Sanitize kernel names.

ext/CUDAExt.jl

@@ -1,7 +1,7 @@
 module CUDAExt
 
 using cuTile
-using cuTile: TileArray, Constant, CGOpts, CuTileResults, emit_code
+using cuTile: TileArray, Constant, CGOpts, CuTileResults, emit_code, sanitize_name
 
 using CompilerCaching: CacheView, method_instance, results
 
@@ -51,7 +51,7 @@ function emit_function(cache::CacheView, mi::Core.MethodInstance)
     res = results(cache, ci)
     res.cuda_func !== nothing && return res.cuda_func
 
-    kernel_name = string(mi.def.name)
+    kernel_name = sanitize_name(string(mi.def.name))
     cumod = CuModule(cubin)
     cufunc = CuFunction(cumod, kernel_name)
     res.cuda_func = cufunc

src/compiler/interface.jl

@@ -337,6 +337,9 @@ function emit_ir(cache::CacheView, mi::Core.MethodInstance)
     return res.julia_ir
 end
 
+# Encode characters outside [a-zA-Z0-9_] as _XX hex escapes for PTX/MLIR compatibility.
+sanitize_name(name::String) = replace(name, r"[^a-zA-Z0-9_]" => c -> "_$(string(UInt8(only(c)); base=16, pad=2))")
+
 """
     emit_code(cache, mi) -> Vector{UInt8}
 
@@ -357,7 +360,7 @@ function emit_code(cache::CacheView, mi::Core.MethodInstance)
     # Generate Tile IR bytecode
     bytecode = write_bytecode!(1) do writer, func_buf
         emit_kernel!(writer, func_buf, sci, rettype;
-            name = string(mi.def.name),
+            name = sanitize_name(string(mi.def.name)),
             sm_arch = opts.sm_arch,
             num_ctas = opts.num_ctas,
             occupancy = opts.occupancy,

src/cuTile.jl

@@ -39,6 +39,6 @@ include("language/operations.jl")
 include("language/atomics.jl")
 
 public launch
-launch() = error("Please import CUDA.jl before using `cuTile.launch`.")
+launch(args...) = error("Please import CUDA.jl before using `cuTile.launch`.")
 
 end # module cuTile

test/execution/basic.jl

@@ -1058,3 +1058,10 @@ const _EXEC_TEST_GLOBAL_CONST = Float32(1 / log(2))
     @test Array(b) ≈ Array(a) .* (scale * _EXEC_TEST_GLOBAL_CONST)
 end
 
+@testset "kernel name with !" begin
+    function kernel!()
+        return
+    end
+    ct.launch(kernel!, 1)
+end
+