|
| 1 | +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5 |
1 | 2 | ; RUN: llc -mtriple=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s |
2 | | -; RUN: llc -mtriple=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s |
| 3 | +; RUN: llc -mtriple=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -check-prefix=GCNF %s |
3 | 4 |
|
4 | | -; GCN-LABEL: {{^}}any_extend_vector_inreg_v16i8_to_v4i32: |
5 | | -; GCN: s_load_dwordx8 |
6 | | -; GCN-DAG: s_load_dword |
7 | 5 |
|
8 | | -; GCN: {{buffer|flat}}_store_byte |
9 | | -; GCN: {{buffer|flat}}_store_byte |
10 | | -; GCN: {{buffer|flat}}_store_byte |
11 | | -; GCN: {{buffer|flat}}_store_byte |
12 | | - |
13 | | -; GCN: {{buffer|flat}}_store_byte |
14 | | -; GCN: {{buffer|flat}}_store_byte |
15 | | -; GCN: {{buffer|flat}}_store_byte |
16 | | -; GCN: {{buffer|flat}}_store_byte |
17 | | - |
18 | | -; GCN: {{buffer|flat}}_store_byte |
19 | | -; GCN: {{buffer|flat}}_store_byte |
20 | | -; GCN: {{buffer|flat}}_store_byte |
21 | | -; GCN: {{buffer|flat}}_store_byte |
22 | | - |
23 | | -; GCN: {{buffer|flat}}_store_byte |
24 | | -; GCN: {{buffer|flat}}_store_byte |
25 | | -; GCN: {{buffer|flat}}_store_byte |
26 | | -; GCN: {{buffer|flat}}_store_byte |
27 | 6 | define amdgpu_kernel void @any_extend_vector_inreg_v16i8_to_v4i32(ptr addrspace(1) nocapture readonly %arg, ptr addrspace(1) %arg1) local_unnamed_addr #0 { |
| 7 | +; GCN-LABEL: any_extend_vector_inreg_v16i8_to_v4i32: |
| 8 | +; GCN: ; %bb.0: ; %bb |
| 9 | +; GCN-NEXT: s_load_dwordx4 s[12:15], s[4:5], 0x9 |
| 10 | +; GCN-NEXT: s_mov_b32 s3, 0xf000 |
| 11 | +; GCN-NEXT: s_mov_b32 s2, -1 |
| 12 | +; GCN-NEXT: v_mov_b32_e32 v0, 0 |
| 13 | +; GCN-NEXT: s_waitcnt lgkmcnt(0) |
| 14 | +; GCN-NEXT: s_mov_b32 s0, s14 |
| 15 | +; GCN-NEXT: s_mov_b32 s1, s15 |
| 16 | +; GCN-NEXT: s_load_dwordx8 s[4:11], s[12:13], 0x0 |
| 17 | +; GCN-NEXT: s_waitcnt lgkmcnt(0) |
| 18 | +; GCN-NEXT: s_load_dword s4, s[12:13], 0x8 |
| 19 | +; GCN-NEXT: buffer_store_byte v0, off, s[0:3], 0 offset:13 |
| 20 | +; GCN-NEXT: buffer_store_byte v0, off, s[0:3], 0 offset:15 |
| 21 | +; GCN-NEXT: buffer_store_byte v0, off, s[0:3], 0 offset:14 |
| 22 | +; GCN-NEXT: buffer_store_byte v0, off, s[0:3], 0 offset:8 |
| 23 | +; GCN-NEXT: buffer_store_byte v0, off, s[0:3], 0 offset:11 |
| 24 | +; GCN-NEXT: buffer_store_byte v0, off, s[0:3], 0 offset:10 |
| 25 | +; GCN-NEXT: buffer_store_byte v0, off, s[0:3], 0 offset:4 |
| 26 | +; GCN-NEXT: buffer_store_byte v0, off, s[0:3], 0 offset:6 |
| 27 | +; GCN-NEXT: buffer_store_byte v0, off, s[0:3], 0 offset:1 |
| 28 | +; GCN-NEXT: buffer_store_byte v0, off, s[0:3], 0 |
| 29 | +; GCN-NEXT: buffer_store_byte v0, off, s[0:3], 0 offset:3 |
| 30 | +; GCN-NEXT: s_lshr_b32 s8, s9, 16 |
| 31 | +; GCN-NEXT: s_waitcnt expcnt(0) |
| 32 | +; GCN-NEXT: v_mov_b32_e32 v0, s6 |
| 33 | +; GCN-NEXT: s_waitcnt lgkmcnt(0) |
| 34 | +; GCN-NEXT: s_lshl_b64 s[6:7], s[4:5], 8 |
| 35 | +; GCN-NEXT: v_mov_b32_e32 v1, s11 |
| 36 | +; GCN-NEXT: buffer_store_byte v1, off, s[0:3], 0 offset:9 |
| 37 | +; GCN-NEXT: s_waitcnt expcnt(0) |
| 38 | +; GCN-NEXT: v_mov_b32_e32 v1, s5 |
| 39 | +; GCN-NEXT: buffer_store_byte v1, off, s[0:3], 0 offset:2 |
| 40 | +; GCN-NEXT: v_alignbit_b32 v0, s8, v0, 16 |
| 41 | +; GCN-NEXT: s_waitcnt expcnt(0) |
| 42 | +; GCN-NEXT: v_mov_b32_e32 v1, s7 |
| 43 | +; GCN-NEXT: buffer_store_byte v1, off, s[0:3], 0 offset:12 |
| 44 | +; GCN-NEXT: s_waitcnt expcnt(0) |
| 45 | +; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v0 |
| 46 | +; GCN-NEXT: v_lshrrev_b32_e32 v0, 24, v0 |
| 47 | +; GCN-NEXT: buffer_store_byte v1, off, s[0:3], 0 offset:5 |
| 48 | +; GCN-NEXT: buffer_store_byte v0, off, s[0:3], 0 offset:7 |
| 49 | +; GCN-NEXT: s_endpgm |
| 50 | +; |
| 51 | +; GCNF-LABEL: any_extend_vector_inreg_v16i8_to_v4i32: |
| 52 | +; GCNF: ; %bb.0: ; %bb |
| 53 | +; GCNF-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x24 |
| 54 | +; GCNF-NEXT: v_mov_b32_e32 v2, 0 |
| 55 | +; GCNF-NEXT: s_waitcnt lgkmcnt(0) |
| 56 | +; GCNF-NEXT: s_load_dwordx8 s[0:7], s[8:9], 0x0 |
| 57 | +; GCNF-NEXT: s_waitcnt lgkmcnt(0) |
| 58 | +; GCNF-NEXT: s_load_dword s0, s[8:9], 0x20 |
| 59 | +; GCNF-NEXT: s_lshr_b32 s6, s5, 24 |
| 60 | +; GCNF-NEXT: s_lshr_b32 s8, s2, 24 |
| 61 | +; GCNF-NEXT: s_waitcnt lgkmcnt(0) |
| 62 | +; GCNF-NEXT: s_lshl_b64 s[2:3], s[0:1], 8 |
| 63 | +; GCNF-NEXT: s_add_u32 s4, s10, 13 |
| 64 | +; GCNF-NEXT: s_addc_u32 s5, s11, 0 |
| 65 | +; GCNF-NEXT: v_mov_b32_e32 v0, s4 |
| 66 | +; GCNF-NEXT: v_mov_b32_e32 v1, s5 |
| 67 | +; GCNF-NEXT: s_add_u32 s4, s10, 15 |
| 68 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 69 | +; GCNF-NEXT: s_addc_u32 s5, s11, 0 |
| 70 | +; GCNF-NEXT: v_mov_b32_e32 v0, s4 |
| 71 | +; GCNF-NEXT: v_mov_b32_e32 v1, s5 |
| 72 | +; GCNF-NEXT: s_add_u32 s4, s10, 14 |
| 73 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 74 | +; GCNF-NEXT: s_addc_u32 s5, s11, 0 |
| 75 | +; GCNF-NEXT: v_mov_b32_e32 v0, s4 |
| 76 | +; GCNF-NEXT: v_mov_b32_e32 v1, s5 |
| 77 | +; GCNF-NEXT: s_add_u32 s4, s10, 8 |
| 78 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 79 | +; GCNF-NEXT: s_addc_u32 s5, s11, 0 |
| 80 | +; GCNF-NEXT: v_mov_b32_e32 v0, s4 |
| 81 | +; GCNF-NEXT: v_mov_b32_e32 v1, s5 |
| 82 | +; GCNF-NEXT: s_add_u32 s4, s10, 11 |
| 83 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 84 | +; GCNF-NEXT: s_addc_u32 s5, s11, 0 |
| 85 | +; GCNF-NEXT: v_mov_b32_e32 v0, s4 |
| 86 | +; GCNF-NEXT: v_mov_b32_e32 v1, s5 |
| 87 | +; GCNF-NEXT: s_add_u32 s4, s10, 10 |
| 88 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 89 | +; GCNF-NEXT: s_addc_u32 s5, s11, 0 |
| 90 | +; GCNF-NEXT: v_mov_b32_e32 v0, s4 |
| 91 | +; GCNF-NEXT: v_mov_b32_e32 v1, s5 |
| 92 | +; GCNF-NEXT: s_add_u32 s4, s10, 4 |
| 93 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 94 | +; GCNF-NEXT: s_addc_u32 s5, s11, 0 |
| 95 | +; GCNF-NEXT: v_mov_b32_e32 v0, s4 |
| 96 | +; GCNF-NEXT: v_mov_b32_e32 v1, s5 |
| 97 | +; GCNF-NEXT: s_add_u32 s4, s10, 6 |
| 98 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 99 | +; GCNF-NEXT: s_addc_u32 s5, s11, 0 |
| 100 | +; GCNF-NEXT: v_mov_b32_e32 v0, s4 |
| 101 | +; GCNF-NEXT: v_mov_b32_e32 v1, s5 |
| 102 | +; GCNF-NEXT: s_add_u32 s4, s10, 1 |
| 103 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 104 | +; GCNF-NEXT: s_addc_u32 s5, s11, 0 |
| 105 | +; GCNF-NEXT: v_mov_b32_e32 v0, s4 |
| 106 | +; GCNF-NEXT: v_mov_b32_e32 v1, s5 |
| 107 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 108 | +; GCNF-NEXT: v_mov_b32_e32 v0, s10 |
| 109 | +; GCNF-NEXT: v_mov_b32_e32 v1, s11 |
| 110 | +; GCNF-NEXT: s_add_u32 s4, s10, 3 |
| 111 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 112 | +; GCNF-NEXT: s_addc_u32 s5, s11, 0 |
| 113 | +; GCNF-NEXT: v_mov_b32_e32 v0, s4 |
| 114 | +; GCNF-NEXT: v_mov_b32_e32 v1, s5 |
| 115 | +; GCNF-NEXT: s_add_u32 s4, s10, 9 |
| 116 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 117 | +; GCNF-NEXT: s_addc_u32 s5, s11, 0 |
| 118 | +; GCNF-NEXT: v_mov_b32_e32 v0, s4 |
| 119 | +; GCNF-NEXT: v_mov_b32_e32 v1, s5 |
| 120 | +; GCNF-NEXT: v_mov_b32_e32 v2, s7 |
| 121 | +; GCNF-NEXT: s_add_u32 s4, s10, 2 |
| 122 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 123 | +; GCNF-NEXT: s_addc_u32 s5, s11, 0 |
| 124 | +; GCNF-NEXT: v_mov_b32_e32 v0, s4 |
| 125 | +; GCNF-NEXT: v_mov_b32_e32 v1, s5 |
| 126 | +; GCNF-NEXT: v_mov_b32_e32 v2, s1 |
| 127 | +; GCNF-NEXT: s_add_u32 s0, s10, 5 |
| 128 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 129 | +; GCNF-NEXT: s_addc_u32 s1, s11, 0 |
| 130 | +; GCNF-NEXT: v_mov_b32_e32 v0, s0 |
| 131 | +; GCNF-NEXT: v_mov_b32_e32 v1, s1 |
| 132 | +; GCNF-NEXT: v_mov_b32_e32 v2, s8 |
| 133 | +; GCNF-NEXT: s_add_u32 s0, s10, 12 |
| 134 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 135 | +; GCNF-NEXT: s_addc_u32 s1, s11, 0 |
| 136 | +; GCNF-NEXT: v_mov_b32_e32 v0, s0 |
| 137 | +; GCNF-NEXT: v_mov_b32_e32 v1, s1 |
| 138 | +; GCNF-NEXT: v_mov_b32_e32 v2, s3 |
| 139 | +; GCNF-NEXT: s_add_u32 s0, s10, 7 |
| 140 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 141 | +; GCNF-NEXT: s_addc_u32 s1, s11, 0 |
| 142 | +; GCNF-NEXT: v_mov_b32_e32 v0, s0 |
| 143 | +; GCNF-NEXT: v_mov_b32_e32 v1, s1 |
| 144 | +; GCNF-NEXT: v_mov_b32_e32 v2, s6 |
| 145 | +; GCNF-NEXT: flat_store_byte v[0:1], v2 |
| 146 | +; GCNF-NEXT: s_endpgm |
28 | 147 | bb: |
29 | 148 | %tmp2 = load <16 x i8>, ptr addrspace(1) %arg, align 16 |
30 | 149 | %tmp3 = extractelement <16 x i8> %tmp2, i64 4 |
|
0 commit comments