[flang-commits] [flang] [flang][cuda] Pass assumed-shape arrays of bind(c) attributes(global) kernels by base address (PR #199313)

[Valentin Clement バレンタイン クレメン via flang-commits]

================
@@ -0,0 +1,51 @@
+<!--===- docs/CUDA.md
+
+   Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+   See https://llvm.org/LICENSE.txt for license information.
+   SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+-->
+
+# CUDA Fortran lowering notes
+
+```{contents}
+---
+local:
+---
+```
+
+List of CUDA Fortran lowering decisions in Flang that diverge from the
+Fortran 2018 standard, for cases the [CUDA Fortran Programming
+Guide](https://docs.nvidia.com/hpc-sdk/compilers/cuda-fortran-prog-guide/index.html)
+does not specify.
+
+## `BIND(C) ATTRIBUTES(GLOBAL)` assumed-shape and assumed-rank dummies
+
+For a `BIND(C)` procedure with `ATTRIBUTES(GLOBAL)` or
+`ATTRIBUTES(GRID_GLOBAL)`, an assumed-shape (`dimension(:)`) or assumed-rank
+(`dimension(..)`) dummy is passed by base address (`!fir.ref<T>`) instead of by
+`CFI_cdesc_t *` (`!fir.box<T>`). `ALLOCATABLE` and `POINTER` dummies take an
+earlier descriptor-of-mutable path and are unaffected. To deliver a CFI
+descriptor to the kernel, drop `BIND(C)`: a plain `ATTRIBUTES(GLOBAL)` kernel
+keeps the descriptor-passing lowering.
----------------
clementval wrote:

drop BIND(C) might be confusing because if you just remove the kernel then cfi is passed. 

https://github.com/llvm/llvm-project/pull/199313