I think D79300 has fixed the D51892 (`__i686.get_pc_thunk.bx`) issue, so
we can bring back rL330869.
D79300 says `would error undefined symbol instead of the more relevant discarded section`
but it doesn't reproduce now.
This avoids a quirk in `isUndefWeak()`.
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D111365
This option is a subset of -Bsymbolic-functions. It applies to STB_GLOBAL
STT_FUNC definitions.
The address of a vague linkage function (STB_WEAK STT_FUNC, e.g. an inline
function, a template instantiation) seen by a -Bsymbolic-functions linked
shared object may be different from the address seen from outside the shared
object. Such cases are uncommon. (ELF/Mach-O programs may use
`-fvisibility-inlines-hidden` to break such pointer equality. On Windows,
correct dllexport and dllimport are needed to make pointer equality work.
Windows link.exe enables /OPT:ICF by default so different inline functions may
have the same address.)
```
// a.cc -> a.o -> a.so (-Bsymbolic-functions)
inline void f() {}
void *g() { return (void *)&f; }
// b.cc -> b.o -> exe
// The address is different!
inline void f() {}
```
-Bsymbolic-non-weak-functions is a safer (C++ conforming) subset of
-Bsymbolic-functions, which can make such programs work.
Implementations usually emit a vague linkage definition in a COMDAT group. We
could detect the group (with more code) but I feel that we should just check
STB_WEAK for simplicity. A weak definition will thus serve as an escape hatch
for rare cases when users want interposition on definitions.
GNU ld feature request: https://sourceware.org/bugzilla/show_bug.cgi?id=27871
Longer write-up: https://maskray.me/blog/2021-05-16-elf-interposition-and-bsymbolic
If Linux distributions migrate to protected non-vague-linkage external linkage
functions by default, the linker option can still be handy because it allows
rapid experiment without recompilation. Protected function addresses currently
have deep issues in GNU ld.
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D102570
Since D100490 this case is diagnosed for -z rel. This commit implements
R_AARCH64_TLSDESC cases for AArch64::getImplicitAddend() and
AArch64::relocate(). However, there are probably further relocation types
that need to be handled for full support of -z rel.
Fixes https://bugs.llvm.org/show_bug.cgi?id=47009
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D100544
This patch changes the DynamicReloc class to store an enum instead
of the overloaded useSymVA member to make it easier to understand
and fix incorrect addends being written in some corner cases. The
change is motivated by a follow-up review that checks the value of
implicit Elf_Rel addends written to the output file.
This patch fixes an incorrect output when using `-z rela` for i386 files
with R_386_GOT32 relocations (not that this really matters since it's an
unsupported configuration).
Storing the relocation expression kind also addresses an incorrect addend
FIXME in ppc64-abs64-dyn.s introduced in D63383.
DynamicReloc now also has a special case for the MIPS TLS relocations
(DynamicReloc::AgainstSymbolWithTargetVA) since the
R_MIPS_TLS_TPREL{32/64} the symbol VA to the GOT for preemptible
symbols. I'm not sure if the symbol value actually should be written
for R_MIPS_TLS_TPREL32, but this patch does not attempt to change
that behaviour.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D100490
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
Adopt my suggestion in https://reviews.llvm.org/D91426#2653926 ,
generalizing the ppc64 specific code.
GNU ld and glibc ld.so has a contract about the first few entries of .got .
There are somewhat complex conditions when the header is needed. This patch
switches to a simpler approach: add a header unconditionally if
_GLOBAL_OFFSET_TABLE_ is used or the number of entries is more than just the
header.
The original page no longer works, so use a web.archive.org link instead.
Reviewed By: atanasyan
Differential Revision: https://reviews.llvm.org/D100949
This is the same problem as 127176e59e, but for static TLS rather than
dynamic TLS. Although we know the symbol will be the one in our own TLS
segment, and thus the offset of it within that, we don't know where in
the static TLS block our data will be allocated and thus we must emit a
dynamic relocation for this case.
Reviewed By: MaskRay, atanasyan
Differential Revision: https://reviews.llvm.org/D101381
Whilst not wrong (unless using static PIE where the relocations are
likely not implemented by the runtime), this is inefficient, as the TLS
module indices and offsets are independent of the executable's load
address.
Reviewed By: MaskRay, atanasyan
Differential Revision: https://reviews.llvm.org/D101382
An unfetched lazy symbol (undefined weak) should be considered to have its
original versionId which is VER_NDX_GLOBAL, instead of the lazy symbol's
versionId. (The original versionId cannot be non-VER_NDX_GLOBAL because a
undefined versioned symbol is an error.)
The regression was introduced in D77280 when making version scripts work
with lazy symbols fetched by LTO calls.
Fix PR49915
Differential Revision: https://reviews.llvm.org/D100624
From the PowerPC ELFv2 ABI section 4.2.3. Global Offset Table.
```
The GOT consists of an 8-byte header that contains the TOC base (the first TOC
base when multiple TOCs are present), followed by an array of 8-byte addresses.
```
Due to the introduction of PC Relative code it is now possible to require a GOT
without having a .TOC. symbol in the object that is being linked. Since LLD uses
the .TOC. symbol to determine whether or not a GOT is required the GOT header is
not setup correctly and the 8-byte header is missing.
This patch allows the Power PC GOT setup to happen when an element is added to
the GOT instead of at the very begining. When this header is added a .TOC.
symbol is also added.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D91426
Fixes PR48693: --emit-relocs keeps relocation sections. --gdb-index drops
.debug_gnu_pubnames and .debug_gnu_pubtypes but not their relocation sections.
This can cause a null pointer dereference in `getOutputSectionName`.
Also delete debug-gnu-pubnames.s which is covered by gdb-index.s
Reviewed By: grimar
Differential Revision: https://reviews.llvm.org/D94354
As indicated by AArch64 ELF specification, symbols with st_other
marked with STO_AARCH64_VARIANT_PCS indicates it may follow a variant
procedure call standard with different register usage convention
(for instance SVE calls).
Static linkers must preserve the marking and propagate it to the dynamic
symbol table if any reference or definition of the symbol is marked with
STO_AARCH64_VARIANT_PCS, and add a DT_AARCH64_VARIANT_PCS dynamic tag if
there are R_<CLS>_JUMP_SLOT relocations that reference that symbols.
It implements https://bugs.llvm.org/show_bug.cgi?id=48368.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D93045
In the presence of a gap, the st_value field of a STT_SECTION symbol is the
address of the first input section (incorrect if there is a gap). Set it to the
output section address instead.
In -r mode, this bug can cause an incorrect non-zero st_value of a STT_SECTION
symbol (while output sections have zero addresses, input sections may have
non-zero outSecOff). The non-zero st_value can cause the final link to have
incorrect relocation computation (both GNU ld and LLD add st_value of the
STT_SECTION symbol to the output section address).
Reviewed By: grimar
Differential Revision: https://reviews.llvm.org/D90520
The ELF spec says
> If the sh_flags field for this section header includes the attribute SHF_INFO_LINK, then this member represents a section header table index.
Set SHF_INFO_LINK so that binary manipulation tools know that sh_info is
a section header table index instead of (the number of local symbols in the case of SHT_SYMTAB/SHT_DYNSYM).
We have already added SHF_INFO_LINK for --emit-relocs retained SHT_REL[A].
For example, we can teach llvm-objcopy to preserve the section index of the sh_info referenced section if
SHF_INFO_LINK is set. (GNU objcopy recognizes .rel[a].plt and updates
sh_info even if SHF_INFO_LINK is not set).
Reviewed By: grimar, psmith
Differential Revision: https://reviews.llvm.org/D89828
The combination has not been tested before. In the case of ICF,
`e.section->getVA(0)` equals the start address of the output section.
This can cause incorrect overlapping with the actual function at the
start of the output section and potentially trigger a GDB internal error
in `dw2_find_pc_sect_compunit_symtab` (presumably because:
if a short address range incorrectly starts at the start address of the
output section, GDB may pick it instead of the correct longer address
range. When mapping an address within the long address range but
out of the scope of the short address range, the routine may find
nothing - while the code asserts that it can find something).
Note that in the case of ICF there may be duplicate address range entries,
but GDB appears to be fine with them.
Reviewed By: grimar
Differential Revision: https://reviews.llvm.org/D89751
Thunk alignment is added in thie patch when using pc-rel instructions
to avoid crossing the 64 byte boundary.
Patched by: nemanjai, NeHuang
Reviewed By: sfertile, MaskRay
Differential Revision: https://reviews.llvm.org/D85973
-gdwarf-5 -fdebug-types-section may produce multiple .debug_info sections. All
except one are type units (.debug_types before DWARF v5). When constructing
.gdb_index, we should ignore these type units. We use a simple heuristic: the
compile unit does not have the SHF_GROUP flag. (This needs to be revisited if
people place compile unit .debug_info in COMDAT groups.)
This issue manifests as a data race: because an object file may have multiple
.debug_info sections, we may concurrently construct `LLDDwarfObj` for the same
file in multiple threads. The threads may access `InputSectionBase::data()`
concurrently on the same input section. `InputSectionBase::data()` does a lazy
uncompress() and rewrites the member variable `rawData`. A thread running zlib
`inflate()` (transitively called by uncompress()) on a buffer with `rawData`
tampered by another thread may fail with `uncompress failed: zlib error: Z_DATA_ERROR`.
Even if no data race occurred in an optimistic run, if there are N .debug_info,
one CU entry and its address ranges will be replicated N times. The result
.gdb_index can be much larger than a correct one.
The new test gdb-index-dwarf5-type-unit.s actually has two compile units. This
cannot be produced with regular approaches (it can be produced with -r
--unique). This is used to demonstrate that the .gdb_index construction code
only considers the last non-SHF_GROUP .debug_info
Reviewed By: grimar
Differential Revision: https://reviews.llvm.org/D85579
For an InputSection, the `buf` argument of `InputSectionBase::relocate` points
to the content of the containing OutputSection, instead of the content of the
InputSection itself, so `outSecOff` needs to be added in its callees. This is
counter-intuitive and leads to many `- outSecOff` and `+ outSecOff`.
This patch makes `InputSection::writeTo` call `InputSectionBase::relocate` with
`outSecOff` added. relocAlloc/relocNonAlloc/relocateNonAllocForRelocatable can
thus be simplified now.
Updated test:
* non-abs-reloc.s: A minor offset bug is fixed for a diagnostic in `relocateNonAlloc`
Reviewed By: grimar
Differential Revision: https://reviews.llvm.org/D85618
Fix PR36272 and PR46835
A .eh_frame FDE references a text section and (optionally) a LSDA (in
.gcc_except_table). Even if two text sections have identical content and
relocations (e.g. a() and b()), we cannot fold them if their LSDA are different.
```
void foo();
void a() {
try { foo(); } catch (int) { }
}
void b() {
try { foo(); } catch (float) { }
}
```
Scan .eh_frame pieces with LSDA and disallow referenced text sections to be
folded. If two .gcc_except_table have identical semantics (usually identical
content with PC-relative encoding), we will lose folding opportunity.
For ClickHouse (an exception-heavy application), this can reduce --icf=all efficiency
from 9% to 5%. There may be some percentage we can reclaim without affecting
correctness, if we analyze .eh_frame and .gcc_except_table sections.
gold 2.24 implemented a more complex fix (resolution to
https://sourceware.org/bugzilla/show_bug.cgi?id=21066) which combines the
checksum of .eh_frame CIE/FDE pieces.
Reviewed By: grimar
Differential Revision: https://reviews.llvm.org/D84610
The patch adds checking for various potential issues in parsing name
lookup tables and reporting them as recoverable errors, similarly as we
do for other tables.
Differential Revision: https://reviews.llvm.org/D83050
The parsing method did not check reading errors and might easily fall
into an infinite loop on an invalid input because of that.
Differential Revision: https://reviews.llvm.org/D83049
Both the .ARM.exidx and .eh_frame sections have a custom SyntheticSection
that acts as a container for the InputSections. The InputSections are added
to the SyntheticSection prior to /DISCARD/ which limits the affect a
/DISCARD/ can have to the whole SyntheticSection. In the majority of cases
this is sufficient as it is not common to discard subsets of the
InputSections. The Linux kernel has one of these scripts which has something
like:
/DISCARD/ : { *(.ARM.exidx.exit.text) *(.ARM.extab.exit.text) ... }
The .ARM.exidx.exit.text are not discarded because the InputSection has been
transferred to the Synthetic Section. The *(.ARM.extab.exit.text) sections
have not so they are discarded. When we come to write out the .ARM.exidx
sections the dangling references from .ARM.exidx.exit.text to
.ARM.extab.exit.text currently cause relocation out of range errors, but
could as easily cause a fatal error message if we check for dangling
references at relocation time.
This patch attempts to respect the /DISCARD/ command by running it on the
.ARM.exidx InputSections stored in the SyntheticSection.
The .eh_frame is in theory affected by this problem, but I don't think that
there is a dangling reference problem that can happen with these sections.
Fixes remaining part of pr44824
Differential Revision: https://reviews.llvm.org/D79687
Essentially takes the lld/Common/Threads.h wrappers and moves them to
the llvm/Support/Paralle.h algorithm header.
The changes are:
- Remove policy parameter, since all clients use `par`.
- Rename the methods to `parallelSort` etc to match LLVM style, since
they are no longer C++17 pstl compatible.
- Move algorithms from llvm::parallel:: to llvm::, since they have
"parallel" in the name and are no longer overloads of the regular
algorithms.
- Add range overloads
- Use the sequential algorithm directly when 1 thread is requested
(skips task grouping)
- Fix the index type of parallelForEachN to size_t. Nobody in LLVM was
using any other parameter, and it made overload resolution hard for
for_each_n(par, 0, foo.size(), ...) because 0 is int, not size_t.
Remove Threads.h and update LLD for that.
This is a prerequisite for parallel public symbol processing in the PDB
library, which is in LLVM.
Reviewed By: MaskRay, aganea
Differential Revision: https://reviews.llvm.org/D79390
A linker will create .ARM.exidx sections for InputSections that don't
have them. This can cause a relocation out of range error If the
InputSection happens to be extremely far away from the other sections.
This is often the case for the vector table on older ARM CPUs as the only
two places that the table can be placed is 0 or 0xffff0000. We fix this
by removing InputSections that need a linker generated .ARM.exidx
section if that would cause an error.
Differential Revision: https://reviews.llvm.org/D79289
Fixed error detected by msan. The size field of the .ARM.exidx synthetic
section needs to be initialized to at least estimation level before
calling assignAddresses as that will use the size field.
This was previously reverted in 1ca16fc4f5.
Differential Revision: https://reviews.llvm.org/D78422
This reverts commit f969c2aa65.
There are some msan buildbot failures sanitzer-x86_64-linux-fast that
I need to investigate.
Differential Revision: https://reviews.llvm.org/D78422
The contents of the .ARM.exidx section must be ordered by SHF_LINK_ORDER
rules. We don't need to know the precise address for this order, but we
do need to know the relative order of sections. We have been using the
sectionIndex for this purpose, this works when the OutputSection order
has a monotonically increasing virtual address, but it is possible to
write a linker script with non-monotonically increasing virtual address.
For these cases we need to evaluate the base address of the OutputSection
so that we can order the .ARM.exidx sections properly.
This change moves the finalisation of .ARM.exidx till after the first
call to AssignAddresses. This permits us to sort on virtual address which
is linker script safe. It also permits a fix for part of pr44824 where
we generate .ARM.exidx section for the vector table when that table is so
far away it is out of range of the .ARM.exidx section. This fix will come
in a follow up patch.
Differential Revision: https://reviews.llvm.org/D78422
--no-threads is a name copied from gold.
gold has --no-thread, --thread-count and several other --thread-count-*.
There are needs to customize the number of threads (running several lld
processes concurrently or customizing the number of LTO threads).
Having a single --threads=N is a straightforward replacement of gold's
--no-threads + --thread-count.
--no-threads is used rarely. So just delete --no-threads instead of
keeping it for compatibility for a while.
If --threads= is specified (ELF,wasm; COFF /threads: is similar),
--thinlto-jobs= defaults to --threads=,
otherwise all available hardware threads are used.
There is currently no way to override a --threads={1,2,...}. It is still
a debate whether we should use --threads=all.
Reviewed By: rnk, aganea
Differential Revision: https://reviews.llvm.org/D76885
On an internal target,
* Before D74773: time -f '%M' => 18275680
* After D74773: time -f '%M' => 22088964
This patch restores to the status before D74773.
llvm::call_once(initDwarfLine, [this]() { initializeDwarf(); });
Though it is not used in all places.
I need that patch for implementing "Remove obsolete debug info" feature
(D74169). But this caching mechanism is useful by itself, and I think it
would be good to use it without connection to "Remove obsolete debug info"
feature. So this patch changes inplace creation of DWARFContext with
its cached version.
Depends on D74308
Reviewed By: ruiu
Differential Revision: https://reviews.llvm.org/D74773
Summary:
LLD has workaround for the times when SectionIndex was not passed properly:
LT->getFileLineInfoForAddress(
S->getOffsetInFile() + Offset, nullptr,
DILineInfoSpecifier::FileLineInfoKind::AbsoluteFilePath, Info));
S->getOffsetInFile() was added to differentiate offsets between
various sections. Now SectionIndex is properly specified.
Thus it is not necessary to use getOffsetInFile() workaround.
See https://reviews.llvm.org/D58194, https://reviews.llvm.org/D58357.
This patch removes getOffsetInFile() workaround.
Reviewers: ruiu, grimar, MaskRay, espindola
Reviewed By: grimar, MaskRay
Subscribers: emaste, arichardson, llvm-commits
Tags: #llvm, #lld
Differential Revision: https://reviews.llvm.org/D75636
Summary:
Extracted from D74773. Currently, errors happened while parsing
debug info are reported as errors. DebugInfoDWARF library treats such
errors as "Recoverable errors". This patch makes debug info errors
to be reported as warnings, to support DebugInfoDWARF approach.
Reviewers: ruiu, grimar, MaskRay, jhenderson, espindola
Reviewed By: MaskRay, jhenderson
Subscribers: emaste, aprantl, arichardson, arphaman, llvm-commits
Tags: #llvm, #debug-info, #lld
Differential Revision: https://reviews.llvm.org/D75234
Summary:
Generate PAC protected plt only when "-z pac-plt" is passed to the
linker. GNU toolchain generates when it is explicitly requested[1].
When pac-plt is requested then set the GNU_PROPERTY_AARCH64_FEATURE_1_PAC
note even when not all function compiled with PAC but issue a warning.
Harmonizing the warning style for BTI/PAC/IBT.
Generate BTI protected PLT if case of "-z force-bti".
[1] https://www.sourceware.org/ml/binutils/2019-03/msg00021.html
Reviewers: peter.smith, espindola, MaskRay, grimar
Reviewed By: peter.smith, MaskRay
Subscribers: tatyana-krasnukha, emaste, arichardson, kristof.beyls, MaskRay, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D74537
The goal of this patch is to maximize CPU utilization on multi-socket or high core count systems, so that parallel computations such as LLD/ThinLTO can use all hardware threads in the system. Before this patch, on Windows, a maximum of 64 hardware threads could be used at most, in some cases dispatched only on one CPU socket.
== Background ==
Windows doesn't have a flat cpu_set_t like Linux. Instead, it projects hardware CPUs (or NUMA nodes) to applications through a concept of "processor groups". A "processor" is the smallest unit of execution on a CPU, that is, an hyper-thread if SMT is active; a core otherwise. There's a limit of 32-bit processors on older 32-bit versions of Windows, which later was raised to 64-processors with 64-bit versions of Windows. This limit comes from the affinity mask, which historically is represented by the sizeof(void*). Consequently, the concept of "processor groups" was introduced for dealing with systems with more than 64 hyper-threads.
By default, the Windows OS assigns only one "processor group" to each starting application, in a round-robin manner. If the application wants to use more processors, it needs to programmatically enable it, by assigning threads to other "processor groups". This also means that affinity cannot cross "processor group" boundaries; one can only specify a "preferred" group on start-up, but the application is free to allocate more groups if it wants to.
This creates a peculiar situation, where newer CPUs like the AMD EPYC 7702P (64-cores, 128-hyperthreads) are projected by the OS as two (2) "processor groups". This means that by default, an application can only use half of the cores. This situation could only get worse in the years to come, as dies with more cores will appear on the market.
== The problem ==
The heavyweight_hardware_concurrency() API was introduced so that only *one hardware thread per core* was used. Once that API returns, that original intention is lost, only the number of threads is retained. Consider a situation, on Windows, where the system has 2 CPU sockets, 18 cores each, each core having 2 hyper-threads, for a total of 72 hyper-threads. Both heavyweight_hardware_concurrency() and hardware_concurrency() currently return 36, because on Windows they are simply wrappers over std:🧵:hardware_concurrency() -- which can only return processors from the current "processor group".
== The changes in this patch ==
To solve this situation, we capture (and retain) the initial intention until the point of usage, through a new ThreadPoolStrategy class. The number of threads to use is deferred as late as possible, until the moment where the std::threads are created (ThreadPool in the case of ThinLTO).
When using hardware_concurrency(), setting ThreadCount to 0 now means to use all the possible hardware CPU (SMT) threads. Providing a ThreadCount above to the maximum number of threads will have no effect, the maximum will be used instead.
The heavyweight_hardware_concurrency() is similar to hardware_concurrency(), except that only one thread per hardware *core* will be used.
When LLVM_ENABLE_THREADS is OFF, the threading APIs will always return 1, to ensure any caller loops will be exercised at least once.
Differential Revision: https://reviews.llvm.org/D71775
This adds some of LLD specific scopes and picks up optimisation scopes
via LTO/ThinLTO. Makes use of TimeProfiler multi-thread support added in
77e6bb3c.
Differential Revision: https://reviews.llvm.org/D71060
-fno-pie produces a pair of non-GOT-non-PLT relocations R_PPC_ADDR16_{HA,LO} (R_ABS) referencing external
functions.
```
lis 3, func@ha
la 3, func@l(3)
```
In a -no-pie/-pie link, if func is not defined in the executable, a canonical PLT entry (st_value>0, st_shndx=0) will be needed.
References to func in shared objects will be resolved to this address.
-fno-pie -pie should fail with "can't create dynamic relocation ... against ...", so we just need to think about -no-pie.
On x86, the PLT entry passes the JMP_SLOT offset to the rtld PLT resolver.
On x86-64: the PLT entry passes the JUMP_SLOT index to the rtld PLT resolver.
On ARM/AArch64: the PLT entry passes &.got.plt[n]. The PLT header passes &.got.plt[fixed-index]. The rtld PLT resolver can compute the JUMP_SLOT index from the two addresses.
For these targets, the canonical PLT entry can just reuse the regular PLT entry (in PltSection).
On PPC32: PltSection (.glink) consists of `b PLTresolve` instructions and `PLTresolve`. The rtld PLT resolver depends on r11 having been set up to the .plt (GotPltSection) entry.
On PPC64 ELFv2: PltSection (.glink) consists of `__glink_PLTresolve` and `bl __glink_PLTresolve`. The rtld PLT resolver depends on r12 having been set up to the .plt (GotPltSection) entry.
We cannot reuse a `b PLTresolve`/`bl __glink_PLTresolve` in PltSection as a canonical PLT entry. PPC64 ELFv2 avoids the problem by using TOC for any external reference, even in non-pic code, so the canonical PLT entry scenario should not happen in the first place.
For PPC32, we have to create a PLT call stub as the canonical PLT entry. The code sequence sets up r11.
Reviewed By: Bdragon28
Differential Revision: https://reviews.llvm.org/D73399
Symbol information can be used to improve out-of-range/misalignment diagnostics.
It also helps R_ARM_CALL/R_ARM_THM_CALL which has different behaviors with different symbol types.
There are many (67) relocateOne() call sites used in thunks, {Arm,AArch64}errata, PLT, etc.
Rename them to `relocateNoSym()` to be clearer that there is no symbol information.
Reviewed By: grimar, peter.smith
Differential Revision: https://reviews.llvm.org/D73254
Fangrui Song