inputSections temporarily contains EhInputSection objects mainly for
combineEhSections. Place EhInputSection objects into a new vector
ehInputSections instead of inputSections.
A LazySymbol is one that lives in `.a` archive and gets pulled in by a
strong reference. However, weak references to such symbols do not
result in them be loaded from the archive. In this case we want to
treat such symbols at undefined rather then lazy, once symbols
resolution is complete.
This fixes a crash bug in the linker when weakly referenced symbol that
lives in an archive file is live at the end of the link. In the case of
dynamic linking this is expected to turn into an import with (in the
case of a function symbol) a function index.
Differential Revision: https://reviews.llvm.org/D130736
Similarly to -o output directories will not be created so -Map being
copied verbatim will likely cause ld.lld @response.txt to fail.
Differential Revision: https://reviews.llvm.org/D130681
A symbol `$ld$previous$/Another$1.2.3$1$3.0$14.0$_xxx$` means
"pretend symbol `_xxx` is in dylib `/Another` with version `1.2.3`
if the deployment target is between `3.0` and `14.0` and we're
targeting platform `1` (ie macOS)".
This means dylibs can now inject synthetic dylibs into the link, so
DylibFile needs to grow a 3rd constructor.
The only other interesting thing is that such an injected dylib
counts as a use of the original dylib. This patch gets this mostly
right (if _only_ `$ld$previous` symbols are used from a dylib,
we don't add a dep on the dylib itself, matching ld64), but one case
where we don't match ld64 yet is that ld64 even omits the original
dylib when linking it with `-needed-l`. Lld currently still adds a load
command for the original dylib in that case. (That's for a future
patch.)
Fixes#56074.
Differential Revision: https://reviews.llvm.org/D130725
Linking fails when targeting `x86_64-apple-darwin` for runtimes. The issue
is that LLD strictly assumes the target architecture be present in the tbd
files (which isn't always true). For example, when targeting `x86_64h`, it should
work with `x86_64` because they are ABI compatible. This is also inline with what
ld64 does.
An environment variable (which ld64 also supports) is also added to preserve the
existing behavior of strict architecture matching.
Reviewed By: #lld-macho, int3
Differential Revision: https://reviews.llvm.org/D130683
We don't need to recompute the list LLVMConfig.cmake provides us.
When LLVM is being built, the list is two elements long: generated headers and headers from source.
When LLVM is already built,the list is one element long: the installed header directory containing both generated and hand-written sources.
Reviewed By: sebastian-ne
Differential Revision: https://reviews.llvm.org/D130553
We were previously doing it after LTO, which did have the desired effect
of having the un-exported symbols marked as private extern in the final
output binary, but doing it before LTO creates more optimization
opportunities.
One observable difference is that LTO can now elide un-exported symbols
entirely, so they may not even be present as private externs in the
output.
This is also what ld64 implements.
Reviewed By: #lld-macho, thevinster
Differential Revision: https://reviews.llvm.org/D130429
This hint instructs the linker to optimize an adrp+add+ldr sequence used
for loading from a local symbol's address by loading directly if it's
close enough, or with an adrp(p)+ldr sequence if it's not.
This transformation is the same as what's done for ADRP_LDR_GOT_LDR when
the symbol is local. The logic for acting on this hint is therefore
moved to a new function which will be called from the existing
applyAdrpLdrGotLdr() function.
Differential Revision: https://reviews.llvm.org/D130505
In DWARF5, the `DW_AT_name` and `DW_AT_comp_dir` attributes are encoded
using the `strx*` forms, which specify an index into `__debug_str_offs`.
This commit adds that section to DwarfObject, so the debug info parser
can resolve these references.
The test case was manually adapted from stabs-icf.s.
Fixes#51668
Differential Revision: https://reviews.llvm.org/D130559
This is still undocumented and unsupported, but if someone passed it
before you would end up with a missing file error since this takes an
argument that wouldn't be handled.
Differential Revision: https://reviews.llvm.org/D130606
Turning on opaque pointers has uncovered an issue with WPD where we currently pattern match away `assume(type.test)` in WPD so that a later LTT doesn't resolve the type test to undef and introduce an `assume(false)`. The pattern matching can fail in cases where we transform two `assume(type.test)`s into `assume(phi(type.test.1, type.test.2))`.
Currently we create `assume(type.test)` for all virtual calls that might be devirtualized. This is to support `-Wl,--lto-whole-program-visibility`.
To prevent this, all virtual calls that may not be in the same LTO module instead use a new `llvm.public.type.test` intrinsic in place of the `llvm.type.test`. Then when we know if `-Wl,--lto-whole-program-visibility` is passed or not, we can either replace all `llvm.public.type.test` with `llvm.type.test`, or replace all `llvm.public.type.test` with `true`. This prevents WPD from trying to pattern match away `assume(type.test)` for public virtual calls when failing the pattern matching will result in miscompiles.
Reviewed By: tejohnson
Differential Revision: https://reviews.llvm.org/D128955
Fixes a regression from D117973, that used CMAKE_BINARY_DIR instead of
LLVM_BINARY_DIR in some places.
Differential Revision: https://reviews.llvm.org/D130555
Most Arm disassemblers, including GNU objdump and Arm's own `fromelf`,
emit an instruction's raw encoding as a 32-bit words or (for Thumb)
one or two 16-bit halfwords, in logical order rather than according to
their storage endianness. This is generally easier to read: it matches
the encoding diagrams in the architecture spec, it matches the value
you'd write in a `.inst` directive, and it means that fields within
the instruction encoding that span more than one byte (such as branch
offsets or `SVC` immediates) can be read directly in the encoding
without having to mentally reverse the bytes.
llvm-objdump already has a system of PrettyPrinter subclasses which
makes it easy for a target to drop in its own preferred formatting.
This patch adds pretty-printers for all the Arm targets, so that
llvm-objdump will display Arm instruction encodings in their preferred
layout instead of little-endian and bytewise.
Reviewed By: DavidSpickett
Differential Revision: https://reviews.llvm.org/D130358
Similarly to -load_hidden, this flag instructs the linker to not export
symbols from the specified archive. While that flag takes a path,
-hidden-l looks for the specified library name in the search path.
The test changes are needed because -hidden-lfoo resolves to libfoo.a,
not foo.a.
Differential Revision: https://reviews.llvm.org/D130529
Firstly, we we make an additional GNUInstallDirs-style variable. With
NixOS, for example, this is crucial as we want those to go in
`${dev}/lib/cmake` not `${out}/lib/cmake` as that would a cmake subdir
of the "regular" libdir, which is installed even when no one needs to do
any development.
Secondly, we make *Config.cmake robust to absolute package install
paths. We for NixOS will in fact be passing them absolute paths to make
the `${dev}` vs `${out}` distinction mentioned above, and the
GNUInstallDirs-style variables are suposed to support absolute paths in
general so it's good practice besides the NixOS use-case.
Thirdly, we make `${project}_INSTALL_PACKAGE_DIR` CACHE PATHs like other
install dirs are.
Reviewed By: sebastian-ne
Differential Revision: https://reviews.llvm.org/D117973
This flag was introduced in ld64-609. It instructs the linker to link to
a static library while treating its symbols as if they had hidden
visibility. This is useful when building a dylib that links to static
libraries but we don't want the symbols from those to be exported.
Closes#51505
This reland adds bitcode file handling, so we won't get any compile
errors due to BitcodeFile::forceHidden being unused.
Differential Revision: https://reviews.llvm.org/D130473
This flag was introduced in ld64-609. It instructs the linker to link to
a static library while treating its symbols as if they had hidden
visibility. This is useful when building a dylib that links to static
libraries but we don't want the symbols from those to be exported.
Closes#51505
Differential Revision: https://reviews.llvm.org/D130473
If the `-demangle` flag is passed to lld, symbol names will now be
demangled in the "referenced by:" message in addition to the referenced
symbol's name, which was already demangled before this change.
Differential Revision: https://reviews.llvm.org/D130490
Previously, we treated it as a regular ConcatInputSection. However, ld64
actually parses its contents and uses that to synthesize a single image
info struct, generating one 8-byte section instead of `8 * number of
object files with ObjC code`.
I'm not entirely sure what impact this section has on the runtime, so I
just tried to follow ld64's semantics as closely as possible in this
diff. My main motivation though was to reduce binary size.
No significant perf change on chromium_framework on my 16-core Mac Pro:
base diff difference (95% CI)
sys_time 1.764 ± 0.062 1.748 ± 0.032 [ -2.4% .. +0.5%]
user_time 5.112 ± 0.104 5.106 ± 0.046 [ -0.9% .. +0.7%]
wall_time 6.111 ± 0.184 6.085 ± 0.076 [ -1.6% .. +0.8%]
samples 30 32
Reviewed By: #lld-macho, thakis
Differential Revision: https://reviews.llvm.org/D130125
which occurs when there are EH frames present in the object file's weak
def.
Reviewed By: abrachet
Differential Revision: https://reviews.llvm.org/D130409
llvm::sort is beneficial even when we use the iterator-based overload,
since it can optionally shuffle the elements (to detect
non-determinism). However llvm::sort is not usable everywhere, for
example, in compiler-rt.
Reviewed By: nhaehnle
Differential Revision: https://reviews.llvm.org/D130406
First of all, `LLVM_TOOLS_INSTALL_DIR` put there breaks our NixOS
builds, because `LLVM_TOOLS_INSTALL_DIR` defined the same as
`CMAKE_INSTALL_BINDIR` becomes an *absolute* path, and then when
downstream projects try to install there too this breaks because our
builds always install to fresh directories for isolation's sake.
Second of all, note that `LLVM_TOOLS_INSTALL_DIR` stands out against the
other specially crafted `LLVM_CONFIG_*` variables substituted in
`llvm/cmake/modules/LLVMConfig.cmake.in`.
@beanz added it in d0e1c2a550 to fix a
dangling reference in `AddLLVM`, but I am suspicious of how this
variable doesn't follow the pattern.
Those other ones are carefully made to be build-time vs install-time
variables depending on which `LLVMConfig.cmake` is being generated, are
carefully made relative as appropriate, etc. etc. For my NixOS use-case
they are also fine because they are never used as downstream install
variables, only for reading not writing.
To avoid the problems I face, and restore symmetry, I deleted the
exported and arranged to have many `${project}_TOOLS_INSTALL_DIR`s.
`AddLLVM` now instead expects each project to define its own, and they
do so based on `CMAKE_INSTALL_BINDIR`. `LLVMConfig` still exports
`LLVM_TOOLS_BINARY_DIR` which is the location for the tools defined in
the usual way, matching the other remaining exported variables.
For the `AddLLVM` changes, I tried to copy the existing pattern of
internal vs non-internal or for LLVM vs for downstream function/macro
names, but it would good to confirm I did that correctly.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D117977
`advanceSubsection()` didn't account for the possibility that a section
could have no subsections.
Reviewed By: #lld-macho, thakis, BertalanD
Differential Revision: https://reviews.llvm.org/D130288
If there are multiple symbols at the same address, our unwind info
implementation assumes that we always register unwind entries to a
single canonical symbol.
This assumption was violated by the `registerEhFrame` code.
Fixes#56570.
Reviewed By: #lld-macho, thakis
Differential Revision: https://reviews.llvm.org/D130208
It's only used in one branch, so we were unnecessarily calculating the
length of many symbol names.
Tiny speedup when linking chromium_framework on my M1 Mac mini:
x before.txt
+ after.txt
N Min Max Median Avg Stddev
x 10 3.9917109 4.0418 4.0318099 4.0203902 0.021459873
+ 10 3.944725 4.053988 3.9708955 3.9825602 0.037257609
Difference at 95.0% confidence
-0.03783 +/- 0.0285663
-0.940953% +/- 0.710536%
(Student's t, pooled s = 0.0304028)
Differential Revision: https://reviews.llvm.org/D130234
This commit reduces the size of the emitted rebase sections by
generating the REBASE_OPCODE_DO_REBASE_ADD_ADDR_ULEB and
REBASE_OPCODE_DO_REBASE_ULEB_TIMES_SKIPPING_ULEB opcodes.
With this change, chromium_framework's rebase section is a 40% smaller
197 kilobytes, down from the previous 320 kB. That is 6 kB smaller than
what ld64 produces for the same input.
Performance figures from my M1 Mac mini:
x before
+ after
N Min Max Median Avg Stddev
x 10 4.2269349 4.3300061 4.2689675 4.2690016 0.031151669
+ 10 4.219331 4.2914009 4.2398136 4.2448277 0.023817308
No difference proven at 95.0% confidence
Differential Revision: https://reviews.llvm.org/D130180
Similar to cstrings ld64 always deduplicates cfstrings. This was already
being done when enabling ICF, but for debug builds you may want to flip
this on if you cannot eliminate your instances of this, so this change
makes --deduplicate-literals also apply to cfstrings.
Differential Revision: https://reviews.llvm.org/D130134
Print the actual number of symbols that would have been exported
too, which helps assessing the situation.
Differential Revision: https://reviews.llvm.org/D130117
This is a follow-on to {D129556}. I've refactored the code such that
`addFile()` no longer needs to take an extra parameter. Additionally,
the "do we force-load or not" policy logic is now fully contained within
addFile, instead of being split between `addFile` and
`parseLCLinkerOptions`. This also allows us to move the `ForceLoad` (now
`LoadType`) enum out of the header file.
Additionally, we can now correctly report loads induced by
`LC_LINKER_OPTION` in our `-why_load` output.
I've also added another test to check that CLI library non-force-loads
take precedence over `LC_LINKER_OPTION` + `-force_load_swift_libs`. (The
existing logic is correct, just untested.)
Reviewed By: #lld-macho, thakis
Differential Revision: https://reviews.llvm.org/D130137
The new format uses symbol relocations, as described in {D127637}.
Reviewed By: #lld-macho, alx32
Differential Revision: https://reviews.llvm.org/D128938
This fixes https://github.com/llvm/llvm-project/issues/56059 and
https://github.com/llvm/llvm-project/issues/56440. This is inspired by
tapthaker's patch (https://reviews.llvm.org/D127941), and has reused his
test cases. This patch adds an bool "isCommandLineLoad" to indicate
where archives are from. If lld tries to load the same library loaded
previously by LC_LINKER_OPTION from CLI, it will use this
isCommandLineLoad to determine if it should be affected by -all_load &
-ObjC flags. This also prevents -force_load from affecting archives
loaded previously from CLI without such flag, whereas tapthaker's patch
will fail such test case (introduced by
https://reviews.llvm.org/D128025).
Reviewed By: int3, #lld-macho
Differential Revision: https://reviews.llvm.org/D129556
This creates a symbol alias similar to --defsym in the elf linker. This
is used by swiftpm for all executables, so it's useful to support. This
doesn't implement -alias_list but that could be done pretty easily as
needed.
Differential Revision: https://reviews.llvm.org/D129938
To do this, we need to slice away the LSDA pointer, just like we are
slicing away the functionAddress pointer.
No observable difference in perf on chromium_framework:
base diff difference (95% CI)
sys_time 1.769 ± 0.068 1.761 ± 0.065 [ -2.7% .. +1.8%]
user_time 9.517 ± 0.110 9.528 ± 0.116 [ -0.6% .. +0.8%]
wall_time 8.291 ± 0.174 8.307 ± 0.183 [ -1.1% .. +1.5%]
samples 21 25
Reviewed By: #lld-macho, thakis
Differential Revision: https://reviews.llvm.org/D129830
This method is called on each relocation when parsing input files, so
the overhead of using virtual functions ends up being quite large. We
now have a single non-virtual method, which reads from the appropriate
array of relocation attributes set in the TargetInfo constructor.
This change results in a modest 2.3% reduction in link time for
chromium_framework measured on an x86-64 VPS, and 0.7% on an arm64 Mac.
N Min Max Median Avg Stddev
x 10 11.869417 12.032609 11.935041 11.938268 0.045802324
+ 10 11.581526 11.785265 11.649885 11.659507 0.054634834
Difference at 95.0% confidence
-0.278761 +/- 0.0473673
-2.33502% +/- 0.396768%
(Student's t, pooled s = 0.0504124)
Differential Revision: https://reviews.llvm.org/D130000
Clang passes a filename rather than a directory in -lto_object_path when
using FullLTO. Previously, it was always treated it as a directory, so
lld would crash when it attempted to create temporary files inside it.
Fixes#54805
Differential Revision: https://reviews.llvm.org/D129705
While working on {D129830}, I realized that our handling of ICF +
eh_frame combined was untested. Additionally I realized that the comment
explaining why we were safely slicing away the functionAddress reloc
from our compact unwind entries was... insufficient and slightly
misleading. I've tried to clarify it.
Reviewed By: #lld-macho, thevinster
Differential Revision: https://reviews.llvm.org/D129894
We were re-defining the various numeric variables when we actually
intended to check already-defined variables against the value on the
current CHECK line.
Reviewed By: #lld-macho, thakis
Differential Revision: https://reviews.llvm.org/D129831
Use a format more similar to unresolved references from regular object
files. It's probably easier to read for people who are less familiar
with the linker diagnostics.
Reviewed By: ikudrin
Differential Revision: https://reviews.llvm.org/D129790
On my system the date formatting is a bit different from what the test used to
support. I'm using:
Windows 11 version 21H2, build 22000.795 using the English(Canada) region.
ls from BusyBox 1.36
VS 2022 17.2.5
WinSDK 10.0.22000
This just removes the code that gates the logic. The main issue here is
perf impact: without {D122258}, LLD takes a significant perf hit because
it now has to do a lot more work in the input parsing phase. But with
that change to eliminate unnecessary EH frames from input object files,
the perf overhead here is minimal. Concretely, here are the numbers for
some builds as measured on my 16-core Mac Pro:
**chromium_framework**
This is without the use of `-femit-dwarf-unwind=no-compact-unwind`:
base diff difference (95% CI)
sys_time 1.826 ± 0.019 1.962 ± 0.034 [ +6.5% .. +8.4%]
user_time 9.306 ± 0.054 9.926 ± 0.082 [ +6.2% .. +7.1%]
wall_time 8.225 ± 0.068 8.947 ± 0.128 [ +8.0% .. +9.6%]
samples 15 22
With that flag enabled, the regression mostly disappears, as hoped:
base diff difference (95% CI)
sys_time 1.839 ± 0.062 1.866 ± 0.068 [ -0.9% .. +3.8%]
user_time 9.452 ± 0.068 9.490 ± 0.067 [ -0.1% .. +0.9%]
wall_time 8.383 ± 0.127 8.452 ± 0.114 [ -0.1% .. +1.8%]
samples 17 21
**Unnamed internal app**
Without `-femit-dwarf-unwind`, this is the perf hit:
base diff difference (95% CI)
sys_time 1.372 ± 0.029 1.317 ± 0.024 [ -4.6% .. -3.5%]
user_time 2.835 ± 0.028 2.980 ± 0.027 [ +4.8% .. +5.4%]
wall_time 3.205 ± 0.079 3.383 ± 0.066 [ +4.9% .. +6.2%]
samples 102 83
With `-femit-dwarf-unwind`, the perf hit almost disappears:
base diff difference (95% CI)
sys_time 1.274 ± 0.026 1.270 ± 0.025 [ -0.9% .. +0.3%]
user_time 2.812 ± 0.023 2.822 ± 0.035 [ +0.1% .. +0.7%]
wall_time 3.166 ± 0.047 3.174 ± 0.059 [ -0.2% .. +0.7%]
samples 95 97
Just for fun, I measured the impact of `-femit-dwarf-unwind` on ld64
(`base` has the extra DWARF unwind info in the input object files,
`diff` doesn't):
base diff difference (95% CI)
sys_time 1.128 ± 0.010 1.124 ± 0.023 [ -1.3% .. +0.6%]
user_time 7.176 ± 0.030 7.106 ± 0.094 [ -1.5% .. -0.4%]
wall_time 7.874 ± 0.041 7.795 ± 0.121 [ -1.7% .. -0.3%]
samples 16 25
And for LLD:
base diff difference (95% CI)
sys_time 1.315 ± 0.019 1.280 ± 0.019 [ -3.2% .. -2.0%]
user_time 2.980 ± 0.022 2.822 ± 0.016 [ -5.5% .. -5.0%]
wall_time 3.369 ± 0.038 3.175 ± 0.033 [ -6.2% .. -5.3%]
samples 47 47
So parsing the extra EH frames is a lot more expensive for us than for
ld64. But given that we are quite a lot faster than ld64 to begin with,
I guess this isn't entirely unexpected...
Reviewed By: #lld-macho, oontvoo
Differential Revision: https://reviews.llvm.org/D129540
Currently the PPC64R2SaveStub thunk will produce Power 10 code by default.
This produced an issue when linking older code that made use of the st_other=1
bit but was never meant to be linked or run on Power 10.
This patch makes it so that only the R_PPC64_REL24_NOTOC relocation can produce
Power 10 code.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D129580
It's more natural to use uint8_t * (std::byte needs C++17 and llvm has
too much uint8_t *) and most callers use uint8_t * instead of char *.
The functions are recently moved into `llvm::compression::zlib::`, so
downstream projects need to make adaption anyway.
This load command specifies the offset and size of the exports trie.
This information used to be a field in LC_DYLD_INFO, but in newer
libraries, it has a dedicated load command: LC_DYLD_EXPORTS_TRIE.
The format of the trie is the same for both load commands, so the code
for parsing it can be shared.
LLD does not generate this yet; it is mainly useful when chained fixups
are in use, as the other members of LC_DYLD_INFO are unused then, so the
smaller LC_DYLD_EXPORTS_TRIE can be output instead.
LLDB gained support for this in D107673.
Fixes#54550
Differential Revision: https://reviews.llvm.org/D129430
This hint instructs the linker to relax a GOT-indirect load.
If the referenced symbol is external and its GOT entry is within +/- 1
MiB, the GOT entry can be loaded with a single literal ldr instruction.
If the referenced symbol is local, its address may be loaded directly if
it's close enough, or with an adr(p) + ldr pair if it's not.
This type accounts for more than half of all LOHs in chromium_framework.
This commit moves the eligibility checks into helper functions to
improve the readability of the LOH processing code. Ho functional
changes are intended to the previously implemented LOH types.
Differential Revision: https://reviews.llvm.org/D129427
D127611 computed st_value is inaccurate:
* For a backward branch, the destination address may be wrong if there is no
relaxable relocation between it and the current location due to `if (remove)`.
We may incorrectly relax a branch to c.j which ends up an overflow.
* For a forward branch, the destination address may be overestimated
and lose relaxation opportunities.
To fix the issues,
* Don't reset st_value to the original value.
* Save the st_value delta from the previous iteration into valueDelta, and use
`sa[0].d->value -= delta - valueDelta.find(sa[0].d)->second`.
When `--symbol-ordering-file` is specified, the linker today will always put
hot contributions in the middle of cold ones when targeting RISC machine, so
to minimize the chances that branch thunks need be generated for hot code
calling into cold code. This is not necessary when user specifies an ordering
of read-only data (vs. function) symbols, or when output section is small such
that no branch thunk would ever be required. The latter is common for mobile
apps. For example, among all the native ARM64 libraries in Facebook Instagram
App for Android, 80% of them have text section smaller than 64KB and the
largest text section seen is less than 8MB, well below the distance that a
BRANCH26 can reach.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D128382
It seems like the `sed` on Windows is not particularly
smart. It's not actually needed in this place, so I've
removed it's usage and just created an invalid yaml
another way.
This patch adds a new flag vfsoverlay similar to clang’s
ivfsoverlay flag. This is helpful when compiling on case
sensitive file systems when cross compiling to Windows.
Particularly when compiling third party code containing
\#pragma comment(“linker”, “/defaultlib:...”) which
can’t be easily changed.
Differential Revision: https://reviews.llvm.org/D125800
This fixes https://github.com/llvm/llvm-project/issues/56238. ld64.lld currently does not generate __dof section in Mach-O, and -no_dtrace_dof option is on by default. However when there are user-defined dtrace symbols, ld64.lld will treat them as undefined symbols, which causes the linking to fail because lld cannot find their definitions. This patch allows ld64.lld to rewrite the instructions calling dtrace symbols to instructions like nop as what ld64 does; therefore, when encountered with user-provided dtrace probes, the linking can still succeed.
I'm not sure whether support for dtrace is expected in lld, so for now I didn't add codes to make lld emit __dof section like ld64, and only made it possible to link with dtrace symbols provided. If this feature is needed, I can add that part in Dtrace.cpp & Dtrace.h.
Reviewed By: int3, #lld-macho
Differential Revision: https://reviews.llvm.org/D129062
This test was failing on our 32 bit build bot:
https://lab.llvm.org/buildbot/#/builders/178/builds/2463
This happened because in UnwindInfoSectionImpl::finalize
a decision is made whether to write out regular or compressed
unwind info.
One check in this does:
```
if (cuPtr->functionAddress >= functionAddressMax) {
break;
```
Where cuPtr->functionAddress was uint64_t and functionAddressMax
was uintptr_t, which is 4 bytes on a 32 bit system.
Using uint64_t for functionAddressMax fixes this problem.
Presumably because at only 4 bytes, the max is much lower than
we expect. We're targetting 64 bit though so the size of the max
should match the size of the addresses.
Reviewed By: #lld-macho, int3
Differential Revision: https://reviews.llvm.org/D129363
* Refactor compression namespaces across the project, making way for a possible
introduction of alternatives to zlib compression.
Changes are as follows:
* Relocate the `llvm::zlib` namespace to `llvm::compression::zlib`.
Reviewed By: MaskRay, leonardchan, phosek
Differential Revision: https://reviews.llvm.org/D128953
A pair of auipc+jalr relocated by R_RISCV_CALL or R_RISCV_CALL_PLT can be
converted to c.j, c.jal, or jal.
* c.j: RVC and displacement is representable as an int12
* c.jal: RV32C and displacement is representable as an int12
* jal: displacement is representable as an int21
Use the D127581 relaxation framework to implement the relaxation. If a shorter
sequence is satisfied, we record the new relocation type in `relocTypes` and
saves the new instruction into `writes`. Finally let `riscvFinalizeRelax` rewrite the
instruction by setting `skip`.
Differential Revision: https://reviews.llvm.org/D127611
Alternative to D125036. Implement R_RISCV_ALIGN relaxation so that we can handle
-mrelax object files (i.e. -mno-relax is no longer needed) and creates a
framework for future relaxation.
`relaxAux` is placed in a union with InputSectionBase::jumpInstrMod, storing
auxiliary information for relaxation. In the first pass, `relaxAux` is allocated.
The main data structure is `relocDeltas`: when referencing `relocations[i]`, the
actual offset is `r_offset - (i ? relocDeltas[i-1] : 0)`.
`relaxOnce` performs one relaxation pass. It computes `relocDeltas` for all text
section. Then, adjust st_value/st_size for symbols relative to this section
based on `SymbolAnchor`. `bytesDropped` is set so that `assignAddresses` knows
that the size has changed.
Run `relaxOnce` in the `finalizeAddressDependentContent` loop to wait for
convergence of text sections and other address dependent sections (e.g.
SHT_RELR). Note: extrating `relaxOnce` into a separate loop works for many cases
but has issues in some linker script edge cases.
After convergence, compute section contents: shrink the NOP sequence of each
R_RISCV_ALIGN as appropriate. Instead of deleting bytes, we run a sequence of
memcpy on the content delimitered by relocation locations. For R_RISCV_ALIGN let
the next memcpy skip the desired number of bytes. Section content computation is
parallelizable, but let's ensure the implementation is mature before
optimizations. Technically we can save a copy if we interleave some code with
`OutputSection::writeTo`, but let's not pollute the generic code (we don't have
templated relocation resolving, so using conditions can impose overhead to
non-RISCV.)
Tested:
`make ARCH=riscv CROSS_COMPILE=riscv64-linux-gnu- LLVM=1 defconfig all` built Linux kernel using -mrelax is bootable.
FreeBSD RISCV64 system using -mrelax is bootable.
bash/curl/firefox/libevent/vim/tmux using -mrelax works.
Differential Revision: https://reviews.llvm.org/D127581
Allows specific “temps” to be saved, instead of the current all-or-nothing nature of --save-temps. Multiple of these “temps” can be saved by specifying the argument multiple times.
Differential Revision: https://reviews.llvm.org/D127778
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
Add FORCE_LLD_DIAGNOSTICS_CRASH inspired by the existing
FORCE_CLANG_DIAGNOSTICS_CRASH.
This is particularly useful for people customizing LLD as they may
want to modify the crash reporting behavior.
Differential Revision: https://reviews.llvm.org/D128195
This hint instructs the linker to perform the AdrpLdr or AdrpAdd
transformation depending on whether the GOT load has been relaxed to
load a local symbol's address.
Differential Revision: https://reviews.llvm.org/D129059
This linker optimization hint transforms a pair of adrp+ldr (immediate)
instructions into an ldr (literal) load from a PC-relative address if
it is 4-byte aligned and within +/- 1 MiB, as ldr can encode a signed
19-bit offset that gets multiplied by 4.
In the wild, only a small number of these hints are applicable because
not many loads end up close enough to the data segment. However, the
added helper functions will be useful in implementing the rest of the
LOH types.
Differential Revision: https://reviews.llvm.org/D128942
Linker optimization hints mark a sequence of instructions used for
synthesizing an address, like ADRP+ADD. If the referenced symbol ends up
close enough, it can be replaced by a faster sequence of instructions
like ADR+NOP.
This commit adds support for 2 of the 7 defined ARM64 optimization
hints:
- LOH_ARM64_ADRP_ADD, which transforms a pair of ADRP+ADD into ADR+NOP
if the referenced address is within +/- 1 MiB
- LOH_ARM64_ADRP_ADRP, which transforms two ADRP instructions into
ADR+NOP if they reference the same page
These two kinds already cover more than 50% of all LOHs in
chromium_framework.
Differential Review: https://reviews.llvm.org/D128093
An ADD_ADDR rebase opcode's argument can be encoded as an immediate if
the offset is less than 15 * word size. This change reduces the size of
chromium_framework by 100+ KiB.
Differential Revision: https://reviews.llvm.org/D128798
Instead, export `__wasm_apply_data_relocs` and `__wasm_call_ctors`
separately.
This is required since user code in a shared library (such as static
constructors) should not be run until relocations have been applied to
all loaded libraries.
See: https://github.com/emscripten-core/emscripten/issues/17295
Differential Revision: https://reviews.llvm.org/D128515
-dc is deprecated in release/14.x. Remove it for 15.0.
The only usage I know was FreeBSD crungen which was removed by https://reviews.freebsd.org/D34215
glibc just dropped -Wl,-d today. Keep -d for now.
For 1 != 1 <= 1 ? 1 : 2, the current code incorrectly considers that ?
has a higher precedence than != (minPrec).
Also, add a test for right associativity.
GOT references to absolute symbols can't be relaxed to use ADRP/ADD in
position-independent code because these instructions produce a relative
address.
Differential Revision: https://reviews.llvm.org/D128492
`--time-trace=foo` has the same behavior as `--time-trace --time-trace-file=<file>`
had previously.
Also, for mac, make --time-trace-granularity *not* imply --time-trace, to match
behavior of the ELF port.
Differential Revision: https://reviews.llvm.org/D128451
Allows ThinLTO indices to be written to disk on-the-fly/as-part-of “normal” linker execution. Previously ThinLTO indices could be written via --thinlto-index-only but that would cause the linker to exit early. For MLGO specifically, this enables saving the ThinLTO index files without having to restart the linker to collect data only available at later stages (i.e. output of --save-temps) of the linker's execution.
Note, this option does not currently work with:
--thinlto-object-suffix-replace, as this is intended to be used to consume minimized IR bitcode files while --thinlto-emit-index-files is intended to be run together with InProcessThinLTO (which cannot parse minimized IR).
--thinlto-prefix-replace support is left unimplemented but can be implemented if needed
Differential Revision: https://reviews.llvm.org/D127777
Identical literal folding takes ~1.4% of the time, and was missing
from the trace.
Signature computation still needs ~2.2% of the time, so probably worth
explicitly marking its contribution to "Write output file" (9.1%)
Differential Revision: https://reviews.llvm.org/D128343
Similarly to how undefined symbol diagnostics were changed in D128184,
we now show where in the source file duplicate symbols are defined at:
ld64.lld: error: duplicate symbol: _foo
>> defined in bar.c:42
>> /path/to/bar.o
>> defined in baz.c:1
>> /path/to/libbaz.a(baz.o)
For objects that don't contain DWARF data, the format is unchanged.
A slight difference to undefined symbol diagnostics is that we don't
print the name of the symbol on the third line, as it's already
contained on the first line.
Differential Revision: https://reviews.llvm.org/D128425
According to ministat, this is a small but measurable speedup
(using the repro in PR56121):
N Min Max Median Avg Stddev
x 10 3.7439518 3.7783802 3.7730219 3.7655502 0.012375226
+ 10 3.6149218 3.692198 3.6519327 3.6502951 0.025905601
Difference at 95.0% confidence
-0.115255 +/- 0.0190746
-3.06078% +/- 0.506554%
(Student's t, pooled s = 0.0203008)
(Without 858e8b17f7, this change here to use parallelFor is an 18% speedup,
and doing 858e8b17f7 on top of this change is just a 2.55% +/- 0.58% win.
Doing both results in a total speedup of 20.85% +/- 0.44%.)
Differential Revision: https://reviews.llvm.org/D128298
The error used to look like this:
ld64.lld: error: undefined symbol: _foo
>>> referenced by /path/to/bar.o:(symbol _baz+0x4)
If DWARF line information is available, we now show where in the source
the references are coming from:
ld64.lld: error: unreferenced symbol: _foo
>>> referenced by: bar.cpp:42 (/path/to/bar.cpp:42)
>>> /path/to/bar.o:(symbol _baz+0x4)
The reland is identical to the first time this landed. The fix was in D128294.
This reverts commit 0cc7ad4175.
Differential Revision: https://reviews.llvm.org/D128184
If ld64.lld was supplied an object file that had a `__debug_abbrev` or
`__debug_str` section, but didn't have any compile unit DIEs in
`__debug_info`, it would dereference an iterator pointing to the empty
array of DIEs. This underlying issue started causing segmentation faults
when parsing for `__debug_info` was addded in D128184. That commit was
reverted, and this one fixes the invalid dereference to allow relanding
it.
This commit adds an assertion to `filter_iterator_base`'s dereference
operators to catch bugs like this one.
Ran check-llvm, check-clang and check-lld.
Differential Revision: https://reviews.llvm.org/D128294
It's in libSystem, so it doesn't bring in any new deps, and it's
currently much faster than LLVM's current SHA256 implementation.
Makes linking (arm64) Chromium Framework with ld64.lld 17% faster.
See also PR56121.
No behavior change.
Differential Revision: https://reviews.llvm.org/D128290
This reverts commit 9ffeaaa0ea.
This fixes debugging large executables with lldb and gdb.
When StringTableBuilder is used, the string offsets for any string
can point anywhere in the string table - while previously, all strings
were inserted in order (without deduplication and tail merging).
For symbols, there's no complications in encoding the string offset;
the offset is encoded as a raw 32 bit binary number in half of the
symbol name field.
For sections, the string table offset is written as
"/<decimaloffset>", but if the decimal offset would be larger than
7 digits, it's instead written as "//<base64offset>". Tools that
operate on object files can handle the base64 offset format, but
apparently neither lldb nor gdb expect that syntax when locating the
debug information section. Prior to the reverted commit, all long
section names were located at the start of the string table, so
their offset never exceeded the range for the decimal syntax.
Just reverting this change for now, as the actual benefit from it
was fairly modest.
Longer term, lld could write all long section names unoptimized
at the start of the string table, followed by all the strings for
symbol names, with deduplication and tail merging. And lldb and
gdb could be fixed to handle sections with the base64 offset syntax.
This fixes https://github.com/mstorsjo/llvm-mingw/issues/289.
Microsoft does not seem to document the flag. Ignoring it for now is probably
better than getting an unknown flag error.
Reviewed By: thakis
Differential Revision: https://reviews.llvm.org/D128231
The error used to look like this:
ld64.lld: error: undefined symbol: _foo
>>> referenced by /path/to/bar.o:(symbol _baz+0x4)
If DWARF line information is available, we now show where in the source
the references are coming from:
ld64.lld: error: unreferenced symbol: _foo
>>> referenced by: bar.cpp:42 (/path/to/bar.cpp:42)
>>> /path/to/bar.o:(symbol _baz+0x4)
Differential Revision: https://reviews.llvm.org/D128184
Patch created by running:
rg -l parallelForEachN | xargs sed -i '' -c 's/parallelForEachN/parallelFor/'
No behavior change.
Differential Revision: https://reviews.llvm.org/D128140
I realized we'd forgotten to cover this case (though our existing
behavior is indeed correct / matches ld64's).
Reviewed By: #lld-macho, thakis
Differential Revision: https://reviews.llvm.org/D128025
As an optimization for ld64 sometimes it can be useful to not export any
symbols for top level binaries that don't need any exports, to do this
you can pass `-exported_symbols_list /dev/null`, or new with Xcode 14
(ld64 816) there is a `-no_exported_symbols` flag for the same behavior.
This reproduces this behavior where previously an empty exported symbols
list file would have been ignored.
Differential Revision: https://reviews.llvm.org/D127562
It seems to be a bug in `LinkerDriver::findFile`, the file name is not converted
to lowercase when being inserted into `visitedLibs`. This is the only exception
in the file and all other places always convert file names to lowercase when
inserting them into `visitedLibs` (or `visitedFiles`).
Reviewed By: thieta, hans
Differential Revision: https://reviews.llvm.org/D127709
Since binutils 2.36, GNU ld defaults to emitting base relocations,
and that version added the new option --disable-reloc-section to
disable it.
Differential Revision: https://reviews.llvm.org/D127478
ld64.lld used to print the "undefined symbol" line for each reference to
an undefined symbol previously:
ld64.lld: error: undefined symbol: _foo
>>> referenced by /path/to/bar.o:(symbol _baz+0x0)
ld64.lld: error: undefined symbol: _foo
>>> referenced by /path/to/bar.o:(symbol _quux+0x1)
Now they are deduplicated:
ld64.lld: error: undefined symbol: _foo
>>> referenced by /path/to/bar.o:(symbol _baz+0x0)
>>> referenced by /path/to/bar.o:(symbol _quux+0x1)
As with the other lld ports, only the first 3 references are printed.
Differential Revision: https://reviews.llvm.org/D127753
The error used to look like this:
ld64.lld: error: undefined symbol: _foo
>>> referenced by /path/to/bar.o
Now it displays the name of the function that contains the undefined
reference as well:
ld64.lld: error: undefined symbol: _foo
>>> referenced by /path/to/bar.o:(symbol _baz+0x4)
Differential Revision: https://reviews.llvm.org/D127696
This commit fixes the issue that getLocation always printed the name of
the first symbol in the section.
For clarity, upper_bound is used instead of a linear search for finding
the closest symbol name. Note that this change does not affect
performance: this function is only called when printing errors and
`symbols` typically contains a single symbol because of
.subsections_via_symbols.
Differential Revision: https://reviews.llvm.org/D127670
This reverts commit 942f4e3a7c.
The additional change required to avoid the assertion errors seen
previously is:
--- a/lld/MachO/ICF.cpp
+++ b/lld/MachO/ICF.cpp
@@ -443,7 +443,9 @@ void macho::foldIdenticalSections() {
/*relocVA=*/0);
isec->data = copy;
}
- } else {
+ } else if (!isEhFrameSection(isec)) {
+ // EH frames are gathered as hashables from unwindEntry above; give a
+ // unique ID to everything else.
isec->icfEqClass[0] = ++icfUniqueID;
}
}
Differential Revision: https://reviews.llvm.org/D123435
Just matter of enabling the config option.
(Also changed the platform of the input test file to macOS, since that's
the default that we specify in the `%lld` substitution. The conflict was
causing errors when linking with LTO.)
Reviewed By: #lld-macho, thakis
Differential Revision: https://reviews.llvm.org/D127600
This flag suppresses warnings produced by the linker. In ld64 this has
an interesting interaction with -fatal_warnings, it silences the
warnings but the link still fails. Instead of doing that here we still
print the warning and eagerly fail the link in case both are passed,
this seems more reasonable so users can understand why the link fails.
Differential Revision: https://reviews.llvm.org/D127564
First of all, `LLVM_TOOLS_INSTALL_DIR` put there breaks our NixOS
builds, because `LLVM_TOOLS_INSTALL_DIR` defined the same as
`CMAKE_INSTALL_BINDIR` becomes an *absolute* path, and then when
downstream projects try to install there too this breaks because our
builds always install to fresh directories for isolation's sake.
Second of all, note that `LLVM_TOOLS_INSTALL_DIR` stands out against the
other specially crafted `LLVM_CONFIG_*` variables substituted in
`llvm/cmake/modules/LLVMConfig.cmake.in`.
@beanz added it in d0e1c2a550 to fix a
dangling reference in `AddLLVM`, but I am suspicious of how this
variable doesn't follow the pattern.
Those other ones are carefully made to be build-time vs install-time
variables depending on which `LLVMConfig.cmake` is being generated, are
carefully made relative as appropriate, etc. etc. For my NixOS use-case
they are also fine because they are never used as downstream install
variables, only for reading not writing.
To avoid the problems I face, and restore symmetry, I deleted the
exported and arranged to have many `${project}_TOOLS_INSTALL_DIR`s.
`AddLLVM` now instead expects each project to define its own, and they
do so based on `CMAKE_INSTALL_BINDIR`. `LLVMConfig` still exports
`LLVM_TOOLS_BINARY_DIR` which is the location for the tools defined in
the usual way, matching the other remaining exported variables.
For the `AddLLVM` changes, I tried to copy the existing pattern of
internal vs non-internal or for LLVM vs for downstream function/macro
names, but it would good to confirm I did that correctly.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D117977
This silences the following warning:
../tools/lld/ELF/SyntheticSections.cpp:1596:48: warning: suggest parentheses around ‘&&’ within ‘||’ [-Wparentheses]
1596 | assert((index != 0 || type != target->gotRel && type != target->pltRel ||
Differential Revision: https://reviews.llvm.org/D127395
For arm64, llvm-mc emits relocations for the target function
address like so:
ltmp:
<CIE start>
...
<CIE end>
... multiple FDEs ...
<FDE start>
<target function address - (ltmp + pcrel offset)>
...
If any of the FDEs in `multiple FDEs` get dead-stripped, then `FDE start`
will move to an earlier address, and `ltmp + pcrel offset` will no longer
reflect an accurate pcrel value. To avoid this problem, we "canonicalize"
our relocation by adding an `EH_Frame` symbol at `FDE start`, and updating
the reloc to be `target function address - (EH_Frame + new pcrel offset)`.
Reviewed By: #lld-macho, Roger
Differential Revision: https://reviews.llvm.org/D124561
== Background ==
`llvm-mc` generates unwind info in both compact unwind and DWARF
formats. LLD already handles the compact unwind format; this diff gets
us close to handling the DWARF format properly.
== Caveats ==
It's not quite done yet, but I figure it's worth getting this reviewed
and landed first as it's shaping up to be a fairly large code change.
**Known limitations of the current code:**
* Only works for x86_64, for which `llvm-mc` emits "abs-ified"
relocations as described in 618def651b.
`llvm-mc` emits regular relocations for ARM EH frames, which we do not
yet handle correctly.
Since the feature is not ready for real use yet, I've gated it behind a
flag that only gets toggled on during test suite runs. With most of the
new code disabled, we see just a hint of perf regression, so I don't
think it'd be remiss to land this as-is:
base diff difference (95% CI)
sys_time 1.926 ± 0.168 1.979 ± 0.117 [ -1.2% .. +6.6%]
user_time 3.590 ± 0.033 3.606 ± 0.028 [ +0.0% .. +0.9%]
wall_time 7.104 ± 0.184 7.179 ± 0.151 [ -0.2% .. +2.3%]
samples 30 31
== Design ==
Like compact unwind entries, EH frames are also represented as regular
ConcatInputSections that get pointed to via `Defined::unwindEntry`. This
allows them to be handled generically by e.g. the MarkLive and ICF
code. (But note that unlike compact unwind subsections, EH frame
subsections do end up in the final binary.)
In order to make EH frames "look like" a regular ConcatInputSection,
some processing is required. First, we need to split the `__eh_frame`
section along EH frame boundaries rather than along symbol boundaries.
We do this by decoding the length field of each EH frame. Second, the
abs-ified relocations need to be turned into regular Relocs.
== Next Steps ==
In order to support EH frames on ARM targets, we will either have to
teach LLD how to handle EH frames with explicit relocs, or we can try to
make `llvm-mc` emit abs-ified relocs for ARM as well. I'm hoping to do
the latter as I think it will make the LLD implementation both simpler
and faster to execute.
== Misc ==
The `obj-file-with-stabs.s` test had to be updated as the previous
version would trip assertion errors in the code. It appears that in our
attempt to produce a minimal YAML test input, we created a file with
invalid EH frame data. I've fixed this by re-generating the YAML and not
doing any hand-pruning of it.
Reviewed By: #lld-macho, Roger
Differential Revision: https://reviews.llvm.org/D123435
This reduces linking time by ~8% for my project (1.19s -> 0.53s for
writeSections()). writeTo is const, which bodes well for it being
parallelizable, and I've looked through the different overridden versions and
can't see any race conditions. It produces the same byte-for-byte output for my
project.
Differential Revision: https://reviews.llvm.org/D126800
This reverts commit dcf3368e33.
It breaks -DLLVM_ENABLE_ASSERTIONS=on builds. In addition, the description is
incorrect about ld.lld behavior. For wasm, there should be justification to add
the new mode.
As well as ELF linker does, retain all data segments named X referenced
through `__start_X` or `__stop_X`.
For example, `FOO_MD` should not be stripped in the below case, but it's currently mis-stripped
```llvm
@FOO_MD = global [4 x i8] c"bar\00", section "foo_md", align 1
@__start_foo_md = external constant i8*
@__stop_foo_md = external constant i8*
@llvm.used = appending global [1 x i8*] [i8* bitcast (i32 ()* @foo_md_size to i8*)], section "llvm.metadata"
define i32 @foo_md_size() {
entry:
ret i32 sub (
i32 ptrtoint (i8** @__stop_foo_md to i32),
i32 ptrtoint (i8** @__start_foo_md to i32)
)
}
```
This fixes https://github.com/llvm/llvm-project/issues/55839
Reviewed By: sbc100
Differential Revision: https://reviews.llvm.org/D126950
.zdebug is unlikely used any longer: gcc -gz switched from legacy
.zdebug to SHF_COMPRESSED with binutils 2.26 (2016), which has been
several years. clang 14 dropped -gz=zlib-gnu support. According to
Debian Code Search (`gz=zlib-gnu`), no project uses -gz=zlib-gnu.
Remove .zdebug support to (a) simplify code and (b) allow removal of llvm-mc's
--compress-debug-sections=zlib-gnu.
In case the old object file `a.o` uses .zdebug, run `objcopy --decompress-debug-sections a.o`
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D126793
LTO code may end up mixing bitcode files from various sources varying in
their use of opaque pointer types. The current strategy to decide
between opaque / typed pointers upon the first bitcode file loaded does
not work here, since we could be loading a non-opaque bitcode file first
and would then be unable to load any files with opaque pointer types
later.
So for LTO this:
- Adds an `lto::Config::OpaquePointer` option and enforces an upfront
decision between the two modes.
- Adds `-opaque-pointers`/`-no-opaque-pointers` options to the gold
plugin; disabled by default.
- `--opaque-pointers`/`--no-opaque-pointers` options with
`-plugin-opt=-opaque-pointers`/`-plugin-opt=-no-opaque-pointers`
aliases to lld; disabled by default.
- Adds an `-lto-opaque-pointers` option to the `llvm-lto2` tool.
- Changes the clang driver to pass `-plugin-opt=-opaque-pointers` to
the linker in LTO modes when clang was configured with opaque
pointers enabled by default.
This fixes https://github.com/llvm/llvm-project/issues/55377
Differential Revision: https://reviews.llvm.org/D125847
This reduces the time emitStabs() takes by about 275ms, or 3% of overall
linking time for the project I'm on. Although the parent function is run in
parallel, it's one of the slowest tasks in that concurrent batch (I have
another optimization for another slow task as well).
Differential Revision: https://reviews.llvm.org/D126785
Symbols from LTO objects don't contain Wasm signatures, but we need a
signature when we create undefined/stub functions for missing weakly
undefined symbols.
Luckily, after LTO, we know that symbols that are not referenced by a
regular object file must not be needed in the final output so there
is no need to generate undefined/stub function for them.
Differential Revision: https://reviews.llvm.org/D126554
Currently there are 2 duplicate implementation, and I want to add
a use in a 3rd place. Combine them in lib/BinaryFormat so they can
be shared.
Also update toString for symbol and reloc types to use StringRef
Differential Revision: https://reviews.llvm.org/D126553
Fangrui Song