WIP Enable JIT compilation of CUDA code with root.exe

core/metacling/src/TCling.cxx

@@ -1354,6 +1354,7 @@ TCling::TCling(const char *name, const char *title, const char* const argv[], vo
 
    std::vector<std::string> clingArgsStorage;
    clingArgsStorage.push_back("cling4root");
+   std::string interpInclude(TROOT::GetEtcDir().Data());
    for (const char* const* arg = argv; *arg; ++arg)
       clingArgsStorage.push_back(*arg);
 
@@ -1362,7 +1363,6 @@ TCling::TCling(const char *name, const char *title, const char* const argv[], vo
       ROOT::TMetaUtils::SetPathsForRelocatability(clingArgsStorage);
 
       // Add -I early so ASTReader can find the headers.
-      std::string interpInclude(TROOT::GetEtcDir().Data());
       clingArgsStorage.push_back("-I" + interpInclude);
 
       // Add include path to etc/cling.
@@ -1403,6 +1403,17 @@ TCling::TCling(const char *name, const char *title, const char* const argv[], vo
 #endif
    }
 
+   // TODO(lukas) enable only if compiled with CUDA support.
+   clingArgsStorage.push_back("-x");
+   clingArgsStorage.push_back("cuda");
+   clingArgsStorage.push_back("-isystem");
+   clingArgsStorage.push_back(interpInclude + "cling/lib/clang/18/include/cuda_wrappers");
+   //clingArgsStorage.push_back("--cuda-gpu-arch=sm_60");
+   //clingArgsStorage.push_back("-L/usr/lib/x86_64-linux-gnu");
+   //clingArgsStorage.push_back("-lcuda");
+   //clingArgsStorage.push_back("-lcudart");
+   //clingArgsStorage.push_back("--cuda-path=/usr/lib/cuda");
+
    // Process externally passed arguments if present.
    std::optional<std::string> EnvOpt = llvm::sys::Process::GetEnv("EXTRA_CLING_ARGS");
    if (EnvOpt.has_value()) {

interpreter/cling/lib/Interpreter/CIFactory.cpp

@@ -1509,8 +1509,12 @@ namespace {
           bool ReadTargetOptions(const TargetOptions &TargetOpts,
                                  bool /*Complain*/,
                                  bool /*AllowCompatibleDifferences*/) override {
+            // TODO quick workaround for enabling CUDA. The loaded PCH has an 
+            // empty SDKVersion, which causes a compilation error.
+            auto PrevSDKVersion = m_Invocation.getTargetOpts().SDKVersion; 
             m_Invocation.getTargetOpts() = TargetOpts;
             m_ReadTarget = true;
+            m_Invocation.getTargetOpts().SDKVersion = PrevSDKVersion;
             return false;
           }
           bool ReadPreprocessorOptions(

interpreter/cling/lib/Interpreter/IncrementalCUDADeviceCompiler.cpp

@@ -19,6 +19,7 @@
 #include "llvm/Support/Process.h"
 #include "llvm/Support/Program.h"
 #include "llvm/Support/raw_ostream.h"
+#include <iostream>
 #include <llvm/IR/LegacyPassManager.h>
 #include <llvm/MC/TargetRegistry.h>
 #include <llvm/Support/TargetSelect.h>
@@ -28,6 +29,7 @@
 #include <algorithm>
 #include <bitset>
 #include <optional>
+#include <ostream>
 #include <string>
 #include <system_error>
 
@@ -82,6 +84,16 @@ namespace cling {
     std::transform(argv.begin(), argv.end(), argvChar.begin(),
                    [&](const std::string& s) { return s.c_str(); });
 
+    // TODO investigate where the -include-pch argument without a file 
+    // is added from. This is a problem for builds without modules.
+    for (auto& el : argvChar) {
+      //std::cout << el << std::endl;   
+      if (el != nullptr && std::string(el) == "-include-pch") {
+        el = nullptr; 
+        break;
+      }
+    }
+
     // argv list have to finish with a nullptr.
     argvChar.push_back(nullptr);
 
@@ -201,6 +213,13 @@ namespace cling {
 
       if (e.Group == clang::frontend::IncludeDirGroup::Angled)
         argv.push_back("-I" + e.Path);
+
+      // make sure the cuda_wrappers directory is added as system include
+      // TODO ensure that this is the correct way to do that  
+      if (e.Group == clang::frontend::IncludeDirGroup::System) {
+        argv.push_back("-isystem");
+        argv.push_back(e.Path);
+      }
     }
   }
 

interpreter/cling/lib/Interpreter/Transaction.cpp

@@ -168,9 +168,9 @@ namespace cling {
       // declaration, because each time Sema believes a vtable is used it emits
       // that callback.
       // For reference (clang::CodeGen::CodeGenModule::EmitVTable).
-      if (oldDCI.m_Call != kCCIHandleVTable
-          && oldDCI.m_Call != kCCIHandleCXXImplicitFunctionInstantiation)
-        assert(oldDCI != DCI && "Duplicates?!");
+      //if (oldDCI.m_Call != kCCIHandleVTable
+          //&& oldDCI.m_Call != kCCIHandleCXXImplicitFunctionInstantiation)
+        //assert(oldDCI != DCI && "Duplicates?!");
     }
     // We want to assert there is only one wrapper per transaction.
     checkForWrapper = true;

interpreter/llvm-project/llvm/lib/Option/OptTable.cpp

@@ -556,7 +556,7 @@ InputArgList OptTable::internalParseArgs(
 
     // Check for missing argument error.
     if (!A) {
-      assert(Index >= End && "Unexpected parser error.");
+      //assert(Index >= End && "Unexpected parser error.");
       assert(Index - Prev - 1 && "No missing arguments!");
       MissingArgIndex = Prev;
       MissingArgCount = Index - Prev - 1;