Merge branch 'sycl' into urlza428

Author: winstonzhang-intel

.github/workflows/sycl-nightly.yml

@@ -110,7 +110,7 @@ jobs:
       image_options: ${{ matrix.image_options }}
       target_devices: ${{ matrix.target_devices }}
       tests_selector: ${{ matrix.tests_selector }}
-      extra_lit_opts: ${{ matrix.extra_lit_opts }}
+      extra_lit_opts: "--param 'cxx_flags=-D_GLIBCXX_USE_CXX11_ABI=0' ${{ matrix.extra_lit_opts }}"
       reset_intel_gpu: ${{ matrix.reset_intel_gpu }}
       ref: ${{ github.sha }}
       merge_ref: ''

libclc/utils/libclc-remangler/LibclcRemangler.cpp

@@ -105,6 +105,8 @@ class BumpPointerAllocator {
 public:
   BumpPointerAllocator()
       : BlockList(new(InitialBuffer) BlockMeta{nullptr, 0}) {}
+  BumpPointerAllocator(const BumpPointerAllocator &) = delete;
+  BumpPointerAllocator &operator=(const BumpPointerAllocator &) = delete;
 
   void *allocate(size_t N) {
     N = (N + 15u) & ~15u;

llvm/include/llvm/Support/PropertySetIO.h

@@ -212,6 +212,9 @@ class PropertySetRegistry {
   static constexpr char SYCL_VIRTUAL_FUNCTIONS[] = "SYCL/virtual functions";
   static constexpr char SYCL_IMPLICIT_LOCAL_ARG[] = "SYCL/implicit local arg";
 
+  static constexpr char PROPERTY_REQD_WORK_GROUP_SIZE[] =
+      "reqd_work_group_size_uint64_t";
+
   /// Function for bulk addition of an entire property set in the given
   /// \p Category .
   template <typename MapTy> void add(StringRef Category, const MapTy &Props) {
@@ -231,6 +234,17 @@ class PropertySetRegistry {
     PropSet.insert({PropName, PropertyValue(PropVal)});
   }
 
+  void remove(StringRef Category, StringRef PropName) {
+    auto PropertySetIt = PropSetMap.find(Category);
+    if (PropertySetIt == PropSetMap.end())
+      return;
+    auto &PropertySet = PropertySetIt->second;
+    auto PropIt = PropertySet.find(PropName);
+    if (PropIt == PropertySet.end())
+      return;
+    PropertySet.erase(PropIt);
+  }
+
   /// Parses from the given \p Buf a property set registry.
   static Expected<std::unique_ptr<PropertySetRegistry>>
   read(const MemoryBuffer *Buf);

llvm/lib/SYCLLowerIR/SYCLDeviceRequirements.cpp

@@ -40,7 +40,6 @@ SYCLDeviceRequirements
 llvm::computeDeviceRequirements(const Module &M,
                                 const SetVector<Function *> &EntryPoints) {
   SYCLDeviceRequirements Reqs;
-  bool MultipleReqdWGSize = false;
   // Process all functions in the module
   for (const Function &F : M) {
     if (auto *MDN = F.getMetadata("sycl_used_aspects")) {
@@ -81,8 +80,6 @@ llvm::computeDeviceRequirements(const Module &M,
             ExtractUnsignedIntegerFromMDNodeOperand(MDN, I));
       if (!Reqs.ReqdWorkGroupSize.has_value())
         Reqs.ReqdWorkGroupSize = NewReqdWorkGroupSize;
-      if (Reqs.ReqdWorkGroupSize != NewReqdWorkGroupSize)
-        MultipleReqdWGSize = true;
     }
 
     if (auto *MDN = F.getMetadata("sycl_joint_matrix")) {
@@ -119,13 +116,6 @@ llvm::computeDeviceRequirements(const Module &M,
     }
   }
 
-  // Usually, we would only expect one ReqdWGSize, as the module passed to
-  // this function would be split according to that. However, when splitting
-  // is disabled, this cannot be guaranteed. In this case, we reset the value,
-  // which makes so that no value is reqd_work_group_size data is attached in
-  // in the device image.
-  if (MultipleReqdWGSize)
-    Reqs.ReqdWorkGroupSize.reset();
   return Reqs;
 }
 
@@ -152,7 +142,8 @@ std::map<StringRef, util::PropertyValue> SYCLDeviceRequirements::asMap() const {
   // reqd_work_group_size_uint64_t attribute. At the next ABI-breaking
   // window, this can be changed back to reqd_work_group_size.
   if (ReqdWorkGroupSize.has_value())
-    Requirements["reqd_work_group_size_uint64_t"] = *ReqdWorkGroupSize;
+    Requirements[util::PropertySetRegistry::PROPERTY_REQD_WORK_GROUP_SIZE] =
+        *ReqdWorkGroupSize;
 
   if (JointMatrix.has_value())
     Requirements["joint_matrix"] = *JointMatrix;

llvm/lib/Transforms/Instrumentation/AddressSanitizer.cpp

@@ -1379,7 +1379,7 @@ static void ExtendSpirKernelArgs(Module &M, FunctionAnalysisManager &FAM,
     }
 
     // New argument: uintptr_t as(1)*, which is allocated in shared USM buffer
-    Types.push_back(IntptrTy->getPointerTo(kSpirOffloadGlobalAS));
+    Types.push_back(llvm::PointerType::get(IntptrTy, kSpirOffloadGlobalAS));
 
     FunctionType *NewFTy = FunctionType::get(F->getReturnType(), Types, false);
 
@@ -1704,7 +1704,7 @@ void AddressSanitizer::AppendDebugInfoToArgs(Instruction *InsertBefore,
 
   // SPIR constant address space
   PointerType *ConstASPtrTy =
-      Type::getInt8Ty(C)->getPointerTo(kSpirOffloadConstantAS);
+      llvm::PointerType::get(Type::getInt8Ty(C), kSpirOffloadConstantAS);
 
   // File & Line
   if (Loc) {
@@ -1863,9 +1863,9 @@ void AddressSanitizer::instrumentInitAsanLaunchInfo(
   // FIXME: if the initial value of "__AsanLaunchInfo" is zero, we'll not need
   // this step
   initializeCallbacks(TLI);
-  IRB.CreateStore(
-      ConstantPointerNull::get(IntptrTy->getPointerTo(kSpirOffloadGlobalAS)),
-      AsanLaunchInfo);
+  IRB.CreateStore(ConstantPointerNull::get(
+                      llvm::PointerType::get(IntptrTy, kSpirOffloadGlobalAS)),
+                  AsanLaunchInfo);
 }
 
 // Instrument memset/memmove/memcpy
@@ -3457,7 +3457,7 @@ void AddressSanitizer::initializeCallbacks(const TargetLibraryInfo *TLI) {
       // )
       if (TargetTriple.isSPIROrSPIRV()) {
         auto *Int8PtrTy =
-            Type::getInt8Ty(*C)->getPointerTo(kSpirOffloadConstantAS);
+            llvm::PointerType::get(Type::getInt8Ty(*C), kSpirOffloadConstantAS);
 
         Args1.push_back(Int8PtrTy);            // file
         Args1.push_back(Type::getInt32Ty(*C)); // line
@@ -3574,9 +3574,10 @@ void AddressSanitizer::initializeCallbacks(const TargetLibraryInfo *TLI) {
                               IRB.getVoidTy(), IntptrTy, Int32Ty);
 
     AsanLaunchInfo = M.getOrInsertGlobal(
-        "__AsanLaunchInfo", IntptrTy->getPointerTo(kSpirOffloadGlobalAS), [&] {
+        "__AsanLaunchInfo",
+        llvm::PointerType::get(IntptrTy, kSpirOffloadGlobalAS), [&] {
           return new GlobalVariable(
-              M, IntptrTy->getPointerTo(kSpirOffloadGlobalAS), false,
+              M, llvm::PointerType::get(IntptrTy, kSpirOffloadGlobalAS), false,
               GlobalVariable::ExternalLinkage, nullptr, "__AsanLaunchInfo",
               nullptr, GlobalVariable::NotThreadLocal, kSpirOffloadLocalAS);
         });

llvm/tools/sycl-post-link/sycl-post-link.cpp

@@ -310,6 +310,15 @@ std::string saveModuleProperties(module_split::ModuleDesc &MD,
   auto PropSet =
       computeModuleProperties(MD.getModule(), MD.entries(), GlobProps);
 
+  // When the split mode is none, the required work group size will be added
+  // to the whole module, which will make the runtime unable to
+  // launch the other kernels in the module that have different
+  // required work group sizes or no required work group sizes. So we need to
+  // remove the required work group size metadata in this case.
+  if (SplitMode == module_split::SPLIT_NONE)
+    PropSet.remove(PropSetRegTy::SYCL_DEVICE_REQUIREMENTS,
+                   PropSetRegTy::PROPERTY_REQD_WORK_GROUP_SIZE);
+
   std::string NewSuff = Suff.str();
   if (!Target.empty()) {
     PropSet.add(PropSetRegTy::SYCL_DEVICE_REQUIREMENTS, "compile_target",

sycl/cmake/modules/UnifiedRuntimeTag.cmake

@@ -1,7 +1,7 @@
-# commit a172cde4bc2857dcf74b0b2907a5b0f90566e808 (HEAD, origin/main, origin/HEAD)
-# Merge: 1851eff47b0a cde0d4c820b0
+# commit aa72577da55571b15028dd1710e02a7ea4f62338
+# Merge: 89def64486dd 3d2b7210e352
 # Author: Martin Grant <[email protected]>
 # Date:   Thu Dec 5 12:29:44 2024 +0000
 #   Merge pull request #2249 from zhaomaosu/use-device-usm-for-rtl-data
 #   [DeviceASAN] Use device usm to sync asan runtime data instead of shared usm
-set(UNIFIED_RUNTIME_TAG 60e6f5e5e63b06e887aa2a692650b40f2fbc4b66)
+set(UNIFIED_RUNTIME_TAG dbf1b7073174777e2dbe01f40739f72f70e39302)