1414#[ unsafe( no_mangle) ]
1515fn main ( ) {
1616 let mut x = [ 3.0 ; 256 ] ;
17- kernel_1 ( & mut x) ;
17+ let y = [ 1.0 ; 256 ] ;
18+ kernel_1 ( & mut x, & y) ;
1819 core:: hint:: black_box ( & x) ;
20+ core:: hint:: black_box ( & y) ;
1921}
2022
21- pub fn kernel_1 ( x : & mut [ f32 ; 256 ] ) {
22- core:: intrinsics:: offload ( kernel_1 , [ 256 , 1 , 1 ] , [ 32 , 1 , 1 ] , ( x, ) )
23+ pub fn kernel_1 ( x : & mut [ f32 ; 256 ] , y : & [ f32 ; 256 ] ) {
24+ core:: intrinsics:: offload ( _kernel_1 , [ 256 , 1 , 1 ] , [ 32 , 1 , 1 ] , ( x, y ) )
2325}
2426
25- #[ unsafe( no_mangle) ]
2627#[ inline( never) ]
27- pub fn _kernel_1 ( x : & mut [ f32 ; 256 ] ) {
28+ pub fn _kernel_1 ( x : & mut [ f32 ; 256 ] , y : & [ f32 ; 256 ] ) {
2829 for i in 0 ..256 {
29- x[ i] = 21.0 ;
30+ x[ i] = 21.0 + y [ i ] ;
3031 }
3132}
3233
@@ -39,8 +40,10 @@ pub fn _kernel_1(x: &mut [f32; 256]) {
3940
4041// CHECK-DAG: @.omp_offloading.descriptor = internal constant { i32, ptr, ptr, ptr } zeroinitializer
4142// CHECK-DAG: @llvm.global_ctors = appending constant [1 x { i32, ptr, ptr }] [{ i32, ptr, ptr } { i32 101, ptr @.omp_offloading.descriptor_reg, ptr null }]
42- // CHECK-DAG: @.offload_sizes.[[K:[^ ]*kernel_1]] = private unnamed_addr constant [1 x i64] [i64 1024]
43- // CHECK-DAG: @.offload_maptypes.[[K]] = private unnamed_addr constant [1 x i64] [i64 35]
43+ // CHECK-DAG: @.offload_sizes.[[K:[^ ]*kernel_1]] = private unnamed_addr constant [2 x i64] [i64 1024, i64 1024]
44+ // CHECK-DAG: @.offload_maptypes.[[K]].begin = private unnamed_addr constant [2 x i64] [i64 1, i64 1]
45+ // CHECK-DAG: @.offload_maptypes.[[K]].kernel = private unnamed_addr constant [2 x i64] [i64 32, i64 32]
46+ // CHECK-DAG: @.offload_maptypes.[[K]].end = private unnamed_addr constant [2 x i64] [i64 2, i64 0]
4447// CHECK-DAG: @.[[K]].region_id = internal constant i8 0
4548// CHECK-DAG: @.offloading.entry_name.[[K]] = internal unnamed_addr constant [{{[0-9]+}} x i8] c"[[K]]{{\\00}}", section ".llvm.rodata.offloading", align 1
4649// CHECK-DAG: @.offloading.entry.[[K]] = internal constant %struct.__tgt_offload_entry { i64 0, i16 1, i16 1, i32 0, ptr @.[[K]].region_id, ptr @.offloading.entry_name.[[K]], i64 0, i64 0, ptr null }, section "llvm_offload_entries", align 8
@@ -49,28 +52,36 @@ pub fn _kernel_1(x: &mut [f32; 256]) {
4952
5053// CHECK-LABEL: define{{( dso_local)?}} void @main()
5154// CHECK-NEXT: start:
52- // CHECK-NEXT: %0 = alloca [8 x i8], align 8
53- // CHECK-NEXT: %x = alloca [1024 x i8], align 16
54- // CHECK-NEXT: %.offload_baseptrs = alloca [1 x ptr], align 8
55- // CHECK-NEXT: %.offload_ptrs = alloca [1 x ptr], align 8
56- // CHECK-NEXT: %.offload_sizes = alloca [1 x i64], align 8
55+ // CHECK-NEXT: %0 = alloca [8 x i8], align 8
56+ // CHECK-NEXT: %1 = alloca [8 x i8], align 8
57+ // CHECK-NEXT: %y = alloca [1024 x i8], align 16
58+ // CHECK-NEXT: %x = alloca [1024 x i8], align 16
59+ // CHECK-NEXT: %.offload_baseptrs = alloca [2 x ptr], align 8
60+ // CHECK-NEXT: %.offload_ptrs = alloca [2 x ptr], align 8
61+ // CHECK-NEXT: %.offload_sizes = alloca [2 x i64], align 8
5762// CHECK-NEXT: %kernel_args = alloca %struct.__tgt_kernel_arguments, align 8
5863// CHECK: call void @__tgt_init_all_rtls()
5964// CHECK-NEXT: store ptr %x, ptr %.offload_baseptrs, align 8
6065// CHECK-NEXT: store ptr %x, ptr %.offload_ptrs, align 8
6166// CHECK-NEXT: store i64 1024, ptr %.offload_sizes, align 8
62- // CHECK-NEXT: call void @__tgt_target_data_begin_mapper(ptr nonnull @anon.{{.*}}.1, i64 -1, i32 1, ptr nonnull %.offload_baseptrs, ptr nonnull %.offload_ptrs, ptr nonnull %.offload_sizes, ptr nonnull @.offload_maptypes.[[K]], ptr null, ptr null)
67+ // CHECK-NEXT: [[BPTRS_1:%.*]] = getelementptr inbounds nuw i8, ptr %.offload_baseptrs, i64 8
68+ // CHECK-NEXT: store ptr %y, ptr [[BPTRS_1]], align 8
69+ // CHECK-NEXT: [[PTRS_1:%.*]] = getelementptr inbounds nuw i8, ptr %.offload_ptrs, i64 8
70+ // CHECK-NEXT: store ptr %y, ptr [[PTRS_1]], align 8
71+ // CHECK-NEXT: [[SIZES_1:%.*]] = getelementptr inbounds nuw i8, ptr %.offload_sizes, i64 8
72+ // CHECK-NEXT: store i64 1024, ptr [[SIZES_1]], align 8
73+ // CHECK-NEXT: call void @__tgt_target_data_begin_mapper(ptr nonnull @anon.{{.*}}.1, i64 -1, i32 2, ptr nonnull %.offload_baseptrs, ptr nonnull %.offload_ptrs, ptr nonnull %.offload_sizes, ptr nonnull @.offload_maptypes.[[K]].begin, ptr null, ptr null)
6374// CHECK-NEXT: store i32 3, ptr %kernel_args, align 8
6475// CHECK-NEXT: [[P4:%[^ ]+]] = getelementptr inbounds nuw i8, ptr %kernel_args, i64 4
65- // CHECK-NEXT: store i32 1 , ptr [[P4]], align 4
76+ // CHECK-NEXT: store i32 2 , ptr [[P4]], align 4
6677// CHECK-NEXT: [[P8:%[^ ]+]] = getelementptr inbounds nuw i8, ptr %kernel_args, i64 8
6778// CHECK-NEXT: store ptr %.offload_baseptrs, ptr [[P8]], align 8
6879// CHECK-NEXT: [[P16:%[^ ]+]] = getelementptr inbounds nuw i8, ptr %kernel_args, i64 16
6980// CHECK-NEXT: store ptr %.offload_ptrs, ptr [[P16]], align 8
7081// CHECK-NEXT: [[P24:%[^ ]+]] = getelementptr inbounds nuw i8, ptr %kernel_args, i64 24
7182// CHECK-NEXT: store ptr %.offload_sizes, ptr [[P24]], align 8
7283// CHECK-NEXT: [[P32:%[^ ]+]] = getelementptr inbounds nuw i8, ptr %kernel_args, i64 32
73- // CHECK-NEXT: store ptr @.offload_maptypes.[[K]], ptr [[P32]], align 8
84+ // CHECK-NEXT: store ptr @.offload_maptypes.[[K]].kernel , ptr [[P32]], align 8
7485// CHECK-NEXT: [[P40:%[^ ]+]] = getelementptr inbounds nuw i8, ptr %kernel_args, i64 40
7586// CHECK-NEXT: [[P72:%[^ ]+]] = getelementptr inbounds nuw i8, ptr %kernel_args, i64 72
7687// CHECK-NEXT: call void @llvm.memset.p0.i64(ptr noundef nonnull align 8 dereferenceable(32) [[P40]], i8 0, i64 32, i1 false)
@@ -81,9 +92,9 @@ pub fn _kernel_1(x: &mut [f32; 256]) {
8192// CHECK-NEXT: store i32 1, ptr [[P92]], align 4
8293// CHECK-NEXT: [[P96:%[^ ]+]] = getelementptr inbounds nuw i8, ptr %kernel_args, i64 96
8394// CHECK-NEXT: store i32 0, ptr [[P96]], align 8
84- // CHECK-NEXT: {{%[^ ]+}} = call i32 @__tgt_target_kernel(ptr nonnull @anon.{{.*}}.1, i64 -1, i32 256, i32 32, ptr nonnull @.[[K]].region_id, ptr nonnull %kernel_args)
85- // CHECK-NEXT: call void @__tgt_target_data_end_mapper(ptr nonnull @anon.{{.*}}.1, i64 -1, i32 1 , ptr nonnull %.offload_baseptrs, ptr nonnull %.offload_ptrs, ptr nonnull %.offload_sizes, ptr nonnull @.offload_maptypes.[[K]], ptr null, ptr null)
86- // CHECK: ret void
95+ // CHECK-NEXT: [[TGT_RET:%.*]] = call i32 @__tgt_target_kernel(ptr nonnull @anon.{{.*}}.1, i64 -1, i32 256, i32 32, ptr nonnull @.[[K]].region_id, ptr nonnull %kernel_args)
96+ // CHECK-NEXT: call void @__tgt_target_data_end_mapper(ptr nonnull @anon.{{.*}}.1, i64 -1, i32 2 , ptr nonnull %.offload_baseptrs, ptr nonnull %.offload_ptrs, ptr nonnull %.offload_sizes, ptr nonnull @.offload_maptypes.[[K]].end , ptr null, ptr null)
97+ // CHECK: ret void
8798// CHECK-NEXT: }
8899
89100// CHECK: declare void @__tgt_register_lib(ptr) local_unnamed_addr
0 commit comments