Lines Matching full:ci
3 …ched=0 -march=amdgcn -mcpu=bonaire -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,CI %s
48 ; CI-LABEL: frem_f16:
49 ; CI: ; %bb.0:
50 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
51 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
52 ; CI-NEXT: s_mov_b32 s11, 0xf000
53 ; CI-NEXT: s_mov_b32 s10, -1
54 ; CI-NEXT: s_mov_b32 s2, s10
55 ; CI-NEXT: s_waitcnt lgkmcnt(0)
56 ; CI-NEXT: s_mov_b32 s8, s4
57 ; CI-NEXT: s_mov_b32 s9, s5
58 ; CI-NEXT: s_mov_b32 s4, s6
59 ; CI-NEXT: s_mov_b32 s5, s7
60 ; CI-NEXT: s_mov_b32 s3, s11
61 ; CI-NEXT: s_mov_b32 s6, s10
62 ; CI-NEXT: s_mov_b32 s7, s11
63 ; CI-NEXT: buffer_load_ushort v0, off, s[4:7], 0
64 ; CI-NEXT: buffer_load_ushort v1, off, s[0:3], 0 offset:8
65 ; CI-NEXT: s_waitcnt vmcnt(1)
66 ; CI-NEXT: v_cvt_f32_f16_e32 v0, v0
67 ; CI-NEXT: s_waitcnt vmcnt(0)
68 ; CI-NEXT: v_cvt_f32_f16_e32 v1, v1
69 ; CI-NEXT: v_div_scale_f32 v3, s[0:1], v1, v1, v0
70 ; CI-NEXT: v_div_scale_f32 v2, vcc, v0, v1, v0
71 ; CI-NEXT: v_rcp_f32_e32 v4, v3
72 ; CI-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3
73 ; CI-NEXT: v_fma_f32 v5, -v3, v4, 1.0
74 ; CI-NEXT: v_fma_f32 v4, v5, v4, v4
75 ; CI-NEXT: v_mul_f32_e32 v5, v2, v4
76 ; CI-NEXT: v_fma_f32 v6, -v3, v5, v2
77 ; CI-NEXT: v_fma_f32 v5, v6, v4, v5
78 ; CI-NEXT: v_fma_f32 v2, -v3, v5, v2
79 ; CI-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0
80 ; CI-NEXT: v_div_fmas_f32 v2, v2, v4, v5
81 ; CI-NEXT: v_div_fixup_f32 v2, v2, v1, v0
82 ; CI-NEXT: v_trunc_f32_e32 v2, v2
83 ; CI-NEXT: v_fma_f32 v0, -v2, v1, v0
84 ; CI-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0
85 ; CI-NEXT: v_cvt_f16_f32_e32 v0, v0
86 ; CI-NEXT: buffer_store_short v0, off, s[8:11], 0
87 ; CI-NEXT: s_endpgm
155 ; CI-LABEL: fast_frem_f16:
156 ; CI: ; %bb.0:
157 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
158 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
159 ; CI-NEXT: s_mov_b32 s11, 0xf000
160 ; CI-NEXT: s_mov_b32 s10, -1
161 ; CI-NEXT: s_mov_b32 s2, s10
162 ; CI-NEXT: s_mov_b32 s3, s11
163 ; CI-NEXT: s_waitcnt lgkmcnt(0)
164 ; CI-NEXT: buffer_load_ushort v1, off, s[0:3], 0 offset:8
165 ; CI-NEXT: s_mov_b32 s8, s4
166 ; CI-NEXT: s_mov_b32 s9, s5
167 ; CI-NEXT: s_mov_b32 s4, s6
168 ; CI-NEXT: s_mov_b32 s5, s7
169 ; CI-NEXT: s_mov_b32 s6, s10
170 ; CI-NEXT: s_mov_b32 s7, s11
171 ; CI-NEXT: buffer_load_ushort v0, off, s[4:7], 0
172 ; CI-NEXT: s_waitcnt vmcnt(1)
173 ; CI-NEXT: v_cvt_f32_f16_e32 v1, v1
174 ; CI-NEXT: v_rcp_f32_e32 v2, v1
175 ; CI-NEXT: s_waitcnt vmcnt(0)
176 ; CI-NEXT: v_cvt_f32_f16_e32 v0, v0
177 ; CI-NEXT: v_mul_f32_e32 v2, v0, v2
178 ; CI-NEXT: v_trunc_f32_e32 v2, v2
179 ; CI-NEXT: v_fma_f32 v0, -v2, v1, v0
180 ; CI-NEXT: v_cvt_f16_f32_e32 v0, v0
181 ; CI-NEXT: buffer_store_short v0, off, s[8:11], 0
182 ; CI-NEXT: s_endpgm
245 ; CI-LABEL: unsafe_frem_f16:
246 ; CI: ; %bb.0:
247 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
248 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
249 ; CI-NEXT: s_mov_b32 s11, 0xf000
250 ; CI-NEXT: s_mov_b32 s10, -1
251 ; CI-NEXT: s_mov_b32 s2, s10
252 ; CI-NEXT: s_mov_b32 s3, s11
253 ; CI-NEXT: s_waitcnt lgkmcnt(0)
254 ; CI-NEXT: buffer_load_ushort v1, off, s[0:3], 0 offset:8
255 ; CI-NEXT: s_mov_b32 s8, s4
256 ; CI-NEXT: s_mov_b32 s9, s5
257 ; CI-NEXT: s_mov_b32 s4, s6
258 ; CI-NEXT: s_mov_b32 s5, s7
259 ; CI-NEXT: s_mov_b32 s6, s10
260 ; CI-NEXT: s_mov_b32 s7, s11
261 ; CI-NEXT: buffer_load_ushort v0, off, s[4:7], 0
262 ; CI-NEXT: s_waitcnt vmcnt(1)
263 ; CI-NEXT: v_cvt_f32_f16_e32 v1, v1
264 ; CI-NEXT: v_rcp_f32_e32 v2, v1
265 ; CI-NEXT: s_waitcnt vmcnt(0)
266 ; CI-NEXT: v_cvt_f32_f16_e32 v0, v0
267 ; CI-NEXT: v_mul_f32_e32 v2, v0, v2
268 ; CI-NEXT: v_trunc_f32_e32 v2, v2
269 ; CI-NEXT: v_fma_f32 v0, -v2, v1, v0
270 ; CI-NEXT: v_cvt_f16_f32_e32 v0, v0
271 ; CI-NEXT: buffer_store_short v0, off, s[8:11], 0
272 ; CI-NEXT: s_endpgm
342 ; CI-LABEL: frem_f32:
343 ; CI: ; %bb.0:
344 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
345 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
346 ; CI-NEXT: s_mov_b32 s11, 0xf000
347 ; CI-NEXT: s_mov_b32 s10, -1
348 ; CI-NEXT: s_mov_b32 s2, s10
349 ; CI-NEXT: s_waitcnt lgkmcnt(0)
350 ; CI-NEXT: s_mov_b32 s8, s4
351 ; CI-NEXT: s_mov_b32 s9, s5
352 ; CI-NEXT: s_mov_b32 s4, s6
353 ; CI-NEXT: s_mov_b32 s5, s7
354 ; CI-NEXT: s_mov_b32 s6, s10
355 ; CI-NEXT: s_mov_b32 s7, s11
356 ; CI-NEXT: s_mov_b32 s3, s11
357 ; CI-NEXT: buffer_load_dword v0, off, s[4:7], 0
358 ; CI-NEXT: buffer_load_dword v1, off, s[0:3], 0 offset:16
359 ; CI-NEXT: s_waitcnt vmcnt(0)
360 ; CI-NEXT: v_div_scale_f32 v3, s[0:1], v1, v1, v0
361 ; CI-NEXT: v_div_scale_f32 v2, vcc, v0, v1, v0
362 ; CI-NEXT: v_rcp_f32_e32 v4, v3
363 ; CI-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3
364 ; CI-NEXT: v_fma_f32 v5, -v3, v4, 1.0
365 ; CI-NEXT: v_fma_f32 v4, v5, v4, v4
366 ; CI-NEXT: v_mul_f32_e32 v5, v2, v4
367 ; CI-NEXT: v_fma_f32 v6, -v3, v5, v2
368 ; CI-NEXT: v_fma_f32 v5, v6, v4, v5
369 ; CI-NEXT: v_fma_f32 v2, -v3, v5, v2
370 ; CI-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0
371 ; CI-NEXT: v_div_fmas_f32 v2, v2, v4, v5
372 ; CI-NEXT: v_div_fixup_f32 v2, v2, v1, v0
373 ; CI-NEXT: v_trunc_f32_e32 v2, v2
374 ; CI-NEXT: v_fma_f32 v0, -v2, v1, v0
375 ; CI-NEXT: buffer_store_dword v0, off, s[8:11], 0
376 ; CI-NEXT: s_endpgm
446 ; CI-LABEL: fast_frem_f32:
447 ; CI: ; %bb.0:
448 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
449 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
450 ; CI-NEXT: s_mov_b32 s11, 0xf000
451 ; CI-NEXT: s_mov_b32 s10, -1
452 ; CI-NEXT: s_mov_b32 s2, s10
453 ; CI-NEXT: s_waitcnt lgkmcnt(0)
454 ; CI-NEXT: s_mov_b32 s8, s4
455 ; CI-NEXT: s_mov_b32 s9, s5
456 ; CI-NEXT: s_mov_b32 s4, s6
457 ; CI-NEXT: s_mov_b32 s5, s7
458 ; CI-NEXT: s_mov_b32 s6, s10
459 ; CI-NEXT: s_mov_b32 s7, s11
460 ; CI-NEXT: s_mov_b32 s3, s11
461 ; CI-NEXT: buffer_load_dword v0, off, s[4:7], 0
462 ; CI-NEXT: buffer_load_dword v1, off, s[0:3], 0 offset:16
463 ; CI-NEXT: s_waitcnt vmcnt(0)
464 ; CI-NEXT: v_rcp_f32_e32 v2, v1
465 ; CI-NEXT: v_mul_f32_e32 v2, v0, v2
466 ; CI-NEXT: v_trunc_f32_e32 v2, v2
467 ; CI-NEXT: v_fma_f32 v0, -v2, v1, v0
468 ; CI-NEXT: buffer_store_dword v0, off, s[8:11], 0
469 ; CI-NEXT: s_endpgm
528 ; CI-LABEL: unsafe_frem_f32:
529 ; CI: ; %bb.0:
530 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
531 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
532 ; CI-NEXT: s_mov_b32 s11, 0xf000
533 ; CI-NEXT: s_mov_b32 s10, -1
534 ; CI-NEXT: s_mov_b32 s2, s10
535 ; CI-NEXT: s_waitcnt lgkmcnt(0)
536 ; CI-NEXT: s_mov_b32 s8, s4
537 ; CI-NEXT: s_mov_b32 s9, s5
538 ; CI-NEXT: s_mov_b32 s4, s6
539 ; CI-NEXT: s_mov_b32 s5, s7
540 ; CI-NEXT: s_mov_b32 s6, s10
541 ; CI-NEXT: s_mov_b32 s7, s11
542 ; CI-NEXT: s_mov_b32 s3, s11
543 ; CI-NEXT: buffer_load_dword v0, off, s[4:7], 0
544 ; CI-NEXT: buffer_load_dword v1, off, s[0:3], 0 offset:16
545 ; CI-NEXT: s_waitcnt vmcnt(0)
546 ; CI-NEXT: v_rcp_f32_e32 v2, v1
547 ; CI-NEXT: v_mul_f32_e32 v2, v0, v2
548 ; CI-NEXT: v_trunc_f32_e32 v2, v2
549 ; CI-NEXT: v_fma_f32 v0, -v2, v1, v0
550 ; CI-NEXT: buffer_store_dword v0, off, s[8:11], 0
551 ; CI-NEXT: s_endpgm
638 ; CI-LABEL: frem_f64:
639 ; CI: ; %bb.0:
640 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
641 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
642 ; CI-NEXT: s_mov_b32 s11, 0xf000
643 ; CI-NEXT: s_mov_b32 s10, -1
644 ; CI-NEXT: s_mov_b32 s2, s10
645 ; CI-NEXT: s_waitcnt lgkmcnt(0)
646 ; CI-NEXT: s_mov_b32 s8, s4
647 ; CI-NEXT: s_mov_b32 s9, s5
648 ; CI-NEXT: s_mov_b32 s4, s6
649 ; CI-NEXT: s_mov_b32 s5, s7
650 ; CI-NEXT: s_mov_b32 s6, s10
651 ; CI-NEXT: s_mov_b32 s7, s11
652 ; CI-NEXT: s_mov_b32 s3, s11
653 ; CI-NEXT: buffer_load_dwordx2 v[0:1], off, s[4:7], 0
654 ; CI-NEXT: buffer_load_dwordx2 v[2:3], off, s[0:3], 0
655 ; CI-NEXT: s_waitcnt vmcnt(0)
656 ; CI-NEXT: v_div_scale_f64 v[4:5], s[0:1], v[2:3], v[2:3], v[0:1]
657 ; CI-NEXT: v_rcp_f64_e32 v[6:7], v[4:5]
658 ; CI-NEXT: v_fma_f64 v[8:9], -v[4:5], v[6:7], 1.0
659 ; CI-NEXT: v_fma_f64 v[6:7], v[6:7], v[8:9], v[6:7]
660 ; CI-NEXT: v_fma_f64 v[8:9], -v[4:5], v[6:7], 1.0
661 ; CI-NEXT: v_fma_f64 v[6:7], v[6:7], v[8:9], v[6:7]
662 ; CI-NEXT: v_div_scale_f64 v[8:9], vcc, v[0:1], v[2:3], v[0:1]
663 ; CI-NEXT: v_mul_f64 v[10:11], v[8:9], v[6:7]
664 ; CI-NEXT: v_fma_f64 v[4:5], -v[4:5], v[10:11], v[8:9]
665 ; CI-NEXT: s_nop 1
666 ; CI-NEXT: v_div_fmas_f64 v[4:5], v[4:5], v[6:7], v[10:11]
667 ; CI-NEXT: v_div_fixup_f64 v[4:5], v[4:5], v[2:3], v[0:1]
668 ; CI-NEXT: v_trunc_f64_e32 v[4:5], v[4:5]
669 ; CI-NEXT: v_fma_f64 v[0:1], -v[4:5], v[2:3], v[0:1]
670 ; CI-NEXT: buffer_store_dwordx2 v[0:1], off, s[8:11], 0
671 ; CI-NEXT: s_endpgm
765 ; CI-LABEL: fast_frem_f64:
766 ; CI: ; %bb.0:
767 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
768 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
769 ; CI-NEXT: s_mov_b32 s11, 0xf000
770 ; CI-NEXT: s_mov_b32 s10, -1
771 ; CI-NEXT: s_mov_b32 s2, s10
772 ; CI-NEXT: s_waitcnt lgkmcnt(0)
773 ; CI-NEXT: s_mov_b32 s8, s4
774 ; CI-NEXT: s_mov_b32 s9, s5
775 ; CI-NEXT: s_mov_b32 s4, s6
776 ; CI-NEXT: s_mov_b32 s5, s7
777 ; CI-NEXT: s_mov_b32 s6, s10
778 ; CI-NEXT: s_mov_b32 s7, s11
779 ; CI-NEXT: s_mov_b32 s3, s11
780 ; CI-NEXT: buffer_load_dwordx2 v[0:1], off, s[4:7], 0
781 ; CI-NEXT: buffer_load_dwordx2 v[2:3], off, s[0:3], 0
782 ; CI-NEXT: s_waitcnt vmcnt(0)
783 ; CI-NEXT: v_div_scale_f64 v[4:5], s[0:1], v[2:3], v[2:3], v[0:1]
784 ; CI-NEXT: v_rcp_f64_e32 v[6:7], v[4:5]
785 ; CI-NEXT: v_fma_f64 v[8:9], -v[4:5], v[6:7], 1.0
786 ; CI-NEXT: v_fma_f64 v[6:7], v[6:7], v[8:9], v[6:7]
787 ; CI-NEXT: v_fma_f64 v[8:9], -v[4:5], v[6:7], 1.0
788 ; CI-NEXT: v_fma_f64 v[6:7], v[6:7], v[8:9], v[6:7]
789 ; CI-NEXT: v_div_scale_f64 v[8:9], vcc, v[0:1], v[2:3], v[0:1]
790 ; CI-NEXT: v_mul_f64 v[10:11], v[8:9], v[6:7]
791 ; CI-NEXT: v_fma_f64 v[4:5], -v[4:5], v[10:11], v[8:9]
792 ; CI-NEXT: s_nop 1
793 ; CI-NEXT: v_div_fmas_f64 v[4:5], v[4:5], v[6:7], v[10:11]
794 ; CI-NEXT: v_div_fixup_f64 v[4:5], v[4:5], v[2:3], v[0:1]
795 ; CI-NEXT: v_trunc_f64_e32 v[4:5], v[4:5]
796 ; CI-NEXT: v_fma_f64 v[0:1], -v[4:5], v[2:3], v[0:1]
797 ; CI-NEXT: buffer_store_dwordx2 v[0:1], off, s[8:11], 0
798 ; CI-NEXT: s_endpgm
879 ; CI-LABEL: unsafe_frem_f64:
880 ; CI: ; %bb.0:
881 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
882 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
883 ; CI-NEXT: s_mov_b32 s11, 0xf000
884 ; CI-NEXT: s_mov_b32 s10, -1
885 ; CI-NEXT: s_mov_b32 s2, s10
886 ; CI-NEXT: s_waitcnt lgkmcnt(0)
887 ; CI-NEXT: s_mov_b32 s8, s4
888 ; CI-NEXT: s_mov_b32 s9, s5
889 ; CI-NEXT: s_mov_b32 s4, s6
890 ; CI-NEXT: s_mov_b32 s5, s7
891 ; CI-NEXT: s_mov_b32 s6, s10
892 ; CI-NEXT: s_mov_b32 s7, s11
893 ; CI-NEXT: s_mov_b32 s3, s11
894 ; CI-NEXT: buffer_load_dwordx2 v[0:1], off, s[4:7], 0
895 ; CI-NEXT: buffer_load_dwordx2 v[2:3], off, s[0:3], 0
896 ; CI-NEXT: s_waitcnt vmcnt(0)
897 ; CI-NEXT: v_rcp_f64_e32 v[4:5], v[2:3]
898 ; CI-NEXT: v_mul_f64 v[4:5], v[0:1], v[4:5]
899 ; CI-NEXT: v_trunc_f64_e32 v[4:5], v[4:5]
900 ; CI-NEXT: v_fma_f64 v[0:1], -v[4:5], v[2:3], v[0:1]
901 ; CI-NEXT: buffer_store_dwordx2 v[0:1], off, s[8:11], 0
902 ; CI-NEXT: s_endpgm
998 ; CI-LABEL: frem_v2f16:
999 ; CI: ; %bb.0:
1000 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
1001 ; CI-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0xd
1002 ; CI-NEXT: s_mov_b32 s3, 0xf000
1003 ; CI-NEXT: s_mov_b32 s2, -1
1004 ; CI-NEXT: s_mov_b32 s10, s2
1005 ; CI-NEXT: s_waitcnt lgkmcnt(0)
1006 ; CI-NEXT: s_mov_b32 s0, s4
1007 ; CI-NEXT: s_mov_b32 s1, s5
1008 ; CI-NEXT: s_mov_b32 s4, s6
1009 ; CI-NEXT: s_mov_b32 s5, s7
1010 ; CI-NEXT: s_mov_b32 s11, s3
1011 ; CI-NEXT: s_mov_b32 s6, s2
1012 ; CI-NEXT: s_mov_b32 s7, s3
1013 ; CI-NEXT: buffer_load_dword v0, off, s[4:7], 0
1014 ; CI-NEXT: buffer_load_dword v2, off, s[8:11], 0 offset:16
1015 ; CI-NEXT: s_mov_b32 s6, 3
1016 ; CI-NEXT: s_mov_b32 s7, 0
1017 ; CI-NEXT: s_waitcnt vmcnt(1)
1018 ; CI-NEXT: v_cvt_f32_f16_e32 v1, v0
1019 ; CI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
1020 ; CI-NEXT: s_waitcnt vmcnt(0)
1021 ; CI-NEXT: v_cvt_f32_f16_e32 v3, v2
1022 ; CI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
1023 ; CI-NEXT: v_cvt_f32_f16_e32 v0, v0
1024 ; CI-NEXT: v_cvt_f32_f16_e32 v2, v2
1025 ; CI-NEXT: v_div_scale_f32 v5, s[4:5], v2, v2, v0
1026 ; CI-NEXT: v_div_scale_f32 v4, vcc, v0, v2, v0
1027 ; CI-NEXT: v_rcp_f32_e32 v6, v5
1028 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s6
1029 ; CI-NEXT: v_fma_f32 v7, -v5, v6, 1.0
1030 ; CI-NEXT: v_fma_f32 v6, v7, v6, v6
1031 ; CI-NEXT: v_mul_f32_e32 v7, v4, v6
1032 ; CI-NEXT: v_fma_f32 v8, -v5, v7, v4
1033 ; CI-NEXT: v_fma_f32 v7, v8, v6, v7
1034 ; CI-NEXT: v_fma_f32 v4, -v5, v7, v4
1035 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s7
1036 ; CI-NEXT: v_div_fmas_f32 v4, v4, v6, v7
1037 ; CI-NEXT: v_div_fixup_f32 v4, v4, v2, v0
1038 ; CI-NEXT: v_trunc_f32_e32 v4, v4
1039 ; CI-NEXT: v_fma_f32 v0, -v4, v2, v0
1040 ; CI-NEXT: v_div_scale_f32 v4, s[4:5], v3, v3, v1
1041 ; CI-NEXT: v_div_scale_f32 v2, vcc, v1, v3, v1
1042 ; CI-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0
1043 ; CI-NEXT: v_cvt_f16_f32_e32 v0, v0
1044 ; CI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
1045 ; CI-NEXT: v_rcp_f32_e32 v5, v4
1046 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s6
1047 ; CI-NEXT: v_fma_f32 v6, -v4, v5, 1.0
1048 ; CI-NEXT: v_fma_f32 v5, v6, v5, v5
1049 ; CI-NEXT: v_mul_f32_e32 v6, v2, v5
1050 ; CI-NEXT: v_fma_f32 v7, -v4, v6, v2
1051 ; CI-NEXT: v_fma_f32 v6, v7, v5, v6
1052 ; CI-NEXT: v_fma_f32 v2, -v4, v6, v2
1053 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s7
1054 ; CI-NEXT: v_div_fmas_f32 v2, v2, v5, v6
1055 ; CI-NEXT: v_div_fixup_f32 v2, v2, v3, v1
1056 ; CI-NEXT: v_trunc_f32_e32 v2, v2
1057 ; CI-NEXT: v_fma_f32 v1, -v2, v3, v1
1058 ; CI-NEXT: v_cvt_f16_f32_e32 v1, v1
1059 ; CI-NEXT: v_or_b32_e32 v0, v1, v0
1060 ; CI-NEXT: buffer_store_dword v0, off, s[0:3], 0
1061 ; CI-NEXT: s_endpgm
1217 ; CI-LABEL: frem_v4f16:
1218 ; CI: ; %bb.0:
1219 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
1220 ; CI-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0xd
1221 ; CI-NEXT: s_mov_b32 s3, 0xf000
1222 ; CI-NEXT: s_mov_b32 s2, -1
1223 ; CI-NEXT: s_mov_b32 s10, s2
1224 ; CI-NEXT: s_waitcnt lgkmcnt(0)
1225 ; CI-NEXT: s_mov_b32 s0, s4
1226 ; CI-NEXT: s_mov_b32 s1, s5
1227 ; CI-NEXT: s_mov_b32 s4, s6
1228 ; CI-NEXT: s_mov_b32 s5, s7
1229 ; CI-NEXT: s_mov_b32 s6, s2
1230 ; CI-NEXT: s_mov_b32 s7, s3
1231 ; CI-NEXT: buffer_load_dwordx2 v[0:1], off, s[4:7], 0
1232 ; CI-NEXT: s_mov_b32 s11, s3
1233 ; CI-NEXT: s_mov_b32 s6, 3
1234 ; CI-NEXT: s_mov_b32 s7, 0
1235 ; CI-NEXT: s_waitcnt vmcnt(0)
1236 ; CI-NEXT: v_cvt_f32_f16_e32 v2, v0
1237 ; CI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
1238 ; CI-NEXT: v_cvt_f32_f16_e32 v3, v0
1239 ; CI-NEXT: v_lshrrev_b32_e32 v0, 16, v1
1240 ; CI-NEXT: v_cvt_f32_f16_e32 v4, v1
1241 ; CI-NEXT: v_cvt_f32_f16_e32 v5, v0
1242 ; CI-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0 offset:32
1243 ; CI-NEXT: s_waitcnt vmcnt(0)
1244 ; CI-NEXT: v_cvt_f32_f16_e32 v7, v1
1245 ; CI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
1246 ; CI-NEXT: v_cvt_f32_f16_e32 v1, v1
1247 ; CI-NEXT: v_cvt_f32_f16_e32 v6, v0
1248 ; CI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
1249 ; CI-NEXT: v_cvt_f32_f16_e32 v0, v0
1250 ; CI-NEXT: v_div_scale_f32 v9, s[4:5], v1, v1, v5
1251 ; CI-NEXT: v_div_scale_f32 v8, vcc, v5, v1, v5
1252 ; CI-NEXT: v_rcp_f32_e32 v10, v9
1253 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s6
1254 ; CI-NEXT: v_fma_f32 v11, -v9, v10, 1.0
1255 ; CI-NEXT: v_fma_f32 v10, v11, v10, v10
1256 ; CI-NEXT: v_mul_f32_e32 v11, v8, v10
1257 ; CI-NEXT: v_fma_f32 v12, -v9, v11, v8
1258 ; CI-NEXT: v_fma_f32 v11, v12, v10, v11
1259 ; CI-NEXT: v_fma_f32 v8, -v9, v11, v8
1260 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s7
1261 ; CI-NEXT: v_div_fmas_f32 v8, v8, v10, v11
1262 ; CI-NEXT: v_div_fixup_f32 v8, v8, v1, v5
1263 ; CI-NEXT: v_trunc_f32_e32 v8, v8
1264 ; CI-NEXT: v_fma_f32 v1, -v8, v1, v5
1265 ; CI-NEXT: v_div_scale_f32 v8, s[4:5], v7, v7, v4
1266 ; CI-NEXT: v_div_scale_f32 v5, vcc, v4, v7, v4
1267 ; CI-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0
1268 ; CI-NEXT: v_cvt_f16_f32_e32 v1, v1
1269 ; CI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
1270 ; CI-NEXT: v_rcp_f32_e32 v9, v8
1271 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s6
1272 ; CI-NEXT: v_fma_f32 v10, -v8, v9, 1.0
1273 ; CI-NEXT: v_fma_f32 v9, v10, v9, v9
1274 ; CI-NEXT: v_mul_f32_e32 v10, v5, v9
1275 ; CI-NEXT: v_fma_f32 v11, -v8, v10, v5
1276 ; CI-NEXT: v_fma_f32 v10, v11, v9, v10
1277 ; CI-NEXT: v_fma_f32 v5, -v8, v10, v5
1278 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s7
1279 ; CI-NEXT: v_div_fmas_f32 v5, v5, v9, v10
1280 ; CI-NEXT: v_div_fixup_f32 v5, v5, v7, v4
1281 ; CI-NEXT: v_trunc_f32_e32 v5, v5
1282 ; CI-NEXT: v_fma_f32 v4, -v5, v7, v4
1283 ; CI-NEXT: v_div_scale_f32 v5, s[4:5], v0, v0, v3
1284 ; CI-NEXT: v_cvt_f16_f32_e32 v4, v4
1285 ; CI-NEXT: v_or_b32_e32 v1, v4, v1
1286 ; CI-NEXT: v_div_scale_f32 v4, vcc, v3, v0, v3
1287 ; CI-NEXT: v_rcp_f32_e32 v7, v5
1288 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s6
1289 ; CI-NEXT: v_fma_f32 v8, -v5, v7, 1.0
1290 ; CI-NEXT: v_fma_f32 v7, v8, v7, v7
1291 ; CI-NEXT: v_mul_f32_e32 v8, v4, v7
1292 ; CI-NEXT: v_fma_f32 v9, -v5, v8, v4
1293 ; CI-NEXT: v_fma_f32 v8, v9, v7, v8
1294 ; CI-NEXT: v_fma_f32 v4, -v5, v8, v4
1295 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s7
1296 ; CI-NEXT: v_div_fmas_f32 v4, v4, v7, v8
1297 ; CI-NEXT: v_div_fixup_f32 v4, v4, v0, v3
1298 ; CI-NEXT: v_trunc_f32_e32 v4, v4
1299 ; CI-NEXT: v_fma_f32 v0, -v4, v0, v3
1300 ; CI-NEXT: v_div_scale_f32 v4, s[4:5], v6, v6, v2
1301 ; CI-NEXT: v_div_scale_f32 v3, vcc, v2, v6, v2
1302 ; CI-NEXT: v_cvt_f16_f32_e32 v0, v0
1303 ; CI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
1304 ; CI-NEXT: v_rcp_f32_e32 v5, v4
1305 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s6
1306 ; CI-NEXT: v_fma_f32 v7, -v4, v5, 1.0
1307 ; CI-NEXT: v_fma_f32 v5, v7, v5, v5
1308 ; CI-NEXT: v_mul_f32_e32 v7, v3, v5
1309 ; CI-NEXT: v_fma_f32 v8, -v4, v7, v3
1310 ; CI-NEXT: v_fma_f32 v7, v8, v5, v7
1311 ; CI-NEXT: v_fma_f32 v3, -v4, v7, v3
1312 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s7
1313 ; CI-NEXT: v_div_fmas_f32 v3, v3, v5, v7
1314 ; CI-NEXT: v_div_fixup_f32 v3, v3, v6, v2
1315 ; CI-NEXT: v_trunc_f32_e32 v3, v3
1316 ; CI-NEXT: v_fma_f32 v2, -v3, v6, v2
1317 ; CI-NEXT: v_cvt_f16_f32_e32 v2, v2
1318 ; CI-NEXT: v_or_b32_e32 v0, v2, v0
1319 ; CI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
1320 ; CI-NEXT: s_endpgm
1444 ; CI-LABEL: frem_v2f32:
1445 ; CI: ; %bb.0:
1446 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
1447 ; CI-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0xd
1448 ; CI-NEXT: s_mov_b32 s3, 0xf000
1449 ; CI-NEXT: s_mov_b32 s2, -1
1450 ; CI-NEXT: s_mov_b32 s10, s2
1451 ; CI-NEXT: s_waitcnt lgkmcnt(0)
1452 ; CI-NEXT: s_mov_b32 s0, s4
1453 ; CI-NEXT: s_mov_b32 s1, s5
1454 ; CI-NEXT: s_mov_b32 s4, s6
1455 ; CI-NEXT: s_mov_b32 s5, s7
1456 ; CI-NEXT: s_mov_b32 s6, s2
1457 ; CI-NEXT: s_mov_b32 s7, s3
1458 ; CI-NEXT: s_mov_b32 s11, s3
1459 ; CI-NEXT: buffer_load_dwordx2 v[0:1], off, s[4:7], 0
1460 ; CI-NEXT: buffer_load_dwordx2 v[2:3], off, s[8:11], 0 offset:32
1461 ; CI-NEXT: s_mov_b32 s6, 3
1462 ; CI-NEXT: s_mov_b32 s7, 0
1463 ; CI-NEXT: s_waitcnt vmcnt(0)
1464 ; CI-NEXT: v_div_scale_f32 v5, s[4:5], v3, v3, v1
1465 ; CI-NEXT: v_div_scale_f32 v4, vcc, v1, v3, v1
1466 ; CI-NEXT: v_rcp_f32_e32 v6, v5
1467 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s6
1468 ; CI-NEXT: v_fma_f32 v7, -v5, v6, 1.0
1469 ; CI-NEXT: v_fma_f32 v6, v7, v6, v6
1470 ; CI-NEXT: v_mul_f32_e32 v7, v4, v6
1471 ; CI-NEXT: v_fma_f32 v8, -v5, v7, v4
1472 ; CI-NEXT: v_fma_f32 v7, v8, v6, v7
1473 ; CI-NEXT: v_fma_f32 v4, -v5, v7, v4
1474 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s7
1475 ; CI-NEXT: v_div_fmas_f32 v4, v4, v6, v7
1476 ; CI-NEXT: v_div_fixup_f32 v4, v4, v3, v1
1477 ; CI-NEXT: v_trunc_f32_e32 v4, v4
1478 ; CI-NEXT: v_fma_f32 v1, -v4, v3, v1
1479 ; CI-NEXT: v_div_scale_f32 v4, s[4:5], v2, v2, v0
1480 ; CI-NEXT: v_div_scale_f32 v3, vcc, v0, v2, v0
1481 ; CI-NEXT: v_rcp_f32_e32 v5, v4
1482 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s6
1483 ; CI-NEXT: v_fma_f32 v6, -v4, v5, 1.0
1484 ; CI-NEXT: v_fma_f32 v5, v6, v5, v5
1485 ; CI-NEXT: v_mul_f32_e32 v6, v3, v5
1486 ; CI-NEXT: v_fma_f32 v7, -v4, v6, v3
1487 ; CI-NEXT: v_fma_f32 v6, v7, v5, v6
1488 ; CI-NEXT: v_fma_f32 v3, -v4, v6, v3
1489 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s7
1490 ; CI-NEXT: v_div_fmas_f32 v3, v3, v5, v6
1491 ; CI-NEXT: v_div_fixup_f32 v3, v3, v2, v0
1492 ; CI-NEXT: v_trunc_f32_e32 v3, v3
1493 ; CI-NEXT: v_fma_f32 v0, -v3, v2, v0
1494 ; CI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
1495 ; CI-NEXT: s_endpgm
1640 ; CI-LABEL: frem_v4f32:
1641 ; CI: ; %bb.0:
1642 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
1643 ; CI-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0xd
1644 ; CI-NEXT: s_mov_b32 s3, 0xf000
1645 ; CI-NEXT: s_mov_b32 s2, -1
1646 ; CI-NEXT: s_mov_b32 s10, s2
1647 ; CI-NEXT: s_waitcnt lgkmcnt(0)
1648 ; CI-NEXT: s_mov_b32 s0, s4
1649 ; CI-NEXT: s_mov_b32 s1, s5
1650 ; CI-NEXT: s_mov_b32 s4, s6
1651 ; CI-NEXT: s_mov_b32 s5, s7
1652 ; CI-NEXT: s_mov_b32 s6, s2
1653 ; CI-NEXT: s_mov_b32 s7, s3
1654 ; CI-NEXT: s_mov_b32 s11, s3
1655 ; CI-NEXT: buffer_load_dwordx4 v[0:3], off, s[4:7], 0
1656 ; CI-NEXT: buffer_load_dwordx4 v[4:7], off, s[8:11], 0 offset:64
1657 ; CI-NEXT: s_mov_b32 s6, 3
1658 ; CI-NEXT: s_mov_b32 s7, 0
1659 ; CI-NEXT: s_waitcnt vmcnt(0)
1660 ; CI-NEXT: v_div_scale_f32 v9, s[4:5], v7, v7, v3
1661 ; CI-NEXT: v_div_scale_f32 v8, vcc, v3, v7, v3
1662 ; CI-NEXT: v_rcp_f32_e32 v10, v9
1663 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s6
1664 ; CI-NEXT: v_fma_f32 v11, -v9, v10, 1.0
1665 ; CI-NEXT: v_fma_f32 v10, v11, v10, v10
1666 ; CI-NEXT: v_mul_f32_e32 v11, v8, v10
1667 ; CI-NEXT: v_fma_f32 v12, -v9, v11, v8
1668 ; CI-NEXT: v_fma_f32 v11, v12, v10, v11
1669 ; CI-NEXT: v_fma_f32 v8, -v9, v11, v8
1670 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s7
1671 ; CI-NEXT: v_div_fmas_f32 v8, v8, v10, v11
1672 ; CI-NEXT: v_div_fixup_f32 v8, v8, v7, v3
1673 ; CI-NEXT: v_trunc_f32_e32 v8, v8
1674 ; CI-NEXT: v_fma_f32 v3, -v8, v7, v3
1675 ; CI-NEXT: v_div_scale_f32 v8, s[4:5], v6, v6, v2
1676 ; CI-NEXT: v_div_scale_f32 v7, vcc, v2, v6, v2
1677 ; CI-NEXT: v_rcp_f32_e32 v9, v8
1678 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s6
1679 ; CI-NEXT: v_fma_f32 v10, -v8, v9, 1.0
1680 ; CI-NEXT: v_fma_f32 v9, v10, v9, v9
1681 ; CI-NEXT: v_mul_f32_e32 v10, v7, v9
1682 ; CI-NEXT: v_fma_f32 v11, -v8, v10, v7
1683 ; CI-NEXT: v_fma_f32 v10, v11, v9, v10
1684 ; CI-NEXT: v_fma_f32 v7, -v8, v10, v7
1685 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s7
1686 ; CI-NEXT: v_div_fmas_f32 v7, v7, v9, v10
1687 ; CI-NEXT: v_div_fixup_f32 v7, v7, v6, v2
1688 ; CI-NEXT: v_trunc_f32_e32 v7, v7
1689 ; CI-NEXT: v_fma_f32 v2, -v7, v6, v2
1690 ; CI-NEXT: v_div_scale_f32 v7, s[4:5], v5, v5, v1
1691 ; CI-NEXT: v_div_scale_f32 v6, vcc, v1, v5, v1
1692 ; CI-NEXT: v_rcp_f32_e32 v8, v7
1693 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s6
1694 ; CI-NEXT: v_fma_f32 v9, -v7, v8, 1.0
1695 ; CI-NEXT: v_fma_f32 v8, v9, v8, v8
1696 ; CI-NEXT: v_mul_f32_e32 v9, v6, v8
1697 ; CI-NEXT: v_fma_f32 v10, -v7, v9, v6
1698 ; CI-NEXT: v_fma_f32 v9, v10, v8, v9
1699 ; CI-NEXT: v_fma_f32 v6, -v7, v9, v6
1700 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s7
1701 ; CI-NEXT: v_div_fmas_f32 v6, v6, v8, v9
1702 ; CI-NEXT: v_div_fixup_f32 v6, v6, v5, v1
1703 ; CI-NEXT: v_trunc_f32_e32 v6, v6
1704 ; CI-NEXT: v_fma_f32 v1, -v6, v5, v1
1705 ; CI-NEXT: v_div_scale_f32 v6, s[4:5], v4, v4, v0
1706 ; CI-NEXT: v_div_scale_f32 v5, vcc, v0, v4, v0
1707 ; CI-NEXT: v_rcp_f32_e32 v7, v6
1708 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s6
1709 ; CI-NEXT: v_fma_f32 v8, -v6, v7, 1.0
1710 ; CI-NEXT: v_fma_f32 v7, v8, v7, v7
1711 ; CI-NEXT: v_mul_f32_e32 v8, v5, v7
1712 ; CI-NEXT: v_fma_f32 v9, -v6, v8, v5
1713 ; CI-NEXT: v_fma_f32 v8, v9, v7, v8
1714 ; CI-NEXT: v_fma_f32 v5, -v6, v8, v5
1715 ; CI-NEXT: s_setreg_b32 hwreg(HW_REG_MODE, 4, 2), s7
1716 ; CI-NEXT: v_div_fmas_f32 v5, v5, v7, v8
1717 ; CI-NEXT: v_div_fixup_f32 v5, v5, v4, v0
1718 ; CI-NEXT: v_trunc_f32_e32 v5, v5
1719 ; CI-NEXT: v_fma_f32 v0, -v5, v4, v0
1720 ; CI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
1721 ; CI-NEXT: s_endpgm
1897 ; CI-LABEL: frem_v2f64:
1898 ; CI: ; %bb.0:
1899 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
1900 ; CI-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0xd
1901 ; CI-NEXT: s_mov_b32 s3, 0xf000
1902 ; CI-NEXT: s_mov_b32 s2, -1
1903 ; CI-NEXT: s_mov_b32 s10, s2
1904 ; CI-NEXT: s_waitcnt lgkmcnt(0)
1905 ; CI-NEXT: s_mov_b32 s0, s4
1906 ; CI-NEXT: s_mov_b32 s1, s5
1907 ; CI-NEXT: s_mov_b32 s4, s6
1908 ; CI-NEXT: s_mov_b32 s5, s7
1909 ; CI-NEXT: s_mov_b32 s6, s2
1910 ; CI-NEXT: s_mov_b32 s7, s3
1911 ; CI-NEXT: s_mov_b32 s11, s3
1912 ; CI-NEXT: buffer_load_dwordx4 v[0:3], off, s[4:7], 0
1913 ; CI-NEXT: buffer_load_dwordx4 v[4:7], off, s[8:11], 0 offset:64
1914 ; CI-NEXT: s_waitcnt vmcnt(0)
1915 ; CI-NEXT: v_div_scale_f64 v[8:9], s[4:5], v[6:7], v[6:7], v[2:3]
1916 ; CI-NEXT: v_rcp_f64_e32 v[10:11], v[8:9]
1917 ; CI-NEXT: v_fma_f64 v[12:13], -v[8:9], v[10:11], 1.0
1918 ; CI-NEXT: v_fma_f64 v[10:11], v[10:11], v[12:13], v[10:11]
1919 ; CI-NEXT: v_fma_f64 v[12:13], -v[8:9], v[10:11], 1.0
1920 ; CI-NEXT: v_fma_f64 v[10:11], v[10:11], v[12:13], v[10:11]
1921 ; CI-NEXT: v_div_scale_f64 v[12:13], vcc, v[2:3], v[6:7], v[2:3]
1922 ; CI-NEXT: v_mul_f64 v[14:15], v[12:13], v[10:11]
1923 ; CI-NEXT: v_fma_f64 v[8:9], -v[8:9], v[14:15], v[12:13]
1924 ; CI-NEXT: s_nop 1
1925 ; CI-NEXT: v_div_fmas_f64 v[8:9], v[8:9], v[10:11], v[14:15]
1926 ; CI-NEXT: v_div_fixup_f64 v[8:9], v[8:9], v[6:7], v[2:3]
1927 ; CI-NEXT: v_trunc_f64_e32 v[8:9], v[8:9]
1928 ; CI-NEXT: v_fma_f64 v[2:3], -v[8:9], v[6:7], v[2:3]
1929 ; CI-NEXT: v_div_scale_f64 v[6:7], s[4:5], v[4:5], v[4:5], v[0:1]
1930 ; CI-NEXT: v_rcp_f64_e32 v[8:9], v[6:7]
1931 ; CI-NEXT: v_fma_f64 v[10:11], -v[6:7], v[8:9], 1.0
1932 ; CI-NEXT: v_fma_f64 v[8:9], v[8:9], v[10:11], v[8:9]
1933 ; CI-NEXT: v_fma_f64 v[10:11], -v[6:7], v[8:9], 1.0
1934 ; CI-NEXT: v_fma_f64 v[8:9], v[8:9], v[10:11], v[8:9]
1935 ; CI-NEXT: v_div_scale_f64 v[10:11], vcc, v[0:1], v[4:5], v[0:1]
1936 ; CI-NEXT: v_mul_f64 v[12:13], v[10:11], v[8:9]
1937 ; CI-NEXT: v_fma_f64 v[6:7], -v[6:7], v[12:13], v[10:11]
1938 ; CI-NEXT: s_nop 1
1939 ; CI-NEXT: v_div_fmas_f64 v[6:7], v[6:7], v[8:9], v[12:13]
1940 ; CI-NEXT: v_div_fixup_f64 v[6:7], v[6:7], v[4:5], v[0:1]
1941 ; CI-NEXT: v_trunc_f64_e32 v[6:7], v[6:7]
1942 ; CI-NEXT: v_fma_f64 v[0:1], -v[6:7], v[4:5], v[0:1]
1943 ; CI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
1944 ; CI-NEXT: s_endpgm