[flang-commits] [flang] [flang][cuda] Support predefined conversion in inlined function (PR #185723)

[via flang-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-flang-fir-hlfir

Author: Valentin Clement (バレンタイン クレメン) (clementval)

<details>
<summary>Changes</summary>

Only fir.declare at top level were converted. Update the pass to loop through all fir.declare operations. 

---
Full diff: https://github.com/llvm/llvm-project/pull/185723.diff


2 Files Affected:

- (modified) flang/lib/Optimizer/Transforms/CUDA/CUFPredefinedVarToGPU.cpp (+2-2) 
- (modified) flang/test/Fir/CUDA/predefined-variables.mlir (+48) 


``````````diff
diff --git a/flang/lib/Optimizer/Transforms/CUDA/CUFPredefinedVarToGPU.cpp b/flang/lib/Optimizer/Transforms/CUDA/CUFPredefinedVarToGPU.cpp
index 3eb655980a391..ce97b56fe3497 100644
--- a/flang/lib/Optimizer/Transforms/CUDA/CUFPredefinedVarToGPU.cpp
+++ b/flang/lib/Optimizer/Transforms/CUDA/CUFPredefinedVarToGPU.cpp
@@ -132,7 +132,7 @@ struct CUFPredefinedVarToGPU
                                                         blockdims);
 
         llvm::SmallVector<mlir::Operation *> opsToDelete;
-        for (auto declareOp : funcOp.getOps<fir::DeclareOp>()) {
+        funcOp.walk([&](fir::DeclareOp declareOp) {
           processDeclareOp(builder, loc, declareOp, mangleBuiltin(threadidx),
                            threadids, opsToDelete);
           processDeclareOp(builder, loc, declareOp, mangleBuiltin(blockidx),
@@ -141,7 +141,7 @@ struct CUFPredefinedVarToGPU
                            blockdims, opsToDelete);
           processDeclareOp(builder, loc, declareOp, mangleBuiltin(griddim),
                            griddims, opsToDelete);
-        }
+        });
 
         for (auto op : opsToDelete)
           op->erase();
diff --git a/flang/test/Fir/CUDA/predefined-variables.mlir b/flang/test/Fir/CUDA/predefined-variables.mlir
index 1a9a46d213be4..6c9d080b00ee8 100644
--- a/flang/test/Fir/CUDA/predefined-variables.mlir
+++ b/flang/test/Fir/CUDA/predefined-variables.mlir
@@ -186,3 +186,51 @@ func.func @_QPsub1(%arg0: !fir.ref<i32> {fir.bindc_name = "i", cuf.data_attr = #
 // CHECK: %{{.*}} = arith.addi %[[BASE_BLOCK_ID_X]], %c1{{.*}} : i32
 // CHECK: %{{.*}} = nvvm.read.ptx.sreg.nctaid.y : i32
 // CHECK: %{{.*}} = nvvm.read.ptx.sreg.ntid.x : i32
+
+// -----
+
+func.func @_QMbarPgfoo(%arg0: !fir.ref<i32> {cuf.data_attr = #cuf.cuda<device>, fir.bindc_name = "a"}) attributes {cuf.proc_attr = #cuf.cuda_proc<global>, no_inline} {
+  %c100_i32 = arith.constant 100 : i32
+  %cond = arith.cmpi sle, %c100_i32, %c100_i32 : i32
+  fir.if %cond {
+    %0 = fir.address_of(@_QM__fortran_builtinsE__builtin_threadidx) : !fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>
+    %1 = fir.declare %0 {uniq_name = "_QM__fortran_builtinsE__builtin_threadidx"} : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>
+    %2 = fir.coordinate_of %1, x : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<i32>
+    %3 = fir.load %2 : !fir.ref<i32>
+    fir.store %3 to %arg0 : !fir.ref<i32>
+  }
+  return
+}
+
+// CHECK-LABEL: func.func @_QMbarPgfoo
+// CHECK: %[[THREAD_ID_X:.*]] = nvvm.read.ptx.sreg.tid.x : i32
+// CHECK: %[[ADD:.*]] = arith.addi %[[THREAD_ID_X]], %c1_i32 : i32
+// CHECK: fir.if
+// CHECK: fir.store %[[ADD]] to %{{.*}} : !fir.ref<i32>
+
+// -----
+
+func.func @_QMbarPgfoo2(%arg0: !fir.ref<i32> {cuf.data_attr = #cuf.cuda<device>, fir.bindc_name = "a"}, %arg1: !fir.ref<i32> {cuf.data_attr = #cuf.cuda<device>, fir.bindc_name = "b"}) attributes {cuf.proc_attr = #cuf.cuda_proc<global>, no_inline} {
+  %c100_i32 = arith.constant 100 : i32
+  %0 = fir.address_of(@_QM__fortran_builtinsE__builtin_threadidx) : !fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>
+  %1 = fir.declare %0 {uniq_name = "_QM__fortran_builtinsE__builtin_threadidx"} : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>
+  %2 = fir.coordinate_of %1, x : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<i32>
+  %3 = fir.load %2 : !fir.ref<i32>
+  fir.store %3 to %arg0 : !fir.ref<i32>
+  %cond = arith.cmpi sle, %c100_i32, %c100_i32 : i32
+  fir.if %cond {
+    %4 = fir.address_of(@_QM__fortran_builtinsE__builtin_threadidx) : !fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>
+    %5 = fir.declare %4 {uniq_name = "_QM__fortran_builtinsE__builtin_threadidx"} : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>
+    %6 = fir.coordinate_of %5, x : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<i32>
+    %7 = fir.load %6 : !fir.ref<i32>
+    fir.store %7 to %arg1 : !fir.ref<i32>
+  }
+  return
+}
+
+// CHECK-LABEL: func.func @_QMbarPgfoo2
+// CHECK: %[[THREAD_ID_X:.*]] = nvvm.read.ptx.sreg.tid.x : i32
+// CHECK: %[[ADD:.*]] = arith.addi %[[THREAD_ID_X]], %c1_i32 : i32
+// CHECK: fir.store %[[ADD]] to %{{.*}} : !fir.ref<i32>
+// CHECK: fir.if
+// CHECK: fir.store %[[ADD]] to %{{.*}} : !fir.ref<i32>

``````````

</details>


https://github.com/llvm/llvm-project/pull/185723