[SYCL] Match explicit offload arch for AMD and NVIDIA

[jchlanda]

Fixes: #6792

When specifying multiple SYCL targets make sure that we correctly match offload arch with the target. Normally this is fixed up later on (when calling SYCLActionBuilder::withBoundArchForToolChain), but in case of creating libraries we might end up in a broken state, as the code relies on ordering of the gpu map.
See the phases of the following clang invocation:
clang++ -fsycl -fsycl-targets=nvptx64-nvidia-cuda,amdgcn-amd-amdhsa -Xsycl-target-backend=amdgcn-amd-amdhsa --offload-arch=gfx908 -Xsycl-target-backend=nvptx64-nvidia-cuda --offload-arch=sm_86 -c in.cpp -o out.o -ccc-print-phases

         +- 0: input, "/dash_c_multiple_targets.cpp", c++, (device-sycl, gfx908)
      +- 1: preprocessor, {0}, c++-cpp-output, (device-sycl, gfx908)
   +- 2: compiler, {1}, ir, (device-sycl, gfx908)
+- 3: offload, "device-sycl (nvptx64-nvidia-cuda:gfx908)" {2}, ir
|        +- 4: input, "/dash_c_multiple_targets.cpp", c++, (device-sycl, sm_86)
|     +- 5: preprocessor, {4}, c++-cpp-output, (device-sycl, sm_86)
|  +- 6: compiler, {5}, ir, (device-sycl, sm_86)
|- 7: offload, "device-sycl (amdgcn-amd-amdhsa:sm_86)" {6}, ir
|                 +- 8: input, "/dash_c_multiple_targets.cpp", c++, (host-sycl)
|              +- 9: append-footer, {8}, c++, (host-sycl)
|           +- 10: preprocessor, {9}, c++-cpp-output, (host-sycl)
|        +- 11: offload, "host-sycl (x86_64-unknown-linux-gnu)" {10}, "device-sycl (amdgcn-amd-amdhsa:sm_86)" {6}, c++-cpp-output
|     +- 12: compiler, {11}, ir, (host-sycl)
|  +- 13: backend, {12}, assembler, (host-sycl)
|- 14: assembler, {13}, object, (host-sycl)
15: clang-offload-bundler, {3, 7, 14}, object, (host-sycl)

where we end up in mismatched offload arch.

[jchlanda]

@mdtoguchi I don't have access to an Intel GPU, but I suspect that the creation of Devicename (https://github.com/jchlanda/llvm/blob/jakub/dash_c_arch_fix/clang/lib/Driver/Driver.cpp#L5801) could suffer from the same problem.

[mdtoguchi]

@mdtoguchi I don't have access to an Intel GPU, but I suspect that the creation of Devicename (https://github.com/jchlanda/llvm/blob/jakub/dash_c_arch_fix/clang/lib/Driver/Driver.cpp#L5801) could suffer from the same problem.

Thanks for the pointer - I'll take a look.

[jchlanda]

My bad, sorry. Added now.

[mdtoguchi]

LGTM