llvm/test/CodeGen/AMDGPU/GlobalISel/flat-scratch-init.ll

   1 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,RW-FLAT %s
   2 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=+architected-flat-scratch < %s | FileCheck -check-prefixes=GCN,RO-FLAT %s
   3
   4 ; Make sure flat_scratch_init is set
   5
   6 ; GCN-LABEL: {{^}}stack_object_addrspacecast_in_kernel_no_calls:
   7 ; RW-FLAT:     s_add_u32 flat_scratch_lo, s4, s7
   8 ; RW-FLAT:     s_addc_u32 flat_scratch_hi, s5, 0
   9 ; RO-FLAT-NOT: flat_scratch
  10 ; GCN:         flat_store_dword
  11 ; RO-FLAT-NOT: .amdhsa_user_sgpr_private_segment_buffer
  12 ; RW-FLAT:     .amdhsa_user_sgpr_flat_scratch_init 1
  13 ; RO-FLAT-NOT: .amdhsa_user_sgpr_flat_scratch_init
  14 ; RW-FLAT:     .amdhsa_system_sgpr_private_segment_wavefront_offset
  15 ; RW-FLAT-NOT: .amdhsa_enable_private_segment
  16 ; RO-FLAT-NOT: .amdhsa_system_sgpr_private_segment_wavefront_offset
  17 ; RO-FLAT:     .amdhsa_enable_private_segment 1
  18 ; GCN-NOT:     .amdhsa_reserve_flat_scratch
  19 ; GCN:         COMPUTE_PGM_RSRC2:SCRATCH_EN: 1
  20 ; RW-FLAT:     COMPUTE_PGM_RSRC2:USER_SGPR: 6
  21 ; RO-FLAT:     COMPUTE_PGM_RSRC2:USER_SGPR: 0
  22 define amdgpu_kernel void @stack_object_addrspacecast_in_kernel_no_calls() {
  23   %alloca = alloca i32, addrspace(5)
  24   %cast = addrspacecast i32 addrspace(5)* %alloca to i32*
  25   store volatile i32 0, i32* %cast
  26   ret void
  27 }
  28
  29 ; TODO: Could optimize out in this case
  30 ; GCN-LABEL: {{^}}stack_object_in_kernel_no_calls:
  31 ; RO-FLAT-NOT: flat_scratch
  32 ; RW-FLAT:     buffer_store_dword
  33 ; RO-FLAT:     scratch_store_dword
  34 ; RW-FLAT:     .amdhsa_user_sgpr_private_segment_buffer 1
  35 ; RO-FLAT-NOT: .amdhsa_user_sgpr_private_segment_buffer
  36 ; RW-FLAT:     .amdhsa_user_sgpr_flat_scratch_init 1
  37 ; RO-FLAT-NOT: .amdhsa_user_sgpr_flat_scratch_init
  38 ; RW-FLAT:     .amdhsa_system_sgpr_private_segment_wavefront_offset 1
  39 ; RW-FLAT-NOT: .amdhsa_enable_private_segment
  40 ; RO-FLAT-NOT: .amdhsa_system_sgpr_private_segment_wavefront_offset
  41 ; RO-FLAT:     .amdhsa_enable_private_segment 1
  42 ; RW-FLAT:     .amdhsa_reserve_flat_scratch 0
  43 ; RO-FLAT-NOT: .amdhsa_reserve_flat_scratch
  44 ; GCN:         COMPUTE_PGM_RSRC2:SCRATCH_EN: 1
  45 ; RW-FLAT:     COMPUTE_PGM_RSRC2:USER_SGPR: 6
  46 ; RO-FLAT:     COMPUTE_PGM_RSRC2:USER_SGPR: 0
  47 define amdgpu_kernel void @stack_object_in_kernel_no_calls() {
  48   %alloca = alloca i32, addrspace(5)
  49   store volatile i32 0, i32 addrspace(5)* %alloca
  50   ret void
  51 }
  52
  53 ; GCN-LABEL: {{^}}kernel_no_calls_no_stack:
  54 ; GCN-NOT:    flat_scratch
  55 ; RW-FLAT:     .amdhsa_user_sgpr_private_segment_buffer 1
  56 ; RO-FLAT-NOT: .amdhsa_user_sgpr_private_segment_buffer
  57 ; RW-FLAT:     .amdhsa_user_sgpr_flat_scratch_init 0
  58 ; RO-FLAT-NOT: .amdhsa_user_sgpr_flat_scratch_init
  59 ; RW-FLAT:     .amdhsa_system_sgpr_private_segment_wavefront_offset 0
  60 ; RW-FLAT-NOT: .amdhsa_enable_private_segment
  61 ; RO-FLAT-NOT: .amdhsa_system_sgpr_private_segment_wavefront_offset
  62 ; RO-FLAT:     .amdhsa_enable_private_segment 0
  63 ; RW-FLAT:     .amdhsa_reserve_flat_scratch 0
  64 ; RO-FLAT-NOT: .amdhsa_reserve_flat_scratch 0
  65 ; GCN:         COMPUTE_PGM_RSRC2:SCRATCH_EN: 0
  66 ; RW-FLAT:     COMPUTE_PGM_RSRC2:USER_SGPR: 4
  67 ; RO-FLAT:     COMPUTE_PGM_RSRC2:USER_SGPR: 0
  68 define amdgpu_kernel void @kernel_no_calls_no_stack() {
  69   ret void
  70 }