diff options
Diffstat (limited to 'ref-rvct-neon-nofp16.txt')
| -rw-r--r-- | ref-rvct-neon-nofp16.txt | 1528 |
1 files changed, 764 insertions, 764 deletions
diff --git a/ref-rvct-neon-nofp16.txt b/ref-rvct-neon-nofp16.txt index 6e6d1cb..12c4c5a 100644 --- a/ref-rvct-neon-nofp16.txt +++ b/ref-rvct-neon-nofp16.txt @@ -339,9 +339,9 @@ VGET_LOW:19:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VGET_LOW:20:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VGET_LOW:21:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMLAL_LANE overflow output: -VQDMLAL_LANE:0:vqdmlal_lane_s16 Neon overflow 0 -VQDMLAL_LANE:1:vqdmlal_lane_s32 Neon overflow 0 +VQDMLAL_LANE cumulative saturation output: +VQDMLAL_LANE:0:vqdmlal_lane_s16 Neon cumulative saturation 0 +VQDMLAL_LANE:1:vqdmlal_lane_s32 Neon cumulative saturation 0 VQDMLAL_LANE output: VQDMLAL_LANE:2:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -367,9 +367,9 @@ VQDMLAL_LANE:21:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQDMLAL_LANE:22:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMLAL_LANE:23:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMLAL_LANE (mul with input=0) overflow output: -VQDMLAL_LANE:24:vqdmlal_lane_s16 Neon overflow 0 -VQDMLAL_LANE:25:vqdmlal_lane_s32 Neon overflow 0 +VQDMLAL_LANE (mul with input=0) cumulative saturation output: +VQDMLAL_LANE:24:vqdmlal_lane_s16 Neon cumulative saturation 0 +VQDMLAL_LANE:25:vqdmlal_lane_s32 Neon cumulative saturation 0 VQDMLAL_LANE (mul with input=0) output: VQDMLAL_LANE:26:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -395,11 +395,11 @@ VQDMLAL_LANE:45:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQDMLAL_LANE:46:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMLAL_LANE:47:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMLAL_LANE (check mul overflow) overflow output: -VQDMLAL_LANE:48:vqdmlal_lane_s16 Neon overflow 1 -VQDMLAL_LANE:49:vqdmlal_lane_s32 Neon overflow 1 +VQDMLAL_LANE (check mul cumulative saturation) cumulative saturation output: +VQDMLAL_LANE:48:vqdmlal_lane_s16 Neon cumulative saturation 1 +VQDMLAL_LANE:49:vqdmlal_lane_s32 Neon cumulative saturation 1 -VQDMLAL_LANE (check mul overflow) output: +VQDMLAL_LANE (check mul cumulative saturation) output: VQDMLAL_LANE:50:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMLAL_LANE:51:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQDMLAL_LANE:52:result_int32x2 [] = { 33333333, 33333333, } @@ -423,9 +423,9 @@ VQDMLAL_LANE:69:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQDMLAL_LANE:70:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMLAL_LANE:71:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMLSL_LANE overflow output: -VQDMLSL_LANE:0:vqdmlsl_lane_s16 Neon overflow 0 -VQDMLSL_LANE:1:vqdmlsl_lane_s32 Neon overflow 0 +VQDMLSL_LANE cumulative saturation output: +VQDMLSL_LANE:0:vqdmlsl_lane_s16 Neon cumulative saturation 0 +VQDMLSL_LANE:1:vqdmlsl_lane_s32 Neon cumulative saturation 0 VQDMLSL_LANE output: VQDMLSL_LANE:2:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -451,9 +451,9 @@ VQDMLSL_LANE:21:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQDMLSL_LANE:22:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMLSL_LANE:23:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMLSL_LANE (mul with input=0) overflow output: -VQDMLSL_LANE:24:vqdmlsl_lane_s16 Neon overflow 0 -VQDMLSL_LANE:25:vqdmlsl_lane_s32 Neon overflow 0 +VQDMLSL_LANE (mul with input=0) cumulative saturation output: +VQDMLSL_LANE:24:vqdmlsl_lane_s16 Neon cumulative saturation 0 +VQDMLSL_LANE:25:vqdmlsl_lane_s32 Neon cumulative saturation 0 VQDMLSL_LANE (mul with input=0) output: VQDMLSL_LANE:26:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -479,11 +479,11 @@ VQDMLSL_LANE:45:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQDMLSL_LANE:46:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMLSL_LANE:47:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMLSL_LANE (check mul overflow) overflow output: -VQDMLSL_LANE:48:vqdmlsl_lane_s16 Neon overflow 1 -VQDMLSL_LANE:49:vqdmlsl_lane_s32 Neon overflow 1 +VQDMLSL_LANE (check mul cumulative saturation) cumulative saturation output: +VQDMLSL_LANE:48:vqdmlsl_lane_s16 Neon cumulative saturation 1 +VQDMLSL_LANE:49:vqdmlsl_lane_s32 Neon cumulative saturation 1 -VQDMLSL_LANE (check mul overflow) output: +VQDMLSL_LANE (check mul cumulative saturation) output: VQDMLSL_LANE:50:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMLSL_LANE:51:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQDMLSL_LANE:52:result_int32x2 [] = { 33333333, 33333333, } @@ -507,9 +507,9 @@ VQDMLSL_LANE:69:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQDMLSL_LANE:70:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMLSL_LANE:71:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMLAL_N overflow output: -VQDMLAL_N:0:vqdmlal_n_s16 Neon overflow 0 -VQDMLAL_N:1:vqdmlal_n_s32 Neon overflow 0 +VQDMLAL_N cumulative saturation output: +VQDMLAL_N:0:vqdmlal_n_s16 Neon cumulative saturation 0 +VQDMLAL_N:1:vqdmlal_n_s32 Neon cumulative saturation 0 VQDMLAL_N output: VQDMLAL_N:2:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -535,11 +535,11 @@ VQDMLAL_N:21:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQDMLAL_N:22:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMLAL_N:23:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMLAL_N (check mul overflow) overflow output: -VQDMLAL_N:24:vqdmlal_n_s16 Neon overflow 1 -VQDMLAL_N:25:vqdmlal_n_s32 Neon overflow 1 +VQDMLAL_N (check mul cumulative saturation) cumulative saturation output: +VQDMLAL_N:24:vqdmlal_n_s16 Neon cumulative saturation 1 +VQDMLAL_N:25:vqdmlal_n_s32 Neon cumulative saturation 1 -VQDMLAL_N (check mul overflow) output: +VQDMLAL_N (check mul cumulative saturation) output: VQDMLAL_N:26:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMLAL_N:27:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQDMLAL_N:28:result_int32x2 [] = { 33333333, 33333333, } @@ -563,9 +563,9 @@ VQDMLAL_N:45:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQDMLAL_N:46:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMLAL_N:47:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMLSL_N overflow output: -VQDMLSL_N:0:vqdmlsl_n_s16 Neon overflow 0 -VQDMLSL_N:1:vqdmlsl_n_s32 Neon overflow 0 +VQDMLSL_N cumulative saturation output: +VQDMLSL_N:0:vqdmlsl_n_s16 Neon cumulative saturation 0 +VQDMLSL_N:1:vqdmlsl_n_s32 Neon cumulative saturation 0 VQDMLSL_N output: VQDMLSL_N:2:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -591,11 +591,11 @@ VQDMLSL_N:21:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQDMLSL_N:22:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMLSL_N:23:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMLSL_N (check mul overflow) overflow output: -VQDMLSL_N:24:vqdmlsl_n_s16 Neon overflow 1 -VQDMLSL_N:25:vqdmlsl_n_s32 Neon overflow 1 +VQDMLSL_N (check mul cumulative saturation) cumulative saturation output: +VQDMLSL_N:24:vqdmlsl_n_s16 Neon cumulative saturation 1 +VQDMLSL_N:25:vqdmlsl_n_s32 Neon cumulative saturation 1 -VQDMLSL_N (check mul overflow) output: +VQDMLSL_N (check mul cumulative saturation) output: VQDMLSL_N:26:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMLSL_N:27:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQDMLSL_N:28:result_int32x2 [] = { 33333333, 33333333, } @@ -763,13 +763,13 @@ VRSHRN_N:63:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VRSHRN_N:64:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VRSHRN_N:65:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHRN_N overflow output: -VQRSHRN_N:0:vqrshrn_n_s16 Neon overflow 0 -VQRSHRN_N:1:vqrshrn_n_s32 Neon overflow 0 -VQRSHRN_N:2:vqrshrn_n_s64 Neon overflow 0 -VQRSHRN_N:3:vqrshrn_n_u16 Neon overflow 1 -VQRSHRN_N:4:vqrshrn_n_u32 Neon overflow 1 -VQRSHRN_N:5:vqrshrn_n_u64 Neon overflow 1 +VQRSHRN_N cumulative saturation output: +VQRSHRN_N:0:vqrshrn_n_s16 Neon cumulative saturation 0 +VQRSHRN_N:1:vqrshrn_n_s32 Neon cumulative saturation 0 +VQRSHRN_N:2:vqrshrn_n_s64 Neon cumulative saturation 0 +VQRSHRN_N:3:vqrshrn_n_u16 Neon cumulative saturation 1 +VQRSHRN_N:4:vqrshrn_n_u32 Neon cumulative saturation 1 +VQRSHRN_N:5:vqrshrn_n_u64 Neon cumulative saturation 1 VQRSHRN_N output: VQRSHRN_N:6:result_int8x8 [] = { fffffff8, fffffff9, fffffff9, fffffffa, fffffffa, fffffffb, fffffffb, fffffffc, } @@ -795,13 +795,13 @@ VQRSHRN_N:25:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQRSHRN_N:26:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRSHRN_N:27:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHRN_N (check saturation: shift by 3) overflow output: -VQRSHRN_N:28:vqrshrn_n_s16 Neon overflow 1 -VQRSHRN_N:29:vqrshrn_n_s32 Neon overflow 1 -VQRSHRN_N:30:vqrshrn_n_s64 Neon overflow 1 -VQRSHRN_N:31:vqrshrn_n_u16 Neon overflow 1 -VQRSHRN_N:32:vqrshrn_n_u32 Neon overflow 1 -VQRSHRN_N:33:vqrshrn_n_u64 Neon overflow 1 +VQRSHRN_N (check saturation: shift by 3) cumulative saturation output: +VQRSHRN_N:28:vqrshrn_n_s16 Neon cumulative saturation 1 +VQRSHRN_N:29:vqrshrn_n_s32 Neon cumulative saturation 1 +VQRSHRN_N:30:vqrshrn_n_s64 Neon cumulative saturation 1 +VQRSHRN_N:31:vqrshrn_n_u16 Neon cumulative saturation 1 +VQRSHRN_N:32:vqrshrn_n_u32 Neon cumulative saturation 1 +VQRSHRN_N:33:vqrshrn_n_u64 Neon cumulative saturation 1 VQRSHRN_N (check saturation: shift by 3) output: VQRSHRN_N:34:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -827,13 +827,13 @@ VQRSHRN_N:53:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQRSHRN_N:54:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRSHRN_N:55:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHRN_N (check saturation: shift by max) overflow output: -VQRSHRN_N:56:vqrshrn_n_s16 Neon overflow 1 -VQRSHRN_N:57:vqrshrn_n_s32 Neon overflow 1 -VQRSHRN_N:58:vqrshrn_n_s64 Neon overflow 1 -VQRSHRN_N:59:vqrshrn_n_u16 Neon overflow 1 -VQRSHRN_N:60:vqrshrn_n_u32 Neon overflow 1 -VQRSHRN_N:61:vqrshrn_n_u64 Neon overflow 1 +VQRSHRN_N (check saturation: shift by max) cumulative saturation output: +VQRSHRN_N:56:vqrshrn_n_s16 Neon cumulative saturation 1 +VQRSHRN_N:57:vqrshrn_n_s32 Neon cumulative saturation 1 +VQRSHRN_N:58:vqrshrn_n_s64 Neon cumulative saturation 1 +VQRSHRN_N:59:vqrshrn_n_u16 Neon cumulative saturation 1 +VQRSHRN_N:60:vqrshrn_n_u32 Neon cumulative saturation 1 +VQRSHRN_N:61:vqrshrn_n_u64 Neon cumulative saturation 1 VQRSHRN_N (check saturation: shift by max) output: VQRSHRN_N:62:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -908,23 +908,23 @@ vgetq_lane_p16: fff6 vgetq_lane_f32: c1500000 -VQSUB/VQSUBQ overflow output: -VQSUB/VQSUBQ:0:vqsub_s8 Neon overflow 0 -VQSUB/VQSUBQ:1:vqsub_s16 Neon overflow 0 -VQSUB/VQSUBQ:2:vqsub_s32 Neon overflow 0 -VQSUB/VQSUBQ:3:vqsub_s64 Neon overflow 0 -VQSUB/VQSUBQ:4:vqsub_u8 Neon overflow 0 -VQSUB/VQSUBQ:5:vqsub_u16 Neon overflow 0 -VQSUB/VQSUBQ:6:vqsub_u32 Neon overflow 0 -VQSUB/VQSUBQ:7:vqsub_u64 Neon overflow 0 -VQSUB/VQSUBQ:8:vqsubq_s8 Neon overflow 0 -VQSUB/VQSUBQ:9:vqsubq_s16 Neon overflow 0 -VQSUB/VQSUBQ:10:vqsubq_s32 Neon overflow 0 -VQSUB/VQSUBQ:11:vqsubq_s64 Neon overflow 0 -VQSUB/VQSUBQ:12:vqsubq_u8 Neon overflow 0 -VQSUB/VQSUBQ:13:vqsubq_u16 Neon overflow 0 -VQSUB/VQSUBQ:14:vqsubq_u32 Neon overflow 0 -VQSUB/VQSUBQ:15:vqsubq_u64 Neon overflow 0 +VQSUB/VQSUBQ cumulative saturation output: +VQSUB/VQSUBQ:0:vqsub_s8 Neon cumulative saturation 0 +VQSUB/VQSUBQ:1:vqsub_s16 Neon cumulative saturation 0 +VQSUB/VQSUBQ:2:vqsub_s32 Neon cumulative saturation 0 +VQSUB/VQSUBQ:3:vqsub_s64 Neon cumulative saturation 0 +VQSUB/VQSUBQ:4:vqsub_u8 Neon cumulative saturation 0 +VQSUB/VQSUBQ:5:vqsub_u16 Neon cumulative saturation 0 +VQSUB/VQSUBQ:6:vqsub_u32 Neon cumulative saturation 0 +VQSUB/VQSUBQ:7:vqsub_u64 Neon cumulative saturation 0 +VQSUB/VQSUBQ:8:vqsubq_s8 Neon cumulative saturation 0 +VQSUB/VQSUBQ:9:vqsubq_s16 Neon cumulative saturation 0 +VQSUB/VQSUBQ:10:vqsubq_s32 Neon cumulative saturation 0 +VQSUB/VQSUBQ:11:vqsubq_s64 Neon cumulative saturation 0 +VQSUB/VQSUBQ:12:vqsubq_u8 Neon cumulative saturation 0 +VQSUB/VQSUBQ:13:vqsubq_u16 Neon cumulative saturation 0 +VQSUB/VQSUBQ:14:vqsubq_u32 Neon cumulative saturation 0 +VQSUB/VQSUBQ:15:vqsubq_u64 Neon cumulative saturation 0 VQSUB/VQSUBQ output: VQSUB/VQSUBQ:16:result_int8x8 [] = { ffffffdf, ffffffe0, ffffffe1, ffffffe2, ffffffe3, ffffffe4, ffffffe5, ffffffe6, } @@ -950,11 +950,11 @@ VQSUB/VQSUBQ:35:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQSUB/VQSUBQ:36:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSUB/VQSUBQ:37:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSUB/VQSUBQ 64 bits saturation overflow output: -VQSUB/VQSUBQ:38:vqsub_s64 Neon overflow 0 -VQSUB/VQSUBQ:39:vqsub_u64 Neon overflow 0 -VQSUB/VQSUBQ:40:vqsubq_s64 Neon overflow 0 -VQSUB/VQSUBQ:41:vqsubq_u64 Neon overflow 0 +VQSUB/VQSUBQ 64 bits saturation cumulative saturation output: +VQSUB/VQSUBQ:38:vqsub_s64 Neon cumulative saturation 0 +VQSUB/VQSUBQ:39:vqsub_u64 Neon cumulative saturation 0 +VQSUB/VQSUBQ:40:vqsubq_s64 Neon cumulative saturation 0 +VQSUB/VQSUBQ:41:vqsubq_u64 Neon cumulative saturation 0 64 bits saturation: VQSUB/VQSUBQ:42:result_int64x1 [] = { fffffffffffffff0, } @@ -962,33 +962,33 @@ VQSUB/VQSUBQ:43:result_uint64x1 [] = { fffffffffffffff0, } VQSUB/VQSUBQ:44:result_int64x2 [] = { fffffffffffffff0, fffffffffffffff1, } VQSUB/VQSUBQ:45:result_uint64x2 [] = { fffffffffffffff0, fffffffffffffff1, } -VQSUB/VQSUBQ 64 bits saturation overflow output: -VQSUB/VQSUBQ:46:vqsub_s64 Neon overflow 0 -VQSUB/VQSUBQ:47:vqsub_u64 Neon overflow 0 -VQSUB/VQSUBQ:48:vqsubq_s64 Neon overflow 0 -VQSUB/VQSUBQ:49:vqsubq_u64 Neon overflow 0 +VQSUB/VQSUBQ 64 bits saturation cumulative saturation output: +VQSUB/VQSUBQ:46:vqsub_s64 Neon cumulative saturation 0 +VQSUB/VQSUBQ:47:vqsub_u64 Neon cumulative saturation 0 +VQSUB/VQSUBQ:48:vqsubq_s64 Neon cumulative saturation 0 +VQSUB/VQSUBQ:49:vqsubq_u64 Neon cumulative saturation 0 VQSUB/VQSUBQ:50:result_int64x1 [] = { ffffffffffffffac, } VQSUB/VQSUBQ:51:result_uint64x1 [] = { ffffffffffffff68, } VQSUB/VQSUBQ:52:result_int64x2 [] = { ffffffffffffffac, ffffffffffffffad, } VQSUB/VQSUBQ:53:result_uint64x2 [] = { ffffffffffffff68, ffffffffffffff69, } -VQSUB/VQSUBQ 64 bits saturation overflow output: -VQSUB/VQSUBQ:54:vqsub_s64 Neon overflow 1 -VQSUB/VQSUBQ:55:vqsub_u64 Neon overflow 1 -VQSUB/VQSUBQ:56:vqsubq_s64 Neon overflow 1 -VQSUB/VQSUBQ:57:vqsubq_u64 Neon overflow 1 +VQSUB/VQSUBQ 64 bits saturation cumulative saturation output: +VQSUB/VQSUBQ:54:vqsub_s64 Neon cumulative saturation 1 +VQSUB/VQSUBQ:55:vqsub_u64 Neon cumulative saturation 1 +VQSUB/VQSUBQ:56:vqsubq_s64 Neon cumulative saturation 1 +VQSUB/VQSUBQ:57:vqsubq_u64 Neon cumulative saturation 1 VQSUB/VQSUBQ:58:result_int64x1 [] = { 8000000000000000, } VQSUB/VQSUBQ:59:result_uint64x1 [] = { 0, } VQSUB/VQSUBQ:60:result_int64x2 [] = { 7fffffffffffffff, 7fffffffffffffff, } VQSUB/VQSUBQ:61:result_uint64x2 [] = { 0, 0, } less than 64 bits saturation: -VQSUB/VQSUBQ:62:vqsub_s8 Neon overflow 1 -VQSUB/VQSUBQ:63:vqsub_s16 Neon overflow 1 -VQSUB/VQSUBQ:64:vqsub_s32 Neon overflow 1 -VQSUB/VQSUBQ:65:vqsubq_s8 Neon overflow 1 -VQSUB/VQSUBQ:66:vqsubq_s16 Neon overflow 1 -VQSUB/VQSUBQ:67:vqsubq_s32 Neon overflow 1 +VQSUB/VQSUBQ:62:vqsub_s8 Neon cumulative saturation 1 +VQSUB/VQSUBQ:63:vqsub_s16 Neon cumulative saturation 1 +VQSUB/VQSUBQ:64:vqsub_s32 Neon cumulative saturation 1 +VQSUB/VQSUBQ:65:vqsubq_s8 Neon cumulative saturation 1 +VQSUB/VQSUBQ:66:vqsubq_s16 Neon cumulative saturation 1 +VQSUB/VQSUBQ:67:vqsubq_s32 Neon cumulative saturation 1 VQSUB/VQSUBQ:68:result_int8x8 [] = { ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, } VQSUB/VQSUBQ:69:result_int16x4 [] = { ffff8000, ffff8000, ffff8000, ffff8000, } VQSUB/VQSUBQ:70:result_int32x2 [] = { 80000000, 80000000, } @@ -996,13 +996,13 @@ VQSUB/VQSUBQ:71:result_int8x16 [] = { ffffff80, ffffff80, ffffff80, ffffff80, ff VQSUB/VQSUBQ:72:result_int16x8 [] = { ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, } VQSUB/VQSUBQ:73:result_int32x4 [] = { 80000000, 80000000, 80000000, 80000000, } -VQSUB/VQSUBQ less than 64 bits saturation overflow output: -VQSUB/VQSUBQ:74:vqsub_u8 Neon overflow 1 -VQSUB/VQSUBQ:75:vqsub_u16 Neon overflow 1 -VQSUB/VQSUBQ:76:vqsub_u32 Neon overflow 1 -VQSUB/VQSUBQ:77:vqsubq_u8 Neon overflow 1 -VQSUB/VQSUBQ:78:vqsubq_u16 Neon overflow 1 -VQSUB/VQSUBQ:79:vqsubq_u32 Neon overflow 1 +VQSUB/VQSUBQ less than 64 bits saturation cumulative saturation output: +VQSUB/VQSUBQ:74:vqsub_u8 Neon cumulative saturation 1 +VQSUB/VQSUBQ:75:vqsub_u16 Neon cumulative saturation 1 +VQSUB/VQSUBQ:76:vqsub_u32 Neon cumulative saturation 1 +VQSUB/VQSUBQ:77:vqsubq_u8 Neon cumulative saturation 1 +VQSUB/VQSUBQ:78:vqsubq_u16 Neon cumulative saturation 1 +VQSUB/VQSUBQ:79:vqsubq_u32 Neon cumulative saturation 1 VQSUB/VQSUBQ:80:result_uint8x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } VQSUB/VQSUBQ:81:result_uint16x4 [] = { 0, 0, 0, 0, } VQSUB/VQSUBQ:82:result_uint32x2 [] = { 0, 0, } @@ -1010,11 +1010,11 @@ VQSUB/VQSUBQ:83:result_uint8x16 [] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, VQSUB/VQSUBQ:84:result_uint16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } VQSUB/VQSUBQ:85:result_uint32x4 [] = { 0, 0, 0, 0, } -VQDMULH overflow output: -VQDMULH:0:vqdmulh_s16 Neon overflow 0 -VQDMULH:1:vqdmulh_s32 Neon overflow 0 -VQDMULH:2:vqdmulhq_s16 Neon overflow 0 -VQDMULH:3:vqdmulhq_s32 Neon overflow 0 +VQDMULH cumulative saturation output: +VQDMULH:0:vqdmulh_s16 Neon cumulative saturation 0 +VQDMULH:1:vqdmulh_s32 Neon cumulative saturation 0 +VQDMULH:2:vqdmulhq_s16 Neon cumulative saturation 0 +VQDMULH:3:vqdmulhq_s32 Neon cumulative saturation 0 VQDMULH output: VQDMULH:4:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -1040,11 +1040,11 @@ VQDMULH:23:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33 VQDMULH:24:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMULH:25:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMULH overflow output: -VQDMULH:26:vqdmulh_s16 Neon overflow 1 -VQDMULH:27:vqdmulh_s32 Neon overflow 1 -VQDMULH:28:vqdmulhq_s16 Neon overflow 1 -VQDMULH:29:vqdmulhq_s32 Neon overflow 1 +VQDMULH cumulative saturation output: +VQDMULH:26:vqdmulh_s16 Neon cumulative saturation 1 +VQDMULH:27:vqdmulh_s32 Neon cumulative saturation 1 +VQDMULH:28:vqdmulhq_s16 Neon cumulative saturation 1 +VQDMULH:29:vqdmulhq_s32 Neon cumulative saturation 1 VQDMULH output: VQDMULH:30:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -1070,11 +1070,11 @@ VQDMULH:49:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33 VQDMULH:50:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMULH:51:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMULH_LANE overflow output: -VQDMULH_LANE:0:vqdmulh_lane_s16 Neon overflow 0 -VQDMULH_LANE:1:vqdmulh_lane_s32 Neon overflow 0 -VQDMULH_LANE:2:vqdmulhq_lane_s16 Neon overflow 0 -VQDMULH_LANE:3:vqdmulhq_lane_s32 Neon overflow 0 +VQDMULH_LANE cumulative saturation output: +VQDMULH_LANE:0:vqdmulh_lane_s16 Neon cumulative saturation 0 +VQDMULH_LANE:1:vqdmulh_lane_s32 Neon cumulative saturation 0 +VQDMULH_LANE:2:vqdmulhq_lane_s16 Neon cumulative saturation 0 +VQDMULH_LANE:3:vqdmulhq_lane_s32 Neon cumulative saturation 0 VQDMULH_LANE output: VQDMULH_LANE:4:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -1100,13 +1100,13 @@ VQDMULH_LANE:23:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQDMULH_LANE:24:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMULH_LANE:25:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMULH_LANE (check mul overflow) overflow output: -VQDMULH_LANE:26:vqdmulh_lane_s16 Neon overflow 1 -VQDMULH_LANE:27:vqdmulh_lane_s32 Neon overflow 1 -VQDMULH_LANE:28:vqdmulhq_lane_s16 Neon overflow 1 -VQDMULH_LANE:29:vqdmulhq_lane_s32 Neon overflow 1 +VQDMULH_LANE (check mul cumulative saturation) cumulative saturation output: +VQDMULH_LANE:26:vqdmulh_lane_s16 Neon cumulative saturation 1 +VQDMULH_LANE:27:vqdmulh_lane_s32 Neon cumulative saturation 1 +VQDMULH_LANE:28:vqdmulhq_lane_s16 Neon cumulative saturation 1 +VQDMULH_LANE:29:vqdmulhq_lane_s32 Neon cumulative saturation 1 -VQDMULH_LANE (check mul overflow) output: +VQDMULH_LANE (check mul cumulative saturation) output: VQDMULH_LANE:30:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMULH_LANE:31:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQDMULH_LANE:32:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -1130,11 +1130,11 @@ VQDMULH_LANE:49:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQDMULH_LANE:50:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMULH_LANE:51:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMULH_N overflow output: -VQDMULH_N:0:vqdmulh_n_s16 Neon overflow 0 -VQDMULH_N:1:vqdmulh_n_s32 Neon overflow 0 -VQDMULH_N:2:vqdmulhq_n_s16 Neon overflow 0 -VQDMULH_N:3:vqdmulhq_n_s32 Neon overflow 0 +VQDMULH_N cumulative saturation output: +VQDMULH_N:0:vqdmulh_n_s16 Neon cumulative saturation 0 +VQDMULH_N:1:vqdmulh_n_s32 Neon cumulative saturation 0 +VQDMULH_N:2:vqdmulhq_n_s16 Neon cumulative saturation 0 +VQDMULH_N:3:vqdmulhq_n_s32 Neon cumulative saturation 0 VQDMULH_N output: VQDMULH_N:4:result_int16x4 [] = { 19, 19, 19, 19, } @@ -1142,13 +1142,13 @@ VQDMULH_N:5:result_int32x2 [] = { 4, 4, } VQDMULH_N:6:result_int16x8 [] = { 10, 10, 10, 10, 10, 10, 10, 10, } VQDMULH_N:7:result_int32x4 [] = { a, a, a, a, } -VQDMULH_N (check mul overflow) overflow output: -VQDMULH_N:8:vqdmulh_n_s16 Neon overflow 1 -VQDMULH_N:9:vqdmulh_n_s32 Neon overflow 1 -VQDMULH_N:10:vqdmulhq_n_s16 Neon overflow 1 -VQDMULH_N:11:vqdmulhq_n_s32 Neon overflow 1 +VQDMULH_N (check mul cumulative saturation) cumulative saturation output: +VQDMULH_N:8:vqdmulh_n_s16 Neon cumulative saturation 1 +VQDMULH_N:9:vqdmulh_n_s32 Neon cumulative saturation 1 +VQDMULH_N:10:vqdmulhq_n_s16 Neon cumulative saturation 1 +VQDMULH_N:11:vqdmulhq_n_s32 Neon cumulative saturation 1 -VQDMULH_N (check mul overflow) output: +VQDMULH_N (check mul cumulative saturation) output: VQDMULH_N:12:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMULH_N:13:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQDMULH_N:14:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -1172,9 +1172,9 @@ VQDMULH_N:31:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQDMULH_N:32:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMULH_N:33:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMULL overflow output: -VQDMULL:0:vqdmull_s16 Neon overflow 0 -VQDMULL:1:vqdmull_s32 Neon overflow 0 +VQDMULL cumulative saturation output: +VQDMULL:0:vqdmull_s16 Neon cumulative saturation 0 +VQDMULL:1:vqdmull_s32 Neon cumulative saturation 0 VQDMULL output: VQDMULL:2:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -1200,11 +1200,11 @@ VQDMULL:21:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33 VQDMULL:22:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMULL:23:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMULL (check mul overflow) overflow output: -VQDMULL:24:vqdmull_s16 Neon overflow 1 -VQDMULL:25:vqdmull_s32 Neon overflow 1 +VQDMULL (check mul cumulative saturation) cumulative saturation output: +VQDMULL:24:vqdmull_s16 Neon cumulative saturation 1 +VQDMULL:25:vqdmull_s32 Neon cumulative saturation 1 -VQDMULL (check mul overflow) output: +VQDMULL (check mul cumulative saturation) output: VQDMULL:26:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMULL:27:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQDMULL:28:result_int32x2 [] = { 33333333, 33333333, } @@ -1228,9 +1228,9 @@ VQDMULL:45:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33 VQDMULL:46:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMULL:47:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMLAL overflow output: -VQDMLAL:0:vqdmlal_s16 Neon overflow 0 -VQDMLAL:1:vqdmlal_s32 Neon overflow 0 +VQDMLAL cumulative saturation output: +VQDMLAL:0:vqdmlal_s16 Neon cumulative saturation 0 +VQDMLAL:1:vqdmlal_s32 Neon cumulative saturation 0 VQDMLAL output: VQDMLAL:2:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -1256,11 +1256,11 @@ VQDMLAL:21:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33 VQDMLAL:22:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMLAL:23:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMLAL (check mul overflow) overflow output: -VQDMLAL:24:vqdmlal_s16 Neon overflow 1 -VQDMLAL:25:vqdmlal_s32 Neon overflow 1 +VQDMLAL (check mul cumulative saturation) cumulative saturation output: +VQDMLAL:24:vqdmlal_s16 Neon cumulative saturation 1 +VQDMLAL:25:vqdmlal_s32 Neon cumulative saturation 1 -VQDMLAL (check mul overflow) output: +VQDMLAL (check mul cumulative saturation) output: VQDMLAL:26:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMLAL:27:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQDMLAL:28:result_int32x2 [] = { 33333333, 33333333, } @@ -1284,9 +1284,9 @@ VQDMLAL:45:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33 VQDMLAL:46:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMLAL:47:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMLSL overflow output: -VQDMLSL:0:vqdmlsl_s16 Neon overflow 0 -VQDMLSL:1:vqdmlsl_s32 Neon overflow 0 +VQDMLSL cumulative saturation output: +VQDMLSL:0:vqdmlsl_s16 Neon cumulative saturation 0 +VQDMLSL:1:vqdmlsl_s32 Neon cumulative saturation 0 VQDMLSL output: VQDMLSL:2:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -1312,11 +1312,11 @@ VQDMLSL:21:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33 VQDMLSL:22:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQDMLSL:23:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQDMLSL (check mul overflow) overflow output: -VQDMLSL:24:vqdmlsl_s16 Neon overflow 1 -VQDMLSL:25:vqdmlsl_s32 Neon overflow 1 +VQDMLSL (check mul cumulative saturation) cumulative saturation output: +VQDMLSL:24:vqdmlsl_s16 Neon cumulative saturation 1 +VQDMLSL:25:vqdmlsl_s32 Neon cumulative saturation 1 -VQDMLSL (check mul overflow) output: +VQDMLSL (check mul cumulative saturation) output: VQDMLSL:26:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMLSL:27:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQDMLSL:28:result_int32x2 [] = { 33333333, 33333333, } @@ -1594,23 +1594,23 @@ VSHL_N:19:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VSHL_N:20:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VSHL_N:21:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHL/VQSHLQ (with input = 0) overflow output: -VQSHL/VQSHLQ:0:vqshl_s8 Neon overflow 0 -VQSHL/VQSHLQ:1:vqshl_s16 Neon overflow 0 -VQSHL/VQSHLQ:2:vqshl_s32 Neon overflow 0 -VQSHL/VQSHLQ:3:vqshl_s64 Neon overflow 0 -VQSHL/VQSHLQ:4:vqshl_u8 Neon overflow 0 -VQSHL/VQSHLQ:5:vqshl_u16 Neon overflow 0 -VQSHL/VQSHLQ:6:vqshl_u32 Neon overflow 0 -VQSHL/VQSHLQ:7:vqshl_u64 Neon overflow 0 -VQSHL/VQSHLQ:8:vqshlq_s8 Neon overflow 0 -VQSHL/VQSHLQ:9:vqshlq_s16 Neon overflow 0 -VQSHL/VQSHLQ:10:vqshlq_s32 Neon overflow 0 -VQSHL/VQSHLQ:11:vqshlq_s64 Neon overflow 0 -VQSHL/VQSHLQ:12:vqshlq_u8 Neon overflow 0 -VQSHL/VQSHLQ:13:vqshlq_u16 Neon overflow 0 -VQSHL/VQSHLQ:14:vqshlq_u32 Neon overflow 0 -VQSHL/VQSHLQ:15:vqshlq_u64 Neon overflow 0 +VQSHL/VQSHLQ (with input = 0) cumulative saturation output: +VQSHL/VQSHLQ:0:vqshl_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:1:vqshl_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:2:vqshl_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:3:vqshl_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:4:vqshl_u8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:5:vqshl_u16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:6:vqshl_u32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:7:vqshl_u64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:8:vqshlq_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:9:vqshlq_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:10:vqshlq_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:11:vqshlq_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:12:vqshlq_u8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:13:vqshlq_u16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:14:vqshlq_u32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:15:vqshlq_u64 Neon cumulative saturation 0 VQSHL/VQSHLQ (with input = 0) output: VQSHL/VQSHLQ:16:result_int8x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } @@ -1636,23 +1636,23 @@ VQSHL/VQSHLQ:35:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQSHL/VQSHLQ:36:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHL/VQSHLQ:37:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHL/VQSHLQ (input 0 and negative shift amount) overflow output: -VQSHL/VQSHLQ:38:vqshl_s8 Neon overflow 0 -VQSHL/VQSHLQ:39:vqshl_s16 Neon overflow 0 -VQSHL/VQSHLQ:40:vqshl_s32 Neon overflow 0 -VQSHL/VQSHLQ:41:vqshl_s64 Neon overflow 0 -VQSHL/VQSHLQ:42:vqshl_u8 Neon overflow 0 -VQSHL/VQSHLQ:43:vqshl_u16 Neon overflow 0 -VQSHL/VQSHLQ:44:vqshl_u32 Neon overflow 0 -VQSHL/VQSHLQ:45:vqshl_u64 Neon overflow 0 -VQSHL/VQSHLQ:46:vqshlq_s8 Neon overflow 0 -VQSHL/VQSHLQ:47:vqshlq_s16 Neon overflow 0 -VQSHL/VQSHLQ:48:vqshlq_s32 Neon overflow 0 -VQSHL/VQSHLQ:49:vqshlq_s64 Neon overflow 0 -VQSHL/VQSHLQ:50:vqshlq_u8 Neon overflow 0 -VQSHL/VQSHLQ:51:vqshlq_u16 Neon overflow 0 -VQSHL/VQSHLQ:52:vqshlq_u32 Neon overflow 0 -VQSHL/VQSHLQ:53:vqshlq_u64 Neon overflow 0 +VQSHL/VQSHLQ (input 0 and negative shift amount) cumulative saturation output: +VQSHL/VQSHLQ:38:vqshl_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:39:vqshl_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:40:vqshl_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:41:vqshl_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:42:vqshl_u8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:43:vqshl_u16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:44:vqshl_u32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:45:vqshl_u64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:46:vqshlq_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:47:vqshlq_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:48:vqshlq_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:49:vqshlq_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:50:vqshlq_u8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:51:vqshlq_u16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:52:vqshlq_u32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:53:vqshlq_u64 Neon cumulative saturation 0 VQSHL/VQSHLQ (input 0 and negative shift amount) output: VQSHL/VQSHLQ:54:result_int8x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } @@ -1678,23 +1678,23 @@ VQSHL/VQSHLQ:73:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQSHL/VQSHLQ:74:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHL/VQSHLQ:75:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHL/VQSHLQ overflow output: -VQSHL/VQSHLQ:76:vqshl_s8 Neon overflow 0 -VQSHL/VQSHLQ:77:vqshl_s16 Neon overflow 0 -VQSHL/VQSHLQ:78:vqshl_s32 Neon overflow 0 -VQSHL/VQSHLQ:79:vqshl_s64 Neon overflow 0 -VQSHL/VQSHLQ:80:vqshl_u8 Neon overflow 1 -VQSHL/VQSHLQ:81:vqshl_u16 Neon overflow 1 -VQSHL/VQSHLQ:82:vqshl_u32 Neon overflow 1 -VQSHL/VQSHLQ:83:vqshl_u64 Neon overflow 0 -VQSHL/VQSHLQ:84:vqshlq_s8 Neon overflow 1 -VQSHL/VQSHLQ:85:vqshlq_s16 Neon overflow 1 -VQSHL/VQSHLQ:86:vqshlq_s32 Neon overflow 1 -VQSHL/VQSHLQ:87:vqshlq_s64 Neon overflow 1 -VQSHL/VQSHLQ:88:vqshlq_u8 Neon overflow 1 -VQSHL/VQSHLQ:89:vqshlq_u16 Neon overflow 1 -VQSHL/VQSHLQ:90:vqshlq_u32 Neon overflow 1 -VQSHL/VQSHLQ:91:vqshlq_u64 Neon overflow 1 +VQSHL/VQSHLQ cumulative saturation output: +VQSHL/VQSHLQ:76:vqshl_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:77:vqshl_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:78:vqshl_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:79:vqshl_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:80:vqshl_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:81:vqshl_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:82:vqshl_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:83:vqshl_u64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:84:vqshlq_s8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:85:vqshlq_s16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:86:vqshlq_s32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:87:vqshlq_s64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:88:vqshlq_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:89:vqshlq_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:90:vqshlq_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:91:vqshlq_u64 Neon cumulative saturation 1 VQSHL/VQSHLQ output: VQSHL/VQSHLQ:92:result_int8x8 [] = { ffffffe0, ffffffe2, ffffffe4, ffffffe6, ffffffe8, ffffffea, ffffffec, ffffffee, } @@ -1720,23 +1720,23 @@ VQSHL/VQSHLQ:111:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQSHL/VQSHLQ:112:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHL/VQSHLQ:113:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHL/VQSHLQ (negative shift amount) overflow output: -VQSHL/VQSHLQ:114:vqshl_s8 Neon overflow 0 -VQSHL/VQSHLQ:115:vqshl_s16 Neon overflow 0 -VQSHL/VQSHLQ:116:vqshl_s32 Neon overflow 0 -VQSHL/VQSHLQ:117:vqshl_s64 Neon overflow 0 -VQSHL/VQSHLQ:118:vqshl_u8 Neon overflow 0 -VQSHL/VQSHLQ:119:vqshl_u16 Neon overflow 0 -VQSHL/VQSHLQ:120:vqshl_u32 Neon overflow 0 -VQSHL/VQSHLQ:121:vqshl_u64 Neon overflow 0 -VQSHL/VQSHLQ:122:vqshlq_s8 Neon overflow 0 -VQSHL/VQSHLQ:123:vqshlq_s16 Neon overflow 0 -VQSHL/VQSHLQ:124:vqshlq_s32 Neon overflow 0 -VQSHL/VQSHLQ:125:vqshlq_s64 Neon overflow 0 -VQSHL/VQSHLQ:126:vqshlq_u8 Neon overflow 0 -VQSHL/VQSHLQ:127:vqshlq_u16 Neon overflow 0 -VQSHL/VQSHLQ:128:vqshlq_u32 Neon overflow 0 -VQSHL/VQSHLQ:129:vqshlq_u64 Neon overflow 0 +VQSHL/VQSHLQ (negative shift amount) cumulative saturation output: +VQSHL/VQSHLQ:114:vqshl_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:115:vqshl_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:116:vqshl_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:117:vqshl_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:118:vqshl_u8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:119:vqshl_u16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:120:vqshl_u32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:121:vqshl_u64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:122:vqshlq_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:123:vqshlq_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:124:vqshlq_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:125:vqshlq_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:126:vqshlq_u8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:127:vqshlq_u16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:128:vqshlq_u32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:129:vqshlq_u64 Neon cumulative saturation 0 VQSHL/VQSHLQ (negative shift amount) output: VQSHL/VQSHLQ:130:result_int8x8 [] = { fffffff8, fffffff8, fffffff9, fffffff9, fffffffa, fffffffa, fffffffb, fffffffb, } @@ -1762,23 +1762,23 @@ VQSHL/VQSHLQ:149:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQSHL/VQSHLQ:150:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHL/VQSHLQ:151:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHL/VQSHLQ (large shift amount, negative input) overflow output: -VQSHL/VQSHLQ:152:vqshl_s8 Neon overflow 1 -VQSHL/VQSHLQ:153:vqshl_s16 Neon overflow 1 -VQSHL/VQSHLQ:154:vqshl_s32 Neon overflow 1 -VQSHL/VQSHLQ:155:vqshl_s64 Neon overflow 1 -VQSHL/VQSHLQ:156:vqshl_u8 Neon overflow 1 -VQSHL/VQSHLQ:157:vqshl_u16 Neon overflow 1 -VQSHL/VQSHLQ:158:vqshl_u32 Neon overflow 1 -VQSHL/VQSHLQ:159:vqshl_u64 Neon overflow 1 -VQSHL/VQSHLQ:160:vqshlq_s8 Neon overflow 1 -VQSHL/VQSHLQ:161:vqshlq_s16 Neon overflow 1 -VQSHL/VQSHLQ:162:vqshlq_s32 Neon overflow 1 -VQSHL/VQSHLQ:163:vqshlq_s64 Neon overflow 1 -VQSHL/VQSHLQ:164:vqshlq_u8 Neon overflow 1 -VQSHL/VQSHLQ:165:vqshlq_u16 Neon overflow 1 -VQSHL/VQSHLQ:166:vqshlq_u32 Neon overflow 1 -VQSHL/VQSHLQ:167:vqshlq_u64 Neon overflow 1 +VQSHL/VQSHLQ (large shift amount, negative input) cumulative saturation output: +VQSHL/VQSHLQ:152:vqshl_s8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:153:vqshl_s16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:154:vqshl_s32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:155:vqshl_s64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:156:vqshl_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:157:vqshl_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:158:vqshl_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:159:vqshl_u64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:160:vqshlq_s8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:161:vqshlq_s16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:162:vqshlq_s32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:163:vqshlq_s64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:164:vqshlq_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:165:vqshlq_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:166:vqshlq_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:167:vqshlq_u64 Neon cumulative saturation 1 VQSHL/VQSHLQ (large shift amount, negative input) output: VQSHL/VQSHLQ:168:result_int8x8 [] = { ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, } @@ -1804,25 +1804,25 @@ VQSHL/VQSHLQ:187:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQSHL/VQSHLQ:188:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHL/VQSHLQ:189:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHL/VQSHLQ (check saturation/overflow) overflow output: -VQSHL/VQSHLQ:190:vqshl_s8 Neon overflow 0 -VQSHL/VQSHLQ:191:vqshl_s16 Neon overflow 0 -VQSHL/VQSHLQ:192:vqshl_s32 Neon overflow 0 -VQSHL/VQSHLQ:193:vqshl_s64 Neon overflow 0 -VQSHL/VQSHLQ:194:vqshl_u8 Neon overflow 0 -VQSHL/VQSHLQ:195:vqshl_u16 Neon overflow 0 -VQSHL/VQSHLQ:196:vqshl_u32 Neon overflow 0 -VQSHL/VQSHLQ:197:vqshl_u64 Neon overflow 0 -VQSHL/VQSHLQ:198:vqshlq_s8 Neon overflow 0 -VQSHL/VQSHLQ:199:vqshlq_s16 Neon overflow 0 -VQSHL/VQSHLQ:200:vqshlq_s32 Neon overflow 0 -VQSHL/VQSHLQ:201:vqshlq_s64 Neon overflow 0 -VQSHL/VQSHLQ:202:vqshlq_u8 Neon overflow 0 -VQSHL/VQSHLQ:203:vqshlq_u16 Neon overflow 0 -VQSHL/VQSHLQ:204:vqshlq_u32 Neon overflow 0 -VQSHL/VQSHLQ:205:vqshlq_u64 Neon overflow 0 - -VQSHL/VQSHLQ (check saturation/overflow) output: +VQSHL/VQSHLQ (check cumulative saturation) cumulative saturation output: +VQSHL/VQSHLQ:190:vqshl_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:191:vqshl_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:192:vqshl_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:193:vqshl_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:194:vqshl_u8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:195:vqshl_u16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:196:vqshl_u32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:197:vqshl_u64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:198:vqshlq_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:199:vqshlq_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:200:vqshlq_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:201:vqshlq_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:202:vqshlq_u8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:203:vqshlq_u16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:204:vqshlq_u32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:205:vqshlq_u64 Neon cumulative saturation 0 + +VQSHL/VQSHLQ (check cumulative saturation) output: VQSHL/VQSHLQ:206:result_int8x8 [] = { 3f, 3f, 3f, 3f, 3f, 3f, 3f, 3f, } VQSHL/VQSHLQ:207:result_int16x4 [] = { 3fff, 3fff, 3fff, 3fff, } VQSHL/VQSHLQ:208:result_int32x2 [] = { 3fffffff, 3fffffff, } @@ -1846,23 +1846,23 @@ VQSHL/VQSHLQ:225:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQSHL/VQSHLQ:226:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHL/VQSHLQ:227:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHL/VQSHLQ (large shift amount, positive input) overflow output: -VQSHL/VQSHLQ:228:vqshl_s8 Neon overflow 1 -VQSHL/VQSHLQ:229:vqshl_s16 Neon overflow 1 -VQSHL/VQSHLQ:230:vqshl_s32 Neon overflow 1 -VQSHL/VQSHLQ:231:vqshl_s64 Neon overflow 1 -VQSHL/VQSHLQ:232:vqshl_u8 Neon overflow 1 -VQSHL/VQSHLQ:233:vqshl_u16 Neon overflow 1 -VQSHL/VQSHLQ:234:vqshl_u32 Neon overflow 1 -VQSHL/VQSHLQ:235:vqshl_u64 Neon overflow 1 -VQSHL/VQSHLQ:236:vqshlq_s8 Neon overflow 1 -VQSHL/VQSHLQ:237:vqshlq_s16 Neon overflow 1 -VQSHL/VQSHLQ:238:vqshlq_s32 Neon overflow 1 -VQSHL/VQSHLQ:239:vqshlq_s64 Neon overflow 1 -VQSHL/VQSHLQ:240:vqshlq_u8 Neon overflow 1 -VQSHL/VQSHLQ:241:vqshlq_u16 Neon overflow 1 -VQSHL/VQSHLQ:242:vqshlq_u32 Neon overflow 1 -VQSHL/VQSHLQ:243:vqshlq_u64 Neon overflow 1 +VQSHL/VQSHLQ (large shift amount, positive input) cumulative saturation output: +VQSHL/VQSHLQ:228:vqshl_s8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:229:vqshl_s16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:230:vqshl_s32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:231:vqshl_s64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:232:vqshl_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:233:vqshl_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:234:vqshl_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:235:vqshl_u64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:236:vqshlq_s8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:237:vqshlq_s16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:238:vqshlq_s32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:239:vqshlq_s64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:240:vqshlq_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:241:vqshlq_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:242:vqshlq_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:243:vqshlq_u64 Neon cumulative saturation 1 VQSHL/VQSHLQ (large shift amount, positive input) output: VQSHL/VQSHLQ:244:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -1888,23 +1888,23 @@ VQSHL/VQSHLQ:263:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQSHL/VQSHLQ:264:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHL/VQSHLQ:265:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHL/VQSHLQ (check saturation on 64 bits) overflow output: -VQSHL/VQSHLQ:266:vqshl_s8 Neon overflow 1 -VQSHL/VQSHLQ:267:vqshl_s16 Neon overflow 1 -VQSHL/VQSHLQ:268:vqshl_s32 Neon overflow 1 -VQSHL/VQSHLQ:269:vqshl_s64 Neon overflow 1 -VQSHL/VQSHLQ:270:vqshl_u8 Neon overflow 1 -VQSHL/VQSHLQ:271:vqshl_u16 Neon overflow 1 -VQSHL/VQSHLQ:272:vqshl_u32 Neon overflow 1 -VQSHL/VQSHLQ:273:vqshl_u64 Neon overflow 1 -VQSHL/VQSHLQ:274:vqshlq_s8 Neon overflow 1 -VQSHL/VQSHLQ:275:vqshlq_s16 Neon overflow 1 -VQSHL/VQSHLQ:276:vqshlq_s32 Neon overflow 1 -VQSHL/VQSHLQ:277:vqshlq_s64 Neon overflow 1 -VQSHL/VQSHLQ:278:vqshlq_u8 Neon overflow 1 -VQSHL/VQSHLQ:279:vqshlq_u16 Neon overflow 1 -VQSHL/VQSHLQ:280:vqshlq_u32 Neon overflow 1 -VQSHL/VQSHLQ:281:vqshlq_u64 Neon overflow 1 +VQSHL/VQSHLQ (check saturation on 64 bits) cumulative saturation output: +VQSHL/VQSHLQ:266:vqshl_s8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:267:vqshl_s16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:268:vqshl_s32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:269:vqshl_s64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:270:vqshl_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:271:vqshl_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:272:vqshl_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:273:vqshl_u64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:274:vqshlq_s8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:275:vqshlq_s16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:276:vqshlq_s32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:277:vqshlq_s64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:278:vqshlq_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:279:vqshlq_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:280:vqshlq_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:281:vqshlq_u64 Neon cumulative saturation 1 VQSHL/VQSHLQ (check saturation on 64 bits) output: VQSHL/VQSHLQ:282:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -1930,23 +1930,23 @@ VQSHL/VQSHLQ:301:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQSHL/VQSHLQ:302:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHL/VQSHLQ:303:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHL_N/VQSHLQ_N overflow output: -VQSHL_N/VQSHLQ_N:0:vqshl_n_s8 Neon overflow 0 -VQSHL_N/VQSHLQ_N:1:vqshl_n_s16 Neon overflow 0 -VQSHL_N/VQSHLQ_N:2:vqshl_n_s32 Neon overflow 0 -VQSHL_N/VQSHLQ_N:3:vqshl_n_s64 Neon overflow 0 -VQSHL_N/VQSHLQ_N:4:vqshl_n_u8 Neon overflow 1 -VQSHL_N/VQSHLQ_N:5:vqshl_n_u16 Neon overflow 1 -VQSHL_N/VQSHLQ_N:6:vqshl_n_u32 Neon overflow 1 -VQSHL_N/VQSHLQ_N:7:vqshl_n_u64 Neon overflow 1 -VQSHL_N/VQSHLQ_N:8:vqshlq_n_s8 Neon overflow 0 -VQSHL_N/VQSHLQ_N:9:vqshlq_n_s16 Neon overflow 0 -VQSHL_N/VQSHLQ_N:10:vqshlq_n_s32 Neon overflow 0 -VQSHL_N/VQSHLQ_N:11:vqshlq_n_s64 Neon overflow 0 -VQSHL_N/VQSHLQ_N:12:vqshlq_n_u8 Neon overflow 1 -VQSHL_N/VQSHLQ_N:13:vqshlq_n_u16 Neon overflow 1 -VQSHL_N/VQSHLQ_N:14:vqshlq_n_u32 Neon overflow 1 -VQSHL_N/VQSHLQ_N:15:vqshlq_n_u64 Neon overflow 1 +VQSHL_N/VQSHLQ_N cumulative saturation output: +VQSHL_N/VQSHLQ_N:0:vqshl_n_s8 Neon cumulative saturation 0 +VQSHL_N/VQSHLQ_N:1:vqshl_n_s16 Neon cumulative saturation 0 +VQSHL_N/VQSHLQ_N:2:vqshl_n_s32 Neon cumulative saturation 0 +VQSHL_N/VQSHLQ_N:3:vqshl_n_s64 Neon cumulative saturation 0 +VQSHL_N/VQSHLQ_N:4:vqshl_n_u8 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:5:vqshl_n_u16 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:6:vqshl_n_u32 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:7:vqshl_n_u64 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:8:vqshlq_n_s8 Neon cumulative saturation 0 +VQSHL_N/VQSHLQ_N:9:vqshlq_n_s16 Neon cumulative saturation 0 +VQSHL_N/VQSHLQ_N:10:vqshlq_n_s32 Neon cumulative saturation 0 +VQSHL_N/VQSHLQ_N:11:vqshlq_n_s64 Neon cumulative saturation 0 +VQSHL_N/VQSHLQ_N:12:vqshlq_n_u8 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:13:vqshlq_n_u16 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:14:vqshlq_n_u32 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:15:vqshlq_n_u64 Neon cumulative saturation 1 VQSHL_N/VQSHLQ_N output: VQSHL_N/VQSHLQ_N:16:result_int8x8 [] = { ffffffc0, ffffffc4, ffffffc8, ffffffcc, ffffffd0, ffffffd4, ffffffd8, ffffffdc, } @@ -1972,23 +1972,23 @@ VQSHL_N/VQSHLQ_N:35:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQSHL_N/VQSHLQ_N:36:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHL_N/VQSHLQ_N:37:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHL_N/VQSHLQ_N (check saturation with large positive input) overflow output: -VQSHL_N/VQSHLQ_N:38:vqshl_n_s8 Neon overflow 1 -VQSHL_N/VQSHLQ_N:39:vqshl_n_s16 Neon overflow 1 -VQSHL_N/VQSHLQ_N:40:vqshl_n_s32 Neon overflow 1 -VQSHL_N/VQSHLQ_N:41:vqshl_n_s64 Neon overflow 1 -VQSHL_N/VQSHLQ_N:42:vqshl_n_u8 Neon overflow 1 -VQSHL_N/VQSHLQ_N:43:vqshl_n_u16 Neon overflow 1 -VQSHL_N/VQSHLQ_N:44:vqshl_n_u32 Neon overflow 1 -VQSHL_N/VQSHLQ_N:45:vqshl_n_u64 Neon overflow 1 -VQSHL_N/VQSHLQ_N:46:vqshlq_n_s8 Neon overflow 1 -VQSHL_N/VQSHLQ_N:47:vqshlq_n_s16 Neon overflow 1 -VQSHL_N/VQSHLQ_N:48:vqshlq_n_s32 Neon overflow 1 -VQSHL_N/VQSHLQ_N:49:vqshlq_n_s64 Neon overflow 1 -VQSHL_N/VQSHLQ_N:50:vqshlq_n_u8 Neon overflow 1 -VQSHL_N/VQSHLQ_N:51:vqshlq_n_u16 Neon overflow 1 -VQSHL_N/VQSHLQ_N:52:vqshlq_n_u32 Neon overflow 1 -VQSHL_N/VQSHLQ_N:53:vqshlq_n_u64 Neon overflow 1 +VQSHL_N/VQSHLQ_N (check saturation with large positive input) cumulative saturation output: +VQSHL_N/VQSHLQ_N:38:vqshl_n_s8 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:39:vqshl_n_s16 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:40:vqshl_n_s32 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:41:vqshl_n_s64 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:42:vqshl_n_u8 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:43:vqshl_n_u16 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:44:vqshl_n_u32 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:45:vqshl_n_u64 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:46:vqshlq_n_s8 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:47:vqshlq_n_s16 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:48:vqshlq_n_s32 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:49:vqshlq_n_s64 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:50:vqshlq_n_u8 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:51:vqshlq_n_u16 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:52:vqshlq_n_u32 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:53:vqshlq_n_u64 Neon cumulative saturation 1 VQSHL_N/VQSHLQ_N (check saturation with large positive input) output: VQSHL_N/VQSHLQ_N:54:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -2470,35 +2470,35 @@ VDUP_LANE/VDUP_LANEQ:19:result_poly8x16 [] = { f5, f5, f5, f5, f5, f5, f5, f5, f VDUP_LANE/VDUP_LANEQ:20:result_poly16x8 [] = { fff1, fff1, fff1, fff1, fff1, fff1, fff1, fff1, } VDUP_LANE/VDUP_LANEQ:21:result_float32x4 [] = { c1700000 -0x1.e000000p+3 -15, c1700000 -0x1.e000000p+3 -15, c1700000 -0x1.e000000p+3 -15, c1700000 -0x1.e000000p+3 -15, } -VQDMULL_LANE overflow output: -VQDMULL_LANE:0:vqdmull_lane_s16 Neon overflow 0 -VQDMULL_LANE:1:vqdmull_lane_s32 Neon overflow 0 +VQDMULL_LANE cumulative saturation output: +VQDMULL_LANE:0:vqdmull_lane_s16 Neon cumulative saturation 0 +VQDMULL_LANE:1:vqdmull_lane_s32 Neon cumulative saturation 0 VQDMULL_LANE output: VQDMULL_LANE:2:result_int32x4 [] = { 8000, 8000, 8000, 8000, } VQDMULL_LANE:3:result_int64x2 [] = { 4000, 4000, } -VQDMULL_LANE (check mul overflow) overflow output: -VQDMULL_LANE:4:vqdmull_lane_s16 Neon overflow 1 -VQDMULL_LANE:5:vqdmull_lane_s32 Neon overflow 1 +VQDMULL_LANE (check mul cumulative saturation) cumulative saturation output: +VQDMULL_LANE:4:vqdmull_lane_s16 Neon cumulative saturation 1 +VQDMULL_LANE:5:vqdmull_lane_s32 Neon cumulative saturation 1 -VQDMULL_LANE (check mul overflow) output: +VQDMULL_LANE (check mul cumulative saturation) output: VQDMULL_LANE:6:result_int32x4 [] = { 7fffffff, 7fffffff, 7fffffff, 7fffffff, } VQDMULL_LANE:7:result_int64x2 [] = { 7fffffffffffffff, 7fffffffffffffff, } -VQDMULL_N overflow output: -VQDMULL_N:0:vqdmull_n_s16 Neon overflow 0 -VQDMULL_N:1:vqdmull_n_s32 Neon overflow 0 +VQDMULL_N cumulative saturation output: +VQDMULL_N:0:vqdmull_n_s16 Neon cumulative saturation 0 +VQDMULL_N:1:vqdmull_n_s32 Neon cumulative saturation 0 VQDMULL_N output: VQDMULL_N:2:result_int32x4 [] = { 44000, 44000, 44000, 44000, } VQDMULL_N:3:result_int64x2 [] = { aa000, aa000, } -VQDMULL_N (check mul overflow) overflow output: -VQDMULL_N:4:vqdmull_n_s16 Neon overflow 1 -VQDMULL_N:5:vqdmull_n_s32 Neon overflow 1 +VQDMULL_N (check mul cumulative saturation) cumulative saturation output: +VQDMULL_N:4:vqdmull_n_s16 Neon cumulative saturation 1 +VQDMULL_N:5:vqdmull_n_s32 Neon cumulative saturation 1 -VQDMULL_N (check mul overflow) output: +VQDMULL_N (check mul cumulative saturation) output: VQDMULL_N:6:result_int32x4 [] = { 7fffffff, 7fffffff, 7fffffff, 7fffffff, } VQDMULL_N:7:result_int64x2 [] = { 7fffffffffffffff, 7fffffffffffffff, } @@ -2554,23 +2554,23 @@ float32: VSUB/VSUBQ:22:result_float32x2 [] = { c00ccccd -0x1.19999a0p+1 -2.2, c00ccccd -0x1.19999a0p+1 -2.2, } VSUB/VSUBQ:23:result_float32x4 [] = { c00ccccc -0x1.1999980p+1 -2.2, c00ccccc -0x1.1999980p+1 -2.2, c00ccccc -0x1.1999980p+1 -2.2, c00ccccc -0x1.1999980p+1 -2.2, } -VQADD/VQADDQ overflow output: -VQADD/VQADDQ:0:vqadd_s8 Neon overflow 0 -VQADD/VQADDQ:1:vqadd_s16 Neon overflow 0 -VQADD/VQADDQ:2:vqadd_s32 Neon overflow 0 -VQADD/VQADDQ:3:vqadd_s64 Neon overflow 0 -VQADD/VQADDQ:4:vqadd_u8 Neon overflow 1 -VQADD/VQADDQ:5:vqadd_u16 Neon overflow 1 -VQADD/VQADDQ:6:vqadd_u32 Neon overflow 1 -VQADD/VQADDQ:7:vqadd_u64 Neon overflow 1 -VQADD/VQADDQ:8:vqaddq_s8 Neon overflow 0 -VQADD/VQADDQ:9:vqaddq_s16 Neon overflow 0 -VQADD/VQADDQ:10:vqaddq_s32 Neon overflow 0 -VQADD/VQADDQ:11:vqaddq_s64 Neon overflow 0 -VQADD/VQADDQ:12:vqaddq_u8 Neon overflow 1 -VQADD/VQADDQ:13:vqaddq_u16 Neon overflow 1 -VQADD/VQADDQ:14:vqaddq_u32 Neon overflow 1 -VQADD/VQADDQ:15:vqaddq_u64 Neon overflow 1 +VQADD/VQADDQ cumulative saturation output: +VQADD/VQADDQ:0:vqadd_s8 Neon cumulative saturation 0 +VQADD/VQADDQ:1:vqadd_s16 Neon cumulative saturation 0 +VQADD/VQADDQ:2:vqadd_s32 Neon cumulative saturation 0 +VQADD/VQADDQ:3:vqadd_s64 Neon cumulative saturation 0 +VQADD/VQADDQ:4:vqadd_u8 Neon cumulative saturation 1 +VQADD/VQADDQ:5:vqadd_u16 Neon cumulative saturation 1 +VQADD/VQADDQ:6:vqadd_u32 Neon cumulative saturation 1 +VQADD/VQADDQ:7:vqadd_u64 Neon cumulative saturation 1 +VQADD/VQADDQ:8:vqaddq_s8 Neon cumulative saturation 0 +VQADD/VQADDQ:9:vqaddq_s16 Neon cumulative saturation 0 +VQADD/VQADDQ:10:vqaddq_s32 Neon cumulative saturation 0 +VQADD/VQADDQ:11:vqaddq_s64 Neon cumulative saturation 0 +VQADD/VQADDQ:12:vqaddq_u8 Neon cumulative saturation 1 +VQADD/VQADDQ:13:vqaddq_u16 Neon cumulative saturation 1 +VQADD/VQADDQ:14:vqaddq_u32 Neon cumulative saturation 1 +VQADD/VQADDQ:15:vqaddq_u64 Neon cumulative saturation 1 VQADD/VQADDQ output: VQADD/VQADDQ:16:result_int8x8 [] = { 1, 2, 3, 4, 5, 6, 7, 8, } @@ -2596,11 +2596,11 @@ VQADD/VQADDQ:35:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQADD/VQADDQ:36:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQADD/VQADDQ:37:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQADD/VQADDQ 64 bits saturation overflow output: -VQADD/VQADDQ:38:vqadd_s64 Neon overflow 0 -VQADD/VQADDQ:39:vqadd_u64 Neon overflow 0 -VQADD/VQADDQ:40:vqaddq_s64 Neon overflow 0 -VQADD/VQADDQ:41:vqaddq_u64 Neon overflow 0 +VQADD/VQADDQ 64 bits saturation cumulative saturation output: +VQADD/VQADDQ:38:vqadd_s64 Neon cumulative saturation 0 +VQADD/VQADDQ:39:vqadd_u64 Neon cumulative saturation 0 +VQADD/VQADDQ:40:vqaddq_s64 Neon cumulative saturation 0 +VQADD/VQADDQ:41:vqaddq_u64 Neon cumulative saturation 0 64 bits saturation: VQADD/VQADDQ:42:result_int64x1 [] = { fffffffffffffff0, } @@ -2608,33 +2608,33 @@ VQADD/VQADDQ:43:result_uint64x1 [] = { fffffffffffffff0, } VQADD/VQADDQ:44:result_int64x2 [] = { fffffffffffffff0, fffffffffffffff1, } VQADD/VQADDQ:45:result_uint64x2 [] = { fffffffffffffff0, fffffffffffffff1, } -VQADD/VQADDQ 64 bits saturation overflow output: -VQADD/VQADDQ:46:vqadd_s64 Neon overflow 0 -VQADD/VQADDQ:47:vqadd_u64 Neon overflow 1 -VQADD/VQADDQ:48:vqaddq_s64 Neon overflow 0 -VQADD/VQADDQ:49:vqaddq_u64 Neon overflow 1 +VQADD/VQADDQ 64 bits saturation cumulative saturation output: +VQADD/VQADDQ:46:vqadd_s64 Neon cumulative saturation 0 +VQADD/VQADDQ:47:vqadd_u64 Neon cumulative saturation 1 +VQADD/VQADDQ:48:vqaddq_s64 Neon cumulative saturation 0 +VQADD/VQADDQ:49:vqaddq_u64 Neon cumulative saturation 1 VQADD/VQADDQ:50:result_int64x1 [] = { 34, } VQADD/VQADDQ:51:result_uint64x1 [] = { ffffffffffffffff, } VQADD/VQADDQ:52:result_int64x2 [] = { 34, 35, } VQADD/VQADDQ:53:result_uint64x2 [] = { ffffffffffffffff, ffffffffffffffff, } -VQADD/VQADDQ 64 bits saturation overflow output: -VQADD/VQADDQ:54:vqadd_s64 Neon overflow 1 -VQADD/VQADDQ:55:vqadd_u64 Neon overflow 1 -VQADD/VQADDQ:56:vqaddq_s64 Neon overflow 1 -VQADD/VQADDQ:57:vqaddq_u64 Neon overflow 1 +VQADD/VQADDQ 64 bits saturation cumulative saturation output: +VQADD/VQADDQ:54:vqadd_s64 Neon cumulative saturation 1 +VQADD/VQADDQ:55:vqadd_u64 Neon cumulative saturation 1 +VQADD/VQADDQ:56:vqaddq_s64 Neon cumulative saturation 1 +VQADD/VQADDQ:57:vqaddq_u64 Neon cumulative saturation 1 VQADD/VQADDQ:58:result_int64x1 [] = { 8000000000000000, } VQADD/VQADDQ:59:result_uint64x1 [] = { ffffffffffffffff, } VQADD/VQADDQ:60:result_int64x2 [] = { 7fffffffffffffff, 7fffffffffffffff, } VQADD/VQADDQ:61:result_uint64x2 [] = { ffffffffffffffff, ffffffffffffffff, } less than 64 bits saturation: -VQADD/VQADDQ:62:vqadd_s8 Neon overflow 1 -VQADD/VQADDQ:63:vqadd_s16 Neon overflow 1 -VQADD/VQADDQ:64:vqadd_s32 Neon overflow 1 -VQADD/VQADDQ:65:vqaddq_s8 Neon overflow 1 -VQADD/VQADDQ:66:vqaddq_s16 Neon overflow 1 -VQADD/VQADDQ:67:vqaddq_s32 Neon overflow 1 +VQADD/VQADDQ:62:vqadd_s8 Neon cumulative saturation 1 +VQADD/VQADDQ:63:vqadd_s16 Neon cumulative saturation 1 +VQADD/VQADDQ:64:vqadd_s32 Neon cumulative saturation 1 +VQADD/VQADDQ:65:vqaddq_s8 Neon cumulative saturation 1 +VQADD/VQADDQ:66:vqaddq_s16 Neon cumulative saturation 1 +VQADD/VQADDQ:67:vqaddq_s32 Neon cumulative saturation 1 VQADD/VQADDQ:68:result_int8x8 [] = { ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, } VQADD/VQADDQ:69:result_int16x4 [] = { ffff8000, ffff8000, ffff8000, ffff8000, } VQADD/VQADDQ:70:result_int32x2 [] = { 80000000, 80000000, } @@ -2642,13 +2642,13 @@ VQADD/VQADDQ:71:result_int8x16 [] = { ffffff80, ffffff80, ffffff80, ffffff80, ff VQADD/VQADDQ:72:result_int16x8 [] = { ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, } VQADD/VQADDQ:73:result_int32x4 [] = { 80000000, 80000000, 80000000, 80000000, } -VQADD/VQADDQ less than 64 bits saturation overflow output: -VQADD/VQADDQ:74:vqadd_u8 Neon overflow 1 -VQADD/VQADDQ:75:vqadd_u16 Neon overflow 1 -VQADD/VQADDQ:76:vqadd_u32 Neon overflow 1 -VQADD/VQADDQ:77:vqaddq_u8 Neon overflow 1 -VQADD/VQADDQ:78:vqaddq_u16 Neon overflow 1 -VQADD/VQADDQ:79:vqaddq_u32 Neon overflow 1 +VQADD/VQADDQ less than 64 bits saturation cumulative saturation output: +VQADD/VQADDQ:74:vqadd_u8 Neon cumulative saturation 1 +VQADD/VQADDQ:75:vqadd_u16 Neon cumulative saturation 1 +VQADD/VQADDQ:76:vqadd_u32 Neon cumulative saturation 1 +VQADD/VQADDQ:77:vqaddq_u8 Neon cumulative saturation 1 +VQADD/VQADDQ:78:vqaddq_u16 Neon cumulative saturation 1 +VQADD/VQADDQ:79:vqaddq_u32 Neon cumulative saturation 1 VQADD/VQADDQ:80:result_uint8x8 [] = { ff, ff, ff, ff, ff, ff, ff, ff, } VQADD/VQADDQ:81:result_uint16x4 [] = { ffff, ffff, ffff, ffff, } VQADD/VQADDQ:82:result_uint32x2 [] = { ffffffff, ffffffff, } @@ -2684,13 +2684,13 @@ float32: VABS/VABSQ:22:result_float32x2 [] = { 40133333 0x1.2666660p+1 2.3, 40133333 0x1.2666660p+1 2.3, } VABS/VABSQ:23:result_float32x4 [] = { 4059999a 0x1.b333340p+1 3.4, 4059999a 0x1.b333340p+1 3.4, 4059999a 0x1.b333340p+1 3.4, 4059999a 0x1.b333340p+1 3.4, } -VQABS/VQABSQ overflow output: -VQABS/VQABSQ:0:vqabs_s8 Neon overflow 0 -VQABS/VQABSQ:1:vqabs_s16 Neon overflow 0 -VQABS/VQABSQ:2:vqabs_s32 Neon overflow 0 -VQABS/VQABSQ:3:vqabsq_s8 Neon overflow 0 -VQABS/VQABSQ:4:vqabsq_s16 Neon overflow 0 -VQABS/VQABSQ:5:vqabsq_s32 Neon overflow 0 +VQABS/VQABSQ cumulative saturation output: +VQABS/VQABSQ:0:vqabs_s8 Neon cumulative saturation 0 +VQABS/VQABSQ:1:vqabs_s16 Neon cumulative saturation 0 +VQABS/VQABSQ:2:vqabs_s32 Neon cumulative saturation 0 +VQABS/VQABSQ:3:vqabsq_s8 Neon cumulative saturation 0 +VQABS/VQABSQ:4:vqabsq_s16 Neon cumulative saturation 0 +VQABS/VQABSQ:5:vqabsq_s32 Neon cumulative saturation 0 VQABS/VQABSQ output: VQABS/VQABSQ:6:result_int8x8 [] = { 10, f, e, d, c, b, a, 9, } @@ -2716,13 +2716,13 @@ VQABS/VQABSQ:25:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQABS/VQABSQ:26:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQABS/VQABSQ:27:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQABS/VQABSQ overflow output: -VQABS/VQABSQ:0:vqabs_s8 Neon overflow 1 -VQABS/VQABSQ:1:vqabs_s16 Neon overflow 1 -VQABS/VQABSQ:2:vqabs_s32 Neon overflow 1 -VQABS/VQABSQ:3:vqabsq_s8 Neon overflow 1 -VQABS/VQABSQ:4:vqabsq_s16 Neon overflow 1 -VQABS/VQABSQ:5:vqabsq_s32 Neon overflow 1 +VQABS/VQABSQ cumulative saturation output: +VQABS/VQABSQ:0:vqabs_s8 Neon cumulative saturation 1 +VQABS/VQABSQ:1:vqabs_s16 Neon cumulative saturation 1 +VQABS/VQABSQ:2:vqabs_s32 Neon cumulative saturation 1 +VQABS/VQABSQ:3:vqabsq_s8 Neon cumulative saturation 1 +VQABS/VQABSQ:4:vqabsq_s16 Neon cumulative saturation 1 +VQABS/VQABSQ:5:vqabsq_s32 Neon cumulative saturation 1 VQABS/VQABSQ output: VQABS/VQABSQ:6:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -2860,13 +2860,13 @@ float32: VNEG/VNEGQ:22:result_float32x2 [] = { c0133333 -0x1.2666660p+1 -2.3, c0133333 -0x1.2666660p+1 -2.3, } VNEG/VNEGQ:23:result_float32x4 [] = { c059999a -0x1.b333340p+1 -3.4, c059999a -0x1.b333340p+1 -3.4, c059999a -0x1.b333340p+1 -3.4, c059999a -0x1.b333340p+1 -3.4, } -VQNEG/VQNEGQ overflow output: -VQNEG/VQNEGQ:0:vqneg_s8 Neon overflow 0 -VQNEG/VQNEGQ:1:vqneg_s16 Neon overflow 0 -VQNEG/VQNEGQ:2:vqneg_s32 Neon overflow 0 -VQNEG/VQNEGQ:3:vqnegq_s8 Neon overflow 0 -VQNEG/VQNEGQ:4:vqnegq_s16 Neon overflow 0 -VQNEG/VQNEGQ:5:vqnegq_s32 Neon overflow 0 +VQNEG/VQNEGQ cumulative saturation output: +VQNEG/VQNEGQ:0:vqneg_s8 Neon cumulative saturation 0 +VQNEG/VQNEGQ:1:vqneg_s16 Neon cumulative saturation 0 +VQNEG/VQNEGQ:2:vqneg_s32 Neon cumulative saturation 0 +VQNEG/VQNEGQ:3:vqnegq_s8 Neon cumulative saturation 0 +VQNEG/VQNEGQ:4:vqnegq_s16 Neon cumulative saturation 0 +VQNEG/VQNEGQ:5:vqnegq_s32 Neon cumulative saturation 0 VQNEG/VQNEGQ output: VQNEG/VQNEGQ:6:result_int8x8 [] = { 10, f, e, d, c, b, a, 9, } @@ -2892,13 +2892,13 @@ VQNEG/VQNEGQ:25:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQNEG/VQNEGQ:26:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQNEG/VQNEGQ:27:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQNEG/VQNEGQ overflow output: -VQNEG/VQNEGQ:0:vqneg_s8 Neon overflow 1 -VQNEG/VQNEGQ:1:vqneg_s16 Neon overflow 1 -VQNEG/VQNEGQ:2:vqneg_s32 Neon overflow 1 -VQNEG/VQNEGQ:3:vqnegq_s8 Neon overflow 1 -VQNEG/VQNEGQ:4:vqnegq_s16 Neon overflow 1 -VQNEG/VQNEGQ:5:vqnegq_s32 Neon overflow 1 +VQNEG/VQNEGQ cumulative saturation output: +VQNEG/VQNEGQ:0:vqneg_s8 Neon cumulative saturation 1 +VQNEG/VQNEGQ:1:vqneg_s16 Neon cumulative saturation 1 +VQNEG/VQNEGQ:2:vqneg_s32 Neon cumulative saturation 1 +VQNEG/VQNEGQ:3:vqnegq_s8 Neon cumulative saturation 1 +VQNEG/VQNEGQ:4:vqnegq_s16 Neon cumulative saturation 1 +VQNEG/VQNEGQ:5:vqnegq_s32 Neon cumulative saturation 1 VQNEG/VQNEGQ output: VQNEG/VQNEGQ:6:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -3608,11 +3608,11 @@ VREINTERPRET/VREINTERPRETQ:199:result_uint64x2 [] = { c1700000c1800000, c1500000 VREINTERPRET/VREINTERPRETQ:200:result_poly8x16 [] = { 0, 0, 80, c1, 0, 0, 70, c1, 0, 0, 60, c1, 0, 0, 50, c1, } VREINTERPRET/VREINTERPRETQ:201:result_poly16x8 [] = { 0, c180, 0, c170, 0, c160, 0, c150, } -VQRDMULH overflow output: -VQRDMULH:0:vqrdmulh_s16 Neon overflow 0 -VQRDMULH:1:vqrdmulh_s32 Neon overflow 0 -VQRDMULH:2:vqrdmulhq_s16 Neon overflow 0 -VQRDMULH:3:vqrdmulhq_s32 Neon overflow 0 +VQRDMULH cumulative saturation output: +VQRDMULH:0:vqrdmulh_s16 Neon cumulative saturation 0 +VQRDMULH:1:vqrdmulh_s32 Neon cumulative saturation 0 +VQRDMULH:2:vqrdmulhq_s16 Neon cumulative saturation 0 +VQRDMULH:3:vqrdmulhq_s32 Neon cumulative saturation 0 VQRDMULH output: VQRDMULH:4:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -3638,13 +3638,13 @@ VQRDMULH:23:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQRDMULH:24:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRDMULH:25:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRDMULH (check mul overflow) overflow output: -VQRDMULH:26:vqrdmulh_s16 Neon overflow 1 -VQRDMULH:27:vqrdmulh_s32 Neon overflow 1 -VQRDMULH:28:vqrdmulhq_s16 Neon overflow 1 -VQRDMULH:29:vqrdmulhq_s32 Neon overflow 1 +VQRDMULH (check mul cumulative saturation) cumulative saturation output: +VQRDMULH:26:vqrdmulh_s16 Neon cumulative saturation 1 +VQRDMULH:27:vqrdmulh_s32 Neon cumulative saturation 1 +VQRDMULH:28:vqrdmulhq_s16 Neon cumulative saturation 1 +VQRDMULH:29:vqrdmulhq_s32 Neon cumulative saturation 1 -VQRDMULH (check mul overflow) output: +VQRDMULH (check mul cumulative saturation) output: VQRDMULH:30:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRDMULH:31:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQRDMULH:32:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -3668,13 +3668,13 @@ VQRDMULH:49:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQRDMULH:50:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRDMULH:51:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRDMULH (check rounding overflow) overflow output: -VQRDMULH:52:vqrdmulh_s16 Neon overflow 0 -VQRDMULH:53:vqrdmulh_s32 Neon overflow 0 -VQRDMULH:54:vqrdmulhq_s16 Neon overflow 0 -VQRDMULH:55:vqrdmulhq_s32 Neon overflow 0 +VQRDMULH (check rounding cumulative saturation) cumulative saturation output: +VQRDMULH:52:vqrdmulh_s16 Neon cumulative saturation 0 +VQRDMULH:53:vqrdmulh_s32 Neon cumulative saturation 0 +VQRDMULH:54:vqrdmulhq_s16 Neon cumulative saturation 0 +VQRDMULH:55:vqrdmulhq_s32 Neon cumulative saturation 0 -VQRDMULH (check rounding overflow) output: +VQRDMULH (check rounding cumulative saturation) output: VQRDMULH:56:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRDMULH:57:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQRDMULH:58:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -3698,11 +3698,11 @@ VQRDMULH:75:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQRDMULH:76:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRDMULH:77:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRDMULH_LANE overflow output: -VQRDMULH_LANE:0:vqrdmulh_lane_s16 Neon overflow 0 -VQRDMULH_LANE:1:vqrdmulh_lane_s32 Neon overflow 0 -VQRDMULH_LANE:2:vqrdmulhq_lane_s16 Neon overflow 0 -VQRDMULH_LANE:3:vqrdmulhq_lane_s32 Neon overflow 0 +VQRDMULH_LANE cumulative saturation output: +VQRDMULH_LANE:0:vqrdmulh_lane_s16 Neon cumulative saturation 0 +VQRDMULH_LANE:1:vqrdmulh_lane_s32 Neon cumulative saturation 0 +VQRDMULH_LANE:2:vqrdmulhq_lane_s16 Neon cumulative saturation 0 +VQRDMULH_LANE:3:vqrdmulhq_lane_s32 Neon cumulative saturation 0 VQRDMULH_LANE output: VQRDMULH_LANE:4:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -3728,13 +3728,13 @@ VQRDMULH_LANE:23:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQRDMULH_LANE:24:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRDMULH_LANE:25:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRDMULH_LANE (check mul overflow) overflow output: -VQRDMULH_LANE:26:vqrdmulh_lane_s16 Neon overflow 1 -VQRDMULH_LANE:27:vqrdmulh_lane_s32 Neon overflow 1 -VQRDMULH_LANE:28:vqrdmulhq_lane_s16 Neon overflow 1 -VQRDMULH_LANE:29:vqrdmulhq_lane_s32 Neon overflow 1 +VQRDMULH_LANE (check mul cumulative saturation) cumulative saturation output: +VQRDMULH_LANE:26:vqrdmulh_lane_s16 Neon cumulative saturation 1 +VQRDMULH_LANE:27:vqrdmulh_lane_s32 Neon cumulative saturation 1 +VQRDMULH_LANE:28:vqrdmulhq_lane_s16 Neon cumulative saturation 1 +VQRDMULH_LANE:29:vqrdmulhq_lane_s32 Neon cumulative saturation 1 -VQRDMULH_LANE (check mul overflow) output: +VQRDMULH_LANE (check mul cumulative saturation) output: VQRDMULH_LANE:30:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRDMULH_LANE:31:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQRDMULH_LANE:32:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -3758,13 +3758,13 @@ VQRDMULH_LANE:49:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQRDMULH_LANE:50:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRDMULH_LANE:51:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRDMULH_LANE (check rounding overflow) overflow output: -VQRDMULH_LANE:52:vqrdmulh_lane_s16 Neon overflow 0 -VQRDMULH_LANE:53:vqrdmulh_lane_s32 Neon overflow 0 -VQRDMULH_LANE:54:vqrdmulhq_lane_s16 Neon overflow 0 -VQRDMULH_LANE:55:vqrdmulhq_lane_s32 Neon overflow 0 +VQRDMULH_LANE (check rounding cumulative saturation) cumulative saturation output: +VQRDMULH_LANE:52:vqrdmulh_lane_s16 Neon cumulative saturation 0 +VQRDMULH_LANE:53:vqrdmulh_lane_s32 Neon cumulative saturation 0 +VQRDMULH_LANE:54:vqrdmulhq_lane_s16 Neon cumulative saturation 0 +VQRDMULH_LANE:55:vqrdmulhq_lane_s32 Neon cumulative saturation 0 -VQRDMULH_LANE (check rounding overflow) output: +VQRDMULH_LANE (check rounding cumulative saturation) output: VQRDMULH_LANE:56:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRDMULH_LANE:57:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQRDMULH_LANE:58:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -3788,11 +3788,11 @@ VQRDMULH_LANE:75:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQRDMULH_LANE:76:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRDMULH_LANE:77:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRDMULH_N overflow output: -VQRDMULH_N:0:vqrdmulh_n_s16 Neon overflow 0 -VQRDMULH_N:1:vqrdmulh_n_s32 Neon overflow 0 -VQRDMULH_N:2:vqrdmulhq_n_s16 Neon overflow 0 -VQRDMULH_N:3:vqrdmulhq_n_s32 Neon overflow 0 +VQRDMULH_N cumulative saturation output: +VQRDMULH_N:0:vqrdmulh_n_s16 Neon cumulative saturation 0 +VQRDMULH_N:1:vqrdmulh_n_s32 Neon cumulative saturation 0 +VQRDMULH_N:2:vqrdmulhq_n_s16 Neon cumulative saturation 0 +VQRDMULH_N:3:vqrdmulhq_n_s32 Neon cumulative saturation 0 VQRDMULH_N output: VQRDMULH_N:4:result_int16x4 [] = { fffffffc, fffffffc, fffffffc, fffffffd, } @@ -3800,13 +3800,13 @@ VQRDMULH_N:5:result_int32x2 [] = { fffffffe, fffffffe, } VQRDMULH_N:6:result_int16x8 [] = { 6, 6, 6, 5, 5, 4, 4, 4, } VQRDMULH_N:7:result_int32x4 [] = { fffffffe, fffffffe, fffffffe, fffffffe, } -VQRDMULH_N (check mul overflow) overflow output: -VQRDMULH_N:8:vqrdmulh_n_s16 Neon overflow 1 -VQRDMULH_N:9:vqrdmulh_n_s32 Neon overflow 1 -VQRDMULH_N:10:vqrdmulhq_n_s16 Neon overflow 1 -VQRDMULH_N:11:vqrdmulhq_n_s32 Neon overflow 1 +VQRDMULH_N (check mul cumulative saturation) cumulative saturation output: +VQRDMULH_N:8:vqrdmulh_n_s16 Neon cumulative saturation 1 +VQRDMULH_N:9:vqrdmulh_n_s32 Neon cumulative saturation 1 +VQRDMULH_N:10:vqrdmulhq_n_s16 Neon cumulative saturation 1 +VQRDMULH_N:11:vqrdmulhq_n_s32 Neon cumulative saturation 1 -VQRDMULH_N (check mul overflow) output: +VQRDMULH_N (check mul cumulative saturation) output: VQRDMULH_N:12:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRDMULH_N:13:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQRDMULH_N:14:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -3830,13 +3830,13 @@ VQRDMULH_N:31:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQRDMULH_N:32:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRDMULH_N:33:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRDMULH_N (check rounding overflow) overflow output: -VQRDMULH_N:34:vqrdmulh_n_s16 Neon overflow 0 -VQRDMULH_N:35:vqrdmulh_n_s32 Neon overflow 0 -VQRDMULH_N:36:vqrdmulhq_n_s16 Neon overflow 0 -VQRDMULH_N:37:vqrdmulhq_n_s32 Neon overflow 0 +VQRDMULH_N (check rounding cumulative saturation) cumulative saturation output: +VQRDMULH_N:34:vqrdmulh_n_s16 Neon cumulative saturation 0 +VQRDMULH_N:35:vqrdmulh_n_s32 Neon cumulative saturation 0 +VQRDMULH_N:36:vqrdmulhq_n_s16 Neon cumulative saturation 0 +VQRDMULH_N:37:vqrdmulhq_n_s32 Neon cumulative saturation 0 -VQRDMULH_N (check rounding overflow) output: +VQRDMULH_N (check rounding cumulative saturation) output: VQRDMULH_N:38:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRDMULH_N:39:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQRDMULH_N:40:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -3860,23 +3860,23 @@ VQRDMULH_N:57:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQRDMULH_N:58:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRDMULH_N:59:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHL/VQRSHLQ (with input = 0) overflow output: -VQRSHL/VQRSHLQ:0:vqrshl_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:1:vqrshl_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:2:vqrshl_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:3:vqrshl_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:4:vqrshl_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:5:vqrshl_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:6:vqrshl_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:7:vqrshl_u64 Neon overflow 0 -VQRSHL/VQRSHLQ:8:vqrshlq_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:9:vqrshlq_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:10:vqrshlq_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:11:vqrshlq_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:12:vqrshlq_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:13:vqrshlq_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:14:vqrshlq_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:15:vqrshlq_u64 Neon overflow 0 +VQRSHL/VQRSHLQ (with input = 0) cumulative saturation output: +VQRSHL/VQRSHLQ:0:vqrshl_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:1:vqrshl_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:2:vqrshl_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:3:vqrshl_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:4:vqrshl_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:5:vqrshl_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:6:vqrshl_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:7:vqrshl_u64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:8:vqrshlq_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:9:vqrshlq_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:10:vqrshlq_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:11:vqrshlq_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:12:vqrshlq_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:13:vqrshlq_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:14:vqrshlq_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:15:vqrshlq_u64 Neon cumulative saturation 0 VQRSHL/VQRSHLQ (with input = 0) output: VQRSHL/VQRSHLQ:16:result_int8x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } @@ -3902,23 +3902,23 @@ VQRSHL/VQRSHLQ:35:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQRSHL/VQRSHLQ:36:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRSHL/VQRSHLQ:37:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHL/VQRSHLQ (input 0 and negative shift amount) overflow output: -VQRSHL/VQRSHLQ:38:vqrshl_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:39:vqrshl_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:40:vqrshl_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:41:vqrshl_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:42:vqrshl_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:43:vqrshl_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:44:vqrshl_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:45:vqrshl_u64 Neon overflow 0 -VQRSHL/VQRSHLQ:46:vqrshlq_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:47:vqrshlq_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:48:vqrshlq_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:49:vqrshlq_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:50:vqrshlq_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:51:vqrshlq_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:52:vqrshlq_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:53:vqrshlq_u64 Neon overflow 0 +VQRSHL/VQRSHLQ (input 0 and negative shift amount) cumulative saturation output: +VQRSHL/VQRSHLQ:38:vqrshl_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:39:vqrshl_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:40:vqrshl_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:41:vqrshl_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:42:vqrshl_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:43:vqrshl_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:44:vqrshl_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:45:vqrshl_u64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:46:vqrshlq_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:47:vqrshlq_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:48:vqrshlq_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:49:vqrshlq_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:50:vqrshlq_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:51:vqrshlq_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:52:vqrshlq_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:53:vqrshlq_u64 Neon cumulative saturation 0 VQRSHL/VQRSHLQ (input 0 and negative shift amount) output: VQRSHL/VQRSHLQ:54:result_int8x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } @@ -3944,23 +3944,23 @@ VQRSHL/VQRSHLQ:73:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQRSHL/VQRSHLQ:74:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRSHL/VQRSHLQ:75:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHL/VQRSHLQ overflow output: -VQRSHL/VQRSHLQ:76:vqrshl_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:77:vqrshl_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:78:vqrshl_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:79:vqrshl_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:80:vqrshl_u8 Neon overflow 1 -VQRSHL/VQRSHLQ:81:vqrshl_u16 Neon overflow 1 -VQRSHL/VQRSHLQ:82:vqrshl_u32 Neon overflow 1 -VQRSHL/VQRSHLQ:83:vqrshl_u64 Neon overflow 1 -VQRSHL/VQRSHLQ:84:vqrshlq_s8 Neon overflow 1 -VQRSHL/VQRSHLQ:85:vqrshlq_s16 Neon overflow 1 -VQRSHL/VQRSHLQ:86:vqrshlq_s32 Neon overflow 1 -VQRSHL/VQRSHLQ:87:vqrshlq_s64 Neon overflow 1 -VQRSHL/VQRSHLQ:88:vqrshlq_u8 Neon overflow 1 -VQRSHL/VQRSHLQ:89:vqrshlq_u16 Neon overflow 1 -VQRSHL/VQRSHLQ:90:vqrshlq_u32 Neon overflow 1 -VQRSHL/VQRSHLQ:91:vqrshlq_u64 Neon overflow 1 +VQRSHL/VQRSHLQ cumulative saturation output: +VQRSHL/VQRSHLQ:76:vqrshl_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:77:vqrshl_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:78:vqrshl_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:79:vqrshl_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:80:vqrshl_u8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:81:vqrshl_u16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:82:vqrshl_u32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:83:vqrshl_u64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:84:vqrshlq_s8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:85:vqrshlq_s16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:86:vqrshlq_s32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:87:vqrshlq_s64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:88:vqrshlq_u8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:89:vqrshlq_u16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:90:vqrshlq_u32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:91:vqrshlq_u64 Neon cumulative saturation 1 VQRSHL/VQRSHLQ output: VQRSHL/VQRSHLQ:92:result_int8x8 [] = { ffffffe0, ffffffe2, ffffffe4, ffffffe6, ffffffe8, ffffffea, ffffffec, ffffffee, } @@ -3986,23 +3986,23 @@ VQRSHL/VQRSHLQ:111:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33 VQRSHL/VQRSHLQ:112:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRSHL/VQRSHLQ:113:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHL/VQRSHLQ (negative shift amount) overflow output: -VQRSHL/VQRSHLQ:114:vqrshl_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:115:vqrshl_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:116:vqrshl_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:117:vqrshl_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:118:vqrshl_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:119:vqrshl_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:120:vqrshl_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:121:vqrshl_u64 Neon overflow 0 -VQRSHL/VQRSHLQ:122:vqrshlq_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:123:vqrshlq_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:124:vqrshlq_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:125:vqrshlq_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:126:vqrshlq_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:127:vqrshlq_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:128:vqrshlq_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:129:vqrshlq_u64 Neon overflow 0 +VQRSHL/VQRSHLQ (negative shift amount) cumulative saturation output: +VQRSHL/VQRSHLQ:114:vqrshl_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:115:vqrshl_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:116:vqrshl_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:117:vqrshl_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:118:vqrshl_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:119:vqrshl_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:120:vqrshl_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:121:vqrshl_u64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:122:vqrshlq_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:123:vqrshlq_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:124:vqrshlq_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:125:vqrshlq_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:126:vqrshlq_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:127:vqrshlq_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:128:vqrshlq_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:129:vqrshlq_u64 Neon cumulative saturation 0 VQRSHL/VQRSHLQ (negative shift amount) output: VQRSHL/VQRSHLQ:130:result_int8x8 [] = { fffffffc, fffffffc, fffffffd, fffffffd, fffffffd, fffffffd, fffffffe, fffffffe, } @@ -4028,25 +4028,25 @@ VQRSHL/VQRSHLQ:149:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33 VQRSHL/VQRSHLQ:150:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRSHL/VQRSHLQ:151:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHL/VQRSHLQ (checking overflow: shift by -1) overflow output: -VQRSHL/VQRSHLQ:152:vqrshl_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:153:vqrshl_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:154:vqrshl_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:155:vqrshl_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:156:vqrshl_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:157:vqrshl_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:158:vqrshl_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:159:vqrshl_u64 Neon overflow 0 -VQRSHL/VQRSHLQ:160:vqrshlq_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:161:vqrshlq_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:162:vqrshlq_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:163:vqrshlq_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:164:vqrshlq_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:165:vqrshlq_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:166:vqrshlq_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:167:vqrshlq_u64 Neon overflow 0 - -VQRSHL/VQRSHLQ (checking overflow: shift by -1) output: +VQRSHL/VQRSHLQ (checking cumulative saturation: shift by -1) cumulative saturation output: +VQRSHL/VQRSHLQ:152:vqrshl_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:153:vqrshl_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:154:vqrshl_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:155:vqrshl_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:156:vqrshl_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:157:vqrshl_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:158:vqrshl_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:159:vqrshl_u64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:160:vqrshlq_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:161:vqrshlq_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:162:vqrshlq_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:163:vqrshlq_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:164:vqrshlq_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:165:vqrshlq_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:166:vqrshlq_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:167:vqrshlq_u64 Neon cumulative saturation 0 + +VQRSHL/VQRSHLQ (checking cumulative saturation: shift by -1) output: VQRSHL/VQRSHLQ:168:result_int8x8 [] = { 40, 40, 40, 40, 40, 40, 40, 40, } VQRSHL/VQRSHLQ:169:result_int16x4 [] = { 4000, 4000, 4000, 4000, } VQRSHL/VQRSHLQ:170:result_int32x2 [] = { 40000000, 40000000, } @@ -4070,25 +4070,25 @@ VQRSHL/VQRSHLQ:187:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33 VQRSHL/VQRSHLQ:188:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRSHL/VQRSHLQ:189:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHL/VQRSHLQ (checking overflow: shift by -3) overflow output: -VQRSHL/VQRSHLQ:190:vqrshl_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:191:vqrshl_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:192:vqrshl_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:193:vqrshl_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:194:vqrshl_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:195:vqrshl_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:196:vqrshl_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:197:vqrshl_u64 Neon overflow 0 -VQRSHL/VQRSHLQ:198:vqrshlq_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:199:vqrshlq_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:200:vqrshlq_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:201:vqrshlq_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:202:vqrshlq_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:203:vqrshlq_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:204:vqrshlq_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:205:vqrshlq_u64 Neon overflow 0 - -VQRSHL/VQRSHLQ (checking overflow: shift by -3) output: +VQRSHL/VQRSHLQ (checking cumulative saturation: shift by -3) cumulative saturation output: +VQRSHL/VQRSHLQ:190:vqrshl_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:191:vqrshl_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:192:vqrshl_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:193:vqrshl_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:194:vqrshl_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:195:vqrshl_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:196:vqrshl_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:197:vqrshl_u64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:198:vqrshlq_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:199:vqrshlq_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:200:vqrshlq_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:201:vqrshlq_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:202:vqrshlq_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:203:vqrshlq_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:204:vqrshlq_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:205:vqrshlq_u64 Neon cumulative saturation 0 + +VQRSHL/VQRSHLQ (checking cumulative saturation: shift by -3) output: VQRSHL/VQRSHLQ:206:result_int8x8 [] = { 10, 10, 10, 10, 10, 10, 10, 10, } VQRSHL/VQRSHLQ:207:result_int16x4 [] = { 1000, 1000, 1000, 1000, } VQRSHL/VQRSHLQ:208:result_int32x2 [] = { 10000000, 10000000, } @@ -4112,25 +4112,25 @@ VQRSHL/VQRSHLQ:225:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33 VQRSHL/VQRSHLQ:226:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRSHL/VQRSHLQ:227:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHL/VQRSHLQ (checking overflow: large shift amount) overflow output: -VQRSHL/VQRSHLQ:228:vqrshl_s8 Neon overflow 1 -VQRSHL/VQRSHLQ:229:vqrshl_s16 Neon overflow 1 -VQRSHL/VQRSHLQ:230:vqrshl_s32 Neon overflow 1 -VQRSHL/VQRSHLQ:231:vqrshl_s64 Neon overflow 1 -VQRSHL/VQRSHLQ:232:vqrshl_u8 Neon overflow 1 -VQRSHL/VQRSHLQ:233:vqrshl_u16 Neon overflow 1 -VQRSHL/VQRSHLQ:234:vqrshl_u32 Neon overflow 1 -VQRSHL/VQRSHLQ:235:vqrshl_u64 Neon overflow 1 -VQRSHL/VQRSHLQ:236:vqrshlq_s8 Neon overflow 1 -VQRSHL/VQRSHLQ:237:vqrshlq_s16 Neon overflow 1 -VQRSHL/VQRSHLQ:238:vqrshlq_s32 Neon overflow 1 -VQRSHL/VQRSHLQ:239:vqrshlq_s64 Neon overflow 1 -VQRSHL/VQRSHLQ:240:vqrshlq_u8 Neon overflow 1 -VQRSHL/VQRSHLQ:241:vqrshlq_u16 Neon overflow 1 -VQRSHL/VQRSHLQ:242:vqrshlq_u32 Neon overflow 1 -VQRSHL/VQRSHLQ:243:vqrshlq_u64 Neon overflow 1 - -VQRSHL/VQRSHLQ (checking overflow: large shift amount) output: +VQRSHL/VQRSHLQ (checking cumulative saturation: large shift amount) cumulative saturation output: +VQRSHL/VQRSHLQ:228:vqrshl_s8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:229:vqrshl_s16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:230:vqrshl_s32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:231:vqrshl_s64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:232:vqrshl_u8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:233:vqrshl_u16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:234:vqrshl_u32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:235:vqrshl_u64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:236:vqrshlq_s8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:237:vqrshlq_s16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:238:vqrshlq_s32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:239:vqrshlq_s64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:240:vqrshlq_u8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:241:vqrshlq_u16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:242:vqrshlq_u32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:243:vqrshlq_u64 Neon cumulative saturation 1 + +VQRSHL/VQRSHLQ (checking cumulative saturation: large shift amount) output: VQRSHL/VQRSHLQ:244:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } VQRSHL/VQRSHLQ:245:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQRSHL/VQRSHLQ:246:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -4154,25 +4154,25 @@ VQRSHL/VQRSHLQ:263:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33 VQRSHL/VQRSHLQ:264:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRSHL/VQRSHLQ:265:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHL/VQRSHLQ (checking overflow: large shift amount with negative input) overflow output: -VQRSHL/VQRSHLQ:266:vqrshl_s8 Neon overflow 1 -VQRSHL/VQRSHLQ:267:vqrshl_s16 Neon overflow 1 -VQRSHL/VQRSHLQ:268:vqrshl_s32 Neon overflow 1 -VQRSHL/VQRSHLQ:269:vqrshl_s64 Neon overflow 1 -VQRSHL/VQRSHLQ:270:vqrshl_u8 Neon overflow 1 -VQRSHL/VQRSHLQ:271:vqrshl_u16 Neon overflow 1 -VQRSHL/VQRSHLQ:272:vqrshl_u32 Neon overflow 1 -VQRSHL/VQRSHLQ:273:vqrshl_u64 Neon overflow 1 -VQRSHL/VQRSHLQ:274:vqrshlq_s8 Neon overflow 1 -VQRSHL/VQRSHLQ:275:vqrshlq_s16 Neon overflow 1 -VQRSHL/VQRSHLQ:276:vqrshlq_s32 Neon overflow 1 -VQRSHL/VQRSHLQ:277:vqrshlq_s64 Neon overflow 1 -VQRSHL/VQRSHLQ:278:vqrshlq_u8 Neon overflow 1 -VQRSHL/VQRSHLQ:279:vqrshlq_u16 Neon overflow 1 -VQRSHL/VQRSHLQ:280:vqrshlq_u32 Neon overflow 1 -VQRSHL/VQRSHLQ:281:vqrshlq_u64 Neon overflow 1 - -VQRSHL/VQRSHLQ (checking overflow: large shift amount with negative input) output: +VQRSHL/VQRSHLQ (checking cumulative saturation: large shift amount with negative input) cumulative saturation output: +VQRSHL/VQRSHLQ:266:vqrshl_s8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:267:vqrshl_s16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:268:vqrshl_s32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:269:vqrshl_s64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:270:vqrshl_u8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:271:vqrshl_u16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:272:vqrshl_u32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:273:vqrshl_u64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:274:vqrshlq_s8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:275:vqrshlq_s16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:276:vqrshlq_s32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:277:vqrshlq_s64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:278:vqrshlq_u8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:279:vqrshlq_u16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:280:vqrshlq_u32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:281:vqrshlq_u64 Neon cumulative saturation 1 + +VQRSHL/VQRSHLQ (checking cumulative saturation: large shift amount with negative input) output: VQRSHL/VQRSHLQ:282:result_int8x8 [] = { ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, } VQRSHL/VQRSHLQ:283:result_int16x4 [] = { ffff8000, ffff8000, ffff8000, ffff8000, } VQRSHL/VQRSHLQ:284:result_int32x2 [] = { 80000000, 80000000, } @@ -4196,25 +4196,25 @@ VQRSHL/VQRSHLQ:301:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33 VQRSHL/VQRSHLQ:302:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRSHL/VQRSHLQ:303:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHL/VQRSHLQ (checking overflow: large negative shift amount) overflow output: -VQRSHL/VQRSHLQ:304:vqrshl_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:305:vqrshl_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:306:vqrshl_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:307:vqrshl_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:308:vqrshl_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:309:vqrshl_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:310:vqrshl_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:311:vqrshl_u64 Neon overflow 0 -VQRSHL/VQRSHLQ:312:vqrshlq_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:313:vqrshlq_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:314:vqrshlq_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:315:vqrshlq_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:316:vqrshlq_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:317:vqrshlq_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:318:vqrshlq_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:319:vqrshlq_u64 Neon overflow 0 - -VQRSHL/VQRSHLQ (checking overflow: large negative shift amount) output: +VQRSHL/VQRSHLQ (checking cumulative saturation: large negative shift amount) cumulative saturation output: +VQRSHL/VQRSHLQ:304:vqrshl_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:305:vqrshl_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:306:vqrshl_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:307:vqrshl_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:308:vqrshl_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:309:vqrshl_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:310:vqrshl_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:311:vqrshl_u64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:312:vqrshlq_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:313:vqrshlq_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:314:vqrshlq_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:315:vqrshlq_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:316:vqrshlq_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:317:vqrshlq_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:318:vqrshlq_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:319:vqrshlq_u64 Neon cumulative saturation 0 + +VQRSHL/VQRSHLQ (checking cumulative saturation: large negative shift amount) output: VQRSHL/VQRSHLQ:320:result_int8x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } VQRSHL/VQRSHLQ:321:result_int16x4 [] = { 0, 0, 0, 0, } VQRSHL/VQRSHLQ:322:result_int32x2 [] = { 0, 0, } @@ -4238,25 +4238,25 @@ VQRSHL/VQRSHLQ:339:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33 VQRSHL/VQRSHLQ:340:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRSHL/VQRSHLQ:341:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHL/VQRSHLQ (checking overflow: large shift amount with 0 input) overflow output: -VQRSHL/VQRSHLQ:342:vqrshl_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:343:vqrshl_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:344:vqrshl_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:345:vqrshl_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:346:vqrshl_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:347:vqrshl_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:348:vqrshl_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:349:vqrshl_u64 Neon overflow 0 -VQRSHL/VQRSHLQ:350:vqrshlq_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:351:vqrshlq_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:352:vqrshlq_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:353:vqrshlq_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:354:vqrshlq_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:355:vqrshlq_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:356:vqrshlq_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:357:vqrshlq_u64 Neon overflow 0 - -VQRSHL/VQRSHLQ (checking overflow: large shift amount with 0 input) output: +VQRSHL/VQRSHLQ (checking cumulative saturation: large shift amount with 0 input) cumulative saturation output: +VQRSHL/VQRSHLQ:342:vqrshl_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:343:vqrshl_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:344:vqrshl_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:345:vqrshl_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:346:vqrshl_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:347:vqrshl_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:348:vqrshl_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:349:vqrshl_u64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:350:vqrshlq_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:351:vqrshlq_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:352:vqrshlq_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:353:vqrshlq_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:354:vqrshlq_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:355:vqrshlq_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:356:vqrshlq_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:357:vqrshlq_u64 Neon cumulative saturation 0 + +VQRSHL/VQRSHLQ (checking cumulative saturation: large shift amount with 0 input) output: VQRSHL/VQRSHLQ:358:result_int8x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } VQRSHL/VQRSHLQ:359:result_int16x4 [] = { 0, 0, 0, 0, } VQRSHL/VQRSHLQ:360:result_int32x2 [] = { 0, 0, } @@ -5632,13 +5632,13 @@ VMVN/VMVNQ:19:result_poly8x16 [] = { f, e, d, c, b, a, 9, 8, 7, 6, 5, 4, 3, 2, 1 VMVN/VMVNQ:20:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VMVN/VMVNQ:21:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQMOVN overflow output: -VQMOVN:0:vqmovn_s16 Neon overflow 0 -VQMOVN:1:vqmovn_s32 Neon overflow 0 -VQMOVN:2:vqmovn_s64 Neon overflow 0 -VQMOVN:3:vqmovn_u16 Neon overflow 0 -VQMOVN:4:vqmovn_u32 Neon overflow 0 -VQMOVN:5:vqmovn_u64 Neon overflow 0 +VQMOVN cumulative saturation output: +VQMOVN:0:vqmovn_s16 Neon cumulative saturation 0 +VQMOVN:1:vqmovn_s32 Neon cumulative saturation 0 +VQMOVN:2:vqmovn_s64 Neon cumulative saturation 0 +VQMOVN:3:vqmovn_u16 Neon cumulative saturation 0 +VQMOVN:4:vqmovn_u32 Neon cumulative saturation 0 +VQMOVN:5:vqmovn_u64 Neon cumulative saturation 0 VQMOVN output: VQMOVN:6:result_int8x8 [] = { 12, 12, 12, 12, 12, 12, 12, 12, } @@ -5664,13 +5664,13 @@ VQMOVN:25:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQMOVN:26:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQMOVN:27:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQMOVN overflow output: -VQMOVN:28:vqmovn_s16 Neon overflow 1 -VQMOVN:29:vqmovn_s32 Neon overflow 1 -VQMOVN:30:vqmovn_s64 Neon overflow 1 -VQMOVN:31:vqmovn_u16 Neon overflow 1 -VQMOVN:32:vqmovn_u32 Neon overflow 1 -VQMOVN:33:vqmovn_u64 Neon overflow 1 +VQMOVN cumulative saturation output: +VQMOVN:28:vqmovn_s16 Neon cumulative saturation 1 +VQMOVN:29:vqmovn_s32 Neon cumulative saturation 1 +VQMOVN:30:vqmovn_s64 Neon cumulative saturation 1 +VQMOVN:31:vqmovn_u16 Neon cumulative saturation 1 +VQMOVN:32:vqmovn_u32 Neon cumulative saturation 1 +VQMOVN:33:vqmovn_u64 Neon cumulative saturation 1 VQMOVN output: VQMOVN:34:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -5696,10 +5696,10 @@ VQMOVN:53:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQMOVN:54:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQMOVN:55:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQMOVUN overflow output: -VQMOVUN:0:vqmovun_s16 Neon overflow 0 -VQMOVUN:1:vqmovun_s32 Neon overflow 0 -VQMOVUN:2:vqmovun_s64 Neon overflow 0 +VQMOVUN cumulative saturation output: +VQMOVUN:0:vqmovun_s16 Neon cumulative saturation 0 +VQMOVUN:1:vqmovun_s32 Neon cumulative saturation 0 +VQMOVUN:2:vqmovun_s64 Neon cumulative saturation 0 VQMOVUN output: VQMOVUN:3:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -5725,10 +5725,10 @@ VQMOVUN:22:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33 VQMOVUN:23:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQMOVUN:24:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQMOVUN (negative input) overflow output: -VQMOVUN:25:vqmovun_s16 Neon overflow 1 -VQMOVUN:26:vqmovun_s32 Neon overflow 1 -VQMOVUN:27:vqmovun_s64 Neon overflow 1 +VQMOVUN (negative input) cumulative saturation output: +VQMOVUN:25:vqmovun_s16 Neon cumulative saturation 1 +VQMOVUN:26:vqmovun_s32 Neon cumulative saturation 1 +VQMOVUN:27:vqmovun_s64 Neon cumulative saturation 1 VQMOVUN (negative input) output: VQMOVUN:28:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -6162,15 +6162,15 @@ VPADAL/VPADALQ:19:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VPADAL/VPADALQ:20:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VPADAL/VPADALQ:21:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHLU_N/VQSHLUQ_N (negative input) overflow output: -VQSHLU_N/VQSHLUQ_N:0:vqshlu_n_s8 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:1:vqshlu_n_s16 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:2:vqshlu_n_s32 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:3:vqshlu_n_s64 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:4:vqshluq_n_s8 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:5:vqshluq_n_s16 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:6:vqshluq_n_s32 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:7:vqshluq_n_s64 Neon overflow 1 +VQSHLU_N/VQSHLUQ_N (negative input) cumulative saturation output: +VQSHLU_N/VQSHLUQ_N:0:vqshlu_n_s8 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:1:vqshlu_n_s16 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:2:vqshlu_n_s32 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:3:vqshlu_n_s64 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:4:vqshluq_n_s8 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:5:vqshluq_n_s16 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:6:vqshluq_n_s32 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:7:vqshluq_n_s64 Neon cumulative saturation 1 VQSHLU_N/VQSHLUQ_N (negative input) output: VQSHLU_N/VQSHLUQ_N:8:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -6196,17 +6196,17 @@ VQSHLU_N/VQSHLUQ_N:27:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, VQSHLU_N/VQSHLUQ_N:28:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHLU_N/VQSHLUQ_N:29:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHLU_N/VQSHLUQ_N (check saturation/overflow: shift by 1) overflow output: -VQSHLU_N/VQSHLUQ_N:30:vqshlu_n_s8 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:31:vqshlu_n_s16 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:32:vqshlu_n_s32 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:33:vqshlu_n_s64 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:34:vqshluq_n_s8 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:35:vqshluq_n_s16 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:36:vqshluq_n_s32 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:37:vqshluq_n_s64 Neon overflow 0 - -VQSHLU_N/VQSHLUQ_N (check saturation/overflow: shift by 1) output: +VQSHLU_N/VQSHLUQ_N (check cumulative saturation: shift by 1) cumulative saturation output: +VQSHLU_N/VQSHLUQ_N:30:vqshlu_n_s8 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:31:vqshlu_n_s16 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:32:vqshlu_n_s32 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:33:vqshlu_n_s64 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:34:vqshluq_n_s8 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:35:vqshluq_n_s16 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:36:vqshluq_n_s32 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:37:vqshluq_n_s64 Neon cumulative saturation 0 + +VQSHLU_N/VQSHLUQ_N (check cumulative saturation: shift by 1) output: VQSHLU_N/VQSHLUQ_N:38:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQSHLU_N/VQSHLUQ_N:39:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQSHLU_N/VQSHLUQ_N:40:result_int32x2 [] = { 33333333, 33333333, } @@ -6230,17 +6230,17 @@ VQSHLU_N/VQSHLUQ_N:57:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, VQSHLU_N/VQSHLUQ_N:58:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHLU_N/VQSHLUQ_N:59:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHLU_N/VQSHLUQ_N (check saturation/overflow: shift by 2) overflow output: -VQSHLU_N/VQSHLUQ_N:60:vqshlu_n_s8 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:61:vqshlu_n_s16 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:62:vqshlu_n_s32 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:63:vqshlu_n_s64 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:64:vqshluq_n_s8 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:65:vqshluq_n_s16 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:66:vqshluq_n_s32 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:67:vqshluq_n_s64 Neon overflow 1 - -VQSHLU_N/VQSHLUQ_N (check saturation/overflow: shift by 2) output: +VQSHLU_N/VQSHLUQ_N (check cumulative saturation: shift by 2) cumulative saturation output: +VQSHLU_N/VQSHLUQ_N:60:vqshlu_n_s8 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:61:vqshlu_n_s16 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:62:vqshlu_n_s32 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:63:vqshlu_n_s64 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:64:vqshluq_n_s8 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:65:vqshluq_n_s16 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:66:vqshluq_n_s32 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:67:vqshluq_n_s64 Neon cumulative saturation 1 + +VQSHLU_N/VQSHLUQ_N (check cumulative saturation: shift by 2) output: VQSHLU_N/VQSHLUQ_N:68:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQSHLU_N/VQSHLUQ_N:69:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQSHLU_N/VQSHLUQ_N:70:result_int32x2 [] = { 33333333, 33333333, } @@ -6264,15 +6264,15 @@ VQSHLU_N/VQSHLUQ_N:87:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, VQSHLU_N/VQSHLUQ_N:88:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHLU_N/VQSHLUQ_N:89:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHLU_N/VQSHLUQ_N overflow output: -VQSHLU_N/VQSHLUQ_N:90:vqshlu_n_s8 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:91:vqshlu_n_s16 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:92:vqshlu_n_s32 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:93:vqshlu_n_s64 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:94:vqshluq_n_s8 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:95:vqshluq_n_s16 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:96:vqshluq_n_s32 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:97:vqshluq_n_s64 Neon overflow 0 +VQSHLU_N/VQSHLUQ_N cumulative saturation output: +VQSHLU_N/VQSHLUQ_N:90:vqshlu_n_s8 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:91:vqshlu_n_s16 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:92:vqshlu_n_s32 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:93:vqshlu_n_s64 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:94:vqshluq_n_s8 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:95:vqshluq_n_s16 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:96:vqshluq_n_s32 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:97:vqshluq_n_s64 Neon cumulative saturation 0 VQSHLU_N/VQSHLUQ_N output: VQSHLU_N/VQSHLUQ_N:98:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -6418,13 +6418,13 @@ VCNT/VCNTQ:19:result_poly8x16 [] = { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6 VCNT/VCNTQ:20:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VCNT/VCNTQ:21:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHRN_N overflow output: -VQSHRN_N:0:vqshrn_n_s16 Neon overflow 0 -VQSHRN_N:1:vqshrn_n_s32 Neon overflow 0 -VQSHRN_N:2:vqshrn_n_s64 Neon overflow 0 -VQSHRN_N:3:vqshrn_n_u16 Neon overflow 1 -VQSHRN_N:4:vqshrn_n_u32 Neon overflow 1 -VQSHRN_N:5:vqshrn_n_u64 Neon overflow 1 +VQSHRN_N cumulative saturation output: +VQSHRN_N:0:vqshrn_n_s16 Neon cumulative saturation 0 +VQSHRN_N:1:vqshrn_n_s32 Neon cumulative saturation 0 +VQSHRN_N:2:vqshrn_n_s64 Neon cumulative saturation 0 +VQSHRN_N:3:vqshrn_n_u16 Neon cumulative saturation 1 +VQSHRN_N:4:vqshrn_n_u32 Neon cumulative saturation 1 +VQSHRN_N:5:vqshrn_n_u64 Neon cumulative saturation 1 VQSHRN_N output: VQSHRN_N:6:result_int8x8 [] = { fffffff8, fffffff8, fffffff9, fffffff9, fffffffa, fffffffa, fffffffb, fffffffb, } @@ -6450,13 +6450,13 @@ VQSHRN_N:25:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQSHRN_N:26:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHRN_N:27:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHRN_N (check saturation: shift by 3) overflow output: -VQSHRN_N:28:vqshrn_n_s16 Neon overflow 1 -VQSHRN_N:29:vqshrn_n_s32 Neon overflow 1 -VQSHRN_N:30:vqshrn_n_s64 Neon overflow 1 -VQSHRN_N:31:vqshrn_n_u16 Neon overflow 1 -VQSHRN_N:32:vqshrn_n_u32 Neon overflow 1 -VQSHRN_N:33:vqshrn_n_u64 Neon overflow 1 +VQSHRN_N (check saturation: shift by 3) cumulative saturation output: +VQSHRN_N:28:vqshrn_n_s16 Neon cumulative saturation 1 +VQSHRN_N:29:vqshrn_n_s32 Neon cumulative saturation 1 +VQSHRN_N:30:vqshrn_n_s64 Neon cumulative saturation 1 +VQSHRN_N:31:vqshrn_n_u16 Neon cumulative saturation 1 +VQSHRN_N:32:vqshrn_n_u32 Neon cumulative saturation 1 +VQSHRN_N:33:vqshrn_n_u64 Neon cumulative saturation 1 VQSHRN_N (check saturation: shift by 3) output: VQSHRN_N:34:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -6482,13 +6482,13 @@ VQSHRN_N:53:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 3 VQSHRN_N:54:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHRN_N:55:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHRN_N (check saturation: shift by max) overflow output: -VQSHRN_N:56:vqshrn_n_s16 Neon overflow 0 -VQSHRN_N:57:vqshrn_n_s32 Neon overflow 0 -VQSHRN_N:58:vqshrn_n_s64 Neon overflow 0 -VQSHRN_N:59:vqshrn_n_u16 Neon overflow 0 -VQSHRN_N:60:vqshrn_n_u32 Neon overflow 0 -VQSHRN_N:61:vqshrn_n_u64 Neon overflow 0 +VQSHRN_N (check saturation: shift by max) cumulative saturation output: +VQSHRN_N:56:vqshrn_n_s16 Neon cumulative saturation 0 +VQSHRN_N:57:vqshrn_n_s32 Neon cumulative saturation 0 +VQSHRN_N:58:vqshrn_n_s64 Neon cumulative saturation 0 +VQSHRN_N:59:vqshrn_n_u16 Neon cumulative saturation 0 +VQSHRN_N:60:vqshrn_n_u32 Neon cumulative saturation 0 +VQSHRN_N:61:vqshrn_n_u64 Neon cumulative saturation 0 VQSHRN_N (check saturation: shift by max) output: VQSHRN_N:62:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -6562,10 +6562,10 @@ VPMIN:19:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VPMIN:20:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VPMIN:21:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHRUN_N (negative input) overflow output: -VQSHRUN_N:0:vqshrun_n_s16 Neon overflow 1 -VQSHRUN_N:1:vqshrun_n_s32 Neon overflow 1 -VQSHRUN_N:2:vqshrun_n_s64 Neon overflow 1 +VQSHRUN_N (negative input) cumulative saturation output: +VQSHRUN_N:0:vqshrun_n_s16 Neon cumulative saturation 1 +VQSHRUN_N:1:vqshrun_n_s32 Neon cumulative saturation 1 +VQSHRUN_N:2:vqshrun_n_s64 Neon cumulative saturation 1 VQSHRUN_N (negative input) output: VQSHRUN_N:3:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -6591,12 +6591,12 @@ VQSHRUN_N:22:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQSHRUN_N:23:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHRUN_N:24:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHRUN_N (check saturation/overflow) overflow output: -VQSHRUN_N:25:vqshrun_n_s16 Neon overflow 1 -VQSHRUN_N:26:vqshrun_n_s32 Neon overflow 1 -VQSHRUN_N:27:vqshrun_n_s64 Neon overflow 1 +VQSHRUN_N (check cumulative saturation) cumulative saturation output: +VQSHRUN_N:25:vqshrun_n_s16 Neon cumulative saturation 1 +VQSHRUN_N:26:vqshrun_n_s32 Neon cumulative saturation 1 +VQSHRUN_N:27:vqshrun_n_s64 Neon cumulative saturation 1 -VQSHRUN_N (check saturation/overflow) output: +VQSHRUN_N (check cumulative saturation) output: VQSHRUN_N:28:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQSHRUN_N:29:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQSHRUN_N:30:result_int32x2 [] = { 33333333, 33333333, } @@ -6620,10 +6620,10 @@ VQSHRUN_N:47:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQSHRUN_N:48:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHRUN_N:49:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQSHRUN_N overflow output: -VQSHRUN_N:50:vqshrun_n_s16 Neon overflow 0 -VQSHRUN_N:51:vqshrun_n_s32 Neon overflow 1 -VQSHRUN_N:52:vqshrun_n_s64 Neon overflow 0 +VQSHRUN_N cumulative saturation output: +VQSHRUN_N:50:vqshrun_n_s16 Neon cumulative saturation 0 +VQSHRUN_N:51:vqshrun_n_s32 Neon cumulative saturation 1 +VQSHRUN_N:52:vqshrun_n_s64 Neon cumulative saturation 0 VQSHRUN_N output: VQSHRUN_N:53:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -6649,10 +6649,10 @@ VQSHRUN_N:72:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQSHRUN_N:73:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQSHRUN_N:74:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHRUN_N (negative input) overflow output: -VQRSHRUN_N:0:vqrshrun_n_s16 Neon overflow 0 -VQRSHRUN_N:1:vqrshrun_n_s32 Neon overflow 0 -VQRSHRUN_N:2:vqrshrun_n_s64 Neon overflow 1 +VQRSHRUN_N (negative input) cumulative saturation output: +VQRSHRUN_N:0:vqrshrun_n_s16 Neon cumulative saturation 0 +VQRSHRUN_N:1:vqrshrun_n_s32 Neon cumulative saturation 0 +VQRSHRUN_N:2:vqrshrun_n_s64 Neon cumulative saturation 1 VQRSHRUN_N (negative input) output: VQRSHRUN_N:3:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -6678,12 +6678,12 @@ VQRSHRUN_N:22:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQRSHRUN_N:23:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRSHRUN_N:24:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHRUN_N (check saturation/overflow: shift by 1) overflow output: -VQRSHRUN_N:25:vqrshrun_n_s16 Neon overflow 1 -VQRSHRUN_N:26:vqrshrun_n_s32 Neon overflow 1 -VQRSHRUN_N:27:vqrshrun_n_s64 Neon overflow 1 +VQRSHRUN_N (check cumulative saturation: shift by 1) cumulative saturation output: +VQRSHRUN_N:25:vqrshrun_n_s16 Neon cumulative saturation 1 +VQRSHRUN_N:26:vqrshrun_n_s32 Neon cumulative saturation 1 +VQRSHRUN_N:27:vqrshrun_n_s64 Neon cumulative saturation 1 -VQRSHRUN_N (check saturation/overflow: shift by 1) output: +VQRSHRUN_N (check cumulative saturation: shift by 1) output: VQRSHRUN_N:28:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRSHRUN_N:29:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQRSHRUN_N:30:result_int32x2 [] = { 33333333, 33333333, } @@ -6707,12 +6707,12 @@ VQRSHRUN_N:47:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQRSHRUN_N:48:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRSHRUN_N:49:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHRUN_N (check saturation/overflow: shift by max, positive input) overflow output: -VQRSHRUN_N:50:vqrshrun_n_s16 Neon overflow 0 -VQRSHRUN_N:51:vqrshrun_n_s32 Neon overflow 0 -VQRSHRUN_N:52:vqrshrun_n_s64 Neon overflow 0 +VQRSHRUN_N (check cumulative saturation: shift by max, positive input) cumulative saturation output: +VQRSHRUN_N:50:vqrshrun_n_s16 Neon cumulative saturation 0 +VQRSHRUN_N:51:vqrshrun_n_s32 Neon cumulative saturation 0 +VQRSHRUN_N:52:vqrshrun_n_s64 Neon cumulative saturation 0 -VQRSHRUN_N (check saturation/overflow: shift by max, positive input) output: +VQRSHRUN_N (check cumulative saturation: shift by max, positive input) output: VQRSHRUN_N:53:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRSHRUN_N:54:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQRSHRUN_N:55:result_int32x2 [] = { 33333333, 33333333, } @@ -6736,12 +6736,12 @@ VQRSHRUN_N:72:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQRSHRUN_N:73:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRSHRUN_N:74:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHRUN_N (check saturation/overflow: shift by max, negative input) overflow output: -VQRSHRUN_N:75:vqrshrun_n_s16 Neon overflow 1 -VQRSHRUN_N:76:vqrshrun_n_s32 Neon overflow 1 -VQRSHRUN_N:77:vqrshrun_n_s64 Neon overflow 1 +VQRSHRUN_N (check cumulative saturation: shift by max, negative input) cumulative saturation output: +VQRSHRUN_N:75:vqrshrun_n_s16 Neon cumulative saturation 1 +VQRSHRUN_N:76:vqrshrun_n_s32 Neon cumulative saturation 1 +VQRSHRUN_N:77:vqrshrun_n_s64 Neon cumulative saturation 1 -VQRSHRUN_N (check saturation/overflow: shift by max, negative input) output: +VQRSHRUN_N (check cumulative saturation: shift by max, negative input) output: VQRSHRUN_N:78:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRSHRUN_N:79:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQRSHRUN_N:80:result_int32x2 [] = { 33333333, 33333333, } @@ -6765,10 +6765,10 @@ VQRSHRUN_N:97:result_poly8x16 [] = { 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, VQRSHRUN_N:98:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, } VQRSHRUN_N:99:result_float32x4 [] = { 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, 33333333 0x1.6666660p-25 4.172325e-08, } -VQRSHRUN_N overflow output: -VQRSHRUN_N:100:vqrshrun_n_s16 Neon overflow 0 -VQRSHRUN_N:101:vqrshrun_n_s32 Neon overflow 1 -VQRSHRUN_N:102:vqrshrun_n_s64 Neon overflow 0 +VQRSHRUN_N cumulative saturation output: +VQRSHRUN_N:100:vqrshrun_n_s16 Neon cumulative saturation 0 +VQRSHRUN_N:101:vqrshrun_n_s32 Neon cumulative saturation 1 +VQRSHRUN_N:102:vqrshrun_n_s64 Neon cumulative saturation 0 VQRSHRUN_N output: VQRSHRUN_N:103:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } |