[ThinLTO] Allow usage of all SMT threads in the system

[aganea]

Summary

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.

After this patch, one can 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 N > number-of-hardware-threads-in-the-system, the std::threads will be dispatched equally on all CPU sockets (tested only on Windows).
When N <= number-of-hardware-threads-on-a-CPU-socket, the std::threads will remain on the CPU socket where the process started (only on Windows).

All cmd-line flags and code paths that lead to ThinLTO have been modified by this patch:
-flto-jobs=...
--thinlto-jobs=...
-thinlto-threads=...
--plugin-opt=jobs=...

This is a follow-up for: https://reviews.llvm.org/D71775#1891709 and: https://reviews.llvm.org/D71775#1892013

[aganea]

If a simpler (more yucky?) patch is needed to fix what @thakis was suggesting in https://reviews.llvm.org/D71775#1891709, and if we don't want this extra new flag, we can also check the CPU brand for "AMD Opteron", and keep the old behavior in that case.

[RobRich999]

Based upon the description, I think this patch is more applicable than just targeting a specific AMD proc family since it allows the end-user a choice for maximizing threading with both CMT and SMT on all supported platforms.

BTW, until if/when this patch lands, I just set a static value in source as a local workaround for now.

llvm/lib/Support/Host.cpp

#elif defined(_WIN32)
// Defined in llvm/lib/Support/Windows/Threading.inc
static int computeHostNumPhysicalCores() { return 32; }
#else

[mehdi_amini]

/opt:lldltojobs=N -- limit usage to N threads, but constrained by usage of heavyweight_hardware_concurrency().

I really dislike this behavior: this seems user hostile to me. I would either:

• honor the user request (eventually display a warning), this is in line with other system behavior like ninja -j N for instance.
• reject the user request

If you want such a behavior, then it should be another flag which express it in the name like opt:lldltomaxnumjobs.

[aganea]

@mehdi_amini Agreed. In that case, I just need to slightly change this function to ensure threads are equally dispatched on all CPU sockets regardless of the thread strategy.

[aganea]

Simplify. Revert to previous behavior, where any number of threads can be specified on the cmd-line, as suggested by @mehdi_amini. See the updated Summary for cmd-line usage.

+ @dexonsmith for the Darwin part.
+ @ruiu for the LLD part.

[rnk]

I'm busy and I haven't looked at the code in detail, but I'm OK with going back to the old way of doing things. I think honoring user requests to use more threads than cores is an important use case. We had plans at one point to add indirection points to allow us to distribute these backend actions to other machines. Essentially, this would be done by having LLD invoke $mydiscc clang -c foo.o -fuse-thin-lto-index=foo.thinlto.bc. We've gone a different direction, so that change is not likely to come soon, but it seems like a reasonable use case where one would want to pass -j LARGE and have it be honored.

[aganea]

Thanks Reid! I will leave this open for a few days, in case there are any other feedbacks.

As for -fthinlto-index, we have someone looking at distributing it remotely with Fastbuild. I think it is reasonable on the short-term to let the build system handle that (I assume that's what you did). One adjacent question is how to address the size of the local cache folder (ie. lto.cache when compiling LLVM) which grows big very quickly. Pruning periodically is fine, but I wonder if we could keep it compressed on disk. Maybe do DeviceIoControl (...FSCTL_SET_COMPRESSION..) and let Windows handle that? I'll take a look in the weeks to come.

[rnk]

I think making the build system responsible for running the backend actions is the ideal solution, actually. The main reason we have all this threading logic in the linker is to make it easy for users of traditional build systems to use ThinLTO with a few flag flips. We haven't actually made build system changes for Chrome yet. Instead, we replaced the linker with a script that wraps the thin link + backend jobs + native link steps. https://crbug.com/877722 has a partial record of ideas we had and stuff we tried for Chrome.

[MaskRay]

Does taskset -c 0-3 lld -flavor ... restrict the number of cores?

cpu_set_t cpu;
sched_getaffinity(0, sizeof(cpu), &cpu);
CPU_COUNT(&cpu)

[abrachet]

llvm/include/llvm/Support/Threading.h

201

Nit: Remove inline https://llvm.org/docs/CodingStandards.html#don-t-use-inline-when-defining-a-function-in-a-class-definition

[aganea]

In D75153#1906580, @MaskRay wrote:

Does taskset -c 0-3 lld -flavor ... restrict the number of cores?

cpu_set_t cpu;
sched_getaffinity(0, sizeof(cpu), &cpu);
CPU_COUNT(&cpu)

Thanks for raising this! This does not seem to work (I currently only have WSL at hand, no real Linux machine). I don't think it worked before my patch. The current code in LLVM is written such as: (note the "if" statement)

#if defined(HAVE_SCHED_GETAFFINITY) && defined(HAVE_CPU_COUNT)
  cpu_set_t Set;
  if (sched_getaffinity(0, sizeof(Set), &Set))
    return CPU_COUNT(&Set);
#endif

The doc for sched_getaffinity says:

On success, sched_setaffinity() and sched_getaffinity() return 0. On error, -1 is returned, and errno is set appropriately.

So it would always fall back to std::thread::hardware_concurrency, which apparently does not always take affinity into account, according to @respindola (please see rG8c0ff9508da5f02e8ce6580a126a2018c9bf702a).

I'll write a follow-up patch to test affinity on Linux and Windows.

llvm/include/llvm/Support/Threading.h

201

After discussing offling with @abrachet , I'll leave the inline for now. It makes the symbol weak, removing inline would otherwise fail linking. I can move the function(s) to the .CPP after this patch to save on link time.

[MaskRay]

lld/test/ELF/basic.s

252

This change is not needed. lto/thinlto.ll has already tested the functionally.

basic.s should also be split. I did this in 34bdddf9a13cfdbbb5506dc89cf8e781be53105f

253

-verbose is not needed because verbose just prints input filenames, which has nothing to do with --thinlto-jobs=0`.

[zero9178]

This patch seems to break when disabling threading (aka LLVM_ENABLE_THREADS == 0) as get_threadpool_strategy is undefined therefore creating linker failures in clang. (Tested on Windows)

[aganea]

In D75153#1947956, @zero9178 wrote:

This patch seems to break when disabling threading (aka LLVM_ENABLE_THREADS == 0) as get_threadpool_strategy is undefined therefore creating linker failures in clang. (Tested on Windows)

Taking a look now.

[aganea]

In D75153#1947956, @zero9178 wrote:

This patch seems to break when disabling threading (aka LLVM_ENABLE_THREADS == 0) as get_threadpool_strategy is undefined therefore creating linker failures in clang. (Tested on Windows)

Should be fixed after rG3ab3f3c5d5825476dc1be15992f7c964629de688.

[phosek]

We've started seeing llvm-cov on our Linux bots with this error:

terminating with uncaught exception of type std::__2::system_error: thread constructor failed: Resource temporarily unavailable

Specifically, we're running llvm export which uses heavyweight_hardware_concurrency (we use the default number of threads, i.e. 0): https://github.com/llvm/llvm-project/blob/master/llvm/tools/llvm-cov/CoverageExporterJson.cpp#L169

I'm not yet sure what's the problem, but bisecting is pointing at this change.

[phosek]

In D75153#1987272, @phosek wrote:

We've started seeing llvm-cov on our Linux bots with this error:

terminating with uncaught exception of type std::__2::system_error: thread constructor failed: Resource temporarily unavailable

Specifically, we're running llvm export which uses heavyweight_hardware_concurrency (we use the default number of threads, i.e. 0): https://github.com/llvm/llvm-project/blob/master/llvm/tools/llvm-cov/CoverageExporterJson.cpp#L169

I'm not yet sure what's the problem, but bisecting is pointing at this change.

Also on runs that succeeded, we see the execution times more than doubled.

[aganea]

In D75153#1987320, @phosek wrote:

In D75153#1987272, @phosek wrote:

We've started seeing llvm-cov on our Linux bots with this error:

terminating with uncaught exception of type std::__2::system_error: thread constructor failed: Resource temporarily unavailable

Specifically, we're running llvm export which uses heavyweight_hardware_concurrency (we use the default number of threads, i.e. 0): https://github.com/llvm/llvm-project/blob/master/llvm/tools/llvm-cov/CoverageExporterJson.cpp#L169

I'm not yet sure what's the problem, but bisecting is pointing at this change.

Also on runs that succeeded, we see the execution times more than doubled.

This is caused by a mix of the previous change (rG8404aeb5, see https://github.com/llvm/llvm-project/commit/8404aeb56a73ab24f9b295111de3b37a37f0b841#diff-9c7f49c15e22d38241ccb9d294f880f4L950) which was restraining the ThreadPool to the number of hardware threads, and this patch which makes the number of threads in the ThreadPool limitless.

I will revert the behavior of llvm-cov to previous state, and will check if there are other cases like this.

[MaskRay]

In D75153#1942336, @aganea wrote:

In D75153#1906580, @MaskRay wrote:

Does taskset -c 0-3 lld -flavor ... restrict the number of cores?

cpu_set_t cpu;
sched_getaffinity(0, sizeof(cpu), &cpu);
CPU_COUNT(&cpu)

Thanks for raising this! This does not seem to work (I currently only have WSL at hand, no real Linux machine). I don't think it worked before my patch. The current code in LLVM is written such as: (note the "if" statement)

#if defined(HAVE_SCHED_GETAFFINITY) && defined(HAVE_CPU_COUNT)
  cpu_set_t Set;
  if (sched_getaffinity(0, sizeof(Set), &Set))
    return CPU_COUNT(&Set);
#endif

The doc for sched_getaffinity says:

On success, sched_setaffinity() and sched_getaffinity() return 0. On error, -1 is returned, and errno is set appropriately.

So it would always fall back to std::thread::hardware_concurrency, which apparently does not always take affinity into account, according to @respindola (please see rG8c0ff9508da5f02e8ce6580a126a2018c9bf702a).

I'll write a follow-up patch to test affinity on Linux and Windows.

Created D78324 to fix this pre-existing issue. I remember I saw a related bugs.llvm.org report yesterday but I can't find it now...

[aganea]

In D75153#1987538, @MaskRay wrote:

I remember I saw a related bugs.llvm.org report yesterday but I can't find it now...

This? https://bugs.llvm.org/show_bug.cgi?id=45556

[thakis]

https://clang.llvm.org/docs/ThinLTO.html#controlling-backend-parallelism should probably be updated to mention "all".

clang/lib/Driver/ToolChains/CommonArgs.h

bool addOpenMPRuntime(llvm::opt::ArgStringList &CmdArgs, const ToolChain &TC,
                       bool ForceStaticHostRuntime = false,
                       bool IsOffloadingHost = false, bool GompNeedsRT = false);
 
 llvm::opt::Arg *getLastProfileUseArg(const llvm::opt::ArgList &Args);
 llvm::opt::Arg *getLastProfileSampleUseArg(const llvm::opt::ArgList &Args);
 
 bool isObjCAutoRefCount(const llvm::opt::ArgList &Args);
 
-unsigned getLTOParallelism(const llvm::opt::ArgList &Args, const Driver &D);
+llvm::StringRef getLTOParallelism(const llvm::opt::ArgList &Args,
+                                  const Driver &D);
 
 bool areOptimizationsEnabled(const llvm::opt::ArgList &Args);
 
 bool isUseSeparateSections(const llvm::Triple &Triple);
 
 /// \p EnvVar is split by system delimiter for environment variables.
 /// If \p ArgName is "-I", "-L", or an empty string, each entry from \p EnvVar
 /// is prefixed by \p ArgName then added to \p Args. Otherwise, for each

clang/lib/Driver/ToolChains/CommonArgs.cpp

 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/FileSystem.h"
 #include "llvm/Support/Host.h"
 #include "llvm/Support/Path.h"
 #include "llvm/Support/Process.h"
 #include "llvm/Support/Program.h"
 #include "llvm/Support/ScopedPrinter.h"
 #include "llvm/Support/TargetParser.h"
+#include "llvm/Support/Threading.h"
 #include "llvm/Support/VirtualFileSystem.h"
 #include "llvm/Support/YAMLParser.h"
 
 using namespace clang::driver;
 using namespace clang::driver::tools;
 using namespace clang;
 using namespace llvm::opt;
 
case llvm::Triple::amdgcn:
     return getR600TargetGPU(Args);
 
   case llvm::Triple::wasm32:
   case llvm::Triple::wasm64:
     return std::string(getWebAssemblyTargetCPU(Args));
   }
 }
 
-unsigned tools::getLTOParallelism(const ArgList &Args, const Driver &D) {
-  unsigned Parallelism = 0;
+llvm::StringRef tools::getLTOParallelism(const ArgList &Args, const Driver &D) {
   Arg *LtoJobsArg = Args.getLastArg(options::OPT_flto_jobs_EQ);
-  if (LtoJobsArg &&
-      StringRef(LtoJobsArg->getValue()).getAsInteger(10, Parallelism))
-    D.Diag(diag::err_drv_invalid_int_value) << LtoJobsArg->getAsString(Args)
-                                            << LtoJobsArg->getValue();
-  return Parallelism;
+  if (!LtoJobsArg)
+    return {};
+  if (!llvm::get_threadpool_strategy(LtoJobsArg->getValue()))
+    D.Diag(diag::err_drv_invalid_int_value)
+        << LtoJobsArg->getAsString(Args) << LtoJobsArg->getValue();
+  return LtoJobsArg->getValue();
 }
 
 // CloudABI uses -ffunction-sections and -fdata-sections by default.
 bool tools::isUseSeparateSections(const llvm::Triple &Triple) {
   return Triple.getOS() == llvm::Triple::CloudABI;
 }
 
 void tools::addLTOOptions(const ToolChain &ToolChain, const ArgList &Args,
if (Args.hasArg(options::OPT_gsplit_dwarf)) {
     CmdArgs.push_back(
         Args.MakeArgString(Twine("-plugin-opt=dwo_dir=") +
             Output.getFilename() + "_dwo"));
   }
 
   if (IsThinLTO)
     CmdArgs.push_back("-plugin-opt=thinlto");
 
-  if (unsigned Parallelism = getLTOParallelism(Args, ToolChain.getDriver()))
+  StringRef Parallelism = getLTOParallelism(Args, ToolChain.getDriver());
+  if (!Parallelism.empty())
     CmdArgs.push_back(
         Args.MakeArgString("-plugin-opt=jobs=" + Twine(Parallelism)));
 
   // If an explicit debugger tuning argument appeared, pass it along.
   if (Arg *A = Args.getLastArg(options::OPT_gTune_Group,
                                options::OPT_ggdbN_Group)) {
     if (A->getOption().matches(options::OPT_glldb))
       CmdArgs.push_back("-plugin-opt=-debugger-tune=lldb");

clang/lib/Driver/ToolChains/Darwin.cpp

 #include "clang/Driver/Options.h"
 #include "clang/Driver/SanitizerArgs.h"
 #include "llvm/ADT/StringSwitch.h"
 #include "llvm/Option/ArgList.h"
 #include "llvm/ProfileData/InstrProf.h"
 #include "llvm/Support/Path.h"
 #include "llvm/Support/ScopedPrinter.h"
 #include "llvm/Support/TargetParser.h"
+#include "llvm/Support/Threading.h"
 #include "llvm/Support/VirtualFileSystem.h"
 #include <cstdlib> // ::getenv
 
 using namespace clang::driver;
 using namespace clang::driver::tools;
 using namespace clang::driver::toolchains;
 using namespace clang;
 using namespace llvm::opt;
if (LinkingOutput) {
     CmdArgs.push_back(LinkingOutput);
   }
 
   if (Args.hasArg(options::OPT_fnested_functions))
     CmdArgs.push_back("-allow_stack_execute");
 
   getMachOToolChain().addProfileRTLibs(Args, CmdArgs);
 
-  if (unsigned Parallelism =
-          getLTOParallelism(Args, getToolChain().getDriver())) {
+  StringRef Parallelism = getLTOParallelism(Args, getToolChain().getDriver());
+  if (!Parallelism.empty()) {
     CmdArgs.push_back("-mllvm");
-    CmdArgs.push_back(Args.MakeArgString("-threads=" + Twine(Parallelism)));
+    unsigned NumThreads =
+        llvm::get_threadpool_strategy(Parallelism)->compute_thread_count();
+    CmdArgs.push_back(Args.MakeArgString("-threads=" + Twine(NumThreads)));
   }
 
   if (getToolChain().ShouldLinkCXXStdlib(Args))
     getToolChain().AddCXXStdlibLibArgs(Args, CmdArgs);
 
   bool NoStdOrDefaultLibs =
       Args.hasArg(options::OPT_nostdlib, options::OPT_nodefaultlibs);
   bool ForceLinkBuiltins = Args.hasArg(options::OPT_fapple_link_rtlib);

lld/COFF/Config.h

struct Configuration {
   bool safeSEH = false;
   Symbol *sehTable = nullptr;
   Symbol *sehCount = nullptr;
 
   // Used for /opt:lldlto=N
   unsigned ltoo = 2;
 
   // Used for /opt:lldltojobs=N
-  unsigned thinLTOJobs = 0;
+  std::string thinLTOJobs;
   // Used for /opt:lldltopartitions=N
   unsigned ltoPartitions = 1;
 
   // Used for /opt:lldltocache=path
   StringRef ltoCache;
   // Used for /opt:lldltocachepolicy=policy
   llvm::CachePruningPolicy ltoCachePolicy;
 

lld/COFF/Driver.cpp

for (StringRef s : vec) {
       } else if (s == "nolldtailmerge") {
         tailMerge = 0;
       } else if (s.startswith("lldlto=")) {
         StringRef optLevel = s.substr(7);
         if (optLevel.getAsInteger(10, config->ltoo) || config->ltoo > 3)
           error("/opt:lldlto: invalid optimization level: " + optLevel);
       } else if (s.startswith("lldltojobs=")) {
         StringRef jobs = s.substr(11);
-        if (jobs.getAsInteger(10, config->thinLTOJobs) ||
-            config->thinLTOJobs == 0)
+        if (!get_threadpool_strategy(jobs))
           error("/opt:lldltojobs: invalid job count: " + jobs);
+        config->thinLTOJobs = jobs.str();
       } else if (s.startswith("lldltopartitions=")) {
         StringRef n = s.substr(17);
         if (n.getAsInteger(10, config->ltoPartitions) ||
             config->ltoPartitions == 0)
           error("/opt:lldltopartitions: invalid partition count: " + n);
       } else if (s != "lbr" && s != "nolbr")
         error("/opt: unknown option: " + s);
     }

lld/COFF/LTO.cpp

BitcodeCompiler::BitcodeCompiler() {
   // Initialize ltoObj.
   lto::ThinBackend backend;
   if (config->thinLTOIndexOnly) {
     auto OnIndexWrite = [&](StringRef S) { thinIndices.erase(S); };
     backend = lto::createWriteIndexesThinBackend(
         std::string(config->thinLTOPrefixReplace.first),
         std::string(config->thinLTOPrefixReplace.second),
         config->thinLTOEmitImportsFiles, indexFile.get(), OnIndexWrite);
-  } else if (config->thinLTOJobs != 0) {
-    backend = lto::createInProcessThinBackend(config->thinLTOJobs);
+  } else {
+    backend = lto::createInProcessThinBackend(
+        llvm::heavyweight_hardware_concurrency(config->thinLTOJobs));
   }
 
   ltoObj = std::make_unique<lto::LTO>(createConfig(), backend,
                                        config->ltoPartitions);
 }
 
 BitcodeCompiler::~BitcodeCompiler() = default;
 

lld/ELF/Config.h

struct Configuration {
   llvm::Optional<uint64_t> imageBase;
   uint64_t commonPageSize;
   uint64_t maxPageSize;
   uint64_t mipsGotSize;
   uint64_t zStackSize;
   unsigned ltoPartitions;
   unsigned ltoo;
   unsigned optimize;
-  unsigned thinLTOJobs;
+  StringRef thinLTOJobs;
   unsigned timeTraceGranularity;
   int32_t splitStackAdjustSize;
 
   // The following config options do not directly correspond to any
   // particular command line options.
 
   // True if we need to pass through relocations in input files to the
   // output file. Usually false because we consume relocations.

lld/ELF/Driver.cpp

static void readConfigs(opt::InputArgList &args) {
   config->thinLTOCacheDir = args.getLastArgValue(OPT_thinlto_cache_dir);
   config->thinLTOCachePolicy = CHECK(
       parseCachePruningPolicy(args.getLastArgValue(OPT_thinlto_cache_policy)),
       "--thinlto-cache-policy: invalid cache policy");
   config->thinLTOEmitImportsFiles = args.hasArg(OPT_thinlto_emit_imports_files);
   config->thinLTOIndexOnly = args.hasArg(OPT_thinlto_index_only) ||
                              args.hasArg(OPT_thinlto_index_only_eq);
   config->thinLTOIndexOnlyArg = args.getLastArgValue(OPT_thinlto_index_only_eq);
-  config->thinLTOJobs = args::getInteger(args, OPT_thinlto_jobs, -1u);
+  config->thinLTOJobs = args.getLastArgValue(OPT_thinlto_jobs);
   config->thinLTOObjectSuffixReplace =
       getOldNewOptions(args, OPT_thinlto_object_suffix_replace_eq);
   config->thinLTOPrefixReplace =
       getOldNewOptions(args, OPT_thinlto_prefix_replace_eq);
   config->timeTraceEnabled = args.hasArg(OPT_time_trace);
   config->timeTraceGranularity =
       args::getInteger(args, OPT_time_trace_granularity, 500);
   config->trace = args.hasArg(OPT_trace);
static void readConfigs(opt::InputArgList &args) {
   // Parse -mllvm options.
   for (auto *arg : args.filtered(OPT_mllvm))
     parseClangOption(arg->getValue(), arg->getSpelling());
 
   if (config->ltoo > 3)
     error("invalid optimization level for LTO: " + Twine(config->ltoo));
   if (config->ltoPartitions == 0)
     error("--lto-partitions: number of threads must be > 0");
-  if (config->thinLTOJobs == 0)
-    error("--thinlto-jobs: number of threads must be > 0");
+  if (!get_threadpool_strategy(config->thinLTOJobs))
+    error("--thinlto-jobs: invalid job count: " + config->thinLTOJobs);
 
   if (config->splitStackAdjustSize < 0)
     error("--split-stack-adjust-size: size must be >= 0");
 
   // The text segment is traditionally the first segment, whose address equals
   // the base address. However, lld places the R PT_LOAD first. -Ttext-segment
   // is an old-fashioned option that does not play well with lld's layout.
   // Suggest --image-base as a likely alternative.

lld/ELF/LTO.cpp

BitcodeCompiler::BitcodeCompiler() {
   // Initialize ltoObj.
   lto::ThinBackend backend;
   if (config->thinLTOIndexOnly) {
     auto onIndexWrite = [&](StringRef s) { thinIndices.erase(s); };
     backend = lto::createWriteIndexesThinBackend(
         std::string(config->thinLTOPrefixReplace.first),
         std::string(config->thinLTOPrefixReplace.second),
         config->thinLTOEmitImportsFiles, indexFile.get(), onIndexWrite);
-  } else if (config->thinLTOJobs != -1U) {
-    backend = lto::createInProcessThinBackend(config->thinLTOJobs);
+  } else {
+    backend = lto::createInProcessThinBackend(
+        llvm::heavyweight_hardware_concurrency(config->thinLTOJobs));
   }
 
   ltoObj = std::make_unique<lto::LTO>(createConfig(), backend,
                                        config->ltoPartitions);
 
   // Initialize usedStartStop.
   for (Symbol *sym : symtab->symbols()) {
     StringRef s = sym->getName();

lld/test/COFF/thinlto.ll

 ; REQUIRES: x86
 ; RUN: rm -fr %T/thinlto
 ; RUN: mkdir %T/thinlto
 ; RUN: opt -thinlto-bc -o %T/thinlto/main.obj %s
 ; RUN: opt -thinlto-bc -o %T/thinlto/foo.obj %S/Inputs/lto-dep.ll
 ; RUN: lld-link /lldsavetemps /out:%T/thinlto/main.exe /entry:main /subsystem:console %T/thinlto/main.obj %T/thinlto/foo.obj
 ; RUN: llvm-nm %T/thinlto/main.exe1.lto.obj | FileCheck %s
 
+; Test various possible options for /opt:lldltojobs
+; RUN: lld-link /lldsavetemps /out:%T/thinlto/main.exe /entry:main /subsystem:console %T/thinlto/main.obj %T/thinlto/foo.obj /opt:lldltojobs=1
+; RUN: llvm-nm %T/thinlto/main.exe1.lto.obj | FileCheck %s
+; RUN: lld-link /lldsavetemps /out:%T/thinlto/main.exe /entry:main /subsystem:console %T/thinlto/main.obj %T/thinlto/foo.obj /opt:lldltojobs=all
+; RUN: llvm-nm %T/thinlto/main.exe1.lto.obj | FileCheck %s
+; RUN: lld-link /lldsavetemps /out:%T/thinlto/main.exe /entry:main /subsystem:console %T/thinlto/main.obj %T/thinlto/foo.obj /opt:lldltojobs=1000
+; RUN: llvm-nm %T/thinlto/main.exe1.lto.obj | FileCheck %s
+; RUN: not lld-link /lldsavetemps /out:%T/thinlto/main.exe /entry:main /subsystem:console %T/thinlto/main.obj %T/thinlto/foo.obj /opt:lldltojobs=foo 2>&1 | FileCheck %s --check-prefix=BAD-JOBS
+; BAD-JOBS: error: /opt:lldltojobs: invalid job count: foo
+
 ; This command will store full path to foo.obj in the archive %t.lib
 ; Check that /lldsavetemps is still usable in such case.
 ; RUN: lld-link /lib %T/thinlto/foo.obj /out:%t.lib
 ; RUN: lld-link /lldsavetemps /out:%t.exe /entry:main /subsystem:console %T/thinlto/main.obj %t.lib
 
 ; CHECK-NOT: U foo
 
 target datalayout = "e-m:w-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-pc-windows-msvc"
 
 define i32 @main() {
   call void @foo()
   ret i32 0
 }
 
 declare void @foo()

lld/test/ELF/basic.s

 
 # RUN: not ld.lld %t -o /dev/null -m wrong_emul_fbsd 2>&1 | FileCheck --check-prefix=UNKNOWN_EMUL %s
 # UNKNOWN_EMUL: unknown emulation: wrong_emul_fbsd
 
 # RUN: not ld.lld %t --lto-partitions=0 2>&1 | FileCheck --check-prefix=NOTHREADS %s
 # RUN: not ld.lld %t --plugin-opt=lto-partitions=0 2>&1 | FileCheck --check-prefix=NOTHREADS %s
 # NOTHREADS: --lto-partitions: number of threads must be > 0
 
-# RUN: not ld.lld %t --thinlto-jobs=0 2>&1 | FileCheck --check-prefix=NOTHREADSTHIN %s
-# RUN: not ld.lld %t --plugin-opt=jobs=0 2>&1 | FileCheck --check-prefix=NOTHREADSTHIN %s
-# NOTHREADSTHIN: --thinlto-jobs: number of threads must be > 0
+# RUN: ld.lld %t --thinlto-jobs=0 -verbose 2>&1 | FileCheck --check-prefix=THREADSTHIN %s
+# RUN: ld.lld %t --thinlto-jobs=1 -verbose 2>&1 | FileCheck --check-prefix=THREADSTHIN %s
+# RUN: ld.lld %t --thinlto-jobs=2 -verbose 2>&1 | FileCheck --check-prefix=THREADSTHIN %s
+# RUN: ld.lld %t --thinlto-jobs=all -verbose 2>&1 | FileCheck --check-prefix=THREADSTHIN %s
+# RUN: ld.lld %t --thinlto-jobs=1000 -verbose 2>&1 | FileCheck --check-prefix=THREADSTHIN %s
+# THREADSTHIN: basic.s.tmp
+# RUN: not ld.lld %t --thinlto-jobs=foo -verbose 2>&1 | FileCheck --check-prefix=BADTHREADSTHIN %s
+# BADTHREADSTHIN: error: --thinlto-jobs: invalid job count: foo
+
+# RUN: ld.lld %t --plugin-opt=jobs=0 -verbose 2>&1 | FileCheck --check-prefix=THREADSTHIN %s
+# RUN: ld.lld %t --plugin-opt=jobs=1 -verbose 2>&1 | FileCheck --check-prefix=THREADSTHIN %s
+# RUN: ld.lld %t --plugin-opt=jobs=2 -verbose 2>&1 | FileCheck --check-prefix=THREADSTHIN %s
+# RUN: ld.lld %t --plugin-opt=jobs=all -verbose 2>&1 | FileCheck --check-prefix=THREADSTHIN %s
+# RUN: ld.lld %t --plugin-opt=jobs=1000 -verbose 2>&1 | FileCheck --check-prefix=THREADSTHIN %s
+# RUN: not ld.lld %t --plugin-opt=jobs=foo -verbose 2>&1 | FileCheck --check-prefix=BADTHREADSTHIN %s
 
 # RUN: not ld.lld %t -z ifunc-noplt -z text 2>&1 | FileCheck --check-prefix=NOIFUNCPLTNOTEXTREL %s
 # NOIFUNCPLTNOTEXTREL: -z text and -z ifunc-noplt may not be used together

lld/test/ELF/lto/thinlto.ll

 ; RUN: llvm-nm %t32.lto.o | FileCheck %s --check-prefix=NM2
 
 ; Next force multi-threaded mode
 ; RUN: rm -f %t31.lto.o %t32.lto.o
 ; RUN: ld.lld -save-temps --thinlto-jobs=2 -shared %t1.o %t2.o -o %t3
 ; RUN: llvm-nm %t31.lto.o | FileCheck %s --check-prefix=NM1
 ; RUN: llvm-nm %t32.lto.o | FileCheck %s --check-prefix=NM2
 
-; Then check without --thinlto-jobs (which currently default to hardware_concurrency)
-; RUN: ld.lld -shared %t1.o %t2.o -o %t3
+; Test with all threads, on all cores, on all CPU sockets
+; RUN: rm -f %t31.lto.o %t32.lto.o
+; RUN: ld.lld -save-temps --thinlto-jobs=all -shared %t1.o %t2.o -o %t3
+; RUN: llvm-nm %t31.lto.o | FileCheck %s --check-prefix=NM1
+; RUN: llvm-nm %t32.lto.o | FileCheck %s --check-prefix=NM2
+
+; Test with many more threads than the system has
+; RUN: rm -f %t31.lto.o %t32.lto.o
+; RUN: ld.lld -save-temps --thinlto-jobs=1000 -shared %t1.o %t2.o -o %t3
+; RUN: llvm-nm %t31.lto.o | FileCheck %s --check-prefix=NM1
+; RUN: llvm-nm %t32.lto.o | FileCheck %s --check-prefix=NM2
+
+; Test with a bad value
+; RUN: rm -f %t31.lto.o %t32.lto.o
+; RUN: not ld.lld -save-temps --thinlto-jobs=foo -shared %t1.o %t2.o -o %t3 2>&1 | FileCheck %s --check-prefix=BAD-JOBS
+; BAD-JOBS: error: --thinlto-jobs: invalid job count: foo
+
+; Then check without --thinlto-jobs (which currently defaults to heavyweight_hardware_concurrency, meanning one thread per hardware core -- not SMT)
+; RUN: ld.lld -shared -save-temps %t1.o %t2.o -o %t3
 ; RUN: llvm-nm %t31.lto.o | FileCheck %s --check-prefix=NM1
 ; RUN: llvm-nm %t32.lto.o | FileCheck %s --check-prefix=NM2
 
 ; Check that -save-temps is usable with thin archives
 ; RUN: rm -fr %t.dir
 ; RUN: mkdir -p %t.dir
 ; RUN: cp %t2.o %t.dir/t.o
 ; RUN: llvm-ar rcsT %t.dir/t.a %t.dir/t.o

lld/test/wasm/lto/thinlto.ll

 ; Basic ThinLTO tests.
 ; RUN: opt -module-summary %s -o %t1.o
 ; RUN: opt -module-summary %p/Inputs/thinlto.ll -o %t2.o
 
 ; First force single-threaded mode
 ; RUN: rm -f %t31.lto.o %t32.lto.o
 ; RUN: wasm-ld -r -save-temps --thinlto-jobs=1 %t1.o %t2.o -o %t3
 ; RUN: llvm-nm %t31.lto.o | FileCheck %s --check-prefix=NM1
 ; RUN: llvm-nm %t32.lto.o | FileCheck %s --check-prefix=NM2
 
 ; Next force multi-threaded mode
 ; RUN: rm -f %t31.lto.o %t32.lto.o
 ; RUN: wasm-ld -r -save-temps --thinlto-jobs=2 %t1.o %t2.o -o %t3
 ; RUN: llvm-nm %t31.lto.o | FileCheck %s --check-prefix=NM1
 ; RUN: llvm-nm %t32.lto.o | FileCheck %s --check-prefix=NM2
 
-; Check without --thinlto-jobs (which currently default to hardware_concurrency)
-; RUN: wasm-ld -r %t1.o %t2.o -o %t3
+; Test with all threads, on all cores, on all CPU sockets
+; RUN: rm -f %t31.lto.o %t32.lto.o
+; RUN: wasm-ld -r -save-temps --thinlto-jobs=all %t1.o %t2.o -o %t3
+; RUN: llvm-nm %t31.lto.o | FileCheck %s --check-prefix=NM1
+; RUN: llvm-nm %t32.lto.o | FileCheck %s --check-prefix=NM2
+
+; Test with many more threads than the system has
+; RUN: rm -f %t31.lto.o %t32.lto.o
+; RUN: wasm-ld -r -save-temps --thinlto-jobs=1000 %t1.o %t2.o -o %t3
+; RUN: llvm-nm %t31.lto.o | FileCheck %s --check-prefix=NM1
+; RUN: llvm-nm %t32.lto.o | FileCheck %s --check-prefix=NM2
+
+; Test with a bad value
+; RUN: rm -f %t31.lto.o %t32.lto.o
+; RUN: not wasm-ld -r -save-temps --thinlto-jobs=foo %t1.o %t2.o -o %t3 2>&1 | FileCheck %s --check-prefix=BAD-JOBS
+; BAD-JOBS: error: --thinlto-jobs: invalid job count: foo
+
+; Check without --thinlto-jobs (which currently defaults to heavyweight_hardware_concurrency, meanning one thread per hardware core -- not SMT)
+; RUN: rm -f %t31.lto.o %t32.lto.o
+; RUN: wasm-ld -r -save-temps %t1.o %t2.o -o %t3
 ; RUN: llvm-nm %t31.lto.o | FileCheck %s --check-prefix=NM1
 ; RUN: llvm-nm %t32.lto.o | FileCheck %s --check-prefix=NM2
 
 ; NM1: T f
 ; NM2: T g
 
 target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
 target triple = "wasm32-unknown-unknown"
 
 declare void @g(...)
 
 define void @f() {
 entry:
   call void (...) @g()
   ret void
 }

lld/wasm/Config.h

struct Configuration {
   bool trace;
   uint32_t globalBase;
   uint32_t initialMemory;
   uint32_t maxMemory;
   uint32_t zStackSize;
   unsigned ltoPartitions;
   unsigned ltoo;
   unsigned optimize;
-  unsigned thinLTOJobs;
+  llvm::StringRef thinLTOJobs;
 
   llvm::StringRef entry;
   llvm::StringRef outputFile;
   llvm::StringRef thinLTOCacheDir;
 
   llvm::StringSet<> allowUndefinedSymbols;
   llvm::StringSet<> exportedSymbols;
   std::vector<llvm::StringRef> searchPaths;

lld/wasm/Driver.cpp

static void readConfigs(opt::InputArgList &args) {
   config->stripAll = args.hasArg(OPT_strip_all);
   config->stripDebug = args.hasArg(OPT_strip_debug);
   config->stackFirst = args.hasArg(OPT_stack_first);
   config->trace = args.hasArg(OPT_trace);
   config->thinLTOCacheDir = args.getLastArgValue(OPT_thinlto_cache_dir);
   config->thinLTOCachePolicy = CHECK(
       parseCachePruningPolicy(args.getLastArgValue(OPT_thinlto_cache_policy)),
       "--thinlto-cache-policy: invalid cache policy");
-  config->thinLTOJobs = args::getInteger(args, OPT_thinlto_jobs, -1u);
+  config->thinLTOJobs = args.getLastArgValue(OPT_thinlto_jobs);
   errorHandler().verbose = args.hasArg(OPT_verbose);
   LLVM_DEBUG(errorHandler().verbose = true);
   threadsEnabled = args.hasFlag(OPT_threads, OPT_no_threads, true);
 
   config->initialMemory = args::getInteger(args, OPT_initial_memory, 0);
   config->globalBase = args::getInteger(args, OPT_global_base, 1024);
   config->maxMemory = args::getInteger(args, OPT_max_memory, 0);
   config->zStackSize =
static void checkOptions(opt::InputArgList &args) {
   if (!config->stripDebug && !config->stripAll && config->compressRelocations)
     error("--compress-relocations is incompatible with output debug"
           " information. Please pass --strip-debug or --strip-all");
 
   if (config->ltoo > 3)
     error("invalid optimization level for LTO: " + Twine(config->ltoo));
   if (config->ltoPartitions == 0)
     error("--lto-partitions: number of threads must be > 0");
-  if (config->thinLTOJobs == 0)
-    error("--thinlto-jobs: number of threads must be > 0");
+  if (!get_threadpool_strategy(config->thinLTOJobs))
+    error("--thinlto-jobs: invalid job count: " + config->thinLTOJobs);
 
   if (config->pie && config->shared)
     error("-shared and -pie may not be used together");
 
   if (config->outputFile.empty())
     error("no output file specified");
 
   if (config->importTable && config->exportTable)

lld/wasm/LTO.cpp

static std::unique_ptr<lto::LTO> createLTO() {
   else if (config->isPic)
     c.RelocModel = Reloc::PIC_;
   else
     c.RelocModel = Reloc::Static;
 
   if (config->saveTemps)
     checkError(c.addSaveTemps(config->outputFile.str() + ".",
                               /*UseInputModulePath*/ true));
-
-  lto::ThinBackend backend;
-  if (config->thinLTOJobs != -1U)
-    backend = lto::createInProcessThinBackend(config->thinLTOJobs);
+  lto::ThinBackend backend = lto::createInProcessThinBackend(
+      llvm::heavyweight_hardware_concurrency(config->thinLTOJobs));
   return std::make_unique<lto::LTO>(std::move(c), backend,
                                      config->ltoPartitions);
 }
 
 BitcodeCompiler::BitcodeCompiler() : ltoObj(createLTO()) {}
 
 BitcodeCompiler::~BitcodeCompiler() = default;
 

llvm/include/llvm/LTO/LTO.h

 /// create a ThinBackend using one of the create*ThinBackend() functions below.
 using ThinBackend = std::function<std::unique_ptr<ThinBackendProc>(
     const Config &C, ModuleSummaryIndex &CombinedIndex,
     StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
     AddStreamFn AddStream, NativeObjectCache Cache)>;
 
 /// This ThinBackend runs the individual backend jobs in-process.
 /// The default value means to use one job per hardware core (not hyper-thread).
-ThinBackend createInProcessThinBackend(unsigned ParallelismLevel = 0);
+ThinBackend createInProcessThinBackend(ThreadPoolStrategy Parallelism);
 
 /// This ThinBackend writes individual module indexes to files, instead of
 /// running the individual backend jobs. This backend is for distributed builds
 /// where separate processes will invoke the real backends.
 ///
 /// To find the path to write the index to, the backend checks if the path has a
 /// prefix of OldPrefix; if so, it replaces that prefix with NewPrefix. It then
 /// appends ".thinlto.bc" and writes the index to that path. If

llvm/include/llvm/Support/Threading.h

public:
     /// accounts for affinity masks and takes advantage of all CPU sockets.
     unsigned compute_thread_count() const;
 
     /// Assign the current thread to an ideal hardware CPU or NUMA node. In a
     /// multi-socket system, this ensures threads are assigned to all CPU
     /// sockets. \p ThreadPoolNum represents a number bounded by [0,
     /// compute_thread_count()).
     void apply_thread_strategy(unsigned ThreadPoolNum) const;
+
+    /// Finds the CPU socket where a thread should go. Returns 'None' if the
+    /// thread shall remain on the actual CPU socket.
+    Optional<unsigned> compute_cpu_socket(unsigned ThreadPoolNum) const;
   };
 
+  /// Build a strategy from a number of threads as a string provided in \p Num.
+  /// When Num is above the max number of threads specified by the \p Default
+  /// strategy, we attempt to equally allocate the threads on all CPU sockets.
+  /// "0" or an empty string will return the \p Default strategy.
+  /// "all" for using all hardware threads.
+  Optional<ThreadPoolStrategy>
+  get_threadpool_strategy(StringRef Num, ThreadPoolStrategy Default = {});
+
   /// Returns a thread strategy for tasks requiring significant memory or other
   /// resources. To be used for workloads where hardware_concurrency() proves to
   /// be less efficient. Avoid this strategy if doing lots of I/O. Currently
   /// based on physical cores, if available for the host system, otherwise falls
   /// back to hardware_concurrency(). Returns 1 when LLVM is configured with
   /// LLVM_ENABLE_THREADS = OFF.
   inline ThreadPoolStrategy
   heavyweight_hardware_concurrency(unsigned ThreadCount = 0) {
     ThreadPoolStrategy S;
     S.UseHyperThreads = false;
     S.ThreadsRequested = ThreadCount;
     return S;
   }
 
+  /// Like heavyweight_hardware_concurrency() above, but builds a strategy
+  /// based on the rules described for get_threadpool_strategy().
+  /// If \p Num is invalid, returns a default strategy where one thread per
+  /// hardware core is used.
+  inline ThreadPoolStrategy heavyweight_hardware_concurrency(StringRef Num) {
+    Optional<ThreadPoolStrategy> S =
+        get_threadpool_strategy(Num, heavyweight_hardware_concurrency());
+    if (S)
+      return *S;
+    return heavyweight_hardware_concurrency();
+  }
+
   /// Returns a default thread strategy where all available hardware ressources
   /// are to be used, except for those initially excluded by an affinity mask.
   /// This function takes affinity into consideration. Returns 1 when LLVM is
   /// configured with LLVM_ENABLE_THREADS=OFF.
   inline ThreadPoolStrategy hardware_concurrency(unsigned ThreadCount = 0) {
     ThreadPoolStrategy S;
     S.ThreadsRequested = ThreadCount;
     return S;

llvm/lib/LTO/LTO.cpp

LTO::RegularLTOState::RegularLTOState(unsigned ParallelCodeGenParallelismLevel,
                                       const Config &Conf)
     : ParallelCodeGenParallelismLevel(ParallelCodeGenParallelismLevel),
       Ctx(Conf), CombinedModule(std::make_unique<Module>("ld-temp.o", Ctx)),
       Mover(std::make_unique<IRMover>(*CombinedModule)) {}
 
 LTO::ThinLTOState::ThinLTOState(ThinBackend Backend)
     : Backend(Backend), CombinedIndex(/*HaveGVs*/ false) {
   if (!Backend)
-    this->Backend = createInProcessThinBackend();
+    this->Backend =
+        createInProcessThinBackend(llvm::heavyweight_hardware_concurrency());
 }
 
 LTO::LTO(Config Conf, ThinBackend Backend,
          unsigned ParallelCodeGenParallelismLevel)
     : Conf(std::move(Conf)),
       RegularLTO(ParallelCodeGenParallelismLevel, this->Conf),
       ThinLTO(std::move(Backend)) {}
 
class InProcessThinBackend : public ThinBackendProc {
   std::set<GlobalValue::GUID> CfiFunctionDecls;
 
   Optional<Error> Err;
   std::mutex ErrMu;
 
 public:
   InProcessThinBackend(
       const Config &Conf, ModuleSummaryIndex &CombinedIndex,
-      unsigned ThinLTOParallelismLevel,
+      ThreadPoolStrategy ThinLTOParallelism,
       const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
       AddStreamFn AddStream, NativeObjectCache Cache)
       : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries),
-        BackendThreadPool(
-            heavyweight_hardware_concurrency(ThinLTOParallelismLevel)),
-        AddStream(std::move(AddStream)), Cache(std::move(Cache)) {
+        BackendThreadPool(ThinLTOParallelism), AddStream(std::move(AddStream)),
+        Cache(std::move(Cache)) {
     for (auto &Name : CombinedIndex.cfiFunctionDefs())
       CfiFunctionDefs.insert(
           GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name)));
     for (auto &Name : CombinedIndex.cfiFunctionDecls())
       CfiFunctionDecls.insert(
           GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name)));
   }
 
Error wait() override {
     if (Err)
       return std::move(*Err);
     else
       return Error::success();
   }
 };
 } // end anonymous namespace
 
-ThinBackend lto::createInProcessThinBackend(unsigned ParallelismLevel) {
+ThinBackend lto::createInProcessThinBackend(ThreadPoolStrategy Parallelism) {
   return [=](const Config &Conf, ModuleSummaryIndex &CombinedIndex,
              const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
              AddStreamFn AddStream, NativeObjectCache Cache) {
     return std::make_unique<InProcessThinBackend>(
-        Conf, CombinedIndex, ParallelismLevel, ModuleToDefinedGVSummaries,
-        AddStream, Cache);
+        Conf, CombinedIndex, Parallelism, ModuleToDefinedGVSummaries, AddStream,
+        Cache);
   };
 }
 
 // Given the original \p Path to an output file, replace any path
 // prefix matching \p OldPrefix with \p NewPrefix. Also, create the
 // resulting directory if it does not yet exist.
 std::string lto::getThinLTOOutputFile(const std::string &Path,
                                       const std::string &OldPrefix,

llvm/lib/Support/Threading.cpp

 }
 #endif
 
 #else
 
 int computeHostNumHardwareThreads();
 
 unsigned llvm::ThreadPoolStrategy::compute_thread_count() const {
+  if (ThreadsRequested > 0)
+    return ThreadsRequested;
+
   int MaxThreadCount = UseHyperThreads ? computeHostNumHardwareThreads()
                                        : sys::getHostNumPhysicalCores();
   if (MaxThreadCount <= 0)
     MaxThreadCount = 1;
-
-  // No need to create more threads than there are hardware threads, it would
-  // uselessly induce more context-switching and cache eviction.
-  if (!ThreadsRequested || ThreadsRequested > (unsigned)MaxThreadCount)
-    return MaxThreadCount;
-  return ThreadsRequested;
+}
+
+Optional<ThreadPoolStrategy>
+llvm::get_threadpool_strategy(StringRef Num, ThreadPoolStrategy Default) {
+  if (Num == "all")
+    return llvm::hardware_concurrency();
+  if (Num.empty())
+    return Default;
+  unsigned V;
+  if (Num.getAsInteger(10, V))
+    return None; // malformed 'Num' value
+  if (V == 0)
+    return Default;
+
+  // Do not take the Default into account. This effectively disables
+  // heavyweight_hardware_concurrency() if the user asks for any number of
+  // threads on the cmd-line.
+  ThreadPoolStrategy S = llvm::hardware_concurrency();
+  S.ThreadsRequested = V;
+  return S;
 }
 
 namespace {
 struct SyncThreadInfo {
   void (*UserFn)(void *);
   void *UserData;
 };
 

llvm/lib/Support/Unix/Threading.inc

#endif
   return SetThreadPriorityResult::FAILURE;
 }
 
 #include <thread>
 
 int computeHostNumHardwareThreads() {
 #if defined(HAVE_SCHED_GETAFFINITY) && defined(HAVE_CPU_COUNT)
   cpu_set_t Set;
-  if (sched_getaffinity(0, sizeof(Set), &Set))
+  if (sched_getaffinity(0, sizeof(Set), &Set) == 0)
     return CPU_COUNT(&Set);
 #endif
   // Guard against std::thread::hardware_concurrency() returning 0.
   if (unsigned Val = std::thread::hardware_concurrency())
     return Val;
   return 1;
 }
 

llvm/lib/Support/Windows/Threading.inc

 }
 
 struct ProcessorGroup {
   unsigned ID;
   unsigned AllThreads;
   unsigned UsableThreads;
   unsigned ThreadsPerCore;
   uint64_t Affinity;
+
+  unsigned useableCores() const {
+    return std::max(1U, UsableThreads / ThreadsPerCore);
+  }
 };
 
 template <typename F>
 static bool IterateProcInfo(LOGICAL_PROCESSOR_RELATIONSHIP Relationship, F Fn) {
   DWORD Len = 0;
   BOOL R = ::GetLogicalProcessorInformationEx(Relationship, NULL, &Len);
   if (R || GetLastError() != ERROR_INSUFFICIENT_BUFFER) {
     return false;

 
 int computeHostNumHardwareThreads() {
   static unsigned Threads =
       aggregate(getProcessorGroups(),
                 [](const ProcessorGroup &G) { return G.UsableThreads; });
   return Threads;
 }
 
-// Assign the current thread to a more appropriate CPU socket or CPU group
-void llvm::ThreadPoolStrategy::apply_thread_strategy(
-    unsigned ThreadPoolNum) const {
+// Finds the proper CPU socket where a thread number should go. Returns 'None'
+// if the thread shall remain on the actual CPU socket.
+Optional<unsigned>
+llvm::ThreadPoolStrategy::compute_cpu_socket(unsigned ThreadPoolNum) const {
   ArrayRef<ProcessorGroup> Groups = getProcessorGroups();
+  // Only one CPU socket in the system or process affinity was set, no need to
+  // move the thread(s) to another CPU socket.
+  if (Groups.size() <= 1)
+    return None;
+
+  // We ask for less threads than there are hardware threads per CPU socket, no
+  // need to dispatch threads to other CPU sockets.
+  unsigned MaxThreadsPerSocket =
+      UseHyperThreads ? Groups[0].UsableThreads : Groups[0].useableCores();
+  if (compute_thread_count() <= MaxThreadsPerSocket)
+    return None;
 
   assert(ThreadPoolNum < compute_thread_count() &&
          "The thread index is not within thread strategy's range!");
 
-  // In this mode, the ThreadNumber represents the core number, not the
-  // hyper-thread number. Assumes all NUMA groups have the same amount of
-  // hyper-threads.
-  if (!UseHyperThreads)
-    ThreadPoolNum *= Groups[0].ThreadsPerCore;
-
-  unsigned ThreadRangeStart = 0;
-  for (unsigned I = 0; I < Groups.size(); ++I) {
-    const ProcessorGroup &G = Groups[I];
-    if (ThreadPoolNum >= ThreadRangeStart &&
-        ThreadPoolNum < ThreadRangeStart + G.UsableThreads) {
+  // Assumes the same number of hardware threads per CPU socket.
+  return (ThreadPoolNum * Groups.size()) / compute_thread_count();
+}
 
-      GROUP_AFFINITY Affinity{};
-      Affinity.Group = G.ID;
-      Affinity.Mask = G.Affinity;
-      SetThreadGroupAffinity(GetCurrentThread(), &Affinity, nullptr);
-    }
-    ThreadRangeStart += G.UsableThreads;
-  }
-}
+// Assign the current thread to a more appropriate CPU socket or CPU group
+void llvm::ThreadPoolStrategy::apply_thread_strategy(
+    unsigned ThreadPoolNum) const {
+  Optional<unsigned> Socket = compute_cpu_socket(ThreadPoolNum);
+  if (!Socket)
+    return;
+  ArrayRef<ProcessorGroup> Groups = getProcessorGroups();
+  Affinity.Group = Groups[*Socket].ID;
+  Affinity.Mask = Groups[*Socket].Affinity;
 
 llvm::BitVector llvm::get_thread_affinity_mask() {
   GROUP_AFFINITY Affinity{};
   GetThreadGroupAffinity(GetCurrentThread(), &Affinity);
 
   static unsigned All =
       aggregate(getProcessorGroups(),
                 [](const ProcessorGroup &G) { return G.AllThreads; });

llvm/tools/gold/gold-plugin.cpp

 #include "llvm/Object/Error.h"
 #include "llvm/Support/CachePruning.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/FileSystem.h"
 #include "llvm/Support/ManagedStatic.h"
 #include "llvm/Support/MemoryBuffer.h"
 #include "llvm/Support/Path.h"
 #include "llvm/Support/TargetSelect.h"
+#include "llvm/Support/Threading.h"
 #include "llvm/Support/raw_ostream.h"
 #include <list>
 #include <map>
 #include <plugin-api.h>
 #include <string>
 #include <system_error>
 #include <utility>
 #include <vector>
enum OutputType {
     OT_NORMAL,
     OT_DISABLE,
     OT_BC_ONLY,
     OT_ASM_ONLY,
     OT_SAVE_TEMPS
   };
   static OutputType TheOutputType = OT_NORMAL;
   static unsigned OptLevel = 2;
-  // Default parallelism of 0 used to indicate that user did not specify.
-  // Actual parallelism default value depends on implementation.
   // Currently only affects ThinLTO, where the default is the max cores in the
-  // system.
-  static unsigned Parallelism = 0;
+  // system. See llvm::get_threadpool_strategy() for acceptable values.
+  static std::string Parallelism;
   // Default regular LTO codegen parallelism (number of partitions).
   static unsigned ParallelCodeGenParallelismLevel = 1;
 #ifdef NDEBUG
   static bool DisableVerify = true;
 #else
   static bool DisableVerify = false;
 #endif
   static std::string obj_path;
if (opt.startswith("mcpu=")) {
       cache_dir = std::string(opt.substr(strlen("cache-dir=")));
     } else if (opt.startswith("cache-policy=")) {
       cache_policy = std::string(opt.substr(strlen("cache-policy=")));
     } else if (opt.size() == 2 && opt[0] == 'O') {
       if (opt[1] < '0' || opt[1] > '3')
         message(LDPL_FATAL, "Optimization level must be between 0 and 3");
       OptLevel = opt[1] - '0';
     } else if (opt.startswith("jobs=")) {
-      if (StringRef(opt_ + 5).getAsInteger(10, Parallelism))
-        message(LDPL_FATAL, "Invalid parallelism level: %s", opt_ + 5);
+      StringRef Num(opt_ + 5);
+      if (!get_threadpool_strategy(Num))
+        message(LDPL_FATAL, "Invalid parallelism level: %s", Num.data());
+      Parallelism = Num;
     } else if (opt.startswith("lto-partitions=")) {
       if (opt.substr(strlen("lto-partitions="))
               .getAsInteger(10, ParallelCodeGenParallelismLevel))
         message(LDPL_FATAL, "Invalid codegen partition level: %s", opt_ + 5);
     } else if (opt == "disable-verify") {
       DisableVerify = true;
     } else if (opt.startswith("sample-profile=")) {
       sample_profile = std::string(opt.substr(strlen("sample-profile=")));
static std::unique_ptr<LTO> createLTO(IndexWriteCallback OnIndexWrite,
   Conf.RelocModel = RelocationModel;
   Conf.CodeModel = codegen::getExplicitCodeModel();
   Conf.CGOptLevel = getCGOptLevel();
   Conf.DisableVerify = options::DisableVerify;
   Conf.OptLevel = options::OptLevel;
   Conf.PTO.LoopVectorization = options::OptLevel > 1;
   Conf.PTO.SLPVectorization = options::OptLevel > 1;
 
-  if (options::Parallelism)
-    Backend = createInProcessThinBackend(options::Parallelism);
   if (options::thinlto_index_only) {
     std::string OldPrefix, NewPrefix;
     getThinLTOOldAndNewPrefix(OldPrefix, NewPrefix);
     Backend = createWriteIndexesThinBackend(OldPrefix, NewPrefix,
                                             options::thinlto_emit_imports_files,
                                             LinkedObjectsFile, OnIndexWrite);
+  } else {
+    Backend = createInProcessThinBackend(
+        llvm::heavyweight_hardware_concurrency(options::Parallelism));
   }
 
   Conf.OverrideTriple = options::triple;
   Conf.DefaultTriple = sys::getDefaultTargetTriple();
 
   Conf.DiagHandler = diagnosticHandler;
 
   switch (options::TheOutputType) {

llvm/tools/llvm-lto2/llvm-lto2.cpp

 
 static cl::opt<bool>
     ThinLTODistributedIndexes("thinlto-distributed-indexes", cl::init(false),
                               cl::desc("Write out individual index and "
                                        "import files for the "
                                        "distributed backend case"));
 
 // Default to using all available threads in the system, but using only one
-// thread per core, as indicated by the usage of
-// heavyweight_hardware_concurrency() in the InProcessThinBackend constructor.
-static cl::opt<int> Threads("thinlto-threads", cl::init(0));
+// thread per core (no SMT).
+// Use -thinlto-threads=all to use hardware_concurrency() instead, which means
+// to use all hardware threads or cores in the system.
+static cl::opt<std::string> Threads("thinlto-threads");
 
 static cl::list<std::string> SymbolResolutions(
     "r",
     cl::desc("Specify a symbol resolution: filename,symbolname,resolution\n"
              "where \"resolution\" is a sequence (which may be empty) of the\n"
              "following characters:\n"
              " p - prevailing: the linker has chosen this definition of the\n"
              "     symbol\n"
static int run(int argc, char **argv) {
   ThinBackend Backend;
   if (ThinLTODistributedIndexes)
     Backend = createWriteIndexesThinBackend(/* OldPrefix */ "",
                                             /* NewPrefix */ "",
                                             /* ShouldEmitImportsFiles */ true,
                                             /* LinkedObjectsFile */ nullptr,
                                             /* OnWrite */ {});
   else
-    Backend = createInProcessThinBackend(Threads);
+    Backend = createInProcessThinBackend(
+        llvm::heavyweight_hardware_concurrency(Threads));
   LTO Lto(std::move(Conf), std::move(Backend));
 
   bool HasErrors = false;
   for (std::string F : InputFilenames) {
     std::unique_ptr<MemoryBuffer> MB = check(MemoryBuffer::getFile(F), F);
     std::unique_ptr<InputFile> Input =
         check(InputFile::create(MB->getMemBufferRef()), F);