[Cherry-pick] [Frontend] Run the MLIR verifier after parsing (#7999) (#542)

Summary:
TODO. num_warps isn't supposed to be on ttir but tlx requires that. We added a pass to add that attribute to ttir, but somehow upstream added a verifier before our pass.

Cherry-picked from upstream OAI repository.

Original Commit: 165dd4b
Original Author: Jeff Niu
Original Date: 2025-08-28 12:44:28 -0700

Original commit message:
```
[Frontend] Run the MLIR verifier after parsing (#7999)

The error messages generated aren't perfect, but this at least prevents
the compiler from dumping a reproducer just for verifier errors.
```

This PR was automatically cherry-picked from the upstream triton-lang/triton repository.

Pull Request resolved: #542

Reviewed By: adamomainz

Differential Revision: D85969387

Pulled By: agron911

fbshipit-source-id: ab79045b41d868000272027979586bcbbd0ed8b0

Author: agron911

python/src/ir.cc

@@ -88,6 +88,23 @@ llvm::raw_ostream &mlir_dumps_or_dbgs() {
   }
 }
 
+// Function to parse a comma-separated string into a vector of C-style strings
+llvm::SmallVector<const char *, 3>
+parseCommaSeparatedValues(const std::string &input,
+                          llvm::SmallVector<std::string, 3> &storage) {
+  llvm::SmallVector<StringRef, 3> split;
+  llvm::SmallVector<const char *, 3> result;
+  StringRef(input.c_str()).split(split, ',');
+  llvm::transform(split, std::back_inserter(result), [&storage](StringRef str) {
+    // StringRefs are not always null-terminated.
+    // The purpose for this storage pattern is to
+    // produce a collection of C-strings that are.
+    storage.push_back(str.str());
+    return storage.back().c_str();
+  });
+  return result;
+}
+
 // Run the pass manager under a source manager diagnostic handler, which
 // enables emitted MLIR diagnostics to directly reference Python source
 // code. This diagnostic handler supports filtering diagnostic info by
@@ -124,6 +141,43 @@ struct TritonSourceMgrDiagnosticHandler : public SourceMgrDiagnosticHandler {
   llvm::SourceMgr sourceMgr;
 };
 
+TritonSourceMgrDiagnosticHandler
+setupTritonDiagnosticHandler(MLIRContext *context) {
+  bool showOperations = false, showStacktraces = false, showRemarks = false,
+       showWarnings = false;
+
+  if (auto enableDiagnostics =
+          triton::tools::getStrEnv("MLIR_ENABLE_DIAGNOSTICS");
+      !enableDiagnostics.empty()) {
+    llvm::SmallVector<std::string, 3> storage;
+    parseCommaSeparatedValues(enableDiagnostics, storage);
+    for (auto &str : storage) {
+      if (str == "warnings") {
+        showWarnings = true;
+      } else if (str == "remarks") {
+        showRemarks = true;
+      } else if (str == "stacktraces") {
+        showStacktraces = true;
+      } else if (str == "operations") {
+        showOperations = true;
+      }
+      // we show errors by default, so no need to set it
+    }
+  }
+
+  DiagnosticSeverity minSeverity =
+      showWarnings ? DiagnosticSeverity::Warning : DiagnosticSeverity::Error;
+  minSeverity = showRemarks ? DiagnosticSeverity::Remark : minSeverity;
+
+  context->printOpOnDiagnostic(showOperations);
+  context->printStackTraceOnDiagnostic(showStacktraces);
+  if (showStacktraces) {
+    context->disableMultithreading();
+  }
+
+  return TritonSourceMgrDiagnosticHandler(context, minSeverity);
+}
+
 std::string locationToString(Location loc) {
   std::string str;
   llvm::raw_string_ostream os(str);
@@ -132,23 +186,6 @@ std::string locationToString(Location loc) {
   return str;
 }
 
-// Function to parse a comma-separated string into a vector of C-style strings
-llvm::SmallVector<const char *, 3>
-parseCommaSeparatedValues(const std::string &input,
-                          llvm::SmallVector<std::string, 3> &storage) {
-  llvm::SmallVector<StringRef, 3> split;
-  llvm::SmallVector<const char *, 3> result;
-  StringRef(input.c_str()).split(split, ',');
-  llvm::transform(split, std::back_inserter(result), [&storage](StringRef str) {
-    // StringRefs are not always null-terminated.
-    // The purpose for this storage pattern is to
-    // produce a collection of C-strings that are.
-    storage.push_back(str.str());
-    return storage.back().c_str();
-  });
-  return result;
-}
-
 void outputWarning(Location loc, const std::string &msg) {
   std::string locStr = locationToString(loc);
 
@@ -694,7 +731,12 @@ void init_triton_ir(py::module &&m) {
       .def("walk",
            [](ModuleOp &self, const std::function<void(Operation *)> &fn) {
              self.walk(fn);
-           });
+           })
+      .def("verify_with_diagnostics", [](ModuleOp &self) {
+        TritonSourceMgrDiagnosticHandler handler =
+            setupTritonDiagnosticHandler(self.getContext());
+        return succeeded(verify(self.getOperation()));
+      });
 
   m.def("make_attr", [](const std::vector<int> &values, MLIRContext &context) {
     return mlir::cast<Attribute>(DenseIntElementsAttr::get(
@@ -1923,42 +1965,8 @@ void init_triton_ir(py::module &&m) {
               self.enableTiming();
             }
 
-            // setting up diagnostics
-            bool showOperations = false, showStacktraces = false,
-                 showRemarks = false, showWarnings = false;
-
-            if (auto enableDiagnostics =
-                    triton::tools::getStrEnv("MLIR_ENABLE_DIAGNOSTICS");
-                !enableDiagnostics.empty()) {
-              llvm::SmallVector<std::string, 3> storage;
-              parseCommaSeparatedValues(enableDiagnostics, storage);
-              for (auto &str : storage) {
-                if (str == "warnings") {
-                  showWarnings = true;
-                } else if (str == "remarks") {
-                  showRemarks = true;
-                } else if (str == "stacktraces") {
-                  showStacktraces = true;
-                } else if (str == "operations") {
-                  showOperations = true;
-                }
-                // we show errors by default, so no need to set it
-              }
-            }
-
-            DiagnosticSeverity minSeverity = showWarnings
-                                                 ? DiagnosticSeverity::Warning
-                                                 : DiagnosticSeverity::Error;
-            minSeverity =
-                showRemarks ? DiagnosticSeverity::Remark : minSeverity;
-
-            TritonSourceMgrDiagnosticHandler diagHandler(context, minSeverity);
-
-            context->printOpOnDiagnostic(showOperations);
-            context->printStackTraceOnDiagnostic(showStacktraces);
-            if (showStacktraces) {
-              context->disableMultithreading();
-            }
+            TritonSourceMgrDiagnosticHandler diagHandler =
+                setupTritonDiagnosticHandler(context);
             if (failed(self.run(mod.getOperation())))
               throw std::runtime_error("PassManager::run failed");
           },

python/triton/compiler/code_generator.py

@@ -1673,7 +1673,14 @@ def ast_to_ttir(fn, src, context, options, codegen_fns, module_map, module=None)
         is_gluon=fn.is_gluon(),
     )
     generator.visit(fn.parse())
-    ret = generator.module
+    module = generator.module
     # module takes ownership of the context
-    ret.context = context
-    return ret
+    module.context = context
+    # Facebook begin
+    # TODO. bring following verify back
+    # if not module.verify_with_diagnostics():
+    #     if not fn.is_gluon():
+    #         print(module)
+    #     raise RuntimeError("error encountered during parsing")
+    # Facebook end
+    return module