Colin Hogben / micropython

Dependents:   micropython-repl

Committer:
Colin Hogben
Date:
Wed Apr 27 22:11:29 2016 +0100
Revision:
10:33521d742af1
Parent:
0:5868e8752d44
Update README and version

Who changed what in which revision?

UserRevisionLine numberNew contents of line
pythontech 0:5868e8752d44 1 /*
pythontech 0:5868e8752d44 2 * This file is part of the Micro Python project, http://micropython.org/
pythontech 0:5868e8752d44 3 *
pythontech 0:5868e8752d44 4 * The MIT License (MIT)
pythontech 0:5868e8752d44 5 *
pythontech 0:5868e8752d44 6 * Copyright (c) 2013, 2014 Damien P. George
pythontech 0:5868e8752d44 7 *
pythontech 0:5868e8752d44 8 * Permission is hereby granted, free of charge, to any person obtaining a copy
pythontech 0:5868e8752d44 9 * of this software and associated documentation files (the "Software"), to deal
pythontech 0:5868e8752d44 10 * in the Software without restriction, including without limitation the rights
pythontech 0:5868e8752d44 11 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
pythontech 0:5868e8752d44 12 * copies of the Software, and to permit persons to whom the Software is
pythontech 0:5868e8752d44 13 * furnished to do so, subject to the following conditions:
pythontech 0:5868e8752d44 14 *
pythontech 0:5868e8752d44 15 * The above copyright notice and this permission notice shall be included in
pythontech 0:5868e8752d44 16 * all copies or substantial portions of the Software.
pythontech 0:5868e8752d44 17 *
pythontech 0:5868e8752d44 18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
pythontech 0:5868e8752d44 19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
pythontech 0:5868e8752d44 20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
pythontech 0:5868e8752d44 21 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
pythontech 0:5868e8752d44 22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
pythontech 0:5868e8752d44 23 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
pythontech 0:5868e8752d44 24 * THE SOFTWARE.
pythontech 0:5868e8752d44 25 */
pythontech 0:5868e8752d44 26
pythontech 0:5868e8752d44 27 #include <stdint.h>
pythontech 0:5868e8752d44 28 #include <stdio.h>
pythontech 0:5868e8752d44 29 #include <assert.h>
pythontech 0:5868e8752d44 30 #include <string.h>
pythontech 0:5868e8752d44 31
pythontech 0:5868e8752d44 32 #include "py/mpconfig.h"
pythontech 0:5868e8752d44 33
pythontech 0:5868e8752d44 34 // wrapper around everything in this file
pythontech 0:5868e8752d44 35 #if MICROPY_EMIT_X64
pythontech 0:5868e8752d44 36
pythontech 0:5868e8752d44 37 #include "py/asmx64.h"
pythontech 0:5868e8752d44 38
pythontech 0:5868e8752d44 39 /* all offsets are measured in multiples of 8 bytes */
pythontech 0:5868e8752d44 40 #define WORD_SIZE (8)
pythontech 0:5868e8752d44 41
pythontech 0:5868e8752d44 42 #define OPCODE_NOP (0x90)
pythontech 0:5868e8752d44 43 #define OPCODE_PUSH_R64 (0x50) /* +rq */
pythontech 0:5868e8752d44 44 #define OPCODE_PUSH_I64 (0x68)
pythontech 0:5868e8752d44 45 #define OPCODE_PUSH_M64 (0xff) /* /6 */
pythontech 0:5868e8752d44 46 #define OPCODE_POP_R64 (0x58) /* +rq */
pythontech 0:5868e8752d44 47 #define OPCODE_RET (0xc3)
pythontech 0:5868e8752d44 48 #define OPCODE_MOV_I8_TO_R8 (0xb0) /* +rb */
pythontech 0:5868e8752d44 49 #define OPCODE_MOV_I64_TO_R64 (0xb8) /* +rq */
pythontech 0:5868e8752d44 50 #define OPCODE_MOV_I32_TO_RM32 (0xc7)
pythontech 0:5868e8752d44 51 #define OPCODE_MOV_R8_TO_RM8 (0x88) /* /r */
pythontech 0:5868e8752d44 52 #define OPCODE_MOV_R64_TO_RM64 (0x89) /* /r */
pythontech 0:5868e8752d44 53 #define OPCODE_MOV_RM64_TO_R64 (0x8b) /* /r */
pythontech 0:5868e8752d44 54 #define OPCODE_MOVZX_RM8_TO_R64 (0xb6) /* 0x0f 0xb6/r */
pythontech 0:5868e8752d44 55 #define OPCODE_MOVZX_RM16_TO_R64 (0xb7) /* 0x0f 0xb7/r */
pythontech 0:5868e8752d44 56 #define OPCODE_LEA_MEM_TO_R64 (0x8d) /* /r */
pythontech 0:5868e8752d44 57 #define OPCODE_AND_R64_TO_RM64 (0x21) /* /r */
pythontech 0:5868e8752d44 58 #define OPCODE_OR_R64_TO_RM64 (0x09) /* /r */
pythontech 0:5868e8752d44 59 #define OPCODE_XOR_R64_TO_RM64 (0x31) /* /r */
pythontech 0:5868e8752d44 60 #define OPCODE_ADD_R64_TO_RM64 (0x01) /* /r */
pythontech 0:5868e8752d44 61 #define OPCODE_ADD_I32_TO_RM32 (0x81) /* /0 */
pythontech 0:5868e8752d44 62 #define OPCODE_ADD_I8_TO_RM32 (0x83) /* /0 */
pythontech 0:5868e8752d44 63 #define OPCODE_SUB_R64_FROM_RM64 (0x29)
pythontech 0:5868e8752d44 64 #define OPCODE_SUB_I32_FROM_RM64 (0x81) /* /5 */
pythontech 0:5868e8752d44 65 #define OPCODE_SUB_I8_FROM_RM64 (0x83) /* /5 */
pythontech 0:5868e8752d44 66 //#define OPCODE_SHL_RM32_BY_I8 (0xc1) /* /4 */
pythontech 0:5868e8752d44 67 //#define OPCODE_SHR_RM32_BY_I8 (0xc1) /* /5 */
pythontech 0:5868e8752d44 68 //#define OPCODE_SAR_RM32_BY_I8 (0xc1) /* /7 */
pythontech 0:5868e8752d44 69 #define OPCODE_SHL_RM64_CL (0xd3) /* /4 */
pythontech 0:5868e8752d44 70 #define OPCODE_SAR_RM64_CL (0xd3) /* /7 */
pythontech 0:5868e8752d44 71 //#define OPCODE_CMP_I32_WITH_RM32 (0x81) /* /7 */
pythontech 0:5868e8752d44 72 //#define OPCODE_CMP_I8_WITH_RM32 (0x83) /* /7 */
pythontech 0:5868e8752d44 73 #define OPCODE_CMP_R64_WITH_RM64 (0x39) /* /r */
pythontech 0:5868e8752d44 74 //#define OPCODE_CMP_RM32_WITH_R32 (0x3b)
pythontech 0:5868e8752d44 75 #define OPCODE_TEST_R8_WITH_RM8 (0x84) /* /r */
pythontech 0:5868e8752d44 76 #define OPCODE_JMP_REL8 (0xeb)
pythontech 0:5868e8752d44 77 #define OPCODE_JMP_REL32 (0xe9)
pythontech 0:5868e8752d44 78 #define OPCODE_JCC_REL8 (0x70) /* | jcc type */
pythontech 0:5868e8752d44 79 #define OPCODE_JCC_REL32_A (0x0f)
pythontech 0:5868e8752d44 80 #define OPCODE_JCC_REL32_B (0x80) /* | jcc type */
pythontech 0:5868e8752d44 81 #define OPCODE_SETCC_RM8_A (0x0f)
pythontech 0:5868e8752d44 82 #define OPCODE_SETCC_RM8_B (0x90) /* | jcc type, /0 */
pythontech 0:5868e8752d44 83 #define OPCODE_CALL_REL32 (0xe8)
pythontech 0:5868e8752d44 84 #define OPCODE_CALL_RM32 (0xff) /* /2 */
pythontech 0:5868e8752d44 85 #define OPCODE_LEAVE (0xc9)
pythontech 0:5868e8752d44 86
pythontech 0:5868e8752d44 87 #define MODRM_R64(x) (((x) & 0x7) << 3)
pythontech 0:5868e8752d44 88 #define MODRM_RM_DISP0 (0x00)
pythontech 0:5868e8752d44 89 #define MODRM_RM_DISP8 (0x40)
pythontech 0:5868e8752d44 90 #define MODRM_RM_DISP32 (0x80)
pythontech 0:5868e8752d44 91 #define MODRM_RM_REG (0xc0)
pythontech 0:5868e8752d44 92 #define MODRM_RM_R64(x) ((x) & 0x7)
pythontech 0:5868e8752d44 93
pythontech 0:5868e8752d44 94 #define OP_SIZE_PREFIX (0x66)
pythontech 0:5868e8752d44 95
pythontech 0:5868e8752d44 96 #define REX_PREFIX (0x40)
pythontech 0:5868e8752d44 97 #define REX_W (0x08) // width
pythontech 0:5868e8752d44 98 #define REX_R (0x04) // register
pythontech 0:5868e8752d44 99 #define REX_X (0x02) // index
pythontech 0:5868e8752d44 100 #define REX_B (0x01) // base
pythontech 0:5868e8752d44 101 #define REX_W_FROM_R64(r64) ((r64) >> 0 & 0x08)
pythontech 0:5868e8752d44 102 #define REX_R_FROM_R64(r64) ((r64) >> 1 & 0x04)
pythontech 0:5868e8752d44 103 #define REX_X_FROM_R64(r64) ((r64) >> 2 & 0x02)
pythontech 0:5868e8752d44 104 #define REX_B_FROM_R64(r64) ((r64) >> 3 & 0x01)
pythontech 0:5868e8752d44 105
pythontech 0:5868e8752d44 106 #define IMM32_L0(x) ((x) & 0xff)
pythontech 0:5868e8752d44 107 #define IMM32_L1(x) (((x) >> 8) & 0xff)
pythontech 0:5868e8752d44 108 #define IMM32_L2(x) (((x) >> 16) & 0xff)
pythontech 0:5868e8752d44 109 #define IMM32_L3(x) (((x) >> 24) & 0xff)
pythontech 0:5868e8752d44 110 #define IMM64_L4(x) (((x) >> 32) & 0xff)
pythontech 0:5868e8752d44 111 #define IMM64_L5(x) (((x) >> 40) & 0xff)
pythontech 0:5868e8752d44 112 #define IMM64_L6(x) (((x) >> 48) & 0xff)
pythontech 0:5868e8752d44 113 #define IMM64_L7(x) (((x) >> 56) & 0xff)
pythontech 0:5868e8752d44 114
pythontech 0:5868e8752d44 115 #define UNSIGNED_FIT8(x) (((x) & 0xffffffffffffff00) == 0)
pythontech 0:5868e8752d44 116 #define UNSIGNED_FIT32(x) (((x) & 0xffffffff00000000) == 0)
pythontech 0:5868e8752d44 117 #define SIGNED_FIT8(x) (((x) & 0xffffff80) == 0) || (((x) & 0xffffff80) == 0xffffff80)
pythontech 0:5868e8752d44 118
pythontech 0:5868e8752d44 119 struct _asm_x64_t {
pythontech 0:5868e8752d44 120 uint pass;
pythontech 0:5868e8752d44 121 mp_uint_t code_offset;
pythontech 0:5868e8752d44 122 mp_uint_t code_size;
pythontech 0:5868e8752d44 123 byte *code_base;
pythontech 0:5868e8752d44 124 byte dummy_data[8];
pythontech 0:5868e8752d44 125
pythontech 0:5868e8752d44 126 mp_uint_t max_num_labels;
pythontech 0:5868e8752d44 127 mp_uint_t *label_offsets;
pythontech 0:5868e8752d44 128 int num_locals;
pythontech 0:5868e8752d44 129 };
pythontech 0:5868e8752d44 130
pythontech 0:5868e8752d44 131 asm_x64_t *asm_x64_new(mp_uint_t max_num_labels) {
pythontech 0:5868e8752d44 132 asm_x64_t *as;
pythontech 0:5868e8752d44 133
pythontech 0:5868e8752d44 134 as = m_new0(asm_x64_t, 1);
pythontech 0:5868e8752d44 135 as->max_num_labels = max_num_labels;
pythontech 0:5868e8752d44 136 as->label_offsets = m_new(mp_uint_t, max_num_labels);
pythontech 0:5868e8752d44 137
pythontech 0:5868e8752d44 138 return as;
pythontech 0:5868e8752d44 139 }
pythontech 0:5868e8752d44 140
pythontech 0:5868e8752d44 141 void asm_x64_free(asm_x64_t *as, bool free_code) {
pythontech 0:5868e8752d44 142 if (free_code) {
pythontech 0:5868e8752d44 143 MP_PLAT_FREE_EXEC(as->code_base, as->code_size);
pythontech 0:5868e8752d44 144 }
pythontech 0:5868e8752d44 145 m_del(mp_uint_t, as->label_offsets, as->max_num_labels);
pythontech 0:5868e8752d44 146 m_del_obj(asm_x64_t, as);
pythontech 0:5868e8752d44 147 }
pythontech 0:5868e8752d44 148
pythontech 0:5868e8752d44 149 void asm_x64_start_pass(asm_x64_t *as, uint pass) {
pythontech 0:5868e8752d44 150 if (pass == ASM_X64_PASS_COMPUTE) {
pythontech 0:5868e8752d44 151 // reset all labels
pythontech 0:5868e8752d44 152 memset(as->label_offsets, -1, as->max_num_labels * sizeof(mp_uint_t));
pythontech 0:5868e8752d44 153 } if (pass == ASM_X64_PASS_EMIT) {
pythontech 0:5868e8752d44 154 MP_PLAT_ALLOC_EXEC(as->code_offset, (void**)&as->code_base, &as->code_size);
pythontech 0:5868e8752d44 155 if (as->code_base == NULL) {
pythontech 0:5868e8752d44 156 assert(0);
pythontech 0:5868e8752d44 157 }
pythontech 0:5868e8752d44 158 //printf("code_size: %u\n", as->code_size);
pythontech 0:5868e8752d44 159 }
pythontech 0:5868e8752d44 160 as->pass = pass;
pythontech 0:5868e8752d44 161 as->code_offset = 0;
pythontech 0:5868e8752d44 162 }
pythontech 0:5868e8752d44 163
pythontech 0:5868e8752d44 164 void asm_x64_end_pass(asm_x64_t *as) {
pythontech 0:5868e8752d44 165 // could check labels are resolved...
pythontech 0:5868e8752d44 166 (void)as;
pythontech 0:5868e8752d44 167 }
pythontech 0:5868e8752d44 168
pythontech 0:5868e8752d44 169 // all functions must go through this one to emit bytes
pythontech 0:5868e8752d44 170 STATIC byte *asm_x64_get_cur_to_write_bytes(asm_x64_t *as, int num_bytes_to_write) {
pythontech 0:5868e8752d44 171 //printf("emit %d\n", num_bytes_to_write);
pythontech 0:5868e8752d44 172 if (as->pass < ASM_X64_PASS_EMIT) {
pythontech 0:5868e8752d44 173 as->code_offset += num_bytes_to_write;
pythontech 0:5868e8752d44 174 return as->dummy_data;
pythontech 0:5868e8752d44 175 } else {
pythontech 0:5868e8752d44 176 assert(as->code_offset + num_bytes_to_write <= as->code_size);
pythontech 0:5868e8752d44 177 byte *c = as->code_base + as->code_offset;
pythontech 0:5868e8752d44 178 as->code_offset += num_bytes_to_write;
pythontech 0:5868e8752d44 179 return c;
pythontech 0:5868e8752d44 180 }
pythontech 0:5868e8752d44 181 }
pythontech 0:5868e8752d44 182
pythontech 0:5868e8752d44 183 mp_uint_t asm_x64_get_code_pos(asm_x64_t *as) {
pythontech 0:5868e8752d44 184 return as->code_offset;
pythontech 0:5868e8752d44 185 }
pythontech 0:5868e8752d44 186
pythontech 0:5868e8752d44 187 mp_uint_t asm_x64_get_code_size(asm_x64_t *as) {
pythontech 0:5868e8752d44 188 return as->code_size;
pythontech 0:5868e8752d44 189 }
pythontech 0:5868e8752d44 190
pythontech 0:5868e8752d44 191 void *asm_x64_get_code(asm_x64_t *as) {
pythontech 0:5868e8752d44 192 return as->code_base;
pythontech 0:5868e8752d44 193 }
pythontech 0:5868e8752d44 194
pythontech 0:5868e8752d44 195 STATIC void asm_x64_write_byte_1(asm_x64_t *as, byte b1) {
pythontech 0:5868e8752d44 196 byte* c = asm_x64_get_cur_to_write_bytes(as, 1);
pythontech 0:5868e8752d44 197 c[0] = b1;
pythontech 0:5868e8752d44 198 }
pythontech 0:5868e8752d44 199
pythontech 0:5868e8752d44 200 STATIC void asm_x64_write_byte_2(asm_x64_t *as, byte b1, byte b2) {
pythontech 0:5868e8752d44 201 byte* c = asm_x64_get_cur_to_write_bytes(as, 2);
pythontech 0:5868e8752d44 202 c[0] = b1;
pythontech 0:5868e8752d44 203 c[1] = b2;
pythontech 0:5868e8752d44 204 }
pythontech 0:5868e8752d44 205
pythontech 0:5868e8752d44 206 STATIC void asm_x64_write_byte_3(asm_x64_t *as, byte b1, byte b2, byte b3) {
pythontech 0:5868e8752d44 207 byte* c = asm_x64_get_cur_to_write_bytes(as, 3);
pythontech 0:5868e8752d44 208 c[0] = b1;
pythontech 0:5868e8752d44 209 c[1] = b2;
pythontech 0:5868e8752d44 210 c[2] = b3;
pythontech 0:5868e8752d44 211 }
pythontech 0:5868e8752d44 212
pythontech 0:5868e8752d44 213 STATIC void asm_x64_write_word32(asm_x64_t *as, int w32) {
pythontech 0:5868e8752d44 214 byte* c = asm_x64_get_cur_to_write_bytes(as, 4);
pythontech 0:5868e8752d44 215 c[0] = IMM32_L0(w32);
pythontech 0:5868e8752d44 216 c[1] = IMM32_L1(w32);
pythontech 0:5868e8752d44 217 c[2] = IMM32_L2(w32);
pythontech 0:5868e8752d44 218 c[3] = IMM32_L3(w32);
pythontech 0:5868e8752d44 219 }
pythontech 0:5868e8752d44 220
pythontech 0:5868e8752d44 221 STATIC void asm_x64_write_word64(asm_x64_t *as, int64_t w64) {
pythontech 0:5868e8752d44 222 byte* c = asm_x64_get_cur_to_write_bytes(as, 8);
pythontech 0:5868e8752d44 223 c[0] = IMM32_L0(w64);
pythontech 0:5868e8752d44 224 c[1] = IMM32_L1(w64);
pythontech 0:5868e8752d44 225 c[2] = IMM32_L2(w64);
pythontech 0:5868e8752d44 226 c[3] = IMM32_L3(w64);
pythontech 0:5868e8752d44 227 c[4] = IMM64_L4(w64);
pythontech 0:5868e8752d44 228 c[5] = IMM64_L5(w64);
pythontech 0:5868e8752d44 229 c[6] = IMM64_L6(w64);
pythontech 0:5868e8752d44 230 c[7] = IMM64_L7(w64);
pythontech 0:5868e8752d44 231 }
pythontech 0:5868e8752d44 232
pythontech 0:5868e8752d44 233 // align must be a multiple of 2
pythontech 0:5868e8752d44 234 void asm_x64_align(asm_x64_t* as, mp_uint_t align) {
pythontech 0:5868e8752d44 235 // TODO fill unused data with NOPs?
pythontech 0:5868e8752d44 236 as->code_offset = (as->code_offset + align - 1) & (~(align - 1));
pythontech 0:5868e8752d44 237 }
pythontech 0:5868e8752d44 238
pythontech 0:5868e8752d44 239 void asm_x64_data(asm_x64_t* as, mp_uint_t bytesize, mp_uint_t val) {
pythontech 0:5868e8752d44 240 byte *c = asm_x64_get_cur_to_write_bytes(as, bytesize);
pythontech 0:5868e8752d44 241 // machine is little endian
pythontech 0:5868e8752d44 242 for (uint i = 0; i < bytesize; i++) {
pythontech 0:5868e8752d44 243 *c++ = val;
pythontech 0:5868e8752d44 244 val >>= 8;
pythontech 0:5868e8752d44 245 }
pythontech 0:5868e8752d44 246 }
pythontech 0:5868e8752d44 247
pythontech 0:5868e8752d44 248 /* unused
pythontech 0:5868e8752d44 249 STATIC void asm_x64_write_word32_to(asm_x64_t *as, int offset, int w32) {
pythontech 0:5868e8752d44 250 byte* c;
pythontech 0:5868e8752d44 251 assert(offset + 4 <= as->code_size);
pythontech 0:5868e8752d44 252 c = as->code_base + offset;
pythontech 0:5868e8752d44 253 c[0] = IMM32_L0(w32);
pythontech 0:5868e8752d44 254 c[1] = IMM32_L1(w32);
pythontech 0:5868e8752d44 255 c[2] = IMM32_L2(w32);
pythontech 0:5868e8752d44 256 c[3] = IMM32_L3(w32);
pythontech 0:5868e8752d44 257 }
pythontech 0:5868e8752d44 258 */
pythontech 0:5868e8752d44 259
pythontech 0:5868e8752d44 260 STATIC void asm_x64_write_r64_disp(asm_x64_t *as, int r64, int disp_r64, int disp_offset) {
pythontech 0:5868e8752d44 261 assert(disp_r64 != ASM_X64_REG_RSP);
pythontech 0:5868e8752d44 262
pythontech 0:5868e8752d44 263 if (disp_r64 == ASM_X64_REG_R12) {
pythontech 0:5868e8752d44 264 // special case for r12; not fully implemented
pythontech 0:5868e8752d44 265 assert(SIGNED_FIT8(disp_offset));
pythontech 0:5868e8752d44 266 asm_x64_write_byte_3(as, MODRM_R64(r64) | MODRM_RM_DISP8 | MODRM_RM_R64(disp_r64), 0x24, IMM32_L0(disp_offset));
pythontech 0:5868e8752d44 267 return;
pythontech 0:5868e8752d44 268 }
pythontech 0:5868e8752d44 269
pythontech 0:5868e8752d44 270 if (disp_offset == 0 && disp_r64 != ASM_X64_REG_RBP) {
pythontech 0:5868e8752d44 271 asm_x64_write_byte_1(as, MODRM_R64(r64) | MODRM_RM_DISP0 | MODRM_RM_R64(disp_r64));
pythontech 0:5868e8752d44 272 } else if (SIGNED_FIT8(disp_offset)) {
pythontech 0:5868e8752d44 273 asm_x64_write_byte_2(as, MODRM_R64(r64) | MODRM_RM_DISP8 | MODRM_RM_R64(disp_r64), IMM32_L0(disp_offset));
pythontech 0:5868e8752d44 274 } else {
pythontech 0:5868e8752d44 275 asm_x64_write_byte_1(as, MODRM_R64(r64) | MODRM_RM_DISP32 | MODRM_RM_R64(disp_r64));
pythontech 0:5868e8752d44 276 asm_x64_write_word32(as, disp_offset);
pythontech 0:5868e8752d44 277 }
pythontech 0:5868e8752d44 278 }
pythontech 0:5868e8752d44 279
pythontech 0:5868e8752d44 280 STATIC void asm_x64_generic_r64_r64(asm_x64_t *as, int dest_r64, int src_r64, int op) {
pythontech 0:5868e8752d44 281 asm_x64_write_byte_3(as, REX_PREFIX | REX_W | REX_R_FROM_R64(src_r64) | REX_B_FROM_R64(dest_r64), op, MODRM_R64(src_r64) | MODRM_RM_REG | MODRM_RM_R64(dest_r64));
pythontech 0:5868e8752d44 282 }
pythontech 0:5868e8752d44 283
pythontech 0:5868e8752d44 284 void asm_x64_nop(asm_x64_t *as) {
pythontech 0:5868e8752d44 285 asm_x64_write_byte_1(as, OPCODE_NOP);
pythontech 0:5868e8752d44 286 }
pythontech 0:5868e8752d44 287
pythontech 0:5868e8752d44 288 void asm_x64_push_r64(asm_x64_t *as, int src_r64) {
pythontech 0:5868e8752d44 289 if (src_r64 < 8) {
pythontech 0:5868e8752d44 290 asm_x64_write_byte_1(as, OPCODE_PUSH_R64 | src_r64);
pythontech 0:5868e8752d44 291 } else {
pythontech 0:5868e8752d44 292 asm_x64_write_byte_2(as, REX_PREFIX | REX_B, OPCODE_PUSH_R64 | (src_r64 & 7));
pythontech 0:5868e8752d44 293 }
pythontech 0:5868e8752d44 294 }
pythontech 0:5868e8752d44 295
pythontech 0:5868e8752d44 296 /*
pythontech 0:5868e8752d44 297 void asm_x64_push_i32(asm_x64_t *as, int src_i32) {
pythontech 0:5868e8752d44 298 asm_x64_write_byte_1(as, OPCODE_PUSH_I64);
pythontech 0:5868e8752d44 299 asm_x64_write_word32(as, src_i32); // will be sign extended to 64 bits
pythontech 0:5868e8752d44 300 }
pythontech 0:5868e8752d44 301 */
pythontech 0:5868e8752d44 302
pythontech 0:5868e8752d44 303 /*
pythontech 0:5868e8752d44 304 void asm_x64_push_disp(asm_x64_t *as, int src_r64, int src_offset) {
pythontech 0:5868e8752d44 305 assert(src_r64 < 8);
pythontech 0:5868e8752d44 306 asm_x64_write_byte_1(as, OPCODE_PUSH_M64);
pythontech 0:5868e8752d44 307 asm_x64_write_r64_disp(as, 6, src_r64, src_offset);
pythontech 0:5868e8752d44 308 }
pythontech 0:5868e8752d44 309 */
pythontech 0:5868e8752d44 310
pythontech 0:5868e8752d44 311 void asm_x64_pop_r64(asm_x64_t *as, int dest_r64) {
pythontech 0:5868e8752d44 312 if (dest_r64 < 8) {
pythontech 0:5868e8752d44 313 asm_x64_write_byte_1(as, OPCODE_POP_R64 | dest_r64);
pythontech 0:5868e8752d44 314 } else {
pythontech 0:5868e8752d44 315 asm_x64_write_byte_2(as, REX_PREFIX | REX_B, OPCODE_POP_R64 | (dest_r64 & 7));
pythontech 0:5868e8752d44 316 }
pythontech 0:5868e8752d44 317 }
pythontech 0:5868e8752d44 318
pythontech 0:5868e8752d44 319 STATIC void asm_x64_ret(asm_x64_t *as) {
pythontech 0:5868e8752d44 320 asm_x64_write_byte_1(as, OPCODE_RET);
pythontech 0:5868e8752d44 321 }
pythontech 0:5868e8752d44 322
pythontech 0:5868e8752d44 323 void asm_x64_mov_r64_r64(asm_x64_t *as, int dest_r64, int src_r64) {
pythontech 0:5868e8752d44 324 asm_x64_generic_r64_r64(as, dest_r64, src_r64, OPCODE_MOV_R64_TO_RM64);
pythontech 0:5868e8752d44 325 }
pythontech 0:5868e8752d44 326
pythontech 0:5868e8752d44 327 void asm_x64_mov_r8_to_mem8(asm_x64_t *as, int src_r64, int dest_r64, int dest_disp) {
pythontech 0:5868e8752d44 328 if (src_r64 < 8 && dest_r64 < 8) {
pythontech 0:5868e8752d44 329 asm_x64_write_byte_1(as, OPCODE_MOV_R8_TO_RM8);
pythontech 0:5868e8752d44 330 } else {
pythontech 0:5868e8752d44 331 asm_x64_write_byte_2(as, REX_PREFIX | REX_R_FROM_R64(src_r64) | REX_B_FROM_R64(dest_r64), OPCODE_MOV_R8_TO_RM8);
pythontech 0:5868e8752d44 332 }
pythontech 0:5868e8752d44 333 asm_x64_write_r64_disp(as, src_r64, dest_r64, dest_disp);
pythontech 0:5868e8752d44 334 }
pythontech 0:5868e8752d44 335
pythontech 0:5868e8752d44 336 void asm_x64_mov_r16_to_mem16(asm_x64_t *as, int src_r64, int dest_r64, int dest_disp) {
pythontech 0:5868e8752d44 337 if (src_r64 < 8 && dest_r64 < 8) {
pythontech 0:5868e8752d44 338 asm_x64_write_byte_2(as, OP_SIZE_PREFIX, OPCODE_MOV_R64_TO_RM64);
pythontech 0:5868e8752d44 339 } else {
pythontech 0:5868e8752d44 340 asm_x64_write_byte_3(as, OP_SIZE_PREFIX, REX_PREFIX | REX_R_FROM_R64(src_r64) | REX_B_FROM_R64(dest_r64), OPCODE_MOV_R64_TO_RM64);
pythontech 0:5868e8752d44 341 }
pythontech 0:5868e8752d44 342 asm_x64_write_r64_disp(as, src_r64, dest_r64, dest_disp);
pythontech 0:5868e8752d44 343 }
pythontech 0:5868e8752d44 344
pythontech 0:5868e8752d44 345 void asm_x64_mov_r32_to_mem32(asm_x64_t *as, int src_r64, int dest_r64, int dest_disp) {
pythontech 0:5868e8752d44 346 if (src_r64 < 8 && dest_r64 < 8) {
pythontech 0:5868e8752d44 347 asm_x64_write_byte_1(as, OPCODE_MOV_R64_TO_RM64);
pythontech 0:5868e8752d44 348 } else {
pythontech 0:5868e8752d44 349 asm_x64_write_byte_2(as, REX_PREFIX | REX_R_FROM_R64(src_r64) | REX_B_FROM_R64(dest_r64), OPCODE_MOV_R64_TO_RM64);
pythontech 0:5868e8752d44 350 }
pythontech 0:5868e8752d44 351 asm_x64_write_r64_disp(as, src_r64, dest_r64, dest_disp);
pythontech 0:5868e8752d44 352 }
pythontech 0:5868e8752d44 353
pythontech 0:5868e8752d44 354 void asm_x64_mov_r64_to_mem64(asm_x64_t *as, int src_r64, int dest_r64, int dest_disp) {
pythontech 0:5868e8752d44 355 // use REX prefix for 64 bit operation
pythontech 0:5868e8752d44 356 asm_x64_write_byte_2(as, REX_PREFIX | REX_W | REX_R_FROM_R64(src_r64) | REX_B_FROM_R64(dest_r64), OPCODE_MOV_R64_TO_RM64);
pythontech 0:5868e8752d44 357 asm_x64_write_r64_disp(as, src_r64, dest_r64, dest_disp);
pythontech 0:5868e8752d44 358 }
pythontech 0:5868e8752d44 359
pythontech 0:5868e8752d44 360 void asm_x64_mov_mem8_to_r64zx(asm_x64_t *as, int src_r64, int src_disp, int dest_r64) {
pythontech 0:5868e8752d44 361 assert(src_r64 < 8);
pythontech 0:5868e8752d44 362 if (dest_r64 < 8) {
pythontech 0:5868e8752d44 363 asm_x64_write_byte_2(as, 0x0f, OPCODE_MOVZX_RM8_TO_R64);
pythontech 0:5868e8752d44 364 } else {
pythontech 0:5868e8752d44 365 asm_x64_write_byte_3(as, REX_PREFIX | REX_R, 0x0f, OPCODE_MOVZX_RM8_TO_R64);
pythontech 0:5868e8752d44 366 }
pythontech 0:5868e8752d44 367 asm_x64_write_r64_disp(as, dest_r64, src_r64, src_disp);
pythontech 0:5868e8752d44 368 }
pythontech 0:5868e8752d44 369
pythontech 0:5868e8752d44 370 void asm_x64_mov_mem16_to_r64zx(asm_x64_t *as, int src_r64, int src_disp, int dest_r64) {
pythontech 0:5868e8752d44 371 assert(src_r64 < 8);
pythontech 0:5868e8752d44 372 if (dest_r64 < 8) {
pythontech 0:5868e8752d44 373 asm_x64_write_byte_2(as, 0x0f, OPCODE_MOVZX_RM16_TO_R64);
pythontech 0:5868e8752d44 374 } else {
pythontech 0:5868e8752d44 375 asm_x64_write_byte_3(as, REX_PREFIX | REX_R, 0x0f, OPCODE_MOVZX_RM16_TO_R64);
pythontech 0:5868e8752d44 376 }
pythontech 0:5868e8752d44 377 asm_x64_write_r64_disp(as, dest_r64, src_r64, src_disp);
pythontech 0:5868e8752d44 378 }
pythontech 0:5868e8752d44 379
pythontech 0:5868e8752d44 380 void asm_x64_mov_mem32_to_r64zx(asm_x64_t *as, int src_r64, int src_disp, int dest_r64) {
pythontech 0:5868e8752d44 381 assert(src_r64 < 8);
pythontech 0:5868e8752d44 382 if (dest_r64 < 8) {
pythontech 0:5868e8752d44 383 asm_x64_write_byte_1(as, OPCODE_MOV_RM64_TO_R64);
pythontech 0:5868e8752d44 384 } else {
pythontech 0:5868e8752d44 385 asm_x64_write_byte_2(as, REX_PREFIX | REX_R, OPCODE_MOV_RM64_TO_R64);
pythontech 0:5868e8752d44 386 }
pythontech 0:5868e8752d44 387 asm_x64_write_r64_disp(as, dest_r64, src_r64, src_disp);
pythontech 0:5868e8752d44 388 }
pythontech 0:5868e8752d44 389
pythontech 0:5868e8752d44 390 void asm_x64_mov_mem64_to_r64(asm_x64_t *as, int src_r64, int src_disp, int dest_r64) {
pythontech 0:5868e8752d44 391 // use REX prefix for 64 bit operation
pythontech 0:5868e8752d44 392 asm_x64_write_byte_2(as, REX_PREFIX | REX_W | REX_R_FROM_R64(dest_r64) | REX_B_FROM_R64(src_r64), OPCODE_MOV_RM64_TO_R64);
pythontech 0:5868e8752d44 393 asm_x64_write_r64_disp(as, dest_r64, src_r64, src_disp);
pythontech 0:5868e8752d44 394 }
pythontech 0:5868e8752d44 395
pythontech 0:5868e8752d44 396 STATIC void asm_x64_lea_disp_to_r64(asm_x64_t *as, int src_r64, int src_disp, int dest_r64) {
pythontech 0:5868e8752d44 397 // use REX prefix for 64 bit operation
pythontech 0:5868e8752d44 398 assert(src_r64 < 8);
pythontech 0:5868e8752d44 399 assert(dest_r64 < 8);
pythontech 0:5868e8752d44 400 asm_x64_write_byte_2(as, REX_PREFIX | REX_W, OPCODE_LEA_MEM_TO_R64);
pythontech 0:5868e8752d44 401 asm_x64_write_r64_disp(as, dest_r64, src_r64, src_disp);
pythontech 0:5868e8752d44 402 }
pythontech 0:5868e8752d44 403
pythontech 0:5868e8752d44 404 /*
pythontech 0:5868e8752d44 405 void asm_x64_mov_i8_to_r8(asm_x64_t *as, int src_i8, int dest_r64) {
pythontech 0:5868e8752d44 406 assert(dest_r64 < 8);
pythontech 0:5868e8752d44 407 asm_x64_write_byte_2(as, OPCODE_MOV_I8_TO_R8 | dest_r64, src_i8);
pythontech 0:5868e8752d44 408 }
pythontech 0:5868e8752d44 409 */
pythontech 0:5868e8752d44 410
pythontech 0:5868e8752d44 411 STATIC void asm_x64_mov_i32_to_r64(asm_x64_t *as, int src_i32, int dest_r64) {
pythontech 0:5868e8752d44 412 // cpu defaults to i32 to r64, with zero extension
pythontech 0:5868e8752d44 413 if (dest_r64 < 8) {
pythontech 0:5868e8752d44 414 asm_x64_write_byte_1(as, OPCODE_MOV_I64_TO_R64 | dest_r64);
pythontech 0:5868e8752d44 415 } else {
pythontech 0:5868e8752d44 416 asm_x64_write_byte_2(as, REX_PREFIX | REX_B, OPCODE_MOV_I64_TO_R64 | (dest_r64 & 7));
pythontech 0:5868e8752d44 417 }
pythontech 0:5868e8752d44 418 asm_x64_write_word32(as, src_i32);
pythontech 0:5868e8752d44 419 }
pythontech 0:5868e8752d44 420
pythontech 0:5868e8752d44 421 void asm_x64_mov_i64_to_r64(asm_x64_t *as, int64_t src_i64, int dest_r64) {
pythontech 0:5868e8752d44 422 // cpu defaults to i32 to r64
pythontech 0:5868e8752d44 423 // to mov i64 to r64 need to use REX prefix
pythontech 0:5868e8752d44 424 assert(dest_r64 < 8);
pythontech 0:5868e8752d44 425 asm_x64_write_byte_2(as, REX_PREFIX | REX_W, OPCODE_MOV_I64_TO_R64 | dest_r64);
pythontech 0:5868e8752d44 426 asm_x64_write_word64(as, src_i64);
pythontech 0:5868e8752d44 427 }
pythontech 0:5868e8752d44 428
pythontech 0:5868e8752d44 429 void asm_x64_mov_i64_to_r64_optimised(asm_x64_t *as, int64_t src_i64, int dest_r64) {
pythontech 0:5868e8752d44 430 // TODO use movzx, movsx if possible
pythontech 0:5868e8752d44 431 if (UNSIGNED_FIT32(src_i64)) {
pythontech 0:5868e8752d44 432 // 5 bytes
pythontech 0:5868e8752d44 433 asm_x64_mov_i32_to_r64(as, src_i64 & 0xffffffff, dest_r64);
pythontech 0:5868e8752d44 434 } else {
pythontech 0:5868e8752d44 435 // 10 bytes
pythontech 0:5868e8752d44 436 asm_x64_mov_i64_to_r64(as, src_i64, dest_r64);
pythontech 0:5868e8752d44 437 }
pythontech 0:5868e8752d44 438 }
pythontech 0:5868e8752d44 439
pythontech 0:5868e8752d44 440 // src_i64 is stored as a full word in the code, and aligned to machine-word boundary
pythontech 0:5868e8752d44 441 void asm_x64_mov_i64_to_r64_aligned(asm_x64_t *as, int64_t src_i64, int dest_r64) {
pythontech 0:5868e8752d44 442 // mov instruction uses 2 bytes for the instruction, before the i64
pythontech 0:5868e8752d44 443 while (((as->code_offset + 2) & (WORD_SIZE - 1)) != 0) {
pythontech 0:5868e8752d44 444 asm_x64_nop(as);
pythontech 0:5868e8752d44 445 }
pythontech 0:5868e8752d44 446 asm_x64_mov_i64_to_r64(as, src_i64, dest_r64);
pythontech 0:5868e8752d44 447 }
pythontech 0:5868e8752d44 448
pythontech 0:5868e8752d44 449 void asm_x64_and_r64_r64(asm_x64_t *as, int dest_r64, int src_r64) {
pythontech 0:5868e8752d44 450 asm_x64_generic_r64_r64(as, dest_r64, src_r64, OPCODE_AND_R64_TO_RM64);
pythontech 0:5868e8752d44 451 }
pythontech 0:5868e8752d44 452
pythontech 0:5868e8752d44 453 void asm_x64_or_r64_r64(asm_x64_t *as, int dest_r64, int src_r64) {
pythontech 0:5868e8752d44 454 asm_x64_generic_r64_r64(as, dest_r64, src_r64, OPCODE_OR_R64_TO_RM64);
pythontech 0:5868e8752d44 455 }
pythontech 0:5868e8752d44 456
pythontech 0:5868e8752d44 457 void asm_x64_xor_r64_r64(asm_x64_t *as, int dest_r64, int src_r64) {
pythontech 0:5868e8752d44 458 asm_x64_generic_r64_r64(as, dest_r64, src_r64, OPCODE_XOR_R64_TO_RM64);
pythontech 0:5868e8752d44 459 }
pythontech 0:5868e8752d44 460
pythontech 0:5868e8752d44 461 void asm_x64_shl_r64_cl(asm_x64_t* as, int dest_r64) {
pythontech 0:5868e8752d44 462 asm_x64_generic_r64_r64(as, dest_r64, 4, OPCODE_SHL_RM64_CL);
pythontech 0:5868e8752d44 463 }
pythontech 0:5868e8752d44 464
pythontech 0:5868e8752d44 465 void asm_x64_sar_r64_cl(asm_x64_t* as, int dest_r64) {
pythontech 0:5868e8752d44 466 asm_x64_generic_r64_r64(as, dest_r64, 7, OPCODE_SAR_RM64_CL);
pythontech 0:5868e8752d44 467 }
pythontech 0:5868e8752d44 468
pythontech 0:5868e8752d44 469 void asm_x64_add_r64_r64(asm_x64_t *as, int dest_r64, int src_r64) {
pythontech 0:5868e8752d44 470 asm_x64_generic_r64_r64(as, dest_r64, src_r64, OPCODE_ADD_R64_TO_RM64);
pythontech 0:5868e8752d44 471 }
pythontech 0:5868e8752d44 472
pythontech 0:5868e8752d44 473 void asm_x64_sub_r64_r64(asm_x64_t *as, int dest_r64, int src_r64) {
pythontech 0:5868e8752d44 474 asm_x64_generic_r64_r64(as, dest_r64, src_r64, OPCODE_SUB_R64_FROM_RM64);
pythontech 0:5868e8752d44 475 }
pythontech 0:5868e8752d44 476
pythontech 0:5868e8752d44 477 void asm_x64_mul_r64_r64(asm_x64_t *as, int dest_r64, int src_r64) {
pythontech 0:5868e8752d44 478 // imul reg64, reg/mem64 -- 0x0f 0xaf /r
pythontech 0:5868e8752d44 479 asm_x64_write_byte_1(as, REX_PREFIX | REX_W | REX_R_FROM_R64(dest_r64) | REX_B_FROM_R64(src_r64));
pythontech 0:5868e8752d44 480 asm_x64_write_byte_3(as, 0x0f, 0xaf, MODRM_R64(dest_r64) | MODRM_RM_REG | MODRM_RM_R64(src_r64));
pythontech 0:5868e8752d44 481 }
pythontech 0:5868e8752d44 482
pythontech 0:5868e8752d44 483 /*
pythontech 0:5868e8752d44 484 void asm_x64_sub_i32_from_r32(asm_x64_t *as, int src_i32, int dest_r32) {
pythontech 0:5868e8752d44 485 if (SIGNED_FIT8(src_i32)) {
pythontech 0:5868e8752d44 486 // defaults to 32 bit operation
pythontech 0:5868e8752d44 487 asm_x64_write_byte_2(as, OPCODE_SUB_I8_FROM_RM64, MODRM_R64(5) | MODRM_RM_REG | MODRM_RM_R64(dest_r32));
pythontech 0:5868e8752d44 488 asm_x64_write_byte_1(as, src_i32 & 0xff);
pythontech 0:5868e8752d44 489 } else {
pythontech 0:5868e8752d44 490 // defaults to 32 bit operation
pythontech 0:5868e8752d44 491 asm_x64_write_byte_2(as, OPCODE_SUB_I32_FROM_RM64, MODRM_R64(5) | MODRM_RM_REG | MODRM_RM_R64(dest_r32));
pythontech 0:5868e8752d44 492 asm_x64_write_word32(as, src_i32);
pythontech 0:5868e8752d44 493 }
pythontech 0:5868e8752d44 494 }
pythontech 0:5868e8752d44 495 */
pythontech 0:5868e8752d44 496
pythontech 0:5868e8752d44 497 STATIC void asm_x64_sub_r64_i32(asm_x64_t *as, int dest_r64, int src_i32) {
pythontech 0:5868e8752d44 498 assert(dest_r64 < 8);
pythontech 0:5868e8752d44 499 if (SIGNED_FIT8(src_i32)) {
pythontech 0:5868e8752d44 500 // use REX prefix for 64 bit operation
pythontech 0:5868e8752d44 501 asm_x64_write_byte_3(as, REX_PREFIX | REX_W, OPCODE_SUB_I8_FROM_RM64, MODRM_R64(5) | MODRM_RM_REG | MODRM_RM_R64(dest_r64));
pythontech 0:5868e8752d44 502 asm_x64_write_byte_1(as, src_i32 & 0xff);
pythontech 0:5868e8752d44 503 } else {
pythontech 0:5868e8752d44 504 // use REX prefix for 64 bit operation
pythontech 0:5868e8752d44 505 asm_x64_write_byte_3(as, REX_PREFIX | REX_W, OPCODE_SUB_I32_FROM_RM64, MODRM_R64(5) | MODRM_RM_REG | MODRM_RM_R64(dest_r64));
pythontech 0:5868e8752d44 506 asm_x64_write_word32(as, src_i32);
pythontech 0:5868e8752d44 507 }
pythontech 0:5868e8752d44 508 }
pythontech 0:5868e8752d44 509
pythontech 0:5868e8752d44 510 /*
pythontech 0:5868e8752d44 511 void asm_x64_shl_r32_by_imm(asm_x64_t *as, int r32, int imm) {
pythontech 0:5868e8752d44 512 asm_x64_write_byte_2(as, OPCODE_SHL_RM32_BY_I8, MODRM_R64(4) | MODRM_RM_REG | MODRM_RM_R64(r32));
pythontech 0:5868e8752d44 513 asm_x64_write_byte_1(as, imm);
pythontech 0:5868e8752d44 514 }
pythontech 0:5868e8752d44 515
pythontech 0:5868e8752d44 516 void asm_x64_shr_r32_by_imm(asm_x64_t *as, int r32, int imm) {
pythontech 0:5868e8752d44 517 asm_x64_write_byte_2(as, OPCODE_SHR_RM32_BY_I8, MODRM_R64(5) | MODRM_RM_REG | MODRM_RM_R64(r32));
pythontech 0:5868e8752d44 518 asm_x64_write_byte_1(as, imm);
pythontech 0:5868e8752d44 519 }
pythontech 0:5868e8752d44 520
pythontech 0:5868e8752d44 521 void asm_x64_sar_r32_by_imm(asm_x64_t *as, int r32, int imm) {
pythontech 0:5868e8752d44 522 asm_x64_write_byte_2(as, OPCODE_SAR_RM32_BY_I8, MODRM_R64(7) | MODRM_RM_REG | MODRM_RM_R64(r32));
pythontech 0:5868e8752d44 523 asm_x64_write_byte_1(as, imm);
pythontech 0:5868e8752d44 524 }
pythontech 0:5868e8752d44 525 */
pythontech 0:5868e8752d44 526
pythontech 0:5868e8752d44 527 void asm_x64_cmp_r64_with_r64(asm_x64_t *as, int src_r64_a, int src_r64_b) {
pythontech 0:5868e8752d44 528 asm_x64_generic_r64_r64(as, src_r64_b, src_r64_a, OPCODE_CMP_R64_WITH_RM64);
pythontech 0:5868e8752d44 529 }
pythontech 0:5868e8752d44 530
pythontech 0:5868e8752d44 531 /*
pythontech 0:5868e8752d44 532 void asm_x64_cmp_i32_with_r32(asm_x64_t *as, int src_i32, int src_r32) {
pythontech 0:5868e8752d44 533 if (SIGNED_FIT8(src_i32)) {
pythontech 0:5868e8752d44 534 asm_x64_write_byte_2(as, OPCODE_CMP_I8_WITH_RM32, MODRM_R64(7) | MODRM_RM_REG | MODRM_RM_R64(src_r32));
pythontech 0:5868e8752d44 535 asm_x64_write_byte_1(as, src_i32 & 0xff);
pythontech 0:5868e8752d44 536 } else {
pythontech 0:5868e8752d44 537 asm_x64_write_byte_2(as, OPCODE_CMP_I32_WITH_RM32, MODRM_R64(7) | MODRM_RM_REG | MODRM_RM_R64(src_r32));
pythontech 0:5868e8752d44 538 asm_x64_write_word32(as, src_i32);
pythontech 0:5868e8752d44 539 }
pythontech 0:5868e8752d44 540 }
pythontech 0:5868e8752d44 541 */
pythontech 0:5868e8752d44 542
pythontech 0:5868e8752d44 543 void asm_x64_test_r8_with_r8(asm_x64_t *as, int src_r64_a, int src_r64_b) {
pythontech 0:5868e8752d44 544 // TODO implement for other registers
pythontech 0:5868e8752d44 545 assert(src_r64_a == ASM_X64_REG_RAX);
pythontech 0:5868e8752d44 546 assert(src_r64_b == ASM_X64_REG_RAX);
pythontech 0:5868e8752d44 547 asm_x64_write_byte_2(as, OPCODE_TEST_R8_WITH_RM8, MODRM_R64(src_r64_a) | MODRM_RM_REG | MODRM_RM_R64(src_r64_b));
pythontech 0:5868e8752d44 548 }
pythontech 0:5868e8752d44 549
pythontech 0:5868e8752d44 550 void asm_x64_setcc_r8(asm_x64_t *as, int jcc_type, int dest_r8) {
pythontech 0:5868e8752d44 551 assert(dest_r8 < 8);
pythontech 0:5868e8752d44 552 asm_x64_write_byte_3(as, OPCODE_SETCC_RM8_A, OPCODE_SETCC_RM8_B | jcc_type, MODRM_R64(0) | MODRM_RM_REG | MODRM_RM_R64(dest_r8));
pythontech 0:5868e8752d44 553 }
pythontech 0:5868e8752d44 554
pythontech 0:5868e8752d44 555 void asm_x64_label_assign(asm_x64_t *as, mp_uint_t label) {
pythontech 0:5868e8752d44 556 assert(label < as->max_num_labels);
pythontech 0:5868e8752d44 557 if (as->pass < ASM_X64_PASS_EMIT) {
pythontech 0:5868e8752d44 558 // assign label offset
pythontech 0:5868e8752d44 559 assert(as->label_offsets[label] == (mp_uint_t)-1);
pythontech 0:5868e8752d44 560 as->label_offsets[label] = as->code_offset;
pythontech 0:5868e8752d44 561 } else {
pythontech 0:5868e8752d44 562 // ensure label offset has not changed from PASS_COMPUTE to PASS_EMIT
pythontech 0:5868e8752d44 563 //printf("l%d: (at %ld=%ld)\n", label, as->label_offsets[label], as->code_offset);
pythontech 0:5868e8752d44 564 assert(as->label_offsets[label] == as->code_offset);
pythontech 0:5868e8752d44 565 }
pythontech 0:5868e8752d44 566 }
pythontech 0:5868e8752d44 567
pythontech 0:5868e8752d44 568 STATIC mp_uint_t get_label_dest(asm_x64_t *as, mp_uint_t label) {
pythontech 0:5868e8752d44 569 assert(label < as->max_num_labels);
pythontech 0:5868e8752d44 570 return as->label_offsets[label];
pythontech 0:5868e8752d44 571 }
pythontech 0:5868e8752d44 572
pythontech 0:5868e8752d44 573 void asm_x64_jmp_label(asm_x64_t *as, mp_uint_t label) {
pythontech 0:5868e8752d44 574 mp_uint_t dest = get_label_dest(as, label);
pythontech 0:5868e8752d44 575 mp_int_t rel = dest - as->code_offset;
pythontech 0:5868e8752d44 576 if (dest != (mp_uint_t)-1 && rel < 0) {
pythontech 0:5868e8752d44 577 // is a backwards jump, so we know the size of the jump on the first pass
pythontech 0:5868e8752d44 578 // calculate rel assuming 8 bit relative jump
pythontech 0:5868e8752d44 579 rel -= 2;
pythontech 0:5868e8752d44 580 if (SIGNED_FIT8(rel)) {
pythontech 0:5868e8752d44 581 asm_x64_write_byte_2(as, OPCODE_JMP_REL8, rel & 0xff);
pythontech 0:5868e8752d44 582 } else {
pythontech 0:5868e8752d44 583 rel += 2;
pythontech 0:5868e8752d44 584 goto large_jump;
pythontech 0:5868e8752d44 585 }
pythontech 0:5868e8752d44 586 } else {
pythontech 0:5868e8752d44 587 // is a forwards jump, so need to assume it's large
pythontech 0:5868e8752d44 588 large_jump:
pythontech 0:5868e8752d44 589 rel -= 5;
pythontech 0:5868e8752d44 590 asm_x64_write_byte_1(as, OPCODE_JMP_REL32);
pythontech 0:5868e8752d44 591 asm_x64_write_word32(as, rel);
pythontech 0:5868e8752d44 592 }
pythontech 0:5868e8752d44 593 }
pythontech 0:5868e8752d44 594
pythontech 0:5868e8752d44 595 void asm_x64_jcc_label(asm_x64_t *as, int jcc_type, mp_uint_t label) {
pythontech 0:5868e8752d44 596 mp_uint_t dest = get_label_dest(as, label);
pythontech 0:5868e8752d44 597 mp_int_t rel = dest - as->code_offset;
pythontech 0:5868e8752d44 598 if (dest != (mp_uint_t)-1 && rel < 0) {
pythontech 0:5868e8752d44 599 // is a backwards jump, so we know the size of the jump on the first pass
pythontech 0:5868e8752d44 600 // calculate rel assuming 8 bit relative jump
pythontech 0:5868e8752d44 601 rel -= 2;
pythontech 0:5868e8752d44 602 if (SIGNED_FIT8(rel)) {
pythontech 0:5868e8752d44 603 asm_x64_write_byte_2(as, OPCODE_JCC_REL8 | jcc_type, rel & 0xff);
pythontech 0:5868e8752d44 604 } else {
pythontech 0:5868e8752d44 605 rel += 2;
pythontech 0:5868e8752d44 606 goto large_jump;
pythontech 0:5868e8752d44 607 }
pythontech 0:5868e8752d44 608 } else {
pythontech 0:5868e8752d44 609 // is a forwards jump, so need to assume it's large
pythontech 0:5868e8752d44 610 large_jump:
pythontech 0:5868e8752d44 611 rel -= 6;
pythontech 0:5868e8752d44 612 asm_x64_write_byte_2(as, OPCODE_JCC_REL32_A, OPCODE_JCC_REL32_B | jcc_type);
pythontech 0:5868e8752d44 613 asm_x64_write_word32(as, rel);
pythontech 0:5868e8752d44 614 }
pythontech 0:5868e8752d44 615 }
pythontech 0:5868e8752d44 616
pythontech 0:5868e8752d44 617 void asm_x64_entry(asm_x64_t *as, int num_locals) {
pythontech 0:5868e8752d44 618 asm_x64_push_r64(as, ASM_X64_REG_RBP);
pythontech 0:5868e8752d44 619 asm_x64_mov_r64_r64(as, ASM_X64_REG_RBP, ASM_X64_REG_RSP);
pythontech 0:5868e8752d44 620 if (num_locals < 0) {
pythontech 0:5868e8752d44 621 num_locals = 0;
pythontech 0:5868e8752d44 622 }
pythontech 0:5868e8752d44 623 num_locals |= 1; // make it odd so stack is aligned on 16 byte boundary
pythontech 0:5868e8752d44 624 asm_x64_sub_r64_i32(as, ASM_X64_REG_RSP, num_locals * WORD_SIZE);
pythontech 0:5868e8752d44 625 asm_x64_push_r64(as, ASM_X64_REG_RBX);
pythontech 0:5868e8752d44 626 asm_x64_push_r64(as, ASM_X64_REG_R12);
pythontech 0:5868e8752d44 627 asm_x64_push_r64(as, ASM_X64_REG_R13);
pythontech 0:5868e8752d44 628 as->num_locals = num_locals;
pythontech 0:5868e8752d44 629 }
pythontech 0:5868e8752d44 630
pythontech 0:5868e8752d44 631 void asm_x64_exit(asm_x64_t *as) {
pythontech 0:5868e8752d44 632 asm_x64_pop_r64(as, ASM_X64_REG_R13);
pythontech 0:5868e8752d44 633 asm_x64_pop_r64(as, ASM_X64_REG_R12);
pythontech 0:5868e8752d44 634 asm_x64_pop_r64(as, ASM_X64_REG_RBX);
pythontech 0:5868e8752d44 635 asm_x64_write_byte_1(as, OPCODE_LEAVE);
pythontech 0:5868e8752d44 636 asm_x64_ret(as);
pythontech 0:5868e8752d44 637 }
pythontech 0:5868e8752d44 638
pythontech 0:5868e8752d44 639 // locals:
pythontech 0:5868e8752d44 640 // - stored on the stack in ascending order
pythontech 0:5868e8752d44 641 // - numbered 0 through as->num_locals-1
pythontech 0:5868e8752d44 642 // - RBP points above the last local
pythontech 0:5868e8752d44 643 //
pythontech 0:5868e8752d44 644 // | RBP
pythontech 0:5868e8752d44 645 // v
pythontech 0:5868e8752d44 646 // l0 l1 l2 ... l(n-1)
pythontech 0:5868e8752d44 647 // ^ ^
pythontech 0:5868e8752d44 648 // | low address | high address in RAM
pythontech 0:5868e8752d44 649 //
pythontech 0:5868e8752d44 650 STATIC int asm_x64_local_offset_from_ebp(asm_x64_t *as, int local_num) {
pythontech 0:5868e8752d44 651 return (-as->num_locals + local_num) * WORD_SIZE;
pythontech 0:5868e8752d44 652 }
pythontech 0:5868e8752d44 653
pythontech 0:5868e8752d44 654 void asm_x64_mov_local_to_r64(asm_x64_t *as, int src_local_num, int dest_r64) {
pythontech 0:5868e8752d44 655 asm_x64_mov_mem64_to_r64(as, ASM_X64_REG_RBP, asm_x64_local_offset_from_ebp(as, src_local_num), dest_r64);
pythontech 0:5868e8752d44 656 }
pythontech 0:5868e8752d44 657
pythontech 0:5868e8752d44 658 void asm_x64_mov_r64_to_local(asm_x64_t *as, int src_r64, int dest_local_num) {
pythontech 0:5868e8752d44 659 asm_x64_mov_r64_to_mem64(as, src_r64, ASM_X64_REG_RBP, asm_x64_local_offset_from_ebp(as, dest_local_num));
pythontech 0:5868e8752d44 660 }
pythontech 0:5868e8752d44 661
pythontech 0:5868e8752d44 662 void asm_x64_mov_local_addr_to_r64(asm_x64_t *as, int local_num, int dest_r64) {
pythontech 0:5868e8752d44 663 int offset = asm_x64_local_offset_from_ebp(as, local_num);
pythontech 0:5868e8752d44 664 if (offset == 0) {
pythontech 0:5868e8752d44 665 asm_x64_mov_r64_r64(as, dest_r64, ASM_X64_REG_RBP);
pythontech 0:5868e8752d44 666 } else {
pythontech 0:5868e8752d44 667 asm_x64_lea_disp_to_r64(as, ASM_X64_REG_RBP, offset, dest_r64);
pythontech 0:5868e8752d44 668 }
pythontech 0:5868e8752d44 669 }
pythontech 0:5868e8752d44 670
pythontech 0:5868e8752d44 671 /*
pythontech 0:5868e8752d44 672 void asm_x64_push_local(asm_x64_t *as, int local_num) {
pythontech 0:5868e8752d44 673 asm_x64_push_disp(as, ASM_X64_REG_RBP, asm_x64_local_offset_from_ebp(as, local_num));
pythontech 0:5868e8752d44 674 }
pythontech 0:5868e8752d44 675
pythontech 0:5868e8752d44 676 void asm_x64_push_local_addr(asm_x64_t *as, int local_num, int temp_r64) {
pythontech 0:5868e8752d44 677 asm_x64_mov_r64_r64(as, temp_r64, ASM_X64_REG_RBP);
pythontech 0:5868e8752d44 678 asm_x64_add_i32_to_r32(as, asm_x64_local_offset_from_ebp(as, local_num), temp_r64);
pythontech 0:5868e8752d44 679 asm_x64_push_r64(as, temp_r64);
pythontech 0:5868e8752d44 680 }
pythontech 0:5868e8752d44 681 */
pythontech 0:5868e8752d44 682
pythontech 0:5868e8752d44 683 /*
pythontech 0:5868e8752d44 684 can't use these because code might be relocated when resized
pythontech 0:5868e8752d44 685
pythontech 0:5868e8752d44 686 void asm_x64_call(asm_x64_t *as, void* func) {
pythontech 0:5868e8752d44 687 asm_x64_sub_i32_from_r32(as, 8, ASM_X64_REG_RSP);
pythontech 0:5868e8752d44 688 asm_x64_write_byte_1(as, OPCODE_CALL_REL32);
pythontech 0:5868e8752d44 689 asm_x64_write_word32(as, func - (void*)(as->code_cur + 4));
pythontech 0:5868e8752d44 690 asm_x64_mov_r64_r64(as, ASM_X64_REG_RSP, ASM_X64_REG_RBP);
pythontech 0:5868e8752d44 691 }
pythontech 0:5868e8752d44 692
pythontech 0:5868e8752d44 693 void asm_x64_call_i1(asm_x64_t *as, void* func, int i1) {
pythontech 0:5868e8752d44 694 asm_x64_sub_i32_from_r32(as, 8, ASM_X64_REG_RSP);
pythontech 0:5868e8752d44 695 asm_x64_sub_i32_from_r32(as, 12, ASM_X64_REG_RSP);
pythontech 0:5868e8752d44 696 asm_x64_push_i32(as, i1);
pythontech 0:5868e8752d44 697 asm_x64_write_byte_1(as, OPCODE_CALL_REL32);
pythontech 0:5868e8752d44 698 asm_x64_write_word32(as, func - (void*)(as->code_cur + 4));
pythontech 0:5868e8752d44 699 asm_x64_add_i32_to_r32(as, 16, ASM_X64_REG_RSP);
pythontech 0:5868e8752d44 700 asm_x64_mov_r64_r64(as, ASM_X64_REG_RSP, ASM_X64_REG_RBP);
pythontech 0:5868e8752d44 701 }
pythontech 0:5868e8752d44 702 */
pythontech 0:5868e8752d44 703
pythontech 0:5868e8752d44 704 void asm_x64_call_ind(asm_x64_t *as, void *ptr, int temp_r64) {
pythontech 0:5868e8752d44 705 assert(temp_r64 < 8);
pythontech 0:5868e8752d44 706 #ifdef __LP64__
pythontech 0:5868e8752d44 707 asm_x64_mov_i64_to_r64_optimised(as, (int64_t)ptr, temp_r64);
pythontech 0:5868e8752d44 708 #else
pythontech 0:5868e8752d44 709 // If we get here, sizeof(int) == sizeof(void*).
pythontech 0:5868e8752d44 710 asm_x64_mov_i64_to_r64_optimised(as, (int64_t)(unsigned int)ptr, temp_r64);
pythontech 0:5868e8752d44 711 #endif
pythontech 0:5868e8752d44 712 asm_x64_write_byte_2(as, OPCODE_CALL_RM32, MODRM_R64(2) | MODRM_RM_REG | MODRM_RM_R64(temp_r64));
pythontech 0:5868e8752d44 713 // this reduces code size by 2 bytes per call, but doesn't seem to speed it up at all
pythontech 0:5868e8752d44 714 // doesn't work anymore because calls are 64 bits away
pythontech 0:5868e8752d44 715 /*
pythontech 0:5868e8752d44 716 asm_x64_write_byte_1(as, OPCODE_CALL_REL32);
pythontech 0:5868e8752d44 717 asm_x64_write_word32(as, ptr - (void*)(as->code_base + as->code_offset + 4));
pythontech 0:5868e8752d44 718 */
pythontech 0:5868e8752d44 719 }
pythontech 0:5868e8752d44 720
pythontech 0:5868e8752d44 721 #endif // MICROPY_EMIT_X64