[ThinLTO] pass UnrollLoops/VectorizeLoop/VectorizeSLP in CGOpts down to pass builder in ltobackend

[wmi]

Summary

Currently UnrollLoops/VectorizeLoop/VectorizeSLP in CGOpts are not passed down to pass builder in ltobackend when new pass manager is used. This is inconsistent with the behavior when new pass manager is used and thinlto is not used. Such inconsistency causes slp vectorization pass not being enabled in ltobackend for O3 + thinlto right now. This patch fixes that.

[wmi]

update flag in the test.

[tejohnson]

I think the old PM has some issues as well, although more in the opposite direction. The code in LTOBackend.cpp:runOldPMPasses is simply hardcoding loop and slp vectorization flags to true, but presumably should use whatever is passed through the lto Config as well. Not sure about LoopsInterleaved, ah looks like it is set to the value of the internal option which is enabled by default. That and DisableUnrollLoops should also use whatever is passed through here (similar to how they are set in clang BackendUtil.cpp when configuring the old PM).

clang/lib/CodeGen/BackendUtil.cpp

1443

Looks like the PTO constructor sets this to true, so we were getting lucky here.

1446

Looks like this only affects the values passed to LoopVectorizePass. Is there a way to test that it is being enabled properly now?

1447

Ditto.

1448

Is this enabled by default at -O3? I looked at clang but couldn't tell where that would be happening. If so, then llvm/test/Other/new-pm-thinlto-defaults.ll should be fixed to check that it is added to the postlink pipeline at -O3 to make sure we don't regress here again.

[tejohnson]

Thinking through this some more - this patch fixes the distributed ThinLTO backend case, but not the in-process ThinLTO case. For that, both lld and gold plugin will need to set this up. And since they don't have the clang options, they will just need to set to some reasonable default, probably either hardwired on like in the old PM or set based on the CGOptLevel they set from linker options. However, they both default to -O2 unless a special plugin or lld linker option is set, so we'd probably want to enable these passes there at O2+.

[wmi]

In D72386#1810205, @tejohnson wrote:

Thinking through this some more - this patch fixes the distributed ThinLTO backend case, but not the in-process ThinLTO case. For that, both lld and gold plugin will need to set this up. And since they don't have the clang options, they will just need to set to some reasonable default, probably either hardwired on like in the old PM or set based on the CGOptLevel they set from linker options. However, they both default to -O2 unless a special plugin or lld linker option is set, so we'd probably want to enable these passes there at O2+.

Yes, opt using new pass manager -- NewPMDriver.cpp needs to be fixed too, similar as opt using old PM which is set based on Opt Level.

[wmi]

Teresa, do we have existing tests where I can add some check to see if some passes are executed for the lld and gold plugin changes?

[tejohnson]

In D72386#1810284, @wmi wrote:

Teresa, do we have existing tests where I can add some check to see if some passes are executed for the lld and gold plugin changes?

I don't see any. There is a test for each that -debug-pass-manager and new-pass-manager are passed through correctly:
test/tools/gold/X86/new-pm.ll
lld/test/ELF/lto/new-pass-manager.ll

Perhaps those can be extended to actually check the passes that get enabled, similar to llvm/test/Other/new-pm-thinlto-defaults.ll?

[wmi]

Enable loop and slp vectorization in O2/O3 (Os for opt as well) in gold plugin, lld, opt and llvm-lto2. Add related tests.

There are still some issues like we cannot override the default setting in O2/O3 using internal flags like -vectorize-loops, -vectorize-slp, and the issues like those setting are now hard coded in old pass manager which Teresa mentioned. I leave them as it is for the moment and let this patch just handle the major issue of slp being disabled in thinlto backend.

[wmi]

clang/lib/CodeGen/BackendUtil.cpp

1446

For vectorization, given a function with vectorizable loop, I test it by specifying -force-vector-width=2 and -force-vector-interleave=1 and checking whether the IR will contain llvm.loop.isvectorized metadata at the end of the pipeline.

Similarly for interleaving, I test it by specifying -force-vector-width=1 and -force-vector-interleave=2 and checking whether the IR will contain llvm.loop.isvectorized metadata at the end of the pipeline.

1448

Yes, -vectorize-slp will be added as cc1 option when O2/O3 is specified for clang.

I fixed llvm/test/Other/new-pm-thinlto-defaults.ll, but llvm/test/Other/new-pm-thinlto-defaults.ll depends on the setting in opt instead of clang, so I also enabled slp and loop vectorization in O2, O3 and Os in opt.

[tejohnson]

Thanks! Looks pretty good, I just have a few comments and questions.

clang/test/CodeGen/thinlto-slp-vectorize-pm.c

3

Can you clarify the difference between -vectorize-slp and -mllvm -vectorize-slp=false? Is this related to you comment about not being able to use the internal options to disable it? Worth a comment in that case. Ditto for the vectorize-loops tests below.

lld/test/ELF/lto/slp-vectorize-pm.ll

11

I think the above testing was cloned from another existing lld test - can it be removed from this one since not related to vectorization?

llvm/include/llvm/LTO/Config.h

130

add doxygen comment

llvm/test/Other/new-pm-defaults.ll

244

Does clang run this at -Os?

llvm/test/tools/gold/X86/slp-vectorize-pm.ll

12

Similar comment to new lld test - I think the above lines are from another test and can be removed from here?

[wmi]

clang/test/CodeGen/thinlto-slp-vectorize-pm.c

3

-vectorize-slp is a cc1 option and will be added automatically when O2/O3/Os/Oz is available. Once -vectorize-slp is enabled, "-mllvm -vectorize-slp=false" won't disable it currently. I added "-mllvm -vectorize-slp=false" here in the test to ensure the slp vectorization is executed because the -vectorize-slp cc1 flag is passed down, not because -mllvm -vectorize-slp is enabled.

llvm/test/Other/new-pm-defaults.ll

244

Yes it does. But surprisingly I find clang add "-vectorize-slp" at Oz as well but doesn't not add "-vectorize-loops" at Oz. This is another inconsistency with opt old pass manager, which enable neither slp nor vectorize-loops at Oz. I am not sure how it should behave.

[tejohnson]

clang/test/CodeGen/thinlto-slp-vectorize-pm.c

3

Why add "-mllvm -vectorize-slp=false" to the invocation lines though? If it is to illustrate that the internal option cannot override the cc1 option, then add comments to that effect.

llvm/test/Other/new-pm-defaults.ll

244

Hmm, I think probably best to be consistent with the old PM on this for now, and add a comment where you enable these differently. This should presumably be reevaluated, but probably best if the two are in sync for now.

[wmi]

clang/test/CodeGen/thinlto-slp-vectorize-pm.c

3

I add it to the invocation line because if the default value of "-mllvm -vectorize-slp" is true, even without this patch, the test will still pass. That is why I want to add "-mllvm -vectorize-slp=false" explicitly to ensure the patch is effective.

llvm/test/Other/new-pm-defaults.ll

244

Ok, thanks. I will add a comment at the place where it sets this.

[wmi]

Address Teresa's comment.

[tejohnson]

A few minor suggestions for comment tweaks, but lgtm otherwise.

clang/test/CodeGen/thinlto-slp-vectorize-pm.c

12

maybe add "by default" to the end of the last sentence for clarity? Ditto elsewhere.

llvm/lib/Passes/PassBuilder.cpp

1905

Do you mean "consistent with old pass manager invoked via opt"? It currently reads a little confusing, because the old PM is also invoked via clang.

[wmi]

Thanks for the review!

clang/test/CodeGen/thinlto-slp-vectorize-pm.c

12

Done.

llvm/lib/Passes/PassBuilder.cpp

1905

Fixed.

[xbolva00]

lld/ELF/LTO.cpp

96

COFF?

[tejohnson]

lld/ELF/LTO.cpp

96

This came up recently on discourse: https://discourse.llvm.org/t/unintended-lto-configuration-differences-between-elf-and-coff/62636

This patch is a couple years old now and @wmi isn't working actively on LLVM anymore, maybe someone working on COFF who can test it there can send a patch to make this consistent?

clang/lib/CodeGen/BackendUtil.cpp

static void runThinLTOBackend(ModuleSummaryIndex *CombinedIndex, Module *M,
   Conf.CPU = TOpts.CPU;
   Conf.CodeModel = getCodeModel(CGOpts);
   Conf.MAttrs = TOpts.Features;
   Conf.RelocModel = CGOpts.RelocationModel;
   Conf.CGOptLevel = getCGOptLevel(CGOpts);
   Conf.OptLevel = CGOpts.OptimizationLevel;
   initTargetOptions(Conf.Options, CGOpts, TOpts, LOpts, HeaderOpts);
   Conf.SampleProfile = std::move(SampleProfile);
+  Conf.PTO.LoopUnrolling = CGOpts.UnrollLoops;
+  // For historical reasons, loop interleaving is set to mirror setting for loop
+  // unrolling.
+  Conf.PTO.LoopInterleaving = CGOpts.UnrollLoops;
+  Conf.PTO.LoopVectorization = CGOpts.VectorizeLoop;
+  Conf.PTO.SLPVectorization = CGOpts.VectorizeSLP;
 
   // Context sensitive profile.
   if (CGOpts.hasProfileCSIRInstr()) {
     Conf.RunCSIRInstr = true;
     Conf.CSIRProfile = std::move(CGOpts.InstrProfileOutput);
   } else if (CGOpts.hasProfileCSIRUse()) {
     Conf.RunCSIRInstr = false;
     Conf.CSIRProfile = std::move(CGOpts.ProfileInstrumentUsePath);

clang/test/CodeGen/thinlto-slp-vectorize-pm.c

+// REQUIRES: x86-registered-target
+// Test to ensure the SLP flag is passed down to the ThinLTO backend.
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -emit-obj -O2 -vectorize-slp -mllvm -vectorize-slp=false -o %t2.o -x ir %t.o -fthinlto-index=%t.thinlto.bc -fdebug-pass-manager -fexperimental-new-pass-manager 2>&1 | FileCheck %s --check-prefix=O2-SLP
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -emit-obj -O0 -vectorize-slp -mllvm -vectorize-slp=false -o %t2.o -x ir %t.o -fthinlto-index=%t.thinlto.bc -fdebug-pass-manager -fexperimental-new-pass-manager 2>&1 | FileCheck %s --check-prefix=O0-SLP
+// O2-SLP: Running pass: SLPVectorizerPass
+// O0-SLP-NOT: Running pass: SLPVectorizerPass
+
+// Test to ensure the loop vectorize codegen option is passed down to the ThinLTO backend. The loop vectorize codegen option is enabled if -vectorize-loops is in cc1 options.
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -emit-obj -O2 -vectorize-loops -mllvm -vectorize-loops=false -mllvm -force-vector-width=2 -mllvm -force-vector-interleave=1 -emit-llvm -o - -x ir %t.o -fthinlto-index=%t.thinlto.bc -fexperimental-new-pass-manager 2>&1 | FileCheck %s --check-prefix=O2-LPV
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -emit-obj -O0 -vectorize-loops -mllvm -vectorize-loops=false -mllvm -force-vector-width=2 -mllvm -force-vector-interleave=1 -emit-llvm -o - -x ir %t.o -fthinlto-index=%t.thinlto.bc -fexperimental-new-pass-manager 2>&1 | FileCheck %s --check-prefix=O0-LPV
+// O2-LPV: = !{!"llvm.loop.isvectorized", i32 1}
+// O0-LPV-NOT: = !{!"llvm.loop.isvectorized", i32 1}
+
+// Test to ensure the loop Interleave codegen opt is passed down to the ThinLTO backend. The loop interleave codegen option is enabled if O2 or above are in cc1 options.
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -emit-obj -O2 -vectorize-loops -mllvm -interleave-loops=false -mllvm -force-vector-width=1 -mllvm -force-vector-interleave=2 -emit-llvm -o - -x ir %t.o -fthinlto-index=%t.thinlto.bc -fexperimental-new-pass-manager 2>&1 | FileCheck %s --check-prefix=O2-InterLeave
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -emit-obj -O0 -vectorize-loops -mllvm -interleave-loops=false -mllvm -force-vector-width=1 -mllvm -force-vector-interleave=2 -emit-llvm -o - -x ir %t.o -fthinlto-index=%t.thinlto.bc -fexperimental-new-pass-manager 2>&1 | FileCheck %s --check-prefix=O0-InterLeave
+// O2-InterLeave: = !{!"llvm.loop.isvectorized", i32 1}
+// O0-InterLeave-NOT: = !{!"llvm.loop.isvectorized", i32 1}
+
+void foo(double *a) {
+  for (int i = 0; i < 1000; i++)
+    a[i] = 10;
+}

lld/COFF/CMakeLists.txt

add_lld_library(lldCOFF
   DebugInfoMSF
   DebugInfoPDB
   Demangle
   LibDriver
   LTO
   MC
   Object
   Option
+  Passes
   Support
   WindowsManifest
 
   LINK_LIBS
   lldCommon
   ${LLVM_PTHREAD_LIB}
 
   DEPENDS
   COFFOptionsTableGen
   ${tablegen_deps}
   )

lld/ELF/CMakeLists.txt

add_lld_library(lldELF
   BitWriter
   Core
   DebugInfoDWARF
   Demangle
   LTO
   MC
   Object
   Option
+  Passes
   Support
 
   LINK_LIBS
   lldCommon
   ${LLVM_PTHREAD_LIB}
 
   DEPENDS
   ELFOptionsTableGen
   ${tablegen_deps}
   )

lld/ELF/LTO.cpp

static lto::Config createConfig() {
   c.CodeModel = getCodeModelFromCMModel();
   c.DisableVerify = config->disableVerify;
   c.DiagHandler = diagnosticHandler;
   c.OptLevel = config->ltoo;
   c.CPU = getCPUStr();
   c.MAttrs = getMAttrs();
   c.CGOptLevel = args::getCGOptLevel(config->ltoo);
 
+  c.PTO.LoopVectorization = c.OptLevel > 1;
+  c.PTO.SLPVectorization = c.OptLevel > 1;
+
   // Set up a custom pipeline if we've been asked to.
   c.OptPipeline = config->ltoNewPmPasses;
   c.AAPipeline = config->ltoAAPipeline;
 
   // Set up optimization remarks if we've been asked to.
   c.RemarksFilename = config->optRemarksFilename;
   c.RemarksPasses = config->optRemarksPasses;
   c.RemarksWithHotness = config->optRemarksWithHotness;

lld/test/ELF/lto/slp-vectorize-pm.ll

+; REQUIRES: x86
+; RUN: opt -module-summary %s -o %t.o
+
+; Test new-pass-manager and debug-pass-manager option
+; RUN: ld.lld --plugin-opt=new-pass-manager --plugin-opt=debug-pass-manager -o %t2.o %t.o 2>&1 | FileCheck %s
+; RUN: ld.lld --plugin-opt=new-pass-manager --lto-debug-pass-manager -o %t2.o %t.o 2>&1 | FileCheck %s
+; RUN: ld.lld --lto-new-pass-manager --plugin-opt=debug-pass-manager -o %t2.o %t.o 2>&1 | FileCheck %s
+; RUN: ld.lld --lto-new-pass-manager --lto-debug-pass-manager -o %t2.o %t.o 2>&1 | FileCheck %s
+
+; CHECK: Starting llvm::Module pass manager run
+; CHECK: Finished llvm::Module pass manager run
+
+; Test SLP and Loop Vectorization are enabled by default at O2 and O3.
+; RUN: ld.lld --plugin-opt=new-pass-manager --plugin-opt=debug-pass-manager --plugin-opt=O0 --plugin-opt=save-temps -shared -o %t1.o %t.o 2>&1 | FileCheck %s --check-prefix=CHECK-O0-SLP
+; RUN: llvm-dis %t.o.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-O0-LPV
+
+; RUN: ld.lld --plugin-opt=new-pass-manager --plugin-opt=debug-pass-manager --plugin-opt=O1 --plugin-opt=save-temps -shared -o %t2.o %t.o 2>&1 | FileCheck %s --check-prefix=CHECK-O1-SLP
+; RUN: llvm-dis %t.o.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-O1-LPV
+
+; RUN: ld.lld --plugin-opt=new-pass-manager --plugin-opt=debug-pass-manager --plugin-opt=O2 --plugin-opt=save-temps -shared -o %t3.o %t.o 2>&1 | FileCheck %s --check-prefix=CHECK-O2-SLP
+; RUN: llvm-dis %t.o.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-O2-LPV
+
+; RUN: ld.lld --plugin-opt=new-pass-manager --plugin-opt=debug-pass-manager --plugin-opt=O3 --plugin-opt=save-temps -shared -o %t4.o %t.o 2>&1 | FileCheck %s --check-prefix=CHECK-O3-SLP
+; RUN: llvm-dis %t.o.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-O3-LPV
+
+; CHECK-O0-SLP-NOT: Running pass: SLPVectorizerPass
+; CHECK-O1-SLP-NOT: Running pass: SLPVectorizerPass
+; CHECK-O2-SLP: Running pass: SLPVectorizerPass
+; CHECK-O3-SLP: Running pass: SLPVectorizerPass
+; CHECK-O0-LPV-NOT: = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK-O1-LPV-NOT: = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK-O2-LPV: = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK-O3-LPV: = !{!"llvm.loop.isvectorized", i32 1}
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define i32 @foo(i32* %a) {
+entry:
+  br label %for.body
+
+for.body:
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
+  %red.05 = phi i32 [ 0, %entry ], [ %add, %for.body ]
+  %arrayidx = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
+  %0 = load i32, i32* %arrayidx, align 4
+  %add = add nsw i32 %0, %red.05
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond = icmp eq i64 %indvars.iv.next, 255
+  br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:
+  ret i32 %add
+}
+
+!0 = distinct !{!0, !1}
+!1 = !{!"llvm.loop.unroll.disable", i1 true}

lld/wasm/CMakeLists.txt

add_lld_library(lldWasm
   ${LLVM_TARGETS_TO_BUILD}
   BinaryFormat
   Core
   Demangle
   LTO
   MC
   Object
   Option
+  Passes
   Support
 
   LINK_LIBS
   lldCommon
 
   DEPENDS
   WasmOptionsTableGen
   ${tablegen_deps}
   )

llvm/include/llvm/LTO/Config.h

 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_LTO_CONFIG_H
 #define LLVM_LTO_CONFIG_H
 
 #include "llvm/IR/DiagnosticInfo.h"
 #include "llvm/IR/LLVMContext.h"
+#include "llvm/Passes/PassBuilder.h"
 #include "llvm/Support/CodeGen.h"
 #include "llvm/Target/TargetOptions.h"
 
 #include <functional>
 
 namespace llvm {
 
 class Error;
struct Config {
   DiagnosticHandlerFunction DiagHandler;
 
   /// If this field is set, LTO will write input file paths and symbol
   /// resolutions here in llvm-lto2 command line flag format. This can be
   /// used for testing and for running the LTO pipeline outside of the linker
   /// with llvm-lto2.
   std::unique_ptr<raw_ostream> ResolutionFile;
 
+  PipelineTuningOptions PTO;
+
   /// The following callbacks deal with tasks, which normally represent the
   /// entire optimization and code generation pipeline for what will become a
   /// single native object file. Each task has a unique identifier between 0 and
   /// getMaxTasks()-1, which is supplied to the callback via the Task parameter.
   /// A task represents the entire pipeline for ThinLTO and regular
   /// (non-parallel) LTO, but a parallel code generation task will be split into
   /// N tasks before code generation, where N is the parallelism level.
   ///

llvm/lib/LTO/LTOBackend.cpp

static void runNewPMPasses(Config &Conf, Module &Mod, TargetMachine *TM,
   } else if (!Conf.CSIRProfile.empty()) {
     PGOOpt = PGOOptions(Conf.CSIRProfile, "", Conf.ProfileRemapping,
                         PGOOptions::IRUse, PGOOptions::CSIRUse);
   }
 
   PassInstrumentationCallbacks PIC;
   StandardInstrumentations SI;
   SI.registerCallbacks(PIC);
-  PassBuilder PB(TM, PipelineTuningOptions(),PGOOpt, &PIC);
+  PassBuilder PB(TM, Conf.PTO, PGOOpt, &PIC);
   AAManager AA;
 
   // Parse a custom AA pipeline if asked to.
   if (auto Err = PB.parseAAPipeline(AA, "default"))
     report_fatal_error("Error parsing default AA pipeline");
 
   LoopAnalysisManager LAM(Conf.DebugPassManager);
   FunctionAnalysisManager FAM(Conf.DebugPassManager);

llvm/lib/Passes/PassBuilder.cpp

if (L == O0) {
             MPM, DebugLogging,
             /* RunProfileGen */ (PGOOpt->Action == PGOOptions::IRInstr),
             /* IsCS */ false, PGOOpt->ProfileFile,
             PGOOpt->ProfileRemappingFile);
       // Do nothing else at all!
       return Error::success();
     }
 
+    PTO.LoopVectorization = L > O1 && L < Oz;
+    PTO.SLPVectorization = L > O1 && L < Oz;
+
     if (Matches[1] == "default") {
       MPM.addPass(buildPerModuleDefaultPipeline(L, DebugLogging));
     } else if (Matches[1] == "thinlto-pre-link") {
       MPM.addPass(buildThinLTOPreLinkDefaultPipeline(L, DebugLogging));
     } else if (Matches[1] == "thinlto") {
       MPM.addPass(buildThinLTODefaultPipeline(L, DebugLogging, nullptr));
     } else if (Matches[1] == "lto-pre-link") {
       MPM.addPass(buildLTOPreLinkDefaultPipeline(L, DebugLogging));

llvm/test/Other/new-pm-defaults.ll

 ; CHECK-O-NEXT: Running pass: LoopDistributePass
 ; CHECK-O-NEXT: Running pass: LoopVectorizePass
 ; CHECK-O-NEXT: Running analysis: BlockFrequencyAnalysis
 ; CHECK-O-NEXT: Running analysis: BranchProbabilityAnalysis
 ; CHECK-O-NEXT: Running pass: LoopLoadEliminationPass
 ; CHECK-O-NEXT: Running analysis: LoopAccessAnalysis
 ; CHECK-O-NEXT: Running pass: InstCombinePass
 ; CHECK-O-NEXT: Running pass: SimplifyCFGPass
+; CHECK-O2-NEXT: Running pass: SLPVectorizerPass
+; CHECK-O3-NEXT: Running pass: SLPVectorizerPass
+; CHECK-Os-NEXT: Running pass: SLPVectorizerPass
 ; CHECK-O-NEXT: Running pass: InstCombinePass
 ; CHECK-O-NEXT: Running pass: LoopUnrollPass
 ; CHECK-O-NEXT: Running pass: WarnMissedTransformationsPass
 ; CHECK-O-NEXT: Running pass: InstCombinePass
 ; CHECK-O-NEXT: Running pass: RequireAnalysisPass<{{.*}}OptimizationRemarkEmitterAnalysis
 ; CHECK-O-NEXT: Running pass: FunctionToLoopPassAdaptor<{{.*}}LICMPass
 ; CHECK-O-NEXT: Starting llvm::Function pass manager run.
 ; CHECK-O-NEXT: Running pass: LoopSimplifyPass

llvm/test/Other/new-pm-thinlto-defaults.ll

 ; CHECK-POSTLINK-O-NEXT: Running pass: LoopDistributePass
 ; CHECK-POSTLINK-O-NEXT: Running pass: LoopVectorizePass
 ; CHECK-POSTLINK-O-NEXT: Running analysis: BlockFrequencyAnalysis
 ; CHECK-POSTLINK-O-NEXT: Running analysis: BranchProbabilityAnalysis
 ; CHECK-POSTLINK-O-NEXT: Running pass: LoopLoadEliminationPass
 ; CHECK-POSTLINK-O-NEXT: Running analysis: LoopAccessAnalysis
 ; CHECK-POSTLINK-O-NEXT: Running pass: InstCombinePass
 ; CHECK-POSTLINK-O-NEXT: Running pass: SimplifyCFGPass
+; CHECK-POSTLINK-O2-NEXT: Running pass: SLPVectorizerPass
+; CHECK-POSTLINK-O3-NEXT: Running pass: SLPVectorizerPass
+; CHECK-POSTLINK-Os-NEXT: Running pass: SLPVectorizerPass
 ; CHECK-POSTLINK-O-NEXT: Running pass: InstCombinePass
 ; CHECK-POSTLINK-O-NEXT: Running pass: LoopUnrollPass
 ; CHECK-POSTLINK-O-NEXT: Running pass: WarnMissedTransformationsPass
 ; CHECK-POSTLINK-O-NEXT: Running pass: InstCombinePass
 ; CHECK-POSTLINK-O-NEXT: Running pass: RequireAnalysisPass<{{.*}}OptimizationRemarkEmitterAnalysis
 ; CHECK-POSTLINK-O-NEXT: Running pass: FunctionToLoopPassAdaptor<{{.*}}LICMPass
 ; CHECK-POSTLINK-O-NEXT: Starting llvm::Function pass manager run
 ; CHECK-POSTLINK-O-NEXT: Running pass: LoopSimplifyPass

llvm/test/tools/gold/X86/slp-vectorize-pm.ll

+; Test plugin options new-pass-manager and debug-pass-manager
+; RUN: opt -module-summary %s -o %t.o
+
+; RUN: %gold -m elf_x86_64 -plugin %llvmshlibdir/LLVMgold%shlibext \
+; RUN:     --plugin-opt=thinlto \
+; RUN:     --plugin-opt=new-pass-manager \
+; RUN:     --plugin-opt=debug-pass-manager \
+; RUN:     --plugin-opt=cache-dir=%t.cache \
+; RUN:     -shared \
+; RUN:     -o %t1.o %t.o 2>&1 | FileCheck %s
+
+; CHECK: Starting llvm::Module pass manager run
+
+; Test SLP and Loop Vectorization are enabled by default at O2 and O3.
+; RUN: %gold -m elf_x86_64 -plugin %llvmshlibdir/LLVMgold%shlibext \
+; RUN:     --plugin-opt=thinlto \
+; RUN:     --plugin-opt=new-pass-manager \
+; RUN:     --plugin-opt=debug-pass-manager \
+; RUN:     --plugin-opt=cache-dir=%t.cache \
+; RUN:     --plugin-opt=O0 \
+; RUN:     --plugin-opt=save-temps \
+; RUN:     -shared \
+; RUN:     -o %t2.o %t.o 2>&1 | FileCheck %s --check-prefix=CHECK-O0-SLP
+; RUN: llvm-dis %t.o.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-O0-LPV
+
+; RUN: %gold -m elf_x86_64 -plugin %llvmshlibdir/LLVMgold%shlibext \
+; RUN:     --plugin-opt=thinlto \
+; RUN:     --plugin-opt=new-pass-manager \
+; RUN:     --plugin-opt=debug-pass-manager \
+; RUN:     --plugin-opt=cache-dir=%t.cache \
+; RUN:     --plugin-opt=O1 \
+; RUN:     --plugin-opt=save-temps \
+; RUN:     -shared \
+; RUN:     -o %t3.o %t.o 2>&1 | FileCheck %s --check-prefix=CHECK-O1-SLP
+; RUN: llvm-dis %t.o.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-O1-LPV
+
+; RUN: %gold -m elf_x86_64 -plugin %llvmshlibdir/LLVMgold%shlibext \
+; RUN:     --plugin-opt=thinlto \
+; RUN:     --plugin-opt=new-pass-manager \
+; RUN:     --plugin-opt=debug-pass-manager \
+; RUN:     --plugin-opt=cache-dir=%t.cache \
+; RUN:     --plugin-opt=O2 \
+; RUN:     --plugin-opt=save-temps \
+; RUN:     -shared \
+; RUN:     -o %t4.o %t.o 2>&1 | FileCheck %s --check-prefix=CHECK-O2-SLP
+; RUN: llvm-dis %t.o.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-O2-LPV
+
+; RUN: %gold -m elf_x86_64 -plugin %llvmshlibdir/LLVMgold%shlibext \
+; RUN:     --plugin-opt=thinlto \
+; RUN:     --plugin-opt=new-pass-manager \
+; RUN:     --plugin-opt=debug-pass-manager \
+; RUN:     --plugin-opt=cache-dir=%t.cache \
+; RUN:     --plugin-opt=O3 \
+; RUN:     --plugin-opt=save-temps \
+; RUN:     -shared \
+; RUN:     -o %t5.o %t.o 2>&1 | FileCheck %s --check-prefix=CHECK-O3-SLP
+; RUN: llvm-dis %t.o.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-O3-LPV
+
+; CHECK-O0-SLP-NOT: Running pass: SLPVectorizerPass
+; CHECK-O1-SLP-NOT: Running pass: SLPVectorizerPass
+; CHECK-O2-SLP: Running pass: SLPVectorizerPass
+; CHECK-O3-SLP: Running pass: SLPVectorizerPass
+; CHECK-O0-LPV-NOT: = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK-O1-LPV-NOT: = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK-O2-LPV: = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK-O3-LPV: = !{!"llvm.loop.isvectorized", i32 1}
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define i32 @foo(i32* %a) {
+entry:
+  br label %for.body
+
+for.body:
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
+  %red.05 = phi i32 [ 0, %entry ], [ %add, %for.body ]
+  %arrayidx = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
+  %0 = load i32, i32* %arrayidx, align 4
+  %add = add nsw i32 %0, %red.05
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond = icmp eq i64 %indvars.iv.next, 255
+  br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:
+  ret i32 %add
+}
+
+!0 = distinct !{!0, !1}
+!1 = !{!"llvm.loop.unroll.disable", i1 true}

llvm/test/tools/llvm-lto2/X86/slp-vectorize-pm.ll

+; RUN: opt -module-summary %s -o %t1.bc
+
+; Test SLP and Loop Vectorization are enabled by default at O2 and O3.
+; RUN: llvm-lto2 run %t1.bc -o %t2.o -O0 -r %t1.bc,foo,plx -debug-pass-manager \
+; RUN:  -use-new-pm -save-temps 2>&1 | FileCheck %s --check-prefix=CHECK-O0-SLP
+; RUN: llvm-dis %t2.o.1.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-O0-LPV
+
+; RUN: llvm-lto2 run %t1.bc -o %t3.o -O1 -r %t1.bc,foo,plx -debug-pass-manager \
+; RUN:  -use-new-pm -save-temps 2>&1 | FileCheck %s --check-prefix=CHECK-O1-SLP
+; RUN: llvm-dis %t3.o.1.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-O1-LPV
+
+; RUN: llvm-lto2 run %t1.bc -o %t4.o -O2 -r %t1.bc,foo,plx -debug-pass-manager \
+; RUN:  -use-new-pm -save-temps 2>&1 | FileCheck %s --check-prefix=CHECK-O2-SLP
+; RUN: llvm-dis %t4.o.1.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-O2-LPV
+
+; RUN: llvm-lto2 run %t1.bc -o %t5.o -O3 -r %t1.bc,foo,plx -debug-pass-manager \
+; RUN:  -use-new-pm -save-temps 2>&1 | FileCheck %s --check-prefix=CHECK-O3-SLP
+; RUN: llvm-dis %t5.o.1.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-O3-LPV
+
+; CHECK-O0-SLP-NOT: Running pass: SLPVectorizerPass
+; CHECK-O1-SLP-NOT: Running pass: SLPVectorizerPass
+; CHECK-O2-SLP: Running pass: SLPVectorizerPass
+; CHECK-O3-SLP: Running pass: SLPVectorizerPass
+; CHECK-O0-LPV-NOT: = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK-O1-LPV-NOT: = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK-O2-LPV: = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK-O3-LPV: = !{!"llvm.loop.isvectorized", i32 1}
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define i32 @foo(i32* %a) {
+entry:
+  br label %for.body
+
+for.body:
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
+  %red.05 = phi i32 [ 0, %entry ], [ %add, %for.body ]
+  %arrayidx = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
+  %0 = load i32, i32* %arrayidx, align 4
+  %add = add nsw i32 %0, %red.05
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond = icmp eq i64 %indvars.iv.next, 255
+  br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:
+  ret i32 %add
+}
+
+!0 = distinct !{!0, !1}
+!1 = !{!"llvm.loop.unroll.disable", i1 true}

llvm/tools/gold/gold-plugin.cpp

if (DataSections.getNumOccurrences() == 0)
     Conf.Options.DataSections = SplitSections;
 
   Conf.MAttrs = MAttrs;
   Conf.RelocModel = RelocationModel;
   Conf.CodeModel = getCodeModel();
   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);

llvm/tools/llvm-lto2/CMakeLists.txt

 set(LLVM_LINK_COMPONENTS
   AllTargetsAsmParsers
   AllTargetsCodeGens
   AllTargetsDescs
   AllTargetsInfos
   BitReader
   Core
   Linker
   LTO
   MC
   Object
+  Passes
   Support
   Target
   )
 
 add_llvm_tool(llvm-lto2
   llvm-lto2.cpp
 
   DEPENDS
   intrinsics_gen
   )

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

static int run(int argc, char **argv) {
   }
 
   if (FileType.getNumOccurrences())
     Conf.CGFileType = FileType;
 
   Conf.OverrideTriple = OverrideTriple;
   Conf.DefaultTriple = DefaultTriple;
   Conf.StatsFile = StatsFile;
+  Conf.PTO.LoopVectorization = Conf.OptLevel > 1;
+  Conf.PTO.SLPVectorization = Conf.OptLevel > 1;
 
   ThinBackend Backend;
   if (ThinLTODistributedIndexes)
     Backend = createWriteIndexesThinBackend(/* OldPrefix */ "",
                                             /* NewPrefix */ "",
                                             /* ShouldEmitImportsFiles */ true,
                                             /* LinkedObjectsFile */ nullptr,
                                             /* OnWrite */ {});