https://github.com/vchuravy updated 
https://github.com/llvm/llvm-project/pull/212195

>From f4f02f7f9bb415a7f44e93f36bf0f6f620172367 Mon Sep 17 00:00:00 2001
From: Valentin Churavy <[email protected]>
Date: Wed, 22 Jul 2026 17:44:14 +0200
Subject: [PATCH] [flang] Add a pass-pipeline config-augmentor hook for
 -load'ed plugins

The HLFIR-to-FIR pass pipeline exposes extension points on
MLIRToLLVMPassPipelineConfig, but that config is built inside the frontend, so
a -load'ed plugin has no way to reach it and register passes.
registerDefaultInlinerPass is the only augmentor today, and it is wired in by
hand.

Add a global registry of config augmentors:

  * fir::registerPassPipelineConfigCallback(cb) appends a callback, to be
    called from a plugin's static initializer at -load time.
  * fir::invokePassPipelineConfigCallbacks(config) runs them on the config.

This mirrors what flang already does for -load'ed plugin actions via
FrontendPluginRegistry: a process-global, append-only registry populated from
static initializers, which run before any compilation begins.

Both code generation entry points invoke the callbacks after building their
config and before constructing the pipeline: CodeGenAction::lowerHLFIRToFIR,
which serves -emit-fir, and CodeGenAction::generateLLVMIR, which serves
-emit-llvm/-emit-obj and reaches createHLFIRToFIRPassPipeline through
createMLIRToLLVMPassPipeline. The two are mutually exclusive for a given
compilation, so the callbacks run exactly once either way and a plugin gets the
same behaviour whichever output the user asked for.

No-op unless a callback is registered; tools that build the same pipelines
without consulting the registry (bbc, tco) are unaffected.

Co-Authored-By: Claude Opus 5 <[email protected]>
---
 flang/docs/FlangDriver.md                     |  34 ++++++
 .../flang/Optimizer/Passes/Pipelines.h        |  11 ++
 flang/lib/Frontend/FrontendActions.cpp        |   6 +
 flang/lib/Optimizer/Passes/Pipelines.cpp      |  17 +++
 .../Optimizer/HLFIRExtensionPointsTest.cpp    | 106 +++++++++++++++++-
 5 files changed, 173 insertions(+), 1 deletion(-)

diff --git a/flang/docs/FlangDriver.md b/flang/docs/FlangDriver.md
index 07feba4f4674b..55b4534cbffbe 100644
--- a/flang/docs/FlangDriver.md
+++ b/flang/docs/FlangDriver.md
@@ -549,6 +549,40 @@ config.registerHLFIROptEarlyEPCallbacks(
     });
 ```
 
+### Reaching the Extension Points from a `-load`'ed Plugin
+
+The `MLIRToLLVMPassPipelineConfig` the frontend driver builds is a local of
+`CodeGenAction`, so a shared object loaded with `flang -fc1 -load` cannot get 
at
+it directly. `fir::registerPassPipelineConfigCallback`
+(`flang/include/flang/Optimizer/Passes/Pipelines.h`) provides a process-global
+registry of *config augmentors* for that purpose. Register one from a static
+initializer -- the same idiom `FrontendPluginRegistry` uses for `-load`'ed
+plugin actions -- and the frontend will invoke it once the config has been
+built, before the pipeline is constructed:
+
+```c++
+struct MyPluginRegistration {
+  MyPluginRegistration() {
+    fir::registerPassPipelineConfigCallback(
+        [](MLIRToLLVMPassPipelineConfig &config) {
+          config.registerHLFIROptEarlyEPCallbacks(
+              [](mlir::PassManager &pm, llvm::OptimizationLevel) {
+                pm.addPass(createMyHLFIRPass());
+              });
+        });
+  }
+};
+static MyPluginRegistration myPluginRegistration;
+```
+
+The augmentors are invoked for both `-emit-fir`
+(`CodeGenAction::lowerHLFIRToFIR`) and the `-emit-llvm`/`-emit-obj` path
+(`CodeGenAction::generateLLVMIR`), so a plugin only needs to register once. The
+registry is append-only and callbacks run in registration order; it is intended
+to be populated from static initializers, which run before any compilation
+begins. Tools that build the same pipelines without invoking the registry (for
+example `bbc` and `tco`) are unaffected.
+
 ## LLVM Pass Plugins
 
 Pass plugins are dynamic shared objects that consist of one or more LLVM IR
diff --git a/flang/include/flang/Optimizer/Passes/Pipelines.h 
b/flang/include/flang/Optimizer/Passes/Pipelines.h
index c50d41844941e..1ac73ba02a7cd 100644
--- a/flang/include/flang/Optimizer/Passes/Pipelines.h
+++ b/flang/include/flang/Optimizer/Passes/Pipelines.h
@@ -132,6 +132,17 @@ void addLLVMDialectToLLVMPass(mlir::PassManager &pm, 
llvm::raw_ostream &output);
 /// Use inliner extension point callback to register the default inliner pass.
 void registerDefaultInlinerPass(MLIRToLLVMPassPipelineConfig &config);
 
+/// Register a callback that augments the MLIRToLLVMPassPipelineConfig before
+/// the pass pipeline is built, so that a -load'ed plugin can register passes 
at
+/// the pipeline extension points. Callbacks run in registration order. 
Register
+/// from a static initializer, which runs before any compilation begins.
+void registerPassPipelineConfigCallback(
+    std::function<void(MLIRToLLVMPassPipelineConfig &)> callback);
+
+/// Invoke every callback registered via registerPassPipelineConfigCallback on
+/// \p config.
+void invokePassPipelineConfigCallbacks(MLIRToLLVMPassPipelineConfig &config);
+
 /// Register the passes used in Flang's MLIR pass pipeline
 /// e.g. --mlir-print-ir-before=<pass> and similar.
 void registerFlangPipelinePasses();
diff --git a/flang/lib/Frontend/FrontendActions.cpp 
b/flang/lib/Frontend/FrontendActions.cpp
index 8955a8f61e513..44cffdbfb49ab 100644
--- a/flang/lib/Frontend/FrontendActions.cpp
+++ b/flang/lib/Frontend/FrontendActions.cpp
@@ -633,6 +633,8 @@ void CodeGenAction::lowerHLFIRToFIR() {
       ci.getInvocation().getLoweringOpts().getFPMaxminBehavior();
   if (ci.getInvocation().getLangOpts().OpenMPIsTargetDevice)
     config.EnableOpenMPIsTargetDevice = true;
+  // Let -load'ed plugins augment the pipeline config.
+  fir::invokePassPipelineConfigCallbacks(config);
   // Create the pass pipeline
   fir::createHLFIRToFIRPassPipeline(pm, enableOpenMP, config);
   (void)mlir::applyPassManagerCLOptions(pm);
@@ -751,6 +753,10 @@ void CodeGenAction::generateLLVMIR() {
   config.SkipConvertComplexPow = pipelineTriple.isAMDGCN();
   fir::registerDefaultInlinerPass(config);
 
+  // Let -load'ed plugins augment the pipeline config. This path reaches
+  // createHLFIRToFIRPassPipeline through createMLIRToLLVMPassPipeline.
+  fir::invokePassPipelineConfigCallbacks(config);
+
   if (auto vsr = getVScaleRange(ci)) {
     config.VScaleMin = vsr->first;
     config.VScaleMax = vsr->second;
diff --git a/flang/lib/Optimizer/Passes/Pipelines.cpp 
b/flang/lib/Optimizer/Passes/Pipelines.cpp
index 48ae9f3bea58c..c2d317c02ed37 100644
--- a/flang/lib/Optimizer/Passes/Pipelines.cpp
+++ b/flang/lib/Optimizer/Passes/Pipelines.cpp
@@ -171,6 +171,23 @@ void 
registerDefaultInlinerPass(MLIRToLLVMPassPipelineConfig &config) {
       });
 }
 
+static std::vector<std::function<void(MLIRToLLVMPassPipelineConfig &)>> &
+getPassPipelineConfigCallbacks() {
+  static std::vector<std::function<void(MLIRToLLVMPassPipelineConfig &)>>
+      callbacks;
+  return callbacks;
+}
+
+void registerPassPipelineConfigCallback(
+    std::function<void(MLIRToLLVMPassPipelineConfig &)> callback) {
+  getPassPipelineConfigCallbacks().push_back(std::move(callback));
+}
+
+void invokePassPipelineConfigCallbacks(MLIRToLLVMPassPipelineConfig &config) {
+  for (auto &callback : getPassPipelineConfigCallbacks())
+    callback(config);
+}
+
 /// Create a pass pipeline for running default optimization passes for
 /// incremental conversion of FIR.
 ///
diff --git a/flang/unittests/Optimizer/HLFIRExtensionPointsTest.cpp 
b/flang/unittests/Optimizer/HLFIRExtensionPointsTest.cpp
index 2dc17b458ea9a..275b282d9b91a 100644
--- a/flang/unittests/Optimizer/HLFIRExtensionPointsTest.cpp
+++ b/flang/unittests/Optimizer/HLFIRExtensionPointsTest.cpp
@@ -6,7 +6,8 @@
 //
 
//===----------------------------------------------------------------------===//
 //
-// Tests for the HLFIR extension points of the HLFIR-to-FIR pass pipeline.
+// Tests for the HLFIR extension points of the HLFIR-to-FIR pass pipeline, and
+// for the config augmentor registry that -load'ed plugins use to reach them.
 //
 // The callbacks run when the pipeline is built, so no IR is needed: building
 // the pipeline is enough to observe them.
@@ -138,4 +139,107 @@ TEST(HLFIRExtensionPoint, 
MarkersAreAtTheDocumentedPositions) {
   EXPECT_LT(lastMarker, lowerIntrinsics) << pipeline;
 }
 
+// The registry is process-global and append-only, with no way to unregister, 
so
+// a callback capturing a local by reference would be re-invoked from a later
+// test with the referent destroyed. The tests record into this
+// process-lifetime recorder instead, and each asserts only on the markers it
+// wrote, so they do not depend on execution order.
+struct AugmentorRecorder {
+  std::vector<std::string> order;
+  MLIRToLLVMPassPipelineConfig *seenConfig = nullptr;
+  bool epRan = false;
+
+  void reset() {
+    order.clear();
+    seenConfig = nullptr;
+    epRan = false;
+  }
+  /// Index of \p marker in `order`, or npos.
+  size_t indexOf(llvm::StringRef marker) const {
+    for (size_t i = 0, e = order.size(); i != e; ++i)
+      if (order[i] == marker)
+        return i;
+    return std::string::npos;
+  }
+};
+
+AugmentorRecorder &recorder() {
+  static AugmentorRecorder r;
+  return r;
+}
+
+TEST(PassPipelineConfigCallback, CallbacksRunInRegistrationOrderOnTheConfig) {
+  fir::registerPassPipelineConfigCallback(
+      [](MLIRToLLVMPassPipelineConfig &config) {
+        recorder().order.push_back("order-first");
+        recorder().seenConfig = &config;
+      });
+  fir::registerPassPipelineConfigCallback([](MLIRToLLVMPassPipelineConfig &) {
+    recorder().order.push_back("order-second");
+  });
+
+  recorder().reset();
+  MLIRToLLVMPassPipelineConfig config(llvm::OptimizationLevel::O2);
+  fir::invokePassPipelineConfigCallbacks(config);
+
+  size_t first = recorder().indexOf("order-first");
+  size_t second = recorder().indexOf("order-second");
+  ASSERT_NE(first, std::string::npos);
+  ASSERT_NE(second, std::string::npos);
+  EXPECT_LT(first, second);
+  // The callback receives the config the pipeline will be built from.
+  EXPECT_EQ(recorder().seenConfig, &config);
+}
+
+// The plugin shape: the augmentor registers an extension-point callback, which
+// then contributes a pass when the pipeline is built.
+TEST(PassPipelineConfigCallback, CanRegisterHLFIRExtensionPoints) {
+  fir::registerPassPipelineConfigCallback(
+      [](MLIRToLLVMPassPipelineConfig &config) {
+        config.registerHLFIROptEarlyEPCallbacks(
+            [](mlir::PassManager &pm, llvm::OptimizationLevel) {
+              recorder().epRan = true;
+              pm.addPass(std::make_unique<MarkerPass>());
+            });
+      });
+
+  recorder().reset();
+  mlir::MLIRContext context;
+  mlir::PassManager pm(&context, mlir::ModuleOp::getOperationName());
+  MLIRToLLVMPassPipelineConfig config(llvm::OptimizationLevel::O2);
+  fir::invokePassPipelineConfigCallbacks(config);
+  fir::createHLFIRToFIRPassPipeline(pm, fir::EnableOpenMP::None, config);
+
+  std::string pipeline;
+  llvm::raw_string_ostream os(pipeline);
+  pm.printAsTextualPipeline(os);
+
+  EXPECT_TRUE(recorder().epRan);
+  EXPECT_NE(pipeline.find("ep-marker"), std::string::npos) << pipeline;
+}
+
+// A config never handed to invokePassPipelineConfigCallbacks is unaffected,
+// which is what keeps bbc and tco out of the registry.
+TEST(PassPipelineConfigCallback, NotInvokedMeansNoEffect) {
+  fir::registerPassPipelineConfigCallback(
+      [](MLIRToLLVMPassPipelineConfig &config) {
+        config.registerHLFIROptEarlyEPCallbacks(
+            [](mlir::PassManager &pm, llvm::OptimizationLevel) {
+              pm.addPass(std::make_unique<MarkerPass>());
+            });
+      });
+
+  recorder().reset();
+  mlir::MLIRContext context;
+  mlir::PassManager pm(&context, mlir::ModuleOp::getOperationName());
+  MLIRToLLVMPassPipelineConfig config(llvm::OptimizationLevel::O2);
+  fir::createHLFIRToFIRPassPipeline(pm, fir::EnableOpenMP::None, config);
+
+  std::string pipeline;
+  llvm::raw_string_ostream os(pipeline);
+  pm.printAsTextualPipeline(os);
+
+  EXPECT_EQ(pipeline.find("ep-marker"), std::string::npos) << pipeline;
+}
+
 } // namespace

_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits

Reply via email to