This essentially reverts the code change of D63132 and switches to a simpler approach.
In an executable/shared object, st_shndx of a symbol can be:
1) SHN_UNDEF: undefined symbol (or canonical PLT)
2) SHN_ABS: absolute symbol
3) any other value (usually a regular section index) represents a relative symbol.
The actual value does not matter.
Many ld.so (musl, all archs except MIPS of FreeBSD rtld-elf) even treat 2) and 3)
the same. If .sdata does not exist, it does not matter what value/section
__global_pointer$ has, as long as it is relative (otherwise there will be a pedantic
lld error. See D63132). Just set the st_shndx arbitrarily to 1.
Dummy st_shndx=1 may be used by __rela_iplt_start, linker-script-defined symbols outside a section, __dso_handle, etc.
Reviewed By: ruiu
Differential Revision: https://reviews.llvm.org/D66798
llvm-svn: 370172
Port the D64906 technique to ARM. It deletes 3 alignments at
PT_LOAD boundaries for the default case: the size of an arm binary
decreases by at most 12kb.
Reviewed By: grimar
Differential Revision: https://reviews.llvm.org/D66749
llvm-svn: 370049
EhFrameSection::addSection checks liveness of FDE early. This makes it
infeasible to move combineEhSections() before ICF.
Postpone the check to EhFrameSection::finalizeContents(). This is what
ARMExidxSyntheticSection does and it will make a subsequent patch D66717
simpler.
Reviewed By: ruiu
Differential Revision: https://reviews.llvm.org/D66727
llvm-svn: 369890
PR42990. For `SECTIONS { b = a; . = 0xff00 + (a >> 8); a = .; }`,
we currently set st_value(a)=0xff00 while st_value(b)=0xffff.
The following call tree demonstrates the problem:
```
link<ELF64LE>(Args);
Script->declareSymbols(); // insert a and b as absolute Defined
Writer<ELFT>().run();
Script->processSectionCommands();
addSymbol(cmd); // a and b are re-inserted. LinkerScript::getSymbolValue
// is lazily called by subsequent evaluation
finalizeSections();
forEachRelSec(scanRelocations<ELFT>);
processRelocAux // another problem PR42506, not affected by this patch
finalizeAddressDependentContent(); // loop executed once
script->assignAddresses(); // a = 0, b = 0xff00
script->assignAddresses(); // a = 0xff00, _end = 0xffff
```
We need another assignAddresses() to finalize the value of `a`.
This patch
1) modifies assignAddress() to track the original section/value of each
symbol and return a symbol whose section/value has changed.
2) moves the post-finalizeSections assignAddress() inside the loop
of finalizeAddressDependentContent() and makes it iterative.
Symbol assignment may not converge so we make a few attempts before
bailing out.
Note, assignAddresses() must be called at least twice. The penultimate
call finalized section addresses while the last finalized symbol values.
It is somewhat obscure and there was no comment.
linkerscript/addr-zero.test tests this.
Reviewed By: ruiu
Differential Revision: https://reviews.llvm.org/D66279
llvm-svn: 369889
Reported at https://reviews.llvm.org/D64930#1642223
If the only section of a PT_LOAD is a SHT_NOBITS section (e.g. .bss), we
may not align its sh_offset. p_offset of the PT_LOAD will be set to
sh_offset, and we will get p_offset!=p_vaddr (mod p_align). If such
executable is mapped by the Linux kernel, it will segfault.
After D64906, this may happen the non-linker script case.
The linker script case has had this issue for a long time.
This was fixed by rL321657 (but the test linkerscript/nobits-offset.s
failed to test a SHT_NOBITS section), but broken by rL345154.
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D66658
llvm-svn: 369828
Ported the D64906 technique to EM_386.
If `sh_addralign(.tdata) < sh_addralign(.tbss)`,
we can potentially make `p_vaddr(PT_TLS)%p_align(PT_TLS) != 0`.
ld.so that are known to have problems if p_vaddr%p_align!=0:
* FreeBSD 13.0-CURRENT rtld-elf
* glibc https://sourceware.org/bugzilla/show_bug.cgi?id=24606
New test i386-tls-vaddr-align.s checks our workaround makes p_vaddr%p_align = 0.
Reviewed By: ruiu
Differential Revision: https://reviews.llvm.org/D65865
llvm-svn: 369347
Ported the D64906 technique to AArch64. It deletes 3 alignments at
PT_LOAD boundaries for the default case: the size of an aarch64 binary
decreases by at most 192kb.
If `sh_addralign(.tdata) < sh_addralign(.tbss)`,
we can potentially make `p_vaddr(PT_TLS)%p_align(PT_TLS) != 0`.
ld.so that are known to have problems if p_vaddr%p_align!=0:
* musl<=1.1.22
* FreeBSD 13.0-CURRENT (and before) rtld-elf arm64
New test aarch64-tls-vaddr-align.s checks that our workaround makes p_vaddr%p_align = 0.
Reviewed By: ruiu
Differential Revision: https://reviews.llvm.org/D64930
llvm-svn: 369344
This change affects the non-linker script case (precisely, when the
`SECTIONS` command is not used). It deletes 3 alignments at PT_LOAD
boundaries for the default case: the size of a powerpc64 binary can be
decreased by at most 192kb. The technique can be ported to other
targets.
Let me demonstrate the idea with a maxPageSize=65536 example:
When assigning the address to the first output section of a new PT_LOAD,
if the end p_vaddr of the previous PT_LOAD is 0x10020, we advance to
the next multiple of maxPageSize: 0x20000. The new PT_LOAD will thus
have p_vaddr=0x20000. Because p_offset and p_vaddr are congruent modulo
maxPageSize, p_offset will be 0x20000, leaving a p_offset gap [0x10020,
0x20000) in the output.
Alternatively, if we advance to 0x20020, the new PT_LOAD will have
p_vaddr=0x20020. We can pick either 0x10020 or 0x20020 for p_offset!
Obviously 0x10020 is the choice because it leaves no gap. At runtime,
p_vaddr will be rounded down by pagesize (65536 if
pagesize=maxPageSize). This PT_LOAD will load additional initial
contents from p_offset ranges [0x10000,0x10020), which will also be
loaded by the previous PT_LOAD. This is fine if -z noseparate-code is in
effect or if we are not transiting between executable and non-executable
segments.
ld.bfd -z noseparate-code leverages this technique to keep output small.
This patch implements the technique in lld, which is mostly effective on
targets with large defaultMaxPageSize (AArch64/MIPS/PPC: 65536). The 3
removed alignments can save almost 3*65536 bytes.
Two places that rely on p_vaddr%pagesize = 0 have to be updated.
1) We used to round p_memsz(PT_GNU_RELRO) up to commonPageSize (defaults
to 4096 on all targets). Now p_vaddr%commonPageSize may be non-zero.
The updated formula takes account of that factor.
2) Our TP offsets formulae are only correct if p_vaddr%p_align = 0.
Fix them. See the updated comments in InputSection.cpp for details.
On targets that we enable the technique (only PPC64 now),
we can potentially make `p_vaddr(PT_TLS)%p_align(PT_TLS) != 0`
if `sh_addralign(.tdata) < sh_addralign(.tbss)`
This exposes many problems in ld.so implementations, especially the
offsets of dynamic TLS blocks. Known issues:
FreeBSD 13.0-CURRENT rtld-elf (i386/amd64/powerpc/arm64)
glibc (HEAD) i386 and x86_64 https://sourceware.org/bugzilla/show_bug.cgi?id=24606
musl<=1.1.22 on TLS Variant I architectures (aarch64/powerpc64/...)
So, force p_vaddr%p_align = 0 by rounding dot up to p_align(PT_TLS).
The technique will be enabled (with updated tests) for other targets in
subsequent patches.
Reviewed By: ruiu
Differential Revision: https://reviews.llvm.org/D64906
llvm-svn: 369343
Currently the following 3 relocation types do not trigger the creation
of a canonical PLT (which changes STT_GNU_IFUNC to STT_FUNC and
redirects all references):
1) GOT-generating (`needsGot`)
2) PLT-generating (`needsPlt`)
3) R_ABS with 0 addend in a writable location. This is used for
for ifunc function pointers in writable sections such as .data and .toc.
This patch deletes case 3) to simplify the R_*_IRELATIVE generating
logic added in D57371. Other advantages:
* It is guaranteed no more than 1 R_*_IRELATIVE is created for an ifunc.
* PPC64: no need to special case ifunc in toc-indirect to toc-relative relaxation. See D65755
The deleted elf::addIRelativeRelocs demonstrates that one-pass scan
through relocations makes several optimizations difficult. This is
something we can think about in the future.
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D65995
llvm-svn: 368661
In Writer::includeInDynSym(), exportDynamic is used by a Defined with
protected or default visibility, to record whether it is required to be
exported into .dynsym. It is set when any of the following conditions
hold:
1) There is an interposable symbol from a DSO (Undefined or SharedSymbol with default visibility)
2) If -shared or --export-dynamic is specified, any symbol in an object file/bitcode sets this property, unless suppressed by canBeOmittedFromSymbolTable().
3) --dynamic-list when producing an executable
4) protected symbol from a DSO preempted by copy relocation/canonical PLT when
--ignore-{data,function}-address-equality is specified
5) ifunc is exported when -z ifunc-noplt is specified
Bullet points 4) and 5) are irrelevant in this patch.
Bullet 3) does not play well with 1) and 2). When -shared is specified,
exportDynamic of most symbols is true. This makes it incapable to record
--dynamic-list marked symbols. We thus have obscure:
if (!config->shared)
b->exportDynamic = true;
else if (b->includeInDynsym())
b->isPreemptible = true;
This patch adds another bit `Symbol::inDynamicList` to record
3). We can thus simplify handleDynamicList() by unifying the DSO and
executable cases. It also allows us to simplify isPreemptible - now
the field is only used in finalizeSections() and later stages.
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D66091
llvm-svn: 368659
We prioritize non-* wildcards overs VER_NDX_LOCAL/VER_NDX_GLOBAL "*".
This patch generalizes the rule to "*" of other versions and thus fixes PR40176.
I don't feel strongly about this GNU linkers' behavior but the
generalization simplifies code.
Delete `config->defaultSymbolVersion` which was used to special case
VER_NDX_LOCAL/VER_NDX_GLOBAL "*".
In `SymbolTable::scanVersionScript`, custom versions are handled the same
way as VER_NDX_LOCAL/VER_NDX_GLOBAL. So merge
`config->versionScript{Locals,Globals}` into `config->versionDefinitions`.
Overall this seems to simplify the code.
In `SymbolTable::assign{Exact,Wildcard}Versions`,
`sym->verdefIndex == config->defaultSymbolVersion` is changed to
`verdefIndex == UINT32_C(-1)`.
This allows us to give duplicate assignment diagnostics for
`{ global: foo; };` `V1 { global: foo; };`
In test/linkerscript/version-script.s:
vs_index of an undefined symbol changes from 0 to 1. This doesn't matter (arguably 1 is better because the binding is STB_GLOBAL) because vs_index of an undefined symbol is ignored.
Reviewed By: ruiu
Differential Revision: https://reviews.llvm.org/D65716
llvm-svn: 367869
An R_*_IRELATIVE represents the address of a STT_GNU_IFUNC symbol
(redirected at runtime) which is non-preemptable and is not associated
with a canonical PLT (associated with a symbol with a section index of
SHN_UNDEF but a non-zero st_value).
.rel[a].plt [DT_JMPREL, DT_JMPREL+DT_JMPRELSZ) contains relocations that
can be lazily resolved. R_*_IRELATIVE are always eagerly resolved, so
conceptually they do not belong to .rela.plt. "iplt" is mostly a misnomer.
glibc powerpc and powerpc64 do not resolve R_*_IRELATIVE if they are in .rela.plt.
// a.o - synthesized PLT call stub has an R_*_IRELATIVE
void ifunc(); int main() { ifunc(); }
// b.o
static void real() {}
asm (".type ifunc, %gnu_indirect_function");
void *ifunc() { return ℜ }
The lld-linked executable crashes. ld.bfd places R_*_IRELATIVE in
.rela.dyn and the executable works.
glibc i386, x86_64, and aarch64 have logic
(glibc/sysdeps/*/dl-machine.h:elf_machine_lazy_rel) to eagerly resolve
R_*_IRELATIVE in .rel[a].plt so the lld-linked executable works.
Move R_*_IRELATIVE from .rel[a].plt to .rel[a].dyn to fix the crashes on
glibc powerpc/powerpc64. This also helps simplifying ifunc
implementation in FreeBSD rtld-elf powerpc64.
If --pack-dyn-relocs=android[+relr] is specified, the Android packed
dynamic relocation format is used for .rela.dyn. We cannot name
in.relaIplt ".rela.dyn" because the output section will have mixed
formats. This can be improved in the future.
Reviewed By: pcc
Differential Revision: https://reviews.llvm.org/D65651
llvm-svn: 367745
This patch
1) adds -z separate-code and -z noseparate-code (default).
2) changes the condition that the last page of last PF_X PT_LOAD is
padded with trap instructions.
Current condition (after D33630): if there is no `SECTIONS` commands.
After this change: if -z separate-code is specified.
-z separate-code was introduced to ld.bfd in 2018, to place the text
segment in its own pages. There is no overlap in pages between an
executable segment and a non-executable segment:
1) RX cannot load initial contents from R or RW(or non-SHF_ALLOC).
2) R and RW(or non-SHF_ALLOC) cannot load initial contents from RX.
lld's current status:
- Between R and RX: in `Writer<ELFT>::fixSectionAlignments()`, the start of a
segment is always aligned to maxPageSize, so the initial contents loaded by R
and RX do not overlap. I plan to allow overlaps in D64906 if -z noseparate-code
is in effect.
- Between RX and RW(or non-SHF_ALLOC if RW doesn't exist):
we currently unconditionally pad the last page to commonPageSize
(defaults to 4096 on all targets we support).
This patch will make it effective only if -z separate-code is specified.
-z separate-code is a dubious feature that intends to reduce the number
of ROP gadgets (which is actually ineffective because attackers can find
plenty of gadgets in the text segment, no need to find gadgets in
non-code regions).
With the overlapping PT_LOAD technique D64906, -z noseparate-code
removes two more alignments at segment boundaries than -z separate-code.
This saves at most defaultCommonPageSize*2 bytes, which are significant
on targets with large defaultCommonPageSize (AArch64/MIPS/PPC: 65536).
Issues/feedback on alignment at segment boundaries to help understand
the implication:
* binutils PR24490 (the situation on ld.bfd is worse because they have
two R-- on both sides of R-E so more alignments.)
* In binutils, the 2018-02-27 commit "ld: Add --enable-separate-code" made -z separate-code the default on Linux.
d969dea983
In musl-cross-make, binutils is configured with --disable-separate-code
to address size regressions caused by -z separate-code. (lld actually has the same
issue, which I plan to fix in a future patch. The ld.bfd x86 status is
worse because they default to max-page-size=0x200000).
* https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=237676 people want
smaller code size. This patch will remove one alignment boundary.
* Stef O'Rear: I'm opposed to any kind of page alignment at the
text/rodata line (having a partial page of text aliased as rodata and
vice versa has no demonstrable harm, and I actually care about small
systems).
So, make -z noseparate-code the default.
Reviewed By: ruiu
Differential Revision: https://reviews.llvm.org/D64903
llvm-svn: 367537
Summary:
After D58892 split the RW PT_LOAD on the PT_GNU_RELRO boundary, the new
layout is:
PT_LOAD(PT_GNU_RELRO(.data.rel.ro .bss.rel.ro)) PT_LOAD(.data. .bss)
The two pageAlign() calls at PT_GNU_RELRO boundaries are redundant due
to the existence of PT_LOAD.
Reviewers: grimar, peter.smith, ruiu, espindola
Reviewed By: ruiu
Subscribers: sfertile, atanasyan, emaste, arichardson, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D64854
llvm-svn: 366307
This patch does the same thing as r365595 to other subdirectories,
which completes the naming style change for the entire lld directory.
With this, the naming style conversion is complete for lld.
Differential Revision: https://reviews.llvm.org/D64473
llvm-svn: 365730
This patch is mechanically generated by clang-llvm-rename tool that I wrote
using Clang Refactoring Engine just for creating this patch. You can see the
source code of the tool at https://reviews.llvm.org/D64123. There's no manual
post-processing; you can generate the same patch by re-running the tool against
lld's code base.
Here is the main discussion thread to change the LLVM coding style:
https://lists.llvm.org/pipermail/llvm-dev/2019-February/130083.html
In the discussion thread, I proposed we use lld as a testbed for variable
naming scheme change, and this patch does that.
I chose to rename variables so that they are in camelCase, just because that
is a minimal change to make variables to start with a lowercase letter.
Note to downstream patch maintainers: if you are maintaining a downstream lld
repo, just rebasing ahead of this commit would cause massive merge conflicts
because this patch essentially changes every line in the lld subdirectory. But
there's a remedy.
clang-llvm-rename tool is a batch tool, so you can rename variables in your
downstream repo with the tool. Given that, here is how to rebase your repo to
a commit after the mass renaming:
1. rebase to the commit just before the mass variable renaming,
2. apply the tool to your downstream repo to mass-rename variables locally, and
3. rebase again to the head.
Most changes made by the tool should be identical for a downstream repo and
for the head, so at the step 3, almost all changes should be merged and
disappear. I'd expect that there would be some lines that you need to merge by
hand, but that shouldn't be too many.
Differential Revision: https://reviews.llvm.org/D64121
llvm-svn: 365595
If .sdata is absent, linker synthesized __global_pointer$ gets a section index of SHN_ABS.
(ld.bfd has a similar issue: binutils PR24678)
Scrt1.o may use `lla gp, __global_pointer$` to reference the symbol PC
relatively. In -pie/-shared mode, lld complains if a PC relative
relocation references an absolute symbol (SHN_ABS) but ld.bfd doesn't:
ld.lld: error: relocation R_RISCV_PCREL_HI20 cannot refer to lute symbol: __global_pointer$
Let the reference of __global_pointer$ to force creation of .sdata to
fix the problem. This is similar to _GLOBAL_OFFSET_TABLE_, which forces
creation of .got or .got.plt .
Also, change the visibility from STV_HIDDEN to STV_DEFAULT and don't
define the symbol for -shared. This matches ld.bfd, though I don't
understand why it uses STV_DEFAULT.
Reviewed By: ruiu, jrtc27
Differential Revision: https://reviews.llvm.org/D63132
llvm-svn: 363351
Otherwise the getPartition() accessor may return an OOB pointer. Found
using _GLIBCXX_DEBUG.
The error is benign (we never dereference the pointer for the end marker)
so this wasn't caught by e.g. the sanitizer bots.
llvm-svn: 363026
We create several types of synthetic sections for loadable partitions, including:
- The dynamic symbol table. This allows code outside of the loadable partitions
to find entry points with dlsym.
- Creating a dynamic symbol table also requires the creation of several other
synthetic sections for the partition, such as the dynamic table and hash table
sections.
- The partition's ELF header is represented as a synthetic section in the
combined output file, and will be used by llvm-objcopy to extract partitions.
Differential Revision: https://reviews.llvm.org/D62350
llvm-svn: 362819
Many -static/-no-pie/-shared/-pie applications linked against glibc or musl
should work with this patch. This also helps FreeBSD PowerPC64 to migrate
their lib32 (PR40888).
* Fix default image base and max page size.
* Support new-style Secure PLT (see below). Old-style BSS PLT is not
implemented, so it is not suitable for FreeBSD rtld now because it doesn't
support Secure PLT yet.
* Support more initial relocation types:
R_PPC_ADDR32, R_PPC_REL16*, R_PPC_LOCAL24PC, R_PPC_PLTREL24, and R_PPC_GOT16.
The addend of R_PPC_PLTREL24 is special: it decides the call stub PLT type
but it should be ignored for the computation of target symbol VA.
* Support GNU ifunc
* Support .glink used for lazy PLT resolution in glibc
* Add a new thunk type: PPC32PltCallStub that is similar to PPC64PltCallStub.
It is used by R_PPC_REL24 and R_PPC_PLTREL24.
A PLT stub used in -fPIE/-fPIC usually loads an address relative to
.got2+0x8000 (-fpie/-fpic code uses _GLOBAL_OFFSET_TABLE_ relative
addresses).
Two .got2 sections in two object files have different addresses, thus a PLT stub
can't be shared by two object files. To handle this incompatibility,
change the parameters of Thunk::isCompatibleWith to
`const InputSection &, const Relocation &`.
PowerPC psABI specified an old-style .plt (BSS PLT) that is both
writable and executable. Linkers don't make separate RW- and RWE segments,
which causes all initially writable memory (think .data) executable.
This is a big security concern so a new PLT scheme (secure PLT) was developed to
address the security issue.
TLS will be implemented in D62940.
glibc older than ~2012 requires .rela.dyn to include .rela.plt, it can
not handle the DT_RELA+DT_RELASZ == DT_JMPREL case correctly. A hack
(not included in this patch) in LinkerScript.cpp addOrphanSections() to
work around the issue:
if (Config->EMachine == EM_PPC) {
// Older glibc assumes .rela.dyn includes .rela.plt
Add(In.RelaDyn);
if (In.RelaPlt->isLive() && !In.RelaPlt->Parent)
In.RelaDyn->getParent()->addSection(In.RelaPlt);
}
Reviewed By: ruiu
Differential Revision: https://reviews.llvm.org/D62464
llvm-svn: 362721
We currently (ab)use the Live bit on output sections to track whether
the section has ever had an input section added to it, and then later
use it during orphan placement. This will conflict with one of my upcoming
partition-related changes that will assign all output sections to a partition
(thus marking them as live) so that they can be added to the correct segment
by the code that creates program headers.
Instead of using the Live bit for this purpose, create a new flag and
start using it to track the property explicitly.
Differential Revision: https://reviews.llvm.org/D62348
llvm-svn: 362444
(1) {gcc,clang} -fuse-ld=bfd -pie -fPIE -nostdlib a.c => .interp created
(2) {gcc,clang} -fuse-ld=lld -pie -fPIE -nostdlib a.c => .interp not created
(3) {gcc,clang} -fuse-ld=lld -pie -fPIE -nostdlib a.c a.so => .interp created
The inconsistency of (2) is due to the condition `!Config->SharedFiles.empty()`.
To make lld behave more like ld.bfd, we could change the condition to:
Config->HasDynSymTab && !Config->DynamicLinker.empty() && Script->needsInterpSection();
However, that would bring another inconsistency as can be observed with:
(4) {gcc,clang} -fuse-ld=bfd -no-pie -nostdlib a.c => .interp not created
So instead, use `!Config->DynamicLinker.empty() && Script->needsInterpSection()`,
which is both simple and consistent in these cases.
The inconsistency of (4) likely originated from ld.bfd and gold's choice to have a default --dynamic-linker.
Their condition to create .interp is ANDed with (not -shared).
Since lld doesn't have a default --dynamic-linker,
compiler drivers (gcc/clang) don't pass --dynamic-linker for -shared,
and direct ld users are not supposed to specify --dynamic-linker for -shared,
we do not need the condition !Config->Shared.
Reviewed By: ruiu
Differential Revision: https://reviews.llvm.org/D62765
llvm-svn: 362355
For the Local Dynamic case of TLSDESC, _TLS_MODULE_BASE_ is defined as a
special TLS symbol that makes:
1) Without relaxation: it produces a dynamic TLSDESC relocation that
computes 0. Adding @dtpoff to access a TLS symbol.
2) With LD->LE relaxation: _TLS_MODULE_BASE_@tpoff = 0 (lowest address in
the TLS block). Adding @tpoff to access a TLS symbol.
For 1), this saves dynamic relocations and GOT slots as otherwise
(General Dynamic) we would create an R_X86_64_TLSDESC and reserve two
GOT slots for each symbol.
Add ElfSym::TlsModuleBase and change the signature of getTlsTpOffset()
to special case _TLS_MODULE_BASE_.
Reviewed By: ruiu
Differential Revision: https://reviews.llvm.org/D62577
llvm-svn: 362078
This change causes us to read partition specifications from partition
specification sections and split output sections into partitions according
to their reachability from partition entry points.
This is only the first step towards a full implementation of partitions. Later
changes will add additional synthetic sections to each partition so that
they can be loaded independently.
Differential Revision: https://reviews.llvm.org/D60353
llvm-svn: 361925
This reverts D53906.
D53906 increased p_align of PT_TLS on ARM/AArch64 to 32/64 to make the
static TLS layout compatible with Android Bionic's ELF TLS. However,
this may cause glibc ARM/AArch64 programs to crash (see PR41527).
The faulty PT_TLS in the executable satisfies p_vaddr%p_align != 0. The
remainder is normally 0 but may be non-zero with the hack in place. The
problem is that we increase PT_TLS's p_align after OutputSections'
addresses are fixed (assignAddress()). It is possible that
p_vaddr%old_p_align = 0 while p_vaddr%new_p_align != 0.
For a thread local variable defined in the executable, lld computed TLS
offset (local exec) is different from glibc computed TLS offset from
another module (initial exec/generic dynamic). Note: PR41527 said the
bug affects initial exec but actually generic dynamic is affected as
well.
(glibc is correct in that it compute offsets that satisfy
`offset%p_align == p_vaddr%p_align`, which is a basic ELF requirement.
This hack appears to work on FreeBSD rtld, musl<=1.1.22, and Bionic, but
that is just because they (and lld) incorrectly compute offsets that
satisfy `offset%p_align = 0` instead.)
Android developers are fine to revert this patch, carry this patch in
their tree before figuring out a long-term solution (e.g. a dummy .tdata
with sh_addralign=64 sh_size={0,1} in crtbegin*.o files. The overhead is
now insignificant after D62059).
Reviewed By: rprichard, srhines
Differential Revision: https://reviews.llvm.org/D62055
llvm-svn: 361090
The code was added in r252352, probably to address some layout issues.
Actually PT_TLS's p_memsz doesn't need to be aligned on either variant.
ld.bfd doesn't do that.
In case of larger alignment (e.g. 64 for Android Bionic on AArch64, see
D62055), this may make the overhead smaller.
Reviewed By: ruiu
Differential Revision: https://reviews.llvm.org/D62059
llvm-svn: 361029
This is the last patch of the series of patches to make it possible to
resolve symbols without asking SymbolTable to do so.
The main point of this patch is the introduction of
`elf::resolveSymbol(Symbol *Old, Symbol *New)`. That function resolves
or merges given symbols by examining symbol types and call
replaceSymbol (which memcpy's New to Old) if necessary.
With the new function, we have now separated symbol resolution from
symbol lookup. If you already have a Symbol pointer, you can directly
resolve the symbol without asking SymbolTable to do that.
Now that the nice abstraction become available, I can start working on
performance improvement of the linker. As a starter, I'm thinking of
making --{start,end}-lib faster.
--{start,end}-lib is currently unnecessarily slow because it looks up
the symbol table twice for each symbol.
- The first hash table lookup/insertion occurs when we instantiate a
LazyObject file to insert LazyObject symbols.
- The second hash table lookup/insertion occurs when we create an
ObjFile from LazyObject file. That overwrites LazyObject symbols
with Defined symbols.
I think it is not too hard to see how we can now eliminate the second
hash table lookup. We can keep LazyObject symbols in Step 1, and then
call elf::resolveSymbol() to do Step 2.
Differential Revision: https://reviews.llvm.org/D61898
llvm-svn: 360975
Previously, we handled common symbols as a kind of Defined symbol,
but what we were doing for common symbols is pretty different from
regular defined symbols.
Common symbol and defined symbol are probably as different as shared
symbol and defined symbols are different.
This patch introduces CommonSymbol to represent common symbols.
After symbols are resolved, they are converted to Defined symbols
residing in a .bss section.
Differential Revision: https://reviews.llvm.org/D61895
llvm-svn: 360841
SymbolTable's add-family functions have lots of parameters because
when they have to create a new symbol, they forward given arguments
to Symbol's constructors. Therefore, the functions take at least as
many arguments as their corresponding constructors.
This patch simplifies the add-family functions. Now, the functions
take a symbol instead of arguments to construct a symbol. If there's
no existing symbol, a given symbol is memcpy'ed to the symbol table.
Otherwise, the functions attempt to merge the existing and a given
new symbol.
I also eliminated `CanOmitFromDynSym` parameter, so that the functions
take really one argument.
Symbol classes are trivially constructible, so looks like constructing
them to pass to add-family functions is as cheap as passing a lot of
arguments to the functions. A quick benchmark showed that this patch
seems performance-neutral.
This is a preparation for
http://lists.llvm.org/pipermail/llvm-dev/2019-April/131902.html
Differential Revision: https://reviews.llvm.org/D61855
llvm-svn: 360838
The -n (--nmagic) disables page alignment, and acts as a -Bstatic
The -N (--omagic) does what -n does but also marks the executable segment as
writeable. As page alignment is disabled headers are not allocated unless
explicit in the linker script.
To disable page alignment in LLD we choose to set the page sizes to 1 so
that any alignment based on the page size does nothing. To set the
Target->PageSize to 1 we implement -z common-page-size, which has the side
effect of allowing the user to set the value as well.
Setting the page alignments to 1 does mean that any use of
CONSTANT(MAXPAGESIZE) or CONSTANT(COMMONPAGESIZE) in a linker script will
return 1, unlike in ld.bfd. However given that -n and -N disable paging
these probably shouldn't be used in a linker script where -n or -N is in
use.
Differential Revision: https://reviews.llvm.org/D61688
llvm-svn: 360593
Summary:
While the generic ABI requires notes to be 8-byte aligned in ELF64, many
vendor-specific notes (from Linux, NetBSD, Solaris, etc) use 4-byte
alignment.
In a PT_NOTE segment, if 4-byte aligned notes are followed by an 8-byte
aligned note, the possible 4-byte padding may make consumers fail to
parse the 8-byte aligned note. See PR41000 for a recent report about
.note.gnu.property (NT_GNU_PROPERTY_TYPE_0).
(Note, for NT_GNU_PROPERTY_TYPE_0, the consumers should probably migrate
to PT_GNU_PROPERTY, but the alignment issue affects other notes as well.)
To fix the issue, don't mix notes with different alignments in one
PT_NOTE. If compilers emit 4-byte aligned notes before 8-byte aligned
notes, we'll create at most 2 segments.
sh_size%sh_addralign=0 is actually implied by the rule for linking
unrecognized sections (in generic ABI), so we don't have to check that.
Notes that match in name, type and attribute flags are concatenated into
a single output section. The compilers have to ensure
sh_size%sh_addralign=0 to make concatenated notes parsable.
An alternative approach is to create a PT_NOTE for each SHT_NOTE, but
we'll have to incur the sizeof(Elf64_Phdr)=56 overhead every time a new
note section is introduced.
Reviewers: ruiu, jakehehrlich, phosek, jhenderson, pcc, espindola
Subscribers: emaste, arichardson, krytarowski, fedor.sergeev, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61296
llvm-svn: 359853
This is a follow up to r358979 which made findOrphanPos only consider
live sections. Unfortunately, this required change to getRankProximity,
used by findOrphanPos, was missed.
Differential Revision: https://reviews.llvm.org/D61197
llvm-svn: 359554
This patch changes the behaviour of findOrphanPos to only consider live
sections when placing orphan sections. This used to be how it behaved in
the past.
Differential Revision: https://reviews.llvm.org/D60273
llvm-svn: 358979
Make some small adjustment while touching the code: make parameters
const, use less_first(), etc.
Differential Revision: https://reviews.llvm.org/D60989
llvm-svn: 358943
This generalizes code and also fixes the broken behavior shown in
one of our test cases for some targets, like x86-64.
The issue occurs when the forward declarations are used in the script.
One of the samples is:
SECTIONS {
foo = ADDR(.text) - ABSOLUTE(ADDR(.text));
};
In that case, we have a broken output when output target does
not use thunks. That happens because thunks creating code
(called from maybeAddThunks)
calls Script->assignAddresses() at least one more time,
what fixups the values. As a result final symbols values can
be different on AArch64 and x86, for example.
In this patch, I generalize and rename maybeAddThunks to
finalizeAddressDependentContent and now it is used and called
by all targets.
Differential revision: https://reviews.llvm.org/D55550
llvm-svn: 358646
With partitions, each partition should have the same build id. This means
that the build id needs to be only computed once, otherwise we will end up
with different build ids in each partition as a result of the file contents
changing. This change moves the computation of the build id into Writer so
that it only happens once.
Differential Revision: https://reviews.llvm.org/D60342
llvm-svn: 358536
For partitions I intend to use the same set of version indexes in
each partition for simplicity. Since each partition will need its own
VersionNeedSection this will require moving the verneed tracking out of
VersionNeedSection. The way I've done this is to move most of the tracking
into SharedFile. What will eventually become the per-partition tracking
still lives in VersionNeedSection.
As a bonus the code gets a little simpler and more consistent with how we
handle verdef.
Differential Revision: https://reviews.llvm.org/D60307
llvm-svn: 357926
Previously, we drop symbols starting with .L from the symbol table, so
if there is a relocation that refers a .L symbol, it ended up
referencing a null -- which happened to be interpreted as an absolute
symbol.
This patch copies all symbols including local ones if -emit-reloc is
given.
Fixes https://bugs.llvm.org/show_bug.cgi?id=41385
Differential Revision: https://reviews.llvm.org/D60306
llvm-svn: 357885
And rename the function to combineEhSections(). This makes the processing
of .ARM.exidx even more similar to .eh_frame and means that we can avoid an
additional loop over InputSections.
Differential Revision: https://reviews.llvm.org/D60026
llvm-svn: 357417
Summary:
Some synthetic sections can be empty while still being needed, thus they
can't be removed by removeUnusedSyntheticSections(). Rename this member
function to more appropriate isNeeded() with the opposite meaning.
No functional change intended.
Reviewers: ruiu, espindola
Reviewed By: ruiu
Subscribers: jhenderson, grimar, emaste, arichardson, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59982
llvm-svn: 357377
This change itself doesn't mean anything, but it helps D59780 because
in patch, we don't know whether we need to create a CET-aware PLT or
not until we read all input files.
llvm-svn: 357194
Recommit r356666 with fixes for buildbot failure, as well as handling for
--emit-relocs, which we decide not to emit any relocation sections as the
table is already position independent and an offline tool can deduce the
relocations.
Instead of creating extra Synthetic .ARM.exidx sections to account for
gaps in the table, create a single .ARM.exidx SyntheticSection that can
derive the contents of the gaps from a sorted list of the executable
InputSections. This has the benefit of moving the ARM specific code for
SyntheticSections in SHF_LINK_ORDER processing and the table merging code
into the ARM specific SyntheticSection. This also makes it easier to create
EXIDX_CANTUNWIND table entries for executable InputSections that don't
have an associated .ARM.exidx section.
Fixes pr40277
Differential Revision: https://reviews.llvm.org/D59216
llvm-svn: 357160
Summary:
This should address remaining issues discussed in PR36555.
Currently R_GOT*_FROM_END are exclusively used by x86 and x86_64 to
express relocations types relative to the GOT base. We have
_GLOBAL_OFFSET_TABLE_ (GOT base) = start(.got.plt) but end(.got) !=
start(.got.plt)
This can have problems when _GLOBAL_OFFSET_TABLE_ is used as a symbol, e.g.
glibc dl_machine_dynamic assumes _GLOBAL_OFFSET_TABLE_ is start(.got.plt),
which is not true.
extern const ElfW(Addr) _GLOBAL_OFFSET_TABLE_[] attribute_hidden;
return _GLOBAL_OFFSET_TABLE_[0]; // R_X86_64_GOTPC32
In this patch, we
* Change all GOT*_FROM_END to GOTPLT* to fix the problem.
* Add HasGotPltOffRel to denote whether .got.plt should be kept even if
the section is empty.
* Simplify GotSection::empty and GotPltSection::empty by setting
HasGotOffRel and HasGotPltOffRel according to GlobalOffsetTable early.
The change of R_386_GOTPC makes X86::writePltHeader simpler as we don't
have to compute the offset start(.got.plt) - Ebx (it is constant 0).
We still diverge from ld.bfd (at least in most cases) and gold in that
.got.plt and .got are not adjacent, but the advantage doing that is
unclear.
Reviewers: ruiu, sivachandra, espindola
Subscribers: emaste, mehdi_amini, arichardson, dexonsmith, jdoerfert, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59594
llvm-svn: 356968
There is a reproducible buildbot failure (segfault) on the 2 stage
clang-cmake-armv8-lld bot. Reverting while I investigate.
Differential Revision: https://reviews.llvm.org/D59216
llvm-svn: 356684
Instead of creating extra Synthetic .ARM.exidx sections to account for
gaps in the table, create a single .ARM.exidx SyntheticSection that can
derive the contents of the gaps from a sorted list of the executable
InputSections. This has the benefit of moving the ARM specific code for
SyntheticSections in SHF_LINK_ORDER processing and the table merging code
into the ARM specific SyntheticSection. This also makes it easier to create
EXIDX_CANTUNWIND table entries for executable InputSections that don't
have an associated .ARM.exidx section.
Fixes pr40277
Differential Revision: https://reviews.llvm.org/D59216
llvm-svn: 356666
Summary:
Based on Peter Collingbourne's suggestion in D56828.
Before D56828: PT_LOAD(.data PT_GNU_RELRO(.data.rel.ro .bss.rel.ro) .bss)
Old: PT_LOAD(PT_GNU_RELRO(.data.rel.ro .bss.rel.ro) .data .bss)
New: PT_LOAD(PT_GNU_RELRO(.data.rel.ro .bss.rel.ro)) PT_LOAD(.data. .bss)
The new layout reflects the runtime memory mappings.
By having two PT_LOAD segments, we can utilize the NOBITS part of the
first PT_LOAD and save bytes for .bss.rel.ro.
.bss.rel.ro is currently small and only used by copy relocations of
symbols in read-only segments, but it can be used for other purposes in
the future, e.g. if a relro section's statically relocated data is all
zeros, we can move it to .bss.rel.ro.
Reviewers: espindola, ruiu, pcc
Reviewed By: ruiu
Subscribers: nemanjai, jvesely, nhaehnle, javed.absar, kbarton, emaste, arichardson, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D58892
llvm-svn: 356226
Old: PT_LOAD(.data | PT_GNU_RELRO(.data.rel.ro .bss.rel.ro) | .bss)
New: PT_LOAD(PT_GNU_RELRO(.data.rel.ro .bss.rel.ro) | .data .bss)
The placement of | indicates page alignment caused by PT_GNU_RELRO. The
new layout has simpler rules and saves space for many cases.
Old size: roundup(.data) + roundup(.data.rel.ro)
New size: roundup(.data.rel.ro + .bss.rel.ro) + .data
Other advantages:
* At runtime the 3 memory mappings decrease to 2.
* start(PT_TLS) = start(PT_GNU_RELRO) = start(RW PT_LOAD). This
simplifies binary manipulation tools.
GNU strip before 2.31 discards PT_GNU_RELRO if its
address is not equal to the start of its associated PT_LOAD.
This has been fixed by https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=f2731e0c374e5323ce4cdae2bcc7b7fe22da1a6f
But with this change, we will be compatible with GNU strip before 2.31
* Before, .got.plt (non-relro by default) was placed before .got (relro
by default), which made it impossible to have _GLOBAL_OFFSET_TABLE_
(start of .got.plt on x86-64) equal to the end of .got (R_GOT*_FROM_END)
(https://bugs.llvm.org/show_bug.cgi?id=36555). With the new ordering, we
can improve on this regard if we'd like to.
Reviewers: ruiu, espindola, pcc
Subscribers: emaste, arichardson, llvm-commits, joerg, jdoerfert
Differential Revision: https://reviews.llvm.org/D56828
llvm-svn: 356117
This does not appear to be necessary because StringTableSection does not
need to be finalized, which also means that we can remove the call to
finalizeSynthetic on .dynstr.
Differential Revision: https://reviews.llvm.org/D59240
llvm-svn: 355977
We're going to need a separate VersionNeedSection for each partition, and
the partition data structure won't be templated.
With this the VersionTableSection class no longer needs ELFT, so detemplate it.
Differential Revision: https://reviews.llvm.org/D58808
llvm-svn: 355478
This lets us detect file size overflows when creating a 64-bit binary on
a 32-bit machine.
Differential Revision: https://reviews.llvm.org/D58840
llvm-svn: 355218
r355153 introduced a build failure on a build bot that uses clang natively
on an armv7-a machine. This a temporary fix to use size_t rather than
uint64_t.
llvm-svn: 355195
This lets us remove the special case from Writer::writeSections(), and also
fixes a bug where .eh_frame_hdr isn't necessarily written in the correct
order if a linker script moves .eh_frame and .eh_frame_hdr into the same
output section.
Differential Revision: https://reviews.llvm.org/D58795
llvm-svn: 355153
Three MIPS-specific sections `.reginfo`, `.MIPS.options`, and `.MIPS.abiflags`
are used by loader to read their contents and setup environment for running
a program. Loader looks up these data in the corresponding segments:
`PT_MIPS_REGINFO`, `PT_MIPS_OPTIONS`, and `PT_MIPS_ABIFLAGS` respectively.
This patch put these sections to separate segments like we do already
for ARM `SHT_ARM_EXIDX` section.
Differential Revision: http://reviews.llvm.org/D58381
llvm-svn: 354468
Non-GOT non-PLT relocations to non-preemptible ifuncs result in the
creation of a canonical PLT, which now takes the identity of the IFUNC
in the symbol table. This (a) ensures address consistency inside and
outside the module, and (b) fixes a bug where some of these relocations
end up pointing to the resolver.
Fixes (at least) PR40474 and PR40501.
Differential Revision: https://reviews.llvm.org/D57371
llvm-svn: 353981
A follow up to the intial patch that unblocked linking against libgcc.
For lld we don't need to bother tracking which objects have got based small
code model relocations. This is due to the fact that the compilers on
powerpc64 use the .toc section to generate indirections to symbols (rather then
using got relocations) which keeps the got small. This makes overflowing a
small code model got relocation very unlikely.
Differential Revision: https://reviews.llvm.org/D57245
llvm-svn: 353849
Summary:
In ld.bfd/gold, --no-allow-shlib-undefined is the default when linking
an executable. This patch implements a check to error on undefined
symbols in a shared object, if all of its DT_NEEDED entries are seen.
Our approach resembles the one used in gold, achieves a good balance to
be useful but not too smart (ld.bfd traces all DSOs and emulates the
behavior of a dynamic linker to catch more cases).
The error is issued based on the symbol table, different from undefined
reference errors issued for relocations. It is most effective when there
are DSOs that were not linked with -z defs (e.g. when static sanitizers
runtime is used).
gold has a comment that some system libraries on GNU/Linux may have
spurious undefined references and thus system libraries should be
excluded (https://sourceware.org/bugzilla/show_bug.cgi?id=6811). The
story may have changed now but we make --allow-shlib-undefined the
default for now. Its interaction with -shared can be discussed in the
future.
Reviewers: ruiu, grimar, pcc, espindola
Reviewed By: ruiu
Subscribers: joerg, emaste, arichardson, llvm-commits
Differential Revision: https://reviews.llvm.org/D57385
llvm-svn: 352826
Guessing that the slashes used in the scripts SECTION command was causing the
windows related failures in the added test.
Original commit message:
Small code model global variable access on PPC64 has a very limited range of
addressing. The instructions the relocations are used on add an offset in the
range [-0x8000, 0x7FFC] to the toc pointer which points to .got +0x8000, giving
an addressable range of [.got, .got + 0xFFFC]. While user code can be recompiled
with medium and large code models when the binary grows too large for small code
model, there are small code model relocations in the crt files and libgcc.a
which are typically shipped with the distros, and the ABI dictates that linkers
must allow linking of relocatable object files using different code models.
To minimze the chance of relocation overflow, any file that contains a small
code model relocation should have its .toc section placed closer to the .got
then any .toc from a file without small code model relocations.
Differential Revision: https://reviews.llvm.org/D56920
llvm-svn: 352071
Small code model global variable access on PPC64 has a very limited range of
addressing. The instructions the relocations are used on add an offset in the
range [-0x8000, 0x7FFC] to the toc pointer which points to .got +0x8000, giving
an addressable range of [.got, .got + 0xFFFC]. While user code can be recompiled
with medium and large code models when the binary grows too large for small code
model, there are small code model relocations in the crt files and libgcc.a
which are typically shipped with the distros, and the ABI dictates that linkers
must allow linking of relocatable object files using different code models.
To minimze the chance of relocation overflow, any file that contains a small
code model relocation should have its .toc section placed closer to the .got
then any .toc from a file without small code model relocations.
Differential Revision: https://reviews.llvm.org/D56920
llvm-svn: 351978
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
If .rela.iplt does not exist, we used to emit a corrupt symbol table
that contains two symbols, .rela_iplt_{start,end}, pointing to a
nonexisting section.
This patch fixes the issue by setting section index 0 to the symbols
if .rel.iplt section does not exist.
Differential Revision: https://reviews.llvm.org/D56623
llvm-svn: 351218
ARM and AArch64 use TLS variant 1, where the first two words after the
thread pointer are reserved for the TCB, followed by the executable's TLS
segment. Both the thread pointer and the TLS segment are aligned to at
least the TLS segment's alignment.
Android/Bionic historically has not supported ELF TLS, and it has
allocated memory after the thread pointer for several Bionic TLS slots
(currently 9 but soon only 8). At least one of these allocations
(TLS_SLOT_STACK_GUARD == 5) is widespread throughout Android/AArch64
binaries and can't be changed.
To reconcile this disagreement about TLS memory layout, set the minimum
alignment for executable TLS segments to 8 words on ARM/AArch64, which
reserves at least 8 words of memory after the TP (2 for the ABI-specified
TCB and 6 for alignment padding). For simplicity, and because lld doesn't
know when it's targeting Android, increase the alignment regardless of
operating system.
Differential Revision: https://reviews.llvm.org/D53906
llvm-svn: 350681
Summary:
Other large sections (e.g. .rela.dyn .dynstr) may push .note.* off the
first page. They won't be available in core files if RLIMIT_CORE is
limited.
This patch gives priority to alloctable SHT_NOTE sections so that they
are assuredly in the first page and will be available in core files.
They are small and contain important information (e.g. .note.gnu.build-id
identifies the origin of the core, .note.tag stores NT_FREEBSD_ABI_TAG).
Note: gold Output_section_order has a similar rule:
// Loadable read-only note sections come next so that the PT_NOTE
// segment is on the first page of the executable.
ORDER_RO_NOTE,
Reviewers: ruiu, pcc, espindola
Subscribers: emaste, arichardson, llvm-commits
Differential Revision: https://reviews.llvm.org/D55800
llvm-svn: 349524
At least on Linux, if a file size given to FileOutputBuffer is greater
than 2^63, it fails with "Invalid argument" error, which is not a
user-friendly error message. With this patch, lld prints out "output
file too large" instead.
llvm-svn: 348153
There is no need to check that In.DynSymTab != nullptr,
because `includeInDynsym` already checks for `!Config->HasDynSymTab`
and `HasDynSymTab` is the pre-condition for In.DynSymTab creation.
llvm-svn: 348143
The _GLOBAL_OFFSET_TABLE_ is a linker defined symbol that is placed at
some location relative to the .got, .got.plt or .toc section. On some
targets such as Arm the correctness of some code sequences using a
relocation to _GLOBAL_OFFSET_TABLE_ depend on the value of the symbol
being in the linker defined place. Follow the ld.gold example and give
a multiple symbol definition error. The ld.bfd behaviour is to ignore the
definition in the input object and redefine it, which seems like it could
be more surprising.
fixes pr39587
Differential Revision: https://reviews.llvm.org/D54624
llvm-svn: 347854
Now it returns Symbol. This should be NFC that
avoids doing cast at the caller's sides.
Differential revision: https://reviews.llvm.org/D54627
llvm-svn: 347455
On PowerPC64, when a function call offset is too large to encode in a call
instruction the address is stored in a table in the data segment. A thunk is
used to load the branch target address from the table relative to the
TOC-pointer and indirectly branch to the callee. When linking position-dependent
code the addresses are stored directly in the table, for position-independent
code the table is allocated and filled in at load time by the dynamic linker.
For position-independent code the branch targets could have gone in the .got.plt
but using the .branch_lt section for both position dependent and position
independent binaries keeps it consitent and helps keep this PPC64 specific logic
seperated from the target-independent code handling the .got.plt.
Differential Revision: https://reviews.llvm.org/D53408
llvm-svn: 346877
Fangrui Song