/ld/pcln.go.
• The function offset table, which has one entry for each function and method in the executable, plus one more entry that records the end of the last function. This can be found via a slice in moduledata. The format is runtime.functab, which is just a pair of uint32 offsets. The number of functions can also be found in the PC header, although as noted the slice has one more entry. This appears in the .gopclntab section. It is written out by writePCToFunc in cmd/link/internal/ld/pcln.go.
• The PC to function lookup table. This divides the address space devoted to functions into buckets, where each bucket covers 4096 bytes. See runtime.findfuncbucket for details. It can be found via a pointer in the moduledata. The size doesn't seem to be recorded, it must be computed based on minpc and maxpc in moduledata (or the addresses in the function offset table). It appears in the .rodata section. It is written out by (*pcln).findfunctab.
• The function table. This records information for each function. It can be found via the function offset table, or by the pclntable field in moduledata or the pclnOffset field in pcheader. The format starts with runtime._func, and is followed by variable length arrays containing pcdata and funcdata offsets. This appears in the .gopclntab section. It is written out by writeFuncs in cmd/link/internal/ld/pcln.go.
• The function name table. This is simply the name of each function as a series of NUL terminated strings. The function table has an offset into this table for each function. This can also be found as the funcnametab field of moduledata, and the funcnameOffset field of pcheader. It appears in the .gopclntab section. It is written out by (*pclntab).generateFuncnametab.
• The compilation unit table, which maps file numbers for a compilation unit to offsets into the file name table. See runtime.funcfile for how this is used. Basically it lets each function store small numbers when mapping from PC to file name. This can be found via the cutab field of moduledata or the cuOffset field of pcheader. The format is just a slice of uint32 values. It appears in the .gopclntab section. It is written out by (*pclntab).generateFilenameTabs.
• The file name table, which contains the actual file names, as a series of NUL terminated strings. Entries are found via the compilation unit table. This can also be found via the filetab field of moduledata or the filetabOffset field of pcheader. It appears in the .gopclntab section. It is written out by (*pclntab).generateFilenameTabs.
• The pcdata information, also known as pctab. The function table pcdata offsets point here, as does the function table pcsp field. The table can also be found via the pctab field of moduledata or the pctabOffset field of pcheader. The format of this table is complex and is described at http://go.dev/s/go12symtab in the PC-Value Table Encoding appendix. It appears in the .gopclntab section. The data is created by the compiler.
• The funcdata information. The function table funcdata offsets point here, and it can also be found via the gofunc field of moduledata. The format of this table varies depending on the exact funcdata information being recorded. It appears in the .rodata section (not the .gopclntab section). There doesn't seem to be a way to know the complete size of the data. The data is created by the compiler, though the linker seems to compute some of the inline symbol information.
• Type descriptors used for the reflect package. Besides direct references from code, this can be found via the typelinks field of moduledata, which indexes into the types (and etypes) pointers in moduledata. These appear in the .rodata section.
• The type links used by the reflect package to find type descriptors. This is a sequence of int32 offsets into the types section, sorted by type string. This appears in the .typelink section. It is written by (*Link).typelink.
• Type descriptors have sizes that depend on their kind. Many types also have a GC bitmask. The size of each GC bitmask depends on the type. These appear in the .rodata section.
• Larger types will compute the GC bitmask at runtime as needed. For these types, the program will contain a pointer that is filled out as needed (by runtime.getGCMaskOnDemand). These pointers appear in the .noptrdata section (though they could be .noptrbss; there is a TODO in dgcptrmaskOnDemand in cmd/compile/internal/reflectdata/reflect.go).
• The data and BSS sections have their own GC information. These can be found via the gcdata and gcbss fields of moduledata. They appear in the .rodata section. These are GC programs, not bitmasks; they are expanded into bitmasks by runtime.progToPointerMask.
• Interface tables computed by the compiler. These can be found via the itablink field of moduledata, which is a slice of pointers to runtime.itab structs. itablinks appears in the .itablink section. The itabs themselves appear in the .rodata section. The itabs are variable length, as they have a list of pointers to methods.
• FIPS checking information used by crypto/internal/fips140/check to verify the checksum of FIPS sections when in FIPS mode. This can be found at the go:fipsinfo symbol. It appears in the .go.fipsinfo section. Currently it is 120 bytes.
• Function descriptors are created whenever a top-level function or method expression is converted to a function type. A value of function type is a pointer to memory such that the first word of that memory is the address of the function code. Remaining words are closure or method pointers, but a top-level function or method expression doesn't have any of those. So a function descriptor is just a word of memory pointing to the function code. These appear in the .rodata section. They are created by WriteFuncSyms in cmd/compile/staticdata/data.go.

[gabyhelp]

Related Issues

• cmd/link: the value stored in pcHeader in `runtime.firstmoduledata` is different from `runtime.pclntab` when using external links on `macOS arm64 cgo` #69428 (closed)
• cmd/compile: move DWARF go.info.* symbol generation out of the linker #22094
• cmd/link: add a flag to the linker to do not write function names to runtime.pclntab #36555

Related Code Changes

• [dev.link] cmd/link: use generator symbols for the rest of pclntab
• cmd/compile/internal/gc: add position info to binary export data
• [dev.link] create runtime.funcnametab
• [dev.link] cmd/link: add runtime.pcheader
• [dev.link] cmd/{compile,link}: remove pcdata tables from pclntab_old

Related Discussions

• Do we document the history of .gosymtab and .gopclntab, and the differences with .symtab and .debug_line, .debug_frame?
• Presence of PCLN Table in CGo binaries

_{(Emoji vote if this was helpful or unhelpful; more detailed feedback welcome in this discussion.)}

[ianlancetaylor]

Today, when using ELF, when generating an ordinary executable, the Go linker generates the following Go-specific sections:

• .gopclntab. This holds the function names, the file names, the function table, the function lookup table, the filename offset table, and the pcdata.
• .gosymtab. This is always empty. The symbols .runtime.symtab and .runtime.esymtab point here; they always have the same value, and they don't point to anything meaninful. It's been a long time since anything appeared in this section, maybe before Go 1.
• .go.buildinfo. This holds the go:buildinfo symbol which holds the buildinfo data: the Go version and the module dependencies.
• .go.fipsinfo. This holds the go:fipsinfo symbol used to verify FIPS information in crypto/internal/fips140/check.
• .typelink. This holds the runtime.typelink symbol, which is a series of int32 offsets into the type section of the module. Note that although the field seems to consistently have type int32 the actual use converts to uintptr, so negative numbers are inadvisable.
• .itablink. This holds the runtime.itablink symbol, which is a series of pointers (not offsets) to itab tables.
• .noptrdata. This holds variables that can't contain pointers.
• .noptrbss. This holds zero-initialized variables that can't contain pointers.
• .note.go.buildid. This holds the value of the linker's -buildid option, which is normally computed by cmd/go. This is allocated in memory but I'm not sure why.

[ianlancetaylor]

The current Delve implementation looks for at least the .gopclntab and .noptrdata sections (among other sections not specific to Go). It seems to use the .notptrdata section to find the functab information.

[prattmic]

Thanks for writing all of this down!

This is relevant to #74396, where we've considered putting runtime.moduledata in its own section, or pointing to the moduledata from go:buildinfo, to make it easier to find in a stripped binary.

[ianlancetaylor]

I suggest that we do the following. I'll use ELF names for the sections, and the names can be tweaked as appropriate for other platforms.

• Keep the current .go.buildinfo section and go:buildinfo symbol unchanged, holding the buildinfo.
  • At some point I believe that we should add a pointer from buildinfo to moduledata, so that moduledata can be found without a symbol table.
  • At some point we can set the SHF_GNU_RETAIN section flag and drop go:buildinfo.ref.
• Keep the current .go.fipsinfo section and go:fipsinfo symbol unchanged.
• Keep the current .noptrdata and .noptrbss sections unchanged.
• See if we can make .note.go.buildid an unallocated section.
  • We can't, at least not easily. cmd/internal/buildid expects it to be near the start of the file.
• Remove the .gosymtab section.
  • Done in https://go.dev/cl/717200.
• Keep the .gopclntab section, and keep it starting with the pcHeader information, as Delve uses it.
  • Move the PC to function lookup table into .gopclntab.
    • Done in https://go.dev/cl/719743.
  • Move the funcdata information into .gopclntab.
    • Done in https://go.dev/cl/719440.
  • Add an offset to the funcdata at the end of the current pcHeader.
    • Postponing until something would use this.
• Add a new .go.module section in the data segment.
  • Move the moduledata information there.
    • Done in https://go.dev/cl/720660.
  • Longer term, we should move this to the rodata segment. It is only in the data segment because the next field is updated at runtime as shared libraries and plugins are loaded. We can split that out.
• Add a new .gotype section in the read-only data segment.
  • Move all type descriptors to this section.
  • Move typelinks to this section; remove the .typelink section.
  • Move the itabs to this section.
  • Move itablinks to this section; remove the .itablink section.
• Add a new .gogc section in the read-only data segment.
  • Move all type GC bitmasks to this section.
  • Move the global data and global BSS GC programs to this section.
• Add a new .gogccbss section to BSS.
  • Store the GC program data words there.
  • Or just use .noptrbss?
• Perhaps add a new .gofunction section to hold function descriptors.
  • I don't know whether this is worth it.

CC @golang/compiler @golang/runtime

[ianlancetaylor]

I took a look at relro sections. relro means relocatable read-only. A relro section is initially loaded as a writable section. After dynamic relocations are applied, the section is changed to be read-only. This is a security measure. It permits variables to contain pointers to other variables whose address is not known at link time (perhaps because the code is going into a position independent executable (PIE) or a shared library), and for those variables to then become immutable after they have been initialized at program load time. In general this concept does not apply to Go variables, which are never immutable, but it does apply to some of the compiler/linker generated data.

As a side note, as a lot of compiler/linker generated data uses offsets, relro must be used with care when using external linking. Most object file format don't provide relocations for the necessary offsets, so they must be computed by cmd/link. That means that offsets should not cross sections; if they do, the external linker may rearrange the sections such that the cmd/link computed offsets are incorrect.

Specifically, cmd/link currently puts the following compiler/linker generated data into relro sections when generating a PIE or shared library or plugin or a c-archive:

• type descriptors
  • Type descriptors contain several pointers:
    • GC bitmaps
    • Equality function
    • Pointers to type descriptors for element types, such as slice elements, map keys, etc.
    • Interface methods
    • Struct fields
• type links
  • These no longer contain any pointers, just offsets from the start of the type descriptors.
  • So these don't need to be relro.
• itab links
  • These currently contain pointers to itabs.
• the entire pclntab section
  • Currently I don't think this needs to be relro, but further investigation is needed.
• function descriptors
  • These contain pointers to function code.

[ianlancetaylor]

A quick rundown of what is stored in pcdata and funcdata. Currently pcdata appears in the .gopclntab section, and funcdata appears in the .rodata section. I am suggesting that funcdata move to .gopclntab.

Recall that pcdata points to tables that map from PC to value as described at https://go.dev/s/go12symtab. In pcdata we currently find:

• The _func.pcsp field points to a mapping from PC to the offset between the stack pointer and the frame pointer.
• The _func.pcfile field points to a mapping from PC to the current file number. This is added to _func.cuOffset to get an offset into cutab, which provides an offset into the file name table.
• The _func.pcln field points to a mapping from PC to line number.
• Optional pcdata field 0 is UnsafePoint, a mapping from PC to whether that PC is a safe preemption point. This valid values appear in internal/abi/symtab.go (UnsafePointSafe and so forth).
• Optional pcdata field 1 is StackMapIndex, a mapping from PC to an index into the funcdata entries LocalsPointersMaps and ArgsPointerMaps.
• Optional pcdata field 2 is InlTreeIndex, a mapping from PC to an index into the funcdata entry InlTree.
• Optional pcdata field 3 is ArgLiveIndex, a mapping from PC to an index into the funcdata entry ArgLiveInfo.
• Optional pcdata field 4 is PanicBounds, a mapping from PC to a code that is a compressed record of an out of of range slice or index operation. See internal/abi/bounds.go or https://go.dev/cl/682396 for details.

The funcdata information is always optional. It is not a mapping from PC, it's just any random data.

• 0 ArgsPointerMaps: A stackmap of type runtime.stackmap: number of bitmaps, number of bits in a bitmap, and then that many bitmaps of that size. Indexed by pcdata StackmapIndex. This holds a bitmap of pointers in the arguments, used by the GC when walking the stack. It is also used when copying the stack.
• 1 LocalsPointerMaps: Just like ArgsPointerMaps, but for local variables.
• 2 StackObjects: A number of entries, followed by that many entries of type runtime.stackObjectRecord. These record a GC mask for stack objects, so that the GC can ignore stack objects that nothing refers to. The compiler emits this information for variables on the stack whose address is taken. This is separate from stack liveness in that it refers to variables that may or may not be live. See cmd/compile: compiler can unexpectedly preserve memory #22350.
• 3 InlTree: A list of type runtime.inlinedCall. The number of entries is not specified. This is indexed by pcdata InlTreeIndex. This stores information about inlined calls, used when unwinding the stack.
• 4 OpenCodedDeferInfo: a pair of varints holding the stack offsets to two variables used to run open coded defer statements.
• 5 ArgInfo: a sequence of bytes used to print function arguments in a stack traceback. See TraceArgsEndSeq and friends in internal/abi/type.go.
• 6 ArgLiveInfo: A one byte stack offset followed by a bitmap of which register arguments are live. The bitmap is indexed by both pcdata ArgLiveIndex and a slot index from in the funcdata ArgInfo. See cmd/compile/internal/liveness/arg.go for details.
• 7 WrapInfo: A uint32 text offset that is present for wrapper functions. The text offset points to the wrapped function. See cmd/trace: explicit goroutine function names replaced by synthetic ones #50622.

It's worth noting that none of the pcdata or funcdata fields contain pointers. Some of the funcdata fields contain offsets.

As far as I can tell the only pointer in all of pclntab is that the pcheader holds the address of the text section. Interestingly, Delve does not use that information, with the comment "use the start PC instead of reading from the table, which may be unrelocated". And the runtime doesn't use it either, preferring the moduledata text field. Getting rid of that pointer will let us avoid moving .gopclntab into relro.