[flang-commits] [flang] [mlir] [mlir][GPU] Add NVVM-specific `cf.assert` lowering (PR #120431)

[Matthias Springer via flang-commits]

================
@@ -0,0 +1,37 @@
+// RUN: mlir-opt %s -gpu-lower-to-nvvm-pipeline="cubin-format=%gpu_compilation_format" \
+// RUN: | mlir-cpu-runner \
+// RUN:   --shared-libs=%mlir_cuda_runtime \
+// RUN:   --shared-libs=%mlir_runner_utils \
+// RUN:   --entry-point-result=void 2>&1 \
+// RUN: | FileCheck %s
+
+// CHECK-DAG: thread 0: print after passing assertion
+// CHECK-DAG: thread 1: print after passing assertion
+// CHECK-DAG: mlir/test/Integration/GPU/CUDA/assert.mlir:{{.*}}: (unknown): block: [0,0,0], thread: [0,0,0] Assertion `failing assertion` failed.
+// CHECK-DAG: mlir/test/Integration/GPU/CUDA/assert.mlir:{{.*}}: (unknown): block: [0,0,0], thread: [1,0,0] Assertion `failing assertion` failed.
+// CHECK-NOT: print after failing assertion
+
+module attributes {gpu.container_module} {
+gpu.module @kernels {
+gpu.func @test_assert(%c0: i1, %c1: i1) kernel {
+  %0 = gpu.thread_id x
+  cf.assert %c1, "passing assertion"
+  gpu.printf "thread %lld: print after passing assertion\n" %0 : index
+  cf.assert %c0, "failing assertion"
----------------
matthias-springer wrote:

I'm not sure what `__asserfail` does internally, but it works even if it's inside of an `scf.if` like in your example. There's only one of these:
```
llvm-project/mlir/test/Integration/GPU/CUDA/assert.mlir:24: (unknown): block: [0,0,0], thread: [0,0,0] Assertion `failing assertion` failed.
```
But the kernel stops executing right away.


https://github.com/llvm/llvm-project/pull/120431