Merge remote-tracking branch 'upstream/sycl' into sycl

.github/workflows/sycl-linux-precommit.yml

@@ -242,7 +242,6 @@ jobs:
 
       env: '{"LIT_FILTER":"PerformanceTests/"}'
       extra_lit_opts: -a -j 1 --param enable-perf-tests=True
-      extra_cmake_args: ${{ matrix.extra_cmake_args }}
 
       ref: ${{ github.sha }}
       merge_ref: ''

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

@@ -127,12 +127,9 @@ on:
           - '["cts-cpu"]'
           - '["Linux", "build"]'
       image:
-        description: |
-          Use option ending with ":build" for AMDGPU, ":latest" for the rest.
         type: choice
         options:
-          - 'ghcr.io/intel/llvm/sycl_ubuntu2204_nightly:latest'
-          - 'ghcr.io/intel/llvm/sycl_ubuntu2204_nightly:build'
+          - 'ghcr.io/intel/llvm/sycl_ubuntu2404_nightly:latest'
       image_options:
         description: |
           Use option with "--device=/dev/kfd" for AMDGPU, without it for the rest.
@@ -361,7 +358,9 @@ jobs:
         cts_exclude_filter=""
         # If CTS_TESTS_TO_BUILD is null - use filter
         if [ -z "$CTS_TESTS_TO_BUILD" ]; then
-          if [ "${{ contains(inputs.target_devices, 'opencl:cpu')  }}" = "true" ]; then
+          if [ "${{ contains(inputs.cts_testing_mode, 'build-only')  }}" = "true" ]; then
+            cts_exclude_filter=$PWD/devops/cts_exclude_filter_compfails
+          elif [ "${{ contains(inputs.target_devices, 'opencl:cpu')  }}" = "true" ]; then
             cts_exclude_filter=$PWD/devops/cts_exclude_filter_OCL_CPU
           elif [ "${{ contains(inputs.target_devices, 'level_zero:gpu')  }}" = "true" ]; then
             cts_exclude_filter=$PWD/devops/cts_exclude_filter_L0_GPU

.github/workflows/sycl-nightly.yml

@@ -242,7 +242,7 @@ jobs:
         body: "Daily build ${{ steps.tag.outputs.TAG }}"
         target_commitish: ${{ github.sha }}
 
-  ubuntu2204_docker_build_push:
+  docker_build_push:
     if: github.repository == 'intel/llvm'
     runs-on: [Linux, build]
     permissions:
@@ -254,42 +254,16 @@ jobs:
       with:
         name: sycl_linux_default
         path: devops/
-    - name: Build and Push Container (with drivers)
+    - name: Build and Push Container
       uses: ./devops/actions/build_container
       with:
         push: ${{ github.ref_name == 'sycl' }}
-        file: ubuntu2204_preinstalled
+        file: nightly
         username: ${{ github.repository_owner }}
         password: ${{ secrets.GITHUB_TOKEN }}
         build-args: |
           base_image=ghcr.io/intel/llvm/ubuntu2404_intel_drivers
-          base_tag=latest
+          base_tag=alldeps
         tags: |
-          ghcr.io/${{ github.repository }}/sycl_ubuntu2204_nightly:${{ github.sha }}
-          ghcr.io/${{ github.repository }}/sycl_ubuntu2204_nightly:latest
-    - name: Build and Push Container (no drivers)
-      uses: ./devops/actions/build_container
-      with:
-        push: ${{ github.ref_name == 'sycl' }}
-        file: ubuntu2204_preinstalled
-        username: ${{ github.repository_owner }}
-        password: ${{ secrets.GITHUB_TOKEN }}
-        build-args: |
-          base_image=ghcr.io/intel/llvm/ubuntu2204_base
-          base_tag=latest
-        tags: |
-          ghcr.io/${{ github.repository }}/sycl_ubuntu2204_nightly:no-drivers-${{ github.sha }}
-          ghcr.io/${{ github.repository }}/sycl_ubuntu2204_nightly:no-drivers
-    - name: Build and Push Container (Build image)
-      uses: ./devops/actions/build_container
-      with:
-        push: ${{ github.ref_name == 'sycl' }}
-        file: ubuntu2204_preinstalled
-        username: ${{ github.repository_owner }}
-        password: ${{ secrets.GITHUB_TOKEN }}
-        build-args: |
-          base_image=ghcr.io/intel/llvm/ubuntu2204_build
-          base_tag=latest
-        tags: |
-          ghcr.io/${{ github.repository }}/sycl_ubuntu2204_nightly:build-${{ github.sha }}
-          ghcr.io/${{ github.repository }}/sycl_ubuntu2204_nightly:build
+          ghcr.io/${{ github.repository }}/sycl_ubuntu2404_nightly:${{ github.sha }}
+          ghcr.io/${{ github.repository }}/sycl_ubuntu2404_nightly:latest

devops/containers/ubuntu2204_preinstalled.Dockerfile → devops/containers/nightly.Dockerfile

@@ -1,5 +1,5 @@
-ARG base_tag=latest
-ARG base_image=ghcr.io/intel/llvm/ubuntu2204_intel_drivers
+ARG base_tag=alldeps
+ARG base_image=ghcr.io/intel/llvm/ubuntu2404_intel_drivers
 
 FROM $base_image:$base_tag
 

devops/cts_exclude_filter_compfails

@@ -0,0 +1,4 @@
+# Please use "#" to add comments here.
+# Do not delete the file even if it's empty.
+# See https://github.com/intel/llvm/pull/16615
+handler

devops/dependencies.json

@@ -8,7 +8,7 @@
     },
     "igc": {
       "github_tag": "v2.5.6",
-      "version": "2.5.6",
+      "version": "v2.5.6",
       "url": "https://github.com/intel/intel-graphics-compiler/releases/tag/v2.5.6",
       "root": "{DEPS_ROOT}/opencl/runtime/linux/oclgpu"
     },

libdevice/nativecpu_utils.cpp

@@ -31,16 +31,7 @@ using __nativecpu_state = native_cpu::state;
 
 #define OCL_LOCAL __attribute__((opencl_local))
 #define OCL_GLOBAL __attribute__((opencl_global))
-
-DEVICE_EXTERNAL OCL_LOCAL void *
-__spirv_GenericCastToPtrExplicit_ToLocal(void *p, int) {
-  return (OCL_LOCAL void *)p;
-}
-
-DEVICE_EXTERNAL OCL_GLOBAL void *
-__spirv_GenericCastToPtrExplicit_ToGlobal(void *p, int) {
-  return (OCL_GLOBAL void *)p;
-}
+#define OCL_PRIVATE __attribute__((opencl_private))
 
 DEVICE_EXTERN_C void __mux_work_group_barrier(uint32_t id, uint32_t scope,
                                               uint32_t semantics);
@@ -61,6 +52,23 @@ __spirv_MemoryBarrier(uint32_t Memory, uint32_t Semantics) {
 // Turning clang format off here because it reorders macro invocations
 // making the following code very difficult to read.
 // clang-format off
+
+#define DefGenericCastToPtrExplImpl(sfx, asp, cv)\
+DEVICE_EXTERNAL cv asp void *\
+__spirv_GenericCastToPtrExplicit_##sfx(cv void *p ,int) {\
+  return (cv asp void *)p;\
+}
+
+#define DefGenericCastToPtrExpl(sfx, asp)\
+  DefGenericCastToPtrExplImpl(sfx, asp, )\
+  DefGenericCastToPtrExplImpl(sfx, asp, const)\
+  DefGenericCastToPtrExplImpl(sfx, asp, volatile)\
+  DefGenericCastToPtrExplImpl(sfx, asp, const volatile)
+
+DefGenericCastToPtrExpl(ToPrivate, OCL_PRIVATE)
+DefGenericCastToPtrExpl(ToLocal, OCL_LOCAL)
+DefGenericCastToPtrExpl(ToGlobal, OCL_GLOBAL)
+
 #define DefSubgroupBlockINTEL1(Type, PType)                                    \
   template <>                                                                  \
   __SYCL_CONVERGENT__ DEVICE_EXTERNAL Type                                     \

llvm/docs/requirements-hashed.txt

@@ -151,7 +151,7 @@ imagesize==1.4.1 \
     --hash=sha256:0d8d18d08f840c19d0ee7ca1fd82490fdc3729b7ac93f49870406ddde8ef8d8b \
     --hash=sha256:69150444affb9cb0d5cc5a92b3676f0b2fb7cd9ae39e947a5e11a36b4497cd4a
     # via sphinx
-jinja2==3.1.4 \
+jinja2==3.1.5 \
     --hash=sha256:4a3aee7acbbe7303aede8e9648d13b8bf88a429282aa6122a993f0ac800cb369 \
     --hash=sha256:bc5dd2abb727a5319567b7a813e6a2e7318c39f4f487cfe6c89c6f9c7d25197d
     # via

llvm/lib/SYCLLowerIR/ESIMD/ESIMDOptimizeVecArgCallConv.cpp

@@ -354,9 +354,8 @@ optimizeFunction(Function *OldF,
       // preserve data flow equality to the original.
       unsigned OldArgNo = PI.getFormalParam().getArgNo();
       unsigned NewArgNo = oldArgNo2NewArgNo(OldArgNo, SretInd);
-      Instruction *At = nullptr;
       Value *Val = NewF->getArg(NewArgNo);
-      StoreInst *St = new StoreInst(Val, Alloca, false, Al, At);
+      StoreInst *St = new StoreInst(Val, Alloca, false, Al);
       NewInsts.push_back(St);
     }
   }

llvm/lib/SYCLLowerIR/ESIMD/LowerESIMD.cpp

@@ -1239,7 +1239,7 @@ static Instruction *addCastInstIfNeeded(Instruction *OldI, Instruction *NewI,
   if (OITy != NITy) {
     auto CastOpcode = CastInst::getCastOpcode(NewI, false, OITy, false);
     NewI = CastInst::Create(CastOpcode, NewI, OITy,
-                            NewI->getName() + ".cast.ty", OldI);
+                            NewI->getName() + ".cast.ty", OldI->getIterator());
     NewI->setDebugLoc(OldI->getDebugLoc());
   }
   return NewI;
@@ -1565,7 +1565,7 @@ static void translateESIMDIntrinsicCall(CallInst &CI) {
   CallInst *NewCI = IntrinsicInst::Create(
       NewFDecl, GenXArgs,
       NewFDecl->getReturnType()->isVoidTy() ? "" : CI.getName() + ".esimd",
-      &CI);
+      CI.getIterator());
   NewCI->setDebugLoc(CI.getDebugLoc());
   if (DoesFunctionReturnStructure) {
     IRBuilder<> Builder(&CI);

llvm/lib/SYCLLowerIR/GlobalOffset.cpp

@@ -250,7 +250,7 @@ void GlobalOffsetPass::addImplicitParameterToCallers(
           /* Func= */ CalleeWithImplicitParam,
           /* Args= */ ImplicitOffsets,
           /* NameStr= */ Twine(),
-          /* InsertBefore= */ CallToOld);
+          /* InsertBefore= */ CallToOld->getIterator());
       NewCallInst->setTailCallKind(CallToOld->getTailCallKind());
       NewCallInst->copyMetadata(*CallToOld);
       CallToOld->replaceAllUsesWith(NewCallInst);

llvm/lib/SYCLLowerIR/LocalAccessorToSharedMemory.cpp

@@ -177,7 +177,7 @@ Function *LocalAccessorToSharedMemoryPass::processKernel(Module &M,
               ConstantInt::get(Type::getInt32Ty(M.getContext()), 0, false),
               NFA,
           },
-          /* NameStr= */ Twine{NFA->getName()}, InsertBefore);
+          /* NameStr= */ Twine{NFA->getName()}, InsertBefore->getIterator());
       // Then create a bitcast to make sure the new pointer is the same type
       // as the old one. This will only ever be a `i8 addrspace(3)*` to `i32
       // addrspace(3)*` type of cast.

llvm/lib/SYCLLowerIR/LowerInvokeSimd.cpp

@@ -461,8 +461,8 @@ bool processInvokeSimdCall(CallInst *InvokeSimd,
     NewInvokeSimdArgs.push_back(NewHelper);
     auto ThirdArg = std::next(InvokeSimd->arg_begin(), 2);
     NewInvokeSimdArgs.append(ThirdArg, InvokeSimd->arg_end());
-    CallInst *NewInvokeSimd =
-        CallInst::Create(NewInvokeSimdF, NewInvokeSimdArgs, "", InvokeSimd);
+    CallInst *NewInvokeSimd = CallInst::Create(
+        NewInvokeSimdF, NewInvokeSimdArgs, "", InvokeSimd->getIterator());
     // - transfer flags, attributes (with shrinking), calling convention:
     NewInvokeSimd->copyIRFlags(InvokeSimd);
     NewInvokeSimd->setCallingConv(InvokeSimd->getCallingConv());

llvm/lib/SYCLLowerIR/LowerWGScope.cpp

@@ -981,7 +981,7 @@ Value *spirv::genPseudoLocalID(Instruction &Before, const Triple &TT) {
       Align Alignment = M.getDataLayout().getPreferredAlign(G);
       G->setAlignment(MaybeAlign(Alignment));
     }
-    Value *Res = new LoadInst(G->getValueType(), G, "", &Before);
+    Value *Res = new LoadInst(G->getValueType(), G, "", Before.getIterator());
     return Res;
   }
 }

llvm/lib/SYCLLowerIR/SYCLJointMatrixTransform.cpp

@@ -21,6 +21,7 @@ namespace {
 
 static constexpr char ACCESS_CHAIN[] = "_Z19__spirv_AccessChain";
 static constexpr char MATRIX_TYPE[] = "spirv.CooperativeMatrixKHR";
+static constexpr char MATRIX_LAYOUT[] = "joint_matrix_layout_to_spv";
 
 Type *getInnermostType(Type *Ty) {
   while (auto *ArrayTy = dyn_cast<ArrayType>(Ty))
@@ -184,17 +185,99 @@ bool transformAccessChain(Function *F) {
   }
   return ModuleChanged;
 }
+
+StoreInst *findLastStoreBeforeLoad(Value *Ptr, Instruction *Load) {
+  BasicBlock::iterator It(Load);
+  while (It != Load->getParent()->begin()) {
+    --It;
+    if (auto *Store = dyn_cast<StoreInst>(&*It))
+      if (Store->getPointerOperand() == Ptr)
+        return Store;
+  }
+  return nullptr;
+}
+
+// Per SPIR-V specification Layout of a matrix must be a constant instruction
+// aka a constexpr or specialization constant. Meanwhile in SYCL headers
+// layout is passed as a parameter to joint_matrix_load function, so even if
+// that layout is a constant expression in the user's code - it's not possible
+// to prove that to the compiler, so constant propagation will happen only
+// after inlining, not in AST. That means, that with O0 layout would remain
+// to be a runtime variable in LLVM IR.
+// SYCL matrix layout is being mapped on SPIR-V matrix layout by
+// joint_matrix_layout_to_spv function. The following routine finds calls to
+// this function and replaces them with the found constant.
+// This function also cleans up code, that becomes dead. Pattern of the dead
+// code is stable, as user's code doesn't affect it.
+bool propagateConstexprLayout(Function *F) {
+  llvm::SmallVector<Instruction *, 8> ToErase;
+  for (auto I = F->user_begin(), E = F->user_end(); I != E;) {
+    User *U = *I++;
+    auto *CI = dyn_cast<CallInst>(U);
+    if (!CI)
+      continue;
+    auto *Op = dyn_cast<Instruction>(CI->getArgOperand(0));
+    if (!Op || !isa<LoadInst>(Op))
+      continue;
+    auto *Ptr = dyn_cast<Instruction>(cast<LoadInst>(Op)->getPointerOperand());
+    if (!Ptr)
+      continue;
+
+    ConstantInt *ConstLayout = nullptr;
+    StoreInst *SI = findLastStoreBeforeLoad(Ptr, Op);
+    if (!SI)
+      continue;
+    ConstLayout = dyn_cast<ConstantInt>(SI->getValueOperand());
+    if (ConstLayout) {
+      CI->replaceAllUsesWith(ConstLayout);
+      ToErase.push_back(CI);
+      ToErase.push_back(SI);
+      ToErase.push_back(Op);
+      ToErase.push_back(Ptr);
+      if (auto *Cast = dyn_cast<AddrSpaceCastInst>(Ptr)) {
+        auto *OrigPtr = Cast->getPointerOperand();
+        if (auto *AI = dyn_cast<AllocaInst>(OrigPtr))
+          ToErase.push_back(AI);
+      }
+    }
+  }
+
+  // There are possible cases, when a single instruction result is used multiple
+  // times. For this case we have to use a vector to store such instructions
+  // and keep track if we have removed them before to avoid double free().
+  SmallPtrSet<Instruction *, 8> Erased;
+  for (Instruction *II : ToErase) {
+    if (!II->use_empty())
+      continue;
+    if (Erased.contains(II))
+      continue;
+    II->dropAllReferences();
+    II->eraseFromParent();
+    Erased.insert(II);
+  }
+  return !ToErase.empty();
+}
 } // namespace
 
 PreservedAnalyses
 SYCLJointMatrixTransformPass::run(Module &M, ModuleAnalysisManager &MAM) {
   bool ModuleChanged = false;
+  llvm::SmallVector<Function *, 1> ToErase;
   for (Function &F : M) {
-    if (!F.isDeclaration())
-      continue;
+    if (!F.isDeclaration()) {
+      if (F.getName() == MATRIX_LAYOUT) {
+        ModuleChanged |= propagateConstexprLayout(&F);
+        ToErase.push_back(&F);
+      } else
+        continue;
+    }
     if (F.getName().starts_with(ACCESS_CHAIN))
       ModuleChanged |= transformAccessChain(&F);
   }
 
+  for (auto *F : ToErase)
+    if (F->users().empty())
+      F->eraseFromParent();
+
   return ModuleChanged ? PreservedAnalyses::none() : PreservedAnalyses::all();
 }