When enabled, input sections that would otherwise overflow a memory
region are instead spilled to the next matching output section.
This feature parallels the one in GNU LD, but there are some differences
from its documented behavior:
- /DISCARD/ only matches previously-unmatched sections (i.e., the flag
does not affect it).
- If a section fails to fit at any of its matches, the link fails
instead of discarding the section.
- The flag --enable-non-contiguous-regions-warnings is not implemented,
as it exists to warn about such occurrences.
The implementation places stubs at possible spill locations, and
replaces them with the original input section when effecting spills.
Spilling decisions occur after address assignment. Sections are spilled
in reverse order of assignment, with each spill naively decreasing the
size of the affected memory regions. This continues until the memory
regions are brought back under size. Spilling anything causes another
pass of address assignment, and this continues to fixed point.
Spilling after rather than during assignment allows the algorithm to
consider the size effects of unspillable input sections that appear
later in the assignment. Otherwise, such sections (e.g. thunks) may
force an overflow, even if spilling something earlier could have avoided
it.
A few notable feature interactions occur:
- Stubs affect alignment, ONLY_IF_RO, etc, broadly as if a copy of the
input section were actually placed there.
- SHF_MERGE synthetic sections use the spill list of their first
contained input section (the one that gives the section its name).
- ICF occurs oblivious to spill sections; spill lists for merged-away
sections become inert and are removed after assignment.
- SHF_LINK_ORDER and .ARM.exidx are ordered according to the final
section ordering, after all spilling has completed.
- INSERT BEFORE/AFTER and OVERWRITE_SECTIONS are explicitly disallowed.
When enabled, input sections that would otherwise overflow a memory
region are instead spilled to the next matching output section.
This feature parallels the one in GNU LD, but there are some differences
from its documented behavior:
- /DISCARD/ only matches previously-unmatched sections (i.e., the flag
does not affect it).
- If a section fails to fit at any of its matches, the link fails
instead of discarding the section.
- The flag --enable-non-contiguous-regions-warnings is not implemented,
as it exists to warn about such occurrences.
The implementation places stubs at possible spill locations, and
replaces them with the original input section when effecting spills.
Spilling decisions occur after address assignment. Sections are spilled
in reverse order of assignment, with each spill naively decreasing the
size of the affected memory regions. This continues until the memory
regions are brought back under size. Spilling anything causes another
pass of address assignment, and this continues to fixed point.
Spilling after rather than during assignment allows the algorithm to
consider the size effects of unspillable input sections that appear
later in the assignment. Otherwise, such sections (e.g. thunks) may
force an overflow, even if spilling something earlier could have avoided
it.
A few notable feature interactions occur:
- Stubs affect alignment, ONLY_IF_RO, etc, broadly as if a copy of the
input section were actually placed there.
- SHF_MERGE synthetic sections use the spill list of their first
contained input section (the one that gives the section its name).
- ICF occurs oblivious to spill sections; spill lists for merged-away
sections become inert and are removed after assignment.
- SHF_LINK_ORDER and .ARM.exidx are ordered according to the final
section ordering, after all spilling has completed.
- INSERT BEFORE/AFTER and OVERWRITE_SECTIONS are explicitly disallowed.
zstd excels at scaling from low-ratio-very-fast to
high-ratio-pretty-slow. Some users prioritize speed and prefer disk read
speed, while others focus on achieving the highest compression ratio
possible, similar to traditional high-ratio codecs like LZMA.
Add an optional `level` to `--compress-sections` (#84855) to cater to
these diverse needs. While we initially aimed for a one-size-fits-all
approach, this no longer seems to work.
(https://richg42.blogspot.com/2015/11/the-lossless-decompression-pareto.html)
When --compress-debug-sections is used together, make
--compress-sections take precedence since --compress-sections is usually
more specific.
Remove the level distinction between -O/-O1 and -O2 for
--compress-debug-sections=zlib for a more consistent user experience.
Pull Request: https://github.com/llvm/llvm-project/pull/90567
The function may return Z_MEM_ERROR or Z_STREAM_ERR. The former does not
have a good way of testing. The latter will be possible with a pending
change that allows setting the compression level, which will come with a
test.
https://reviews.llvm.org/D133679 utilizes zstd's multithread API to
create one single frame. This provides a higher compression ratio but is
significantly slower than concatenating multiple frames.
With manual parallelism, it is easier to parallelize memcpy in
OutputSection::writeTo for parallel memcpy.
In addition, as the individual allocated decompression buffers are much
smaller, we can make a wild guess (compressed_size/4) without worrying
about a resize (due to wrong guess) would waste memory.
--compress-sections <section-glib>=[none|zlib|zstd] is similar to
--compress-debug-sections but applies to broader sections without the
SHF_ALLOC flag. lld will report an error if a SHF_ALLOC section is
matched. An interesting use case is to compress `.strtab`/`.symtab`,
which consume a significant portion of the file size (15.1% for a
release build of Clang).
An older revision is available at https://reviews.llvm.org/D154641 .
This patch focuses on non-allocated sections for safety. Moving
`maybeCompress` as D154641 does not handle STT_SECTION symbols for
`-r --compress-debug-sections=zlib` (see `relocatable-section-symbol.s`
from #66804).
Since different output sections may use different compression
algorithms, we need CompressedData::type to generalize
config->compressDebugSections.
GNU ld feature request: https://sourceware.org/bugzilla/show_bug.cgi?id=27452
Link: https://discourse.llvm.org/t/rfc-compress-arbitrary-sections-with-ld-lld-compress-sections/71674
Pull Request: https://github.com/llvm/llvm-project/pull/84855
On Windows when zlib is enabled, zlib header introduced some Windows
headers which defines max as a macro. Since OutputSections.cpp uses
std::max with template argument, this causes compilation error.
Define macro NOMINMAX to avoid this.
.plt and .branch_lt have the type of SHT_NOBITS and may be relocated by dynamic
relocations with non-zero addends. They should be skipped for the
--check-dynamic-relocations check, as --apply-dynamic-relocs does not apply.
A side effect is that -z rel does not work for the two sections.
Added two --apply-dynamic-relocs --check-dynamic-relocations tests. Also checked
linking a PPC64 clang.
Arm has BE8 big endian configuration called a byte-invariant(every byte has the same address on little and big-endian systems).
When in BE8 mode:
1. Instructions are big-endian in relocatable objects but
little-endian in executables and shared objects.
2. Data is big-endian.
3. The data encoding of the ELF file is ELFDATA2MSB.
To support BE8 without an ABI break for relocatable objects,the linker takes on the responsibility of changing the endianness of instructions. At a high level the only difference between BE32 and BE8 in the linker is that for BE8:
1. The linker sets the flag EF_ARM_BE8 in the ELF header.
2. The linker endian reverses the instructions, but not data.
This patch adds BE8 big endian support for Arm. To endian reverse the instructions we'll need access to the mapping symbols. Code sections can contain a mix of Arm, Thumb and literal data. We need to endian reverse Arm instructions as words, Thumb instructions
as half-words and ignore literal data.The only way to find these transitions precisely is by using mapping symbols. The instruction reversal will need to take place after relocation. For Arm BE8 code sections (Section has SHF_EXECINSTR flag ) we inserted a step after relocation to endian reverse the instructions. The implementation strategy i have used here is to write all sections BE32 including SyntheticSections then endian reverse all code in InputSections via mapping symbols.
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D150870
Much of NOLOAD's intended use is to explicitly change the type of an
output section, so we shouldn't flag these as warnings.
Differential Revision: https://reviews.llvm.org/D151144
In a link map, the input section name gives more information. See the updated
merge-entsize.s for an example. The output file is unchanged.
Compiler generated input sections with the SHF_MERGE flag have names such as
.rodata.str1.1 and .rodata.cstN, and are not affected by -fdata-sections.
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D149466
This patch allows to specify that some part of tasks should be
done in sequential order. It makes it possible to not use
condition operator for separating sequential tasks:
TaskGroup tg;
for () {
if(condition) ==> tg.spawn([](){fn();}, condition)
fn();
else
tg.spawn([](){fn();});
}
It also prevents execution on main thread. Which allows adding
checks for getThreadIndex() function discussed in D142318.
The patch also replaces std::stack with std::deque in the
ThreadPoolExecutor to have natural execution order in case
(parallel::strategy.ThreadsRequested == 1).
Differential Revision: https://reviews.llvm.org/D148728
By using emplace_back, as well as converting some loops to for-each, we can do more efficient vectorization.
Make copy constructor for TemporaryFile noexcept.
Reviewed By: #lld-macho, int3
Differential Revision: https://reviews.llvm.org/D139552
See D117853: compressing debug sections is a bottleneck and therefore it
has a large value parallizing the step.
zstd provides multi-threading API and the output is deterministic even with
different numbers of threads (see https://github.com/facebook/zstd/issues/2238).
Therefore we can leverage it instead of using the pigz-style sharding approach.
Also, switch to the default compression level 3. The current level 5
is significantly slower without providing justifying size benefit.
```
'dash b.sh 1' ran
1.05 ± 0.01 times faster than 'dash b.sh 3'
1.18 ± 0.01 times faster than 'dash b.sh 4'
1.29 ± 0.02 times faster than 'dash b.sh 5'
level=1 size: 358946945
level=3 size: 309002145
level=4 size: 307693204
level=5 size: 297828315
```
Reviewed By: andrewng, peter.smith
Differential Revision: https://reviews.llvm.org/D133679
`clang -gz=zstd a.o` passes this option to the linker. This option compresses output
debug sections with zstd and sets ch_type to ELFCOMPRESS_ZSTD. As of today, very
few DWARF consumers recognize ELFCOMPRESS_ZSTD.
Use the llvm::zstd::compress API with level llvm::zstd::DefaultCompression (5),
which we may tune after we have more experience with zstd output.
zstd has built-in parallel compression support (so we don't need to do D117853
for zlib), which is not leveraged yet.
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D133548
We currently process one OutputSection at a time and for each OutputSection
write contained input sections in parallel. This strategy does not leverage
multi-threading well. Instead, parallelize writes of different OutputSections.
The default TaskSize for parallelFor often leads to inferior sharding. We
prepare the task in the caller instead.
* Move llvm::parallel::detail::TaskGroup to llvm::parallel::TaskGroup
* Add llvm::parallel::TaskGroup::execute.
* Change writeSections to declare TaskGroup and pass it to writeTo.
Speed-up with --threads=8:
* clang -DCMAKE_BUILD_TYPE=Release: 1.11x as fast
* clang -DCMAKE_BUILD_TYPE=Debug: 1.10x as fast
* chrome -DCMAKE_BUILD_TYPE=Release: 1.04x as fast
* scylladb build/release: 1.09x as fast
On M1, many benchmarks are a small fraction of a percentage faster. Mozilla showed the largest difference with the patch being about 1.03x as fast.
Differential Revision: https://reviews.llvm.org/D131247
In the majority of cases (e.g. orphan sections), an OutputSection has at most
one InputSectionDescription (isd). By changing the return type to
ArrayRef<InputSection *> we can just reference the isd->sections. For
OutputSections with more than one InputSectionDescription we use a caller
provided SmallVector to copy the elements as before.
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D129111
Patch created by running:
rg -l parallelForEachN | xargs sed -i '' -c 's/parallelForEachN/parallelFor/'
No behavior change.
Differential Revision: https://reviews.llvm.org/D128140
Placing a non-SHT_NOBITS input section in an output section specified with
(NOLOAD) is fishy but used by some projects. D118840 changed the output type to
SHT_PROGBITS, but using the specified type seems to make more sense and improve
GNU ld compatibility: `(NOLOAD)` seems to change the output section type
regardless of input.
I think we should keep the current type mismatch warning as it does indicate an
error-prone usage.
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D125074
Add an OutputDesc class inheriting from SectionCommand. An OutputDesc wraps an
OutputSection. This change allows InputSection::getParent to be inlined.
Differential Revision: https://reviews.llvm.org/D120650
Making a (NOLOAD) section SHT_PROGBITS is fishy (the user may expect all-zero
content, but the linker does not check that), but some projects (e.g. Linux
kernel https://github.com/ClangBuiltLinux/linux/issues/1597) traditionally rely
on the behavior. Issue a warning to not break them.
The current output section type allows to set the ELF section type to
SHT_PROGBITS or SHT_NOLOAD. This patch allows an arbitrary section value
to be specified. Some common SHT_* literal names are supported as well.
```
SECTIONS {
note (TYPE=SHT_NOTE) : { BYTE(8) *(note) }
init_array ( TYPE=14 ) : { QUAD(14) }
fini_array (TYPE = SHT_FINI_ARRAY) : { QUAD(15) }
}
```
When `sh_type` is specified, it is an error if an input section has a different type.
Our syntax is compatible with GNU ld 2.39 (https://sourceware.org/bugzilla/show_bug.cgi?id=28841).
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D118840
lld/ELF/OutputSections.cpp includes llvm/Config/config.h for
LLVM_ENABLE_ZLIB definition, but llvm/Config/config.h doesn't exist in
standalone build.
To fix this, this patch moves LLVM_ENABLE_ZLIB from config.h to
llvm-config.h and updates OutputSections.cpp to include llvm-config.h
instead of config.h
Reviewed By: MaskRay, mgorny
Differential Revision: https://reviews.llvm.org/D119058
To fix
../../chromeclang/bin/../include/c++/v1/__algorithm/min.h:39:1: note: candidate template ignored: deduced conflicting types for parameter '_Tp' ('unsigned long' vs. 'unsigned long long')
on macOS arm64.
When linking a Debug build clang (265MiB SHF_ALLOC sections, 920MiB uncompressed
debug info), in a --threads=1 link "Compress debug sections" takes 2/3 time and
in a --threads=8 link "Compress debug sections" takes ~70% time.
This patch splits a section into 1MiB shards and calls zlib `deflake` parallelly.
DEFLATE blocks are a bit sequence. We need to ensure every shard starts
at a byte boundary for concatenation. We use Z_SYNC_FLUSH for all shards
but the last to flush the output to a byte boundary. (Z_FULL_FLUSH can
be used as well, but Z_FULL_FLUSH clears the hash table which just
wastes time.)
The last block requires the BFINAL flag. We call deflate with Z_FINISH
to set the flag as well as flush the output to a byte boundary. Under
the hood, all of Z_SYNC_FLUSH, Z_FULL_FLUSH, and Z_FINISH emit a
non-compressed block (called stored block in zlib). RFC1951 says "Any
bits of input up to the next byte boundary are ignored."
In a --threads=8 link, "Compress debug sections" is 5.7x as fast and the total
speed is 2.54x. Because the hash table for one shard is not shared with the next
shard, the output is slightly larger. Better compression ratio can be achieved
by preloading the window size from the previous shard as dictionary
(`deflateSetDictionary`), but that is overkill.
```
# 1MiB shards
% bloaty clang.new -- clang.old
FILE SIZE VM SIZE
-------------- --------------
+0.3% +129Ki [ = ] 0 .debug_str
+0.1% +105Ki [ = ] 0 .debug_info
+0.3% +101Ki [ = ] 0 .debug_line
+0.2% +2.66Ki [ = ] 0 .debug_abbrev
+0.0% +1.19Ki [ = ] 0 .debug_ranges
+0.1% +341Ki [ = ] 0 TOTAL
# 2MiB shards
% bloaty clang.new -- clang.old
FILE SIZE VM SIZE
-------------- --------------
+0.2% +74.2Ki [ = ] 0 .debug_line
+0.1% +72.3Ki [ = ] 0 .debug_str
+0.0% +69.9Ki [ = ] 0 .debug_info
+0.1% +976 [ = ] 0 .debug_abbrev
+0.0% +882 [ = ] 0 .debug_ranges
+0.0% +218Ki [ = ] 0 TOTAL
```
Bonus in not using zlib::compress
* we can compress a debug section larger than 4GiB
* peak memory usage is lower because for most shards the output size is less
than 50% input size (all less than 55% for a large binary I tested, but
decreasing the initial output size does not decrease memory usage)
Reviewed By: ikudrin
Differential Revision: https://reviews.llvm.org/D117853
