llvm/test/CodeGen/AMDGPU/insert-delay-alu-wmma.mir

# RUN: llc -mtriple=amdgpu11.70 -start-before=amdgpu-insert-delay-alu %s -o - | FileCheck %s
# RUN: llc -mtriple=amdgpu12.00 -start-before=amdgpu-insert-delay-alu %s -o - | FileCheck %s

---
name: wmma_c_reuse_no_delay
tracksRegLiveness: true
body: |
  bb.0:
    ; CHECK-LABEL: {{^}}wmma_c_reuse_no_delay:
    ; CHECK: %bb.0:
    ; CHECK-NEXT: v_wmma_f32_16x16x16_f16 v[0:7], v[8:11], v[12:15], v[0:7]
    ; CHECK-NEXT: v_wmma_f32_16x16x16_f16 v[0:7], v[16:19], v[20:23], v[0:7]
    liveins: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7, $vgpr8_vgpr9_vgpr10_vgpr11, $vgpr12_vgpr13_vgpr14_vgpr15, $vgpr16_vgpr17_vgpr18_vgpr19, $vgpr20_vgpr21_vgpr22_vgpr23
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = V_WMMA_F32_16X16X16_F16_w32_twoaddr 8, $vgpr8_vgpr9_vgpr10_vgpr11, 8, $vgpr12_vgpr13_vgpr14_vgpr15, 8, $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, implicit $exec
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = V_WMMA_F32_16X16X16_F16_w32_twoaddr 8, $vgpr16_vgpr17_vgpr18_vgpr19, 8, $vgpr20_vgpr21_vgpr22_vgpr23, 8, $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, implicit $exec
...

---
name: wmma_no_c_reuse_when_interrupted
tracksRegLiveness: true
body: |
  bb.0:
    ; CHECK-LABEL: {{^}}wmma_no_c_reuse_when_interrupted:
    ; CHECK: %bb.0:
    ; CHECK-NEXT: v_wmma_f32_16x16x16_f16 v[0:7], v[8:11], v[12:15], v[0:7]
    ; CHECK-NEXT: v_add_nc_u32_e32 v24, v24, v25
    ; CHECK-NEXT: s_delay_alu instid0(VALU_DEP_2)
    ; CHECK-NEXT: v_wmma_f32_16x16x16_f16 v[0:7], v[16:19], v[20:23], v[0:7]
    liveins: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7, $vgpr8_vgpr9_vgpr10_vgpr11, $vgpr12_vgpr13_vgpr14_vgpr15, $vgpr16_vgpr17_vgpr18_vgpr19, $vgpr20_vgpr21_vgpr22_vgpr23, $vgpr24, $vgpr25
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = V_WMMA_F32_16X16X16_F16_w32_twoaddr 8, $vgpr8_vgpr9_vgpr10_vgpr11, 8, $vgpr12_vgpr13_vgpr14_vgpr15, 8, $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, implicit $exec
    $vgpr24 = V_ADD_U32_e32 $vgpr24, $vgpr25, implicit $exec
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = V_WMMA_F32_16X16X16_F16_w32_twoaddr 8, $vgpr16_vgpr17_vgpr18_vgpr19, 8, $vgpr20_vgpr21_vgpr22_vgpr23, 8, $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, implicit $exec
...