test/MC/Disassembler/AMDGPU/wave32.txt

   1 # RUN: llvm-mc -arch=amdgcn -mcpu=gfx1010 -mattr=+wavefrontsize32,-wavefrontsize64 -disassemble -show-encoding < %s | FileCheck -check-prefix=GFX1032 %s
   2 # RUN: llvm-mc -arch=amdgcn -mcpu=gfx1010 -mattr=+wavefrontsize64,-wavefrontsize32 -disassemble -show-encoding < %s | FileCheck -check-prefix=GFX1064 %s
   3
   4 # GFX1032:   v_cmp_lt_f32_e32 vcc_lo, s2, v4
   5 # GFX1064:   v_cmp_lt_f32_e32 vcc, s2, v4
   6 0x02,0x08,0x02,0x7c
   7
   8 # GFX1032:   v_cmp_ge_i32_e64 s2, s0, v2
   9 # GFX1064:   v_cmp_ge_i32_e64 s[2:3], s0, v2
  10 0x02,0x00,0x86,0xd4,0x00,0x04,0x02,0x00
  11
  12 # GFX1032: v_cmp_ge_i32_sdwa vcc_lo, v0, v2 src0_sel:WORD_1 src1_sel:DWORD
  13 # GFX1064: v_cmp_ge_i32_sdwa vcc, v0, v2 src0_sel:WORD_1 src1_sel:DWORD
  14 0xf9,0x04,0x0c,0x7d,0x00,0x00,0x05,0x06
  15
  16 # GFX1032: v_cmp_le_f16_sdwa s0, v3, v4 src0_sel:WORD_1 src1_sel:DWORD
  17 # GFX1064: v_cmp_le_f16_sdwa s[0:1], v3, v4 src0_sel:WORD_1 src1_sel:DWORD
  18 0xf9,0x08,0x96,0x7d,0x03,0x80,0x05,0x06
  19
  20 # GFX1032: v_cmp_class_f32_e32 vcc_lo, s0, v0
  21 # GFX1064: v_cmp_class_f32_e32 vcc, s0, v0
  22 0x00,0x00,0x10,0x7d
  23
  24 # GFX1032: v_cmp_class_f16_sdwa vcc_lo, v1, v2 src0_sel:DWORD src1_sel:DWORD
  25 # GFX1064: v_cmp_class_f16_sdwa vcc, v1, v2 src0_sel:DWORD src1_sel:DWORD
  26 0xf9,0x04,0x1e,0x7d,0x01,0x00,0x06,0x06
  27
  28 # GFX1032: v_cmp_class_f16_sdwa s0, v1, v2 src0_sel:DWORD src1_sel:DWORD
  29 # GFX1064: v_cmp_class_f16_sdwa s[0:1], v1, v2 src0_sel:DWORD src1_sel:DWORD
  30 0xf9,0x04,0x1e,0x7d,0x01,0x80,0x06,0x06
  31
  32 # GFX1032: v_cndmask_b32_e32 v5, 0, v2, vcc_lo
  33 # GFX1064: v_cndmask_b32_e32 v5, 0, v2, vcc ;
  34 0x80,0x04,0x0a,0x02
  35
  36 # GFX1032: v_cndmask_b32_e32 v1, v2, v3, vcc_lo
  37 # GFX1064: v_cndmask_b32_e32 v1, v2, v3, vcc ;
  38 0x02,0x07,0x02,0x02
  39
  40 # GFX1032: v_cndmask_b32_sdwa v5, v1, v2, vcc_lo  dst_sel:DWORD dst_unused:UNUSED_PRESERVE src0_sel:DWORD src1_sel:DWORD
  41 # GFX1064: v_cndmask_b32_sdwa v5, v1, v2, vcc  dst_sel:DWORD dst_unused:UNUSED_PRESERVE src0_sel:DWORD src1_sel:DWORD
  42 0xf9,0x04,0x0a,0x02,0x01,0x16,0x06,0x06
  43
  44 # GFX1032: v_cndmask_b32_dpp v5, v1, v2, vcc_lo  quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0
  45 # GFX1064: v_cndmask_b32_dpp v5, v1, v2, vcc  quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0
  46 0xfa,0x04,0x0a,0x02,0x01,0xe4,0x00,0x00
  47
  48 # GFX1032: v_add_co_u32_e64 v2, vcc_lo, s0, v2
  49 # GFX1064: v_add_co_u32_e64 v2, vcc, s0, v2
  50 0x02,0x6a,0x0f,0xd7,0x00,0x04,0x02,0x00
  51
  52 # GFX1032: v_add_co_ci_u32_e32 v3, vcc_lo, v3, v4, vcc_lo
  53 # GFX1064: v_add_co_ci_u32_e32 v3, vcc, v3, v4, vcc ;
  54 0x03,0x09,0x06,0x50
  55
  56 # GFX1032: v_sub_co_u32_e64 v2, vcc_lo, s0, v2
  57 # GFX1064: v_sub_co_u32_e64 v2, vcc, s0, v2
  58 0x02,0x6a,0x10,0xd7,0x00,0x04,0x02,0x00
  59
  60 # GFX1032: v_subrev_co_u32_e64 v2, vcc_lo, s0, v2
  61 # GFX1064: v_subrev_co_u32_e64 v2, vcc, s0, v2
  62 0x02,0x6a,0x19,0xd7,0x00,0x04,0x02,0x00
  63
  64 # GFX1032: v_sub_co_ci_u32_e32 v3, vcc_lo, v3, v4, vcc_lo
  65 # GFX1064: v_sub_co_ci_u32_e32 v3, vcc, v3, v4, vcc ;
  66 0x03,0x09,0x06,0x52
  67
  68 # GFX1032: v_subrev_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
  69 # GFX1064: v_subrev_co_ci_u32_e32 v1, vcc, 0, v1, vcc ;
  70 0x80,0x02,0x02,0x54
  71
  72 # GFX1032: v_add_co_ci_u32_sdwa v1, vcc_lo, v1, v4, vcc_lo  dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
  73 # GFX1064: v_add_co_ci_u32_sdwa v1, vcc, v1, v4, vcc  dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
  74 0xf9,0x08,0x02,0x50,0x01,0x06,0x00,0x06
  75
  76 # GFX1032: v_sub_co_ci_u32_sdwa v1, vcc_lo, v1, v4, vcc_lo  dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
  77 # GFX1064: v_sub_co_ci_u32_sdwa v1, vcc, v1, v4, vcc  dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
  78 0xf9,0x08,0x02,0x52,0x01,0x06,0x00,0x06
  79
  80 # GFX1032: v_subrev_co_ci_u32_sdwa v1, vcc_lo, v1, v4, vcc_lo  dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
  81 # GFX1064: v_subrev_co_ci_u32_sdwa v1, vcc, v1, v4, vcc  dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
  82 0xf9,0x08,0x02,0x54,0x01,0x06,0x00,0x06
  83
  84 # GFX1032: v_add_co_ci_u32_sdwa v1, vcc_lo, sext(v1), sext(v4), vcc_lo  dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
  85 # GFX1064: v_add_co_ci_u32_sdwa v1, vcc, sext(v1), sext(v4), vcc  dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
  86 0xf9,0x08,0x02,0x50,0x01,0x06,0x08,0x0e
  87
  88 # GFX1032: v_add_nc_u32_dpp v5, v1, v2  quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0
  89 # GFX1064: v_add_nc_u32_dpp v5, v1, v2  quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0
  90 0xfa,0x04,0x0a,0x4a,0x01,0xe4,0x00,0x00
  91
  92 # FIXME: Results in invalid v_subrev_u16_dpp which apparently has the same encoding but does not exist in GFX10
  93
  94 # gfx1032: v_add_co_ci_u32_dpp v5, vcc_lo, v1, v2, vcc_lo  quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0
  95 # gfx1064: v_add_co_ci_u32_dpp v5, vcc, v1, v2, vcc  quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0
  96 # 0xfa,0x04,0x0a,0x50,0x01,0xe4,0x00,0x00
  97
  98 # FIXME: Results in v_mul_lo_u16_dpp
  99
 100 # gfx1032: v_sub_co_ci_u32_dpp v5, vcc_lo, v1, v2, vcc_lo  quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0
 101 # gfx1064: v_sub_co_ci_u32_dpp v5, vcc, v1, v2, vcc  quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0
 102 # 0xfa,0x04,0x0a,0x52,0x01,0xe4,0x00,0x00
 103
 104 # FIXME: gives v_lshlrev_b16_dpp
 105
 106 # gfx1032: v_subrev_co_ci_u32_dpp v5, vcc_lo, v1, v2, vcc_lo  quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0
 107 # gfx1064: v_subrev_co_ci_u32_dpp v5, vcc, v1, v2, vcc  quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0
 108 # 0xfa,0x04,0x0a,0x54,0x01,0xe4,0x00,0x00
 109
 110 # GFX1032: v_add_co_u32_e64 v0, s0, v0, v2
 111 # GFX1064: v_add_co_u32_e64 v0, s[0:1], v0, v2
 112 0x00,0x00,0x0f,0xd7,0x00,0x05,0x02,0x00
 113
 114 # GFX1032: v_add_co_ci_u32_e64 v4, s0, v1, v5, s2
 115 # GFX1064: v_add_co_ci_u32_e64 v4, s[0:1], v1, v5, s[2:3]
 116 0x04,0x00,0x28,0xd5,0x01,0x0b,0x0a,0x00
 117
 118 # GFX1032: v_sub_co_u32_e64 v0, s0, v0, v2
 119 # GFX1064: v_sub_co_u32_e64 v0, s[0:1], v0, v2
 120 0x00,0x00,0x10,0xd7,0x00,0x05,0x02,0x00
 121
 122 # GFX1032: v_sub_co_ci_u32_e64 v4, s0, v1, v5, s2
 123 # GFX1064: v_sub_co_ci_u32_e64 v4, s[0:1], v1, v5, s[2:3]
 124 0x04,0x00,0x29,0xd5,0x01,0x0b,0x0a,0x00
 125
 126 # GFX1032: v_subrev_co_u32_e64 v0, s0, v0, v2
 127 # GFX1064: v_subrev_co_u32_e64 v0, s[0:1], v0, v2
 128 0x00,0x00,0x19,0xd7,0x00,0x05,0x02,0x00
 129
 130 # GFX1032: v_subrev_co_ci_u32_e64 v4, s0, v1, v5, s2
 131 # GFX1064: v_subrev_co_ci_u32_e64 v4, s[0:1], v1, v5, s[2:3]
 132 0x04,0x00,0x2a,0xd5,0x01,0x0b,0x0a,0x00
 133
 134 # GFX1032: v_add_co_ci_u32_e64 v4, vcc_lo, v1, v5, s2
 135 # GFX1064: v_add_co_ci_u32_e64 v4, vcc, v1, v5, s[2:3]
 136 0x04,0x6a,0x28,0xd5,0x01,0x0b,0x0a,0x00
 137
 138 # GFX1032: v_add_co_ci_u32_e64 v4, s0, v1, v5, vcc_lo
 139 # GFX1064: v_add_co_ci_u32_e64 v4, s[0:1], v1, v5, vcc ;
 140 0x04,0x00,0x28,0xd5,0x01,0x0b,0xaa,0x01
 141
 142 # GFX1032: v_div_scale_f32 v2, s2, v0, v0, v2
 143 # GFX1064: v_div_scale_f32 v2, s[2:3], v0, v0, v2
 144 0x02,0x02,0x6d,0xd5,0x00,0x01,0x0a,0x04
 145
 146 # GFX1032: v_div_scale_f64 v[2:3], s2, v[0:1], v[0:1], v[2:3]
 147 # GFX1064: v_div_scale_f64 v[2:3], s[2:3], v[0:1], v[0:1], v[2:3]
 148 0x02,0x02,0x6e,0xd5,0x00,0x01,0x0a,0x04
 149
 150 # GFX1032: v_mad_i64_i32 v[0:1], s6, v0, v1, v[2:3]
 151 # GFX1064: v_mad_i64_i32 v[0:1], s[6:7], v0, v1, v[2:3]
 152 0x00,0x06,0x77,0xd5,0x00,0x03,0x0a,0x04
 153
 154 # GFX1032: v_mad_u64_u32 v[0:1], s6, v0, v1, v[2:3]
 155 # GFX1064: v_mad_u64_u32 v[0:1], s[6:7], v0, v1, v[2:3]
 156 0x00,0x06,0x76,0xd5,0x00,0x03,0x0a,0x04
 157
 158 # GFX1032: v_cmpx_neq_f32_e32 v0, v1
 159 # GFX1064: v_cmpx_neq_f32_e32 v0, v1
 160 0x00,0x03,0x3a,0x7c
 161
 162 # GFX1032: v_cmpx_neq_f32_sdwa v0, v1 src0_sel:WORD_1 src1_sel:DWORD
 163 # GFX1064: v_cmpx_neq_f32_sdwa v0, v1 src0_sel:WORD_1 src1_sel:DWORD
 164 0xf9,0x02,0x3a,0x7c,0x00,0x00,0x05,0x06
 165
 166 # GFX1032: v_cmpx_class_f32_e64 v0, 1
 167 # GFX1064: v_cmpx_class_f32_e64 v0, 1
 168 0x00,0x00,0x98,0xd4,0x00,0x03,0x01,0x00
 169
 170 # GFX1032: v_cmpx_class_f32_sdwa v0, 1 src0_sel:WORD_1 src1_sel:DWORD
 171 # GFX1064: v_cmpx_class_f32_sdwa v0, 1 src0_sel:WORD_1 src1_sel:DWORD
 172 0xf9,0x02,0x31,0x7d,0x00,0x00,0x05,0x86