[SYCL] Fix NVPTX compilation with the new offload driver (#19039)

A typical SYCL compilation for NVPTX with the default
`nvptx(64)?-nvidia-cuda` triple compiles for an older SM_50
architecture, relying on forward compatibility and JIT to run on newer
devices. Thus compilation for NVPTX with the new offload driver relies
on us generating a fat binary containing the textual PTX as well as the
compiled object. This is a kind of LTO though not officially considered
so.

A recent pulldown broke the generation of textual assembly with the `-S`
flag, and so we were later passing a compiled ELF binary into `ptxas`
which predictably went wrong.

Fixes #18432.

Author: frasercrmck

clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp

@@ -1691,7 +1691,7 @@ Expected<StringRef> linkDevice(ArrayRef<StringRef> InputFiles,
   case Triple::ppc64:
   case Triple::ppc64le:
   case Triple::systemz:
-    return generic::clang(InputFiles, Args);
+    return generic::clang(InputFiles, Args, IsSYCLKind);
   case Triple::spirv32:
   case Triple::spirv64:
   case Triple::spir:
@@ -1724,7 +1724,7 @@ Expected<StringRef> linkDevice(ArrayRef<StringRef> InputFiles,
     return generic::clang(InputFiles, Args, IsSYCLKind);
   default:
     if (Triple.str() == "native_cpu" && IsSYCLKind)
-      return generic::clang(InputFiles, Args);
+      return generic::clang(InputFiles, Args, IsSYCLKind);
 
     return createStringError(Triple.getArchName() +
                              " linking is not supported");

sycl/test-e2e/DeviceImageDependencies/NewOffloadDriver/free_function_kernels.cpp

@@ -1,7 +1,6 @@
 // Ensure -fsycl-allow-device-dependencies can work with free function kernels.
 
-// REQUIRES: aspect-usm_shared_allocations, pdtracker
-// PDTRACKER: https://github.com/intel/llvm/issues/18432
+// REQUIRES: aspect-usm_shared_allocations
 // RUN: %{build} -o %t.out --offload-new-driver -fsycl-allow-device-image-dependencies
 // RUN: %{run} %t.out
 

sycl/test-e2e/NewOffloadDriver/multisource.cpp

@@ -5,8 +5,6 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: pdtracker
-// PDTRACKER: https://github.com/intel/llvm/issues/18432
 // Separate kernel sources and host code sources
 // Test with `--offload-new-driver`
 // RUN: %{build} --offload-new-driver -c -o %t.kernel.o -DINIT_KERNEL -DCALC_KERNEL

sycl/test-e2e/NewOffloadDriver/split-per-source-main.cpp

@@ -1,5 +1,3 @@
-// REQUIRES: pdtracker
-// PDTRACKER: https://github.com/intel/llvm/issues/18432
 // RUN: %{build} -Wno-error=unused-command-line-argument -fsycl-device-code-split=per_source -I %S/Inputs -o %t.out %S/Inputs/split-per-source-second-file.cpp \
 // RUN: --offload-new-driver -fsycl-dead-args-optimization
 // RUN: %{run} %t.out

sycl/test-e2e/NewOffloadDriver/sycl-external-with-optional-features.cpp

@@ -1,5 +1,3 @@
-// REQUIRES: pdtracker
-// PDTRACKER: https://github.com/intel/llvm/issues/18432
 // Test with `--offload-new-driver`
 // RUN: %{build} -DSOURCE1 --offload-new-driver -c -o %t1.o
 // RUN: %{build} -DSOURCE2 --offload-new-driver -c -o %t2.o