llvm/test/CodeGen/AMDGPU/hsa-metadata-enqueue-kernel-v3.ll

   1 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 --amdhsa-code-object-version=3 -filetype=obj -o - < %s | llvm-readelf --notes - | FileCheck --check-prefix=CHECK %s
   2 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 --amdhsa-code-object-version=3 -amdgpu-dump-hsa-metadata -amdgpu-verify-hsa-metadata -filetype=obj -o - < %s 2>&1 | FileCheck --check-prefix=PARSER %s
   3
   4 ; CHECK:              ---
   5 ; CHECK:      amdhsa.kernels:
   6 ; CHECK:        - .args:
   7 ; CHECK-NEXT:       - .name:           a
   8 ; CHECK-NEXT:         .offset:         0
   9 ; CHECK-NEXT:         .size:           1
  10 ; CHECK-NEXT:         .type_name:      char
  11 ; CHECK-NEXT:         .value_kind:     by_value
  12 ; CHECK-NEXT:       - .offset:         8
  13 ; CHECK-NEXT:         .size:           8
  14 ; CHECK-NEXT:         .value_kind:     hidden_global_offset_x
  15 ; CHECK-NEXT:       - .offset:         16
  16 ; CHECK-NEXT:         .size:           8
  17 ; CHECK-NEXT:         .value_kind:     hidden_global_offset_y
  18 ; CHECK-NEXT:       - .offset:         24
  19 ; CHECK-NEXT:         .size:           8
  20 ; CHECK-NEXT:         .value_kind:     hidden_global_offset_z
  21 ; CHECK-NOT:        .value_kind:    hidden_default_queue
  22 ; CHECK-NOT:        .value_kind:    hidden_completion_action
  23 ; CHECK:          .language:       OpenCL C
  24 ; CHECK-NEXT:     .language_version:
  25 ; CHECK-NEXT:       - 2
  26 ; CHECK-NEXT:       - 0
  27 ; CHECK:          .name:           test_non_enqueue_kernel_caller
  28 ; CHECK:          .symbol:         test_non_enqueue_kernel_caller.kd
  29 define amdgpu_kernel void @test_non_enqueue_kernel_caller(i8 %a) #0
  30     !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !3
  31     !kernel_arg_base_type !3 !kernel_arg_type_qual !4 {
  32   ret void
  33 }
  34
  35 ; CHECK:        - .args:
  36 ; CHECK-NEXT:       - .name:           a
  37 ; CHECK-NEXT:         .offset:         0
  38 ; CHECK-NEXT:         .size:           1
  39 ; CHECK-NEXT:         .type_name:      char
  40 ; CHECK-NEXT:         .value_kind:     by_value
  41 ; CHECK-NEXT:       - .offset:         8
  42 ; CHECK-NEXT:         .size:           8
  43 ; CHECK-NEXT:         .value_kind:     hidden_global_offset_x
  44 ; CHECK-NEXT:       - .offset:         16
  45 ; CHECK-NEXT:         .size:           8
  46 ; CHECK-NEXT:         .value_kind:     hidden_global_offset_y
  47 ; CHECK-NEXT:       - .offset:         24
  48 ; CHECK-NEXT:         .size:           8
  49 ; CHECK-NEXT:         .value_kind:     hidden_global_offset_z
  50 ; CHECK-NEXT:       - .offset:         32
  51 ; CHECK-NEXT:         .size:           8
  52 ; CHECK-NEXT:         .value_kind:     hidden_hostcall_buffer
  53 ; CHECK-NEXT:       - .offset:         40
  54 ; CHECK-NEXT:         .size:           8
  55 ; CHECK-NEXT:         .value_kind:     hidden_default_queue
  56 ; CHECK-NEXT:       - .offset:         48
  57 ; CHECK-NEXT:         .size:           8
  58 ; CHECK-NEXT:         .value_kind:     hidden_completion_action
  59 ; CHECK:          .language:       OpenCL C
  60 ; CHECK-NEXT:     .language_version:
  61 ; CHECK-NEXT:       - 2
  62 ; CHECK-NEXT:       - 0
  63 ; CHECK:          .name:           test_enqueue_kernel_caller
  64 ; CHECK:          .symbol:         test_enqueue_kernel_caller.kd
  65 define amdgpu_kernel void @test_enqueue_kernel_caller(i8 %a) #1
  66     !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !3
  67     !kernel_arg_base_type !3 !kernel_arg_type_qual !4 {
  68   ret void
  69 }
  70
  71 ; CHECK:  amdhsa.version:
  72 ; CHECK-NEXT: - 1
  73 ; CHECK-NEXT: - 0
  74 ; CHECK-NOT:  amdhsa.printf:
  75
  76 attributes #0 = { optnone noinline "amdgpu-implicitarg-num-bytes"="48" }
  77 attributes #1 = { optnone noinline "calls-enqueue-kernel" "amdgpu-implicitarg-num-bytes"="48" }
  78
  79 !1 = !{i32 0}
  80 !2 = !{!"none"}
  81 !3 = !{!"char"}
  82 !4 = !{!""}
  83
  84 !opencl.ocl.version = !{!90}
  85 !90 = !{i32 2, i32 0}
  86
  87 ; PARSER: AMDGPU HSA Metadata Parser Test: PASS