[flang][tco] Add -emit-final-mlir flag

[mrkajetanp]

Add a flag to tco for emitting the final MLIR, prior to lowering to LLVM IR. This is intended to produce output that can be passed directly to mlir-translate.

[mrkajetanp]

I can also rename to just -emit-mlir for brevity, but I wasn't sure whether that'd be confusing or not on account of it all technically being mlir.

[tblah]

Great to see this, I'm sure I won't be the only one who can save time not having to dump mlir after every pass. I think the name is good.

Please could you add a test that uses this flag and checks that what it gets is LLVM dialect MLIR.

I'll add some other reviewers who often work on LLVMIR codegen as they may also find this useful.

[llvmbot]

@llvm/pr-subscribers-flang-driver

Author: Kajetan Puchalski (mrkajetanp)

Changes

Add a flag to tco for emitting the final MLIR, prior to lowering to LLVM IR. This is intended to produce output that can be passed directly to mlir-translate.

---

Full diff: https://github.com/llvm/llvm-project/pull/146533.diff

2 Files Affected:

• (added) flang/test/Driver/tco-emit-final-mlir.fir (+19)
• (modified) flang/tools/tco/tco.cpp (+9-2)

diff --git a/flang/test/Driver/tco-emit-final-mlir.fir b/flang/test/Driver/tco-emit-final-mlir.fir
new file mode 100644
index 0000000000000..75f8f153127af
--- /dev/null
+++ b/flang/test/Driver/tco-emit-final-mlir.fir
@@ -0,0 +1,19 @@
+// Test tco's -emit-final-mlir option.
+
+// RUN: tco -emit-final-mlir %s 2>&1 | FileCheck %s
+
+// Check that the FIR file is translated into the LLVM IR
+// MLIR dialect, but not into LLVM IR itself.
+
+// CHECK-NOT: func.func
+// CHECK-LABEL: llvm.func @_QPfoo()
+// CHECK-NOT: fir.alloca
+// CHECK: %[[VAL_0:.*]] = llvm.mlir.constant(1 : i64) : i64
+// CHECK: %[[VAL_1:.*]] = llvm.alloca %[[VAL_0]] x i32 : (i64) -> !llvm.ptr
+// CHECK: llvm.return
+// CHECK-NOT: func.func
+
+func.func @_QPfoo() {
+  %1 = fir.alloca i32
+  return
+}
diff --git a/flang/tools/tco/tco.cpp b/flang/tools/tco/tco.cpp
index eaf4bae088454..806cd3148d09f 100644
--- a/flang/tools/tco/tco.cpp
+++ b/flang/tools/tco/tco.cpp
@@ -70,6 +70,11 @@ static cl::opt<bool> codeGenLLVM(
     cl::desc("Run only CodeGen passes and translate FIR to LLVM IR"),
     cl::init(false));
 
+static cl::opt<bool> emitFinalMLIR(
+    "emit-final-mlir",
+    cl::desc("Only translate FIR to MLIR, do not lower to LLVM IR"),
+    cl::init(false));
+
 #include "flang/Optimizer/Passes/CommandLineOpts.h"
 #include "flang/Optimizer/Passes/Pipelines.h"
 
@@ -149,13 +154,15 @@ compileFIR(const mlir::PassPipelineCLParser &passPipeline) {
       fir::registerDefaultInlinerPass(config);
       fir::createMLIRToLLVMPassPipeline(pm, config);
     }
-    fir::addLLVMDialectToLLVMPass(pm, out.os());
+    if (!emitFinalMLIR)
+      fir::addLLVMDialectToLLVMPass(pm, out.os());
   }
 
   // run the pass manager
   if (mlir::succeeded(pm.run(*owningRef))) {
     // passes ran successfully, so keep the output
-    if ((emitFir || passPipeline.hasAnyOccurrences()) && !codeGenLLVM)
+    if ((emitFir || passPipeline.hasAnyOccurrences() || emitFinalMLIR) &&
+        !codeGenLLVM)
       printModule(*owningRef, out.os());
     out.keep();
     return mlir::success();

[tblah]

Thanks

[skatrak]

Thank you, LGTM!

[kaviya2510]

Thanks for the patch, LGTM!