[flang-commits] [flang] 5406ca3 - [flang][cuda] Extend CUF kernel host stub line table to the END statement (#219608)

[via flang-commits]

Author: jiel-nv
Date: 2026-08-31T09:50:37-07:00
New Revision: 5406ca380af4730c07dad3f98eadbc6ecd9f8bea

URL: https://github.com/llvm/llvm-project/commit/5406ca380af4730c07dad3f98eadbc6ecd9f8bea
DIFF: https://github.com/llvm/llvm-project/commit/5406ca380af4730c07dad3f98eadbc6ecd9f8bea.diff

LOG: [flang][cuda] Extend CUF kernel host stub line table to the END statement (#219608)

`CUFDeviceFuncTransform` creates a host stub for a kernel, with an empty
body holding nothing but a `return`. Both the stub and its `return` keep
only the original declaration location, so the stub's line table ends up
with rows only for the declaration line.

This change extends the line table to span the procedure's whole source
line range, by using the location of the original `END` statement for
the stub's `return`. A debugger can then set a breakpoint on a line
inside the kernel body and stop there.

Lit test `cuda-device-func-transform.mlir` is enhanced to guard this
change.

Added: 
    

Modified: 
    flang/lib/Optimizer/Transforms/CUDA/CUFDeviceFuncTransform.cpp
    flang/test/Fir/CUDA/cuda-device-func-transform.mlir

Removed: 
    


################################################################################
diff  --git a/flang/lib/Optimizer/Transforms/CUDA/CUFDeviceFuncTransform.cpp b/flang/lib/Optimizer/Transforms/CUDA/CUFDeviceFuncTransform.cpp
index 307d41cb62a21..a0303a181775e 100644
--- a/flang/lib/Optimizer/Transforms/CUDA/CUFDeviceFuncTransform.cpp
+++ b/flang/lib/Optimizer/Transforms/CUDA/CUFDeviceFuncTransform.cpp
@@ -162,6 +162,11 @@ class CUFDeviceFuncTransform
   static void createHostStub(mlir::func::FuncOp funcOp,
                              mlir::SymbolTable &symTab, mlir::ModuleOp mod) {
     mlir::Location loc = funcOp.getLoc();
+    // Host stub's line table needs to span the procedure body.
+    mlir::Location endLoc = loc;
+    if (!funcOp.getBody().empty())
+      if (mlir::Operation *terminator = funcOp.getBody().back().getTerminator())
+        endLoc = terminator->getLoc();
     mlir::OpBuilder modBuilder(mod.getBodyRegion());
     modBuilder.setInsertionPointToEnd(mod.getBody());
     auto emptyStub = func::FuncOp::create(modBuilder, loc, funcOp.getName(),
@@ -170,7 +175,9 @@ class CUFDeviceFuncTransform
     emptyStub->setAttrs(funcOp->getAttrs());
     auto entryBlock = emptyStub.addEntryBlock();
     modBuilder.setInsertionPointToEnd(entryBlock);
-    func::ReturnOp::create(modBuilder, loc);
+    // Add a return operation at the end of the stub with the location of the
+    // original procedure's terminator.
+    func::ReturnOp::create(modBuilder, endLoc);
 
     symTab.erase(funcOp);
     symTab.insert(emptyStub);

diff  --git a/flang/test/Fir/CUDA/cuda-device-func-transform.mlir b/flang/test/Fir/CUDA/cuda-device-func-transform.mlir
index 427381732630d..ab17f2c705d0c 100644
--- a/flang/test/Fir/CUDA/cuda-device-func-transform.mlir
+++ b/flang/test/Fir/CUDA/cuda-device-func-transform.mlir
@@ -1,4 +1,6 @@
 // RUN: fir-opt --split-input-file --cuf-transform-device-func %s | FileCheck %s
+// RUN: fir-opt --split-input-file --cuf-transform-device-func --mlir-print-debuginfo \
+// RUN:   --mlir-print-local-scope %s | FileCheck %s --check-prefix=LOC
 
 func.func @_QPsub_device1() attributes {cuf.proc_attr = #cuf.cuda_proc<device>} {
   return
@@ -221,3 +223,17 @@ func.func @cuda_global() attributes {cuf.proc_attr = #cuf.cuda_proc<global>} {
 // CHECK-NOT: gpu.func @acc_routine()
 // CHECK: gpu.func @cuda_global() kernel
 // CHECK: fir.call @acc_routine() : () -> ()
+
+// -----
+
+func.func @_QPsub_global_endline() attributes {cuf.proc_attr = #cuf.cuda_proc<global>} {
+  %cst = arith.constant 2.000000e+00 : f32 loc(#loc_body)
+  return loc(#loc_end)
+} loc(#loc_decl)
+#loc_decl = loc("test.cuf":4:3)
+#loc_body = loc("test.cuf":10:5)
+#loc_end = loc("test.cuf":15:3)
+
+// LOC-LABEL: func.func @_QPsub_global_endline()
+// LOC: return loc("test.cuf":15:3)
+// LOC-NEXT: } loc("test.cuf":4:3)