!sum
- go: SetElem
  asm: "VPINSR[BWDQ]"
  in:
  - &t
    class: vreg
    base: $b
  - class: greg
    base: $b
    lanes: 1 # Scalar, darn it!
  - &imm
    class: immediate
    immOffset: 0
    name: index
  out:
  - *t

- go: SetElem
  asm: "VPINSR[DQ]"
  in:
  - &t
    class: vreg
    base: int
    OverwriteBase: float
  - class: greg
    base: int
    OverwriteBase: float
    lanes: 1 # Scalar, darn it!
  - &imm
    class: immediate
    immOffset: 0
    name: index
  out:
  - *t

- go: GetElem
  asm: "VPEXTR[BWDQ]"
  in:
  - class: vreg
    base: $b
    elemBits: $e
  - *imm
  out:
  - class: greg
    base: $b
    bits: $e

- go: GetElem
  asm: "VPEXTR[DQ]"
  in:
  - class: vreg
    base: int
    elemBits: $e
    OverwriteBase: float
  - *imm
  out:
  - class: greg
    base: int
    bits: $e
    OverwriteBase: float

- go: "SetHi|SetLo"
  regexpTag: "move"
  asm: "VINSERTI128|VINSERTI64X4"
  inVariant: []
  in:
  - &i8x2N
    class: vreg
    base: $t
    OverwriteElementBits: 8
  - &i8xN
    class: vreg
    base: $t
    OverwriteElementBits: 8
  - &imm01 # This immediate should be only 0 or 1
    class: immediate
    const: 0 # place holder
    name: index
  out:
  - *i8x2N

- go: "GetHi|GetLo"
  asm: "VEXTRACTI128|VEXTRACTI64X4"
  regexpTag: "move"
  inVariant: []
  in:
  - *i8x2N
  - *imm01
  out:
  - *i8xN

- go: "SetHi|SetLo"
  asm: "VINSERTI128|VINSERTI64X4"
  regexpTag: "move"
  inVariant: []
  in:
  - &i16x2N
    class: vreg
    base: $t
    OverwriteElementBits: 16
  - &i16xN
    class: vreg
    base: $t
    OverwriteElementBits: 16
  - *imm01
  out:
  - *i16x2N

- go: "GetHi|GetLo"
  regexpTag: "move"
  asm: "VEXTRACTI128|VEXTRACTI64X4"
  inVariant: []
  in:
  - *i16x2N
  - *imm01
  out:
  - *i16xN

- go: "SetHi|SetLo"
  regexpTag: "move"
  asm: "VINSERTI128|VINSERTI64X4"
  inVariant: []
  in:
  - &i32x2N
    class: vreg
    base: $t
    OverwriteElementBits: 32
  - &i32xN
    class: vreg
    base: $t
    OverwriteElementBits: 32
  - *imm01
  out:
  - *i32x2N

- go: "GetHi|GetLo"
  regexpTag: "move"
  asm: "VEXTRACTI128|VEXTRACTI64X4"
  inVariant: []
  in:
  - *i32x2N
  - *imm01
  out:
  - *i32xN

- go: "SetHi|SetLo"
  regexpTag: "move"
  asm: "VINSERTI128|VINSERTI64X4"
  inVariant: []
  in:
  - &i64x2N
    class: vreg
    base: $t
    OverwriteElementBits: 64
  - &i64xN
    class: vreg
    base: $t
    OverwriteElementBits: 64
  - *imm01
  out:
  - *i64x2N

- go: "GetHi|GetLo"
  regexpTag: "move"
  asm: "VEXTRACTI128|VEXTRACTI64X4"
  inVariant: []
  in:
  - *i64x2N
  - *imm01
  out:
  - *i64xN

- go: "SetHi|SetLo"
  regexpTag: "move"
  asm: "VINSERTF128|VINSERTF64X4"
  inVariant: []
  in:
  - &f32x2N
    class: vreg
    base: $t
    OverwriteElementBits: 32
  - &f32xN
    class: vreg
    base: $t
    OverwriteElementBits: 32
  - *imm01
  out:
  - *f32x2N

- go: "GetHi|GetLo"
  regexpTag: "move"
  asm: "VEXTRACTF128|VEXTRACTF64X4"
  inVariant: []
  in:
  - *f32x2N
  - *imm01
  out:
  - *f32xN

- go: "SetHi|SetLo"
  regexpTag: "move"
  asm: "VINSERTF128|VINSERTF64X4"
  inVariant: []
  in:
  - &f64x2N
    class: vreg
    base: $t
    OverwriteElementBits: 64
  - &f64xN
    class: vreg
    base: $t
    OverwriteElementBits: 64
  - *imm01
  out:
  - *f64x2N

- go: "GetHi|GetLo"
  regexpTag: "move"
  asm: "VEXTRACTF128|VEXTRACTF64X4"
  inVariant: []
  in:
  - *f64x2N
  - *imm01
  out:
  - *f64xN

- go: Permute
  asm: "VPERMQ|VPERMPD"
  addDoc: !string |-
    // The low 2 bits (values 0-3) of each element of indices is used
  operandOrder: "21Type1"
  in:
  - &anyindices
    go: $t
    name: indices
    overwriteBase: uint
  - &any4
    go: $t
    lanes: 4
  out:
  - &any
    go: $t

- go: Permute
  asm: "VPERM[WDQ]|VPERMP[SD]"
  addDoc: !string |-
    // The low 3 bits (values 0-7) of each element of indices is used
  operandOrder: "21Type1"
  in:
  - *anyindices
  - &any8
    go: $t
    lanes: 8
  out:
  - *any

- go: Permute
  asm: "VPERM[BWD]|VPERMPS"
  addDoc: !string |-
    // The low 4 bits (values 0-15) of each element of indices is used
  operandOrder: "21Type1"
  in:
  - *anyindices
  - &any16
    go: $t
    lanes: 16
  out:
  - *any

- go: Permute
  asm: "VPERM[BW]"
  addDoc: !string |-
    // The low 5 bits (values 0-31) of each element of indices is used
  operandOrder: "21Type1"
  in:
  - *anyindices
  - &any32
    go: $t
    lanes: 32
  out:
  - *any

- go: Permute
  asm: "VPERMB"
  addDoc: !string |-
    // The low 6 bits (values 0-63) of each element of indices is used
  operandOrder: "21Type1"
  in:
  - *anyindices
  - &any64
    go: $t
    lanes: 64
  out:
  - *any

- go: ConcatPermute
  asm: "VPERMI2[BWDQ]|VPERMI2P[SD]"
  # Because we are overwriting the receiver's type, we
  # have to move the receiver to be a parameter so that
  # we can have no duplication.
  operandOrder: "231Type1"
  in:
  - *anyindices # result in arg 0
  - *any
  - *any
  out:
  - *any

- go: Compress
  asm: "VPCOMPRESS[BWDQ]|VCOMPRESSP[SD]"
  in:
    # The mask in Compress is a control mask rather than a write mask, so it's not optional.
  - class: mask
  - *any
  out:
  - *any

# For now a non-public method because
# (1) [OverwriteClass] must be set together with [OverwriteBase]
# (2) "simdgen does not support [OverwriteClass] in inputs".
# That means the signature is wrong.
- go: blend
  asm: VPBLENDVB
  zeroing: false
  in:
  - &v
    go: $t
    class: vreg
    base: int
  - *v
  -
    class: vreg
    base: int
    name: mask
  out:
  - *v

# For AVX512
- go: blend
  asm: VPBLENDM[BWDQ]
  zeroing: false
  in:
  - &v
    go: $t
    bits: 512
    class: vreg
    base: int
  - *v
  inVariant:
  -
    class: mask
  out:
  - *v

  # For AVX512
- go: move
  asm: VMOVDQU(8|16|32|64)
  zeroing: true
  in:
  - &v
    go: $t
    class: vreg
    base: int|uint
  inVariant:
  -
    class: mask
  out:
  - *v

- go: Expand
  asm: "VPEXPAND[BWDQ]|VEXPANDP[SD]"
  in:
    # The mask in Expand is a control mask rather than a write mask, so it's not optional.
  - class: mask
  - *any
  out:
  - *any

- go: Broadcast128
  asm: VPBROADCAST[BWDQ]
  in:
  - class: vreg
    bits: 128
    elemBits: $e
    base: $b
  out:
  - class: vreg
    bits: 128
    elemBits: $e
    base: $b

# weirdly, this one case on AVX2 is memory-operand-only
- go: Broadcast128
  asm: VPBROADCASTQ
  in:
  - class: vreg
    bits: 128
    elemBits: 64
    base: int
    OverwriteBase: float
  out:
  - class: vreg
    bits: 128
    elemBits: 64
    base: int
    OverwriteBase: float

- go: Broadcast256
  asm: VPBROADCAST[BWDQ]
  in:
  - class: vreg
    bits: 128
    elemBits: $e
    base: $b
  out:
  - class: vreg
    bits: 256
    elemBits: $e
    base: $b

- go: Broadcast512
  asm: VPBROADCAST[BWDQ]
  in:
  - class: vreg
    bits: 128
    elemBits: $e
    base: $b
  out:
  - class: vreg
    bits: 512
    elemBits: $e
    base: $b

- go: Broadcast128
  asm: VBROADCASTS[SD]
  in:
  - class: vreg
    bits: 128
    elemBits: $e
    base: $b
  out:
  - class: vreg
    bits: 128
    elemBits: $e
    base: $b

- go: Broadcast256
  asm: VBROADCASTS[SD]
  in:
  - class: vreg
    bits: 128
    elemBits: $e
    base: $b
  out:
  - class: vreg
    bits: 256
    elemBits: $e
    base: $b

- go: Broadcast512
  asm: VBROADCASTS[SD]
  in:
  - class: vreg
    bits: 128
    elemBits: $e
    base: $b
  out:
  - class: vreg
    bits: 512
    elemBits: $e
    base: $b

# VPSHUFB for 128-bit byte shuffles will be picked with higher priority than VPERMB, given its lower CPU feature requirement. (It's AVX)
- go: PermuteOrZero
  asm: VPSHUFB
  addDoc: !string |-
    // The lower four bits of each byte-sized index in indices select an element from x,
    // unless the index's sign bit is set in which case zero is used instead.
  in:
  - &128any
    bits: 128
    go: $t
  - bits: 128
    name: indices
    base: int # always signed
  out:
  - *128any

- go: PermuteOrZeroGrouped
  asm: VPSHUFB
  addDoc: !string |-
    // result = {x_group0[indices[0]], x_group0[indices[1]], ..., x_group1[indices[16]], x_group1[indices[17]], ...}
    // The lower four bits of each byte-sized index in indices select an element from its corresponding group in x,
    // unless the index's sign bit is set in which case zero is used instead.
    // Each group is of size 128-bit.
  in:
  - &256Or512any
    bits: "256|512"
    go: $t
  - bits: "256|512"
    base: int
    name: indices
  out:
  - *256Or512any

- go: permuteScalars
  asm: VPSHUFD
  addDoc: !string |-
    // result = {x[indices[0:2]], x[indices[2:4]], x[indices[4:6]], x[indices[6:8]]}
    // Indices is four 2-bit values packed into a byte, thus indices[0:2] is the first index.
  in:
  - *128any
  - class: immediate
    immOffset: 0
    name: indices
  hideMaskMethods: true
  out:
  - *128any

- go: permuteScalarsGrouped
  asm: VPSHUFD
  addDoc: !string |-
    // result = {x_group0[indices[0:2]], x_group0[indices[2:4]], x_group0[indices[4:6]], x_group0[indices[6:8]], x_group1[indices[0:2]], ...}
    // Indices is four 2-bit values packed into a byte, thus indices[0:2] is the first index.
    // Each group is of size 128-bit.
  in:
  - *256Or512any
  - class: immediate
    immOffset: 0
    name: indices
  hideMaskMethods: true
  out:
  - *256Or512any

- go: permuteScalarsLo
  asm: VPSHUFLW
  addDoc: !string |-
    // result = {x[indices[0:2]], x[indices[2:4]], x[indices[4:6]], x[indices[6:8]], x[4], x[5], x[6], x[7]}
    // Indices is four 2-bit values packed into a byte, thus indices[0:2] is the first index.
  in:
    - &128lanes8
      bits: 128
      go: $t
      elemBits: 16
    - class: immediate
      immOffset: 0
      name: indices
  hideMaskMethods: true
  out:
    - *128lanes8

- go: permuteScalarsLoGrouped
  asm: VPSHUFLW
  addDoc: !string |-
    //
    //   result = {x_group0[indices[0:2]], x_group0[indices[2:4]], x_group0[indices[4:6]], x_group0[indices[6:8]], x[4], x[5], x[6], x[7],
    //    x_group1[indices[0:2]], ...}
    //
    // Indices is four 2-bit values packed into a byte, thus indices[0:2] is the first index.
    // Each group is of size 128-bit.
  in:
  - &256Or512lanes8
    bits: "256|512"
    go: $t
    elemBits: 16
  - class: immediate
    immOffset: 0
    name: indices
  hideMaskMethods: true
  out:
  - *256Or512lanes8

- go: permuteScalarsHi
  asm: VPSHUFHW
  addDoc: !string |-
    // result = {x[0], x[1], x[2], x[3], x[indices[0:2]+4], x[indices[2:4]+4], x[indices[4:6]+4], x[indices[6:8]+4]}
    // Indices is four 2-bit values packed into a byte, thus indices[0:2] is the first index.
  in:
  - *128lanes8
  - class: immediate
    immOffset: 0
    name: indices
  hideMaskMethods: true
  out:
  - *128lanes8

- go: permuteScalarsHiGrouped
  asm: VPSHUFHW
  addDoc: !string |-
    // result =
    //
    //   {x_group0[0], x_group0[1], x_group0[2], x_group0[3], x_group0[indices[0:2]+4], x_group0[indices[2:4]+4], x_group0[indices[4:6]+4], x_group0[indices[6:8]+4],
    //    x_group1[0], x_group1[1], x_group1[2], x_group1[3], x_group1[indices[0:2]+4], ...}
    //
    // Indices is four 2-bit values packed into a byte, thus indices[0:2] is the first index.
    // Each group is of size 128-bit.
  in:
  - *256Or512lanes8
  - class: immediate
    immOffset: 0
    name: indices
  hideMaskMethods: true
  out:
  - *256Or512lanes8

- go: InterleaveHi
  asm: VPUNPCKH(QDQ|DQ|WD|WB)
  in:
  - *128any
  - *128any
  inVariant: []
  out:
  - *128any

- go: InterleaveLo
  asm: VPUNPCKL(QDQ|DQ|WD|WB)
  in:
  - *128any
  - *128any
  inVariant: []
  out:
  - *128any

- go: InterleaveHiGrouped
  asm: VPUNPCKH(QDQ|DQ|WD|WB)
  in:
  - *256Or512any
  - *256Or512any
  inVariant: []
  out:
  - *256Or512any

- go: InterleaveLoGrouped
  asm: VPUNPCKL(QDQ|DQ|WD|WB)
  in:
  - *256Or512any
  - *256Or512any
  inVariant: []
  out:
  - *256Or512any

# These are all described separately to carry the name of the constant parameter

- go: concatSelectedConstant
  asm: VSHUFPS
  width: 32
  in:
  - &v
    go: $t
    class: vreg
    base: float
    bits: 128
  - *v
  - class: immediate
    immOffset: 0
    name: h1h0l1l0
  inVariant: []
  out:
  - *v

- go: concatSelectedConstant
  asm: VSHUFPS
  in:
  - &v
    go: $t
    class: vreg
    base: float
    bits: 128
    OverwriteBase: int
  - *v
  - class: immediate
    immOffset: 0
    name: h1h0l1l0
  inVariant: []
  out:
  - *v

- go: concatSelectedConstant
  asm: VSHUFPS
  in:
  - &v
    go: $t
    class: vreg
    base: float
    bits: 128
    OverwriteBase: uint
  - *v
  - class: immediate
    immOffset: 0
    name: h1h0l1l0
  inVariant: []
  out:
  - *v


- go: concatSelectedConstantGrouped
  asm: VSHUFPS
  in:
  - &v
    go: $t
    class: vreg
    base: float
    bits: "256|512"
  - *v
  - class: immediate
    immOffset: 0
    name: h1h0l1l0
  inVariant: []
  out:
  - *v

- go: concatSelectedConstantGrouped
  asm: VSHUFPS
  in:
  - &v
    go: $t
    class: vreg
    base: float
    bits: "256|512"
    OverwriteBase: int
  - *v
  - class: immediate
    immOffset: 0
    name: h1h0l1l0
  inVariant: []
  out:
  - *v

- go: concatSelectedConstantGrouped
  asm: VSHUFPS
  in:
  - &v
    go: $t
    class: vreg
    base: float
    bits: "256|512"
    OverwriteBase: uint
  - *v
  - class: immediate
    immOffset: 0
    name: h1h0l1l0
  inVariant: []
  out:
  - *v


  # 64 bit versions

- go: concatSelectedConstant
  asm: VSHUFPD
  in:
  - &v
    go: $t
    class: vreg
    base: float
    bits: 128
  - *v
  - class: immediate
    immOffset: 0
    name: hilo
  inVariant: []
  out:
  - *v

- go: concatSelectedConstant
  asm: VSHUFPD
  in:
  - &v
    go: $t
    class: vreg
    base: float
    bits: 128
    OverwriteBase: int
  - *v
  - class: immediate
    immOffset: 0
    name: hilo
  inVariant: []
  out:
  - *v

- go: concatSelectedConstant
  asm: VSHUFPD
  in:
  - &v
    go: $t
    class: vreg
    base: float
    bits: 128
    OverwriteBase: uint
  - *v
  - class: immediate
    immOffset: 0
    name: hilo
  inVariant: []
  out:
  - *v

- go: concatSelectedConstantGrouped
  asm: VSHUFPD
  in:
  - &v
    go: $t
    class: vreg
    base: float
    bits: "256|512"
  - *v
  - class: immediate
    immOffset: 0
    name: hilos
  inVariant: []
  out:
  - *v

- go: concatSelectedConstantGrouped
  asm: VSHUFPD
  in:
  - &v
    go: $t
    class: vreg
    base: float
    bits: "256|512"
    OverwriteBase: int
  - *v
  - class: immediate
    immOffset: 0
    name: hilos
  inVariant: []
  out:
  - *v

- go: concatSelectedConstantGrouped
  asm: VSHUFPD
  in:
  - &v
    go: $t
    class: vreg
    base: float
    bits: "256|512"
    OverwriteBase: uint
  - *v
  - class: immediate
    immOffset: 0
    name: hilos
  inVariant: []
  out:
  - *v

- go: Select128FromPair
  asm: VPERM2F128
  operandOrder: II
  addDoc: !string |-
    // For example,
    //
    //   {40, 41, 50, 51}.NAME(3, 0, {60, 61, 70, 71})
    //
    // returns {70, 71, 40, 41}.
  in:
  - &v
    go: $t
    class: vreg
    base: float
    bits: 256
  - *v
  - class: immediate
    immOffset: 0
    name: "lo, hi"
  inVariant: []
  out:
  - *v

- go: Select128FromPair
  asm: VPERM2F128
  operandOrder: II
  addDoc: !string |-
    // For example,
    //
    //   {40, 41, 42, 43, 50, 51, 52, 53}.NAME(3, 0, {60, 61, 62, 63, 70, 71, 72, 73})
    //
    // returns {70, 71, 72, 73, 40, 41, 42, 43}.
  in:
  - &v
    go: $t
    class: vreg
    base: float
    bits: 256
    OverwriteElementBits: 32
  - *v
  - class: immediate
    immOffset: 0
    name: "lo, hi"
  inVariant: []
  out:
  - *v

- go: Select128FromPair
  asm: VPERM2I128
  operandOrder: II
  addDoc: !string |-
    // For example,
    //
    //   {40, 41, 50, 51}.NAME(3, 0, {60, 61, 70, 71})
    //
    // returns {70, 71, 40, 41}.
  in:
  - &v
    go: $t
    class: vreg
    base: int|uint
    bits: 256
    OverwriteElementBits: 64
  - *v
  - class: immediate
    immOffset: 0
    name: "lo, hi"
  inVariant: []
  out:
  - *v

- go: Select128FromPair
  asm: VPERM2I128
  operandOrder: II
  addDoc: !string |-
    // For example,
    //
    //   {40, 41, 42, 43, 50, 51, 52, 53}.NAME(3, 0, {60, 61, 62, 63, 70, 71, 72, 73})
    //
    // returns {70, 71, 72, 73, 40, 41, 42, 43}.
  in:
  - &v
    go: $t
    class: vreg
    base: int|uint
    bits: 256
    OverwriteElementBits: 32
  - *v
  - class: immediate
    immOffset: 0
    name: "lo, hi"
  inVariant: []
  out:
  - *v

- go: Select128FromPair
  asm: VPERM2I128
  operandOrder: II
  addDoc: !string |-
    // For example,
    //
    //   {40, 41, 42, 43, 44, 45, 46, 47, 50, 51, 52, 53, 54, 55, 56, 57}.NAME(3, 0,
    //    {60, 61, 62, 63, 64, 65, 66, 67, 70, 71, 72, 73, 74, 75, 76, 77})
    //
    // returns {70, 71, 72, 73, 74, 75, 76, 77, 40, 41, 42, 43, 44, 45, 46, 47}.
  in:
  - &v
    go: $t
    class: vreg
    base: int|uint
    bits: 256
    OverwriteElementBits: 16
  - *v
  - class: immediate
    immOffset: 0
    name: "lo, hi"
  inVariant: []
  out:
  - *v

- go: Select128FromPair
  asm: VPERM2I128
  operandOrder: II
  addDoc: !string |-
    // For example,
    //
    //   {0x40, 0x41, ..., 0x4f, 0x50, 0x51, ..., 0x5f}.NAME(3, 0,
    //        {0x60, 0x61, ..., 0x6f, 0x70, 0x71, ..., 0x7f})
    //
    // returns {0x70, 0x71, ..., 0x7f, 0x40, 0x41, ..., 0x4f}.
  in:
  - &v
    go: $t
    class: vreg
    base: int|uint
    bits: 256
    OverwriteElementBits: 8
  - *v
  - class: immediate
    immOffset: 0
    name: "lo, hi"
  inVariant: []
  out:
  - *v

- go: ConcatShiftBytesRight
  asm: VPALIGNR
  in:
  - &uint128
    go: $t
    base: uint
    bits: 128
  - *uint128
  - class: immediate
    immOffset: 0
  out:
  - *uint128

- go: ConcatShiftBytesRightGrouped
  asm: VPALIGNR
  in:
  - &uint256512
    go: $t
    base: uint
    bits: 256|512
  - *uint256512
  - class: immediate
    immOffset: 0
  out:
  - *uint256512