Lines Matching full:s6

30 ; GFX7LESS-NEXT:    v_mbcnt_lo_u32_b32_e64 v0, s6, 0
38 ; GFX7LESS-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
39 ; GFX7LESS-NEXT:    s_mul_i32 s6, s6, 5
44 ; GFX7LESS-NEXT:    v_mov_b32_e32 v1, s6
62 ; GFX8-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
94 ; GFX9-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
127 ; GFX1064-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
133 ; GFX1064-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
135 ; GFX1064-NEXT:    s_mul_i32 s6, s6, 5
137 ; GFX1064-NEXT:    v_mov_b32_e32 v1, s6
159 ; GFX1032-NEXT:    s_mov_b32 s6, exec_lo
161 ; GFX1032-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
166 ; GFX1032-NEXT:    s_bcnt1_i32_b32 s5, s6
194 ; GFX1164-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
201 ; GFX1164-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
203 ; GFX1164-NEXT:    s_mul_i32 s6, s6, 5
205 ; GFX1164-NEXT:    v_mov_b32_e32 v1, s6
227 ; GFX1132-NEXT:    s_mov_b32 s6, exec_lo
229 ; GFX1132-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
235 ; GFX1132-NEXT:    s_bcnt1_i32_b32 s5, s6
263 ; GFX1264-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
271 ; GFX1264-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
274 ; GFX1264-NEXT:    s_mul_i32 s6, s6, 5
277 ; GFX1264-NEXT:    v_mov_b32_e32 v1, s6
298 ; GFX1232-NEXT:    s_mov_b32 s6, exec_lo
300 ; GFX1232-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
306 ; GFX1232-NEXT:    s_bcnt1_i32_b32 s5, s6
339 ; GFX7LESS-NEXT:    v_mbcnt_lo_u32_b32_e64 v0, s6, 0
347 ; GFX7LESS-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
349 ; GFX7LESS-NEXT:    s_mul_i32 s6, s8, s6
353 ; GFX7LESS-NEXT:    v_mov_b32_e32 v1, s6
373 ; GFX8-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
407 ; GFX9-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
443 ; GFX1064-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
449 ; GFX1064-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
452 ; GFX1064-NEXT:    s_mul_i32 s6, s8, s6
454 ; GFX1064-NEXT:    v_mov_b32_e32 v1, s6
476 ; GFX1032-NEXT:    s_load_dword s6, s[4:5], 0x34
487 ; GFX1032-NEXT:    s_mul_i32 s5, s6, s5
501 ; GFX1032-NEXT:    v_mad_u64_u32 v[0:1], s2, s6, v0, s[2:3]
514 ; GFX1164-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
521 ; GFX1164-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
524 ; GFX1164-NEXT:    s_mul_i32 s6, s8, s6
526 ; GFX1164-NEXT:    v_mov_b32_e32 v1, s6
549 ; GFX1132-NEXT:    s_mov_b32 s6, exec_lo
551 ; GFX1132-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
557 ; GFX1132-NEXT:    s_bcnt1_i32_b32 s6, s6
560 ; GFX1132-NEXT:    s_mul_i32 s6, s4, s6
562 ; GFX1132-NEXT:    v_mov_b32_e32 v1, s6
587 ; GFX1264-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
595 ; GFX1264-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
599 ; GFX1264-NEXT:    s_mul_i32 s6, s8, s6
602 ; GFX1264-NEXT:    v_mov_b32_e32 v1, s6
624 ; GFX1232-NEXT:    s_mov_b32 s6, exec_lo
626 ; GFX1232-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
633 ; GFX1232-NEXT:    s_bcnt1_i32_b32 s6, s6
637 ; GFX1232-NEXT:    s_mul_i32 s6, s4, s6
640 ; GFX1232-NEXT:    v_mov_b32_e32 v1, s6
666 ; GFX7LESS_ITERATIVE-NEXT:    s_mov_b32 s6, 0
673 ; GFX7LESS_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, m0
678 ; GFX7LESS_ITERATIVE-NEXT:    s_add_i32 s6, s6, s7
695 ; GFX7LESS_ITERATIVE-NEXT:    v_mov_b32_e32 v0, s6
713 ; GFX8_ITERATIVE-NEXT:    s_mov_b32 s6, 0
721 ; GFX8_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, m0
722 ; GFX8_ITERATIVE-NEXT:    s_add_i32 s6, s6, s7
741 ; GFX8_ITERATIVE-NEXT:    v_mov_b32_e32 v0, s6
758 ; GFX9_ITERATIVE-NEXT:    s_mov_b32 s6, 0
766 ; GFX9_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, m0
767 ; GFX9_ITERATIVE-NEXT:    s_add_i32 s6, s6, s7
786 ; GFX9_ITERATIVE-NEXT:    v_mov_b32_e32 v0, s6
803 ; GFX1064_ITERATIVE-NEXT:    s_mov_b32 s6, 0
810 ; GFX1064_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, s7
812 ; GFX1064_ITERATIVE-NEXT:    s_add_i32 s6, s6, s8
825 ; GFX1064_ITERATIVE-NEXT:    v_mov_b32_e32 v0, s6
849 ; GFX1032_ITERATIVE-NEXT:    s_mov_b32 s6, 0
856 ; GFX1032_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, s1
858 ; GFX1032_ITERATIVE-NEXT:    s_add_i32 s6, s6, s2
870 ; GFX1032_ITERATIVE-NEXT:    v_mov_b32_e32 v0, s6
895 ; GFX1164_ITERATIVE-NEXT:    s_mov_b32 s6, 0
903 ; GFX1164_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s7
906 ; GFX1164_ITERATIVE-NEXT:    s_add_i32 s6, s6, s8
921 ; GFX1164_ITERATIVE-NEXT:    v_mov_b32_e32 v1, s6
946 ; GFX1132_ITERATIVE-NEXT:    s_mov_b32 s6, 0
954 ; GFX1132_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s1
957 ; GFX1132_ITERATIVE-NEXT:    s_add_i32 s6, s6, s2
970 ; GFX1132_ITERATIVE-NEXT:    v_mov_b32_e32 v1, s6
995 ; GFX1264_ITERATIVE-NEXT:    s_mov_b32 s6, 0
1003 ; GFX1264_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s7
1006 ; GFX1264_ITERATIVE-NEXT:    s_add_co_i32 s6, s6, s8
1022 ; GFX1264_ITERATIVE-NEXT:    v_mov_b32_e32 v1, s6
1046 ; GFX1232_ITERATIVE-NEXT:    s_mov_b32 s6, 0
1055 ; GFX1232_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s1
1058 ; GFX1232_ITERATIVE-NEXT:    s_add_co_i32 s6, s6, s2
1072 ; GFX1232_ITERATIVE-NEXT:    v_mov_b32_e32 v1, s6
1096 ; GFX7LESS_DPP-NEXT:    s_mov_b32 s6, -1
1097 ; GFX7LESS_DPP-NEXT:    s_mov_b32 s10, s6
1132 ; GFX8_DPP-NEXT:    v_readlane_b32 s6, v2, 63
1146 ; GFX8_DPP-NEXT:    v_mov_b32_e32 v0, s6
1183 ; GFX9_DPP-NEXT:    v_readlane_b32 s6, v2, 63
1197 ; GFX9_DPP-NEXT:    v_mov_b32_e32 v0, s6
1226 ; GFX1064_DPP-NEXT:    v_readlane_b32 s6, v1, 15
1232 ; GFX1064_DPP-NEXT:    v_writelane_b32 v3, s6, 16
1246 ; GFX1064_DPP-NEXT:    s_mov_b32 s6, -1
1268 ; GFX1064_DPP-NEXT:    s_mov_b32 s2, s6
1286 ; GFX1032_DPP-NEXT:    v_readlane_b32 s6, v1, 31
1294 ; GFX1032_DPP-NEXT:    s_mov_b32 s4, s6
1295 ; GFX1032_DPP-NEXT:    s_mov_b32 s6, -1
1318 ; GFX1032_DPP-NEXT:    s_mov_b32 s2, s6
1343 ; GFX1164_DPP-NEXT:    v_readlane_b32 s6, v1, 15
1349 ; GFX1164_DPP-NEXT:    v_writelane_b32 v3, s6, 16
1365 ; GFX1164_DPP-NEXT:    s_mov_b32 s6, -1
1387 ; GFX1164_DPP-NEXT:    s_mov_b32 s2, s6
1411 ; GFX1132_DPP-NEXT:    v_readlane_b32 s6, v1, 31
1420 ; GFX1132_DPP-NEXT:    s_mov_b32 s4, s6
1421 ; GFX1132_DPP-NEXT:    s_mov_b32 s6, -1
1444 ; GFX1132_DPP-NEXT:    s_mov_b32 s2, s6
1469 ; GFX1264_DPP-NEXT:    v_readlane_b32 s6, v1, 15
1475 ; GFX1264_DPP-NEXT:    v_writelane_b32 v3, s6, 16
1492 ; GFX1264_DPP-NEXT:    s_mov_b32 s6, -1
1514 ; GFX1264_DPP-NEXT:    s_mov_b32 s2, s6
1538 ; GFX1232_DPP-NEXT:    v_readlane_b32 s6, v1, 31
1548 ; GFX1232_DPP-NEXT:    s_mov_b32 s4, s6
1549 ; GFX1232_DPP-NEXT:    s_mov_b32 s6, -1
1572 ; GFX1232_DPP-NEXT:    s_mov_b32 s2, s6
1587 ; GFX7LESS-NEXT:    v_mbcnt_lo_u32_b32_e64 v0, s6, 0
1595 ; GFX7LESS-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
1596 ; GFX7LESS-NEXT:    s_mul_i32 s6, s6, 5
1601 ; GFX7LESS-NEXT:    v_mov_b32_e32 v0, s6
1626 ; GFX8-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
1663 ; GFX9-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
1700 ; GFX1064-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
1707 ; GFX1064-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
1709 ; GFX1064-NEXT:    s_mul_i32 s6, s6, 5
1711 ; GFX1064-NEXT:    v_mov_b32_e32 v0, s6
1735 ; GFX1032-NEXT:    s_mov_b32 s6, exec_lo
1737 ; GFX1032-NEXT:    v_mbcnt_lo_u32_b32 v2, s6, 0
1742 ; GFX1032-NEXT:    s_bcnt1_i32_b32 s5, s6
1772 ; GFX1164-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
1779 ; GFX1164-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
1781 ; GFX1164-NEXT:    s_mul_i32 s6, s6, 5
1783 ; GFX1164-NEXT:    v_mov_b32_e32 v0, s6
1807 ; GFX1132-NEXT:    s_mov_b32 s6, exec_lo
1809 ; GFX1132-NEXT:    v_mbcnt_lo_u32_b32 v2, s6, 0
1815 ; GFX1132-NEXT:    s_bcnt1_i32_b32 s5, s6
1844 ; GFX1264-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
1858 ; GFX1264-NEXT:    v_mov_b32_e32 v0, s6
1884 ; GFX1232-NEXT:    s_mov_b32 s6, exec_lo
1904 ; GFX1232-NEXT:    s_or_b32 exec_lo, exec_lo, s6
1953 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s6, v1
1960 ; GFX7LESS-NEXT:    v_mov_b32_e32 v3, s6
2024 ; GFX9-NEXT:    s_mul_hi_u32 s8, s6, s2
2026 ; GFX9-NEXT:    s_mul_i32 s2, s6, s2
2041 ; GFX9-NEXT:    v_mad_u64_u32 v[0:1], s[2:3], s6, v2, v[0:1]
2066 ; GFX1064-NEXT:    s_mul_hi_u32 s10, s6, s8
2067 ; GFX1064-NEXT:    s_mul_i32 s8, s6, s8
2084 ; GFX1064-NEXT:    v_mad_u64_u32 v[0:1], s[2:3], s6, v2, s[2:3]
2107 ; GFX1032-NEXT:    s_mul_hi_u32 s9, s6, s5
2108 ; GFX1032-NEXT:    s_mul_i32 s5, s6, s5
2125 ; GFX1032-NEXT:    v_mad_u64_u32 v[0:1], s2, s6, v2, s[2:3]
2183 ; GFX1132-NEXT:    s_mov_b32 s6, exec_lo
2207 ; GFX1132-NEXT:    s_or_b32 exec_lo, exec_lo, s6
2268 ; GFX1232-NEXT:    s_mov_b32 s6, exec_lo
2270 ; GFX1232-NEXT:    v_mbcnt_lo_u32_b32 v2, s6, 0
2278 ; GFX1232-NEXT:    s_bcnt1_i32_b32 s6, s6
2285 ; GFX1232-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s7
2323 ; GFX7LESS_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, m0
2324 ; GFX7LESS_ITERATIVE-NEXT:    s_add_u32 s6, s6, s8
2346 ; GFX7LESS_ITERATIVE-NEXT:    v_mov_b32_e32 v3, s6
2377 ; GFX8_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, m0
2378 ; GFX8_ITERATIVE-NEXT:    s_add_u32 s6, s6, s8
2395 ; GFX8_ITERATIVE-NEXT:    v_mov_b32_e32 v3, s6
2430 ; GFX9_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, m0
2431 ; GFX9_ITERATIVE-NEXT:    s_add_u32 s6, s6, s8
2448 ; GFX9_ITERATIVE-NEXT:    v_mov_b32_e32 v3, s6
2482 ; GFX1064_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, s2
2484 ; GFX1064_ITERATIVE-NEXT:    s_add_u32 s6, s6, s3
2500 ; GFX1064_ITERATIVE-NEXT:    v_mov_b32_e32 v3, s6
2535 ; GFX1032_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, s1
2537 ; GFX1032_ITERATIVE-NEXT:    s_add_u32 s6, s6, s2
2552 ; GFX1032_ITERATIVE-NEXT:    v_mov_b32_e32 v3, s6
2589 ; GFX1164_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s2
2591 ; GFX1164_ITERATIVE-NEXT:    s_add_u32 s6, s6, s3
2611 ; GFX1164_ITERATIVE-NEXT:    v_mov_b32_e32 v2, s6
2647 ; GFX1132_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s1
2650 ; GFX1132_ITERATIVE-NEXT:    s_add_u32 s6, s6, s2
2667 ; GFX1132_ITERATIVE-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s7
2705 ; GFX1264_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s10
2723 ; GFX1264_ITERATIVE-NEXT:    v_mov_b32_e32 v2, s6
2761 ; GFX1232_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s1
2777 ; GFX1232_ITERATIVE-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s7
2803 ; GFX7LESS_DPP-NEXT:    s_mov_b32 s6, -1
2805 ; GFX7LESS_DPP-NEXT:    s_mov_b32 s10, s6
2867 ; GFX8_DPP-NEXT:    v_readlane_b32 s6, v3, 63
2876 ; GFX8_DPP-NEXT:    v_mov_b32_e32 v6, s6
2950 ; GFX9_DPP-NEXT:    v_readlane_b32 s6, v3, 63
2959 ; GFX9_DPP-NEXT:    v_mov_b32_e32 v6, s6
3039 ; GFX1064_DPP-NEXT:    v_readlane_b32 s6, v2, 15
3044 ; GFX1064_DPP-NEXT:    v_writelane_b32 v7, s6, 16
3046 ; GFX1064_DPP-NEXT:    v_readlane_b32 s6, v1, 63
3059 ; GFX1064_DPP-NEXT:    s_mov_b32 s6, -1
3083 ; GFX1064_DPP-NEXT:    s_mov_b32 s2, s6
3128 ; GFX1032_DPP-NEXT:    s_or_saveexec_b32 s6, -1
3137 ; GFX1032_DPP-NEXT:    s_mov_b32 exec_lo, s6
3139 ; GFX1032_DPP-NEXT:    s_or_saveexec_b32 s6, -1
3142 ; GFX1032_DPP-NEXT:    s_mov_b32 exec_lo, s6
3143 ; GFX1032_DPP-NEXT:    s_mov_b32 s6, -1
3168 ; GFX1032_DPP-NEXT:    s_mov_b32 s2, s6
3228 ; GFX1164_DPP-NEXT:    v_readlane_b32 s6, v2, 15
3233 ; GFX1164_DPP-NEXT:    v_writelane_b32 v4, s6, 16
3234 ; GFX1164_DPP-NEXT:    v_readlane_b32 s6, v2, 63
3249 ; GFX1164_DPP-NEXT:    s_mov_b32 s6, -1
3274 ; GFX1164_DPP-NEXT:    s_mov_b32 s2, s6
3319 ; GFX1132_DPP-NEXT:    s_or_saveexec_b32 s6, -1
3328 ; GFX1132_DPP-NEXT:    s_mov_b32 exec_lo, s6
3331 ; GFX1132_DPP-NEXT:    s_or_saveexec_b32 s6, -1
3334 ; GFX1132_DPP-NEXT:    s_mov_b32 exec_lo, s6
3335 ; GFX1132_DPP-NEXT:    s_mov_b32 s6, -1
3359 ; GFX1132_DPP-NEXT:    s_mov_b32 s2, s6
3419 ; GFX1264_DPP-NEXT:    v_readlane_b32 s6, v2, 15
3424 ; GFX1264_DPP-NEXT:    v_writelane_b32 v4, s6, 16
3425 ; GFX1264_DPP-NEXT:    v_readlane_b32 s6, v2, 63
3441 ; GFX1264_DPP-NEXT:    s_mov_b32 s6, -1
3466 ; GFX1264_DPP-NEXT:    s_mov_b32 s2, s6
3511 ; GFX1232_DPP-NEXT:    s_or_saveexec_b32 s6, -1
3520 ; GFX1232_DPP-NEXT:    s_mov_b32 exec_lo, s6
3523 ; GFX1232_DPP-NEXT:    s_or_saveexec_b32 s6, -1
3527 ; GFX1232_DPP-NEXT:    s_mov_b32 exec_lo, s6
3528 ; GFX1232_DPP-NEXT:    s_mov_b32 s6, -1
3552 ; GFX1232_DPP-NEXT:    s_mov_b32 s2, s6
3570 ; GFX7LESS-NEXT:    v_mbcnt_lo_u32_b32_e64 v0, s6, 0
3578 ; GFX7LESS-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
3579 ; GFX7LESS-NEXT:    s_mul_i32 s6, s6, 5
3584 ; GFX7LESS-NEXT:    v_mov_b32_e32 v1, s6
3603 ; GFX8-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
3636 ; GFX9-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
3670 ; GFX1064-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
3676 ; GFX1064-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
3678 ; GFX1064-NEXT:    s_mul_i32 s6, s6, 5
3680 ; GFX1064-NEXT:    v_mov_b32_e32 v1, s6
3703 ; GFX1032-NEXT:    s_mov_b32 s6, exec_lo
3705 ; GFX1032-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
3710 ; GFX1032-NEXT:    s_bcnt1_i32_b32 s5, s6
3739 ; GFX1164-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
3746 ; GFX1164-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
3748 ; GFX1164-NEXT:    s_mul_i32 s6, s6, 5
3750 ; GFX1164-NEXT:    v_mov_b32_e32 v1, s6
3773 ; GFX1132-NEXT:    s_mov_b32 s6, exec_lo
3775 ; GFX1132-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
3781 ; GFX1132-NEXT:    s_bcnt1_i32_b32 s5, s6
3810 ; GFX1264-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
3818 ; GFX1264-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
3821 ; GFX1264-NEXT:    s_mul_i32 s6, s6, 5
3824 ; GFX1264-NEXT:    v_mov_b32_e32 v1, s6
3846 ; GFX1232-NEXT:    s_mov_b32 s6, exec_lo
3848 ; GFX1232-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
3854 ; GFX1232-NEXT:    s_bcnt1_i32_b32 s5, s6
3888 ; GFX7LESS-NEXT:    v_mbcnt_lo_u32_b32_e64 v0, s6, 0
3896 ; GFX7LESS-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
3898 ; GFX7LESS-NEXT:    s_mul_i32 s6, s8, s6
3902 ; GFX7LESS-NEXT:    v_mov_b32_e32 v1, s6
3922 ; GFX8-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
3956 ; GFX9-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
3992 ; GFX1064-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
3998 ; GFX1064-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
4001 ; GFX1064-NEXT:    s_mul_i32 s6, s8, s6
4003 ; GFX1064-NEXT:    v_mov_b32_e32 v1, s6
4026 ; GFX1032-NEXT:    s_load_dword s6, s[4:5], 0x34
4037 ; GFX1032-NEXT:    s_mul_i32 s5, s6, s5
4050 ; GFX1032-NEXT:    v_mul_lo_u32 v0, s6, v0
4065 ; GFX1164-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
4072 ; GFX1164-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
4075 ; GFX1164-NEXT:    s_mul_i32 s6, s8, s6
4077 ; GFX1164-NEXT:    v_mov_b32_e32 v1, s6
4101 ; GFX1132-NEXT:    s_mov_b32 s6, exec_lo
4103 ; GFX1132-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
4109 ; GFX1132-NEXT:    s_bcnt1_i32_b32 s6, s6
4112 ; GFX1132-NEXT:    s_mul_i32 s6, s4, s6
4114 ; GFX1132-NEXT:    v_mov_b32_e32 v1, s6
4140 ; GFX1264-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
4148 ; GFX1264-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
4152 ; GFX1264-NEXT:    s_mul_i32 s6, s8, s6
4155 ; GFX1264-NEXT:    v_mov_b32_e32 v1, s6
4178 ; GFX1232-NEXT:    s_mov_b32 s6, exec_lo
4180 ; GFX1232-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
4187 ; GFX1232-NEXT:    s_bcnt1_i32_b32 s6, s6
4191 ; GFX1232-NEXT:    s_mul_i32 s6, s4, s6
4194 ; GFX1232-NEXT:    v_mov_b32_e32 v1, s6
4221 ; GFX7LESS_ITERATIVE-NEXT:    s_mov_b32 s6, 0
4228 ; GFX7LESS_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, m0
4233 ; GFX7LESS_ITERATIVE-NEXT:    s_add_i32 s6, s6, s7
4250 ; GFX7LESS_ITERATIVE-NEXT:    v_mov_b32_e32 v0, s6
4268 ; GFX8_ITERATIVE-NEXT:    s_mov_b32 s6, 0
4276 ; GFX8_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, m0
4277 ; GFX8_ITERATIVE-NEXT:    s_add_i32 s6, s6, s7
4296 ; GFX8_ITERATIVE-NEXT:    v_mov_b32_e32 v0, s6
4313 ; GFX9_ITERATIVE-NEXT:    s_mov_b32 s6, 0
4321 ; GFX9_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, m0
4322 ; GFX9_ITERATIVE-NEXT:    s_add_i32 s6, s6, s7
4341 ; GFX9_ITERATIVE-NEXT:    v_mov_b32_e32 v0, s6
4358 ; GFX1064_ITERATIVE-NEXT:    s_mov_b32 s6, 0
4365 ; GFX1064_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, s7
4367 ; GFX1064_ITERATIVE-NEXT:    s_add_i32 s6, s6, s8
4380 ; GFX1064_ITERATIVE-NEXT:    v_mov_b32_e32 v0, s6
4404 ; GFX1032_ITERATIVE-NEXT:    s_mov_b32 s6, 0
4411 ; GFX1032_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, s1
4413 ; GFX1032_ITERATIVE-NEXT:    s_add_i32 s6, s6, s2
4425 ; GFX1032_ITERATIVE-NEXT:    v_mov_b32_e32 v0, s6
4450 ; GFX1164_ITERATIVE-NEXT:    s_mov_b32 s6, 0
4458 ; GFX1164_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s7
4461 ; GFX1164_ITERATIVE-NEXT:    s_add_i32 s6, s6, s8
4476 ; GFX1164_ITERATIVE-NEXT:    v_mov_b32_e32 v1, s6
4501 ; GFX1132_ITERATIVE-NEXT:    s_mov_b32 s6, 0
4509 ; GFX1132_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s1
4512 ; GFX1132_ITERATIVE-NEXT:    s_add_i32 s6, s6, s2
4525 ; GFX1132_ITERATIVE-NEXT:    v_mov_b32_e32 v1, s6
4550 ; GFX1264_ITERATIVE-NEXT:    s_mov_b32 s6, 0
4558 ; GFX1264_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s7
4561 ; GFX1264_ITERATIVE-NEXT:    s_add_co_i32 s6, s6, s8
4577 ; GFX1264_ITERATIVE-NEXT:    v_mov_b32_e32 v1, s6
4601 ; GFX1232_ITERATIVE-NEXT:    s_mov_b32 s6, 0
4610 ; GFX1232_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s1
4613 ; GFX1232_ITERATIVE-NEXT:    s_add_co_i32 s6, s6, s2
4627 ; GFX1232_ITERATIVE-NEXT:    v_mov_b32_e32 v1, s6
4651 ; GFX7LESS_DPP-NEXT:    s_mov_b32 s6, -1
4652 ; GFX7LESS_DPP-NEXT:    s_mov_b32 s10, s6
4687 ; GFX8_DPP-NEXT:    v_readlane_b32 s6, v2, 63
4701 ; GFX8_DPP-NEXT:    v_mov_b32_e32 v0, s6
4738 ; GFX9_DPP-NEXT:    v_readlane_b32 s6, v2, 63
4752 ; GFX9_DPP-NEXT:    v_mov_b32_e32 v0, s6
4781 ; GFX1064_DPP-NEXT:    v_readlane_b32 s6, v1, 15
4787 ; GFX1064_DPP-NEXT:    v_writelane_b32 v3, s6, 16
4801 ; GFX1064_DPP-NEXT:    s_mov_b32 s6, -1
4823 ; GFX1064_DPP-NEXT:    s_mov_b32 s2, s6
4841 ; GFX1032_DPP-NEXT:    v_readlane_b32 s6, v1, 31
4849 ; GFX1032_DPP-NEXT:    s_mov_b32 s4, s6
4850 ; GFX1032_DPP-NEXT:    s_mov_b32 s6, -1
4873 ; GFX1032_DPP-NEXT:    s_mov_b32 s2, s6
4898 ; GFX1164_DPP-NEXT:    v_readlane_b32 s6, v1, 15
4904 ; GFX1164_DPP-NEXT:    v_writelane_b32 v3, s6, 16
4920 ; GFX1164_DPP-NEXT:    s_mov_b32 s6, -1
4942 ; GFX1164_DPP-NEXT:    s_mov_b32 s2, s6
4966 ; GFX1132_DPP-NEXT:    v_readlane_b32 s6, v1, 31
4975 ; GFX1132_DPP-NEXT:    s_mov_b32 s4, s6
4976 ; GFX1132_DPP-NEXT:    s_mov_b32 s6, -1
4999 ; GFX1132_DPP-NEXT:    s_mov_b32 s2, s6
5024 ; GFX1264_DPP-NEXT:    v_readlane_b32 s6, v1, 15
5030 ; GFX1264_DPP-NEXT:    v_writelane_b32 v3, s6, 16
5047 ; GFX1264_DPP-NEXT:    s_mov_b32 s6, -1
5069 ; GFX1264_DPP-NEXT:    s_mov_b32 s2, s6
5093 ; GFX1232_DPP-NEXT:    v_readlane_b32 s6, v1, 31
5103 ; GFX1232_DPP-NEXT:    s_mov_b32 s4, s6
5104 ; GFX1232_DPP-NEXT:    s_mov_b32 s6, -1
5127 ; GFX1232_DPP-NEXT:    s_mov_b32 s2, s6
5142 ; GFX7LESS-NEXT:    v_mbcnt_lo_u32_b32_e64 v0, s6, 0
5150 ; GFX7LESS-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
5151 ; GFX7LESS-NEXT:    s_mul_i32 s6, s6, 5
5156 ; GFX7LESS-NEXT:    v_mov_b32_e32 v0, s6
5181 ; GFX8-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
5219 ; GFX9-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
5257 ; GFX1064-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
5264 ; GFX1064-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
5266 ; GFX1064-NEXT:    s_mul_i32 s6, s6, 5
5268 ; GFX1064-NEXT:    v_mov_b32_e32 v0, s6
5295 ; GFX1032-NEXT:    s_mov_b32 s6, exec_lo
5297 ; GFX1032-NEXT:    v_mbcnt_lo_u32_b32 v2, s6, 0
5302 ; GFX1032-NEXT:    s_bcnt1_i32_b32 s5, s6
5335 ; GFX1164-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
5342 ; GFX1164-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
5344 ; GFX1164-NEXT:    s_mul_i32 s6, s6, 5
5346 ; GFX1164-NEXT:    v_mov_b32_e32 v0, s6
5373 ; GFX1132-NEXT:    s_mov_b32 s6, exec_lo
5375 ; GFX1132-NEXT:    v_mbcnt_lo_u32_b32 v2, s6, 0
5381 ; GFX1132-NEXT:    s_bcnt1_i32_b32 s5, s6
5413 ; GFX1264-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
5427 ; GFX1264-NEXT:    v_mov_b32_e32 v0, s6
5456 ; GFX1232-NEXT:    s_mov_b32 s6, exec_lo
5476 ; GFX1232-NEXT:    s_or_b32 exec_lo, exec_lo, s6
5528 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s6, v1
5535 ; GFX7LESS-NEXT:    v_mov_b32_e32 v3, s6
5599 ; GFX9-NEXT:    s_mul_hi_u32 s8, s6, s2
5601 ; GFX9-NEXT:    s_mul_i32 s2, s6, s2
5612 ; GFX9-NEXT:    v_mad_u64_u32 v[3:4], s[2:3], s6, v2, 0
5642 ; GFX1064-NEXT:    s_mul_hi_u32 s10, s6, s8
5643 ; GFX1064-NEXT:    s_mul_i32 s8, s6, s8
5658 ; GFX1064-NEXT:    v_mad_u64_u32 v[3:4], s[2:3], s6, v2, 0
5686 ; GFX1032-NEXT:    s_mul_hi_u32 s9, s6, s5
5687 ; GFX1032-NEXT:    s_mul_i32 s5, s6, s5
5702 ; GFX1032-NEXT:    v_mad_u64_u32 v[3:4], s2, s6, v2, 0
5767 ; GFX1132-NEXT:    s_mov_b32 s6, exec_lo
5791 ; GFX1132-NEXT:    s_or_b32 exec_lo, exec_lo, s6
5858 ; GFX1232-NEXT:    s_mov_b32 s6, exec_lo
5860 ; GFX1232-NEXT:    v_mbcnt_lo_u32_b32 v2, s6, 0
5868 ; GFX1232-NEXT:    s_bcnt1_i32_b32 s6, s6
5875 ; GFX1232-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s7
5917 ; GFX7LESS_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, m0
5918 ; GFX7LESS_ITERATIVE-NEXT:    s_add_u32 s6, s6, s8
5940 ; GFX7LESS_ITERATIVE-NEXT:    v_mov_b32_e32 v3, s6
5971 ; GFX8_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, m0
5972 ; GFX8_ITERATIVE-NEXT:    s_add_u32 s6, s6, s8
5989 ; GFX8_ITERATIVE-NEXT:    v_mov_b32_e32 v3, s6
6024 ; GFX9_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, m0
6025 ; GFX9_ITERATIVE-NEXT:    s_add_u32 s6, s6, s8
6042 ; GFX9_ITERATIVE-NEXT:    v_mov_b32_e32 v3, s6
6076 ; GFX1064_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, s2
6078 ; GFX1064_ITERATIVE-NEXT:    s_add_u32 s6, s6, s3
6094 ; GFX1064_ITERATIVE-NEXT:    v_mov_b32_e32 v3, s6
6129 ; GFX1032_ITERATIVE-NEXT:    v_writelane_b32 v1, s6, s1
6131 ; GFX1032_ITERATIVE-NEXT:    s_add_u32 s6, s6, s2
6146 ; GFX1032_ITERATIVE-NEXT:    v_mov_b32_e32 v3, s6
6183 ; GFX1164_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s2
6185 ; GFX1164_ITERATIVE-NEXT:    s_add_u32 s6, s6, s3
6205 ; GFX1164_ITERATIVE-NEXT:    v_mov_b32_e32 v2, s6
6241 ; GFX1132_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s1
6244 ; GFX1132_ITERATIVE-NEXT:    s_add_u32 s6, s6, s2
6261 ; GFX1132_ITERATIVE-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s7
6299 ; GFX1264_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s10
6317 ; GFX1264_ITERATIVE-NEXT:    v_mov_b32_e32 v2, s6
6355 ; GFX1232_ITERATIVE-NEXT:    v_writelane_b32 v0, s6, s1
6371 ; GFX1232_ITERATIVE-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s7
6397 ; GFX7LESS_DPP-NEXT:    s_mov_b32 s6, -1
6399 ; GFX7LESS_DPP-NEXT:    s_mov_b32 s10, s6
6461 ; GFX8_DPP-NEXT:    v_readlane_b32 s6, v3, 63
6470 ; GFX8_DPP-NEXT:    v_mov_b32_e32 v6, s6
6544 ; GFX9_DPP-NEXT:    v_readlane_b32 s6, v3, 63
6553 ; GFX9_DPP-NEXT:    v_mov_b32_e32 v6, s6
6633 ; GFX1064_DPP-NEXT:    v_readlane_b32 s6, v2, 15
6638 ; GFX1064_DPP-NEXT:    v_writelane_b32 v7, s6, 16
6640 ; GFX1064_DPP-NEXT:    v_readlane_b32 s6, v1, 63
6653 ; GFX1064_DPP-NEXT:    s_mov_b32 s6, -1
6677 ; GFX1064_DPP-NEXT:    s_mov_b32 s2, s6
6722 ; GFX1032_DPP-NEXT:    s_or_saveexec_b32 s6, -1
6731 ; GFX1032_DPP-NEXT:    s_mov_b32 exec_lo, s6
6733 ; GFX1032_DPP-NEXT:    s_or_saveexec_b32 s6, -1
6736 ; GFX1032_DPP-NEXT:    s_mov_b32 exec_lo, s6
6737 ; GFX1032_DPP-NEXT:    s_mov_b32 s6, -1
6762 ; GFX1032_DPP-NEXT:    s_mov_b32 s2, s6
6822 ; GFX1164_DPP-NEXT:    v_readlane_b32 s6, v2, 15
6827 ; GFX1164_DPP-NEXT:    v_writelane_b32 v4, s6, 16
6828 ; GFX1164_DPP-NEXT:    v_readlane_b32 s6, v2, 63
6843 ; GFX1164_DPP-NEXT:    s_mov_b32 s6, -1
6868 ; GFX1164_DPP-NEXT:    s_mov_b32 s2, s6
6913 ; GFX1132_DPP-NEXT:    s_or_saveexec_b32 s6, -1
6922 ; GFX1132_DPP-NEXT:    s_mov_b32 exec_lo, s6
6925 ; GFX1132_DPP-NEXT:    s_or_saveexec_b32 s6, -1
6928 ; GFX1132_DPP-NEXT:    s_mov_b32 exec_lo, s6
6929 ; GFX1132_DPP-NEXT:    s_mov_b32 s6, -1
6953 ; GFX1132_DPP-NEXT:    s_mov_b32 s2, s6
7013 ; GFX1264_DPP-NEXT:    v_readlane_b32 s6, v2, 15
7018 ; GFX1264_DPP-NEXT:    v_writelane_b32 v4, s6, 16
7019 ; GFX1264_DPP-NEXT:    v_readlane_b32 s6, v2, 63
7035 ; GFX1264_DPP-NEXT:    s_mov_b32 s6, -1
7060 ; GFX1264_DPP-NEXT:    s_mov_b32 s2, s6
7105 ; GFX1232_DPP-NEXT:    s_or_saveexec_b32 s6, -1
7114 ; GFX1232_DPP-NEXT:    s_mov_b32 exec_lo, s6
7117 ; GFX1232_DPP-NEXT:    s_or_saveexec_b32 s6, -1
7121 ; GFX1232_DPP-NEXT:    s_mov_b32 exec_lo, s6
7122 ; GFX1232_DPP-NEXT:    s_mov_b32 s6, -1
7146 ; GFX1232_DPP-NEXT:    s_mov_b32 s2, s6