[SYCL][CUDA] Enable memcpy optimizations for NVPTX (#18598)

The NVPTX backend doesn't support libcalls but does support the memset
and memcpy intrinsics, so this flag enables optimizations that use these
intrinsics.

This was added for the CUDA path in https://reviews.llvm.org/D106401

Author: npmiller

clang/lib/Driver/ToolChains/Cuda.cpp

@@ -966,6 +966,8 @@ void CudaToolChain::addClangTargetOptions(
     if (FastRelaxedMath || UnsafeMathOpt)
       CC1Args.append({"-mllvm", "--nvptx-prec-divf32=0", "-mllvm",
                       "--nvptx-prec-sqrtf32=0"});
+
+    CC1Args.append({"-mllvm", "-enable-memcpyopt-without-libcalls"});
   } else {
     CC1Args.append({"-fcuda-is-device", "-mllvm",
                     "-enable-memcpyopt-without-libcalls",

clang/test/Driver/sycl-nvptx-memcpy-opt.cpp

@@ -0,0 +1,5 @@
+// RUN: %clang -### -nocudalib \
+// RUN:   -fsycl -fsycl-targets=nvptx64-nvidia-cuda %s 2>&1 \
+// RUN: | FileCheck --check-prefix=CHECK-DEFAULT %s
+
+// CHECK-DEFAULT: "-enable-memcpyopt-without-libcalls"