[clang] [AMDGPU][clang] Fix clang driver check for multiple sanitizer arguments (PR #166851)

[Thorsten Blaß via cfe-commits]

================
@@ -1074,27 +1074,32 @@ ROCMToolChain::getCommonDeviceLibNames(
       getSanitizerArgs(DriverArgs).needsAsanRt());
 }
 
-bool AMDGPUToolChain::shouldSkipSanitizeOption(
+std::optional<std::string> AMDGPUToolChain::filterSanitizeOption(
     const ToolChain &TC, const llvm::opt::ArgList &DriverArgs,
     StringRef TargetID, const llvm::opt::Arg *A) const {
   // For actions without targetID, do nothing.
   if (TargetID.empty())
-    return false;
+    return std::nullopt;
   Option O = A->getOption();
 
   if (!O.matches(options::OPT_fsanitize_EQ))
-    return false;
+    return std::nullopt;
 
   if (!DriverArgs.hasFlag(options::OPT_fgpu_sanitize,
                           options::OPT_fno_gpu_sanitize, true))
-    return true;
+    return "";
 
   auto &Diags = TC.getDriver().getDiags();
 
-  // For simplicity, we only allow -fsanitize=address
-  SanitizerMask K = parseSanitizerValue(A->getValue(), /*AllowGroups=*/false);
-  if (K != SanitizerKind::Address)
-    return true;
+  // We only allow the address sanitizer and ignore all other sanitizers.
----------------
ThorBl wrote:

Centralizing the handling of unsupported sanitizers in diagnoseUnsupportedSanitizers is a good idea. However, in its current form, it only issues a warning and doesn’t prevent unsupported sanitizers from being forwarded to the device compilation. For this issue, the logic needs to not just warn but also actively filter (drop) unsupported sanitizers from the device job to avoid the crash. Extending this function to perform the filtering would help reduce code duplication across the HIP and OpenMP toolchains.

https://github.com/llvm/llvm-project/pull/166851