CUDA: Filter out non-static libraries during device linking

Since commit v3.12.0-rc1~278^2 (CUDA: Pass more link libraries to device
linking, 2018-03-27) we consider every link library during device
linking and use `-Xnvlink` to pass those that do not end in `.a`.
However, nvlink breaks on versioned shared library names such as
`.so.1`.  Work around this problem by not passing library paths that do
not end in `.a` or `.lib`.  nvlink would not find device symbols in them
anyway.

Fixes: #18504

1 files changed, 8 insertions, 8 deletions

diff --git a/Source/cmLinkLineDeviceComputer.cxx b/Source/cmLinkLineDeviceComputer.cxx
index 025c6e4ea2..a93ec1271d 100644
--- a/Source/cmLinkLineDeviceComputer.cxx
+++ b/Source/cmLinkLineDeviceComputer.cxx
@@ -77,15 +77,15 @@ std::string cmLinkLineDeviceComputer::ComputeLinkLibraries(
 
     std::string out;
     if (item.IsPath) {
-      // nvcc understands absolute paths to libraries ending in '.a' should
-      // be passed to nvlink.  Other extensions like '.so' or '.dylib' are
-      // rejected by the nvcc front-end even though nvlink knows to ignore
-      // them.  Bypass the front-end via '-Xnvlink'.
-      if (!cmHasLiteralSuffix(item.Value, ".a")) {
-        out += "-Xnvlink ";
+      // nvcc understands absolute paths to libraries ending in '.a' or '.lib'.
+      // These should be passed to nvlink.  Other extensions need to be left
+      // out because nvlink may not understand or need them.  Even though it
+      // can tolerate '.so' or '.dylib' it cannot tolerate '.so.1'.
+      if (cmHasLiteralSuffix(item.Value, ".a") ||
+          cmHasLiteralSuffix(item.Value, ".lib")) {
+        out += this->ConvertToOutputFormat(
+          this->ConvertToLinkReference(item.Value));
       }
-      out +=
-        this->ConvertToOutputFormat(this->ConvertToLinkReference(item.Value));
     } else if (cmLinkItemValidForDevice(item.Value)) {
       out += item.Value;
     }