Lines Matching refs:FLAT

3 …dgcn-- -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs | FileCheck %s --check-prefix=FLAT
4 …mdgcn-- -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs | FileCheck %s --check-prefix=FLAT
33 ; FLAT-LABEL: s_brev_i16:
34 ; FLAT: ; %bb.0:
35 ; FLAT-NEXT: s_load_dword s4, s[0:1], 0x2c
36 ; FLAT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
37 ; FLAT-NEXT: s_mov_b32 s3, 0xf000
38 ; FLAT-NEXT: s_mov_b32 s2, -1
39 ; FLAT-NEXT: s_waitcnt lgkmcnt(0)
40 ; FLAT-NEXT: s_brev_b32 s4, s4
41 ; FLAT-NEXT: s_lshr_b32 s4, s4, 16
42 ; FLAT-NEXT: v_mov_b32_e32 v0, s4
43 ; FLAT-NEXT: buffer_store_short v0, off, s[0:3], 0
44 ; FLAT-NEXT: s_endpgm
82 ; FLAT-LABEL: v_brev_i16:
83 ; FLAT: ; %bb.0:
84 ; FLAT-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x2c
85 ; FLAT-NEXT: s_mov_b32 s3, 0xf000
86 ; FLAT-NEXT: s_mov_b32 s2, -1
87 ; FLAT-NEXT: s_mov_b32 s6, s2
88 ; FLAT-NEXT: s_mov_b32 s7, s3
89 ; FLAT-NEXT: s_waitcnt lgkmcnt(0)
90 ; FLAT-NEXT: buffer_load_ushort v0, off, s[4:7], 0
91 ; FLAT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
92 ; FLAT-NEXT: s_waitcnt vmcnt(0)
93 ; FLAT-NEXT: v_bfrev_b32_e32 v0, v0
94 ; FLAT-NEXT: v_lshrrev_b32_e32 v0, 16, v0
95 ; FLAT-NEXT: s_waitcnt lgkmcnt(0)
96 ; FLAT-NEXT: buffer_store_short v0, off, s[0:3], 0
97 ; FLAT-NEXT: s_endpgm
133 ; FLAT-LABEL: s_brev_i32:
134 ; FLAT: ; %bb.0:
135 ; FLAT-NEXT: s_load_dword s4, s[0:1], 0x2c
136 ; FLAT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
137 ; FLAT-NEXT: s_mov_b32 s3, 0xf000
138 ; FLAT-NEXT: s_mov_b32 s2, -1
139 ; FLAT-NEXT: s_waitcnt lgkmcnt(0)
140 ; FLAT-NEXT: s_brev_b32 s4, s4
141 ; FLAT-NEXT: v_mov_b32_e32 v0, s4
142 ; FLAT-NEXT: buffer_store_dword v0, off, s[0:3], 0
143 ; FLAT-NEXT: s_endpgm
180 ; FLAT-LABEL: v_brev_i32:
181 ; FLAT: ; %bb.0:
182 ; FLAT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2c
183 ; FLAT-NEXT: v_lshlrev_b32_e32 v0, 2, v0
184 ; FLAT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
185 ; FLAT-NEXT: s_waitcnt lgkmcnt(0)
186 ; FLAT-NEXT: v_mov_b32_e32 v1, s3
187 ; FLAT-NEXT: v_add_u32_e32 v0, vcc, s2, v0
188 ; FLAT-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
189 ; FLAT-NEXT: flat_load_dword v0, v[0:1]
190 ; FLAT-NEXT: s_mov_b32 s3, 0xf000
191 ; FLAT-NEXT: s_mov_b32 s2, -1
192 ; FLAT-NEXT: s_waitcnt vmcnt(0)
193 ; FLAT-NEXT: v_bfrev_b32_e32 v0, v0
194 ; FLAT-NEXT: buffer_store_dword v0, off, s[0:3], 0
195 ; FLAT-NEXT: s_endpgm
237 ; FLAT-LABEL: s_brev_v2i32:
238 ; FLAT: ; %bb.0:
239 ; FLAT-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x2c
240 ; FLAT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
241 ; FLAT-NEXT: s_mov_b32 s3, 0xf000
242 ; FLAT-NEXT: s_mov_b32 s2, -1
243 ; FLAT-NEXT: s_waitcnt lgkmcnt(0)
244 ; FLAT-NEXT: s_brev_b32 s5, s5
245 ; FLAT-NEXT: s_brev_b32 s4, s4
246 ; FLAT-NEXT: v_mov_b32_e32 v0, s4
247 ; FLAT-NEXT: v_mov_b32_e32 v1, s5
248 ; FLAT-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
249 ; FLAT-NEXT: s_endpgm
289 ; FLAT-LABEL: v_brev_v2i32:
290 ; FLAT: ; %bb.0:
291 ; FLAT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2c
292 ; FLAT-NEXT: v_lshlrev_b32_e32 v0, 3, v0
293 ; FLAT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
294 ; FLAT-NEXT: s_waitcnt lgkmcnt(0)
295 ; FLAT-NEXT: v_mov_b32_e32 v1, s3
296 ; FLAT-NEXT: v_add_u32_e32 v0, vcc, s2, v0
297 ; FLAT-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
298 ; FLAT-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
299 ; FLAT-NEXT: s_mov_b32 s3, 0xf000
300 ; FLAT-NEXT: s_mov_b32 s2, -1
301 ; FLAT-NEXT: s_waitcnt vmcnt(0)
302 ; FLAT-NEXT: v_bfrev_b32_e32 v1, v1
303 ; FLAT-NEXT: v_bfrev_b32_e32 v0, v0
304 ; FLAT-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
305 ; FLAT-NEXT: s_endpgm
347 ; FLAT-LABEL: s_brev_i64:
348 ; FLAT: ; %bb.0:
349 ; FLAT-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x2c
350 ; FLAT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
351 ; FLAT-NEXT: s_mov_b32 s3, 0xf000
352 ; FLAT-NEXT: s_mov_b32 s2, -1
353 ; FLAT-NEXT: s_waitcnt lgkmcnt(0)
354 ; FLAT-NEXT: s_brev_b64 s[4:5], s[4:5]
355 ; FLAT-NEXT: v_mov_b32_e32 v0, s4
356 ; FLAT-NEXT: v_mov_b32_e32 v1, s5
357 ; FLAT-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
358 ; FLAT-NEXT: s_endpgm
397 ; FLAT-LABEL: v_brev_i64:
398 ; FLAT: ; %bb.0:
399 ; FLAT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2c
400 ; FLAT-NEXT: v_lshlrev_b32_e32 v0, 3, v0
401 ; FLAT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
402 ; FLAT-NEXT: s_waitcnt lgkmcnt(0)
403 ; FLAT-NEXT: v_mov_b32_e32 v1, s3
404 ; FLAT-NEXT: v_add_u32_e32 v0, vcc, s2, v0
405 ; FLAT-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
406 ; FLAT-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
407 ; FLAT-NEXT: s_mov_b32 s3, 0xf000
408 ; FLAT-NEXT: s_mov_b32 s2, -1
409 ; FLAT-NEXT: s_waitcnt vmcnt(0)
410 ; FLAT-NEXT: v_bfrev_b32_e32 v2, v0
411 ; FLAT-NEXT: v_bfrev_b32_e32 v1, v1
412 ; FLAT-NEXT: buffer_store_dwordx2 v[1:2], off, s[0:3], 0
413 ; FLAT-NEXT: s_endpgm
458 ; FLAT-LABEL: s_brev_v2i64:
459 ; FLAT: ; %bb.0:
460 ; FLAT-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x34
461 ; FLAT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
462 ; FLAT-NEXT: s_mov_b32 s3, 0xf000
463 ; FLAT-NEXT: s_mov_b32 s2, -1
464 ; FLAT-NEXT: s_waitcnt lgkmcnt(0)
465 ; FLAT-NEXT: s_brev_b64 s[6:7], s[6:7]
466 ; FLAT-NEXT: s_brev_b64 s[4:5], s[4:5]
467 ; FLAT-NEXT: v_mov_b32_e32 v0, s4
468 ; FLAT-NEXT: v_mov_b32_e32 v1, s5
469 ; FLAT-NEXT: v_mov_b32_e32 v2, s6
470 ; FLAT-NEXT: v_mov_b32_e32 v3, s7
471 ; FLAT-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
472 ; FLAT-NEXT: s_endpgm
516 ; FLAT-LABEL: v_brev_v2i64:
517 ; FLAT: ; %bb.0:
518 ; FLAT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2c
519 ; FLAT-NEXT: v_lshlrev_b32_e32 v0, 4, v0
520 ; FLAT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
521 ; FLAT-NEXT: s_waitcnt lgkmcnt(0)
522 ; FLAT-NEXT: v_mov_b32_e32 v1, s3
523 ; FLAT-NEXT: v_add_u32_e32 v0, vcc, s2, v0
524 ; FLAT-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
525 ; FLAT-NEXT: flat_load_dwordx4 v[0:3], v[0:1]
526 ; FLAT-NEXT: s_mov_b32 s3, 0xf000
527 ; FLAT-NEXT: s_mov_b32 s2, -1
528 ; FLAT-NEXT: s_waitcnt vmcnt(0)
529 ; FLAT-NEXT: v_bfrev_b32_e32 v4, v2
530 ; FLAT-NEXT: v_bfrev_b32_e32 v3, v3
531 ; FLAT-NEXT: v_bfrev_b32_e32 v2, v0
532 ; FLAT-NEXT: v_bfrev_b32_e32 v1, v1
533 ; FLAT-NEXT: buffer_store_dwordx4 v[1:4], off, s[0:3], 0
534 ; FLAT-NEXT: s_endpgm
573 ; FLAT-LABEL: missing_truncate_promote_bitreverse:
574 ; FLAT: ; %bb.0: ; %bb
575 ; FLAT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
576 ; FLAT-NEXT: v_bfrev_b32_e32 v0, v0
577 ; FLAT-NEXT: v_cvt_f32_f16_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
578 ; FLAT-NEXT: s_setpc_b64 s[30:31]