| 169 | } |
| 170 | |
| 171 | void JitCompilerA64::generateProgramLight(Program& program, ProgramConfiguration& config, uint32_t datasetOffset) |
| 172 | { |
| 173 | uint32_t codePos = MainLoopBegin + 4; |
| 174 | |
| 175 | // and w16, w10, ScratchpadL3Mask64 |
| 176 | emit32(0x121A0000 | 16 | (10 << 5) | ((Log2(RANDOMX_SCRATCHPAD_L3) - 7) << 10), code, codePos); |
| 177 | |
| 178 | // and w17, w18, ScratchpadL3Mask64 |
| 179 | emit32(0x121A0000 | 17 | (18 << 5) | ((Log2(RANDOMX_SCRATCHPAD_L3) - 7) << 10), code, codePos); |
| 180 | |
| 181 | codePos = PrologueSize; |
| 182 | literalPos = ImulRcpLiteralsEnd; |
| 183 | num32bitLiterals = 0; |
| 184 | |
| 185 | for (uint32_t i = 0; i < RegistersCount; ++i) |
| 186 | reg_changed_offset[i] = codePos; |
| 187 | |
| 188 | for (uint32_t i = 0; i < program.getSize(); ++i) |
| 189 | { |
| 190 | Instruction& instr = program(i); |
| 191 | instr.src %= RegistersCount; |
| 192 | instr.dst %= RegistersCount; |
| 193 | (this->*engine[instr.opcode])(instr, codePos); |
| 194 | } |
| 195 | |
| 196 | // Update spMix2 |
| 197 | // eor w18, config.readReg2, config.readReg3 |
| 198 | emit32(ARMV8A::EOR32 | 18 | (IntRegMap[config.readReg2] << 5) | (IntRegMap[config.readReg3] << 16), code, codePos); |
| 199 | |
| 200 | // Jump back to the main loop |
| 201 | const uint32_t offset = (((uint8_t*)randomx_program_aarch64_vm_instructions_end_light) - ((uint8_t*)randomx_program_aarch64)) - codePos; |
| 202 | emit32(ARMV8A::B | (offset / 4), code, codePos); |
| 203 | |
| 204 | // and w2, w9, CacheLineAlignMask |
| 205 | codePos = (((uint8_t*)randomx_program_aarch64_light_cacheline_align_mask) - ((uint8_t*)randomx_program_aarch64)); |
| 206 | emit32(0x121A0000 | 2 | (9 << 5) | ((Log2(RANDOMX_DATASET_BASE_SIZE) - 7) << 10), code, codePos); |
| 207 | |
| 208 | // Update spMix1 |
| 209 | // eor x10, config.readReg0, config.readReg1 |
| 210 | codePos = ((uint8_t*)randomx_program_aarch64_update_spMix1) - ((uint8_t*)randomx_program_aarch64); |
| 211 | emit32(ARMV8A::EOR | 10 | (IntRegMap[config.readReg0] << 5) | (IntRegMap[config.readReg1] << 16), code, codePos); |
| 212 | |
| 213 | // Apply dataset offset |
| 214 | codePos = ((uint8_t*)randomx_program_aarch64_light_dataset_offset) - ((uint8_t*)randomx_program_aarch64); |
| 215 | |
| 216 | datasetOffset /= CacheLineSize; |
| 217 | const uint32_t imm_lo = datasetOffset & ((1 << 12) - 1); |
| 218 | const uint32_t imm_hi = datasetOffset >> 12; |
| 219 | |
| 220 | emit32(ARMV8A::ADD_IMM_LO | 2 | (2 << 5) | (imm_lo << 10), code, codePos); |
| 221 | emit32(ARMV8A::ADD_IMM_HI | 2 | (2 << 5) | (imm_hi << 10), code, codePos); |
| 222 | |
| 223 | #ifdef __GNUC__ |
| 224 | __builtin___clear_cache(reinterpret_cast<char*>(code + MainLoopBegin), reinterpret_cast<char*>(code + codePos)); |
| 225 | #endif |
| 226 | } |
| 227 | |
| 228 | template<size_t N> |