ObjFile::parse combines symbol initialization and resolution. Many tasks
unrelated to symbol resolution can be postponed and parallelized. This patch
extracts local symbol initialization and parallelizes it.
Technically the new function initializeLocalSymbols can be merged into
ObjFile::postParse, but functions like getSrcMsg may access the
uninitialized (all nullptr) local part of InputFile::symbols.
Linking chrome: 1.02x as fast with glibc malloc, 1.04x as fast with mimalloc
Reviewed By: ikudrin
Differential Revision: https://reviews.llvm.org/D119909
Symbol.h depends on InputFiles.h. This change moves us toward dropping the
weird dependency.
The call sites will become slightly uglier (`cast<SharedFile>(s->file)`), but
the compromise is acceptable.
https://discourse.llvm.org/t/parallel-input-file-parsing/60164
To decouple symbol initialization and section initialization, `Defined::section`
assignment should be postponed after input file parsing. To avoid spurious
duplicate definition error due to two definitions in COMDAT groups of the same
signature, we should postpone the duplicate symbol check.
The function is called postScan instead of a more specific name like
checkDuplicateSymbols, because we may merge Symbol::mergeProperties into
postScan. It is placed after compileBitcodeFiles to apply to ET_REL files
produced by LTO. This causes minor diagnostic regression
for skipLinkedOutput configurations: ld.lld --thinlto-index-only a.bc b.o
(bitcode definition prevails) won't detect duplicate symbol error. I think this
is an acceptable compromise. The important cases where (a) both files are
bitcode or (b) --thinlto-index-only is unused are still detected.
Reviewed By: ikudrin
Differential Revision: https://reviews.llvm.org/D119908
In many call sites we know uncompression cannot happen (non-SHF_ALLOC, or the
data (even if compressed) must have been uncompressed by a previous pass).
Prefer rawData in these cases. data() increases code size and prevents
optimization on rawData.
https://maskray.me/blog/2022-01-16-archives-and-start-lib
For every definition in an extracted archive member, we intern the symbol twice,
once for the archive index entry, once for the .o symbol table after extraction.
This is inefficient.
Symbols in a --start-lib ObjFile/BitcodeFile are only interned once because the
result is cached in symbols[i].
Just handle an archive using the --start-lib code path. We can therefore remove
ArchiveFile and LazyArchive. For many projects, archive member extraction ratio
is high and it is a net performance win. Linking a Release build of clang is
1.01x as fast.
Note: --start-lib scans symbols in the same order that llvm-ar adds them to the
index, so in the common case the semantics should be identical. If the archive
symbol table was created in a different order, or is incomplete, this strategy
may have different semantics. Such cases are considered user error.
The `is neither ET_REL nor LLVM bitcode` error is changed to a warning.
Previously an archive may have such members without a diagnostic. Using a
warning prevents breakage.
* For some tests, the diagnostics get improved where we did not consider
the archive member name: `b.a:` => `b.a(b.o):`.
* `no-obj.s`: the link is now allowed, matching GNU ld
* `archive-no-index.s`: the `is neither ET_REL nor LLVM bitcode` diagnostic is
demoted to a warning.
* `incompatible.s`: even when an archive is unextracted, we may report an
"incompatible with" error.
---
I recently decreased sizeof(SymbolUnion) by 8 and decreased memory usage quite a
bit, so retaining `symbols` for un-extracted archive members should not cause a
memory usage problem.
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D119074
https://maskray.me/blog/2022-02-06-all-about-common-symbols#no-define-common
In GNU ld, -dc only affects -r links and causes COMMON symbols to be allocated.
--no-define-common is defined to make COMMON symbols undefined for -shared.
AIUI --no-define-common is a workaround around glibc 2.1 time and not really useful.
gold confuses --define-common with -d/FORCE_COMMON_ALLOCATION and implements
--define-common with -d semantics. Its --no-define-common is incompatible with
GNU ld.
In ld.lld, b2a23cf3c08cee45614f27eb2c6d044e506aa6a6 fixed the default -r
behavior for COMMON symbols but ported the incompatible gold
--[no-]define-common. To the best of my knowledge, no project uses -dp
--[no-]define-common. So just remove these options.
-d/-dc are used by the following projects:
* grub grub-core/genmod.sh.in uses -Wl,-r,-d (https://lists.gnu.org/archive/html/grub-devel/2022-02/msg00088.html)
* FreeBSD crunchgen uses -Wl,-dc (https://reviews.freebsd.org/D34215)
A no-op implementation works for them. Only when a program inspects relocatable
output by itself and does not recognize COMMON symbols, there may be a problem.
This is an extremely unlikely case.
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D119108
Currently `this->getName() == newSym.getName()`.
By keeping the old nameData/nameSize, newSym's nameData/nameSize will be
ignored. The call sites can avoid calling getName().
printTraceSymbol needs to take the symbol name since `other`'s name is empty.
There's a couple of motivations here:
* LLD 12 (which I was originally testing with) was adding an undefined
symbol to the symbol table if you attempted to wrap an unreferenced
lazy symbol, which would later break `--no-allow-shlib-undefined`. LLD
on main actually produces a weak undefined symbol, so this doesn't
break anyway, but it's cleaner to not have the weak undefined symbol
as well. The new behavior also matches bfd and gold.
* PROVIDE in a linker script referencing a wrapped symbol would think
that an otherwise-unreferenced lazy symbol which was wrapped was
actually referenced, and therefore proceed with the definition, which
goes against expectations. The new behavior also matches bfd and gold.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D118756
Replace `f<ELFT>(x)` with `InvokeELFT(f, x)`.
The size reduction comes from turning `link` from 4 specializations into 1.
My x86-64 lld executable is 26KiB smaller.
Reviewed By: ikudrin
Differential Revision: https://reviews.llvm.org/D118551
Previously an InputSectionBase is dead (`partition==0`) by default.
SyntheticSection calls markLive and BssSection overrides that with markDead.
It is more natural to make InputSectionBase live by default and let
--gc-sections mark InputSectionBase dead.
When linking a Release build of clang:
* --no-gc-sections:, the removed `inputSections` loop decreases markLive time from 4ms to 1ms.
* --gc-sections: the extra `inputSections` loop increases markLive time from 0.181296s to 0.188526s.
This is as of we lose the removing one `inputSections` loop optimization (4374824ccf6e7ae68264d996a9ae5bb5e3be7fc5).
I believe the loss can be mitigated if we refactor markLive.
When an archive with an empty index contains only bitcode files, it is
handled as a group of lazy (--start-lib) object files. If there is a
non-bitcode file, there will be a diagnostic a la GNU ld.
For some programs, the archive member extraction ratio is high (e.g. for chrome,
79% archive members are extracted according to --print-archive-stats=). Because
symbol interning is cached for ObjFile::parseLazy but not for ArchiveFile,
parsing an archive as a group of --start-lib object files may be faster.
If the linker speculatively creates section representations for archive members,
the archive index will not be used.
If we take the above view, the archive index is essentially useless. If a user
wants a fast build without using --start-lib, they may just build thin archives
without index (`ar rcS --thin`).
Therefore, I suggest that we no longer treat the code as a hack, instead as a
supported feature. I believe we will do this anyway if we add parallel symbol
interning (parallel symbol interning for lazy object files is simpler than that
for archives).
Ecosystem issues:
* parseLazy actually has nearly the same behavior as ArchiveFile::parse, but the symbol order may be different.
* users may get addicted to the behavior and build archives not working with GNU ld and gold. I think it is easy to rebuild archives to be compatible.
Reviewed By: ikudrin
Differential Revision: https://reviews.llvm.org/D117284
Move all variables at file-scope or function-static-scope into a hosting structure (lld::CommonLinkerContext) that lives at lldMain()-scope. Drivers will inherit from this structure and add their own global state, in the same way as for the existing COFFLinkerContext.
See discussion in https://lists.llvm.org/pipermail/llvm-dev/2021-June/151184.html
Differential Revision: https://reviews.llvm.org/D108850
to decrease sizeof(SymbolUnion) by 8 on ELF64 platforms.
Symbols needing such information are typically 1% or fewer (5134 out of 560520
when linking clang, 19898 out of 5550705 when linking chrome). Storing them
elsewhere can decrease memory usage and symbol initialization time.
There is a ~0.8% saving on max RSS when linking a large program.
Future direction:
* Move some of dynsymIndex/verdefIndex/versionId to SymbolAux
* Support mixed TLSDESC and TLS GD without increasing sizeof(SymbolUnion)
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D116281
This complements D111365.
D111365 did not demote isUsedInRegularObj lazy symbols just to work around
a --symbol-ordering-file diagnostic quirk.
The quirk was dropped by 00dd2d15a40b0fe9916bb55a48f264498d8fe910,
so we can demote all lazy symbols now, not just the isUsedInRegularObj ones.
@tejohnson noticed that freeing MemoryBuffer instances right before
`lto->compile` can save RSS, likely because the memory can be reused by
LTO indexing (e.g. ThinLTO import/export lists).).
For ELFFileBase instances, symbol and section names are backed by MemoryBuffer,
so destroying MemoryBuffer would make some infrequent passes (parseSymbolVersion,
reportBackrefs) crash and make debugging difficult.
For a BitcodeFile, its content is completely unused, but destroying its
MemoryBuffer makes the buffer identifier inaccessible and may introduce
constraints for future changes.
This patch leverages madvise(MADV_DONTNEED) which achieves the major gain
without the latent issues.
`Maximum resident set size (kbytes): ` for a large --thinlto-index-only link:
* current behavior: 10146104KiB
* destroy MemoryBuffer instances: 8555240KiB
* madvise(MADV_DONTNEED) just bitcodeFiles and lazyBitcodeFiles: 8737372KiB
* madvise(MADV_DONTNEED) all MemoryBuffers: 8739796KiB (16% decrease)
Depends on D116366
Reviewed By: tejohnson
Differential Revision: https://reviews.llvm.org/D116367
and remove associated make<XXX> calls.
gnuHash and sysvHash are unchanged, otherwise LinkerScript::discard would
destroy the objects which may be referenced by input section descriptions.
My x86-64 lld executable is 121+KiB smaller.
Placeholders (-y and redirectSymbols removed versioned symbols) are very rare and
the check just makes symbol table iteration slower. Most iterations filter out
placeholders anyway, so this change just drops the filter behavior.
For "Add symbols to symtabs", we need to ensure that redirectSymbols sets
isUsedInRegularObj to false when making a symbol placeholder, to avoid an
assertion failure in SymbolTableSection<ELFT>::writeTo.
My .text is 2KiB smaller. The speed-up linking chrome is 0.x%.
"Process symbol versions" may take 2+% time.
"Redirect symbols" may take 0.6% time.
This change speeds up the two passes and makes `*sym.getVersionSuffix()
== '@'` in the `undefined reference` diagnostic cleaner.
Linking chrome (no debug info) and another large program is 1.5% faster.
For empty-ver2.s: the behavior now matches GNU ld, though I'd consider the input
invalid and the exact behavior does not matter.
This reverts commit e48b1c8a27f0fbd791edc8e45756b268caadfa66.
This reverts commit d019de23a1d761225fdaf0c47394ba58143aea9a.
The changes caused memory leaks (non-final classes cannot use unique_ptr).
The new `lazy` state is the inverse of the previous `LazyObjFile::extracted`.
There are many advantages:
* previously when a LazyObjFile was extracted, a new ObjFile/BitcodeFile was created; now the file is reused, just with `lazy` cleared
* avoid the confusing transfer of `symbols` from LazyObjFile to the new file
* the `incompatible file:` diagnostic is unified with `is incompatible with`
* simpler code, smaller executable (6200+ bytes smaller on x86-64)
* make eager parsing feasible (for parallel section/symbol table initialization)
Currently the singleton `config` is assigned by `config = make<Configuration>()`
and (if `canExitEarly` is false) destroyed by `lld::freeArena`.
`make<Configuration>` allocates a stab with `malloc(4096)`. This both wastes
memory and bloats the executable (every type instantiates `BumpPtrAllocator`
which costs more than 1KiB code on x86-64).
(No need to worry about `clang::no_destroy`. Regular invocations (`canExitEarly`
is true) call `_Exit` via llvm::sys::Process::ExitNoCleanup.)
Reviewed By: lichray
Differential Revision: https://reviews.llvm.org/D116143
There is a trend of having more optional options (usually security
hardening related) like -z cet-report=, -z bti-report=, -z force-bti.
If ld.lld 14.0.0 uses a warning, in 15/16/17/... timeframe when people
add new options to software, they can worry less about linker errors on ld.lld 14.0.0.
In some cases `-z foo` does essential work where a silent ignore can be
problematic, but the user has received a warning. From my observation, the
doing-essential-work `-z foo` is much fewer than the converse. In addition,
the user who cares can use `--fatal-warnings` (Note: GNU ld doesn't upgrade warnings to errors).
It is unclear whether we need something like `clang -Wunknown-warning-option`.
If we ever run into unfortunate transition like `-z start-stop-gc`, the
affected software (e.g. ldc is a compiler which passes linker options to the underlying ld)
can blindly add the `-z` option, without worrying it may cause a linker error to LLD 14.0.0.
Reviewed By: jrtc27, peter.smith
Differential Revision: https://reviews.llvm.org/D114748
Fix r285764: there is no guarantee that Out::first is placed before other
static data members of `struct Out`. After `bufferStart` was introduced, this
out-of-bounds write is destined in many compilers. It is likely benign, though.
And move `Out::elfHeader->size` assignment beside `Out::elfHeader->sectionIndex`
This allows --power10-stubs= and --[no-]power10-stubs to override each other
(they are position dependent in GNU ld).
Also improve --help messages and the manpage.
Note: GNU ld's default "auto" mode uses heuristics to decide whether Power10
instructions are used. Arguably it is a design mistake of R_PPC64_REL24_NOTOC
(acked by the relevant folks on a libc-alpha discussion). We don't implement
"auto", so the default --power10-stubs is the same as "yes".
