Our implementation draws heavily from LLD-ELF's, which in turn delegates
its string deduplication to llvm-mc's StringTableBuilder. The messiness of
this diff is largely due to the fact that we've previously assumed that
all InputSections get concatenated together to form the output. This is
no longer true with CStringInputSections, which split their contents into
StringPieces. StringPieces are much more lightweight than InputSections,
which is important as we create a lot of them. They may also overlap in
the output, which makes it possible for strings to be tail-merged. In
fact, the initial version of this diff implemented tail merging, but
I've dropped it for reasons I'll explain later.
**Alignment Issues**
Mergeable cstring literals are found under the `__TEXT,__cstring`
section. In contrast to ELF, which puts strings that need different
alignments into different sections, clang's Mach-O backend puts them all
in one section. Strings that need to be aligned have the `.p2align`
directive emitted before them, which simply translates into zero padding
in the object file.
I *think* ld64 extracts the desired per-string alignment from this data
by preserving each string's offset from the last section-aligned
address. I'm not entirely certain since it doesn't seem consistent about
doing this; but perhaps this can be chalked up to cases where ld64 has
to deduplicate strings with different offset/alignment combos -- it
seems to pick one of their alignments to preserve. This doesn't seem
correct in general; we can in fact can induce ld64 to produce a crashing
binary just by linking in an additional object file that only contains
cstrings and no code. See PR50563 for details.
Moreover, this scheme seems rather inefficient: since unaligned and
aligned strings are all put in the same section, which has a single
alignment value, it doesn't seem possible to tell whether a given string
doesn't have any alignment requirements. Preserving offset+alignments
for strings that don't need it is wasteful.
In practice, the crashes seen so far seem to stem from x86_64 SIMD
operations on cstrings. X86_64 requires SIMD accesses to be
16-byte-aligned. So for now, I'm thinking of just aligning all strings
to 16 bytes on x86_64. This is indeed wasteful, but implementation-wise
it's simpler than preserving per-string alignment+offsets. It also
avoids the aforementioned crash after deduplication of
differently-aligned strings. Finally, the overhead is not huge: using
16-byte alignment (vs no alignment) is only a 0.5% size overhead when
linking chromium_framework.
With these alignment requirements, it doesn't make sense to attempt tail
merging -- most strings will not be eligible since their overlaps aren't
likely to start at a 16-byte boundary. Tail-merging (with alignment) for
chromium_framework only improves size by 0.3%.
It's worth noting that LLD-ELF only does tail merging at `-O2`. By
default (at `-O1`), it just deduplicates w/o tail merging. @thakis has
also mentioned that they saw it regress compressed size in some cases
and therefore turned it off. `ld64` does not seem to do tail merging at
all.
**Performance Numbers**
CString deduplication reduces chromium_framework from 250MB to 242MB, or
about a 3.2% reduction.
Numbers for linking chromium_framework on my 3.2 GHz 16-Core Intel Xeon W:
N Min Max Median Avg Stddev
x 20 3.91 4.03 3.935 3.95 0.034641016
+ 20 3.99 4.14 4.015 4.0365 0.0492336
Difference at 95.0% confidence
0.0865 +/- 0.027245
2.18987% +/- 0.689746%
(Student's t, pooled s = 0.0425673)
As expected, cstring merging incurs some non-trivial overhead.
When passing `--no-literal-merge`, it seems that performance is the
same, i.e. the refactoring in this diff didn't cost us.
N Min Max Median Avg Stddev
x 20 3.91 4.03 3.935 3.95 0.034641016
+ 20 3.89 4.02 3.935 3.9435 0.043197831
No difference proven at 95.0% confidence
Reviewed By: #lld-macho, gkm
Differential Revision: https://reviews.llvm.org/D102964
When a library "host"'s reexports change their installName with
`$ld$os10.11$install_name$host`, we used to write a load command for "host" but
write the version numbers of the reexport instead of "host". This fixes that.
I first thought that the rule is to take the version numbers from the library
that originally had that install name (implemented in D103819), but that's not
what ld64 seems to be doing: It takes the version number from the first dylib
with that install name it loads, and it loads the reexporting library before
the reexports. We already did most of that, we just added reexports before the
reexporter. After this change, we add the reexporter before the reexports.
Addresses https://bugs.llvm.org/show_bug.cgi?id=49800#c11 part 1.
(ld64 seems to add reexports after processing _all_ files on the command line,
while we add them right after the reexporter. For the common case of reexport +
$ld$ symbol changing back to the exporter name, this doesn't make a difference,
but you can construct a case where it does. I expect this to not make a
difference in practice though.)
Differential Revision: https://reviews.llvm.org/D103821
Our behavior here already matched ld64, now we have a test for it.
(ld64 even strips the library here if you also pass -needed_library bar.dylib.
That seems wrong to me, and lld honors needed_library in that case.)
Differential Revision: https://reviews.llvm.org/D103812
Also adjust a few comments, and move the DylibFile comment talking about
umbrella next to the parameter again.
Differential Revision: https://reviews.llvm.org/D103783
This diff adds first bits to support special symbols $ld$previous* in LLD.
$ld$* symbols modify properties/behavior of the library
(e.g. its install name, compatibility version or hide/add symbols)
for specific target versions.
Test plan: make check-lld-macho
Differential revision: https://reviews.llvm.org/D103505
D103423 neglected to call `parseReexports()` for nested TBD
documents, leading to symbol resolution failures when trying to look up
a symbol nested more than one level deep in a TBD file. This fixes the
regression and adds a test.
It also appears that `umbrella` wasn't being set properly when calling
`parseLoadCommands` -- it's supposed to resolve to `this` if `nullptr`
is passed. I didn't write a failing test case for this but I've made
`umbrella` a member so the previous behavior should be preserved.
Reviewed By: #lld-macho, thakis
Differential Revision: https://reviews.llvm.org/D103586
Make sure that comdat symbols also have a non-null dummy
SectionChunk associated.
This requires moving around an existing FIXME regarding comdats in
LTO.
Differential Revision: https://reviews.llvm.org/D103012
Also adds support for live_support sections, no_dead_strip sections,
.no_dead_strip symbols.
Chromium Framework 345MB unstripped -> 250MB stripped
(vs 290MB unstripped -> 236M stripped with ld64).
Doing dead stripping is a bit faster than not, because so much less
data needs to be processed:
% ministat lld_*
x lld_nostrip.txt
+ lld_strip.txt
N Min Max Median Avg Stddev
x 10 3.929414 4.07692 4.0269079 4.0089678 0.044214794
+ 10 3.8129408 3.9025559 3.8670411 3.8642573 0.024779651
Difference at 95.0% confidence
-0.144711 +/- 0.0336749
-3.60967% +/- 0.839989%
(Student's t, pooled s = 0.0358398)
This interacts with many parts of the linker. I tried to add test coverage
for all added `isLive()` checks, so that some test will fail if any of them
is removed. I checked that the test expectations for the most part match
ld64's behavior (except for live-support-iterations.s, see the comment
in the test). Interacts with:
- debug info
- export tries
- import opcodes
- flags like -exported_symbol(s_list)
- -U / dynamic_lookup
- mod_init_funcs, mod_term_funcs
- weak symbol handling
- unwind info
- stubs
- map files
- -sectcreate
- undefined, dylib, common, defined (both absolute and normal) symbols
It's possible it interacts with more features I didn't think of,
of course.
I also did some manual testing:
- check-llvm check-clang check-lld work with lld with this patch
as host linker and -dead_strip enabled
- Chromium still starts
- Chromium's base_unittests still pass, including unwind tests
Implemenation-wise, this is InputSection-based, so it'll work for
object files with .subsections_via_symbols (which includes all
object files generated by clang). I first based this on the COFF
implementation, but later realized that things are more similar to ELF.
I think it'd be good to refactor MarkLive.cpp to look more like the ELF
part at some point, but I'd like to get a working state checked in first.
Mechanical parts:
- Rename canOmitFromOutput to wasCoalesced (no behavior change)
since it really is for weak coalesced symbols
- Add noDeadStrip to Defined, corresponding to N_NO_DEAD_STRIP
(`.no_dead_strip` in asm)
Fixes PR49276.
Differential Revision: https://reviews.llvm.org/D103324
We used to not print dylibs referenced by other dylibs in `-t` mode. This
affected reexports, and with `-flat_namespace` also just dylibs loaded by
dylibs. Now we print them.
Fixes PR49514.
Differential Revision: https://reviews.llvm.org/D103428
In some cases, we end up with several distinct DylibFiles that
have the same install name. Only emit a single LC_LOAD_DYLIB in
those cases.
This happens in 3 cases I know of:
1. Some tbd files are symlinks. libpthread.tbd is a symlink against
libSystem.tbd for example, so `-lSystem -lpthread` loads
libSystem.tbd twice. We could (and maybe should) cache loaded
dylibs by realpath() to catch this.
2. Some tbd files are copies of each other. For example,
CFNetwork.framework/CFNetwork.tbd and
CFNetwork.framework/Versions/A/CFNetwork.tbd are two distinct
copies of the same file. The former is found by
`-framework CFNetwork` and the latter by the reexport in
CoreServices.tbd. We could conceivably catch this by
making `-framework` search look in `Versions/Current` instead
of in the root, and/or by using a content hash to cache
tbd files, but that's starting to sound complicated.
3. Magic $ld$ symbol processing can change the install name of
a dylib based on the target platform_version. Here, two
truly distinct dylibs can have the same install name.
So we need this code to deal with (3) anyways. Might as well use
it for 1 and 2, at least for now :)
With this (and D103430), clang-format links in the same dylibs
when linked with lld and ld64.
Differential Revision: https://reviews.llvm.org/D103488
This omits load commands for unreferenced dylibs if:
- the dylib was loaded implicitly,
- it is marked MH_DEAD_STRIPPABLE_DYLIB
- or -dead_strip_dylibs is passed
This matches ld64.
Currently, the "is dylib referenced" state is computed before dead code
stripping and is not updated after dead code stripping. This too matches ld64.
We should do better here.
With this, clang-format linked with lld (like with ld64) no longer has
libobjc.A.dylib in `otool -L` output. (It was implicitly loaded as a reexport
of CoreFoundation.framework, but it's not needed.)
Differential Revision: https://reviews.llvm.org/D103430
.s files with `-g` generate __debug_aranges on darwin/arm64 for some
reason, and those lead to `nullptr` symbols. Don't crash on that.
Fixes PR50517.
Differential Revision: https://reviews.llvm.org/D103350
Ghashing is probably going to be faster in most cases, even without
precomputed ghashes in object files.
Here is my table of results linking clang.pdb:
-------------------------------
| threads | GHASH | NOGHASH |
-------------------------------
| j1 | 51.031s | 25.141s |
| j2 | 31.079s | 22.109s |
| j4 | 18.609s | 23.156s |
| j8 | 11.938s | 21.984s |
| j28 | 8.375s | 18.391s |
-------------------------------
This shows that ghashing is faster if at least four cores are available.
This may make the linker slower if most cores are busy in the middle of
a build, but in that case, the linker probably isn't on the critical
path of the build. Incremental build performance is arguably more
important than highly contended batch build link performance.
The -time output indicates that ghash computation is the dominant
factor:
Input File Reading: 924 ms ( 1.8%)
GC: 689 ms ( 1.3%)
ICF: 527 ms ( 1.0%)
Code Layout: 414 ms ( 0.8%)
Commit Output File: 24 ms ( 0.0%)
PDB Emission (Cumulative): 49938 ms ( 94.8%)
Add Objects: 46783 ms ( 88.8%)
Global Type Hashing: 38983 ms ( 74.0%)
GHash Type Merging: 5640 ms ( 10.7%)
Symbol Merging: 2154 ms ( 4.1%)
Publics Stream Layout: 188 ms ( 0.4%)
TPI Stream Layout: 18 ms ( 0.0%)
Commit to Disk: 2818 ms ( 5.4%)
--------------------------------------------------
Total Link Time: 52669 ms (100.0%)
We can speed that up with a faster content hash (not SHA1).
Differential Revision: https://reviews.llvm.org/D102888
This diff paves the way for {D102964} which adds a new kind of
InputSection.
We previously maintained section ordering implicitly: we created
InputSections as we parsed each file in command-line order, and passed
on this ordering when we created OutputSections and OutputSegments by
iterating over these InputSections. The implicitness of the ordering
made it difficult to refactor the code to e.g. handle a new type of
InputSection. As such, I've codified the ordering explicitly via
`inputOrder` fields. This also allows us to use `sort` instead of
`stable_sort`.
Benchmarking chromium_framework on my 3.2 GHz 16-Core Intel Xeon W:
N Min Max Median Avg Stddev
x 20 4.23 4.35 4.27 4.274 0.030157481
+ 20 4.24 4.38 4.27 4.2815 0.033759989
No difference proven at 95.0% confidence
Reviewed By: #lld-macho, alexshap
Differential Revision: https://reviews.llvm.org/D102972
Given the following scenario:
```
// Cat.cpp
struct Animal { virtual void makeNoise() const = 0; };
struct Cat : Animal { void makeNoise() const override; };
extern "C" int puts(char const *);
void Cat::makeNoise() const { puts("Meow"); }
void doThingWithCat(Animal *a) { static_cast<Cat *>(a)->makeNoise(); }
// CatUser.cpp
struct Animal { virtual void makeNoise() const = 0; };
struct Cat : Animal { void makeNoise() const override; };
void doThingWithCat(Animal *a);
void useDoThingWithCat() {
Cat *d = new Cat;
doThingWithCat(d);
}
// cat.ver
{
global: _Z17useDoThingWithCatv;
local: *;
};
$ clang++ Cat.cpp CatUser.cpp -fpic -flto=thin -fwhole-program-vtables
-shared -O3 -fuse-ld=lld -Wl,--lto-whole-program-visibility
-Wl,--version-script,cat.ver
```
We cannot devirtualize `Cat::makeNoise`. The issue is complex:
Due to `-fsplit-lto-unit` and usage of type metadata, we place the Cat
vtable declaration into module 0 and the Cat vtable definition with type
metadata into module 1, causing duplicate entries (Undefined followed by
Defined) in the `lto::InputFile::symbols()` output.
In `BitcodeFile::parse`, after processing the `Undefined` then the
`Defined`, the final state is `Defined`.
In `BitcodeCompiler::add`, for the first symbol, `computeBinding`
returns `STB_LOCAL`, then we reset it to `Undefined` because it is
prevailing (`versionId` is `preserved`). For the second symbol, because
the state is now `Undefined`, `computeBinding` returns `STB_GLOBAL`,
causing `ExportDynamic` to be true and suppressing devirtualization.
In D77280, the `computeBinding` change used a stricter `isDefined()`
condition to make weak``Lazy` symbol work.
This patch relaxes the condition to weaker `!isLazy()` to keep it
working while making the devirtualization work as well.
Differential Revision: https://reviews.llvm.org/D98686
If we support local signature symbols (PR43094), these tests would fail.
When the support is added, new tests (local signature symbol specific) should be developed.
Prior to this change build with `-shared/-pie` and using TLS (but
without -shared-memory) would hit this assert:
"Currenly only a single data segment is supported in PIC mode"
This is because we were not including TLS data when merging data
segments. However, when we build without shared-memory (i.e. without
threads) we effectively lower away TLS into a normal active data
segment.. so we were ending up with two active data segments: the merged
data, and the lowered TLS data.
To fix this problem we can instead avoid combining data segments at
all when running in shared memory mode (because in this case all
segment initialization is passive). And then in non-shared memory
mode we know that TLS has been lowered and therefore we can can
and should combine all segments.
So with this new behavior we have two different modes:
1. With shared memory / mutli-threaded: Never combine data segments
since it is not necessary. (All data segments as passive already).
2. Wihout shared memory / single-threaded: Combine *all* data segments
since we treat TLS as normal data. (We end up with a single
active data segment).
Differential Revision: https://reviews.llvm.org/D102937
The COFF driver produces an ABSOLUTE relocation base for an ADDR32
relocation type and the system is 64 bits (machine=AMD64). The
relocation information won't be added in the output and could
produce an incorrect address access during run-time. This change
set checks if the relocation type is IMAGE_REL_AMD64_ADDR32 and
if so, adds the relocated symbol as IMAGE_REL_BASED_HIGHLOW base.
Differential Revision: https://reviews.llvm.org/D96619
Previously we simply didn't check this. Prereq to make the test suite
pass with ghash enabled by default.
Differential Revision: https://reviews.llvm.org/D102885
__table_base is know 64-bit, since in LLVM it represents a function pointer offset
__table_base32 is a copy in wasm32 for use in elem init expr, since no truncation may be used there.
New reloc R_WASM_TABLE_INDEX_REL_SLEB64 added
Differential Revision: https://reviews.llvm.org/D101784
These symbols are long, and they tend to cause the PDB file size to
overflow. They are generally not necessary when debugging problems in
user code.
This change reduces the size of chrome.dll.pdb with coverage from
6,937,108,480 bytes to 4,690,210,816 bytes.
Differential Revision: https://reviews.llvm.org/D102719
lld/MachO/Driver.cpp and lld/MachO/SyntheticSections.cpp include
llvm/Config/config.h which doesn't exist when building standalone lld.
This patch replaces llvm/Config/config.h include with llvm/Config/llvm-config.h
just like it is in lld/ELF/Driver.cpp and HAVE_LIBXAR with LLVM_HAVE_LIXAR and
moves LLVM_HAVE_LIBXAR from config.h to llvm-config.h
Also it adds LLVM_HAVE_LIBXAR to LLVMConfig.cmake and links liblldMachO2.so
with XAR_LIB if LLVM_HAVE_LIBXAR is set.
Differential Revision: https://reviews.llvm.org/D102084
In LC_DYSYMTAB, private externs were still emitted as exported symbols instead
of as locals.
Fixes PR50373. See bug for details.
Differential Revision: https://reviews.llvm.org/D102662
Besides -Bdynamic and -Bstatic, ld documents additional aliases for both of these options. Instead of -Bstatic, one may write -dn, -non_shared or -static. Instead of -Bdynamic one may write -dy or -call_shared. Source: https://sourceware.org/binutils/docs-2.36/ld/Options.html
This patch adds those aliases to the MinGW driver of lld for the sake of ld compatibility.
Encountered this case while compiling a static Qt 6.1 distribution and got build failures as -static was passed directly to the linker, instead of through the compiler driver.
Differential Revision: https://reviews.llvm.org/D102637
Has the effect that `__mh_execute_header` stays in the symbol table of
outputs even after running `strip` on the output. I don't know if that's
important for anything -- my motivation for the patch is just is to make
the output more similar to ld64.
(Corresponds to symbolTableInAndNeverStrip in ld64.)
Differential Revision: https://reviews.llvm.org/D102619
D62727 removed GotEntrySize and GotPltEntrySize with a comment that they
are always equal to wordsize(), but that is not entirely true: X32 has a
word size of 4, but needs 8-byte GOT entries. This restores gotEntrySize
for both, adjusted for current naming conventions, but defaults it to
config->wordsize to keep things simple for architectures other than
x86_64.
This partially reverts D62727.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D102509
This option will be available in GNU ld 2.27 (https://sourceware.org/bugzilla/show_bug.cgi?id=27834).
This option can cancel previously specified -Bsymbolic and
-Bsymbolic-functions. This is useful for excluding some links when the
default uses -Bsymbolic-functions.
Reviewed By: jhenderson, peter.smith
Differential Revision: https://reviews.llvm.org/D102383
Previously there was no test checking that -Bsymbolic-functions only applies to STT_FUNC symbols.
Reviewed By: jhenderson
Differential Revision: https://reviews.llvm.org/D102461
This fixes a bug with string merging with string symbols that contain
NULLs, as is the case in the `merge-string.s` test.
The bug only showed when we run with `--relocatable` and then try read
the resulting object back in. In this case we would end up with string
symbols that extend past the end of the segment in which they live.
The problem comes from the fact that sections which are flagged as
string mergable assume that all strings are NULL terminated. The
merging algorithm will drop trailing chars that follow a NULL since they
are essentially unreachable. However, the "size" attribute (in the
symbol table) of such a truncated symbol is not updated resulting a
symbol size that can overlap the end of the segment.
I verified that this can happen in ELF too given the right conditions
and the its harmless enough. In practice Strings that contain embedded
null should not be part of a mergable section.
Differential Revision: https://reviews.llvm.org/D102281
Jez Ng