diff options
author | Christophe Lyon <christophe.lyon@linaro.org> | 2014-06-03 22:47:52 +0200 |
---|---|---|
committer | Christophe Lyon <christophe.lyon@linaro.org> | 2014-06-03 22:47:52 +0200 |
commit | 4a6e5cc95176769664bc9e56d4e7c92b7a2fc816 (patch) | |
tree | 614baf4a3939bd6a255da33aebf5cd12a43e8021 /ref-rvct-neon.txt | |
parent | fad316adcad1edd1cbb1e01769768abcffcb1070 (diff) | |
download | platform_external_arm-neon-tests-4a6e5cc95176769664bc9e56d4e7c92b7a2fc816.tar.gz platform_external_arm-neon-tests-4a6e5cc95176769664bc9e56d4e7c92b7a2fc816.tar.bz2 platform_external_arm-neon-tests-4a6e5cc95176769664bc9e56d4e7c92b7a2fc816.zip |
Renamed Neon overflow to Neon cumulative saturation, since it's the real name.
Diffstat (limited to 'ref-rvct-neon.txt')
-rw-r--r-- | ref-rvct-neon.txt | 1528 |
1 files changed, 764 insertions, 764 deletions
diff --git a/ref-rvct-neon.txt b/ref-rvct-neon.txt index 2dd22e8..5915ec2 100644 --- a/ref-rvct-neon.txt +++ b/ref-rvct-neon.txt @@ -367,9 +367,9 @@ VGET_LOW:21:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 333 VGET_LOW:22: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, } VGET_LOW:23:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -397,9 +397,9 @@ VQDMLAL_LANE:23:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQDMLAL_LANE: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, } VQDMLAL_LANE:25:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQDMLAL_LANE (mul with input=0) overflow output: -VQDMLAL_LANE:26:vqdmlal_lane_s16 Neon overflow 0 -VQDMLAL_LANE:27:vqdmlal_lane_s32 Neon overflow 0 +VQDMLAL_LANE (mul with input=0) cumulative saturation output: +VQDMLAL_LANE:26:vqdmlal_lane_s16 Neon cumulative saturation 0 +VQDMLAL_LANE:27:vqdmlal_lane_s32 Neon cumulative saturation 0 VQDMLAL_LANE (mul with input=0) output: VQDMLAL_LANE:28:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -427,11 +427,11 @@ VQDMLAL_LANE:49:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQDMLAL_LANE:50: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:51:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQDMLAL_LANE (check mul overflow) overflow output: -VQDMLAL_LANE:52:vqdmlal_lane_s16 Neon overflow 1 -VQDMLAL_LANE:53:vqdmlal_lane_s32 Neon overflow 1 +VQDMLAL_LANE (check mul cumulative saturation) cumulative saturation output: +VQDMLAL_LANE:52:vqdmlal_lane_s16 Neon cumulative saturation 1 +VQDMLAL_LANE:53:vqdmlal_lane_s32 Neon cumulative saturation 1 -VQDMLAL_LANE (check mul overflow) output: +VQDMLAL_LANE (check mul cumulative saturation) output: VQDMLAL_LANE:54:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMLAL_LANE:55:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQDMLAL_LANE:56:result_int32x2 [] = { 33333333, 33333333, } @@ -457,9 +457,9 @@ VQDMLAL_LANE:75:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQDMLAL_LANE:76: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:77:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -487,9 +487,9 @@ VQDMLSL_LANE:23:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQDMLSL_LANE: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, } VQDMLSL_LANE:25:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQDMLSL_LANE (mul with input=0) overflow output: -VQDMLSL_LANE:26:vqdmlsl_lane_s16 Neon overflow 0 -VQDMLSL_LANE:27:vqdmlsl_lane_s32 Neon overflow 0 +VQDMLSL_LANE (mul with input=0) cumulative saturation output: +VQDMLSL_LANE:26:vqdmlsl_lane_s16 Neon cumulative saturation 0 +VQDMLSL_LANE:27:vqdmlsl_lane_s32 Neon cumulative saturation 0 VQDMLSL_LANE (mul with input=0) output: VQDMLSL_LANE:28:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -517,11 +517,11 @@ VQDMLSL_LANE:49:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQDMLSL_LANE:50: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:51:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQDMLSL_LANE (check mul overflow) overflow output: -VQDMLSL_LANE:52:vqdmlsl_lane_s16 Neon overflow 1 -VQDMLSL_LANE:53:vqdmlsl_lane_s32 Neon overflow 1 +VQDMLSL_LANE (check mul cumulative saturation) cumulative saturation output: +VQDMLSL_LANE:52:vqdmlsl_lane_s16 Neon cumulative saturation 1 +VQDMLSL_LANE:53:vqdmlsl_lane_s32 Neon cumulative saturation 1 -VQDMLSL_LANE (check mul overflow) output: +VQDMLSL_LANE (check mul cumulative saturation) output: VQDMLSL_LANE:54:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMLSL_LANE:55:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQDMLSL_LANE:56:result_int32x2 [] = { 33333333, 33333333, } @@ -547,9 +547,9 @@ VQDMLSL_LANE:75:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQDMLSL_LANE:76: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:77:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -577,11 +577,11 @@ VQDMLAL_N:23:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 33 VQDMLAL_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, } VQDMLAL_N:25:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQDMLAL_N (check mul overflow) overflow output: -VQDMLAL_N:26:vqdmlal_n_s16 Neon overflow 1 -VQDMLAL_N:27:vqdmlal_n_s32 Neon overflow 1 +VQDMLAL_N (check mul cumulative saturation) cumulative saturation output: +VQDMLAL_N:26:vqdmlal_n_s16 Neon cumulative saturation 1 +VQDMLAL_N:27:vqdmlal_n_s32 Neon cumulative saturation 1 -VQDMLAL_N (check mul overflow) output: +VQDMLAL_N (check mul cumulative saturation) output: VQDMLAL_N:28:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMLAL_N:29:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQDMLAL_N:30:result_int32x2 [] = { 33333333, 33333333, } @@ -607,9 +607,9 @@ VQDMLAL_N:49:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 33 VQDMLAL_N:50: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:51:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -637,11 +637,11 @@ VQDMLSL_N:23:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 33 VQDMLSL_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, } VQDMLSL_N:25:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQDMLSL_N (check mul overflow) overflow output: -VQDMLSL_N:26:vqdmlsl_n_s16 Neon overflow 1 -VQDMLSL_N:27:vqdmlsl_n_s32 Neon overflow 1 +VQDMLSL_N (check mul cumulative saturation) cumulative saturation output: +VQDMLSL_N:26:vqdmlsl_n_s16 Neon cumulative saturation 1 +VQDMLSL_N:27:vqdmlsl_n_s32 Neon cumulative saturation 1 -VQDMLSL_N (check mul overflow) output: +VQDMLSL_N (check mul cumulative saturation) output: VQDMLSL_N:28:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMLSL_N:29:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQDMLSL_N:30:result_int32x2 [] = { 33333333, 33333333, } @@ -823,13 +823,13 @@ VRSHRN_N:69:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 333 VRSHRN_N:70: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, } VRSHRN_N:71:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -857,13 +857,13 @@ VQRSHRN_N:27:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 33 VQRSHRN_N:28: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:29:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRSHRN_N (check saturation: shift by 3) overflow output: -VQRSHRN_N:30:vqrshrn_n_s16 Neon overflow 1 -VQRSHRN_N:31:vqrshrn_n_s32 Neon overflow 1 -VQRSHRN_N:32:vqrshrn_n_s64 Neon overflow 1 -VQRSHRN_N:33:vqrshrn_n_u16 Neon overflow 1 -VQRSHRN_N:34:vqrshrn_n_u32 Neon overflow 1 -VQRSHRN_N:35:vqrshrn_n_u64 Neon overflow 1 +VQRSHRN_N (check saturation: shift by 3) cumulative saturation output: +VQRSHRN_N:30:vqrshrn_n_s16 Neon cumulative saturation 1 +VQRSHRN_N:31:vqrshrn_n_s32 Neon cumulative saturation 1 +VQRSHRN_N:32:vqrshrn_n_s64 Neon cumulative saturation 1 +VQRSHRN_N:33:vqrshrn_n_u16 Neon cumulative saturation 1 +VQRSHRN_N:34:vqrshrn_n_u32 Neon cumulative saturation 1 +VQRSHRN_N:35:vqrshrn_n_u64 Neon cumulative saturation 1 VQRSHRN_N (check saturation: shift by 3) output: VQRSHRN_N:36:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -891,13 +891,13 @@ VQRSHRN_N:57:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 33 VQRSHRN_N:58: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:59:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRSHRN_N (check saturation: shift by max) overflow output: -VQRSHRN_N:60:vqrshrn_n_s16 Neon overflow 1 -VQRSHRN_N:61:vqrshrn_n_s32 Neon overflow 1 -VQRSHRN_N:62:vqrshrn_n_s64 Neon overflow 1 -VQRSHRN_N:63:vqrshrn_n_u16 Neon overflow 1 -VQRSHRN_N:64:vqrshrn_n_u32 Neon overflow 1 -VQRSHRN_N:65:vqrshrn_n_u64 Neon overflow 1 +VQRSHRN_N (check saturation: shift by max) cumulative saturation output: +VQRSHRN_N:60:vqrshrn_n_s16 Neon cumulative saturation 1 +VQRSHRN_N:61:vqrshrn_n_s32 Neon cumulative saturation 1 +VQRSHRN_N:62:vqrshrn_n_s64 Neon cumulative saturation 1 +VQRSHRN_N:63:vqrshrn_n_u16 Neon cumulative saturation 1 +VQRSHRN_N:64:vqrshrn_n_u32 Neon cumulative saturation 1 +VQRSHRN_N:65:vqrshrn_n_u64 Neon cumulative saturation 1 VQRSHRN_N (check saturation: shift by max) output: VQRSHRN_N:66:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -976,23 +976,23 @@ 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, } @@ -1020,11 +1020,11 @@ VQSUB/VQSUBQ:37:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQSUB/VQSUBQ:38: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:39:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSUB/VQSUBQ 64 bits saturation overflow output: -VQSUB/VQSUBQ:40:vqsub_s64 Neon overflow 0 -VQSUB/VQSUBQ:41:vqsub_u64 Neon overflow 0 -VQSUB/VQSUBQ:42:vqsubq_s64 Neon overflow 0 -VQSUB/VQSUBQ:43:vqsubq_u64 Neon overflow 0 +VQSUB/VQSUBQ 64 bits saturation cumulative saturation output: +VQSUB/VQSUBQ:40:vqsub_s64 Neon cumulative saturation 0 +VQSUB/VQSUBQ:41:vqsub_u64 Neon cumulative saturation 0 +VQSUB/VQSUBQ:42:vqsubq_s64 Neon cumulative saturation 0 +VQSUB/VQSUBQ:43:vqsubq_u64 Neon cumulative saturation 0 64 bits saturation: VQSUB/VQSUBQ:44:result_int64x1 [] = { fffffffffffffff0, } @@ -1032,33 +1032,33 @@ VQSUB/VQSUBQ:45:result_uint64x1 [] = { fffffffffffffff0, } VQSUB/VQSUBQ:46:result_int64x2 [] = { fffffffffffffff0, fffffffffffffff1, } VQSUB/VQSUBQ:47:result_uint64x2 [] = { fffffffffffffff0, fffffffffffffff1, } -VQSUB/VQSUBQ 64 bits saturation overflow output: -VQSUB/VQSUBQ:48:vqsub_s64 Neon overflow 0 -VQSUB/VQSUBQ:49:vqsub_u64 Neon overflow 0 -VQSUB/VQSUBQ:50:vqsubq_s64 Neon overflow 0 -VQSUB/VQSUBQ:51:vqsubq_u64 Neon overflow 0 +VQSUB/VQSUBQ 64 bits saturation cumulative saturation output: +VQSUB/VQSUBQ:48:vqsub_s64 Neon cumulative saturation 0 +VQSUB/VQSUBQ:49:vqsub_u64 Neon cumulative saturation 0 +VQSUB/VQSUBQ:50:vqsubq_s64 Neon cumulative saturation 0 +VQSUB/VQSUBQ:51:vqsubq_u64 Neon cumulative saturation 0 VQSUB/VQSUBQ:52:result_int64x1 [] = { ffffffffffffffac, } VQSUB/VQSUBQ:53:result_uint64x1 [] = { ffffffffffffff68, } VQSUB/VQSUBQ:54:result_int64x2 [] = { ffffffffffffffac, ffffffffffffffad, } VQSUB/VQSUBQ:55:result_uint64x2 [] = { ffffffffffffff68, ffffffffffffff69, } -VQSUB/VQSUBQ 64 bits saturation overflow output: -VQSUB/VQSUBQ:56:vqsub_s64 Neon overflow 1 -VQSUB/VQSUBQ:57:vqsub_u64 Neon overflow 1 -VQSUB/VQSUBQ:58:vqsubq_s64 Neon overflow 1 -VQSUB/VQSUBQ:59:vqsubq_u64 Neon overflow 1 +VQSUB/VQSUBQ 64 bits saturation cumulative saturation output: +VQSUB/VQSUBQ:56:vqsub_s64 Neon cumulative saturation 1 +VQSUB/VQSUBQ:57:vqsub_u64 Neon cumulative saturation 1 +VQSUB/VQSUBQ:58:vqsubq_s64 Neon cumulative saturation 1 +VQSUB/VQSUBQ:59:vqsubq_u64 Neon cumulative saturation 1 VQSUB/VQSUBQ:60:result_int64x1 [] = { 8000000000000000, } VQSUB/VQSUBQ:61:result_uint64x1 [] = { 0, } VQSUB/VQSUBQ:62:result_int64x2 [] = { 7fffffffffffffff, 7fffffffffffffff, } VQSUB/VQSUBQ:63:result_uint64x2 [] = { 0, 0, } less than 64 bits saturation: -VQSUB/VQSUBQ:64:vqsub_s8 Neon overflow 1 -VQSUB/VQSUBQ:65:vqsub_s16 Neon overflow 1 -VQSUB/VQSUBQ:66:vqsub_s32 Neon overflow 1 -VQSUB/VQSUBQ:67:vqsubq_s8 Neon overflow 1 -VQSUB/VQSUBQ:68:vqsubq_s16 Neon overflow 1 -VQSUB/VQSUBQ:69:vqsubq_s32 Neon overflow 1 +VQSUB/VQSUBQ:64:vqsub_s8 Neon cumulative saturation 1 +VQSUB/VQSUBQ:65:vqsub_s16 Neon cumulative saturation 1 +VQSUB/VQSUBQ:66:vqsub_s32 Neon cumulative saturation 1 +VQSUB/VQSUBQ:67:vqsubq_s8 Neon cumulative saturation 1 +VQSUB/VQSUBQ:68:vqsubq_s16 Neon cumulative saturation 1 +VQSUB/VQSUBQ:69:vqsubq_s32 Neon cumulative saturation 1 VQSUB/VQSUBQ:70:result_int8x8 [] = { ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, } VQSUB/VQSUBQ:71:result_int16x4 [] = { ffff8000, ffff8000, ffff8000, ffff8000, } VQSUB/VQSUBQ:72:result_int32x2 [] = { 80000000, 80000000, } @@ -1066,13 +1066,13 @@ VQSUB/VQSUBQ:73:result_int8x16 [] = { ffffff80, ffffff80, ffffff80, ffffff80, ff VQSUB/VQSUBQ:74:result_int16x8 [] = { ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, } VQSUB/VQSUBQ:75:result_int32x4 [] = { 80000000, 80000000, 80000000, 80000000, } -VQSUB/VQSUBQ less than 64 bits saturation overflow output: -VQSUB/VQSUBQ:76:vqsub_u8 Neon overflow 1 -VQSUB/VQSUBQ:77:vqsub_u16 Neon overflow 1 -VQSUB/VQSUBQ:78:vqsub_u32 Neon overflow 1 -VQSUB/VQSUBQ:79:vqsubq_u8 Neon overflow 1 -VQSUB/VQSUBQ:80:vqsubq_u16 Neon overflow 1 -VQSUB/VQSUBQ:81:vqsubq_u32 Neon overflow 1 +VQSUB/VQSUBQ less than 64 bits saturation cumulative saturation output: +VQSUB/VQSUBQ:76:vqsub_u8 Neon cumulative saturation 1 +VQSUB/VQSUBQ:77:vqsub_u16 Neon cumulative saturation 1 +VQSUB/VQSUBQ:78:vqsub_u32 Neon cumulative saturation 1 +VQSUB/VQSUBQ:79:vqsubq_u8 Neon cumulative saturation 1 +VQSUB/VQSUBQ:80:vqsubq_u16 Neon cumulative saturation 1 +VQSUB/VQSUBQ:81:vqsubq_u32 Neon cumulative saturation 1 VQSUB/VQSUBQ:82:result_uint8x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } VQSUB/VQSUBQ:83:result_uint16x4 [] = { 0, 0, 0, 0, } VQSUB/VQSUBQ:84:result_uint32x2 [] = { 0, 0, } @@ -1080,11 +1080,11 @@ VQSUB/VQSUBQ:85:result_uint8x16 [] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, VQSUB/VQSUBQ:86:result_uint16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } VQSUB/VQSUBQ:87: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, } @@ -1112,11 +1112,11 @@ VQDMULH:25:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQDMULH:26: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:27:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQDMULH overflow output: -VQDMULH:28:vqdmulh_s16 Neon overflow 1 -VQDMULH:29:vqdmulh_s32 Neon overflow 1 -VQDMULH:30:vqdmulhq_s16 Neon overflow 1 -VQDMULH:31:vqdmulhq_s32 Neon overflow 1 +VQDMULH cumulative saturation output: +VQDMULH:28:vqdmulh_s16 Neon cumulative saturation 1 +VQDMULH:29:vqdmulh_s32 Neon cumulative saturation 1 +VQDMULH:30:vqdmulhq_s16 Neon cumulative saturation 1 +VQDMULH:31:vqdmulhq_s32 Neon cumulative saturation 1 VQDMULH output: VQDMULH:32:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -1144,11 +1144,11 @@ VQDMULH:53:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQDMULH:54: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:55:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -1176,13 +1176,13 @@ VQDMULH_LANE:25:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQDMULH_LANE:26: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:27:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQDMULH_LANE (check mul overflow) overflow output: -VQDMULH_LANE:28:vqdmulh_lane_s16 Neon overflow 1 -VQDMULH_LANE:29:vqdmulh_lane_s32 Neon overflow 1 -VQDMULH_LANE:30:vqdmulhq_lane_s16 Neon overflow 1 -VQDMULH_LANE:31:vqdmulhq_lane_s32 Neon overflow 1 +VQDMULH_LANE (check mul cumulative saturation) cumulative saturation output: +VQDMULH_LANE:28:vqdmulh_lane_s16 Neon cumulative saturation 1 +VQDMULH_LANE:29:vqdmulh_lane_s32 Neon cumulative saturation 1 +VQDMULH_LANE:30:vqdmulhq_lane_s16 Neon cumulative saturation 1 +VQDMULH_LANE:31:vqdmulhq_lane_s32 Neon cumulative saturation 1 -VQDMULH_LANE (check mul overflow) output: +VQDMULH_LANE (check mul cumulative saturation) output: VQDMULH_LANE:32:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMULH_LANE:33:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQDMULH_LANE:34:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -1208,11 +1208,11 @@ VQDMULH_LANE:53:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQDMULH_LANE:54: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:55:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -1220,13 +1220,13 @@ 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, } @@ -1252,9 +1252,9 @@ VQDMULH_N:33:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 33 VQDMULH_N:34: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:35:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -1282,11 +1282,11 @@ VQDMULL:23:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQDMULL: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, } VQDMULL:25:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQDMULL (check mul overflow) overflow output: -VQDMULL:26:vqdmull_s16 Neon overflow 1 -VQDMULL:27:vqdmull_s32 Neon overflow 1 +VQDMULL (check mul cumulative saturation) cumulative saturation output: +VQDMULL:26:vqdmull_s16 Neon cumulative saturation 1 +VQDMULL:27:vqdmull_s32 Neon cumulative saturation 1 -VQDMULL (check mul overflow) output: +VQDMULL (check mul cumulative saturation) output: VQDMULL:28:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMULL:29:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQDMULL:30:result_int32x2 [] = { 33333333, 33333333, } @@ -1312,9 +1312,9 @@ VQDMULL:49:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQDMULL:50: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:51:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -1342,11 +1342,11 @@ VQDMLAL:23:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQDMLAL: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, } VQDMLAL:25:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQDMLAL (check mul overflow) overflow output: -VQDMLAL:26:vqdmlal_s16 Neon overflow 1 -VQDMLAL:27:vqdmlal_s32 Neon overflow 1 +VQDMLAL (check mul cumulative saturation) cumulative saturation output: +VQDMLAL:26:vqdmlal_s16 Neon cumulative saturation 1 +VQDMLAL:27:vqdmlal_s32 Neon cumulative saturation 1 -VQDMLAL (check mul overflow) output: +VQDMLAL (check mul cumulative saturation) output: VQDMLAL:28:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMLAL:29:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQDMLAL:30:result_int32x2 [] = { 33333333, 33333333, } @@ -1372,9 +1372,9 @@ VQDMLAL:49:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQDMLAL:50: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:51:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -1402,11 +1402,11 @@ VQDMLSL:23:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQDMLSL: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, } VQDMLSL:25:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQDMLSL (check mul overflow) overflow output: -VQDMLSL:26:vqdmlsl_s16 Neon overflow 1 -VQDMLSL:27:vqdmlsl_s32 Neon overflow 1 +VQDMLSL (check mul cumulative saturation) cumulative saturation output: +VQDMLSL:26:vqdmlsl_s16 Neon cumulative saturation 1 +VQDMLSL:27:vqdmlsl_s32 Neon cumulative saturation 1 -VQDMLSL (check mul overflow) output: +VQDMLSL (check mul cumulative saturation) output: VQDMLSL:28:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQDMLSL:29:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQDMLSL:30:result_int32x2 [] = { 33333333, 33333333, } @@ -1696,23 +1696,23 @@ VSHL_N:21:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, VSHL_N:22: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, } VSHL_N:23:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -1740,23 +1740,23 @@ VQSHL/VQSHLQ:37:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQSHL/VQSHLQ:38: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:39:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSHL/VQSHLQ (input 0 and negative shift amount) overflow output: -VQSHL/VQSHLQ:40:vqshl_s8 Neon overflow 0 -VQSHL/VQSHLQ:41:vqshl_s16 Neon overflow 0 -VQSHL/VQSHLQ:42:vqshl_s32 Neon overflow 0 -VQSHL/VQSHLQ:43:vqshl_s64 Neon overflow 0 -VQSHL/VQSHLQ:44:vqshl_u8 Neon overflow 0 -VQSHL/VQSHLQ:45:vqshl_u16 Neon overflow 0 -VQSHL/VQSHLQ:46:vqshl_u32 Neon overflow 0 -VQSHL/VQSHLQ:47:vqshl_u64 Neon overflow 0 -VQSHL/VQSHLQ:48:vqshlq_s8 Neon overflow 0 -VQSHL/VQSHLQ:49:vqshlq_s16 Neon overflow 0 -VQSHL/VQSHLQ:50:vqshlq_s32 Neon overflow 0 -VQSHL/VQSHLQ:51:vqshlq_s64 Neon overflow 0 -VQSHL/VQSHLQ:52:vqshlq_u8 Neon overflow 0 -VQSHL/VQSHLQ:53:vqshlq_u16 Neon overflow 0 -VQSHL/VQSHLQ:54:vqshlq_u32 Neon overflow 0 -VQSHL/VQSHLQ:55:vqshlq_u64 Neon overflow 0 +VQSHL/VQSHLQ (input 0 and negative shift amount) cumulative saturation output: +VQSHL/VQSHLQ:40:vqshl_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:41:vqshl_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:42:vqshl_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:43:vqshl_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:44:vqshl_u8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:45:vqshl_u16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:46:vqshl_u32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:47:vqshl_u64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:48:vqshlq_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:49:vqshlq_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:50:vqshlq_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:51:vqshlq_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:52:vqshlq_u8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:53:vqshlq_u16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:54:vqshlq_u32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:55:vqshlq_u64 Neon cumulative saturation 0 VQSHL/VQSHLQ (input 0 and negative shift amount) output: VQSHL/VQSHLQ:56:result_int8x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } @@ -1784,23 +1784,23 @@ VQSHL/VQSHLQ:77:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQSHL/VQSHLQ:78: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:79:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSHL/VQSHLQ overflow output: -VQSHL/VQSHLQ:80:vqshl_s8 Neon overflow 0 -VQSHL/VQSHLQ:81:vqshl_s16 Neon overflow 0 -VQSHL/VQSHLQ:82:vqshl_s32 Neon overflow 0 -VQSHL/VQSHLQ:83:vqshl_s64 Neon overflow 0 -VQSHL/VQSHLQ:84:vqshl_u8 Neon overflow 1 -VQSHL/VQSHLQ:85:vqshl_u16 Neon overflow 1 -VQSHL/VQSHLQ:86:vqshl_u32 Neon overflow 1 -VQSHL/VQSHLQ:87:vqshl_u64 Neon overflow 0 -VQSHL/VQSHLQ:88:vqshlq_s8 Neon overflow 1 -VQSHL/VQSHLQ:89:vqshlq_s16 Neon overflow 1 -VQSHL/VQSHLQ:90:vqshlq_s32 Neon overflow 1 -VQSHL/VQSHLQ:91:vqshlq_s64 Neon overflow 1 -VQSHL/VQSHLQ:92:vqshlq_u8 Neon overflow 1 -VQSHL/VQSHLQ:93:vqshlq_u16 Neon overflow 1 -VQSHL/VQSHLQ:94:vqshlq_u32 Neon overflow 1 -VQSHL/VQSHLQ:95:vqshlq_u64 Neon overflow 1 +VQSHL/VQSHLQ cumulative saturation output: +VQSHL/VQSHLQ:80:vqshl_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:81:vqshl_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:82:vqshl_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:83:vqshl_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:84:vqshl_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:85:vqshl_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:86:vqshl_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:87:vqshl_u64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:88:vqshlq_s8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:89:vqshlq_s16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:90:vqshlq_s32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:91:vqshlq_s64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:92:vqshlq_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:93:vqshlq_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:94:vqshlq_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:95:vqshlq_u64 Neon cumulative saturation 1 VQSHL/VQSHLQ output: VQSHL/VQSHLQ:96:result_int8x8 [] = { ffffffe0, ffffffe2, ffffffe4, ffffffe6, ffffffe8, ffffffea, ffffffec, ffffffee, } @@ -1828,23 +1828,23 @@ VQSHL/VQSHLQ:117:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQSHL/VQSHLQ:118: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:119:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSHL/VQSHLQ (negative shift amount) overflow output: -VQSHL/VQSHLQ:120:vqshl_s8 Neon overflow 0 -VQSHL/VQSHLQ:121:vqshl_s16 Neon overflow 0 -VQSHL/VQSHLQ:122:vqshl_s32 Neon overflow 0 -VQSHL/VQSHLQ:123:vqshl_s64 Neon overflow 0 -VQSHL/VQSHLQ:124:vqshl_u8 Neon overflow 0 -VQSHL/VQSHLQ:125:vqshl_u16 Neon overflow 0 -VQSHL/VQSHLQ:126:vqshl_u32 Neon overflow 0 -VQSHL/VQSHLQ:127:vqshl_u64 Neon overflow 0 -VQSHL/VQSHLQ:128:vqshlq_s8 Neon overflow 0 -VQSHL/VQSHLQ:129:vqshlq_s16 Neon overflow 0 -VQSHL/VQSHLQ:130:vqshlq_s32 Neon overflow 0 -VQSHL/VQSHLQ:131:vqshlq_s64 Neon overflow 0 -VQSHL/VQSHLQ:132:vqshlq_u8 Neon overflow 0 -VQSHL/VQSHLQ:133:vqshlq_u16 Neon overflow 0 -VQSHL/VQSHLQ:134:vqshlq_u32 Neon overflow 0 -VQSHL/VQSHLQ:135:vqshlq_u64 Neon overflow 0 +VQSHL/VQSHLQ (negative shift amount) cumulative saturation output: +VQSHL/VQSHLQ:120:vqshl_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:121:vqshl_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:122:vqshl_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:123:vqshl_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:124:vqshl_u8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:125:vqshl_u16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:126:vqshl_u32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:127:vqshl_u64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:128:vqshlq_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:129:vqshlq_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:130:vqshlq_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:131:vqshlq_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:132:vqshlq_u8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:133:vqshlq_u16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:134:vqshlq_u32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:135:vqshlq_u64 Neon cumulative saturation 0 VQSHL/VQSHLQ (negative shift amount) output: VQSHL/VQSHLQ:136:result_int8x8 [] = { fffffff8, fffffff8, fffffff9, fffffff9, fffffffa, fffffffa, fffffffb, fffffffb, } @@ -1872,23 +1872,23 @@ VQSHL/VQSHLQ:157:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQSHL/VQSHLQ:158: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:159:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSHL/VQSHLQ (large shift amount, negative input) overflow output: -VQSHL/VQSHLQ:160:vqshl_s8 Neon overflow 1 -VQSHL/VQSHLQ:161:vqshl_s16 Neon overflow 1 -VQSHL/VQSHLQ:162:vqshl_s32 Neon overflow 1 -VQSHL/VQSHLQ:163:vqshl_s64 Neon overflow 1 -VQSHL/VQSHLQ:164:vqshl_u8 Neon overflow 1 -VQSHL/VQSHLQ:165:vqshl_u16 Neon overflow 1 -VQSHL/VQSHLQ:166:vqshl_u32 Neon overflow 1 -VQSHL/VQSHLQ:167:vqshl_u64 Neon overflow 1 -VQSHL/VQSHLQ:168:vqshlq_s8 Neon overflow 1 -VQSHL/VQSHLQ:169:vqshlq_s16 Neon overflow 1 -VQSHL/VQSHLQ:170:vqshlq_s32 Neon overflow 1 -VQSHL/VQSHLQ:171:vqshlq_s64 Neon overflow 1 -VQSHL/VQSHLQ:172:vqshlq_u8 Neon overflow 1 -VQSHL/VQSHLQ:173:vqshlq_u16 Neon overflow 1 -VQSHL/VQSHLQ:174:vqshlq_u32 Neon overflow 1 -VQSHL/VQSHLQ:175:vqshlq_u64 Neon overflow 1 +VQSHL/VQSHLQ (large shift amount, negative input) cumulative saturation output: +VQSHL/VQSHLQ:160:vqshl_s8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:161:vqshl_s16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:162:vqshl_s32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:163:vqshl_s64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:164:vqshl_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:165:vqshl_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:166:vqshl_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:167:vqshl_u64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:168:vqshlq_s8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:169:vqshlq_s16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:170:vqshlq_s32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:171:vqshlq_s64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:172:vqshlq_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:173:vqshlq_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:174:vqshlq_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:175:vqshlq_u64 Neon cumulative saturation 1 VQSHL/VQSHLQ (large shift amount, negative input) output: VQSHL/VQSHLQ:176:result_int8x8 [] = { ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, } @@ -1916,25 +1916,25 @@ VQSHL/VQSHLQ:197:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQSHL/VQSHLQ:198: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:199:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSHL/VQSHLQ (check saturation/overflow) overflow output: -VQSHL/VQSHLQ:200:vqshl_s8 Neon overflow 0 -VQSHL/VQSHLQ:201:vqshl_s16 Neon overflow 0 -VQSHL/VQSHLQ:202:vqshl_s32 Neon overflow 0 -VQSHL/VQSHLQ:203:vqshl_s64 Neon overflow 0 -VQSHL/VQSHLQ:204:vqshl_u8 Neon overflow 0 -VQSHL/VQSHLQ:205:vqshl_u16 Neon overflow 0 -VQSHL/VQSHLQ:206:vqshl_u32 Neon overflow 0 -VQSHL/VQSHLQ:207:vqshl_u64 Neon overflow 0 -VQSHL/VQSHLQ:208:vqshlq_s8 Neon overflow 0 -VQSHL/VQSHLQ:209:vqshlq_s16 Neon overflow 0 -VQSHL/VQSHLQ:210:vqshlq_s32 Neon overflow 0 -VQSHL/VQSHLQ:211:vqshlq_s64 Neon overflow 0 -VQSHL/VQSHLQ:212:vqshlq_u8 Neon overflow 0 -VQSHL/VQSHLQ:213:vqshlq_u16 Neon overflow 0 -VQSHL/VQSHLQ:214:vqshlq_u32 Neon overflow 0 -VQSHL/VQSHLQ:215:vqshlq_u64 Neon overflow 0 - -VQSHL/VQSHLQ (check saturation/overflow) output: +VQSHL/VQSHLQ (check cumulative saturation) cumulative saturation output: +VQSHL/VQSHLQ:200:vqshl_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:201:vqshl_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:202:vqshl_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:203:vqshl_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:204:vqshl_u8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:205:vqshl_u16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:206:vqshl_u32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:207:vqshl_u64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:208:vqshlq_s8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:209:vqshlq_s16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:210:vqshlq_s32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:211:vqshlq_s64 Neon cumulative saturation 0 +VQSHL/VQSHLQ:212:vqshlq_u8 Neon cumulative saturation 0 +VQSHL/VQSHLQ:213:vqshlq_u16 Neon cumulative saturation 0 +VQSHL/VQSHLQ:214:vqshlq_u32 Neon cumulative saturation 0 +VQSHL/VQSHLQ:215:vqshlq_u64 Neon cumulative saturation 0 + +VQSHL/VQSHLQ (check cumulative saturation) output: VQSHL/VQSHLQ:216:result_int8x8 [] = { 3f, 3f, 3f, 3f, 3f, 3f, 3f, 3f, } VQSHL/VQSHLQ:217:result_int16x4 [] = { 3fff, 3fff, 3fff, 3fff, } VQSHL/VQSHLQ:218:result_int32x2 [] = { 3fffffff, 3fffffff, } @@ -1960,23 +1960,23 @@ VQSHL/VQSHLQ:237:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQSHL/VQSHLQ:238: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:239:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSHL/VQSHLQ (large shift amount, positive input) overflow output: -VQSHL/VQSHLQ:240:vqshl_s8 Neon overflow 1 -VQSHL/VQSHLQ:241:vqshl_s16 Neon overflow 1 -VQSHL/VQSHLQ:242:vqshl_s32 Neon overflow 1 -VQSHL/VQSHLQ:243:vqshl_s64 Neon overflow 1 -VQSHL/VQSHLQ:244:vqshl_u8 Neon overflow 1 -VQSHL/VQSHLQ:245:vqshl_u16 Neon overflow 1 -VQSHL/VQSHLQ:246:vqshl_u32 Neon overflow 1 -VQSHL/VQSHLQ:247:vqshl_u64 Neon overflow 1 -VQSHL/VQSHLQ:248:vqshlq_s8 Neon overflow 1 -VQSHL/VQSHLQ:249:vqshlq_s16 Neon overflow 1 -VQSHL/VQSHLQ:250:vqshlq_s32 Neon overflow 1 -VQSHL/VQSHLQ:251:vqshlq_s64 Neon overflow 1 -VQSHL/VQSHLQ:252:vqshlq_u8 Neon overflow 1 -VQSHL/VQSHLQ:253:vqshlq_u16 Neon overflow 1 -VQSHL/VQSHLQ:254:vqshlq_u32 Neon overflow 1 -VQSHL/VQSHLQ:255:vqshlq_u64 Neon overflow 1 +VQSHL/VQSHLQ (large shift amount, positive input) cumulative saturation output: +VQSHL/VQSHLQ:240:vqshl_s8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:241:vqshl_s16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:242:vqshl_s32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:243:vqshl_s64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:244:vqshl_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:245:vqshl_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:246:vqshl_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:247:vqshl_u64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:248:vqshlq_s8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:249:vqshlq_s16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:250:vqshlq_s32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:251:vqshlq_s64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:252:vqshlq_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:253:vqshlq_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:254:vqshlq_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:255:vqshlq_u64 Neon cumulative saturation 1 VQSHL/VQSHLQ (large shift amount, positive input) output: VQSHL/VQSHLQ:256:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -2004,23 +2004,23 @@ VQSHL/VQSHLQ:277:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQSHL/VQSHLQ:278: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:279:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSHL/VQSHLQ (check saturation on 64 bits) overflow output: -VQSHL/VQSHLQ:280:vqshl_s8 Neon overflow 1 -VQSHL/VQSHLQ:281:vqshl_s16 Neon overflow 1 -VQSHL/VQSHLQ:282:vqshl_s32 Neon overflow 1 -VQSHL/VQSHLQ:283:vqshl_s64 Neon overflow 1 -VQSHL/VQSHLQ:284:vqshl_u8 Neon overflow 1 -VQSHL/VQSHLQ:285:vqshl_u16 Neon overflow 1 -VQSHL/VQSHLQ:286:vqshl_u32 Neon overflow 1 -VQSHL/VQSHLQ:287:vqshl_u64 Neon overflow 1 -VQSHL/VQSHLQ:288:vqshlq_s8 Neon overflow 1 -VQSHL/VQSHLQ:289:vqshlq_s16 Neon overflow 1 -VQSHL/VQSHLQ:290:vqshlq_s32 Neon overflow 1 -VQSHL/VQSHLQ:291:vqshlq_s64 Neon overflow 1 -VQSHL/VQSHLQ:292:vqshlq_u8 Neon overflow 1 -VQSHL/VQSHLQ:293:vqshlq_u16 Neon overflow 1 -VQSHL/VQSHLQ:294:vqshlq_u32 Neon overflow 1 -VQSHL/VQSHLQ:295:vqshlq_u64 Neon overflow 1 +VQSHL/VQSHLQ (check saturation on 64 bits) cumulative saturation output: +VQSHL/VQSHLQ:280:vqshl_s8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:281:vqshl_s16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:282:vqshl_s32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:283:vqshl_s64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:284:vqshl_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:285:vqshl_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:286:vqshl_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:287:vqshl_u64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:288:vqshlq_s8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:289:vqshlq_s16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:290:vqshlq_s32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:291:vqshlq_s64 Neon cumulative saturation 1 +VQSHL/VQSHLQ:292:vqshlq_u8 Neon cumulative saturation 1 +VQSHL/VQSHLQ:293:vqshlq_u16 Neon cumulative saturation 1 +VQSHL/VQSHLQ:294:vqshlq_u32 Neon cumulative saturation 1 +VQSHL/VQSHLQ:295:vqshlq_u64 Neon cumulative saturation 1 VQSHL/VQSHLQ (check saturation on 64 bits) output: VQSHL/VQSHLQ:296:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -2048,23 +2048,23 @@ VQSHL/VQSHLQ:317:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQSHL/VQSHLQ:318: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:319:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -2092,23 +2092,23 @@ VQSHL_N/VQSHLQ_N:37:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3 VQSHL_N/VQSHLQ_N:38: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:39:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSHL_N/VQSHLQ_N (check saturation with large positive input) overflow output: -VQSHL_N/VQSHLQ_N:40:vqshl_n_s8 Neon overflow 1 -VQSHL_N/VQSHLQ_N:41:vqshl_n_s16 Neon overflow 1 -VQSHL_N/VQSHLQ_N:42:vqshl_n_s32 Neon overflow 1 -VQSHL_N/VQSHLQ_N:43:vqshl_n_s64 Neon overflow 1 -VQSHL_N/VQSHLQ_N:44:vqshl_n_u8 Neon overflow 1 -VQSHL_N/VQSHLQ_N:45:vqshl_n_u16 Neon overflow 1 -VQSHL_N/VQSHLQ_N:46:vqshl_n_u32 Neon overflow 1 -VQSHL_N/VQSHLQ_N:47:vqshl_n_u64 Neon overflow 1 -VQSHL_N/VQSHLQ_N:48:vqshlq_n_s8 Neon overflow 1 -VQSHL_N/VQSHLQ_N:49:vqshlq_n_s16 Neon overflow 1 -VQSHL_N/VQSHLQ_N:50:vqshlq_n_s32 Neon overflow 1 -VQSHL_N/VQSHLQ_N:51:vqshlq_n_s64 Neon overflow 1 -VQSHL_N/VQSHLQ_N:52:vqshlq_n_u8 Neon overflow 1 -VQSHL_N/VQSHLQ_N:53:vqshlq_n_u16 Neon overflow 1 -VQSHL_N/VQSHLQ_N:54:vqshlq_n_u32 Neon overflow 1 -VQSHL_N/VQSHLQ_N:55:vqshlq_n_u64 Neon overflow 1 +VQSHL_N/VQSHLQ_N (check saturation with large positive input) cumulative saturation output: +VQSHL_N/VQSHLQ_N:40:vqshl_n_s8 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:41:vqshl_n_s16 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:42:vqshl_n_s32 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:43:vqshl_n_s64 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:44:vqshl_n_u8 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:45:vqshl_n_u16 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:46:vqshl_n_u32 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:47:vqshl_n_u64 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:48:vqshlq_n_s8 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:49:vqshlq_n_s16 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:50:vqshlq_n_s32 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:51:vqshlq_n_s64 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:52:vqshlq_n_u8 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:53:vqshlq_n_u16 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:54:vqshlq_n_u32 Neon cumulative saturation 1 +VQSHL_N/VQSHLQ_N:55:vqshlq_n_u64 Neon cumulative saturation 1 VQSHL_N/VQSHLQ_N (check saturation with large positive input) output: VQSHL_N/VQSHLQ_N:56:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -2630,35 +2630,35 @@ VDUP_LANE/VDUP_LANEQ:21:result_poly16x8 [] = { fff1, fff1, fff1, fff1, fff1, fff VDUP_LANE/VDUP_LANEQ:22:result_float32x4 [] = { c1700000 -0x1.e000000p+3 -15, c1700000 -0x1.e000000p+3 -15, c1700000 -0x1.e000000p+3 -15, c1700000 -0x1.e000000p+3 -15, } VDUP_LANE/VDUP_LANEQ:23:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -2718,23 +2718,23 @@ float32: VSUB/VSUBQ:24:result_float32x2 [] = { c00ccccd -0x1.19999a0p+1 -2.2, c00ccccd -0x1.19999a0p+1 -2.2, } VSUB/VSUBQ:25: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, } @@ -2762,11 +2762,11 @@ VQADD/VQADDQ:37:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQADD/VQADDQ:38: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:39:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQADD/VQADDQ 64 bits saturation overflow output: -VQADD/VQADDQ:40:vqadd_s64 Neon overflow 0 -VQADD/VQADDQ:41:vqadd_u64 Neon overflow 0 -VQADD/VQADDQ:42:vqaddq_s64 Neon overflow 0 -VQADD/VQADDQ:43:vqaddq_u64 Neon overflow 0 +VQADD/VQADDQ 64 bits saturation cumulative saturation output: +VQADD/VQADDQ:40:vqadd_s64 Neon cumulative saturation 0 +VQADD/VQADDQ:41:vqadd_u64 Neon cumulative saturation 0 +VQADD/VQADDQ:42:vqaddq_s64 Neon cumulative saturation 0 +VQADD/VQADDQ:43:vqaddq_u64 Neon cumulative saturation 0 64 bits saturation: VQADD/VQADDQ:44:result_int64x1 [] = { fffffffffffffff0, } @@ -2774,33 +2774,33 @@ VQADD/VQADDQ:45:result_uint64x1 [] = { fffffffffffffff0, } VQADD/VQADDQ:46:result_int64x2 [] = { fffffffffffffff0, fffffffffffffff1, } VQADD/VQADDQ:47:result_uint64x2 [] = { fffffffffffffff0, fffffffffffffff1, } -VQADD/VQADDQ 64 bits saturation overflow output: -VQADD/VQADDQ:48:vqadd_s64 Neon overflow 0 -VQADD/VQADDQ:49:vqadd_u64 Neon overflow 1 -VQADD/VQADDQ:50:vqaddq_s64 Neon overflow 0 -VQADD/VQADDQ:51:vqaddq_u64 Neon overflow 1 +VQADD/VQADDQ 64 bits saturation cumulative saturation output: +VQADD/VQADDQ:48:vqadd_s64 Neon cumulative saturation 0 +VQADD/VQADDQ:49:vqadd_u64 Neon cumulative saturation 1 +VQADD/VQADDQ:50:vqaddq_s64 Neon cumulative saturation 0 +VQADD/VQADDQ:51:vqaddq_u64 Neon cumulative saturation 1 VQADD/VQADDQ:52:result_int64x1 [] = { 34, } VQADD/VQADDQ:53:result_uint64x1 [] = { ffffffffffffffff, } VQADD/VQADDQ:54:result_int64x2 [] = { 34, 35, } VQADD/VQADDQ:55:result_uint64x2 [] = { ffffffffffffffff, ffffffffffffffff, } -VQADD/VQADDQ 64 bits saturation overflow output: -VQADD/VQADDQ:56:vqadd_s64 Neon overflow 1 -VQADD/VQADDQ:57:vqadd_u64 Neon overflow 1 -VQADD/VQADDQ:58:vqaddq_s64 Neon overflow 1 -VQADD/VQADDQ:59:vqaddq_u64 Neon overflow 1 +VQADD/VQADDQ 64 bits saturation cumulative saturation output: +VQADD/VQADDQ:56:vqadd_s64 Neon cumulative saturation 1 +VQADD/VQADDQ:57:vqadd_u64 Neon cumulative saturation 1 +VQADD/VQADDQ:58:vqaddq_s64 Neon cumulative saturation 1 +VQADD/VQADDQ:59:vqaddq_u64 Neon cumulative saturation 1 VQADD/VQADDQ:60:result_int64x1 [] = { 8000000000000000, } VQADD/VQADDQ:61:result_uint64x1 [] = { ffffffffffffffff, } VQADD/VQADDQ:62:result_int64x2 [] = { 7fffffffffffffff, 7fffffffffffffff, } VQADD/VQADDQ:63:result_uint64x2 [] = { ffffffffffffffff, ffffffffffffffff, } less than 64 bits saturation: -VQADD/VQADDQ:64:vqadd_s8 Neon overflow 1 -VQADD/VQADDQ:65:vqadd_s16 Neon overflow 1 -VQADD/VQADDQ:66:vqadd_s32 Neon overflow 1 -VQADD/VQADDQ:67:vqaddq_s8 Neon overflow 1 -VQADD/VQADDQ:68:vqaddq_s16 Neon overflow 1 -VQADD/VQADDQ:69:vqaddq_s32 Neon overflow 1 +VQADD/VQADDQ:64:vqadd_s8 Neon cumulative saturation 1 +VQADD/VQADDQ:65:vqadd_s16 Neon cumulative saturation 1 +VQADD/VQADDQ:66:vqadd_s32 Neon cumulative saturation 1 +VQADD/VQADDQ:67:vqaddq_s8 Neon cumulative saturation 1 +VQADD/VQADDQ:68:vqaddq_s16 Neon cumulative saturation 1 +VQADD/VQADDQ:69:vqaddq_s32 Neon cumulative saturation 1 VQADD/VQADDQ:70:result_int8x8 [] = { ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, } VQADD/VQADDQ:71:result_int16x4 [] = { ffff8000, ffff8000, ffff8000, ffff8000, } VQADD/VQADDQ:72:result_int32x2 [] = { 80000000, 80000000, } @@ -2808,13 +2808,13 @@ VQADD/VQADDQ:73:result_int8x16 [] = { ffffff80, ffffff80, ffffff80, ffffff80, ff VQADD/VQADDQ:74:result_int16x8 [] = { ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, ffff8000, } VQADD/VQADDQ:75:result_int32x4 [] = { 80000000, 80000000, 80000000, 80000000, } -VQADD/VQADDQ less than 64 bits saturation overflow output: -VQADD/VQADDQ:76:vqadd_u8 Neon overflow 1 -VQADD/VQADDQ:77:vqadd_u16 Neon overflow 1 -VQADD/VQADDQ:78:vqadd_u32 Neon overflow 1 -VQADD/VQADDQ:79:vqaddq_u8 Neon overflow 1 -VQADD/VQADDQ:80:vqaddq_u16 Neon overflow 1 -VQADD/VQADDQ:81:vqaddq_u32 Neon overflow 1 +VQADD/VQADDQ less than 64 bits saturation cumulative saturation output: +VQADD/VQADDQ:76:vqadd_u8 Neon cumulative saturation 1 +VQADD/VQADDQ:77:vqadd_u16 Neon cumulative saturation 1 +VQADD/VQADDQ:78:vqadd_u32 Neon cumulative saturation 1 +VQADD/VQADDQ:79:vqaddq_u8 Neon cumulative saturation 1 +VQADD/VQADDQ:80:vqaddq_u16 Neon cumulative saturation 1 +VQADD/VQADDQ:81:vqaddq_u32 Neon cumulative saturation 1 VQADD/VQADDQ:82:result_uint8x8 [] = { ff, ff, ff, ff, ff, ff, ff, ff, } VQADD/VQADDQ:83:result_uint16x4 [] = { ffff, ffff, ffff, ffff, } VQADD/VQADDQ:84:result_uint32x2 [] = { ffffffff, ffffffff, } @@ -2852,13 +2852,13 @@ float32: VABS/VABSQ:24:result_float32x2 [] = { 40133333 0x1.2666660p+1 2.3, 40133333 0x1.2666660p+1 2.3, } VABS/VABSQ:25: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, } @@ -2886,13 +2886,13 @@ VQABS/VQABSQ:27:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQABS/VQABSQ:28: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:29:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -3040,13 +3040,13 @@ float32: VNEG/VNEGQ:24:result_float32x2 [] = { c0133333 -0x1.2666660p+1 -2.3, c0133333 -0x1.2666660p+1 -2.3, } VNEG/VNEGQ:25: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, } @@ -3074,13 +3074,13 @@ VQNEG/VQNEGQ:27:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQNEG/VQNEGQ:28: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:29:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -3876,11 +3876,11 @@ VREINTERPRET/VREINTERPRETQ:243:result_poly8x16 [] = { 0, cc, 80, cb, 0, cb, 80, VREINTERPRET/VREINTERPRETQ:244:result_poly16x8 [] = { cc00, cb80, cb00, ca80, ca00, c980, c900, c880, } VREINTERPRET/VREINTERPRETQ:245:result_float32x4 [] = { cb80cc00 -0x1.0198000p+24 -1.688166e+07, ca80cb00 -0x1.0196000p+22 -4220288, c980ca00 -0x1.0194000p+20 -1055040, c880c900 -0x1.0192000p+18 -263752, } -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, } @@ -3908,13 +3908,13 @@ VQRDMULH:25:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 333 VQRDMULH:26: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:27:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRDMULH (check mul overflow) overflow output: -VQRDMULH:28:vqrdmulh_s16 Neon overflow 1 -VQRDMULH:29:vqrdmulh_s32 Neon overflow 1 -VQRDMULH:30:vqrdmulhq_s16 Neon overflow 1 -VQRDMULH:31:vqrdmulhq_s32 Neon overflow 1 +VQRDMULH (check mul cumulative saturation) cumulative saturation output: +VQRDMULH:28:vqrdmulh_s16 Neon cumulative saturation 1 +VQRDMULH:29:vqrdmulh_s32 Neon cumulative saturation 1 +VQRDMULH:30:vqrdmulhq_s16 Neon cumulative saturation 1 +VQRDMULH:31:vqrdmulhq_s32 Neon cumulative saturation 1 -VQRDMULH (check mul overflow) output: +VQRDMULH (check mul cumulative saturation) output: VQRDMULH:32:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRDMULH:33:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQRDMULH:34:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -3940,13 +3940,13 @@ VQRDMULH:53:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 333 VQRDMULH:54: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:55:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRDMULH (check rounding overflow) overflow output: -VQRDMULH:56:vqrdmulh_s16 Neon overflow 0 -VQRDMULH:57:vqrdmulh_s32 Neon overflow 0 -VQRDMULH:58:vqrdmulhq_s16 Neon overflow 0 -VQRDMULH:59:vqrdmulhq_s32 Neon overflow 0 +VQRDMULH (check rounding cumulative saturation) cumulative saturation output: +VQRDMULH:56:vqrdmulh_s16 Neon cumulative saturation 0 +VQRDMULH:57:vqrdmulh_s32 Neon cumulative saturation 0 +VQRDMULH:58:vqrdmulhq_s16 Neon cumulative saturation 0 +VQRDMULH:59:vqrdmulhq_s32 Neon cumulative saturation 0 -VQRDMULH (check rounding overflow) output: +VQRDMULH (check rounding cumulative saturation) output: VQRDMULH:60:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRDMULH:61:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQRDMULH:62:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -3972,11 +3972,11 @@ VQRDMULH:81:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 333 VQRDMULH:82: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:83:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -4004,13 +4004,13 @@ VQRDMULH_LANE:25:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQRDMULH_LANE:26: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:27:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRDMULH_LANE (check mul overflow) overflow output: -VQRDMULH_LANE:28:vqrdmulh_lane_s16 Neon overflow 1 -VQRDMULH_LANE:29:vqrdmulh_lane_s32 Neon overflow 1 -VQRDMULH_LANE:30:vqrdmulhq_lane_s16 Neon overflow 1 -VQRDMULH_LANE:31:vqrdmulhq_lane_s32 Neon overflow 1 +VQRDMULH_LANE (check mul cumulative saturation) cumulative saturation output: +VQRDMULH_LANE:28:vqrdmulh_lane_s16 Neon cumulative saturation 1 +VQRDMULH_LANE:29:vqrdmulh_lane_s32 Neon cumulative saturation 1 +VQRDMULH_LANE:30:vqrdmulhq_lane_s16 Neon cumulative saturation 1 +VQRDMULH_LANE:31:vqrdmulhq_lane_s32 Neon cumulative saturation 1 -VQRDMULH_LANE (check mul overflow) output: +VQRDMULH_LANE (check mul cumulative saturation) output: VQRDMULH_LANE:32:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRDMULH_LANE:33:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQRDMULH_LANE:34:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -4036,13 +4036,13 @@ VQRDMULH_LANE:53:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQRDMULH_LANE:54: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:55:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRDMULH_LANE (check rounding overflow) overflow output: -VQRDMULH_LANE:56:vqrdmulh_lane_s16 Neon overflow 0 -VQRDMULH_LANE:57:vqrdmulh_lane_s32 Neon overflow 0 -VQRDMULH_LANE:58:vqrdmulhq_lane_s16 Neon overflow 0 -VQRDMULH_LANE:59:vqrdmulhq_lane_s32 Neon overflow 0 +VQRDMULH_LANE (check rounding cumulative saturation) cumulative saturation output: +VQRDMULH_LANE:56:vqrdmulh_lane_s16 Neon cumulative saturation 0 +VQRDMULH_LANE:57:vqrdmulh_lane_s32 Neon cumulative saturation 0 +VQRDMULH_LANE:58:vqrdmulhq_lane_s16 Neon cumulative saturation 0 +VQRDMULH_LANE:59:vqrdmulhq_lane_s32 Neon cumulative saturation 0 -VQRDMULH_LANE (check rounding overflow) output: +VQRDMULH_LANE (check rounding cumulative saturation) output: VQRDMULH_LANE:60:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRDMULH_LANE:61:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQRDMULH_LANE:62:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -4068,11 +4068,11 @@ VQRDMULH_LANE:81:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQRDMULH_LANE:82: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:83:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -4080,13 +4080,13 @@ 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, } @@ -4112,13 +4112,13 @@ VQRDMULH_N:33:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3 VQRDMULH_N:34: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:35:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRDMULH_N (check rounding overflow) overflow output: -VQRDMULH_N:36:vqrdmulh_n_s16 Neon overflow 0 -VQRDMULH_N:37:vqrdmulh_n_s32 Neon overflow 0 -VQRDMULH_N:38:vqrdmulhq_n_s16 Neon overflow 0 -VQRDMULH_N:39:vqrdmulhq_n_s32 Neon overflow 0 +VQRDMULH_N (check rounding cumulative saturation) cumulative saturation output: +VQRDMULH_N:36:vqrdmulh_n_s16 Neon cumulative saturation 0 +VQRDMULH_N:37:vqrdmulh_n_s32 Neon cumulative saturation 0 +VQRDMULH_N:38:vqrdmulhq_n_s16 Neon cumulative saturation 0 +VQRDMULH_N:39:vqrdmulhq_n_s32 Neon cumulative saturation 0 -VQRDMULH_N (check rounding overflow) output: +VQRDMULH_N (check rounding cumulative saturation) output: VQRDMULH_N:40:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRDMULH_N:41:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQRDMULH_N:42:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -4144,23 +4144,23 @@ VQRDMULH_N:61:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3 VQRDMULH_N:62: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:63:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -4188,23 +4188,23 @@ VQRSHL/VQRSHLQ:37:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 333 VQRSHL/VQRSHLQ:38: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:39:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRSHL/VQRSHLQ (input 0 and negative shift amount) overflow output: -VQRSHL/VQRSHLQ:40:vqrshl_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:41:vqrshl_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:42:vqrshl_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:43:vqrshl_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:44:vqrshl_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:45:vqrshl_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:46:vqrshl_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:47:vqrshl_u64 Neon overflow 0 -VQRSHL/VQRSHLQ:48:vqrshlq_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:49:vqrshlq_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:50:vqrshlq_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:51:vqrshlq_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:52:vqrshlq_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:53:vqrshlq_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:54:vqrshlq_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:55:vqrshlq_u64 Neon overflow 0 +VQRSHL/VQRSHLQ (input 0 and negative shift amount) cumulative saturation output: +VQRSHL/VQRSHLQ:40:vqrshl_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:41:vqrshl_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:42:vqrshl_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:43:vqrshl_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:44:vqrshl_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:45:vqrshl_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:46:vqrshl_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:47:vqrshl_u64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:48:vqrshlq_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:49:vqrshlq_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:50:vqrshlq_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:51:vqrshlq_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:52:vqrshlq_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:53:vqrshlq_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:54:vqrshlq_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:55:vqrshlq_u64 Neon cumulative saturation 0 VQRSHL/VQRSHLQ (input 0 and negative shift amount) output: VQRSHL/VQRSHLQ:56:result_int8x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } @@ -4232,23 +4232,23 @@ VQRSHL/VQRSHLQ:77:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 333 VQRSHL/VQRSHLQ:78: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:79:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRSHL/VQRSHLQ overflow output: -VQRSHL/VQRSHLQ:80:vqrshl_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:81:vqrshl_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:82:vqrshl_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:83:vqrshl_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:84:vqrshl_u8 Neon overflow 1 -VQRSHL/VQRSHLQ:85:vqrshl_u16 Neon overflow 1 -VQRSHL/VQRSHLQ:86:vqrshl_u32 Neon overflow 1 -VQRSHL/VQRSHLQ:87:vqrshl_u64 Neon overflow 1 -VQRSHL/VQRSHLQ:88:vqrshlq_s8 Neon overflow 1 -VQRSHL/VQRSHLQ:89:vqrshlq_s16 Neon overflow 1 -VQRSHL/VQRSHLQ:90:vqrshlq_s32 Neon overflow 1 -VQRSHL/VQRSHLQ:91:vqrshlq_s64 Neon overflow 1 -VQRSHL/VQRSHLQ:92:vqrshlq_u8 Neon overflow 1 -VQRSHL/VQRSHLQ:93:vqrshlq_u16 Neon overflow 1 -VQRSHL/VQRSHLQ:94:vqrshlq_u32 Neon overflow 1 -VQRSHL/VQRSHLQ:95:vqrshlq_u64 Neon overflow 1 +VQRSHL/VQRSHLQ cumulative saturation output: +VQRSHL/VQRSHLQ:80:vqrshl_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:81:vqrshl_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:82:vqrshl_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:83:vqrshl_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:84:vqrshl_u8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:85:vqrshl_u16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:86:vqrshl_u32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:87:vqrshl_u64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:88:vqrshlq_s8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:89:vqrshlq_s16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:90:vqrshlq_s32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:91:vqrshlq_s64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:92:vqrshlq_u8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:93:vqrshlq_u16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:94:vqrshlq_u32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:95:vqrshlq_u64 Neon cumulative saturation 1 VQRSHL/VQRSHLQ output: VQRSHL/VQRSHLQ:96:result_int8x8 [] = { ffffffe0, ffffffe2, ffffffe4, ffffffe6, ffffffe8, ffffffea, ffffffec, ffffffee, } @@ -4276,23 +4276,23 @@ VQRSHL/VQRSHLQ:117:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 33 VQRSHL/VQRSHLQ:118: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:119:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRSHL/VQRSHLQ (negative shift amount) overflow output: -VQRSHL/VQRSHLQ:120:vqrshl_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:121:vqrshl_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:122:vqrshl_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:123:vqrshl_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:124:vqrshl_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:125:vqrshl_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:126:vqrshl_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:127:vqrshl_u64 Neon overflow 0 -VQRSHL/VQRSHLQ:128:vqrshlq_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:129:vqrshlq_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:130:vqrshlq_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:131:vqrshlq_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:132:vqrshlq_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:133:vqrshlq_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:134:vqrshlq_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:135:vqrshlq_u64 Neon overflow 0 +VQRSHL/VQRSHLQ (negative shift amount) cumulative saturation output: +VQRSHL/VQRSHLQ:120:vqrshl_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:121:vqrshl_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:122:vqrshl_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:123:vqrshl_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:124:vqrshl_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:125:vqrshl_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:126:vqrshl_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:127:vqrshl_u64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:128:vqrshlq_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:129:vqrshlq_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:130:vqrshlq_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:131:vqrshlq_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:132:vqrshlq_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:133:vqrshlq_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:134:vqrshlq_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:135:vqrshlq_u64 Neon cumulative saturation 0 VQRSHL/VQRSHLQ (negative shift amount) output: VQRSHL/VQRSHLQ:136:result_int8x8 [] = { fffffffc, fffffffc, fffffffd, fffffffd, fffffffd, fffffffd, fffffffe, fffffffe, } @@ -4320,25 +4320,25 @@ VQRSHL/VQRSHLQ:157:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 33 VQRSHL/VQRSHLQ:158: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:159:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRSHL/VQRSHLQ (checking overflow: shift by -1) overflow output: -VQRSHL/VQRSHLQ:160:vqrshl_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:161:vqrshl_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:162:vqrshl_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:163:vqrshl_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:164:vqrshl_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:165:vqrshl_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:166:vqrshl_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:167:vqrshl_u64 Neon overflow 0 -VQRSHL/VQRSHLQ:168:vqrshlq_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:169:vqrshlq_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:170:vqrshlq_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:171:vqrshlq_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:172:vqrshlq_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:173:vqrshlq_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:174:vqrshlq_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:175: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:160:vqrshl_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:161:vqrshl_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:162:vqrshl_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:163:vqrshl_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:164:vqrshl_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:165:vqrshl_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:166:vqrshl_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:167:vqrshl_u64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:168:vqrshlq_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:169:vqrshlq_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:170:vqrshlq_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:171:vqrshlq_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:172:vqrshlq_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:173:vqrshlq_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:174:vqrshlq_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:175:vqrshlq_u64 Neon cumulative saturation 0 + +VQRSHL/VQRSHLQ (checking cumulative saturation: shift by -1) output: VQRSHL/VQRSHLQ:176:result_int8x8 [] = { 40, 40, 40, 40, 40, 40, 40, 40, } VQRSHL/VQRSHLQ:177:result_int16x4 [] = { 4000, 4000, 4000, 4000, } VQRSHL/VQRSHLQ:178:result_int32x2 [] = { 40000000, 40000000, } @@ -4364,25 +4364,25 @@ VQRSHL/VQRSHLQ:197:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 33 VQRSHL/VQRSHLQ:198: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:199:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRSHL/VQRSHLQ (checking overflow: shift by -3) overflow output: -VQRSHL/VQRSHLQ:200:vqrshl_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:201:vqrshl_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:202:vqrshl_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:203:vqrshl_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:204:vqrshl_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:205:vqrshl_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:206:vqrshl_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:207:vqrshl_u64 Neon overflow 0 -VQRSHL/VQRSHLQ:208:vqrshlq_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:209:vqrshlq_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:210:vqrshlq_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:211:vqrshlq_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:212:vqrshlq_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:213:vqrshlq_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:214:vqrshlq_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:215: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:200:vqrshl_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:201:vqrshl_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:202:vqrshl_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:203:vqrshl_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:204:vqrshl_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:205:vqrshl_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:206:vqrshl_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:207:vqrshl_u64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:208:vqrshlq_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:209:vqrshlq_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:210:vqrshlq_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:211:vqrshlq_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:212:vqrshlq_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:213:vqrshlq_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:214:vqrshlq_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:215:vqrshlq_u64 Neon cumulative saturation 0 + +VQRSHL/VQRSHLQ (checking cumulative saturation: shift by -3) output: VQRSHL/VQRSHLQ:216:result_int8x8 [] = { 10, 10, 10, 10, 10, 10, 10, 10, } VQRSHL/VQRSHLQ:217:result_int16x4 [] = { 1000, 1000, 1000, 1000, } VQRSHL/VQRSHLQ:218:result_int32x2 [] = { 10000000, 10000000, } @@ -4408,25 +4408,25 @@ VQRSHL/VQRSHLQ:237:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 33 VQRSHL/VQRSHLQ:238: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:239:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRSHL/VQRSHLQ (checking overflow: large shift amount) overflow output: -VQRSHL/VQRSHLQ:240:vqrshl_s8 Neon overflow 1 -VQRSHL/VQRSHLQ:241:vqrshl_s16 Neon overflow 1 -VQRSHL/VQRSHLQ:242:vqrshl_s32 Neon overflow 1 -VQRSHL/VQRSHLQ:243:vqrshl_s64 Neon overflow 1 -VQRSHL/VQRSHLQ:244:vqrshl_u8 Neon overflow 1 -VQRSHL/VQRSHLQ:245:vqrshl_u16 Neon overflow 1 -VQRSHL/VQRSHLQ:246:vqrshl_u32 Neon overflow 1 -VQRSHL/VQRSHLQ:247:vqrshl_u64 Neon overflow 1 -VQRSHL/VQRSHLQ:248:vqrshlq_s8 Neon overflow 1 -VQRSHL/VQRSHLQ:249:vqrshlq_s16 Neon overflow 1 -VQRSHL/VQRSHLQ:250:vqrshlq_s32 Neon overflow 1 -VQRSHL/VQRSHLQ:251:vqrshlq_s64 Neon overflow 1 -VQRSHL/VQRSHLQ:252:vqrshlq_u8 Neon overflow 1 -VQRSHL/VQRSHLQ:253:vqrshlq_u16 Neon overflow 1 -VQRSHL/VQRSHLQ:254:vqrshlq_u32 Neon overflow 1 -VQRSHL/VQRSHLQ:255: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:240:vqrshl_s8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:241:vqrshl_s16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:242:vqrshl_s32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:243:vqrshl_s64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:244:vqrshl_u8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:245:vqrshl_u16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:246:vqrshl_u32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:247:vqrshl_u64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:248:vqrshlq_s8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:249:vqrshlq_s16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:250:vqrshlq_s32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:251:vqrshlq_s64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:252:vqrshlq_u8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:253:vqrshlq_u16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:254:vqrshlq_u32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:255:vqrshlq_u64 Neon cumulative saturation 1 + +VQRSHL/VQRSHLQ (checking cumulative saturation: large shift amount) output: VQRSHL/VQRSHLQ:256:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } VQRSHL/VQRSHLQ:257:result_int16x4 [] = { 7fff, 7fff, 7fff, 7fff, } VQRSHL/VQRSHLQ:258:result_int32x2 [] = { 7fffffff, 7fffffff, } @@ -4452,25 +4452,25 @@ VQRSHL/VQRSHLQ:277:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 33 VQRSHL/VQRSHLQ:278: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:279:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRSHL/VQRSHLQ (checking overflow: large shift amount with negative input) overflow output: -VQRSHL/VQRSHLQ:280:vqrshl_s8 Neon overflow 1 -VQRSHL/VQRSHLQ:281:vqrshl_s16 Neon overflow 1 -VQRSHL/VQRSHLQ:282:vqrshl_s32 Neon overflow 1 -VQRSHL/VQRSHLQ:283:vqrshl_s64 Neon overflow 1 -VQRSHL/VQRSHLQ:284:vqrshl_u8 Neon overflow 1 -VQRSHL/VQRSHLQ:285:vqrshl_u16 Neon overflow 1 -VQRSHL/VQRSHLQ:286:vqrshl_u32 Neon overflow 1 -VQRSHL/VQRSHLQ:287:vqrshl_u64 Neon overflow 1 -VQRSHL/VQRSHLQ:288:vqrshlq_s8 Neon overflow 1 -VQRSHL/VQRSHLQ:289:vqrshlq_s16 Neon overflow 1 -VQRSHL/VQRSHLQ:290:vqrshlq_s32 Neon overflow 1 -VQRSHL/VQRSHLQ:291:vqrshlq_s64 Neon overflow 1 -VQRSHL/VQRSHLQ:292:vqrshlq_u8 Neon overflow 1 -VQRSHL/VQRSHLQ:293:vqrshlq_u16 Neon overflow 1 -VQRSHL/VQRSHLQ:294:vqrshlq_u32 Neon overflow 1 -VQRSHL/VQRSHLQ:295: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:280:vqrshl_s8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:281:vqrshl_s16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:282:vqrshl_s32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:283:vqrshl_s64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:284:vqrshl_u8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:285:vqrshl_u16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:286:vqrshl_u32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:287:vqrshl_u64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:288:vqrshlq_s8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:289:vqrshlq_s16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:290:vqrshlq_s32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:291:vqrshlq_s64 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:292:vqrshlq_u8 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:293:vqrshlq_u16 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:294:vqrshlq_u32 Neon cumulative saturation 1 +VQRSHL/VQRSHLQ:295:vqrshlq_u64 Neon cumulative saturation 1 + +VQRSHL/VQRSHLQ (checking cumulative saturation: large shift amount with negative input) output: VQRSHL/VQRSHLQ:296:result_int8x8 [] = { ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, ffffff80, } VQRSHL/VQRSHLQ:297:result_int16x4 [] = { ffff8000, ffff8000, ffff8000, ffff8000, } VQRSHL/VQRSHLQ:298:result_int32x2 [] = { 80000000, 80000000, } @@ -4496,25 +4496,25 @@ VQRSHL/VQRSHLQ:317:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 33 VQRSHL/VQRSHLQ:318: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:319:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRSHL/VQRSHLQ (checking overflow: large negative shift amount) overflow output: -VQRSHL/VQRSHLQ:320:vqrshl_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:321:vqrshl_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:322:vqrshl_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:323:vqrshl_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:324:vqrshl_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:325:vqrshl_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:326:vqrshl_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:327:vqrshl_u64 Neon overflow 0 -VQRSHL/VQRSHLQ:328:vqrshlq_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:329:vqrshlq_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:330:vqrshlq_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:331:vqrshlq_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:332:vqrshlq_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:333:vqrshlq_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:334:vqrshlq_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:335: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:320:vqrshl_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:321:vqrshl_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:322:vqrshl_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:323:vqrshl_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:324:vqrshl_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:325:vqrshl_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:326:vqrshl_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:327:vqrshl_u64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:328:vqrshlq_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:329:vqrshlq_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:330:vqrshlq_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:331:vqrshlq_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:332:vqrshlq_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:333:vqrshlq_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:334:vqrshlq_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:335:vqrshlq_u64 Neon cumulative saturation 0 + +VQRSHL/VQRSHLQ (checking cumulative saturation: large negative shift amount) output: VQRSHL/VQRSHLQ:336:result_int8x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } VQRSHL/VQRSHLQ:337:result_int16x4 [] = { 0, 0, 0, 0, } VQRSHL/VQRSHLQ:338:result_int32x2 [] = { 0, 0, } @@ -4540,25 +4540,25 @@ VQRSHL/VQRSHLQ:357:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 33 VQRSHL/VQRSHLQ:358: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:359:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRSHL/VQRSHLQ (checking overflow: large shift amount with 0 input) overflow output: -VQRSHL/VQRSHLQ:360:vqrshl_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:361:vqrshl_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:362:vqrshl_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:363:vqrshl_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:364:vqrshl_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:365:vqrshl_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:366:vqrshl_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:367:vqrshl_u64 Neon overflow 0 -VQRSHL/VQRSHLQ:368:vqrshlq_s8 Neon overflow 0 -VQRSHL/VQRSHLQ:369:vqrshlq_s16 Neon overflow 0 -VQRSHL/VQRSHLQ:370:vqrshlq_s32 Neon overflow 0 -VQRSHL/VQRSHLQ:371:vqrshlq_s64 Neon overflow 0 -VQRSHL/VQRSHLQ:372:vqrshlq_u8 Neon overflow 0 -VQRSHL/VQRSHLQ:373:vqrshlq_u16 Neon overflow 0 -VQRSHL/VQRSHLQ:374:vqrshlq_u32 Neon overflow 0 -VQRSHL/VQRSHLQ:375: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:360:vqrshl_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:361:vqrshl_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:362:vqrshl_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:363:vqrshl_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:364:vqrshl_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:365:vqrshl_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:366:vqrshl_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:367:vqrshl_u64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:368:vqrshlq_s8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:369:vqrshlq_s16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:370:vqrshlq_s32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:371:vqrshlq_s64 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:372:vqrshlq_u8 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:373:vqrshlq_u16 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:374:vqrshlq_u32 Neon cumulative saturation 0 +VQRSHL/VQRSHLQ:375:vqrshlq_u64 Neon cumulative saturation 0 + +VQRSHL/VQRSHLQ (checking cumulative saturation: large shift amount with 0 input) output: VQRSHL/VQRSHLQ:376:result_int8x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } VQRSHL/VQRSHLQ:377:result_int16x4 [] = { 0, 0, 0, 0, } VQRSHL/VQRSHLQ:378:result_int32x2 [] = { 0, 0, } @@ -6048,13 +6048,13 @@ VMVN/VMVNQ:21:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3 VMVN/VMVNQ:22: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, } VMVN/VMVNQ:23:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -6082,13 +6082,13 @@ VQMOVN:27:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQMOVN:28: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:29:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQMOVN overflow output: -VQMOVN:30:vqmovn_s16 Neon overflow 1 -VQMOVN:31:vqmovn_s32 Neon overflow 1 -VQMOVN:32:vqmovn_s64 Neon overflow 1 -VQMOVN:33:vqmovn_u16 Neon overflow 1 -VQMOVN:34:vqmovn_u32 Neon overflow 1 -VQMOVN:35:vqmovn_u64 Neon overflow 1 +VQMOVN cumulative saturation output: +VQMOVN:30:vqmovn_s16 Neon cumulative saturation 1 +VQMOVN:31:vqmovn_s32 Neon cumulative saturation 1 +VQMOVN:32:vqmovn_s64 Neon cumulative saturation 1 +VQMOVN:33:vqmovn_u16 Neon cumulative saturation 1 +VQMOVN:34:vqmovn_u32 Neon cumulative saturation 1 +VQMOVN:35:vqmovn_u64 Neon cumulative saturation 1 VQMOVN output: VQMOVN:36:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -6116,10 +6116,10 @@ VQMOVN:57:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQMOVN:58: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:59:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -6147,10 +6147,10 @@ VQMOVUN:24:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333 VQMOVUN: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, } VQMOVUN:26:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQMOVUN (negative input) overflow output: -VQMOVUN:27:vqmovun_s16 Neon overflow 1 -VQMOVUN:28:vqmovun_s32 Neon overflow 1 -VQMOVUN:29:vqmovun_s64 Neon overflow 1 +VQMOVUN (negative input) cumulative saturation output: +VQMOVUN:27:vqmovun_s16 Neon cumulative saturation 1 +VQMOVUN:28:vqmovun_s32 Neon cumulative saturation 1 +VQMOVUN:29:vqmovun_s64 Neon cumulative saturation 1 VQMOVUN (negative input) output: VQMOVUN:30:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -6620,15 +6620,15 @@ VPADAL/VPADALQ:21:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 333 VPADAL/VPADALQ:22: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, } VPADAL/VPADALQ:23:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -6656,17 +6656,17 @@ VQSHLU_N/VQSHLUQ_N:29:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, VQSHLU_N/VQSHLUQ_N:30: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:31:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSHLU_N/VQSHLUQ_N (check saturation/overflow: shift by 1) overflow output: -VQSHLU_N/VQSHLUQ_N:32:vqshlu_n_s8 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:33:vqshlu_n_s16 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:34:vqshlu_n_s32 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:35:vqshlu_n_s64 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:36:vqshluq_n_s8 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:37:vqshluq_n_s16 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:38:vqshluq_n_s32 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:39: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:32:vqshlu_n_s8 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:33:vqshlu_n_s16 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:34:vqshlu_n_s32 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:35:vqshlu_n_s64 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:36:vqshluq_n_s8 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:37:vqshluq_n_s16 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:38:vqshluq_n_s32 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:39:vqshluq_n_s64 Neon cumulative saturation 0 + +VQSHLU_N/VQSHLUQ_N (check cumulative saturation: shift by 1) output: VQSHLU_N/VQSHLUQ_N:40:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQSHLU_N/VQSHLUQ_N:41:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQSHLU_N/VQSHLUQ_N:42:result_int32x2 [] = { 33333333, 33333333, } @@ -6692,17 +6692,17 @@ VQSHLU_N/VQSHLUQ_N:61:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, VQSHLU_N/VQSHLUQ_N:62: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:63:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSHLU_N/VQSHLUQ_N (check saturation/overflow: shift by 2) overflow output: -VQSHLU_N/VQSHLUQ_N:64:vqshlu_n_s8 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:65:vqshlu_n_s16 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:66:vqshlu_n_s32 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:67:vqshlu_n_s64 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:68:vqshluq_n_s8 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:69:vqshluq_n_s16 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:70:vqshluq_n_s32 Neon overflow 1 -VQSHLU_N/VQSHLUQ_N:71: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:64:vqshlu_n_s8 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:65:vqshlu_n_s16 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:66:vqshlu_n_s32 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:67:vqshlu_n_s64 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:68:vqshluq_n_s8 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:69:vqshluq_n_s16 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:70:vqshluq_n_s32 Neon cumulative saturation 1 +VQSHLU_N/VQSHLUQ_N:71:vqshluq_n_s64 Neon cumulative saturation 1 + +VQSHLU_N/VQSHLUQ_N (check cumulative saturation: shift by 2) output: VQSHLU_N/VQSHLUQ_N:72:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQSHLU_N/VQSHLUQ_N:73:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQSHLU_N/VQSHLUQ_N:74:result_int32x2 [] = { 33333333, 33333333, } @@ -6728,15 +6728,15 @@ VQSHLU_N/VQSHLUQ_N:93:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, VQSHLU_N/VQSHLUQ_N:94: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:95:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSHLU_N/VQSHLUQ_N overflow output: -VQSHLU_N/VQSHLUQ_N:96:vqshlu_n_s8 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:97:vqshlu_n_s16 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:98:vqshlu_n_s32 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:99:vqshlu_n_s64 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:100:vqshluq_n_s8 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:101:vqshluq_n_s16 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:102:vqshluq_n_s32 Neon overflow 0 -VQSHLU_N/VQSHLUQ_N:103:vqshluq_n_s64 Neon overflow 0 +VQSHLU_N/VQSHLUQ_N cumulative saturation output: +VQSHLU_N/VQSHLUQ_N:96:vqshlu_n_s8 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:97:vqshlu_n_s16 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:98:vqshlu_n_s32 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:99:vqshlu_n_s64 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:100:vqshluq_n_s8 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:101:vqshluq_n_s16 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:102:vqshluq_n_s32 Neon cumulative saturation 0 +VQSHLU_N/VQSHLUQ_N:103:vqshluq_n_s64 Neon cumulative saturation 0 VQSHLU_N/VQSHLUQ_N output: VQSHLU_N/VQSHLUQ_N:104:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -6894,13 +6894,13 @@ VCNT/VCNTQ:21:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3 VCNT/VCNTQ:22: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, } VCNT/VCNTQ:23:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -6928,13 +6928,13 @@ VQSHRN_N:27:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 333 VQSHRN_N:28: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:29:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSHRN_N (check saturation: shift by 3) overflow output: -VQSHRN_N:30:vqshrn_n_s16 Neon overflow 1 -VQSHRN_N:31:vqshrn_n_s32 Neon overflow 1 -VQSHRN_N:32:vqshrn_n_s64 Neon overflow 1 -VQSHRN_N:33:vqshrn_n_u16 Neon overflow 1 -VQSHRN_N:34:vqshrn_n_u32 Neon overflow 1 -VQSHRN_N:35:vqshrn_n_u64 Neon overflow 1 +VQSHRN_N (check saturation: shift by 3) cumulative saturation output: +VQSHRN_N:30:vqshrn_n_s16 Neon cumulative saturation 1 +VQSHRN_N:31:vqshrn_n_s32 Neon cumulative saturation 1 +VQSHRN_N:32:vqshrn_n_s64 Neon cumulative saturation 1 +VQSHRN_N:33:vqshrn_n_u16 Neon cumulative saturation 1 +VQSHRN_N:34:vqshrn_n_u32 Neon cumulative saturation 1 +VQSHRN_N:35:vqshrn_n_u64 Neon cumulative saturation 1 VQSHRN_N (check saturation: shift by 3) output: VQSHRN_N:36:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -6962,13 +6962,13 @@ VQSHRN_N:57:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 333 VQSHRN_N:58: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:59:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSHRN_N (check saturation: shift by max) overflow output: -VQSHRN_N:60:vqshrn_n_s16 Neon overflow 0 -VQSHRN_N:61:vqshrn_n_s32 Neon overflow 0 -VQSHRN_N:62:vqshrn_n_s64 Neon overflow 0 -VQSHRN_N:63:vqshrn_n_u16 Neon overflow 0 -VQSHRN_N:64:vqshrn_n_u32 Neon overflow 0 -VQSHRN_N:65:vqshrn_n_u64 Neon overflow 0 +VQSHRN_N (check saturation: shift by max) cumulative saturation output: +VQSHRN_N:60:vqshrn_n_s16 Neon cumulative saturation 0 +VQSHRN_N:61:vqshrn_n_s32 Neon cumulative saturation 0 +VQSHRN_N:62:vqshrn_n_s64 Neon cumulative saturation 0 +VQSHRN_N:63:vqshrn_n_u16 Neon cumulative saturation 0 +VQSHRN_N:64:vqshrn_n_u32 Neon cumulative saturation 0 +VQSHRN_N:65:vqshrn_n_u64 Neon cumulative saturation 0 VQSHRN_N (check saturation: shift by max) output: VQSHRN_N:66:result_int8x8 [] = { 7f, 7f, 7f, 7f, 7f, 7f, 7f, 7f, } @@ -7048,10 +7048,10 @@ VPMIN:21:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3333, VPMIN:22: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, } VPMIN:23:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -7079,12 +7079,12 @@ VQSHRUN_N:24:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 33 VQSHRUN_N: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, } VQSHRUN_N:26:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSHRUN_N (check saturation/overflow) overflow output: -VQSHRUN_N:27:vqshrun_n_s16 Neon overflow 1 -VQSHRUN_N:28:vqshrun_n_s32 Neon overflow 1 -VQSHRUN_N:29:vqshrun_n_s64 Neon overflow 1 +VQSHRUN_N (check cumulative saturation) cumulative saturation output: +VQSHRUN_N:27:vqshrun_n_s16 Neon cumulative saturation 1 +VQSHRUN_N:28:vqshrun_n_s32 Neon cumulative saturation 1 +VQSHRUN_N:29:vqshrun_n_s64 Neon cumulative saturation 1 -VQSHRUN_N (check saturation/overflow) output: +VQSHRUN_N (check cumulative saturation) output: VQSHRUN_N:30:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQSHRUN_N:31:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQSHRUN_N:32:result_int32x2 [] = { 33333333, 33333333, } @@ -7110,10 +7110,10 @@ VQSHRUN_N:51:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 33 VQSHRUN_N:52: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:53:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQSHRUN_N overflow output: -VQSHRUN_N:54:vqshrun_n_s16 Neon overflow 0 -VQSHRUN_N:55:vqshrun_n_s32 Neon overflow 1 -VQSHRUN_N:56:vqshrun_n_s64 Neon overflow 0 +VQSHRUN_N cumulative saturation output: +VQSHRUN_N:54:vqshrun_n_s16 Neon cumulative saturation 0 +VQSHRUN_N:55:vqshrun_n_s32 Neon cumulative saturation 1 +VQSHRUN_N:56:vqshrun_n_s64 Neon cumulative saturation 0 VQSHRUN_N output: VQSHRUN_N:57:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } @@ -7141,10 +7141,10 @@ VQSHRUN_N:78:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 33 VQSHRUN_N:79: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:80:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -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, } @@ -7172,12 +7172,12 @@ VQRSHRUN_N:24:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3 VQRSHRUN_N: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, } VQRSHRUN_N:26:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRSHRUN_N (check saturation/overflow: shift by 1) overflow output: -VQRSHRUN_N:27:vqrshrun_n_s16 Neon overflow 1 -VQRSHRUN_N:28:vqrshrun_n_s32 Neon overflow 1 -VQRSHRUN_N:29:vqrshrun_n_s64 Neon overflow 1 +VQRSHRUN_N (check cumulative saturation: shift by 1) cumulative saturation output: +VQRSHRUN_N:27:vqrshrun_n_s16 Neon cumulative saturation 1 +VQRSHRUN_N:28:vqrshrun_n_s32 Neon cumulative saturation 1 +VQRSHRUN_N:29: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:30:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRSHRUN_N:31:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQRSHRUN_N:32:result_int32x2 [] = { 33333333, 33333333, } @@ -7203,12 +7203,12 @@ VQRSHRUN_N:51:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3 VQRSHRUN_N:52: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:53:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRSHRUN_N (check saturation/overflow: shift by max, positive input) overflow output: -VQRSHRUN_N:54:vqrshrun_n_s16 Neon overflow 0 -VQRSHRUN_N:55:vqrshrun_n_s32 Neon overflow 0 -VQRSHRUN_N:56:vqrshrun_n_s64 Neon overflow 0 +VQRSHRUN_N (check cumulative saturation: shift by max, positive input) cumulative saturation output: +VQRSHRUN_N:54:vqrshrun_n_s16 Neon cumulative saturation 0 +VQRSHRUN_N:55:vqrshrun_n_s32 Neon cumulative saturation 0 +VQRSHRUN_N:56: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:57:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRSHRUN_N:58:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQRSHRUN_N:59:result_int32x2 [] = { 33333333, 33333333, } @@ -7234,12 +7234,12 @@ VQRSHRUN_N:78:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, 3 VQRSHRUN_N:79: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:80:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRSHRUN_N (check saturation/overflow: shift by max, negative input) overflow output: -VQRSHRUN_N:81:vqrshrun_n_s16 Neon overflow 1 -VQRSHRUN_N:82:vqrshrun_n_s32 Neon overflow 1 -VQRSHRUN_N:83:vqrshrun_n_s64 Neon overflow 1 +VQRSHRUN_N (check cumulative saturation: shift by max, negative input) cumulative saturation output: +VQRSHRUN_N:81:vqrshrun_n_s16 Neon cumulative saturation 1 +VQRSHRUN_N:82:vqrshrun_n_s32 Neon cumulative saturation 1 +VQRSHRUN_N:83: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:84:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } VQRSHRUN_N:85:result_int16x4 [] = { 3333, 3333, 3333, 3333, } VQRSHRUN_N:86:result_int32x2 [] = { 33333333, 33333333, } @@ -7265,10 +7265,10 @@ VQRSHRUN_N:105:result_poly16x8 [] = { 3333, 3333, 3333, 3333, 3333, 3333, 3333, VQRSHRUN_N:106: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:107:result_float16x8 [] = { 0, 0, 0, 0, 0, 0, 0, 0, } -VQRSHRUN_N overflow output: -VQRSHRUN_N:108:vqrshrun_n_s16 Neon overflow 0 -VQRSHRUN_N:109:vqrshrun_n_s32 Neon overflow 1 -VQRSHRUN_N:110:vqrshrun_n_s64 Neon overflow 0 +VQRSHRUN_N cumulative saturation output: +VQRSHRUN_N:108:vqrshrun_n_s16 Neon cumulative saturation 0 +VQRSHRUN_N:109:vqrshrun_n_s32 Neon cumulative saturation 1 +VQRSHRUN_N:110:vqrshrun_n_s64 Neon cumulative saturation 0 VQRSHRUN_N output: VQRSHRUN_N:111:result_int8x8 [] = { 33, 33, 33, 33, 33, 33, 33, 33, } |