|
5 | 5 | define amdgpu_kernel void @raw_ptr_atomic_buffer_ptr_load_i32(ptr addrspace(8) %ptr) {
|
6 | 6 | ; CHECK-LABEL: raw_ptr_atomic_buffer_ptr_load_i32:
|
7 | 7 | ; CHECK: ; %bb.0: ; %bb
|
8 |
| -; CHECK-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 |
| 8 | +; CHECK-NEXT: s_load_b128 s[0:3], s[2:3], 0x24 |
| 9 | +; CHECK-NEXT: v_and_b32_e32 v0, 0x3ff, v0 |
9 | 10 | ; CHECK-NEXT: s_mov_b32 s4, 0
|
10 | 11 | ; CHECK-NEXT: .LBB0_1: ; %bb1
|
11 | 12 | ; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1
|
|
33 | 34 | define amdgpu_kernel void @raw_ptr_atomic_buffer_load_i32_off(ptr addrspace(8) %ptr) {
|
34 | 35 | ; CHECK-LABEL: raw_ptr_atomic_buffer_load_i32_off:
|
35 | 36 | ; CHECK: ; %bb.0: ; %bb
|
36 |
| -; CHECK-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 |
| 37 | +; CHECK-NEXT: s_load_b128 s[0:3], s[2:3], 0x24 |
| 38 | +; CHECK-NEXT: v_and_b32_e32 v0, 0x3ff, v0 |
37 | 39 | ; CHECK-NEXT: s_mov_b32 s4, 0
|
38 | 40 | ; CHECK-NEXT: .LBB1_1: ; %bb1
|
39 | 41 | ; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1
|
|
60 | 62 | define amdgpu_kernel void @raw_ptr_atomic_buffer_load_i32_soff(ptr addrspace(8) %ptr) {
|
61 | 63 | ; CHECK-LABEL: raw_ptr_atomic_buffer_load_i32_soff:
|
62 | 64 | ; CHECK: ; %bb.0: ; %bb
|
63 |
| -; CHECK-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 |
| 65 | +; CHECK-NEXT: s_load_b128 s[0:3], s[2:3], 0x24 |
| 66 | +; CHECK-NEXT: v_and_b32_e32 v0, 0x3ff, v0 |
64 | 67 | ; CHECK-NEXT: s_mov_b32 s4, 0
|
65 | 68 | ; CHECK-NEXT: .LBB2_1: ; %bb1
|
66 | 69 | ; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1
|
|
87 | 90 | define amdgpu_kernel void @raw_ptr_atomic_buffer_load_i32_dlc(ptr addrspace(8) %ptr) {
|
88 | 91 | ; CHECK-LABEL: raw_ptr_atomic_buffer_load_i32_dlc:
|
89 | 92 | ; CHECK: ; %bb.0: ; %bb
|
90 |
| -; CHECK-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 |
| 93 | +; CHECK-NEXT: s_load_b128 s[0:3], s[2:3], 0x24 |
| 94 | +; CHECK-NEXT: v_and_b32_e32 v0, 0x3ff, v0 |
91 | 95 | ; CHECK-NEXT: s_mov_b32 s4, 0
|
92 | 96 | ; CHECK-NEXT: .LBB3_1: ; %bb1
|
93 | 97 | ; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1
|
|
115 | 119 | define amdgpu_kernel void @raw_nonptr_atomic_buffer_load_i32(ptr addrspace(8) %ptr) {
|
116 | 120 | ; CHECK-LABEL: raw_nonptr_atomic_buffer_load_i32:
|
117 | 121 | ; CHECK: ; %bb.0: ; %bb
|
118 |
| -; CHECK-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 |
| 122 | +; CHECK-NEXT: s_load_b128 s[0:3], s[2:3], 0x24 |
| 123 | +; CHECK-NEXT: v_and_b32_e32 v0, 0x3ff, v0 |
119 | 124 | ; CHECK-NEXT: s_waitcnt lgkmcnt(0)
|
120 | 125 | ; CHECK-NEXT: buffer_load_b32 v1, off, s[0:3], 0 offset:4 glc
|
121 | 126 | ; CHECK-NEXT: s_mov_b32 s0, 0
|
|
144 | 149 | define amdgpu_kernel void @raw_ptr_atomic_buffer_load_i64(ptr addrspace(8) %ptr) {
|
145 | 150 | ; CHECK-LABEL: raw_ptr_atomic_buffer_load_i64:
|
146 | 151 | ; CHECK: ; %bb.0: ; %bb
|
147 |
| -; CHECK-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 |
148 |
| -; CHECK-NEXT: v_mov_b32_e32 v1, 0 |
| 152 | +; CHECK-NEXT: s_load_b128 s[0:3], s[2:3], 0x24 |
| 153 | +; CHECK-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_and_b32 v0, 0x3ff, v0 |
149 | 154 | ; CHECK-NEXT: s_mov_b32 s4, 0
|
150 | 155 | ; CHECK-NEXT: .LBB5_1: ; %bb1
|
151 | 156 | ; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1
|
|
174 | 179 | define amdgpu_kernel void @raw_ptr_atomic_buffer_load_v2i16(ptr addrspace(8) %ptr) {
|
175 | 180 | ; CHECK-LABEL: raw_ptr_atomic_buffer_load_v2i16:
|
176 | 181 | ; CHECK: ; %bb.0: ; %bb
|
177 |
| -; CHECK-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 |
| 182 | +; CHECK-NEXT: s_load_b128 s[0:3], s[2:3], 0x24 |
| 183 | +; CHECK-NEXT: v_and_b32_e32 v0, 0x3ff, v0 |
178 | 184 | ; CHECK-NEXT: s_mov_b32 s4, 0
|
179 | 185 | ; CHECK-NEXT: .LBB6_1: ; %bb1
|
180 | 186 | ; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1
|
|
203 | 209 | define amdgpu_kernel void @raw_ptr_atomic_buffer_load_v4i16(ptr addrspace(8) %ptr) {
|
204 | 210 | ; CHECK-LABEL: raw_ptr_atomic_buffer_load_v4i16:
|
205 | 211 | ; CHECK: ; %bb.0: ; %bb
|
206 |
| -; CHECK-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 |
| 212 | +; CHECK-NEXT: s_load_b128 s[0:3], s[2:3], 0x24 |
| 213 | +; CHECK-NEXT: v_and_b32_e32 v0, 0x3ff, v0 |
207 | 214 | ; CHECK-NEXT: s_mov_b32 s4, 0
|
208 | 215 | ; CHECK-NEXT: .LBB7_1: ; %bb1
|
209 | 216 | ; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1
|
|
236 | 243 | define amdgpu_kernel void @raw_ptr_atomic_buffer_load_v4i32(ptr addrspace(8) %ptr) {
|
237 | 244 | ; CHECK-LABEL: raw_ptr_atomic_buffer_load_v4i32:
|
238 | 245 | ; CHECK: ; %bb.0: ; %bb
|
239 |
| -; CHECK-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 |
| 246 | +; CHECK-NEXT: s_load_b128 s[0:3], s[2:3], 0x24 |
| 247 | +; CHECK-NEXT: v_and_b32_e32 v0, 0x3ff, v0 |
240 | 248 | ; CHECK-NEXT: s_mov_b32 s4, 0
|
241 | 249 | ; CHECK-NEXT: .LBB8_1: ; %bb1
|
242 | 250 | ; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1
|
|
265 | 273 | define amdgpu_kernel void @raw_ptr_atomic_buffer_load_ptr(ptr addrspace(8) %ptr) {
|
266 | 274 | ; CHECK-LABEL: raw_ptr_atomic_buffer_load_ptr:
|
267 | 275 | ; CHECK: ; %bb.0: ; %bb
|
268 |
| -; CHECK-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 |
| 276 | +; CHECK-NEXT: s_load_b128 s[0:3], s[2:3], 0x24 |
| 277 | +; CHECK-NEXT: v_and_b32_e32 v0, 0x3ff, v0 |
269 | 278 | ; CHECK-NEXT: s_mov_b32 s4, 0
|
270 | 279 | ; CHECK-NEXT: .LBB9_1: ; %bb1
|
271 | 280 | ; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1
|
|
0 commit comments