input_desc = [
  "v128:data",
  "v128:fill",
  "u32:offset",
  "u32:dummy",
]

output_desc = [
  "v128:lhv_e0",
  "v128:lhv_e1",
  "v128:lhv_e2",
  "v128:lhv_e3",
  "v128:lhv_e4",
  "v128:lhv_e5",
  "v128:lhv_e6",
  "v128:lhv_e7",
  "v128:lhv_e8",
  "v128:lhv_e9",
  "v128:lhv_e10",
  "v128:lhv_e11",
  "v128:lhv_e12",
  "v128:lhv_e13",
  "v128:lhv_e14",
  "v128:lhv_e15",
  "v128:shv_e0",
  "v128:shv_e1",
  "v128:shv_e2",
  "v128:shv_e3",
  "v128:shv_e4",
  "v128:shv_e5",
  "v128:shv_e6",
  "v128:shv_e7",
  "v128:shv_e8",
  "v128:shv_e9",
  "v128:shv_e10",
  "v128:shv_e11",
  "v128:shv_e12",
  "v128:shv_e13",
  "v128:shv_e14",
  "v128:shv_e15",
]

rsp_code = """
  li a0,$0
  li a1,$800
  lw t4,$20(a0) // input: offset

  // This test is very sensitive to possible read/write mistakes to memory
  // and we want to make sure that writes actually happen.
  // So clear all registers and also output memory area to make sure
  // tests do not shadow previous results.
  vxor v0,v0
  vxor v1,v1
  vxor v2,v2
  vxor v3,v3
  vxor v4,v4
  vxor v5,v5
  vxor v6,v6
  vxor v7,v7
  vxor v8,v8
  vxor v9,v9
  vxor v10,v10
  vxor v11,v11
  vxor v12,v12
  vxor v13,v13
  vxor v14,v14
  vxor v15,v15

  lqv v0[e0],$10(a0) // input: fill
  add a2,a1,0
  addi a2,$1F0
ClearLoop:
  sqv v0[e0],$00(a2)
  bne a1,a2,ClearLoop
  subi a2,$10

  vxor v0,v0

  add a0,t4 // add offset to disalign
  lhv v0[e0],$00(a0)
  lhv v1[e1],$00(a0)
  lhv v2[e2],$00(a0)
  lhv v3[e3],$00(a0)
  lhv v4[e4],$00(a0)
  lhv v5[e5],$00(a0)
  lhv v6[e6],$00(a0)
  lhv v7[e7],$00(a0)
  lhv v8[e8],$00(a0)
  lhv v9[e9],$00(a0)
  lhv v10[e10],$00(a0)
  lhv v11[e11],$00(a0)
  lhv v12[e12],$00(a0)
  lhv v13[e13],$00(a0)
  lhv v14[e14],$00(a0)
  lhv v15[e15],$00(a0)

  sqv v0[e0],$00(a1)
  sqv v1[e0],$10(a1)
  sqv v2[e0],$20(a1)
  sqv v3[e0],$30(a1)
  sqv v4[e0],$40(a1)
  sqv v5[e0],$50(a1)
  sqv v6[e0],$60(a1)
  sqv v7[e0],$70(a1)
  sqv v8[e0],$80(a1)
  sqv v9[e0],$90(a1)
  sqv v10[e0],$A0(a1)
  sqv v11[e0],$B0(a1)
  sqv v12[e0],$C0(a1)
  sqv v13[e0],$D0(a1)
  sqv v14[e0],$E0(a1)
  sqv v15[e0],$F0(a1)

  li a0,$0
  lqv v0[e0],$00(a0) // input: v0

  add a1,t4 // add offset to disalign
  addi a2,a1,$1F0
  shv v0[e15],$00(a2)
  addi a2,a1,$1E0
  shv v0[e14],$00(a2)
  addi a2,a1,$1D0
  shv v0[e13],$00(a2)
  addi a2,a1,$1C0
  shv v0[e12],$00(a2)
  addi a2,a1,$1B0
  shv v0[e11],$00(a2)
  addi a2,a1,$1A0
  shv v0[e10],$00(a2)
  addi a2,a1,$190
  shv v0[e9],$00(a2)
  addi a2,a1,$180
  shv v0[e8],$00(a2)
  addi a2,a1,$170
  shv v0[e7],$00(a2)
  addi a2,a1,$160
  shv v0[e6],$00(a2)
  addi a2,a1,$150
  shv v0[e5],$00(a2)
  addi a2,a1,$140
  shv v0[e4],$00(a2)
  addi a2,a1,$130
  shv v0[e3],$00(a2)
  addi a2,a1,$120
  shv v0[e2],$00(a2)
  addi a2,a1,$110
  shv v0[e1],$00(a2)
  addi a2,a1,$100
  shv v0[e0],$00(a2)

  break
"""

[[test]]
name = "offset0"
input = [
  0x1122_3344, 0x5A6B_7182, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  0, # offset
  0, # dummy
]
[[test]]
name = "offset1"
input = [
  0x1122_3344, 0x5566_7788, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  1, # offset
  0, # dummy
]
[[test]]
name = "offset2"
input = [
  0x1122_3344, 0x5566_7788, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  2, # offset
  0, # dummy
]
[[test]]
name = "offset3"
input = [
  0x1122_3344, 0x5566_7788, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  3, # offset
  0, # dummy
]
[[test]]
name = "offset4"
input = [
  0x1122_3344, 0x5566_7788, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  4, # offset
  0, # dummy
]
[[test]]
name = "offset5"
input = [
  0x1122_3344, 0x5566_7788, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  5, # offset
  0, # dummy
]
[[test]]
name = "offset6"
input = [
  0x1122_3344, 0x5566_7788, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  6, # offset
  0, # dummy
]
[[test]]
name = "offset7"
input = [
  0x1122_3344, 0x5566_7788, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  7, # offset
  0, # dummy
]
[[test]]
name = "offset8"
input = [
  0x1122_3344, 0x5566_7788, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  8, # offset
  0, # dummy
]
[[test]]
name = "offset9"
input = [
  0x1122_3344, 0x5566_7788, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  9, # offset
  0, # dummy
]
[[test]]
name = "offset10"
input = [
  0x1122_3344, 0x5566_7788, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  10, # offset
  0, # dummy
]
[[test]]
name = "offset11"
input = [
  0x1122_3344, 0x5566_7788, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  11, # offset
  0, # dummy
]
[[test]]
name = "offset12"
input = [
  0x1122_3344, 0x5566_7788, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  12, # offset
  0, # dummy
]
[[test]]
name = "offset13"
input = [
  0x1122_3344, 0x5566_7788, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  13, # offset
  0, # dummy
]
[[test]]
name = "offset14"
input = [
  0x1122_3344, 0x5566_7788, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  14, # offset
  0, # dummy
]
[[test]]
name = "offset15"
input = [
  0x1122_3344, 0x5566_7788, 0x99AA_BBCC, 0xDDEE_FFAB, # v0
  0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, 0xAAAA_AAAA, # fill
  15, # offset
  0, # dummy
]