llvm/test/CodeGen/AMDGPU/implicit-kernel-argument-alignment.ll

   1 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx906 < %s | FileCheck --check-prefixes=CHECK %s
   2
   3
   4 ; CHECK-LABEL: test_unaligned_to_eight:
   5 ; CHECK: .amdhsa_kernarg_size 264
   6 define amdgpu_kernel void @test_unaligned_to_eight(i32 %four)  {
   7   %implicitarg.ptr = call ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()
   8   store volatile ptr addrspace(4) %implicitarg.ptr, ptr addrspace(1) undef
   9   ret void
  10 }
  11
  12
  13 ; CHECK-LABEL: test_aligned_to_eight:
  14 ; CHECK: .amdhsa_kernarg_size 264
  15 define amdgpu_kernel void @test_aligned_to_eight(i64 %eight)  {
  16   %implicitarg.ptr = call ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()
  17   store volatile ptr addrspace(4) %implicitarg.ptr, ptr addrspace(1) undef
  18   ret void
  19 }
  20
  21 ; CHECK-LABEL: amdhsa.kernels:
  22 ; CHECK:  - .args:
  23 ; CHECK-NEXT:      - .name:           four
  24 ; CHECK-NEXT:         .offset:         0
  25 ; CHECK-NEXT:         .size:           4
  26 ; CHECK-NEXT:         .value_kind:     by_value
  27 ; CHECK-NEXT:       - .offset:         8
  28 ; CHECK-NEXT:         .size:           4
  29 ; CHECK-NEXT:         .value_kind:     hidden_block_count_x
  30 ; CHECK-NEXT:       - .offset:         12
  31 ; CHECK-NEXT:         .size:           4
  32 ; CHECK-NEXT:         .value_kind:     hidden_block_count_y
  33 ; CHECK-NEXT:       - .offset:         16
  34 ; CHECK-NEXT:         .size:           4
  35 ; CHECK-NEXT:         .value_kind:     hidden_block_count_z
  36 ; CHECK:              .kernarg_segment_align: 8
  37 ; CHECK-NEXT:         .kernarg_segment_size: 264
  38 ; CHECK-LABEL:        .name:           test_unaligned_to_eight
  39
  40 ; CHECK:  - .args:
  41 ; CHECK-NEXT:      - .name:           eight
  42 ; CHECK-NEXT:         .offset:         0
  43 ; CHECK-NEXT:         .size:           8
  44 ; CHECK-NEXT:         .value_kind:     by_value
  45 ; CHECK-NEXT:       - .offset:         8
  46 ; CHECK-NEXT:         .size:           4
  47 ; CHECK-NEXT:         .value_kind:     hidden_block_count_x
  48 ; CHECK-NEXT:       - .offset:         12
  49 ; CHECK-NEXT:         .size:           4
  50 ; CHECK-NEXT:         .value_kind:     hidden_block_count_y
  51 ; CHECK-NEXT:       - .offset:         16
  52 ; CHECK-NEXT:         .size:           4
  53 ; CHECK-NEXT:         .value_kind:     hidden_block_count_z
  54 ; CHECK:              .kernarg_segment_align: 8
  55 ; CHECK-NEXT:         .kernarg_segment_size: 264
  56 ; CHECK-LABEL:        .name:           test_aligned_to_eight
  57
  58 declare ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()
  59
  60 !llvm.module.flags = !{!0}
  61 !0 = !{i32 1, !"amdgpu_code_object_version", i32 500}