Merge branch 'sycl' into maxime/l0-perf-improvement

.clang-tidy

@@ -1,4 +1,4 @@
-Checks: '-*,clang-diagnostic-*,llvm-*,misc-*,-misc-const-correctness,-misc-unused-parameters,-misc-non-private-member-variables-in-classes,-misc-no-recursion,-misc-use-anonymous-namespace,readability-identifier-naming'
+Checks: '-*,clang-diagnostic-*,llvm-*,misc-*,-misc-const-correctness,-misc-unused-parameters,-misc-non-private-member-variables-in-classes,-misc-no-recursion,-misc-use-anonymous-namespace,readability-identifier-naming,-misc-include-cleaner'
 CheckOptions:
   - key:             readability-identifier-naming.ClassCase
     value:           CamelCase

.github/new-prs-labeler.yml

@@ -869,6 +869,9 @@ backend:SystemZ:
 third-party:unittests:
   - third-party/unittests/**
 
+third-party:benchmark:
+  - third-party/benchmark/**
+
 llvm:binary-utilities:
   - llvm/docs/CommandGuide/llvm-*
   - llvm/include/llvm/BinaryFormat/**

.github/workflows/build-ci-container.yml

@@ -77,14 +77,18 @@ jobs:
           cp ./.github/workflows/containers/github-action-ci/storage.conf ~/.config/containers/storage.conf
           podman info
 
+        # Download the container image into /mnt/podman rather than
+        # $GITHUB_WORKSPACE to avoid space limitations on the default drive
+        # and use the permissions setup for /mnt/podman.
       - name: Download stage1-toolchain
         uses: actions/download-artifact@v4
         with:
           name: stage1-toolchain
+          path: /mnt/podman
 
       - name: Load stage1-toolchain
         run: |
-          podman load -i stage1-toolchain.tar
+          podman load -i /mnt/podman/stage1-toolchain.tar
 
       - name: Build Container
         working-directory: ./.github/workflows/containers/github-action-ci/

.github/workflows/containers/github-action-ci/stage2.Dockerfile

@@ -12,11 +12,13 @@ COPY --from=stage2-toolchain $LLVM_SYSROOT $LLVM_SYSROOT
 # Need to install curl for hendrikmuhs/ccache-action
 # Need nodejs for some of the GitHub actions.
 # Need perl-modules for clang analyzer tests.
+# Need git for SPIRV-Tools tests.
 RUN apt-get update && \
     apt-get install -y \
     binutils \
     cmake \
     curl \
+    git \
     libstdc++-11-dev \
     ninja-build \
     nodejs \

.github/workflows/sycl-linux-run-tests.yml

@@ -297,17 +297,39 @@ jobs:
       if: inputs.tests_selector == 'cts'
       env:
         ONEAPI_DEVICE_SELECTOR: ${{ inputs.target_devices }}
+      # By-default GitHub actions execute the "run" shell script with -e option,
+      # so the execution terminates if any command returns a non-zero status.
+      # Since we're using a loop to run all test-binaries separately, some test
+      # may fail and terminate the execution. Setting "shell" value to override
+      # the default behavior.
+      # https://docs.github.com/en/actions/using-workflows/workflow-syntax-for-github-actions#custom-shell
+      shell: bash {0}
       # FIXME: For some reason the "sub_group api" test-case is failing with
       # SIGSEGV while running test_all, so running each binary separately.
       # BTW test_all requires a lot of resources to build it, so probably it'll
       # be better to build each binary also separately.
       # run: |
       #   ./build-cts/bin/test_all $CTS_TEST_ARGS
       run: |
+        status=""
+        failed_suites=""
         for i in `ls -1 ./build-cts/bin`; do
           if [ "$i" != "test_all" ]; then
             echo "::group::Running $i"
             build-cts/bin/$i
+            if [ $? -ne 0 ]; then
+              status=1
+              if [$failed_suites == ""]; then
+                failed_suites=$i
+              else
+                failed_suites="$failed_suites, $i"
+              fi
+            fi
             echo "::endgroup::"
           fi
         done
+        if [ -n "$status" ]; then
+          echo "Failed suite(s): $failed_suites"
+          exit 1
+        fi
+        exit 0

.github/workflows/sycl-nightly.yml

@@ -73,7 +73,7 @@ jobs:
             target_devices: ext_oneapi_cuda:gpu
 
           - name: SYCL-CTS
-            runner: '["cts-cpu"]'
+            runner: '["Linux", "gen12"]'
             image: ghcr.io/intel/llvm/ubuntu2204_intel_drivers:latest
             image_options: -u 1001 --device=/dev/dri --privileged --cap-add SYS_ADMIN
             target_devices: opencl:cpu

bolt/include/bolt/Core/BinaryFunction.h

@@ -2056,6 +2056,14 @@ class BinaryFunction {
   /// Returns false if disassembly failed.
   Error disassemble();
 
+  /// An external interface to register a branch while the function is in
+  /// disassembled state. Allows to make custom modifications to the
+  /// disassembler. E.g., a pre-CFG pass can add an instruction and register
+  /// a branch that will later be used during the CFG construction.
+  ///
+  /// Return a label at the branch destination.
+  MCSymbol *registerBranch(uint64_t Src, uint64_t Dst);
+
   Error handlePCRelOperand(MCInst &Instruction, uint64_t Address,
                            uint64_t Size);
 

bolt/include/bolt/Core/ParallelUtilities.h

@@ -18,6 +18,7 @@
 
 #include "bolt/Core/MCPlusBuilder.h"
 #include "llvm/Support/CommandLine.h"
+#include "llvm/Support/ThreadPool.h"
 
 using namespace llvm;
 
@@ -28,8 +29,6 @@ extern cl::opt<unsigned> TaskCount;
 } // namespace opts
 
 namespace llvm {
-class ThreadPool;
-
 namespace bolt {
 class BinaryContext;
 class BinaryFunction;
@@ -50,8 +49,8 @@ enum SchedulingPolicy {
   SP_BB_QUADRATIC,   /// cost is estimated by the square of the BB count
 };
 
-/// Return the managed thread pool and initialize it if not initiliazed.
-ThreadPool &getThreadPool();
+/// Return the managed thread pool and initialize it if not initialized.
+ThreadPoolInterface &getThreadPool();
 
 /// Perform the work on each BinaryFunction except those that are accepted
 /// by SkipPredicate, scheduling heuristic is based on SchedPolicy.

bolt/lib/Core/BinaryFunction.cpp

@@ -1445,6 +1445,16 @@ Error BinaryFunction::disassemble() {
   return Error::success();
 }
 
+MCSymbol *BinaryFunction::registerBranch(uint64_t Src, uint64_t Dst) {
+  assert(CurrentState == State::Disassembled &&
+         "Cannot register branch unless function is in disassembled state.");
+  assert(containsAddress(Src) && containsAddress(Dst) &&
+         "Cannot register external branch.");
+  MCSymbol *Target = getOrCreateLocalLabel(Dst);
+  TakenBranches.emplace_back(Src - getAddress(), Dst - getAddress());
+  return Target;
+}
+
 bool BinaryFunction::scanExternalRefs() {
   bool Success = true;
   bool DisassemblyFailed = false;
@@ -1759,13 +1769,6 @@ void BinaryFunction::postProcessJumpTables() {
       }
     }
   }
-
-  // Remove duplicates branches. We can get a bunch of them from jump tables.
-  // Without doing jump table value profiling we don't have use for extra
-  // (duplicate) branches.
-  llvm::sort(TakenBranches);
-  auto NewEnd = std::unique(TakenBranches.begin(), TakenBranches.end());
-  TakenBranches.erase(NewEnd, TakenBranches.end());
 }
 
 bool BinaryFunction::validateExternallyReferencedOffsets() {
@@ -2128,6 +2131,13 @@ Error BinaryFunction::buildCFG(MCPlusBuilder::AllocatorIdTy AllocatorId) {
   // e.g. exit(3), etc. Otherwise we'll see a false fall-through
   // blocks.
 
+  // Remove duplicates branches. We can get a bunch of them from jump tables.
+  // Without doing jump table value profiling we don't have a use for extra
+  // (duplicate) branches.
+  llvm::sort(TakenBranches);
+  auto NewEnd = std::unique(TakenBranches.begin(), TakenBranches.end());
+  TakenBranches.erase(NewEnd, TakenBranches.end());
+
   for (std::pair<uint32_t, uint32_t> &Branch : TakenBranches) {
     LLVM_DEBUG(dbgs() << "registering branch [0x"
                       << Twine::utohexstr(Branch.first) << "] -> [0x"

bolt/lib/Core/ParallelUtilities.cpp

@@ -49,7 +49,7 @@ namespace ParallelUtilities {
 
 namespace {
 /// A single thread pool that is used to run parallel tasks
-std::unique_ptr<ThreadPool> ThreadPoolPtr;
+std::unique_ptr<DefaultThreadPool> ThreadPoolPtr;
 
 unsigned computeCostFor(const BinaryFunction &BF,
                         const PredicateTy &SkipPredicate,
@@ -102,11 +102,11 @@ inline unsigned estimateTotalCost(const BinaryContext &BC,
 
 } // namespace
 
-ThreadPool &getThreadPool() {
+ThreadPoolInterface &getThreadPool() {
   if (ThreadPoolPtr.get())
     return *ThreadPoolPtr;
 
-  ThreadPoolPtr = std::make_unique<ThreadPool>(
+  ThreadPoolPtr = std::make_unique<DefaultThreadPool>(
       llvm::hardware_concurrency(opts::ThreadCount));
   return *ThreadPoolPtr;
 }
@@ -145,7 +145,7 @@ void runOnEachFunction(BinaryContext &BC, SchedulingPolicy SchedPolicy,
       TotalCost > BlocksCount ? TotalCost / BlocksCount : 1;
 
   // Divide work into blocks of equal cost
-  ThreadPool &Pool = getThreadPool();
+  ThreadPoolInterface &Pool = getThreadPool();
   auto BlockBegin = BC.getBinaryFunctions().begin();
   unsigned CurrentCost = 0;
 
@@ -202,7 +202,7 @@ void runOnEachFunctionWithUniqueAllocId(
       TotalCost > BlocksCount ? TotalCost / BlocksCount : 1;
 
   // Divide work into blocks of equal cost
-  ThreadPool &Pool = getThreadPool();
+  ThreadPoolInterface &Pool = getThreadPool();
   auto BlockBegin = BC.getBinaryFunctions().begin();
   unsigned CurrentCost = 0;
   unsigned AllocId = 1;

bolt/lib/Core/Relocation.cpp

@@ -381,6 +381,14 @@ static uint64_t encodeValueAArch64(uint64_t Type, uint64_t Value, uint64_t PC) {
     // OP 1001_01 goes in bits 31:26 of BL.
     Value = ((Value >> 2) & 0x3ffffff) | 0x94000000ULL;
     break;
+  case ELF::R_AARCH64_JUMP26:
+    Value -= PC;
+    assert(isInt<28>(Value) &&
+           "only PC +/- 128MB is allowed for direct branch");
+    // Immediate goes in bits 25:0 of B.
+    // OP 0001_01 goes in bits 31:26 of B.
+    Value = ((Value >> 2) & 0x3ffffff) | 0x14000000ULL;
+    break;
   }
   return Value;
 }

bolt/lib/Passes/IdenticalCodeFolding.cpp

@@ -397,7 +397,7 @@ Error IdenticalCodeFolding::runOnFunctions(BinaryContext &BC) {
     Timer SinglePass("single fold pass", "single fold pass");
     LLVM_DEBUG(SinglePass.startTimer());
 
-    ThreadPool *ThPool;
+    ThreadPoolInterface *ThPool;
     if (!opts::NoThreads)
       ThPool = &ParallelUtilities::getThreadPool();
 

bolt/lib/Rewrite/DWARFRewriter.cpp

@@ -784,7 +784,7 @@ void DWARFRewriter::updateDebugInfo() {
     }
   } else {
     // Update unit debug info in parallel
-    ThreadPool &ThreadPool = ParallelUtilities::getThreadPool();
+    ThreadPoolInterface &ThreadPool = ParallelUtilities::getThreadPool();
     for (std::unique_ptr<DWARFUnit> &CU : BC.DwCtx->compile_units())
       ThreadPool.async(processUnitDIE, CU.get(), &DIEBlder);
     ThreadPool.wait();