[Mlir-commits] [mlir] [MLIR][NVVM] Add `cmd-options` to pass flags to the downstream compiler (PR #127457)

[Guray Ozen]

================
@@ -321,6 +321,25 @@ std::optional<std::string> NVPTXSerializer::findTool(StringRef tool) {
   return std::nullopt;
 }
 
+/// Adds optional command-line arguments to existing arguments.
+template <typename T>
+static void setOptionalCommandlineArguments(NVVMTargetAttr target,
+                                            SmallVectorImpl<T> &ptxasArgs) {
+  if (!target.hasCmdOptions())
+    return;
+
+  std::optional<mlir::NamedAttribute> cmdOptions = target.getCmdOptions();
+  for (Attribute attr : cast<ArrayAttr>(cmdOptions->getValue())) {
+    if (auto strAttr = dyn_cast<StringAttr>(attr)) {
+      if constexpr (std::is_same_v<T, StringRef>) {
+        ptxasArgs.push_back(strAttr.getValue());
+      } else if constexpr (std::is_same_v<T, const char *>) {
+        ptxasArgs.push_back(strAttr.getValue().data());
+      }
----------------
grypp wrote:

good point. It's for `compileToBinaryNVPTX` 
```
  std::pair<llvm::BumpPtrAllocator, SmallVector<const char *>> cmdOpts =
      targetOptions.tokenizeCmdOptions();
  cmdOpts.second.append(
      {"-arch", getTarget().getChip().data(), "--opt-level", optLevel.c_str()});

  // Set optional command line arguments
  setOptionalCommandlineArguments(getTarget(), cmdOpts.second);
```

https://github.com/llvm/llvm-project/pull/127457