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] v4, -|v6| 56 // SICI: v_cmp_lt_f32_e64 s[2:3], v4, -|v6| ; encoding: [0x02,0x02,0x02,0xd0,0x04,0x0d,0x02,0x40] 57 // VI: v_cmp_lt_f32_e64 s[2:3], v4, -|v6| ; encoding: [0x02,0x02,0x41,0xd0,0x04,0x0d,0x02,0x40] 58 59 v_cmp_lt_f32 s[2:3] -|v4|, -|v6| 60 // SICI: v_cmp_lt_f32_e64 s[2:3], -|v4|, -|v6| ; encoding: [0x02,0x03,0x02,0xd0,0x04,0x0d,0x02,0x60] 61 // VI: v_cmp_lt_f32_e64 s[2:3], -|v4|, -|v6| ; encoding: [0x02,0x03,0x41,0xd0,0x04,0x0d,0x02,0x60] 62 63 // 64 // Instruction tests: 65 // 66 67 v_cmp_f_f32 s[2:3], v4, v6 68 // SICI: v_cmp_f_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x00,0xd0,0x04,0x0d,0x02,0x00] 69 // VI: v_cmp_f_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x40,0xd0,0x04,0x0d,0x02,0x00] 70 71 v_cmp_lt_f32 s[2:3], v4, v6 72 // SICI: v_cmp_lt_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x02,0xd0,0x04,0x0d,0x02,0x00] 73 // VI: v_cmp_lt_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x41,0xd0,0x04,0x0d,0x02,0x00] 74 75 v_cmp_eq_f32 s[2:3], v4, v6 76 // SICI: v_cmp_eq_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x04,0xd0,0x04,0x0d,0x02,0x00] 77 // VI: v_cmp_eq_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x42,0xd0,0x04,0x0d,0x02,0x00] 78 79 v_cmp_le_f32 s[2:3], v4, v6 80 // SICI: v_cmp_le_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x06,0xd0,0x04,0x0d,0x02,0x00] 81 // VI: v_cmp_le_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x43,0xd0,0x04,0x0d,0x02,0x00] 82 83 v_cmp_gt_f32 s[2:3], v4, v6 84 // SICI: v_cmp_gt_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x08,0xd0,0x04,0x0d,0x02,0x00] 85 // VI: v_cmp_gt_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x44,0xd0,0x04,0x0d,0x02,0x00] 86 87 v_cmp_lg_f32 s[2:3], v4, v6 88 // SICI: v_cmp_lg_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x0a,0xd0,0x04,0x0d,0x02,0x00] 89 // VI: v_cmp_lg_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x45,0xd0,0x04,0x0d,0x02,0x00] 90 91 v_cmp_ge_f32 s[2:3], v4, v6 92 // SICI: v_cmp_ge_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x0c,0xd0,0x04,0x0d,0x02,0x00] 93 // VI: v_cmp_ge_f32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x46,0xd0,0x04,0x0d,0x02,0x00] 94 95 // TODO: Add tests for the rest of v_cmp_*_f32 96 // TODO: Add tests for v_cmpx_*_f32 97 98 v_cmp_f_f64 s[2:3], v[4:5], v[6:7] 99 // SICI: v_cmp_f_f64_e64 s[2:3], v[4:5], v[6:7] ; encoding: [0x02,0x00,0x40,0xd0,0x04,0x0d,0x02,0x00] 100 // VI: v_cmp_f_f64_e64 s[2:3], v[4:5], v[6:7] ; encoding: [0x02,0x00,0x60,0xd0,0x04,0x0d,0x02,0x00] 101 102 // TODO: Add tests for the rest of v_cmp_*_f64 103 // TODO: Add tests for the rest of the floating-point comparision instructions. 104 105 v_cmp_f_i32 s[2:3], v4, v6 106 // SICI: v_cmp_f_i32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0x00,0xd1,0x04,0x0d,0x02,0x00] 107 // VI: v_cmp_f_i32_e64 s[2:3], v4, v6 ; encoding: [0x02,0x00,0xc0,0xd0,0x04,0x0d,0x02,0x00] 108 109 // TODO: Add test for the rest of v_cmp_*_i32 110 111 v_cmp_f_i64 s[2:3], v[4:5], v[6:7] 112 // SICI: v_cmp_f_i64_e64 s[2:3], v[4:5], v[6:7] ; encoding: [0x02,0x00,0x40,0xd1,0x04,0x0d,0x02,0x00] 113 // VI: v_cmp_f_i64_e64 s[2:3], v[4:5], v[6:7] ; encoding: [0x02,0x00,0xe0,0xd0,0x04,0x0d,0x02,0x00] 114 115 // TODO: Add tests for the rest of the instructions. 116 117 //===----------------------------------------------------------------------===// 118 // VOP1 Instructions 119 //===----------------------------------------------------------------------===// 120 121 // Test forced e64 encoding with e32 operands 122 123 v_mov_b32_e64 v1, v2 124 // SICI: v_mov_b32_e64 v1, v2 ; encoding: [0x01,0x00,0x02,0xd3,0x02,0x01,0x00,0x00] 125 // VI: v_mov_b32_e64 v1, v2 ; encoding: [0x01,0x00,0x41,0xd1,0x02,0x01,0x00,0x00] 126 127 // Force e64 encoding for special instructions. 128 // FIXME, we should be printing the _e64 suffix for v_nop and v_clrexcp. 129 130 v_nop_e64 131 // SICI: v_nop ; encoding: [0x00,0x00,0x00,0xd3,0x00,0x00,0x00,0x00] 132 // VI: v_nop ; encoding: [0x00,0x00,0x40,0xd1,0x00,0x00,0x00,0x00] 133 134 v_clrexcp_e64 135 // SICI: v_clrexcp ; encoding: [0x00,0x00,0x82,0xd3,0x00,0x00,0x00,0x00] 136 // VI: v_clrexcp ; encoding: [0x00,0x00,0x75,0xd1,0x00,0x00,0x00,0x00] 137 138 // 139 // Modifier tests: 140 // 141 142 v_fract_f32 v1, -v2 143 // SICI: v_fract_f32_e64 v1, -v2 ; encoding: [0x01,0x00,0x40,0xd3,0x02,0x01,0x00,0x20] 144 // VI: v_fract_f32_e64 v1, -v2 ; encoding: [0x01,0x00,0x5b,0xd1,0x02,0x01,0x00,0x20] 145 146 v_fract_f32 v1, |v2| 147 // SICI: v_fract_f32_e64 v1, |v2| ; encoding: [0x01,0x01,0x40,0xd3,0x02,0x01,0x00,0x00] 148 // VI: v_fract_f32_e64 v1, |v2| ; encoding: [0x01,0x01,0x5b,0xd1,0x02,0x01,0x00,0x00] 149 150 v_fract_f32 v1, -|v2| 151 // SICI: v_fract_f32_e64 v1, -|v2| ; encoding: [0x01,0x01,0x40,0xd3,0x02,0x01,0x00,0x20] 152 // VI: v_fract_f32_e64 v1, -|v2| ; encoding: [0x01,0x01,0x5b,0xd1,0x02,0x01,0x00,0x20] 153 154 v_fract_f32 v1, v2 clamp 155 // SICI: v_fract_f32_e64 v1, v2 clamp ; encoding: [0x01,0x08,0x40,0xd3,0x02,0x01,0x00,0x00] 156 // VI: v_fract_f32_e64 v1, v2 clamp ; encoding: [0x01,0x80,0x5b,0xd1,0x02,0x01,0x00,0x00] 157 158 v_fract_f32 v1, v2 mul:2 159 // SICI: v_fract_f32_e64 v1, v2 mul:2 ; encoding: [0x01,0x00,0x40,0xd3,0x02,0x01,0x00,0x08] 160 // VI: v_fract_f32_e64 v1, v2 mul:2 ; encoding: [0x01,0x00,0x5b,0xd1,0x02,0x01,0x00,0x08] 161 162 v_fract_f32 v1, v2, div:2 clamp 163 // SICI: v_fract_f32_e64 v1, v2 clamp div:2 ; encoding: [0x01,0x08,0x40,0xd3,0x02,0x01,0x00,0x18] 164 // VI: v_fract_f32_e64 v1, v2 clamp div:2 ; encoding: [0x01,0x80,0x5b,0xd1,0x02,0x01,0x00,0x18] 165 166 // TODO: Finish VOP1 167 168 ///===---------------------------------------------------------------------===// 169 // VOP2 Instructions 170 ///===---------------------------------------------------------------------===// 171 172 // Test forced e64 encoding with e32 operands 173 174 v_add_f32_e64 v1, v3, v5 175 // SICI: v_add_f32_e64 v1, v3, v5 ; encoding: [0x01,0x00,0x06,0xd2,0x03,0x0b,0x02,0x00] 176 // VI: v_add_f32_e64 v1, v3, v5 ; encoding: [0x01,0x00,0x01,0xd1,0x03,0x0b,0x02,0x00] 177 178 179 // TODO: Modifier tests 180 181 v_cndmask_b32 v1, v3, v5, s[4:5] 182 // SICI: v_cndmask_b32_e64 v1, v3, v5, s[4:5] ; encoding: [0x01,0x00,0x00,0xd2,0x03,0x0b,0x12,0x00] 183 // VI: v_cndmask_b32_e64 v1, v3, v5, s[4:5] ; encoding: [0x01,0x00,0x00,0xd1,0x03,0x0b,0x12,0x00] 184 185 //TODO: readlane, writelane 186 187 v_add_f32 v1, v3, s5 188 // SICI: v_add_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x06,0xd2,0x03,0x0b,0x00,0x00] 189 // VI: v_add_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x01,0xd1,0x03,0x0b,0x00,0x00] 190 191 v_sub_f32 v1, v3, s5 192 // SICI: v_sub_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x08,0xd2,0x03,0x0b,0x00,0x00] 193 // VI: v_sub_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x02,0xd1,0x03,0x0b,0x00,0x00] 194 195 v_subrev_f32 v1, v3, s5 196 // SICI: v_subrev_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x0a,0xd2,0x03,0x0b,0x00,0x00] 197 // VI: v_subrev_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x03,0xd1,0x03,0x0b,0x00,0x00] 198 199 v_mac_legacy_f32 v1, v3, s5 200 // SICI: v_mac_legacy_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x0c,0xd2,0x03,0x0b,0x00,0x00] 201 // FIXME: The error message should be: error: instruction not supported on this GPU 202 // NOVI: error: invalid operand for instruction 203 204 v_mul_legacy_f32 v1, v3, s5 205 // SICI: v_mul_legacy_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x0e,0xd2,0x03,0x0b,0x00,0x00] 206 // VI: v_mul_legacy_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x04,0xd1,0x03,0x0b,0x00,0x00] 207 208 v_mul_f32 v1, v3, s5 209 // SICI: v_mul_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x10,0xd2,0x03,0x0b,0x00,0x00] 210 // VI: v_mul_f32_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x05,0xd1,0x03,0x0b,0x00,0x00] 211 212 v_mul_i32_i24 v1, v3, s5 213 // SICI: v_mul_i32_i24_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x12,0xd2,0x03,0x0b,0x00,0x00] 214 // VI: v_mul_i32_i24_e64 v1, v3, s5 ; encoding: [0x01,0x00,0x06,0xd1,0x03,0x0b,0x00,0x00] 215 216 ///===---------------------------------------------------------------------===// 217 // VOP3 Instructions 218 ///===---------------------------------------------------------------------===// 219 220 // TODO: Modifier tests 221 222 v_mad_legacy_f32 v2, v4, v6, v8 223 // SICI: v_mad_legacy_f32 v2, v4, v6, v8 ; encoding: [0x02,0x00,0x80,0xd2,0x04,0x0d,0x22,0x04] 224 // VI: v_mad_legacy_f32 v2, v4, v6, v8 ; encoding: [0x02,0x00,0xc0,0xd1,0x04,0x0d,0x22,0x04] 225 226 227 228 229 230
