use compilesig_invokes=false

src/interface.jl

@@ -615,7 +615,13 @@ function inference_params(@nospecialize(job::CompilerJob))
 end
 
 # the optimization parameters to use when constructing the GPUInterpreter
-optimization_params(@nospecialize(job::CompilerJob)) = CC.OptimizationParams()
+#
+# Emit `:invoke`s targeting the exact inferred specialization instead of the
+# compileable (vararg-widened) signature: GPU back-ends cannot dispatch at run
+# time, and widened signatures may have no inferred code in our cache (e.g.
+# `Base._throw_boundserror_indices(A, i1, I...)` from bounds checks).
+optimization_params(@nospecialize(job::CompilerJob)) =
+    CC.OptimizationParams(; compilesig_invokes=false)
 
 # how much debuginfo to emit
 function llvm_debug_info(@nospecialize(job::CompilerJob))

test/gcn.jl

@@ -493,9 +493,10 @@ end
 
     @test @filecheck begin
         @check_label "define void @{{(julia|j)_kernel_[0-9]+}}"
-        # box: a jl_box_float32 call <1.14; 1.14+ inlines it into the devirt'd ctor
-        @check cond=(VERSION < v"1.14-")  "jl_box_float32"
-        @check cond=(VERSION >= v"1.14-") "gpu_gc_pool_alloc"
+        # box: a jl_box_float32 call on 1.10, which ignores compilesig_invokes;
+        # 1.11+ devirtualizes the ctor, boxing through the GC pool allocator instead
+        @check cond=(VERSION < v"1.11-")  "jl_box_float32"
+        @check cond=(VERSION >= v"1.11-") "gpu_gc_pool_alloc"
         GCN.code_llvm(mod.kernel, Tuple{Float32,Ptr{Float32}}; dump_module=true)
     end
     GCN.code_native(devnull, mod.kernel, Tuple{Float32,Ptr{Float32}})

test/native.jl

@@ -1437,8 +1437,8 @@ end
         (occursin(GPUCompiler.RUNTIME_FUNCTION, msg) ||
          occursin(GPUCompiler.UNKNOWN_FUNCTION, msg) ||
          occursin(GPUCompiler.DYNAMIC_CALL, msg)) &&
-        occursin("[1] println", msg) &&
-        occursin("[2] foobar", msg)
+        occursin(r"\[\d+\] println", msg) &&
+        occursin(r"\[\d+\] foobar", msg)
     end
 end
 
@@ -1561,26 +1561,40 @@ end
         kernel(a, b) = (unsafe_store!(b, nospecialize_child(a)); return)
     end
 
-    @test_throws_message(InvalidIRError,
-                         Native.code_execution(mod.kernel, Tuple{Int,Ptr{Int}})) do msg
-        occursin("invalid LLVM IR", msg) &&
-        occursin(GPUCompiler.DYNAMIC_CALL, msg) &&
-        occursin("call to nospecialize_child", msg) &&
-        occursin("[1] kernel", msg)
+    if VERSION >= v"1.11-"
+        # with compilesig_invokes=false, the call targets the exact specialization
+        # instead of a (possibly uncached) vararg-widened compileable signature,
+        # so no dynamic call remains
+        @test @filecheck begin
+            @check_label "define {{.*}} @{{(julia|j)_kernel_[0-9]+}}"
+            @check_not "jl_invoke"
+            @check_not "apply_generic"
+            Native.code_llvm(mod.kernel, Tuple{Int,Ptr{Int}}; dump_module=true)
+        end
+        Native.code_execution(mod.kernel, Tuple{Int,Ptr{Int}})
+    else
+        # 1.10 ignores compilesig_invokes and emits a dynamic invoke
+        @test_throws_message(InvalidIRError,
+                             Native.code_execution(mod.kernel, Tuple{Int,Ptr{Int}})) do msg
+            occursin("invalid LLVM IR", msg) &&
+            occursin(GPUCompiler.DYNAMIC_CALL, msg) &&
+            occursin("call to nospecialize_child", msg) &&
+            occursin(r"\[\d+\] kernel", msg)
+        end
     end
 end
 
 @testset "dynamic call (apply)" begin
     mod = @eval module $(gensym())
-        func() = println(1)
+        func(a) = (print(Base.inferencebarrier(a)); return)
     end
 
     @test_throws_message(InvalidIRError,
-                         Native.code_execution(mod.func, Tuple{})) do msg
+                         Native.code_execution(mod.func, Tuple{Int})) do msg
         occursin("invalid LLVM IR", msg) &&
         occursin(GPUCompiler.DYNAMIC_CALL, msg) &&
         occursin("call to print", msg) &&
-        occursin("[2] func", msg)
+        occursin(r"\[\d+\] func", msg)
     end
 end
 

test/ptx.jl

@@ -493,9 +493,10 @@ end
 
     @test @filecheck begin
         @check_label "define void @{{(julia|j)_kernel_[0-9]+}}"
-        # box: a jl_box_float32 call <1.14; 1.14+ inlines it into the devirt'd ctor
-        @check cond=(VERSION < v"1.14-")  "jl_box_float32"
-        @check cond=(VERSION >= v"1.14-") "gpu_gc_pool_alloc"
+        # box: a jl_box_float32 call on 1.10, which ignores compilesig_invokes;
+        # 1.11+ devirtualizes the ctor, boxing through the GC pool allocator instead
+        @check cond=(VERSION < v"1.11-")  "jl_box_float32"
+        @check cond=(VERSION >= v"1.11-") "gpu_gc_pool_alloc"
         PTX.code_llvm(mod.kernel, Tuple{Float32,Ptr{Float32}}; dump_module=true)
     end
     PTX.code_native(devnull, mod.kernel, Tuple{Float32,Ptr{Float32}})