[LTO] Split the options for ThinLTO jobs and Regular LTO partitions

[ruiu]

ELF/Driver.cpp

486

It is better to include the option name here.

"--lto-partitions: number of thread must be >0"

[pcc]

ELF/Driver.cpp

489

I think this would result in a default parallelism level of 1. Can we avoid passing a ThinBackend if the user did not specify a parallelism level?

[pcc]

LGTM

ELF/Driver.cpp

489

Maybe -1u here and below?

[mehdi_amini]

ELF/Driver.cpp

489

Doesn't it mean that if the user does not specify --thinlto-jobs on the command line, ThinLTO won't work? I think it would even lead to the LTO library crashing if the bitcode has been generated with ThinLTO? Is there a test that shows what happen in this case? (I don't see any, but I didn't look closely)
The default should be something like std::hardware_concurrency.

[pcc]

ELF/Driver.cpp

489

No, if the ThinBackend is not supplied, LTO will create one for you. See: http://llvm-cs.pcc.me.uk/lib/LTO/LTO.cpp#229

I think we will eventually want to replace the thread::hardware_concurrency() default with something else that will lead to better utilisation of the hardware. In Chromium at least we've observed that hardware_concurrency leads to poor utilisation on machines with a large number of cores (https://bugs.chromium.org/p/chromium/issues/detail?id=645295#c20). So I'd prefer to keep any default out of the clients.

You're right, we should have a test that does not pass a --thinlto-jobs flag. @davide can you please add one?

test/ELF/lto/thinlto.ll

2

I think you need to use opt -module-summary here.

[mehdi_amini]

ELF/Driver.cpp

489

No, if the ThinBackend is not supplied, LTO will create one for you. See: http://llvm-cs.pcc.me.uk/lib/LTO/LTO.cpp#229

OK! Great :)

I think we will eventually want to replace the thread::hardware_concurrency() default with something else that will lead to better utilisation of the hardware. In Chromium at least we've observed that hardware_concurrency leads to poor utilisation on machines with a large number of cores (https://bugs.chromium.org/p/chromium/issues/detail?id=645295#c20). So I'd prefer to keep any default out of the clients.

Yes, Apple's clang is using the number of physical cores because of that, but my rational was to limit the memory consumption at a time where debug info where even more costly than now for ThinLTO.
So indeed we plan to add support for this in LLVM.

On the other hand, David and Teresa benchmarks building Chromium on a 16-cores Xeon machine (32 hyper threaded cores) and they reported 2m18s for the ThinLTO plugin with 32 threads vs 3m38s with 32 threads. So your link is puzzling...

ELF/Config.h

struct Configuration {
   BuildIdKind BuildId = BuildIdKind::None;
   ELFKind EKind = ELFNoneKind;
   uint16_t DefaultSymbolVersion = llvm::ELF::VER_NDX_GLOBAL;
   uint16_t EMachine = llvm::ELF::EM_NONE;
   uint64_t EntryAddr = 0;
   uint64_t ImageBase;
   uint64_t MaxPageSize;
   uint64_t ZStackSize;
-  unsigned LtoJobs;
+  unsigned LtoPartitions;
   unsigned LtoO;
   unsigned Optimize;
+  unsigned ThinLtoJobs;
 };
 
 // The only instance of Configuration struct.
 extern Configuration *Config;
 
 } // namespace elf
 } // namespace lld
 
 #endif

ELF/Driver.cpp

void LinkerDriver::readConfigs(opt::InputArgList &Args) {
   Config->OutputFile = getString(Args, OPT_o);
   Config->SoName = getString(Args, OPT_soname);
   Config->Sysroot = getString(Args, OPT_sysroot);
 
   Config->Optimize = getInteger(Args, OPT_O, 1);
   Config->LtoO = getInteger(Args, OPT_lto_O, 2);
   if (Config->LtoO > 3)
     error("invalid optimization level for LTO: " + getString(Args, OPT_lto_O));
-  Config->LtoJobs = getInteger(Args, OPT_lto_jobs, 1);
-  if (Config->LtoJobs == 0)
+  Config->LtoPartitions = getInteger(Args, OPT_lto_partitions, 1);
+  if (Config->LtoPartitions == 0)
     error("number of threads must be > 0");
+  Config->ThinLtoJobs = getInteger(Args, OPT_thinlto_jobs, 1);
 
   Config->ZCombreloc = !hasZOption(Args, "nocombreloc");
   Config->ZExecStack = hasZOption(Args, "execstack");
   Config->ZNodelete = hasZOption(Args, "nodelete");
   Config->ZNow = hasZOption(Args, "now");
   Config->ZOrigin = hasZOption(Args, "origin");
   Config->ZRelro = !hasZOption(Args, "norelro");
 
   if (!Config->Relocatable)
     Config->Strip = getStripOption(Args);

ELF/LTO.cpp

static std::unique_ptr<lto::LTO> createLTO() {
   Conf.OptPipeline = Config->LtoNewPmPasses;
   Conf.AAPipeline = Config->LtoAAPipeline;
 
   if (Config->SaveTemps)
     checkError(Conf.addSaveTemps(std::string(Config->OutputFile) + ".",
                             /*UseInputModulePath*/ true));
 
   lto::ThinBackend Backend;
-  if (Config->LtoJobs)
-    Backend = lto::createInProcessThinBackend(Config->LtoJobs);
-  return llvm::make_unique<lto::LTO>(std::move(Conf), Backend, Config->LtoJobs);
+  if (Config->ThinLtoJobs)
+    Backend = lto::createInProcessThinBackend(Config->ThinLtoJobs);
+  return llvm::make_unique<lto::LTO>(std::move(Conf), Backend,
+                                     Config->LtoPartitions);
 }
 
 BitcodeCompiler::BitcodeCompiler() : LtoObj(createLTO()) {}
 
 BitcodeCompiler::~BitcodeCompiler() {}
 
 static void undefine(Symbol *S) {
   replaceBody<Undefined>(S, S->body()->getName(), STV_DEFAULT, S->body()->Type,

ELF/Options.td

 // Aliases for ignored options
 def alias_define_common_d: Flag<["-"], "d">, Alias<define_common>;
 def alias_define_common_dc: F<"dc">, Alias<define_common>;
 def alias_define_common_dp: F<"dp">, Alias<define_common>;
 def alias_version_script_version_script: J<"version-script=">,
   Alias<version_script>;
 
 // LTO-related options.
-def lto_jobs: J<"lto-jobs=">, HelpText<"Number of threads to run codegen">;
 def lto_aa_pipeline: J<"lto-aa-pipeline=">,
   HelpText<"AA pipeline to run during LTO. Used in conjunction with -lto-newpm-passes">;
 def lto_newpm_passes: J<"lto-newpm-passes=">,
   HelpText<"Passes to run during LTO">;
+def lto_partitions: J<"lto-partitions=">,
+  HelpText<"Number of LTO codegen partitions">;
 def disable_verify: F<"disable-verify">;
 def mllvm: S<"mllvm">;
 def save_temps: F<"save-temps">;
+def thinlto_jobs: J<"thinlto-jobs=">, HelpText<"Number of ThinLTO jobs">;

test/ELF/basic.s

 
 # RUN: llvm-mc -filetype=obj -triple=x86_64-unknown-linux %s -o %t
 # RUN: not ld.lld %t %t -o %t2 2>&1 | FileCheck --check-prefix=DUP %s
 # DUP: duplicate symbol: _start in {{.*}} and {{.*}}
 
 # RUN: not ld.lld %t -o %t -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-jobs=0 2>&1 | FileCheck --check-prefix=NOTHREADS %s
+# RUN: not ld.lld %t --lto-partitions=0 2>&1 | FileCheck --check-prefix=NOTHREADS %s
 # NOTHREADS: number of threads must be > 0

test/ELF/lto/parallel-internalize.ll

 ; REQUIRES: x86
 ; RUN: llvm-as -o %t.bc %s
-; RUN: ld.lld -m elf_x86_64 --lto-jobs=2 -save-temps -o %t %t.bc -e foo --lto-O0
+; RUN: ld.lld -m elf_x86_64 --lto-partitions=2 -save-temps -o %t %t.bc \
+; RUN:   -e foo --lto-O0
 ; RUN: llvm-readobj -t -dyn-symbols %t | FileCheck %s
 ; RUN: llvm-nm %t0.lto.o | FileCheck --check-prefix=CHECK0 %s
 ; RUN: llvm-nm %t1.lto.o | FileCheck --check-prefix=CHECK1 %s
 
 ; CHECK:      Symbols [
 ; CHECK-NEXT:   Symbol {
 ; CHECK-NEXT:     Name:  (0)
 ; CHECK-NEXT:     Value: 0x0

test/ELF/lto/parallel.ll

 ; REQUIRES: x86
 ; RUN: llvm-as -o %t.bc %s
-; RUN: ld.lld -m elf_x86_64 --lto-jobs=2 -save-temps -o %t %t.bc -shared
+; RUN: ld.lld -m elf_x86_64 --lto-partitions=2 -save-temps -o %t %t.bc -shared
 ; RUN: llvm-nm %t0.lto.o | FileCheck --check-prefix=CHECK0 %s
 ; RUN: llvm-nm %t1.lto.o | FileCheck --check-prefix=CHECK1 %s
 
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 ; CHECK0-NOT: bar
 ; CHECK0: T foo

test/ELF/lto/thinlto.ll

 ; Basic ThinLTO tests.
 ; RUN: llvm-as %s -o %t.o
 ; RUN: llvm-as %p/Inputs/thinlto.ll -o %t2.o
 
 ; First force single-threaded mode
-; RUN: ld.lld -save-temps --lto-jobs=1 -shared %t.o %t2.o -o %t
+; RUN: ld.lld -save-temps --thinlto-jobs=1 -shared %t.o %t2.o -o %t
 ; RUN: llvm-nm %t.lto.o | FileCheck %s --check-prefix=NM
 
 ; NM: T f
 ; NM: T g
 
 ; Next force multi-threaded mode
-; RUN: ld.lld -save-temps --lto-jobs=2 -shared %t.o %t2.o -o %t2
+; RUN: ld.lld -save-temps --thinlto-jobs=2 -shared %t.o %t2.o -o %t2
 ; RUN: llvm-nm %t20.lto.o | FileCheck %s --check-prefix=NM1
 ; RUN: llvm-nm %t21.lto.o | FileCheck %s --check-prefix=NM2
 
 ; NM1: T g
 ; NM2: T f
 
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 declare void @g(...)
 
 define void @f() {
 entry:
   call void (...) @g()
   ret void
 }