Before Windows 11 and Windows Server 2022, only one 'processor group' is assigned by default to a starting process, then the program is responsible for dispatching its own threads on more 'processor groups'. That is what 8404aeb56a73ab24f9b295111de3b37a37f0b841 was doing, allowing LLVM tools to automatically use all hardware threads in the machine.
After Windows 11 and Windows Server 2022, the OS takes care of that. This has an adverse effect reported in #56618 which is that using `GetProcessAffinityMask()` API in some edge cases seems buggy now. That API is used to detect if an affinity mask was set, and adjust accordingly the available threads for a ThreadPool.
With this patch, on one hand, we let the OS dispatch threads on all 'processor groups', but only for Windows 11 & Windows Server 2022 and after. We retain the old behavior for older OS versions. On the other hand, a workaround was added to mitigate the `GetProcessAffinityMask()` issue described above (see Threading.inc, L226).
Differential Revision: https://reviews.llvm.org/D138747
This reverts commit 5e50b8089aee249d77542ea858d956568ec6581f.
This commit breaks the build for BOLT:
bolt/lib/Profile/DataAggregator.cpp:264:66: error: no viable
conversion from 'Optional<StringRef>[3]' to
'ArrayRef<std::optional<StringRef>>'
This change is focussed on simplifying `Support/Host.h` to only do
target detection. In this case, this function is close in usage to
existing functions in `Support/Threading.h`, so I moved it into there.
The function is also renamed to `llvm::get_physical_cores()` to match
the style of threading's functions.
The big change here is that now if you have threading disabled,
`llvm::get_physical_cores()` will return -1, as if it had not been able
to work out the right info. This is due to how Threading.cpp includes
OS-specific code/headers. This seems ok, as if threading is disabled,
LLVM should not need to know the number of physical cores.
Differential Revision: https://reviews.llvm.org/D137836
Apply clang-format on llvm/lib/Support/Windows/ and llvm/lib/Support/Unix/ since .inc files in these folders aren't picked up by default. Eventually we need to add this extension in the monorepo .clang-format file.
Differential Revision: https://reviews.llvm.org/D138714
This reverts commit 9969ceb36b440eaafa17c486f29a69c7a7da3b3b.
On Windows:
lld-link: error: undefined symbol: int __cdecl computeHostNumPhysicalCores(void)
>>> referenced by LLVMSupport.lib(Support.Host.obj):(int __cdecl llvm::sys::getHostNumPhysicalCores(void))
On Apple Silicon Macs, using a Darwin thread priority of PRIO_DARWIN_BG seems to
map directly to the QoS class Background. With this priority, the thread is
confined to efficiency cores only, which makes background indexing take forever.
Introduce a new ThreadPriority "Low" that sits in the middle between Background
and Default, and maps to QoS class "Utility" on Mac. Make this new priority the
default for indexing. This makes the thread run on all cores, but still lowers
priority enough to keep the machine responsive, and not interfere with
user-initiated actions.
I didn't change the implementations for Windows and Linux; on these systems,
both ThreadPriority::Background and ThreadPriority::Low map to the same thread
priority. This could be changed as a followup (e.g. by using SCHED_BATCH for Low
on Linux).
See also https://github.com/clangd/clangd/issues/1119.
Reviewed By: sammccall, dgoldman
Differential Revision: https://reviews.llvm.org/D124715
FreeBSD's condvar.h (included by user.h in Threading.inc) uses a "struct
thread" that conflicts with llvm::thread if both are visible when it's
included.
So this moves our #include after the FreeBSD code.
This adds a new llvm::thread class with the same interface as std::thread
except there is an extra constructor that allows us to set the new thread's
stack size. On Darwin even the default size is boosted to 8MB to match the main
thread.
It also switches all users of the older C-style `llvm_execute_on_thread` API
family over to `llvm::thread` followed by either a `detach` or `join` call and
removes the old API.
Moved definition of DefaultStackSize into the .cpp file to hopefully
fix the build on some (GCC-6?) machines.
This adds a new llvm::thread class with the same interface as std::thread
except there is an extra constructor that allows us to set the new thread's
stack size. On Darwin even the default size is boosted to 8MB to match the main
thread.
It also switches all users of the older C-style `llvm_execute_on_thread` API
family over to `llvm::thread` followed by either a `detach` or `join` call and
removes the old API.
The number of hardware threads available to a ThreadPool can be limited if setting an affinity mask.
For example:
> start /B /AFFINITY 0xF lld-link.exe ...
Would let LLD only use 4 hyper-threads.
Previously, there was an outstanding issue on Windows Server 2019 on dual-CPU machines, which was preventing from using both CPU sockets. In normal conditions, when no affinity mask was set, ProcessorGroup::AllThreads was different from ProcessorGroup::UsableThreads. The previous code in llvm/lib/Support/Windows/Threading.inc L201 was improperly assuming those two values to be equal, and consequently was limiting the execution to only one CPU socket.
Differential Revision: https://reviews.llvm.org/D92419
The number of hardware threads available to a ThreadPool can be limited if setting an affinity mask.
For example:
> start /B /AFFINITY 0xF lld-link.exe ...
Would let LLD only use 4 hyper-threads.
Previously, there was an outstanding issue on Windows Server 2019 on dual-CPU machines, which was preventing from using both CPU sockets. In normal conditions, when no affinity mask was set, ProcessorGroup::AllThreads was different from ProcessorGroup::UsableThreads. The previous code in llvm/lib/Support/Windows/Threading.inc L201 was improperly assuming those two values to be equal, and consequently was limiting the execution to only one CPU socket.
Differential Revision: https://reviews.llvm.org/D92419
Before this patch, it wasn't possible to extend the ThinLTO threads to all SMT/CMT threads in the system. Only one thread per core was allowed, instructed by usage of llvm::heavyweight_hardware_concurrency() in the ThinLTO code. Any number passed to the LLD flag /opt:lldltojobs=..., or any other ThinLTO-specific flag, was previously interpreted in the context of llvm::heavyweight_hardware_concurrency(), which means SMT disabled.
One can now say in LLD:
/opt:lldltojobs=0 -- Use one std::thread / hardware core in the system (no SMT). Default value if flag not specified.
/opt:lldltojobs=N -- Limit usage to N threads, regardless of usage of heavyweight_hardware_concurrency().
/opt:lldltojobs=all -- Use all hardware threads in the system. Equivalent to /opt:lldltojobs=$(nproc) on Linux and /opt:lldltojobs=%NUMBER_OF_PROCESSORS% on Windows. When an affinity mask is set for the process, threads will be created only for the cores selected by the mask.
When N > number-of-hardware-threads-in-the-system, the threads in the thread pool will be dispatched equally on all CPU sockets (tested only on Windows).
When N <= number-of-hardware-threads-on-a-CPU-socket, the threads will remain on the CPU socket where the process started (only on Windows).
Differential Revision: https://reviews.llvm.org/D75153
llvm-ar is using CompareStringOrdinal which is available
only starting with Windows Vista (WINVER 0x600).
Fix this by hoising WindowsSupport.h, which sets _WIN32_WINNT
to 0x0601, up to llvm/include/llvm/Support and use it in llvm-ar.
Patch by Cristian Adam!
Differential revision: https://reviews.llvm.org/D74599
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 reverts commit 40668abca4d307e02b33345cfdb7271549ff48d0.
This causes clang tests to fail, as stacksize=0 is being explicitly passed and
is no longer a no-op.
This roughly mimics `std::thread(...).detach()` except it allows to
customize the stack size. Required for https://reviews.llvm.org/D50993.
I've decided against reusing the existing `llvm_execute_on_thread` because
it's not obvious what to do with the ownership of the passed
function/arguments:
1. If we pass possibly owning functions data to `llvm_execute_on_thread`,
we'll lose the ability to pass small non-owning non-allocating functions
for the joining case (as it's used now). Is it important enough?
2. If we use the non-owning interface in the new use case, we'll force
clients to transfer ownership to the spawned thread manually, but
similar code would still have to exist inside
`llvm_execute_on_thread(_async)` anyway (as we can't just pass the same
non-owning pointer to pthreads and Windows implementations, and would be
forced to wrap it in some structure, and deal with its ownership.
Patch by Dmitry Kozhevnikov!
Differential Revision: https://reviews.llvm.org/D51103
Summary:
We have a multi-platform thread priority setting function(last piece
landed with D58683), I wanted to make this available to all llvm community,
there seem to be other users of such functionality with portability fixmes:
lib/Support/CrashRecoveryContext.cpp
tools/clang/tools/libclang/CIndex.cpp
Reviewers: gribozavr, ioeric
Subscribers: krytarowski, jfb, kristina, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59130
llvm-svn: 358494
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
Move the __try/__except block outside of the set_thread_name function to avoid a conflict with object unwinding due to the use of the llvm::Storage.
Differential Revision: https://reviews.llvm.org/D30707
llvm-svn: 297192
Applications often need the current thread id when making
system calls, and some operating systems provide the notion
of a thread name, which can be useful in enabling better
diagnostics when debugging or logging.
This patch adds an accessor for the thread id, and "best effort"
getters and setters for the thread name. Since this is
non critical functionality, no error is returned to indicate
that a platform doesn't support thread names.
Differential Revision: https://reviews.llvm.org/D30526
llvm-svn: 296887
This reverts commit r221331 and reinstate r220932 as discussed in D19271.
Original commit message was:
This patch adds an llvm_call_once which is a wrapper around
std::call_once on platforms where it is available and devoid
of bugs. The patch also migrates the ManagedStatic mutex to
be allocated using llvm_call_once.
These changes are philosophically equivalent to the changes
added in r219638, which were reverted due to a hang on Win32
which was the result of a bug in the Windows implementation
of std::call_once.
Differential Revision: http://reviews.llvm.org/D5922
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 269577
Commit 220932 caused crash when building clang-tblgen on aarch64 debian target,
so it's blocking all daily tests.
The std::call_once implementation in pthread has bug for aarch64 debian.
llvm-svn: 221331
1>C:\Program Files (x86)\Windows Kits\8.1\Include\um\minwinbase.h(46):
error C2146: syntax error : missing ';' before identifier 'nLength'
1>C:\Program Files (x86)\Windows Kits\8.1\Include\um\minwinbase.h(46):
error C4430: missing type specifier - int assumed. Note: C++ does not support default-int
...
including <windows.h> is actually required.
llvm-svn: 221244
Summary:
This patch adds an llvm_call_once which is a wrapper around std::call_once on platforms where it is available and devoid of bugs. The patch also migrates the ManagedStatic mutex to be allocated using llvm_call_once.
These changes are philosophically equivalent to the changes added in r219638, which were reverted due to a hang on Win32 which was the result of a bug in the Windows implementation of std::call_once.
Reviewers: aaron.ballman, chapuni, chandlerc, rnk
Reviewed By: rnk
Subscribers: majnemer, llvm-commits
Differential Revision: http://reviews.llvm.org/D5922
llvm-svn: 220932
