[SYCL] Fix devicelib identification for NVPTX (#15357)

Recent upstream changes removed the workaround that required libraries
to be named ".cubin".

We also renamed the NVPTX libdevice library when we unified it into a
single bytecode library.

Unfortunately both of these things broke the driver's brittle logic that
determined whether or not an object was a device library. This resulted
in us not stripping any symbols from the final object, which is both
slower and broke some SYCL CTS tests.

This commit more clearly delineates logic between NVPTX and NativeCPU,
as the latter target didn't actually link against anything containing
'libdevice' so was too lenient.

Author: frasercrmck

clang/lib/Driver/ToolChains/SYCL.cpp

@@ -592,16 +592,15 @@ const char *SYCL::Linker::constructLLVMLinkCommand(
         NewLibPostfix = ".new.obj";
       std::string FileName = this->getToolChain().getInputFilename(II);
       StringRef InputFilename = llvm::sys::path::filename(FileName);
-      if (IsNVPTX || IsSYCLNativeCPU) {
-        // Linking SYCL Device libs requires libclc as well as libdevice
-        if ((InputFilename.find("libspirv") != InputFilename.npos ||
-             InputFilename.find("libdevice") != InputFilename.npos))
-          return true;
-        if (IsNVPTX) {
-          LibPostfix = ".cubin";
-          NewLibPostfix = ".new.cubin";
-        }
-      }
+      // NativeCPU links against libclc (libspirv)
+      if (IsSYCLNativeCPU && InputFilename.contains("libspirv"))
+        return true;
+      // NVPTX links against our libclc (libspirv), our libdevice (devicelib),
+      // and the CUDA libdevice
+      if (IsNVPTX && (InputFilename.starts_with("devicelib-") ||
+                      InputFilename.contains("libspirv") ||
+                      InputFilename.contains("libdevice")))
+        return true;
       StringRef LibSyclPrefix("libsycl-");
       if (!InputFilename.starts_with(LibSyclPrefix) ||
           !InputFilename.ends_with(LibPostfix) ||

clang/test/Driver/sycl-nvptx-link.cpp

@@ -0,0 +1,41 @@
+// Check that we correctly determine that the final link command links
+// devicelibs together, as far as the driver is concerned. This results in the
+// -only-needed flag.
+//
+// Note we check the names of the various device libraries because that's the
+// logic the driver uses.
+
+// Older CUDA versions had versioned libdevice files. We don't support CUDA
+// this old in SYCL, but we still test the driver's ability to pick out the
+// correctly versioned libdevice. We use Inputs/CUDA_80 which has a full set of
+// libdevice files.
+// RUN: %clang -### -fsycl -fsycl-targets=nvptx64-nvidia-cuda \
+// RUN:    -Xsycl-target-backend --cuda-gpu-arch=sm_30 \
+// RUN:    --sysroot=%S/Inputs/SYCL --cuda-path=%S/Inputs/CUDA_80/usr/local/cuda %s 2>&1 \
+// RUN:    | FileCheck %s --check-prefixes=CHECK,LIBDEVICE30
+// RUN: %clang -### -fsycl -fsycl-targets=nvptx64-nvidia-cuda \
+// RUN:    -Xsycl-target-backend --cuda-gpu-arch=sm_35 \
+// RUN:    --sysroot=%S/Inputs/SYCL --cuda-path=%S/Inputs/CUDA_80/usr/local/cuda %s 2>&1 \
+// RUN:    | FileCheck %s --check-prefixes=CHECK,LIBDEVICE35
+// RUN: %clang -### -fsycl -fsycl-targets=nvptx64-nvidia-cuda \
+// RUN:    -Xsycl-target-backend --cuda-gpu-arch=sm_50 \
+// RUN:    --sysroot=%S/Inputs/SYCL --cuda-path=%S/Inputs/CUDA_80/usr/local/cuda %s 2>&1 \
+// RUN:    | FileCheck %s --check-prefixes=CHECK,LIBDEVICE50
+
+// CUDA-9+ uses the same libdevice for all GPU variants
+// RUN: %clang -### -fsycl -fsycl-targets=nvptx64-nvidia-cuda \
+// RUN:    -Xsycl-target-backend --cuda-gpu-arch=sm_35 \
+// RUN:    --sysroot=%S/Inputs/SYCL --cuda-path=%S/Inputs/CUDA_90/usr/local/cuda %s 2>&1 \
+// RUN:    | FileCheck %s --check-prefixes=CHECK,LIBDEVICE10
+
+// First link command: ignored
+// CHECK: llvm-link
+
+// CHECK: llvm-link
+// CHECK-SAME: -only-needed
+// CHECK-SAME: devicelib--cuda.bc
+// CHECK-SAME: libspirv-nvptx64-nvidia-cuda.bc
+// LIBDEVICE10-SAME: libdevice.10.bc
+// LIBDEVICE30-SAME: libdevice.compute_30.10.bc
+// LIBDEVICE35-SAME: libdevice.compute_35.10.bc
+// LIBDEVICE50-SAME: libdevice.compute_50.10.bc