Reimplement the fix as a compiler pass

Author: fineg74

llvm/include/llvm/SYCLLowerIR/ESIMD/ESIMDUtils.h

@@ -16,6 +16,7 @@ namespace llvm {
 namespace esimd {
 
 constexpr char ATTR_DOUBLE_GRF[] = "esimd-double-grf";
+constexpr char ATTR_ESIMD_KERNEL[] = "sycl_explicit_simd";
 
 using CallGraphNodeAction = std::function<void(Function *)>;
 void traverseCallgraphUp(llvm::Function *F, CallGraphNodeAction NodeF,
@@ -34,6 +35,8 @@ void traverseCallgraphUp(Function *F, CallGraphNodeActionF ActionF,
 
 // Tells whether given function is a ESIMD kernel.
 bool isESIMDKernel(const Function &F);
+// Tells whether given function is a ESIMD function.
+bool isESIMD(const Function &F);
 
 /// Reports and error with the message \p Msg concatenated with the optional
 /// \p OptMsg if \p Condition is false.

llvm/include/llvm/SYCLLowerIR/ESIMD/LowerESIMD.h

@@ -76,6 +76,12 @@ class SYCLLowerESIMDKernelPropsPass
   PreservedAnalyses run(Module &M, ModuleAnalysisManager &);
 };
 
+// Fixes ESIMD Kernel attributes for wrapper functions for ESIMD kernels
+class SYCLLowerESIMDKernelAttrPass
+    : public PassInfoMixin<SYCLLowerESIMDKernelPropsPass> {
+public:
+  PreservedAnalyses run(Module &M, ModuleAnalysisManager &);
+};
 } // namespace llvm
 
 #endif // LLVM_SYCLLOWERIR_LOWERESIMD_H

llvm/lib/Passes/PassRegistry.def

@@ -1,4 +1,4 @@
-if(${CMAKE_VERSION} VERSION_LESS 3.14)
+if (${CMAKE_VERSION} VERSION_LESS 3.14)
     macro(FetchContent_MakeAvailable NAME)
         FetchContent_GetProperties(${NAME})
         if(NOT ${NAME}_POPULATED)
@@ -8,14 +8,14 @@ if(${CMAKE_VERSION} VERSION_LESS 3.14)
     endmacro()
 endif()
 
-# Lowering of SYCL ESIMD kernels depends on vc-intrinsics
-# NOTE: could have been added earlier from llvm/projects
+#Lowering of SYCL ESIMD kernels depends on vc - intrinsics
+#NOTE : could have been added earlier from llvm / projects
 if (NOT TARGET LLVMGenXIntrinsics)
   if (NOT DEFINED LLVMGenXIntrinsics_SOURCE_DIR)
     set(LLVMGenXIntrinsics_GIT_REPO https://github.com/intel/vc-intrinsics.git)
-    # Author: Haohai Wen <[email protected]>
-    # Date:   Fri Jun 10 03:46:22 2022 +0000
-    # Format InstructionSimplify names
+#Author : Haohai Wen < haohai.wen @intel.com>
+#Date : Fri Jun 10 03 : 46 : 22 2022 + 0000
+#Format InstructionSimplify names
     set(LLVMGenXIntrinsics_GIT_TAG abce9184b7a3a7fe1b02289b9285610d9dc45465)
 
     message(STATUS "vc-intrinsics repo is missing. Will try to download it from ${LLVMGenXIntrinsics_GIT_REPO}")
@@ -30,7 +30,7 @@ if (NOT TARGET LLVMGenXIntrinsics)
     set(LLVMGenXIntrinsics_SOURCE_DIR ${vc-intrinsics_SOURCE_DIR})
     set(LLVMGenXIntrinsics_BINARY_DIR ${vc-intrinsics_BINARY_DIR})
   else (NOT DEFINED LLVMGenXIntrinsics_SOURCE_DIR)
-    # -DLLVMGenXIntrinsics_SOURCE_DIR is provided
+#- DLLVMGenXIntrinsics_SOURCE_DIR is provided
     message(STATUS "vc-intrinsics are added manually ${LLVMGenXIntrinsics_SOURCE_DIR}")
 
     set(LLVMGenXIntrinsics_BINARY_DIR ${CMAKE_BINARY_DIR}/vc-intrinsics-build)
@@ -53,6 +53,7 @@ add_llvm_component_library(LLVMSYCLLowerIR
   ESIMD/LowerESIMDVecArg.cpp
   ESIMD/ESIMDUtils.cpp
   ESIMD/ESIMDVerifier.cpp
+  ESIMD/LowerESIMDKernelAttrs.cpp
   LowerInvokeSimd.cpp
   LowerWGScope.cpp
   LowerWGLocalMemory.cpp
@@ -77,7 +78,7 @@ add_llvm_component_library(LLVMSYCLLowerIR
   LLVMGenXIntrinsics
   LLVMDemangle
   LLVMTransformUtils
-  
+
   LINK_COMPONENTS
   Analysis
   Core

llvm/lib/SYCLLowerIR/CMakeLists.txt

@@ -69,9 +69,12 @@ void traverseCallgraphUp(llvm::Function *F, CallGraphNodeAction ActionF,
   }
 }
 
+bool isESIMD(const Function &F) {
+  return F.getMetadata(ATTR_ESIMD_KERNEL) != nullptr;
+}
+
 bool isESIMDKernel(const Function &F) {
-  return (F.getCallingConv() == CallingConv::SPIR_KERNEL) &&
-         (F.getMetadata("sycl_explicit_simd") != nullptr);
+  return (F.getCallingConv() == CallingConv::SPIR_KERNEL) && isESIMD(F);
 }
 
 } // namespace esimd

llvm/lib/SYCLLowerIR/ESIMD/ESIMDUtils.cpp

@@ -0,0 +1,39 @@
+//===---- LowerESIMDKernelAttrs - lower __esimd_set_kernel_attributes ---===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+// Finds and adds  sycl_explicit_simd attributes to wrapper functions that wrap
+// ESIMD kernel functions
+
+#include "llvm/SYCLLowerIR/ESIMD/ESIMDUtils.h"
+#include "llvm/SYCLLowerIR/ESIMD/LowerESIMD.h"
+
+#include "llvm/IR/Module.h"
+#include "llvm/Pass.h"
+
+#define DEBUG_TYPE "LowerESIMDKernelAttrs"
+
+using namespace llvm;
+
+namespace llvm {
+PreservedAnalyses
+SYCLLowerESIMDKernelAttrPass::run(Module &M, ModuleAnalysisManager &MAM) {
+  bool Modified = false;
+  for (Function &F : M) {
+    if (llvm::esimd::isESIMD(F)) {
+      llvm::esimd::traverseCallgraphUp(&F, [&](Function *GraphNode) {
+        if (!llvm::esimd::isESIMD(*GraphNode)) {
+          GraphNode->setMetadata(
+              llvm::esimd::ATTR_ESIMD_KERNEL,
+              llvm::MDNode::get(GraphNode->getContext(), {}));
+          Modified = true;
+        }
+      });
+    }
+  }
+  return Modified ? PreservedAnalyses::none() : PreservedAnalyses::all();
+}
+} // namespace llvm

llvm/lib/SYCLLowerIR/ESIMD/LowerESIMDKernelAttrs.cpp

@@ -136,63 +136,6 @@ bool isEntryPoint(const Function &F, bool EmitOnlyKernelsAsEntryPoints) {
 bool isESIMDFunction(const Function &F) {
   return F.getMetadata(ESIMD_MARKER_MD) != nullptr;
 }
-// Represents a call graph between functions in a module. Nodes are functions,
-// edges are "calls" relation.
-class CallGraph {
-public:
-  using FunctionSet = SmallPtrSet<const Function *, 16>;
-
-private:
-  std::unordered_map<const Function *, FunctionSet> Graph;
-  SmallPtrSet<const Function *, 1> EmptySet;
-  FunctionSet AddrTakenFunctions;
-
-public:
-  CallGraph(const Module &M) {
-    for (const auto &F : M) {
-      for (const Value *U : F.users()) {
-        if (const auto *I = dyn_cast<CallInst>(U)) {
-          if (I->getCalledFunction() == &F) {
-            const Function *F1 = I->getFunction();
-            Graph[F1].insert(&F);
-          }
-        }
-      }
-      if (F.hasAddressTaken()) {
-        AddrTakenFunctions.insert(&F);
-      }
-    }
-  }
-
-  iterator_range<FunctionSet::const_iterator>
-  successors(const Function *F) const {
-    auto It = Graph.find(F);
-    return (It == Graph.end())
-               ? make_range(EmptySet.begin(), EmptySet.end())
-               : make_range(It->second.begin(), It->second.end());
-  }
-
-  iterator_range<FunctionSet::const_iterator> addrTakenFunctions() const {
-    return make_range(AddrTakenFunctions.begin(), AddrTakenFunctions.end());
-  }
-};
-
-bool isESIMDFunctionInCallChain(const Function *F, const CallGraph &Deps) {
-  if (isESIMDFunction(*F)) {
-    return true;
-  }
-  for (const Function *F1 : Deps.successors(F)) {
-    if (isESIMDFunctionInCallChain(F1, Deps)) {
-      return true;
-    }
-    // Optimization to reduce recursion depth.
-    // No need to check function calls beyond kernel function.
-    if (isEntryPoint(*F1, true)) {
-      return false;
-    }
-  }
-  return false;
-}
 
 // This function makes one or two groups depending on kernel types (SYCL, ESIMD)
 EntryPointGroupVec
@@ -208,7 +151,7 @@ groupEntryPointsByKernelType(const ModuleDesc &MD,
         !MD.isEntryPointCandidate(F))
       continue;
 
-    if (isESIMDFunctionInCallChain(&F, CallGraph{M}))
+    if (isESIMDFunction(F))
       EntryPointMap[ESIMD_SCOPE_NAME].insert(&F);
     else
       EntryPointMap[SYCL_SCOPE_NAME].insert(&F);
@@ -337,6 +280,47 @@ groupEntryPointsByAttribute(const ModuleDesc &MD, StringRef AttrName,
   return EntryPointGroups;
 }
 
+// Represents a call graph between functions in a module. Nodes are functions,
+// edges are "calls" relation.
+class CallGraph {
+public:
+  using FunctionSet = SmallPtrSet<const Function *, 16>;
+
+private:
+  std::unordered_map<const Function *, FunctionSet> Graph;
+  SmallPtrSet<const Function *, 1> EmptySet;
+  FunctionSet AddrTakenFunctions;
+
+public:
+  CallGraph(const Module &M) {
+    for (const auto &F : M) {
+      for (const Value *U : F.users()) {
+        if (const auto *I = dyn_cast<CallInst>(U)) {
+          if (I->getCalledFunction() == &F) {
+            const Function *F1 = I->getFunction();
+            Graph[F1].insert(&F);
+          }
+        }
+      }
+      if (F.hasAddressTaken()) {
+        AddrTakenFunctions.insert(&F);
+      }
+    }
+  }
+
+  iterator_range<FunctionSet::const_iterator>
+  successors(const Function *F) const {
+    auto It = Graph.find(F);
+    return (It == Graph.end())
+               ? make_range(EmptySet.begin(), EmptySet.end())
+               : make_range(It->second.begin(), It->second.end());
+  }
+
+  iterator_range<FunctionSet::const_iterator> addrTakenFunctions() const {
+    return make_range(AddrTakenFunctions.begin(), AddrTakenFunctions.end());
+  }
+};
+
 void collectFunctionsToExtract(SetVector<const GlobalValue *> &GVs,
                                const EntryPointGroup &ModuleEntryPoints,
                                const CallGraph &Deps) {

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

@@ -665,6 +665,10 @@ processInputModule(std::unique_ptr<Module> M) {
   // if none were made.
   bool Modified = false;
 
+  // Propagate ESIMD attribute to wrapper functions to prevent
+  // spurious splits and kernel link errors.
+  Modified |= runModulePass<SYCLLowerESIMDKernelAttrPass>(*M);
+
   // After linking device bitcode "llvm.used" holds references to the kernels
   // that are defined in the device image. But after splitting device image into
   // separate kernels we may end up with having references to kernel declaration