[llvm-branch-commits] [clang] [Clang] Enable UBSan for AMDGPU device offload (PR #219973)

[Joseph Huber via llvm-branch-commits]

================
@@ -1673,8 +1673,33 @@ void tools::linkSanitizerRuntimeDeps(const ToolChain &TC,
     CmdArgs.push_back("-lresolv");
 }
 
+// Host interceptor library for offload UBSan. Enabled if we are
+// offloading to a target that supports UBSan.
+static bool hostNeedsUbsanOffloadRt(Compilation &C, const ToolChain &HostTC) {
+  if (HostTC.getTriple().isGPU())
+    return false;
+
+  static constexpr Action::OffloadKind Kinds[] = {
+      Action::OFK_Cuda, Action::OFK_OpenMP, Action::OFK_HIP, Action::OFK_SYCL};
+  for (Action::OffloadKind Kind : Kinds) {
+    for (const auto &Entry : llvm::make_range(C.getOffloadToolChains(Kind))) {
+      const ToolChain *DevTC = Entry.second;
+      // FIXME: CUDA/HIPSPV copy the host mask and ignore device sanitizers.
+      if (DevTC->getTriple().isNVPTX() || DevTC->getTriple().isSPIROrSPIRV())
+        continue;
+
+      const ArgList &DevArgs = C.getArgsForToolChain(DevTC, {}, Kind);
----------------
jhuber6 wrote:

Done. `hostNeedsUbsanOffloadRt()` now walks each bound architecture from `getOffloadArchs()`, so `-Xarch_gfx908 -fsanitize=undefined` pulls in the host interceptor.

https://github.com/llvm/llvm-project/pull/219973