---

Intrinsics wrapped for AARCH64

Notations are as follows:

• T for trick usually using other intrinsics

• E for scalar emulation

• NOOP for no operation

• NA means the operator does not exist for the given type

• intrinsic for the actual wrapped intrinsic

abs

• abs on i8: vabsq_s8

• abs on u8: NOOP

• abs on i16: vabsq_s16

• abs on u16: NOOP

• abs on i32: vabsq_s32

• abs on u32: NOOP

• abs on i64: vabsq_s64

• abs on u64: NOOP

• abs on f16: T

• abs on f32: vabsq_f32

• abs on f64: vabsq_f64

add

• add on i8: vaddq_s8

• add on u8: vaddq_u8

• add on i16: vaddq_s16

• add on u16: vaddq_u16

• add on i32: vaddq_s32

• add on u32: vaddq_u32

• add on i64: vaddq_s64

• add on u64: vaddq_u64

• add on f16: T

• add on f32: vaddq_f32

• add on f64: vaddq_f64

adds

• adds on i8: vqaddq_s8

• adds on u8: vqaddq_u8

• adds on i16: vqaddq_s16

• adds on u16: vqaddq_u16

• adds on i32: vqaddq_s32

• adds on u32: vqaddq_u32

• adds on i64: vqaddq_s64

• adds on u64: vqaddq_u64

• adds on f16: T

• adds on f32: T

• adds on f64: T

addv

• addv on i8: NA

• addv on u8: NA

• addv on i16: NA

• addv on u16: NA

• addv on i32: NA

• addv on u32: NA

• addv on i64: NA

• addv on u64: NA

• addv on f16: T

• addv on f32: vaddvq_f32

• addv on f64: vaddvq_f64

all

• all on i8: T

• all on u8: T

• all on i16: T

• all on u16: T

• all on i32: T

• all on u32: T

• all on i64: T

• all on u64: T

• all on f16: T

• all on f32: T

• all on f64: T

andb

• andb on i8: vandq_s8

• andb on u8: vandq_u8

• andb on i16: vandq_s16

• andb on u16: vandq_u16

• andb on i32: vandq_s32

• andb on u32: vandq_u32

• andb on i64: vandq_s64

• andb on u64: vandq_u64

• andb on f16: T

• andb on f32: vandq_u32

• andb on f64: vandq_u64

andl

• andl on i8: vandq_u8

• andl on u8: vandq_u8

• andl on i16: vandq_u16

• andl on u16: vandq_u16

• andl on i32: vandq_u32

• andl on u32: vandq_u32

• andl on i64: vandq_u64

• andl on u64: vandq_u64

• andl on f16: T

• andl on f32: vandq_u32

• andl on f64: vandq_u64

andnotb

• andnotb on i8: vbicq_s8

• andnotb on u8: vbicq_u8

• andnotb on i16: vbicq_s16

• andnotb on u16: vbicq_u16

• andnotb on i32: vbicq_s32

• andnotb on u32: vbicq_u32

• andnotb on i64: vbicq_s64

• andnotb on u64: vbicq_u64

• andnotb on f16: T

• andnotb on f32: vbicq_u32

• andnotb on f64: vbicq_u64

andnotl

• andnotl on i8: vbicq_u8

• andnotl on u8: vbicq_u8

• andnotl on i16: vbicq_u16

• andnotl on u16: vbicq_u16

• andnotl on i32: vbicq_u32

• andnotl on u32: vbicq_u32

• andnotl on i64: vbicq_u64

• andnotl on u64: vbicq_u64

• andnotl on f16: T

• andnotl on f32: vbicq_u32

• andnotl on f64: vbicq_u64

any

• any on i8: T

• any on u8: T

• any on i16: T

• any on u16: T

• any on i32: T

• any on u32: T

• any on i64: T

• any on u64: T

• any on f16: T

• any on f32: T

• any on f64: T

ceil

• ceil on i8: NOOP

• ceil on u8: NOOP

• ceil on i16: NOOP

• ceil on u16: NOOP

• ceil on i32: NOOP

• ceil on u32: NOOP

• ceil on i64: NOOP

• ceil on u64: NOOP

• ceil on f16: T

• ceil on f32: vrndpq_f32

• ceil on f64: vrndpq_f64

div

• div on i8: E

• div on u8: E

• div on i16: E

• div on u16: E

• div on i32: E

• div on u32: E

• div on i64: E

• div on u64: E

• div on f16: T

• div on f32: vdivq_f32

• div on f64: vdivq_f64

eq

• eq on i8: vceqq_s8

• eq on u8: vceqq_u8

• eq on i16: vceqq_s16

• eq on u16: vceqq_u16

• eq on i32: vceqq_s32

• eq on u32: vceqq_u32

• eq on i64: vceqq_s64

• eq on u64: vceqq_u64

• eq on f16: T

• eq on f32: vceqq_f32

• eq on f64: vceqq_f64

floor

• floor on i8: NOOP

• floor on u8: NOOP

• floor on i16: NOOP

• floor on u16: NOOP

• floor on i32: NOOP

• floor on u32: NOOP

• floor on i64: NOOP

• floor on u64: NOOP

• floor on f16: T

• floor on f32: vrndmq_f32

• floor on f64: vrndmq_f64

fma

• fma on i8: vmlaq_s8

• fma on u8: vmlaq_u8

• fma on i16: vmlaq_s16

• fma on u16: vmlaq_u16

• fma on i32: vmlaq_s32

• fma on u32: vmlaq_u32

• fma on i64: E

• fma on u64: E

• fma on f16: E

• fma on f32: vfmaq_f32

• fma on f64: vfmaq_f64

fms

• fms on i8: T

• fms on u8: T

• fms on i16: T

• fms on u16: T

• fms on i32: T

• fms on u32: T

• fms on i64: T

• fms on u64: T

• fms on f16: T

• fms on f32: T

• fms on f64: T

fnma

• fnma on i8: vmlsq_s8

• fnma on u8: vmlsq_u8

• fnma on i16: vmlsq_s16

• fnma on u16: vmlsq_u16

• fnma on i32: vmlsq_s32

• fnma on u32: vmlsq_u32

• fnma on i64: E

• fnma on u64: E

• fnma on f16: E

• fnma on f32: vfmsq_f32

• fnma on f64: vfmsq_f64

fnms

• fnms on i8: T

• fnms on u8: T

• fnms on i16: T

• fnms on u16: T

• fnms on i32: T

• fnms on u32: T

• fnms on i64: T

• fnms on u64: T

• fnms on f16: T

• fnms on f32: T

• fnms on f64: T

gather

• gather on i8: NA

• gather on u8: NA

• gather on i16: T

• gather on u16: T

• gather on i32: T

• gather on u32: T

• gather on i64: T

• gather on u64: T

• gather on f16: T

• gather on f32: T

• gather on f64: T

gather_linear

• gather_linear on i8: T

• gather_linear on u8: T

• gather_linear on i16: T

• gather_linear on u16: T

• gather_linear on i32: T

• gather_linear on u32: T

• gather_linear on i64: T

• gather_linear on u64: T

• gather_linear on f16: E

• gather_linear on f32: T

• gather_linear on f64: T

ge

• ge on i8: vcgeq_s8

• ge on u8: vcgeq_u8

• ge on i16: vcgeq_s16

• ge on u16: vcgeq_u16

• ge on i32: vcgeq_s32

• ge on u32: vcgeq_u32

• ge on i64: vcgeq_s64

• ge on u64: vcgeq_u64

• ge on f16: T

• ge on f32: vcgeq_f32

• ge on f64: vcgeq_f64

gt

• gt on i8: vcgtq_s8

• gt on u8: vcgtq_u8

• gt on i16: vcgtq_s16

• gt on u16: vcgtq_u16

• gt on i32: vcgtq_s32

• gt on u32: vcgtq_u32

• gt on i64: vcgtq_s64

• gt on u64: vcgtq_u64

• gt on f16: T

• gt on f32: vcgtq_f32

• gt on f64: vcgtq_f64

if_else1

• if_else1 on i8: vbslq_s8

• if_else1 on u8: vbslq_u8

• if_else1 on i16: vbslq_s16

• if_else1 on u16: vbslq_u16

• if_else1 on i32: vbslq_s32

• if_else1 on u32: vbslq_u32

• if_else1 on i64: vbslq_s64

• if_else1 on u64: vbslq_u64

• if_else1 on f16: T

• if_else1 on f32: vbslq_f32

• if_else1 on f64: vbslq_f64

iota

• iota on i8: T

• iota on u8: T

• iota on i16: T

• iota on u16: T

• iota on i32: T

• iota on u32: T

• iota on i64: T

• iota on u64: T

• iota on f16: T

• iota on f32: T

• iota on f64: T

le

• le on i8: vcleq_s8

• le on u8: vcleq_u8

• le on i16: vcleq_s16

• le on u16: vcleq_u16

• le on i32: vcleq_s32

• le on u32: vcleq_u32

• le on i64: vcleq_s64

• le on u64: vcleq_u64

• le on f16: T

• le on f32: vcleq_f32

• le on f64: vcleq_f64

len

• len on i8: NOOP

• len on u8: NOOP

• len on i16: NOOP

• len on u16: NOOP

• len on i32: NOOP

• len on u32: NOOP

• len on i64: NOOP

• len on u64: NOOP

• len on f16: NOOP

• len on f32: NOOP

• len on f64: NOOP

load2a

• load2a on i8: T

• load2a on u8: T

• load2a on i16: T

• load2a on u16: T

• load2a on i32: T

• load2a on u32: T

• load2a on i64: T

• load2a on u64: T

• load2a on f16: T

• load2a on f32: T

• load2a on f64: T

load2u

• load2u on i8: T

• load2u on u8: T

• load2u on i16: T

• load2u on u16: T

• load2u on i32: T

• load2u on u32: T

• load2u on i64: T

• load2u on u64: T

• load2u on f16: T

• load2u on f32: T

• load2u on f64: T

load3a

• load3a on i8: T

• load3a on u8: T

• load3a on i16: T

• load3a on u16: T

• load3a on i32: T

• load3a on u32: T

• load3a on i64: T

• load3a on u64: T

• load3a on f16: T

• load3a on f32: T

• load3a on f64: T

load3u

• load3u on i8: T

• load3u on u8: T

• load3u on i16: T

• load3u on u16: T

• load3u on i32: T

• load3u on u32: T

• load3u on i64: T

• load3u on u64: T

• load3u on f16: T

• load3u on f32: T

• load3u on f64: T

load4a

• load4a on i8: T

• load4a on u8: T

• load4a on i16: T

• load4a on u16: T

• load4a on i32: T

• load4a on u32: T

• load4a on i64: T

• load4a on u64: T

• load4a on f16: T

• load4a on f32: T

• load4a on f64: T

load4u

• load4u on i8: T

• load4u on u8: T

• load4u on i16: T

• load4u on u16: T

• load4u on i32: T

• load4u on u32: T

• load4u on i64: T

• load4u on u64: T

• load4u on f16: T

• load4u on f32: T

• load4u on f64: T

loada

• loada on i8: vld1q_s8

• loada on u8: vld1q_u8

• loada on i16: vld1q_s16

• loada on u16: vld1q_u16

• loada on i32: vld1q_s32

• loada on u32: vld1q_u32

• loada on i64: vld1q_s64

• loada on u64: vld1q_u64

• loada on f16: T

• loada on f32: vld1q_f32

• loada on f64: vld1q_f64

loadla

• loadla on i8: T

• loadla on u8: T

• loadla on i16: T

• loadla on u16: T

• loadla on i32: T

• loadla on u32: T

• loadla on i64: T

• loadla on u64: T

• loadla on f16: T

• loadla on f32: T

• loadla on f64: T

loadlu

• loadlu on i8: T

• loadlu on u8: T

• loadlu on i16: T

• loadlu on u16: T

• loadlu on i32: T

• loadlu on u32: T

• loadlu on i64: T

• loadlu on u64: T

• loadlu on f16: T

• loadlu on f32: T

• loadlu on f64: T

loadu

• loadu on i8: vld1q_s8

• loadu on u8: vld1q_u8

• loadu on i16: vld1q_s16

• loadu on u16: vld1q_u16

• loadu on i32: vld1q_s32

• loadu on u32: vld1q_u32

• loadu on i64: vld1q_s64

• loadu on u64: vld1q_u64

• loadu on f16: T

• loadu on f32: vld1q_f32

• loadu on f64: vld1q_f64

lt

• lt on i8: vcltq_s8

• lt on u8: vcltq_u8

• lt on i16: vcltq_s16

• lt on u16: vcltq_u16

• lt on i32: vcltq_s32

• lt on u32: vcltq_u32

• lt on i64: vcltq_s64

• lt on u64: vcltq_u64

• lt on f16: T

• lt on f32: vcltq_f32

• lt on f64: vcltq_f64

mask_for_loop_tail

• mask_for_loop_tail on i8: T

• mask_for_loop_tail on u8: T

• mask_for_loop_tail on i16: T

• mask_for_loop_tail on u16: T

• mask_for_loop_tail on i32: T

• mask_for_loop_tail on u32: T

• mask_for_loop_tail on i64: T

• mask_for_loop_tail on u64: T

• mask_for_loop_tail on f16: T

• mask_for_loop_tail on f32: T

• mask_for_loop_tail on f64: T

mask_storea1

• mask_storea1 on i8: E

• mask_storea1 on u8: E

• mask_storea1 on i16: E

• mask_storea1 on u16: E

• mask_storea1 on i32: E

• mask_storea1 on u32: E

• mask_storea1 on i64: E

• mask_storea1 on u64: E

• mask_storea1 on f16: E

• mask_storea1 on f32: E

• mask_storea1 on f64: E

mask_storeu1

• mask_storeu1 on i8: E

• mask_storeu1 on u8: E

• mask_storeu1 on i16: E

• mask_storeu1 on u16: E

• mask_storeu1 on i32: E

• mask_storeu1 on u32: E

• mask_storeu1 on i64: E

• mask_storeu1 on u64: E

• mask_storeu1 on f16: E

• mask_storeu1 on f32: E

• mask_storeu1 on f64: E

masko_loada1

• masko_loada1 on i8: E

• masko_loada1 on u8: E

• masko_loada1 on i16: E

• masko_loada1 on u16: E

• masko_loada1 on i32: E

• masko_loada1 on u32: E

• masko_loada1 on i64: E

• masko_loada1 on u64: E

• masko_loada1 on f16: E

• masko_loada1 on f32: E

• masko_loada1 on f64: E

masko_loadu1

• masko_loadu1 on i8: E

• masko_loadu1 on u8: E

• masko_loadu1 on i16: E

• masko_loadu1 on u16: E

• masko_loadu1 on i32: E

• masko_loadu1 on u32: E

• masko_loadu1 on i64: E

• masko_loadu1 on u64: E

• masko_loadu1 on f16: E

• masko_loadu1 on f32: E

• masko_loadu1 on f64: E

maskz_loada1

• maskz_loada1 on i8: E

• maskz_loada1 on u8: E

• maskz_loada1 on i16: E

• maskz_loada1 on u16: E

• maskz_loada1 on i32: E

• maskz_loada1 on u32: E

• maskz_loada1 on i64: E

• maskz_loada1 on u64: E

• maskz_loada1 on f16: E

• maskz_loada1 on f32: E

• maskz_loada1 on f64: E

maskz_loadu1

• maskz_loadu1 on i8: E

• maskz_loadu1 on u8: E

• maskz_loadu1 on i16: E

• maskz_loadu1 on u16: E

• maskz_loadu1 on i32: E

• maskz_loadu1 on u32: E

• maskz_loadu1 on i64: E

• maskz_loadu1 on u64: E

• maskz_loadu1 on f16: E

• maskz_loadu1 on f32: E

• maskz_loadu1 on f64: E

max

• max on i8: vmaxq_s8

• max on u8: vmaxq_u8

• max on i16: vmaxq_s16

• max on u16: vmaxq_u16

• max on i32: vmaxq_s32

• max on u32: vmaxq_u32

• max on i64: T

• max on u64: T

• max on f16: T

• max on f32: vmaxq_f32

• max on f64: vmaxq_f64

min

• min on i8: vminq_s8

• min on u8: vminq_u8

• min on i16: vminq_s16

• min on u16: vminq_u16

• min on i32: vminq_s32

• min on u32: vminq_u32

• min on i64: T

• min on u64: T

• min on f16: T

• min on f32: vminq_f32

• min on f64: vminq_f64

mul

• mul on i8: vmulq_s8

• mul on u8: vmulq_u8

• mul on i16: vmulq_s16

• mul on u16: vmulq_u16

• mul on i32: vmulq_s32

• mul on u32: vmulq_u32

• mul on i64: E

• mul on u64: E

• mul on f16: T

• mul on f32: vmulq_f32

• mul on f64: vmulq_f64

nbtrue

• nbtrue on i8: vaddvq_s8

• nbtrue on u8: vaddvq_s8

• nbtrue on i16: vaddvq_s16

• nbtrue on u16: vaddvq_s16

• nbtrue on i32: vaddvq_s32

• nbtrue on u32: vaddvq_s32

• nbtrue on i64: T

• nbtrue on u64: T

• nbtrue on f16: T

• nbtrue on f32: vaddvq_s32

• nbtrue on f64: T

ne

• ne on i8: T

• ne on u8: T

• ne on i16: T

• ne on u16: T

• ne on i32: T

• ne on u32: T

• ne on i64: T

• ne on u64: T

• ne on f16: T

• ne on f32: T

• ne on f64: T

neg

• neg on i8: vnegq_s8

• neg on u8: vnegq_s8

• neg on i16: vnegq_s16

• neg on u16: vnegq_s16

• neg on i32: vnegq_s32

• neg on u32: vnegq_s32

• neg on i64: vnegq_s64

• neg on u64: vnegq_s64

• neg on f16: T

• neg on f32: vnegq_f32

• neg on f64: vnegq_f64

notb

• notb on i8: vmvnq_s8

• notb on u8: vmvnq_u8

• notb on i16: vmvnq_s16

• notb on u16: vmvnq_u16

• notb on i32: vmvnq_s32

• notb on u32: vmvnq_u32

• notb on i64: vmvnq_u32

• notb on u64: vmvnq_u32

• notb on f16: T

• notb on f32: vmvnq_u32

• notb on f64: vmvnq_u32

notl

• notl on i8: vmvnq_u8

• notl on u8: vmvnq_u8

• notl on i16: vmvnq_u16

• notl on u16: vmvnq_u16

• notl on i32: vmvnq_u32

• notl on u32: vmvnq_u32

• notl on i64: vmvnq_u32

• notl on u64: vmvnq_u32

• notl on f16: T

• notl on f32: vmvnq_u32

• notl on f64: vmvnq_u32

orb

• orb on i8: vorrq_s8

• orb on u8: vorrq_u8

• orb on i16: vorrq_s16

• orb on u16: vorrq_u16

• orb on i32: vorrq_s32

• orb on u32: vorrq_u32

• orb on i64: vorrq_s64

• orb on u64: vorrq_u64

• orb on f16: T

• orb on f32: vorrq_u32

• orb on f64: vorrq_u64

orl

• orl on i8: vorrq_u8

• orl on u8: vorrq_u8

• orl on i16: vorrq_u16

• orl on u16: vorrq_u16

• orl on i32: vorrq_u32

• orl on u32: vorrq_u32

• orl on i64: vorrq_u64

• orl on u64: vorrq_u64

• orl on f16: T

• orl on f32: vorrq_u32

• orl on f64: vorrq_u64

rec

• rec on i8: NA

• rec on u8: NA

• rec on i16: NA

• rec on u16: NA

• rec on i32: NA

• rec on u32: NA

• rec on i64: NA

• rec on u64: NA

• rec on f16: T

• rec on f32: T

• rec on f64: T

rec11

• rec11 on i8: NA

• rec11 on u8: NA

• rec11 on i16: NA

• rec11 on u16: NA

• rec11 on i32: NA

• rec11 on u32: NA

• rec11 on i64: NA

• rec11 on u64: NA

• rec11 on f16: T

• rec11 on f32: T

• rec11 on f64: T

rec8

• rec8 on i8: NA

• rec8 on u8: NA

• rec8 on i16: NA

• rec8 on u16: NA

• rec8 on i32: NA

• rec8 on u32: NA

• rec8 on i64: NA

• rec8 on u64: NA

• rec8 on f16: T

• rec8 on f32: vrecpeq_f32

• rec8 on f64: vrecpeq_f64

round_to_even

• round_to_even on i8: NOOP

• round_to_even on u8: NOOP

• round_to_even on i16: NOOP

• round_to_even on u16: NOOP

• round_to_even on i32: NOOP

• round_to_even on u32: NOOP

• round_to_even on i64: NOOP

• round_to_even on u64: NOOP

• round_to_even on f16: T

• round_to_even on f32: vrndnq_f32

• round_to_even on f64: vrndnq_f64

rsqrt11

• rsqrt11 on i8: NA

• rsqrt11 on u8: NA

• rsqrt11 on i16: NA

• rsqrt11 on u16: NA

• rsqrt11 on i32: NA

• rsqrt11 on u32: NA

• rsqrt11 on i64: NA

• rsqrt11 on u64: NA

• rsqrt11 on f16: T

• rsqrt11 on f32: T

• rsqrt11 on f64: T

rsqrt8

• rsqrt8 on i8: NA

• rsqrt8 on u8: NA

• rsqrt8 on i16: NA

• rsqrt8 on u16: NA

• rsqrt8 on i32: NA

• rsqrt8 on u32: NA

• rsqrt8 on i64: NA

• rsqrt8 on u64: NA

• rsqrt8 on f16: T

• rsqrt8 on f32: vrsqrteq_f32

• rsqrt8 on f64: vrsqrteq_f64

scatter

• scatter on i8: NA

• scatter on u8: NA

• scatter on i16: T

• scatter on u16: T

• scatter on i32: T

• scatter on u32: T

• scatter on i64: T

• scatter on u64: T

• scatter on f16: T

• scatter on f32: T

• scatter on f64: T

scatter_linear

• scatter_linear on i8: T

• scatter_linear on u8: T

• scatter_linear on i16: T

• scatter_linear on u16: T

• scatter_linear on i32: T

• scatter_linear on u32: T

• scatter_linear on i64: T

• scatter_linear on u64: T

• scatter_linear on f16: E

• scatter_linear on f32: T

• scatter_linear on f64: T

set1

• set1 on i8: vdupq_n_s8

• set1 on u8: vdupq_n_u8

• set1 on i16: vdupq_n_s16

• set1 on u16: vdupq_n_u16

• set1 on i32: vdupq_n_s32

• set1 on u32: vdupq_n_u32

• set1 on i64: vdupq_n_s64

• set1 on u64: vdupq_n_u64

• set1 on f16: T

• set1 on f32: vdupq_n_f32

• set1 on f64: vdupq_n_f64

set1l

• set1l on i8: T

• set1l on u8: T

• set1l on i16: T

• set1l on u16: T

• set1l on i32: T

• set1l on u32: T

• set1l on i64: T

• set1l on u64: T

• set1l on f16: T

• set1l on f32: T

• set1l on f64: T

shl

• shl on i8: T

• shl on u8: T

• shl on i16: T

• shl on u16: T

• shl on i32: T

• shl on u32: T

• shl on i64: T

• shl on u64: T

• shl on f16: NA

• shl on f32: NA

• shl on f64: NA

shr

• shr on i8: T

• shr on u8: T

• shr on i16: T

• shr on u16: T

• shr on i32: T

• shr on u32: T

• shr on i64: T

• shr on u64: T

• shr on f16: NA

• shr on f32: NA

• shr on f64: NA

shra

• shra on i8: T

• shra on u8: T

• shra on i16: T

• shra on u16: T

• shra on i32: T

• shra on u32: T

• shra on i64: T

• shra on u64: T

• shra on f16: NA

• shra on f32: NA

• shra on f64: NA

sqrt

• sqrt on i8: NA

• sqrt on u8: NA

• sqrt on i16: NA

• sqrt on u16: NA

• sqrt on i32: NA

• sqrt on u32: NA

• sqrt on i64: NA

• sqrt on u64: NA

• sqrt on f16: T

• sqrt on f32: vsqrtq_f32

• sqrt on f64: vsqrtq_f64

store2a

• store2a on i8: T

• store2a on u8: T

• store2a on i16: T

• store2a on u16: T

• store2a on i32: T

• store2a on u32: T

• store2a on i64: T

• store2a on u64: T

• store2a on f16: T

• store2a on f32: T

• store2a on f64: T

store2u

• store2u on i8: T

• store2u on u8: T

• store2u on i16: T

• store2u on u16: T

• store2u on i32: T

• store2u on u32: T

• store2u on i64: T

• store2u on u64: T

• store2u on f16: T

• store2u on f32: T

• store2u on f64: T

store3a

• store3a on i8: T

• store3a on u8: T

• store3a on i16: T

• store3a on u16: T

• store3a on i32: T

• store3a on u32: T

• store3a on i64: T

• store3a on u64: T

• store3a on f16: T

• store3a on f32: T

• store3a on f64: T

store3u

• store3u on i8: T

• store3u on u8: T

• store3u on i16: T

• store3u on u16: T

• store3u on i32: T

• store3u on u32: T

• store3u on i64: T

• store3u on u64: T

• store3u on f16: T

• store3u on f32: T

• store3u on f64: T

store4a

• store4a on i8: T

• store4a on u8: T

• store4a on i16: T

• store4a on u16: T

• store4a on i32: T

• store4a on u32: T

• store4a on i64: T

• store4a on u64: T

• store4a on f16: T

• store4a on f32: T

• store4a on f64: T

store4u

• store4u on i8: T

• store4u on u8: T

• store4u on i16: T

• store4u on u16: T

• store4u on i32: T

• store4u on u32: T

• store4u on i64: T

• store4u on u64: T

• store4u on f16: T

• store4u on f32: T

• store4u on f64: T

storea

• storea on i8: vst1q_s8

• storea on u8: vst1q_u8

• storea on i16: vst1q_s16

• storea on u16: vst1q_u16

• storea on i32: vst1q_s32

• storea on u32: vst1q_u32

• storea on i64: vst1q_s64

• storea on u64: vst1q_u64

• storea on f16: T

• storea on f32: vst1q_f32

• storea on f64: vst1q_f64

storela

• storela on i8: T

• storela on u8: T

• storela on i16: T

• storela on u16: T

• storela on i32: T

• storela on u32: T

• storela on i64: T

• storela on u64: T

• storela on f16: T

• storela on f32: T

• storela on f64: T

storelu

• storelu on i8: T

• storelu on u8: T

• storelu on i16: T

• storelu on u16: T

• storelu on i32: T

• storelu on u32: T

• storelu on i64: T

• storelu on u64: T

• storelu on f16: T

• storelu on f32: T

• storelu on f64: T

storeu

• storeu on i8: vst1q_s8

• storeu on u8: vst1q_u8

• storeu on i16: vst1q_s16

• storeu on u16: vst1q_u16

• storeu on i32: vst1q_s32

• storeu on u32: vst1q_u32

• storeu on i64: vst1q_s64

• storeu on u64: vst1q_u64

• storeu on f16: T

• storeu on f32: vst1q_f32

• storeu on f64: vst1q_f64

sub

• sub on i8: vsubq_s8

• sub on u8: vsubq_u8

• sub on i16: vsubq_s16

• sub on u16: vsubq_u16

• sub on i32: vsubq_s32

• sub on u32: vsubq_u32

• sub on i64: vsubq_s64

• sub on u64: vsubq_u64

• sub on f16: T

• sub on f32: vsubq_f32

• sub on f64: vsubq_f64

subs

• subs on i8: vqsubq_s8

• subs on u8: vqsubq_u8

• subs on i16: vqsubq_s16

• subs on u16: vqsubq_u16

• subs on i32: vqsubq_s32

• subs on u32: vqsubq_u32

• subs on i64: vqsubq_s64

• subs on u64: vqsubq_u64

• subs on f16: T

• subs on f32: T

• subs on f64: T

to_logical

• to_logical on i8: T

• to_logical on u8: T

• to_logical on i16: T

• to_logical on u16: T

• to_logical on i32: T

• to_logical on u32: T

• to_logical on i64: T

• to_logical on u64: T

• to_logical on f16: T

• to_logical on f32: T

• to_logical on f64: T

to_mask

• to_mask on i8: NOOP

• to_mask on u8: NOOP

• to_mask on i16: NOOP

• to_mask on u16: NOOP

• to_mask on i32: NOOP

• to_mask on u32: NOOP

• to_mask on i64: NOOP

• to_mask on u64: NOOP

• to_mask on f16: T

• to_mask on f32: NOOP

• to_mask on f64: NOOP

trunc

• trunc on i8: NOOP

• trunc on u8: NOOP

• trunc on i16: NOOP

• trunc on u16: NOOP

• trunc on i32: NOOP

• trunc on u32: NOOP

• trunc on i64: NOOP

• trunc on u64: NOOP

• trunc on f16: T

• trunc on f32: vrndq_f32

• trunc on f64: vrndq_f64

unzip

• unzip on i8: T

• unzip on u8: T

• unzip on i16: T

• unzip on u16: T

• unzip on i32: T

• unzip on u32: T

• unzip on i64: T

• unzip on u64: T

• unzip on f16: T

• unzip on f32: T

• unzip on f64: T

unziphi

• unziphi on i8: vuzp2q_s8

• unziphi on u8: vuzp2q_u8

• unziphi on i16: vuzp2q_s16

• unziphi on u16: vuzp2q_u16

• unziphi on i32: vuzp2q_s32

• unziphi on u32: vuzp2q_u32

• unziphi on i64: vuzp2q_s64

• unziphi on u64: vuzp2q_u64

• unziphi on f16: T

• unziphi on f32: vuzp2q_f32

• unziphi on f64: vuzp2q_f64

unziplo

• unziplo on i8: vuzp1q_s8

• unziplo on u8: vuzp1q_u8

• unziplo on i16: vuzp1q_s16

• unziplo on u16: vuzp1q_u16

• unziplo on i32: vuzp1q_s32

• unziplo on u32: vuzp1q_u32

• unziplo on i64: vuzp1q_s64

• unziplo on u64: vuzp1q_u64

• unziplo on f16: T

• unziplo on f32: vuzp1q_f32

• unziplo on f64: vuzp1q_f64

xorb

• xorb on i8: veorq_s8

• xorb on u8: veorq_u8

• xorb on i16: veorq_s16

• xorb on u16: veorq_u16

• xorb on i32: veorq_s32

• xorb on u32: veorq_u32

• xorb on i64: veorq_s64

• xorb on u64: veorq_u64

• xorb on f16: T

• xorb on f32: veorq_u32

• xorb on f64: veorq_u64

xorl

• xorl on i8: veorq_u8

• xorl on u8: veorq_u8

• xorl on i16: veorq_u16

• xorl on u16: veorq_u16

• xorl on i32: veorq_u32

• xorl on u32: veorq_u32

• xorl on i64: veorq_u64

• xorl on u64: veorq_u64

• xorl on f16: T

• xorl on f32: veorq_u32

• xorl on f64: veorq_u64

zip

• zip on i8: T

• zip on u8: T

• zip on i16: T

• zip on u16: T

• zip on i32: T

• zip on u32: T

• zip on i64: T

• zip on u64: T

• zip on f16: T

• zip on f32: T

• zip on f64: T

ziphi

• ziphi on i8: vzip2q_s8

• ziphi on u8: vzip2q_u8

• ziphi on i16: vzip2q_s16

• ziphi on u16: vzip2q_u16

• ziphi on i32: vzip2q_s32

• ziphi on u32: vzip2q_u32

• ziphi on i64: vzip2q_s64

• ziphi on u64: vzip2q_u64

• ziphi on f16: T

• ziphi on f32: vzip2q_f32

• ziphi on f64: vzip2q_f64

ziplo

• ziplo on i8: vzip1q_s8

• ziplo on u8: vzip1q_u8

• ziplo on i16: vzip1q_s16

• ziplo on u16: vzip1q_u16

• ziplo on i32: vzip1q_s32

• ziplo on u32: vzip1q_u32

• ziplo on i64: vzip1q_s64

• ziplo on u64: vzip1q_u64

• ziplo on f16: T

• ziplo on f32: vzip1q_f32

• ziplo on f64: vzip1q_f64

cvt

• cvt from i8 to i8: NOOP

• cvt from i8 to u8: NOOP

• cvt from u8 to i8: NOOP

• cvt from u8 to u8: NOOP

• cvt from i16 to i16: NOOP

• cvt from i16 to u16: NOOP

• cvt from i16 to f16: T

• cvt from u16 to i16: NOOP

• cvt from u16 to u16: NOOP

• cvt from u16 to f16: T

• cvt from i32 to i32: NOOP

• cvt from i32 to u32: NOOP

• cvt from i32 to f32: vcvtq_f32_s32

• cvt from u32 to i32: NOOP

• cvt from u32 to u32: NOOP

• cvt from u32 to f32: vcvtq_f32_u32

• cvt from i64 to i64: NOOP

• cvt from i64 to u64: NOOP

• cvt from i64 to f64: vcvtq_f64_s64

• cvt from u64 to i64: NOOP

• cvt from u64 to u64: NOOP

• cvt from u64 to f64: vcvtq_f64_u64

• cvt from f16 to i16: T

• cvt from f16 to u16: T

• cvt from f16 to f16: NOOP

• cvt from f32 to i32: vcvtq_s32_f32

• cvt from f32 to u32: vcvtq_u32_f32

• cvt from f32 to f32: NOOP

• cvt from f64 to i64: vcvtq_s64_f64

• cvt from f64 to u64: vcvtq_u64_f64

• cvt from f64 to f64: NOOP

reinterpret

• reinterpret from i8 to i8: NOOP

• reinterpret from i8 to u8: NOOP

• reinterpret from u8 to i8: NOOP

• reinterpret from u8 to u8: NOOP

• reinterpret from i16 to i16: NOOP

• reinterpret from i16 to u16: NOOP

• reinterpret from i16 to f16: T

• reinterpret from u16 to i16: NOOP

• reinterpret from u16 to u16: NOOP

• reinterpret from u16 to f16: T

• reinterpret from i32 to i32: NOOP

• reinterpret from i32 to u32: NOOP

• reinterpret from i32 to f32: NOOP

• reinterpret from u32 to i32: NOOP

• reinterpret from u32 to u32: NOOP

• reinterpret from u32 to f32: NOOP

• reinterpret from i64 to i64: NOOP

• reinterpret from i64 to u64: NOOP

• reinterpret from i64 to f64: NOOP

• reinterpret from u64 to i64: NOOP

• reinterpret from u64 to u64: NOOP

• reinterpret from u64 to f64: NOOP

• reinterpret from f16 to i16: T

• reinterpret from f16 to u16: T

• reinterpret from f16 to f16: NOOP

• reinterpret from f32 to i32: NOOP

• reinterpret from f32 to u32: NOOP

• reinterpret from f32 to f32: NOOP

• reinterpret from f64 to i64: NOOP

• reinterpret from f64 to u64: NOOP

• reinterpret from f64 to f64: NOOP

reinterpretl

• reinterpretl from i8 to i8: NOOP

• reinterpretl from i8 to u8: NOOP

• reinterpretl from u8 to i8: NOOP

• reinterpretl from u8 to u8: NOOP

• reinterpretl from i16 to i16: NOOP

• reinterpretl from i16 to u16: NOOP

• reinterpretl from i16 to f16: T

• reinterpretl from u16 to i16: NOOP

• reinterpretl from u16 to u16: NOOP

• reinterpretl from u16 to f16: T

• reinterpretl from i32 to i32: NOOP

• reinterpretl from i32 to u32: NOOP

• reinterpretl from i32 to f32: NOOP

• reinterpretl from u32 to i32: NOOP

• reinterpretl from u32 to u32: NOOP

• reinterpretl from u32 to f32: NOOP

• reinterpretl from i64 to i64: NOOP

• reinterpretl from i64 to u64: NOOP

• reinterpretl from i64 to f64: NOOP

• reinterpretl from u64 to i64: NOOP

• reinterpretl from u64 to u64: NOOP

• reinterpretl from u64 to f64: NOOP

• reinterpretl from f16 to i16: T

• reinterpretl from f16 to u16: T

• reinterpretl from f16 to f16: NOOP

• reinterpretl from f32 to i32: NOOP

• reinterpretl from f32 to u32: NOOP

• reinterpretl from f32 to f32: NOOP

• reinterpretl from f64 to i64: NOOP

• reinterpretl from f64 to u64: NOOP

• reinterpretl from f64 to f64: NOOP

upcvt

• upcvt from i8 to i16: T

• upcvt from i8 to u16: T

• upcvt from i8 to f16: T

• upcvt from u8 to i16: T

• upcvt from u8 to u16: T

• upcvt from u8 to f16: T

• upcvt from i16 to i32: T

• upcvt from i16 to u32: T

• upcvt from i16 to f32: T

• upcvt from u16 to i32: T

• upcvt from u16 to u32: T

• upcvt from u16 to f32: T

• upcvt from i32 to i64: T

• upcvt from i32 to u64: T

• upcvt from i32 to f64: T

• upcvt from u32 to i64: T

• upcvt from u32 to u64: T

• upcvt from u32 to f64: T

• upcvt from f16 to i32: T

• upcvt from f16 to u32: T

• upcvt from f16 to f32: T

• upcvt from f32 to i64: T

• upcvt from f32 to u64: T

• upcvt from f32 to f64: T

downcvt

• downcvt from i16 to i8: T

• downcvt from i16 to u8: T

• downcvt from u16 to i8: T

• downcvt from u16 to u8: T

• downcvt from i32 to i16: T

• downcvt from i32 to u16: T

• downcvt from i32 to f16: T

• downcvt from u32 to i16: T

• downcvt from u32 to u16: T

• downcvt from u32 to f16: T

• downcvt from i64 to i32: T

• downcvt from i64 to u32: T

• downcvt from i64 to f32: T

• downcvt from u64 to i32: T

• downcvt from u64 to u32: T

• downcvt from u64 to f32: T

• downcvt from f16 to i8: T

• downcvt from f16 to u8: T

• downcvt from f32 to i16: T

• downcvt from f32 to u16: T

• downcvt from f32 to f16: T

• downcvt from f64 to i32: T

• downcvt from f64 to u32: T

• downcvt from f64 to f32: T