The new Darwin backend for LLD is now able to link reasonably large
real-world programs on x86_64. For instance, we have achieved
self-hosting for the X86_64 target, where all LLD tests pass when
building lld with itself on macOS. As such, we would like to make it the
default back-end.
The new port is now named `ld64.lld`, and the old port remains
accessible as `ld64.lld.darwinold`
This [annoucement email][1] has some context. (But note that, unlike
what the email says, we are no longer doing this as part of the LLVM 12
branch cut -- instead we will go into LLVM 13.)
Numerous mechanical test changes were required to make this change; in
the interest of creating something that's reviewable on Phabricator,
I've split out the boring changes into a separate diff (D95905). I plan to
merge its contents with those in this diff before landing.
(@gkm made the original draft of this diff, and he has agreed to let me
take over.)
[1]: https://lists.llvm.org/pipermail/llvm-dev/2021-January/147665.html
Reviewed By: #lld-macho, thakis
Differential Revision: https://reviews.llvm.org/D95204
With this, a simple hello world links against libSystem.tbd and the
old ld64.lld linker kind of works again with newer SDKs.
The motivation here is to have an arm64 cross linker that's good
enough to be able to run simple configure link checks on non-mac
systems for generating config.h files. Once -flavor darwinnew can
link arm64, we'll switch to that.
Drops the behavior from rL217112.
Use the Gnu driver mode by default for all platforms when ld is
invoked. Other names for the program (such as link or ld64) continue
working as before.
Reviewed By: MaskRay, srhines, smeenai, ruiu
Differential Revision: https://reviews.llvm.org/D78837
Summary:
The following patch avoids segfaulting if the section list is empty when writing a mach-o MH_OBJECT. I ran into this case from a more complicated example trying to dead_strip while using '-r' in lld.
I'm not sure if having empty sections is a legal mach-o, but it does seem that other llvm-binutils tools can ingest such a boring object with out issue. Would it be better to emit an error, emit a warning, or do nothing? It seems that adding a warning diagnostic might be helpful to users, as I did not expect to have a section-less object when the linker was done.
Reviewers: kledzik, ruiu
Subscribers: llvm-commits, jrm
Tags: #lld, #llvm
Differential Revision: https://reviews.llvm.org/D67735
llvm-svn: 372995
Also change some options that have different semantics (cause confusion) in llvm-readelf mode:
-s => -S
-t => --symbols
-sd => --section-data
llvm-svn: 359651
Patch by Nicholas Allegra.
The Mach-O writer calculates the size of load commands multiple times.
First, Util::assignAddressesToSections() (in MachONormalizedFileFromAtoms.cpp)
calculates the size using headerAndLoadCommandsSize() (in
MachONormalizedFileBinaryWriter.cpp), which creates a temporary
MachOFileLayout for the NormalizedFile, only to retrieve its
headerAndLoadCommandsSize. Later, writeBinary() (in
MachONormalizedFileBinaryWriter.cpp) creates a new layout and uses the offsets
from that layout to actually write out everything in the NormalizedFile.
But the NormalizedFile changes between the first computation and the second.
When Util::assignAddressesToSections is called, file.functionStarts is always
empty because Util::addFunctionStarts has not yet been called. Yet
MachOFileLayout decides whether to include a LC_FUNCTION_STARTS command based
on whether file.functionStarts is nonempty. Therefore, the initial computation
always omits it.
Because padding for the __TEXT segment (to make its size a multiple of the
page size) is added between the load commands and the first section, LLD still
generates a valid binary as long as the amount of padding happens to be large
enough to fit LC_FUNCTION_STARTS command, which it usually is.
However, it's easy to reproduce the issue by adding a section of a precise
size. Given foo.c:
__attribute__((section("__TEXT,__foo")))
char foo[0xd78] = {0};
Run:
clang -dynamiclib -o foo.dylib foo.c -fuse-ld=lld -install_name
/usr/lib/foo.dylib
otool -lvv foo.dylib
This should produce:
truncated or malformed object (offset field of section 1 in LC_SEGMENT_64
command 0 not past the headers of the file)
This commit:
- Changes MachOFileLayout to always assume LC_FUNCTION_STARTS is present for
the initial computation, as long as generating LC_FUNCTION_STARTS is
enabled. It would be slightly better to check whether there are actually
any functions, since no LC_FUNCTION_STARTS will be generated if not, but it
doesn't cause a problem if the initial computation is too high.
- Adds a test.
- Adds an assert in MachOFileLayout::writeSectionContent() that we are not
writing section content into the load commands region (which would happen
if the offset was calculated too low due to the initial load commands size
calculation being too low). Adds an assert in
MachOFileLayout::writeLoadCommands to validate a similar situation where
two size-of-load-commands computations are expected to be equivalent.
llvm-svn: 358545
If you have the string /usr/bin, prior to this patch it would not
be quoted by our YAML serializer. But a string like C:\src would
be, due to the presence of a backslash. This makes the quoting
rules of basically every single file path different depending on
the path syntax (posix vs. Windows).
While technically not required by the YAML specification to quote
forward slashes, when the behavior of paths is inconsistent it
makes it difficult to portably write FileCheck lines that will
work with either kind of path.
Differential Revision: https://reviews.llvm.org/D53169
llvm-svn: 344359
This reverts commit b86c16ad8c97dadc1f529da72a5bb74e9eaed344.
This is being reverted because I forgot to write a useful
commit message, so I'm going to resubmit it with an actual
commit message.
llvm-svn: 344358
Summary:
The issue with the python path is that the path to python on Windows can contain spaces. To make the tests always work, the path to python needs to be surrounded by quotes.
This is a companion change to: https://reviews.llvm.org/D50206
Reviewers: asmith, zturner, espindola
Subscribers: emaste, sbc100, arichardson, aheejin, steven_wu, dexonsmith, llvm-commits
Differential Revision: https://reviews.llvm.org/D50282
llvm-svn: 339075
Summary:
This fixes two failing tests on Windows with an installed version of python that has spaces in the path.
* elf/lto/cache.ll
* mach-o/dependency_info.yaml
Reviewers: zturner, llvm-commits, stella.stamenova
Subscribers: emaste, arichardson
Differential Revision: https://reviews.llvm.org/D43265
llvm-svn: 325650
Patch by Patricio Villalobos.
I discovered that lld for darwin is generating the wrong code for lazy
bindings in the __stub_helper section (at least for osx 10.12). This is
the way i can reproduce this problem, using this program:
#include <stdio.h>
int main(int argc, char **argv) {
printf("C: printf!\n");
puts("C: puts!\n");
return 0;
}
Then I link it using i have tested it in 3.9, 4.0 and 4.1 versions:
$ clang -c hello.c
$ lld -flavor darwin hello.o -o h1 -lc
When i execute the binary h1 the system gives me the following error:
C: printf!
dyld: lazy symbol binding failed:
BIND_OPCODE_SET_SEGMENT_AND_OFFSET_ULEB
has segment 4 which is too large (0..3)
dyld: BIND_OPCODE_SET_SEGMENT_AND_OFFSET_ULEB has segment 4 which is too
large (0..3)
Trace/BPT trap: 5
Investigating the code, it seems that the problem is that the asm code
generated in the file StubPass.cpp, specifically in the line 323,when it
adds, what it seems an arbitrary number (12) to the offset into the lazy
bind opcodes section, but it should be calculated depending on the
MachONormalizedFileBinaryWrite::lazyBindingInfo result.
I confirmed this bug by patching the code manually in the binary and
writing the right offset in the asm code (__stub_helper).
This patch fixes the content of the atom that contains the assembly code
when the offset is known.
Differential Revision: https://reviews.llvm.org/D35387
llvm-svn: 311734
Summary: This test fails if you're building with BUILD_SHARED_LIBS because the LLVM libraries are found via @rpath. Symlinking instead of copying should be sufficiently robust for the test case.
Reviewers: llvm-commits
Subscribers: davide
Differential Revision: https://reviews.llvm.org/D24476
llvm-svn: 281271
It only makes sense to set on N_NO_DEAD_STRIP on a relocatable object file. Otherwise the bits aren't useful for anything. Matches the ld64 behaviour.
llvm-svn: 278419
We should be using one of BIND_OPCODE_SET_DYLIB_SPECIAL_IMM, BIND_OPCODE_SET_DYLIB_ORDINAL_IMM,
and BIND_OPCODE_SET_DYLIB_ORDINAL_ULEB depending on whether ordinals are <= 0, <= 15, > 15.
This matches the behaviour of ld64.
llvm-svn: 278407
We already had logic for binding opcodes had the same addend as last time. This adds
the cases where the ordinal, symbol name, type, and segment offsets are the same as
the last emitted ordinal.
This gets us one step closer to emitting rebase opcodes as compressed as ld64 can manage.
llvm-svn: 278405
Currently we do this when an atom is used, but we need to do it when a
dylib is referenced on the cmdline as this matches ld64.
This fixes much confusion over which maps are indexed with installName
vs path. There is likely other confusion so i'll be seeing if i can remove
path() completely in a future commit as path() shouldn't really be needed by anyone.
llvm-svn: 278396
A future commit will change when we choose to add an LC_LOAD_DYLIB to the final image. Currently
we do this when an atom is used, but we need to do it when a dylib is referenced on the cmdline as
this matches ld64.
To allow this change, libsystem (and other future yaml files representing dylibs) need to be dylibs
so that the loader can see to add them to the referenced dylib list.
llvm-svn: 278386
A version of 0x1000 is 0.16.0, not 1.0.0 as the comment said. Fix the
value to match the comment, and also the one test case which had this
wrong.
llvm-svn: 278381
An upcoming commit will change how we choose to reference a dylib. Currently
dylibs are only given an LC_LOAD_DYLIB in the final image if an atom is used.
This is different from ld64 which adds the load command when the dylib is referenced
on the cmdline.
In order to change this behaviour, we need libSystem.yaml to actually contain a mach header
so that it is parsed as a dylib, instead of currently being parsed as a normalised file.
To get a mach header, we also require an arch, so now we have one libsystem per arch and
all the tests have been updated to choose the correct one.
llvm-svn: 278372
Using vmsize to populate this file works when outputing MachO images, but fails
when outputting relocatable objects. This patch fixes the computation to use
file offsets, which works for both output types.
Fixes <rdar://problem/27727666>
llvm-svn: 278297
The export trie was being emitted in the order the nodes were
added to the vector, but instead needs to be visited in the order
that the nodes are traversed. This matches the behaviour of ld64.
llvm-svn: 277869
The previous run line depended on libSystem.dylib being present, which it's not
on non-darwin platforms. The new run line uses libSystem.yaml instead.
llvm-svn: 276999
This patch causes LLD to build stabs debugging symbols for files containing
DWARF debug info, and to propagate existing stabs symbols for object files
built using '-r' mode. This enables debugging of binaries generated by LLD
from MachO objects.
llvm-svn: 276921
These references are used to implement MachO/x64-64 subtractor relocations
where the minuend is being fixed up, rather than the subtrahend. The 64-bit
version was not previously supported, the 32-bit version was partially
implemented but contained bugs not caught by existing test cases. This
patch fixes both functionality and test coverage.
llvm-svn: 273759
The .o path always makes sure to store a power of 2 value in the
Section alignment. However, the YAML code didn't verify this.
Added verification and updated all the tests which had a 3 but meant
to have 2^3.
llvm-svn: 264228
The current code for processCIE and processFDE returns out if it sees
any references. The problem with this is that some references could be
explicit in the binary, while others are implicit as they can be
inferred from the content of the EHFrame itself.
This change walks the references we have against the references we
need, and verifies that all explicit references are in the correct place,
and generates any missing implicit ones.
Reviewed by Lang Hames and Nick Kledzik.
Differential Revision: http://reviews.llvm.org/D15439
llvm-svn: 263590
The non lazy atoms generated in the stubs pass use an image cache to
hold all of the pointers. On arm archs, this is the __got section,
but on x86 archs it should be __nl_symbol_ptr.
rdar://problem/24572729
llvm-svn: 260271
ld64 aligns most of the stub's to 2 byte alignment, expect for
the stub helper common atoms which are 4 byte aligned.
This adds a new field to StubInfo which tracks this alignment
and ensures that this is the alignment we get in the final image.
rdar://problem/24570220
llvm-svn: 260248
Jez Ng