[SYCL][NVPTX][NFC] Refactor NVPTX target configuration (#3535)

Avoid using `-sycldevice` for configuring NVPTX target.

Author: bader

clang/lib/Basic/Targets/NVPTX.cpp

@@ -62,8 +62,7 @@ NVPTXTargetInfo::NVPTXTargetInfo(const llvm::Triple &Triple,
                      .Default(32);
   }
 
-  // FIXME: Needed for compiling SYCL to PTX.
-  TLSSupported = Triple.getEnvironment() == llvm::Triple::SYCLDevice;
+  TLSSupported = false;
   VLASupported = false;
   AddrSpaceMap = &NVPTXAddrSpaceMap;
   GridValues = llvm::omp::NVPTXGpuGridValues;

clang/lib/Basic/Targets/NVPTX.h

@@ -171,6 +171,12 @@ class LLVM_LIBRARY_VISIBILITY NVPTXTargetInfo : public TargetInfo {
     return CCCR_Warning;
   }
 
+  void adjust(LangOptions &Opts) override {
+    TargetInfo::adjust(Opts);
+    // FIXME: Needed for compiling SYCL to PTX.
+    TLSSupported = TLSSupported || Opts.SYCLIsDevice;
+  }
+
   bool hasExtIntType() const override { return true; }
 };
 } // namespace targets

llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp

@@ -304,8 +304,9 @@ void NVPTXPassConfig::addIRPasses() {
   const NVPTXSubtarget &ST = *getTM<NVPTXTargetMachine>().getSubtargetImpl();
   addPass(createNVVMReflectPass(ST.getSmVersion()));
 
-  if (getTM<NVPTXTargetMachine>().getTargetTriple().getOS() == Triple::CUDA &&
-      getTM<NVPTXTargetMachine>().getTargetTriple().getEnvironment() == Triple::SYCLDevice) {
+  // FIXME: should the target triple check be done by the pass itself?
+  // See createNVPTXLowerArgsPass as an example
+  if (getTM<NVPTXTargetMachine>().getTargetTriple().getOS() == Triple::CUDA) {
     addPass(createGlobalOffsetPass());
     addPass(createLocalAccessorToSharedMemoryPass());
   }

llvm/lib/Target/NVPTX/SYCL/LocalAccessorToSharedMemory.cpp

@@ -38,7 +38,8 @@ class LocalAccessorToSharedMemory : public ModulePass {
 
   bool runOnModule(Module &M) override {
     // Invariant: This pass is only intended to operate on SYCL kernels being
-    // compiled to the `nvptx{,64}-nvidia-cuda-sycldevice` triple.
+    // compiled to the `nvptx{,64}-nvidia-cuda` triple.
+    // TODO: make sure that non-SYCL kernels are not impacted.
     if (skipModule(M))
       return false;
 
@@ -48,7 +49,9 @@ class LocalAccessorToSharedMemory : public ModulePass {
     // Access `nvvm.annotations` to determine which functions are kernel entry
     // points.
     auto NvvmMetadata = M.getNamedMetadata("nvvm.annotations");
-    assert(NvvmMetadata && "IR compiled to PTX must have nvvm.annotations");
+    if (!NvvmMetadata)
+      return false;
+
     for (auto MetadataNode : NvvmMetadata->operands()) {
       if (MetadataNode->getNumOperands() != 3)
         continue;