Home | History | Annotate | Download | only in AMDGPU 
      1 // RUN: llvm-mc -arch=amdgcn -show-encoding %s | FileCheck %s --check-prefix=SICI
      2 // RUN: llvm-mc -arch=amdgcn -mcpu=SI -show-encoding %s | FileCheck %s --check-prefix=SICI
      3 // RUN: not llvm-mc -arch=amdgcn -mcpu=tonga -show-encoding %s | FileCheck %s --check-prefix=VI
      4 // RUN: not llvm-mc -arch=amdgcn -mcpu=tonga -show-encoding %s 2>&1 | FileCheck %s -check-prefix=NOVI
      5 
      6 
      7 //===----------------------------------------------------------------------===//
      8 // VOPC Instructions
      9 //===----------------------------------------------------------------------===//
     10 
     11 // Test forced e64 encoding
     12 
     13 v_cmp_lt_f32_e64 s[2:3], v4, -v6
     14 // SICI: v_cmp_lt_f32_e64 s[2:3], v4, -v6 ; encoding: [0x02,0x00,0x02,0xd0,0x04,0x0d,0x02,0x40]
     15 // VI:   v_cmp_lt_f32_e64 s[2:3], v4, -v6 ; encoding: [0x02,0x00,0x41,0xd0,0x04,0x0d,0x02,0x40]
     16 
     17 // Test forcing e64 with vcc dst
     18 
     19 v_cmp_lt_f32_e64 vcc, v4, v6
     20 // SICI: v_cmp_lt_f32_e64 vcc, v4, v6 ; encoding: [0x6a,0x00,0x02,0xd0,0x04,0x0d,0x02,0x00]
     21 // VI: v_cmp_lt_f32_e64 vcc, v4, v6 ; encoding: [0x6a,0x00,0x41,0xd0,0x04,0x0d,0x02,0x00]
     22 
     23 //
     24 // Modifier tests:
     25 //
     26 
     27 v_cmp_lt_f32 s[2:3] -v4, v6
     28 // SICI: v_cmp_lt_f32_e64 s[2:3], -v4, v6 ; encoding: [0x02,0x00,0x02,0xd0,0x04,0x0d,0x02,0x20]
     29 // VI:   v_cmp_lt_f32_e64 s[2:3], -v4, v6 ; encoding: [0x02,0x00,0x41,0xd0,0x04,0x0d,0x02,0x20]
     30 
     31 v_cmp_lt_f32 s[2:3]  v4, -v6
     32 // SICI: v_cmp_lt_f32_e64 s[2:3], v4, -v6 ; encoding: [0x02,0x00,0x02,0xd0,0x04,0x0d,0x02,0x40]
     33 // VI:   v_cmp_lt_f32_e64 s[2:3], v4, -v6 ; encoding: [0x02,0x00,0x41,0xd0,0x04,0x0d,0x02,0x40]
     34 
     35 v_cmp_lt_f32 s[2:3] -v4, -v6
     36 // SICI: v_cmp_lt_f32_e64 s[2:3], -v4, -v6 ; encoding: [0x02,0x00,0x02,0xd0,0x04,0x0d,0x02,0x60]
     37 // VI:   v_cmp_lt_f32_e64 s[2:3], -v4, -v6 ; encoding: [0x02,0x00,0x41,0xd0,0x04,0x0d,0x02,0x60]
     38 
     39 v_cmp_lt_f32 s[2:3] |v4|, v6
     40 // SICI: v_cmp_lt_f32_e64 s[2:3], |v4|, v6 ; encoding: [0x02,0x01,0x02,0xd0,0x04,0x0d,0x02,0x00]
     41 // VI:   v_cmp_lt_f32_e64 s[2:3], |v4|, v6 ; encoding: [0x02,0x01,0x41,0xd0,0x04,0x0d,0x02,0x00]
     42 
     43 v_cmp_lt_f32 s[2:3] v4, |v6|
     44 // SICI: v_cmp_lt_f32_e64 s[2:3], v4, |v6| ; encoding: [0x02,0x02,0x02,0xd0,0x04,0x0d,0x02,0x00]
     45 // VI:   v_cmp_lt_f32_e64 s[2:3], v4, |v6| ; encoding: [0x02,0x02,0x41,0xd0,0x04,0x0d,0x02,0x00]
     46 
     47 v_cmp_lt_f32 s[2:3] |v4|, |v6|
     48 // SICI: v_cmp_lt_f32_e64 s[2:3], |v4|, |v6| ; encoding: [0x02,0x03,0x02,0xd0,0x04,0x0d,0x02,0x00]
     49 // VI:   v_cmp_lt_f32_e64 s[2:3], |v4|, |v6| ; encoding: [0x02,0x03,0x41,0xd0,0x04,0x0d,0x02,0x00]
     50 
     51 v_cmp_lt_f32 s[2:3] -|v4|, v6
     52 // SICI: v_cmp_lt_f32_e64 s[2:3], -|v4|, v6 ; encoding: [0x02,0x01,0x02,0xd0,0x04,0x0d,0x02,0x20]
     53 // VI:   v_cmp_lt_f32_e64 s[2:3], -|v4|, v6 ; encoding: [0x02,0x01,0x41,0xd0,0x04,0x0d,0x02,0x20]
     54 
     55 v_cmp_lt_f32 s[2:3] -abs(v4), v6
     56 // SICI: v_cmp_lt_f32_e64 s[2:3], -|v4|, v6 ; encoding: [0x02,0x01,0x02,0xd0,0x04,0x0d,0x02,0x20]
     57 // VI:   v_cmp_lt_f32_e64 s[2:3], -|v4|, v6 ; encoding: [0x02,0x01,0x41,0xd0,0x04,0x0d,0x02,0x20]
     58 
     59 v_cmp_lt_f32 s[2:3] v4, -|v6|
     60 // SICI: v_cmp_lt_f32_e64 s[2:3], v4, -|v6| ; encoding: [0x02,0x02,0x02,0xd0,0x04,0x0d,0x02,0x40]
     61 // VI:   v_cmp_lt_f32_e64 s[2:3], v4, -|v6| ; encoding: [0x02,0x02,0x41,0xd0,0x04,0x0d,0x02,0x40]
     62 
     63 v_cmp_lt_f32 s[2:3] v4, -abs(v6)
     64 // SICI: v_cmp_lt_f32_e64 s[2:3], v4, -|v6| ; encoding: [0x02,0x02,0x02,0xd0,0x04,0x0d,0x02,0x40]
     65 // VI:   v_cmp_lt_f32_e64 s[2:3], v4, -|v6| ; encoding: [0x02,0x02,0x41,0xd0,0x04,0x0d,0x02,0x40]
     66 
     67 v_cmp_lt_f32 s[2:3] -|v4|, -|v6|
     68 // SICI: v_cmp_lt_f32_e64 s[2:3], -|v4|, -|v6| ; encoding: [0x02,0x03,0x02,0xd0,0x04,0x0d,0x02,0x60]
     69 // VI:   v_cmp_lt_f32_e64 s[2:3], -|v4|, -|v6| ; encoding: [0x02,0x03,0x41,0xd0,0x04,0x0d,0x02,0x60]
     70 
     71 v_cmp_lt_f32 s[2:3] -abs(v4), -abs(v6)
     72 // SICI: v_cmp_lt_f32_e64 s[2:3], -|v4|, -|v6| ; encoding: [0x02,0x03,0x02,0xd0,0x04,0x0d,0x02,0x60]
     73 // VI:   v_cmp_lt_f32_e64 s[2:3], -|v4|, -|v6| ; encoding: [0x02,0x03,0x41,0xd0,0x04,0x0d,0x02,0x60]
     74 
     75 //
     76 // Instruction tests:
     77 //
     78 
     79 v_cmp_f_f32 s[2:3], v4, v6
     80 // SICI: v_cmp_f_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x00,0xd0,0x04,0x0d,0x02,0x00]
     81 // VI:   v_cmp_f_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x40,0xd0,0x04,0x0d,0x02,0x00]
     82 
     83 v_cmp_lt_f32 s[2:3], v4, v6
     84 // SICI: v_cmp_lt_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x02,0xd0,0x04,0x0d,0x02,0x00]
     85 // VI:   v_cmp_lt_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x41,0xd0,0x04,0x0d,0x02,0x00]
     86 
     87 v_cmp_eq_f32 s[2:3], v4, v6
     88 // SICI: v_cmp_eq_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x04,0xd0,0x04,0x0d,0x02,0x00]
     89 // VI:   v_cmp_eq_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x42,0xd0,0x04,0x0d,0x02,0x00]
     90 
     91 v_cmp_le_f32 s[2:3], v4, v6
     92 // SICI: v_cmp_le_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x06,0xd0,0x04,0x0d,0x02,0x00]
     93 // VI:   v_cmp_le_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x43,0xd0,0x04,0x0d,0x02,0x00]
     94 
     95 v_cmp_gt_f32 s[2:3], v4, v6
     96 // SICI: v_cmp_gt_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x08,0xd0,0x04,0x0d,0x02,0x00]
     97 // VI:   v_cmp_gt_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x44,0xd0,0x04,0x0d,0x02,0x00]
     98 
     99 v_cmp_lg_f32 s[2:3], v4, v6
    100 // SICI: v_cmp_lg_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x0a,0xd0,0x04,0x0d,0x02,0x00]
    101 // VI:   v_cmp_lg_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x45,0xd0,0x04,0x0d,0x02,0x00]
    102 
    103 v_cmp_ge_f32 s[2:3], v4, v6
    104 // SICI: v_cmp_ge_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x0c,0xd0,0x04,0x0d,0x02,0x00]
    105 // VI:   v_cmp_ge_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x46,0xd0,0x04,0x0d,0x02,0x00]
    106 
    107 // TODO: Add tests for the rest of v_cmp_*_f32
    108 // TODO: Add tests for v_cmpx_*_f32
    109 
    110 v_cmp_f_f64 s[2:3], v[4:5], v[6:7]
    111 // SICI: v_cmp_f_f64_e64 s[2:3], v[4:5], v[6:7] ; encoding: [0x02,0x00,0x40,0xd0,0x04,0x0d,0x02,0x00]
    112 // VI:   v_cmp_f_f64_e64 s[2:3], v[4:5], v[6:7] ; encoding: [0x02,0x00,0x60,0xd0,0x04,0x0d,0x02,0x00]
    113 
    114 // TODO: Add tests for the rest of v_cmp_*_f64
    115 // TODO: Add tests for the rest of the floating-point comparision instructions.
    116 
    117 v_cmp_f_i32 s[2:3], v4, v6
    118 // SICI: v_cmp_f_i32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x00,0xd1,0x04,0x0d,0x02,0x00]
    119 // VI:   v_cmp_f_i32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0xc0,0xd0,0x04,0x0d,0x02,0x00]
    120 
    121 // TODO: Add test for the rest of v_cmp_*_i32
    122 
    123 v_cmp_f_i64 s[2:3], v[4:5], v[6:7]
    124 // SICI: v_cmp_f_i64_e64 s[2:3], v[4:5], v[6:7] ; encoding: [0x02,0x00,0x40,0xd1,0x04,0x0d,0x02,0x00]
    125 // VI:   v_cmp_f_i64_e64 s[2:3], v[4:5], v[6:7] ; encoding: [0x02,0x00,0xe0,0xd0,0x04,0x0d,0x02,0x00]
    126 
    127 // TODO: Add tests for the rest of the instructions.
    128 
    129 //===----------------------------------------------------------------------===//
    130 // VOP1 Instructions
    131 //===----------------------------------------------------------------------===//
    132 
    133 // Test forced e64 encoding with e32 operands
    134 
    135 v_mov_b32_e64 v1, v2
    136 // SICI: v_mov_b32_e64 v1, v2 ; encoding: [0x01,0x00,0x02,0xd3,0x02,0x01,0x00,0x00]
    137 // VI:   v_mov_b32_e64 v1, v2 ; encoding: [0x01,0x00,0x41,0xd1,0x02,0x01,0x00,0x00]
    138 
    139 // Force e64 encoding for special instructions.
    140 // FIXME, we should be printing the _e64 suffix for v_nop and v_clrexcp.
    141 
    142 v_nop_e64
    143 // SICI: v_nop ; encoding: [0x00,0x00,0x00,0xd3,0x00,0x00,0x00,0x00]
    144 // VI:   v_nop ; encoding: [0x00,0x00,0x40,0xd1,0x00,0x00,0x00,0x00]
    145 
    146 v_clrexcp_e64
    147 // SICI: v_clrexcp ; encoding: [0x00,0x00,0x82,0xd3,0x00,0x00,0x00,0x00]
    148 // VI:   v_clrexcp ; encoding: [0x00,0x00,0x75,0xd1,0x00,0x00,0x00,0x00]
    149 
    150 //
    151 // Modifier tests:
    152 //
    153 
    154 v_fract_f32 v1, -v2
    155 // SICI: v_fract_f32_e64 v1, -v2 ; encoding: [0x01,0x00,0x40,0xd3,0x02,0x01,0x00,0x20]
    156 // VI:   v_fract_f32_e64 v1, -v2 ; encoding: [0x01,0x00,0x5b,0xd1,0x02,0x01,0x00,0x20]
    157 
    158 v_fract_f32 v1, |v2|
    159 // SICI: v_fract_f32_e64 v1, |v2| ; encoding: [0x01,0x01,0x40,0xd3,0x02,0x01,0x00,0x00]
    160 // VI:   v_fract_f32_e64 v1, |v2| ; encoding: [0x01,0x01,0x5b,0xd1,0x02,0x01,0x00,0x00]
    161 
    162 v_fract_f32 v1, abs(v2)
    163 // SICI: v_fract_f32_e64 v1, |v2| ; encoding: [0x01,0x01,0x40,0xd3,0x02,0x01,0x00,0x00]
    164 // VI:   v_fract_f32_e64 v1, |v2| ; encoding: [0x01,0x01,0x5b,0xd1,0x02,0x01,0x00,0x00]
    165 
    166 v_fract_f32 v1, -|v2|
    167 // SICI: v_fract_f32_e64 v1, -|v2| ; encoding: [0x01,0x01,0x40,0xd3,0x02,0x01,0x00,0x20]
    168 // VI:   v_fract_f32_e64 v1, -|v2| ; encoding: [0x01,0x01,0x5b,0xd1,0x02,0x01,0x00,0x20]
    169 
    170 v_fract_f32 v1, -abs(v2)
    171 // SICI: v_fract_f32_e64 v1, -|v2| ; encoding: [0x01,0x01,0x40,0xd3,0x02,0x01,0x00,0x20]
    172 // VI:   v_fract_f32_e64 v1, -|v2| ; encoding: [0x01,0x01,0x5b,0xd1,0x02,0x01,0x00,0x20]
    173 
    174 v_fract_f32 v1, v2 clamp
    175 // SICI: v_fract_f32_e64 v1, v2 clamp ; encoding: [0x01,0x08,0x40,0xd3,0x02,0x01,0x00,0x00]
    176 // VI:   v_fract_f32_e64 v1, v2 clamp ; encoding: [0x01,0x80,0x5b,0xd1,0x02,0x01,0x00,0x00]
    177 
    178 v_fract_f32 v1, v2 mul:2
    179 // SICI: v_fract_f32_e64 v1, v2 mul:2 ; encoding: [0x01,0x00,0x40,0xd3,0x02,0x01,0x00,0x08]
    180 // VI:   v_fract_f32_e64 v1, v2 mul:2 ; encoding: [0x01,0x00,0x5b,0xd1,0x02,0x01,0x00,0x08]
    181 
    182 v_fract_f32 v1, v2, clamp div:2
    183 // SICI: v_fract_f32_e64 v1, v2 clamp div:2 ; encoding: [0x01,0x08,0x40,0xd3,0x02,0x01,0x00,0x18]
    184 // VI:   v_fract_f32_e64 v1, v2 clamp div:2 ; encoding: [0x01,0x80,0x5b,0xd1,0x02,0x01,0x00,0x18]
    185 
    186 // TODO: Finish VOP1
    187 
    188 ///===---------------------------------------------------------------------===//
    189 // VOP2 Instructions
    190 ///===---------------------------------------------------------------------===//
    191 
    192 // Test forced e64 encoding with e32 operands
    193 
    194 v_add_f32_e64 v1, v3, v5
    195 // SICI: v_add_f32_e64 v1, v3, v5 ; encoding: [0x01,0x00,0x06,0xd2,0x03,0x0b,0x02,0x00]
    196 // VI:   v_add_f32_e64 v1, v3, v5 ; encoding: [0x01,0x00,0x01,0xd1,0x03,0x0b,0x02,0x00]
    197 
    198 
    199 // TODO: Modifier tests
    200 
    201 v_cndmask_b32 v1, v3, v5, s[4:5]
    202 // SICI: v_cndmask_b32_e64 v1, v3, v5, s[4:5] ; encoding: [0x01,0x00,0x00,0xd2,0x03,0x0b,0x12,0x00]
    203 // VI:   v_cndmask_b32_e64 v1, v3, v5, s[4:5] ; encoding: [0x01,0x00,0x00,0xd1,0x03,0x0b,0x12,0x00]
    204 
    205 v_cndmask_b32_e64 v1, v3, v5, s[4:5]
    206 // SICI: v_cndmask_b32_e64 v1, v3, v5, s[4:5] ; encoding: [0x01,0x00,0x00,0xd2,0x03,0x0b,0x12,0x00]
    207 // VI:   v_cndmask_b32_e64 v1, v3, v5, s[4:5] ; encoding: [0x01,0x00,0x00,0xd1,0x03,0x0b,0x12,0x00]
    208 
    209 v_cndmask_b32_e64 v1, v3, v5, vcc
    210 // SICI: v_cndmask_b32_e64 v1, v3, v5, vcc ; encoding: [0x01,0x00,0x00,0xd2,0x03,0x0b,0xaa,0x01]
    211 // VI:   v_cndmask_b32_e64 v1, v3, v5, vcc ; encoding: [0x01,0x00,0x00,0xd1,0x03,0x0b,0xaa,0x01]
    212 
    213 //TODO: readlane, writelane
    214 
    215 v_add_f32 v1, v3, s5
    216 // SICI: v_add_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x06,0xd2,0x03,0x0b,0x00,0x00]
    217 // VI:   v_add_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x01,0xd1,0x03,0x0b,0x00,0x00]
    218 
    219 v_sub_f32 v1, v3, s5
    220 // SICI: v_sub_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x08,0xd2,0x03,0x0b,0x00,0x00]
    221 // VI:   v_sub_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x02,0xd1,0x03,0x0b,0x00,0x00]
    222 
    223 v_subrev_f32 v1, v3, s5
    224 // SICI: v_subrev_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x0a,0xd2,0x03,0x0b,0x00,0x00]
    225 // VI:   v_subrev_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x03,0xd1,0x03,0x0b,0x00,0x00]
    226 
    227 v_mac_legacy_f32 v1, v3, s5
    228 // SICI: v_mac_legacy_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x0c,0xd2,0x03,0x0b,0x00,0x00]
    229 // NOVI: error: instruction not supported on this GPU
    230 
    231 v_mul_legacy_f32 v1, v3, s5
    232 // SICI: v_mul_legacy_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x0e,0xd2,0x03,0x0b,0x00,0x00]
    233 // VI:   v_mul_legacy_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x04,0xd1,0x03,0x0b,0x00,0x00]
    234 
    235 v_mul_f32 v1, v3, s5
    236 // SICI: v_mul_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x10,0xd2,0x03,0x0b,0x00,0x00]
    237 // VI:   v_mul_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x05,0xd1,0x03,0x0b,0x00,0x00]
    238 
    239 v_mul_i32_i24 v1, v3, s5
    240 // SICI: v_mul_i32_i24_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x12,0xd2,0x03,0x0b,0x00,0x00]
    241 // VI:   v_mul_i32_i24_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x06,0xd1,0x03,0x0b,0x00,0x00]
    242 
    243 ///===---------------------------------------------------------------------===//
    244 // VOP3 Instructions
    245 ///===---------------------------------------------------------------------===//
    246 
    247 // TODO: Modifier tests
    248 
    249 v_mad_legacy_f32 v2, v4, v6, v8
    250 // SICI: v_mad_legacy_f32 v2, v4, v6, v8 ; encoding: [0x02,0x00,0x80,0xd2,0x04,0x0d,0x22,0x04]
    251 // VI:   v_mad_legacy_f32 v2, v4, v6, v8 ; encoding: [0x02,0x00,0xc0,0xd1,0x04,0x0d,0x22,0x04]
    252 
    253 v_add_f64 v[0:1], v[2:3], v[5:6]
    254 // SICI: v_add_f64 v[0:1], v[2:3], v[5:6] ; encoding: [0x00,0x00,0xc8,0xd2,0x02,0x0b,0x02,0x00]
    255 // VI:   v_add_f64 v[0:1], v[2:3], v[5:6] ; encoding: [0x00,0x00,0x80,0xd2,0x02,0x0b,0x02,0x00]
    256 
    257 v_add_f64_e64 v[0:1], v[2:3], v[5:6]
    258 // SICI: v_add_f64 v[0:1], v[2:3], v[5:6] ; encoding: [0x00,0x00,0xc8,0xd2,0x02,0x0b,0x02,0x00]
    259 // VI:   v_add_f64 v[0:1], v[2:3], v[5:6] ; encoding: [0x00,0x00,0x80,0xd2,0x02,0x0b,0x02,0x00]
    260 
    261 v_add_f64 v[0:1], -v[2:3], v[5:6]
    262 // SICI: v_add_f64 v[0:1], -v[2:3], v[5:6] ; encoding: [0x00,0x00,0xc8,0xd2,0x02,0x0b,0x02,0x20]
    263 // VI:   v_add_f64 v[0:1], -v[2:3], v[5:6] ; encoding: [0x00,0x00,0x80,0xd2,0x02,0x0b,0x02,0x20]
    264 
    265 v_add_f64_e64 v[0:1], -v[2:3], v[5:6]
    266 // SICI: v_add_f64 v[0:1], -v[2:3], v[5:6] ; encoding: [0x00,0x00,0xc8,0xd2,0x02,0x0b,0x02,0x20]
    267 // VI:   v_add_f64 v[0:1], -v[2:3], v[5:6] ; encoding: [0x00,0x00,0x80,0xd2,0x02,0x0b,0x02,0x20]
    268 
    269 v_add_f64 v[0:1], v[2:3], -v[5:6]
    270 // SICI: v_add_f64 v[0:1], v[2:3], -v[5:6] ; encoding: [0x00,0x00,0xc8,0xd2,0x02,0x0b,0x02,0x40]
    271 // VI:   v_add_f64 v[0:1], v[2:3], -v[5:6] ; encoding: [0x00,0x00,0x80,0xd2,0x02,0x0b,0x02,0x40]
    272 
    273 v_add_f64_e64 v[0:1], v[2:3], -v[5:6]
    274 // SICI: v_add_f64 v[0:1], v[2:3], -v[5:6] ; encoding: [0x00,0x00,0xc8,0xd2,0x02,0x0b,0x02,0x40]
    275 // VI:   v_add_f64 v[0:1], v[2:3], -v[5:6] ; encoding: [0x00,0x00,0x80,0xd2,0x02,0x0b,0x02,0x40]
    276 
    277 v_add_f64 v[0:1], |v[2:3]|, v[5:6]
    278 // SICI: v_add_f64 v[0:1], |v[2:3]|, v[5:6] ; encoding: [0x00,0x01,0xc8,0xd2,0x02,0x0b,0x02,0x00]
    279 // VI:   v_add_f64 v[0:1], |v[2:3]|, v[5:6] ; encoding: [0x00,0x01,0x80,0xd2,0x02,0x0b,0x02,0x00]
    280 
    281 v_add_f64 v[0:1], abs(v[2:3]), v[5:6]
    282 // SICI: v_add_f64 v[0:1], |v[2:3]|, v[5:6] ; encoding: [0x00,0x01,0xc8,0xd2,0x02,0x0b,0x02,0x00]
    283 // VI:   v_add_f64 v[0:1], |v[2:3]|, v[5:6] ; encoding: [0x00,0x01,0x80,0xd2,0x02,0x0b,0x02,0x00]
    284 
    285 v_add_f64_e64 v[0:1], |v[2:3]|, v[5:6]
    286 // SICI: v_add_f64 v[0:1], |v[2:3]|, v[5:6] ; encoding: [0x00,0x01,0xc8,0xd2,0x02,0x0b,0x02,0x00]
    287 // VI:   v_add_f64 v[0:1], |v[2:3]|, v[5:6] ; encoding: [0x00,0x01,0x80,0xd2,0x02,0x0b,0x02,0x00]
    288 
    289 v_add_f64_e64 v[0:1], abs(v[2:3]), v[5:6]
    290 // SICI: v_add_f64 v[0:1], |v[2:3]|, v[5:6] ; encoding: [0x00,0x01,0xc8,0xd2,0x02,0x0b,0x02,0x00]
    291 // VI:   v_add_f64 v[0:1], |v[2:3]|, v[5:6] ; encoding: [0x00,0x01,0x80,0xd2,0x02,0x0b,0x02,0x00]
    292 
    293 v_add_f64 v[0:1], v[2:3], |v[5:6]|
    294 // SICI: v_add_f64 v[0:1], v[2:3], |v[5:6]| ; encoding: [0x00,0x02,0xc8,0xd2,0x02,0x0b,0x02,0x00]
    295 // VI:   v_add_f64 v[0:1], v[2:3], |v[5:6]| ; encoding: [0x00,0x02,0x80,0xd2,0x02,0x0b,0x02,0x00]
    296 
    297 v_add_f64 v[0:1], v[2:3], abs(v[5:6])
    298 // SICI: v_add_f64 v[0:1], v[2:3], |v[5:6]| ; encoding: [0x00,0x02,0xc8,0xd2,0x02,0x0b,0x02,0x00]
    299 // VI:   v_add_f64 v[0:1], v[2:3], |v[5:6]| ; encoding: [0x00,0x02,0x80,0xd2,0x02,0x0b,0x02,0x00]
    300 
    301 v_add_f64_e64 v[0:1], v[2:3], |v[5:6]|
    302 // SICI: v_add_f64 v[0:1], v[2:3], |v[5:6]| ; encoding: [0x00,0x02,0xc8,0xd2,0x02,0x0b,0x02,0x00]
    303 // VI:   v_add_f64 v[0:1], v[2:3], |v[5:6]| ; encoding: [0x00,0x02,0x80,0xd2,0x02,0x0b,0x02,0x00]
    304 
    305 v_add_f64_e64 v[0:1], v[2:3], abs(v[5:6])
    306 // SICI: v_add_f64 v[0:1], v[2:3], |v[5:6]| ; encoding: [0x00,0x02,0xc8,0xd2,0x02,0x0b,0x02,0x00]
    307 // VI:   v_add_f64 v[0:1], v[2:3], |v[5:6]| ; encoding: [0x00,0x02,0x80,0xd2,0x02,0x0b,0x02,0x00]
    308 
    309 v_add_f64 v[0:1], -v[2:3], |v[5:6]| clamp mul:4
    310 // SICI: v_add_f64 v[0:1], -v[2:3], |v[5:6]| clamp mul:4 ; encoding: [0x00,0x0a,0xc8,0xd2,0x02,0x0b,0x02,0x30]
    311 // VI:   v_add_f64 v[0:1], -v[2:3], |v[5:6]| clamp mul:4 ; encoding: [0x00,0x82,0x80,0xd2,0x02,0x0b,0x02,0x30]
    312 
    313 v_add_f64 v[0:1], -v[2:3], abs(v[5:6]) clamp mul:4
    314 // SICI: v_add_f64 v[0:1], -v[2:3], |v[5:6]| clamp mul:4 ; encoding: [0x00,0x0a,0xc8,0xd2,0x02,0x0b,0x02,0x30]
    315 // VI:   v_add_f64 v[0:1], -v[2:3], |v[5:6]| clamp mul:4 ; encoding: [0x00,0x82,0x80,0xd2,0x02,0x0b,0x02,0x30]
    316 
    317 v_add_f64_e64 v[0:1], -v[2:3], |v[5:6]| clamp mul:4
    318 // SICI: v_add_f64 v[0:1], -v[2:3], |v[5:6]| clamp mul:4 ; encoding: [0x00,0x0a,0xc8,0xd2,0x02,0x0b,0x02,0x30]
    319 // VI:   v_add_f64 v[0:1], -v[2:3], |v[5:6]| clamp mul:4 ; encoding: [0x00,0x82,0x80,0xd2,0x02,0x0b,0x02,0x30]
    320 
    321 v_add_f64_e64 v[0:1], -v[2:3], abs(v[5:6]) clamp mul:4
    322 // SICI: v_add_f64 v[0:1], -v[2:3], |v[5:6]| clamp mul:4 ; encoding: [0x00,0x0a,0xc8,0xd2,0x02,0x0b,0x02,0x30]
    323 // VI:   v_add_f64 v[0:1], -v[2:3], |v[5:6]| clamp mul:4 ; encoding: [0x00,0x82,0x80,0xd2,0x02,0x0b,0x02,0x30]
    324 
    325 v_div_scale_f64  v[24:25], vcc, v[22:23], v[22:23], v[20:21]
    326 // SICI: v_div_scale_f64 v[24:25], vcc, v[22:23], v[22:23], v[20:21] ; encoding: [0x18,0x6a,0xdc,0xd2,0x16,0x2d,0x52,0x04]
    327 // VI:   v_div_scale_f64 v[24:25], vcc, v[22:23], v[22:23], v[20:21] ; encoding: [0x18,0x6a,0xe1,0xd1,0x16,0x2d,0x52,0x04]
    328 
    329 v_div_scale_f64  v[24:25], s[10:11], v[22:23], v[20:21], v[20:21]
    330 // SICI: v_div_scale_f64 v[24:25], s[10:11], v[22:23], v[20:21], v[20:21] ; encoding: [0x18,0x0a,0xdc,0xd2,0x16,0x29,0x52,0x04]
    331 // VI:   v_div_scale_f64 v[24:25], s[10:11], v[22:23], v[20:21], v[20:21] ; encoding: [0x18,0x0a,0xe1,0xd1,0x16,0x29,0x52,0x04]
    332 
    333 v_div_scale_f32  v24, vcc, v22, v22, v20
    334 // SICI: v_div_scale_f32 v24, vcc, v22, v22, v20 ; encoding: [0x18,0x6a,0xda,0xd2,0x16,0x2d,0x52,0x04]
    335 // VI:   v_div_scale_f32 v24, vcc, v22, v22, v20 ; encoding: [0x18,0x6a,0xe0,0xd1,0x16,0x2d,0x52,0x04]
    336 
    337 v_div_scale_f32  v24, vcc, s[10:11], v22, v20
    338 // SICI: v_div_scale_f32 v24, vcc, s[10:11], v22, v20 ; encoding: [0x18,0x6a,0xda,0xd2,0x0a,0x2c,0x52,0x04]
    339 // VI:   v_div_scale_f32 v24, vcc, s[10:11], v22, v20 ; encoding: [0x18,0x6a,0xe0,0xd1,0x0a,0x2c,0x52,0x04]
    340 
    341 v_div_scale_f32  v24, s[10:11], v22, v22, v20
    342 // SICI: v_div_scale_f32 v24, s[10:11], v22, v22, v20 ; encoding: [0x18,0x0a,0xda,0xd2,0x16,0x2d,0x52,0x04]
    343 // VI:   v_div_scale_f32 v24, s[10:11], v22, v22, v20 ; encoding: [0x18,0x0a,0xe0,0xd1,0x16,0x2d,0x52,0x04]
    344 
    345 v_div_scale_f32  v24, vcc, v22, 1.0, v22
    346 // SICI: v_div_scale_f32 v24, vcc, v22, 1.0, v22 ; encoding: [0x18,0x6a,0xda,0xd2,0x16,0xe5,0x59,0x04]
    347 // VI:   v_div_scale_f32 v24, vcc, v22, 1.0, v22 ; encoding: [0x18,0x6a,0xe0,0xd1,0x16,0xe5,0x59,0x04]
    348 
    349 v_div_scale_f32  v24, vcc, v22, v22, -2.0
    350 // SICI: v_div_scale_f32 v24, vcc, v22, v22, -2.0 ; encoding: [0x18,0x6a,0xda,0xd2,0x16,0x2d,0xd2,0x83]
    351 // VI:   v_div_scale_f32 v24, vcc, v22, v22, -2.0 ; encoding: [0x18,0x6a,0xe0,0xd1,0x16,0x2d,0xd2,0x83]
    352 
    353 v_div_scale_f32 v24, vcc, v22, v22, 0xc0000000
    354 // SICI: v_div_scale_f32 v24, vcc, v22, v22, -2.0 ; encoding: [0x18,0x6a,0xda,0xd2,0x16,0x2d,0xd6,0x03]
    355 // VI:   v_div_scale_f32 v24, vcc, v22, v22, -2.0 ; encoding: [0x18,0x6a,0xe0,0xd1,0x16,0x2d,0xd6,0x03]
    356 
    357 v_mad_f32 v9, 0.5, v5, -v8
    358 // SICI: v_mad_f32 v9, 0.5, v5, -v8      ; encoding: [0x09,0x00,0x82,0xd2,0xf0,0x0a,0x22,0x84]
    359 // VI:   v_mad_f32 v9, 0.5, v5, -v8      ; encoding: [0x09,0x00,0xc1,0xd1,0xf0,0x0a,0x22,0x84]
    360