root / tcg / optimize.c @ 1ff8c541
History | View | Annotate | Download (22.6 kB)
| 1 |
/*
|
|---|---|
| 2 |
* Optimizations for Tiny Code Generator for QEMU
|
| 3 |
*
|
| 4 |
* Copyright (c) 2010 Samsung Electronics.
|
| 5 |
* Contributed by Kirill Batuzov <batuzovk@ispras.ru>
|
| 6 |
*
|
| 7 |
* Permission is hereby granted, free of charge, to any person obtaining a copy
|
| 8 |
* of this software and associated documentation files (the "Software"), to deal
|
| 9 |
* in the Software without restriction, including without limitation the rights
|
| 10 |
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
| 11 |
* copies of the Software, and to permit persons to whom the Software is
|
| 12 |
* furnished to do so, subject to the following conditions:
|
| 13 |
*
|
| 14 |
* The above copyright notice and this permission notice shall be included in
|
| 15 |
* all copies or substantial portions of the Software.
|
| 16 |
*
|
| 17 |
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 18 |
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
| 19 |
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
|
| 20 |
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
| 21 |
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 22 |
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
|
| 23 |
* THE SOFTWARE.
|
| 24 |
*/
|
| 25 |
|
| 26 |
#include "config.h" |
| 27 |
|
| 28 |
#include <stdlib.h> |
| 29 |
#include <stdio.h> |
| 30 |
|
| 31 |
#include "qemu-common.h" |
| 32 |
#include "tcg-op.h" |
| 33 |
|
| 34 |
#define CASE_OP_32_64(x) \
|
| 35 |
glue(glue(case INDEX_op_, x), _i32): \
|
| 36 |
glue(glue(case INDEX_op_, x), _i64)
|
| 37 |
|
| 38 |
typedef enum { |
| 39 |
TCG_TEMP_UNDEF = 0,
|
| 40 |
TCG_TEMP_CONST, |
| 41 |
TCG_TEMP_COPY, |
| 42 |
} tcg_temp_state; |
| 43 |
|
| 44 |
struct tcg_temp_info {
|
| 45 |
tcg_temp_state state; |
| 46 |
uint16_t prev_copy; |
| 47 |
uint16_t next_copy; |
| 48 |
tcg_target_ulong val; |
| 49 |
}; |
| 50 |
|
| 51 |
static struct tcg_temp_info temps[TCG_MAX_TEMPS]; |
| 52 |
|
| 53 |
/* Reset TEMP's state to TCG_TEMP_UNDEF. If TEMP only had one copy, remove
|
| 54 |
the copy flag from the left temp. */
|
| 55 |
static void reset_temp(TCGArg temp) |
| 56 |
{
|
| 57 |
if (temps[temp].state == TCG_TEMP_COPY) {
|
| 58 |
if (temps[temp].prev_copy == temps[temp].next_copy) {
|
| 59 |
temps[temps[temp].next_copy].state = TCG_TEMP_UNDEF; |
| 60 |
} else {
|
| 61 |
temps[temps[temp].next_copy].prev_copy = temps[temp].prev_copy; |
| 62 |
temps[temps[temp].prev_copy].next_copy = temps[temp].next_copy; |
| 63 |
} |
| 64 |
} |
| 65 |
temps[temp].state = TCG_TEMP_UNDEF; |
| 66 |
} |
| 67 |
|
| 68 |
static int op_bits(TCGOpcode op) |
| 69 |
{
|
| 70 |
const TCGOpDef *def = &tcg_op_defs[op];
|
| 71 |
return def->flags & TCG_OPF_64BIT ? 64 : 32; |
| 72 |
} |
| 73 |
|
| 74 |
static TCGOpcode op_to_movi(TCGOpcode op)
|
| 75 |
{
|
| 76 |
switch (op_bits(op)) {
|
| 77 |
case 32: |
| 78 |
return INDEX_op_movi_i32;
|
| 79 |
case 64: |
| 80 |
return INDEX_op_movi_i64;
|
| 81 |
default:
|
| 82 |
fprintf(stderr, "op_to_movi: unexpected return value of "
|
| 83 |
"function op_bits.\n");
|
| 84 |
tcg_abort(); |
| 85 |
} |
| 86 |
} |
| 87 |
|
| 88 |
static TCGArg find_better_copy(TCGContext *s, TCGArg temp)
|
| 89 |
{
|
| 90 |
TCGArg i; |
| 91 |
|
| 92 |
/* If this is already a global, we can't do better. */
|
| 93 |
if (temp < s->nb_globals) {
|
| 94 |
return temp;
|
| 95 |
} |
| 96 |
|
| 97 |
/* Search for a global first. */
|
| 98 |
for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) {
|
| 99 |
if (i < s->nb_globals) {
|
| 100 |
return i;
|
| 101 |
} |
| 102 |
} |
| 103 |
|
| 104 |
/* If it is a temp, search for a temp local. */
|
| 105 |
if (!s->temps[temp].temp_local) {
|
| 106 |
for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) {
|
| 107 |
if (s->temps[i].temp_local) {
|
| 108 |
return i;
|
| 109 |
} |
| 110 |
} |
| 111 |
} |
| 112 |
|
| 113 |
/* Failure to find a better representation, return the same temp. */
|
| 114 |
return temp;
|
| 115 |
} |
| 116 |
|
| 117 |
static bool temps_are_copies(TCGArg arg1, TCGArg arg2) |
| 118 |
{
|
| 119 |
TCGArg i; |
| 120 |
|
| 121 |
if (arg1 == arg2) {
|
| 122 |
return true; |
| 123 |
} |
| 124 |
|
| 125 |
if (temps[arg1].state != TCG_TEMP_COPY
|
| 126 |
|| temps[arg2].state != TCG_TEMP_COPY) {
|
| 127 |
return false; |
| 128 |
} |
| 129 |
|
| 130 |
for (i = temps[arg1].next_copy ; i != arg1 ; i = temps[i].next_copy) {
|
| 131 |
if (i == arg2) {
|
| 132 |
return true; |
| 133 |
} |
| 134 |
} |
| 135 |
|
| 136 |
return false; |
| 137 |
} |
| 138 |
|
| 139 |
static void tcg_opt_gen_mov(TCGContext *s, TCGArg *gen_args, |
| 140 |
TCGArg dst, TCGArg src) |
| 141 |
{
|
| 142 |
reset_temp(dst); |
| 143 |
assert(temps[src].state != TCG_TEMP_CONST); |
| 144 |
|
| 145 |
if (s->temps[src].type == s->temps[dst].type) {
|
| 146 |
if (temps[src].state != TCG_TEMP_COPY) {
|
| 147 |
temps[src].state = TCG_TEMP_COPY; |
| 148 |
temps[src].next_copy = src; |
| 149 |
temps[src].prev_copy = src; |
| 150 |
} |
| 151 |
temps[dst].state = TCG_TEMP_COPY; |
| 152 |
temps[dst].next_copy = temps[src].next_copy; |
| 153 |
temps[dst].prev_copy = src; |
| 154 |
temps[temps[dst].next_copy].prev_copy = dst; |
| 155 |
temps[src].next_copy = dst; |
| 156 |
} |
| 157 |
|
| 158 |
gen_args[0] = dst;
|
| 159 |
gen_args[1] = src;
|
| 160 |
} |
| 161 |
|
| 162 |
static void tcg_opt_gen_movi(TCGArg *gen_args, TCGArg dst, TCGArg val) |
| 163 |
{
|
| 164 |
reset_temp(dst); |
| 165 |
temps[dst].state = TCG_TEMP_CONST; |
| 166 |
temps[dst].val = val; |
| 167 |
gen_args[0] = dst;
|
| 168 |
gen_args[1] = val;
|
| 169 |
} |
| 170 |
|
| 171 |
static TCGOpcode op_to_mov(TCGOpcode op)
|
| 172 |
{
|
| 173 |
switch (op_bits(op)) {
|
| 174 |
case 32: |
| 175 |
return INDEX_op_mov_i32;
|
| 176 |
case 64: |
| 177 |
return INDEX_op_mov_i64;
|
| 178 |
default:
|
| 179 |
fprintf(stderr, "op_to_mov: unexpected return value of "
|
| 180 |
"function op_bits.\n");
|
| 181 |
tcg_abort(); |
| 182 |
} |
| 183 |
} |
| 184 |
|
| 185 |
static TCGArg do_constant_folding_2(TCGOpcode op, TCGArg x, TCGArg y)
|
| 186 |
{
|
| 187 |
switch (op) {
|
| 188 |
CASE_OP_32_64(add): |
| 189 |
return x + y;
|
| 190 |
|
| 191 |
CASE_OP_32_64(sub): |
| 192 |
return x - y;
|
| 193 |
|
| 194 |
CASE_OP_32_64(mul): |
| 195 |
return x * y;
|
| 196 |
|
| 197 |
CASE_OP_32_64(and): |
| 198 |
return x & y;
|
| 199 |
|
| 200 |
CASE_OP_32_64(or): |
| 201 |
return x | y;
|
| 202 |
|
| 203 |
CASE_OP_32_64(xor): |
| 204 |
return x ^ y;
|
| 205 |
|
| 206 |
case INDEX_op_shl_i32:
|
| 207 |
return (uint32_t)x << (uint32_t)y;
|
| 208 |
|
| 209 |
case INDEX_op_shl_i64:
|
| 210 |
return (uint64_t)x << (uint64_t)y;
|
| 211 |
|
| 212 |
case INDEX_op_shr_i32:
|
| 213 |
return (uint32_t)x >> (uint32_t)y;
|
| 214 |
|
| 215 |
case INDEX_op_shr_i64:
|
| 216 |
return (uint64_t)x >> (uint64_t)y;
|
| 217 |
|
| 218 |
case INDEX_op_sar_i32:
|
| 219 |
return (int32_t)x >> (int32_t)y;
|
| 220 |
|
| 221 |
case INDEX_op_sar_i64:
|
| 222 |
return (int64_t)x >> (int64_t)y;
|
| 223 |
|
| 224 |
case INDEX_op_rotr_i32:
|
| 225 |
x = ((uint32_t)x << (32 - y)) | ((uint32_t)x >> y);
|
| 226 |
return x;
|
| 227 |
|
| 228 |
case INDEX_op_rotr_i64:
|
| 229 |
x = ((uint64_t)x << (64 - y)) | ((uint64_t)x >> y);
|
| 230 |
return x;
|
| 231 |
|
| 232 |
case INDEX_op_rotl_i32:
|
| 233 |
x = ((uint32_t)x << y) | ((uint32_t)x >> (32 - y));
|
| 234 |
return x;
|
| 235 |
|
| 236 |
case INDEX_op_rotl_i64:
|
| 237 |
x = ((uint64_t)x << y) | ((uint64_t)x >> (64 - y));
|
| 238 |
return x;
|
| 239 |
|
| 240 |
CASE_OP_32_64(not): |
| 241 |
return ~x;
|
| 242 |
|
| 243 |
CASE_OP_32_64(neg): |
| 244 |
return -x;
|
| 245 |
|
| 246 |
CASE_OP_32_64(andc): |
| 247 |
return x & ~y;
|
| 248 |
|
| 249 |
CASE_OP_32_64(orc): |
| 250 |
return x | ~y;
|
| 251 |
|
| 252 |
CASE_OP_32_64(eqv): |
| 253 |
return ~(x ^ y);
|
| 254 |
|
| 255 |
CASE_OP_32_64(nand): |
| 256 |
return ~(x & y);
|
| 257 |
|
| 258 |
CASE_OP_32_64(nor): |
| 259 |
return ~(x | y);
|
| 260 |
|
| 261 |
CASE_OP_32_64(ext8s): |
| 262 |
return (int8_t)x;
|
| 263 |
|
| 264 |
CASE_OP_32_64(ext16s): |
| 265 |
return (int16_t)x;
|
| 266 |
|
| 267 |
CASE_OP_32_64(ext8u): |
| 268 |
return (uint8_t)x;
|
| 269 |
|
| 270 |
CASE_OP_32_64(ext16u): |
| 271 |
return (uint16_t)x;
|
| 272 |
|
| 273 |
case INDEX_op_ext32s_i64:
|
| 274 |
return (int32_t)x;
|
| 275 |
|
| 276 |
case INDEX_op_ext32u_i64:
|
| 277 |
return (uint32_t)x;
|
| 278 |
|
| 279 |
default:
|
| 280 |
fprintf(stderr, |
| 281 |
"Unrecognized operation %d in do_constant_folding.\n", op);
|
| 282 |
tcg_abort(); |
| 283 |
} |
| 284 |
} |
| 285 |
|
| 286 |
static TCGArg do_constant_folding(TCGOpcode op, TCGArg x, TCGArg y)
|
| 287 |
{
|
| 288 |
TCGArg res = do_constant_folding_2(op, x, y); |
| 289 |
if (op_bits(op) == 32) { |
| 290 |
res &= 0xffffffff;
|
| 291 |
} |
| 292 |
return res;
|
| 293 |
} |
| 294 |
|
| 295 |
static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x,
|
| 296 |
TCGArg y, TCGCond c) |
| 297 |
{
|
| 298 |
switch (op_bits(op)) {
|
| 299 |
case 32: |
| 300 |
switch (c) {
|
| 301 |
case TCG_COND_EQ:
|
| 302 |
return (uint32_t)x == (uint32_t)y;
|
| 303 |
case TCG_COND_NE:
|
| 304 |
return (uint32_t)x != (uint32_t)y;
|
| 305 |
case TCG_COND_LT:
|
| 306 |
return (int32_t)x < (int32_t)y;
|
| 307 |
case TCG_COND_GE:
|
| 308 |
return (int32_t)x >= (int32_t)y;
|
| 309 |
case TCG_COND_LE:
|
| 310 |
return (int32_t)x <= (int32_t)y;
|
| 311 |
case TCG_COND_GT:
|
| 312 |
return (int32_t)x > (int32_t)y;
|
| 313 |
case TCG_COND_LTU:
|
| 314 |
return (uint32_t)x < (uint32_t)y;
|
| 315 |
case TCG_COND_GEU:
|
| 316 |
return (uint32_t)x >= (uint32_t)y;
|
| 317 |
case TCG_COND_LEU:
|
| 318 |
return (uint32_t)x <= (uint32_t)y;
|
| 319 |
case TCG_COND_GTU:
|
| 320 |
return (uint32_t)x > (uint32_t)y;
|
| 321 |
} |
| 322 |
break;
|
| 323 |
case 64: |
| 324 |
switch (c) {
|
| 325 |
case TCG_COND_EQ:
|
| 326 |
return (uint64_t)x == (uint64_t)y;
|
| 327 |
case TCG_COND_NE:
|
| 328 |
return (uint64_t)x != (uint64_t)y;
|
| 329 |
case TCG_COND_LT:
|
| 330 |
return (int64_t)x < (int64_t)y;
|
| 331 |
case TCG_COND_GE:
|
| 332 |
return (int64_t)x >= (int64_t)y;
|
| 333 |
case TCG_COND_LE:
|
| 334 |
return (int64_t)x <= (int64_t)y;
|
| 335 |
case TCG_COND_GT:
|
| 336 |
return (int64_t)x > (int64_t)y;
|
| 337 |
case TCG_COND_LTU:
|
| 338 |
return (uint64_t)x < (uint64_t)y;
|
| 339 |
case TCG_COND_GEU:
|
| 340 |
return (uint64_t)x >= (uint64_t)y;
|
| 341 |
case TCG_COND_LEU:
|
| 342 |
return (uint64_t)x <= (uint64_t)y;
|
| 343 |
case TCG_COND_GTU:
|
| 344 |
return (uint64_t)x > (uint64_t)y;
|
| 345 |
} |
| 346 |
break;
|
| 347 |
} |
| 348 |
|
| 349 |
fprintf(stderr, |
| 350 |
"Unrecognized bitness %d or condition %d in "
|
| 351 |
"do_constant_folding_cond.\n", op_bits(op), c);
|
| 352 |
tcg_abort(); |
| 353 |
} |
| 354 |
|
| 355 |
/* Propagate constants and copies, fold constant expressions. */
|
| 356 |
static TCGArg *tcg_constant_folding(TCGContext *s, uint16_t *tcg_opc_ptr,
|
| 357 |
TCGArg *args, TCGOpDef *tcg_op_defs) |
| 358 |
{
|
| 359 |
int i, nb_ops, op_index, nb_temps, nb_globals, nb_call_args;
|
| 360 |
TCGOpcode op; |
| 361 |
const TCGOpDef *def;
|
| 362 |
TCGArg *gen_args; |
| 363 |
TCGArg tmp; |
| 364 |
TCGCond cond; |
| 365 |
|
| 366 |
/* Array VALS has an element for each temp.
|
| 367 |
If this temp holds a constant then its value is kept in VALS' element.
|
| 368 |
If this temp is a copy of other ones then the other copies are
|
| 369 |
available through the doubly linked circular list. */
|
| 370 |
|
| 371 |
nb_temps = s->nb_temps; |
| 372 |
nb_globals = s->nb_globals; |
| 373 |
memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info)); |
| 374 |
|
| 375 |
nb_ops = tcg_opc_ptr - gen_opc_buf; |
| 376 |
gen_args = args; |
| 377 |
for (op_index = 0; op_index < nb_ops; op_index++) { |
| 378 |
op = gen_opc_buf[op_index]; |
| 379 |
def = &tcg_op_defs[op]; |
| 380 |
/* Do copy propagation */
|
| 381 |
if (op == INDEX_op_call) {
|
| 382 |
int nb_oargs = args[0] >> 16; |
| 383 |
int nb_iargs = args[0] & 0xffff; |
| 384 |
for (i = nb_oargs + 1; i < nb_oargs + nb_iargs + 1; i++) { |
| 385 |
if (temps[args[i]].state == TCG_TEMP_COPY) {
|
| 386 |
args[i] = find_better_copy(s, args[i]); |
| 387 |
} |
| 388 |
} |
| 389 |
} else {
|
| 390 |
for (i = def->nb_oargs; i < def->nb_oargs + def->nb_iargs; i++) {
|
| 391 |
if (temps[args[i]].state == TCG_TEMP_COPY) {
|
| 392 |
args[i] = find_better_copy(s, args[i]); |
| 393 |
} |
| 394 |
} |
| 395 |
} |
| 396 |
|
| 397 |
/* For commutative operations make constant second argument */
|
| 398 |
switch (op) {
|
| 399 |
CASE_OP_32_64(add): |
| 400 |
CASE_OP_32_64(mul): |
| 401 |
CASE_OP_32_64(and): |
| 402 |
CASE_OP_32_64(or): |
| 403 |
CASE_OP_32_64(xor): |
| 404 |
CASE_OP_32_64(eqv): |
| 405 |
CASE_OP_32_64(nand): |
| 406 |
CASE_OP_32_64(nor): |
| 407 |
if (temps[args[1]].state == TCG_TEMP_CONST) { |
| 408 |
tmp = args[1];
|
| 409 |
args[1] = args[2]; |
| 410 |
args[2] = tmp;
|
| 411 |
} |
| 412 |
break;
|
| 413 |
CASE_OP_32_64(brcond): |
| 414 |
if (temps[args[0]].state == TCG_TEMP_CONST |
| 415 |
&& temps[args[1]].state != TCG_TEMP_CONST) {
|
| 416 |
tmp = args[0];
|
| 417 |
args[0] = args[1]; |
| 418 |
args[1] = tmp;
|
| 419 |
args[2] = tcg_swap_cond(args[2]); |
| 420 |
} |
| 421 |
break;
|
| 422 |
CASE_OP_32_64(setcond): |
| 423 |
if (temps[args[1]].state == TCG_TEMP_CONST |
| 424 |
&& temps[args[2]].state != TCG_TEMP_CONST) {
|
| 425 |
tmp = args[1];
|
| 426 |
args[1] = args[2]; |
| 427 |
args[2] = tmp;
|
| 428 |
args[3] = tcg_swap_cond(args[3]); |
| 429 |
} |
| 430 |
break;
|
| 431 |
CASE_OP_32_64(movcond): |
| 432 |
cond = args[5];
|
| 433 |
if (temps[args[1]].state == TCG_TEMP_CONST |
| 434 |
&& temps[args[2]].state != TCG_TEMP_CONST) {
|
| 435 |
tmp = args[1];
|
| 436 |
args[1] = args[2]; |
| 437 |
args[2] = tmp;
|
| 438 |
cond = tcg_swap_cond(cond); |
| 439 |
} |
| 440 |
/* For movcond, we canonicalize the "false" input reg to match
|
| 441 |
the destination reg so that the tcg backend can implement
|
| 442 |
a "move if true" operation. */
|
| 443 |
if (args[0] == args[3]) { |
| 444 |
tmp = args[3];
|
| 445 |
args[3] = args[4]; |
| 446 |
args[4] = tmp;
|
| 447 |
cond = tcg_invert_cond(cond); |
| 448 |
} |
| 449 |
args[5] = cond;
|
| 450 |
default:
|
| 451 |
break;
|
| 452 |
} |
| 453 |
|
| 454 |
/* Simplify expressions for "shift/rot r, 0, a => movi r, 0" */
|
| 455 |
switch (op) {
|
| 456 |
CASE_OP_32_64(shl): |
| 457 |
CASE_OP_32_64(shr): |
| 458 |
CASE_OP_32_64(sar): |
| 459 |
CASE_OP_32_64(rotl): |
| 460 |
CASE_OP_32_64(rotr): |
| 461 |
if (temps[args[1]].state == TCG_TEMP_CONST |
| 462 |
&& temps[args[1]].val == 0) { |
| 463 |
gen_opc_buf[op_index] = op_to_movi(op); |
| 464 |
tcg_opt_gen_movi(gen_args, args[0], 0); |
| 465 |
args += 3;
|
| 466 |
gen_args += 2;
|
| 467 |
continue;
|
| 468 |
} |
| 469 |
break;
|
| 470 |
default:
|
| 471 |
break;
|
| 472 |
} |
| 473 |
|
| 474 |
/* Simplify expression for "op r, a, 0 => mov r, a" cases */
|
| 475 |
switch (op) {
|
| 476 |
CASE_OP_32_64(add): |
| 477 |
CASE_OP_32_64(sub): |
| 478 |
CASE_OP_32_64(shl): |
| 479 |
CASE_OP_32_64(shr): |
| 480 |
CASE_OP_32_64(sar): |
| 481 |
CASE_OP_32_64(rotl): |
| 482 |
CASE_OP_32_64(rotr): |
| 483 |
CASE_OP_32_64(or): |
| 484 |
CASE_OP_32_64(xor): |
| 485 |
if (temps[args[1]].state == TCG_TEMP_CONST) { |
| 486 |
/* Proceed with possible constant folding. */
|
| 487 |
break;
|
| 488 |
} |
| 489 |
if (temps[args[2]].state == TCG_TEMP_CONST |
| 490 |
&& temps[args[2]].val == 0) { |
| 491 |
if (temps_are_copies(args[0], args[1])) { |
| 492 |
gen_opc_buf[op_index] = INDEX_op_nop; |
| 493 |
} else {
|
| 494 |
gen_opc_buf[op_index] = op_to_mov(op); |
| 495 |
tcg_opt_gen_mov(s, gen_args, args[0], args[1]); |
| 496 |
gen_args += 2;
|
| 497 |
} |
| 498 |
args += 3;
|
| 499 |
continue;
|
| 500 |
} |
| 501 |
break;
|
| 502 |
default:
|
| 503 |
break;
|
| 504 |
} |
| 505 |
|
| 506 |
/* Simplify expression for "op r, a, 0 => movi r, 0" cases */
|
| 507 |
switch (op) {
|
| 508 |
CASE_OP_32_64(and): |
| 509 |
CASE_OP_32_64(mul): |
| 510 |
if ((temps[args[2]].state == TCG_TEMP_CONST |
| 511 |
&& temps[args[2]].val == 0)) { |
| 512 |
gen_opc_buf[op_index] = op_to_movi(op); |
| 513 |
tcg_opt_gen_movi(gen_args, args[0], 0); |
| 514 |
args += 3;
|
| 515 |
gen_args += 2;
|
| 516 |
continue;
|
| 517 |
} |
| 518 |
break;
|
| 519 |
default:
|
| 520 |
break;
|
| 521 |
} |
| 522 |
|
| 523 |
/* Simplify expression for "op r, a, a => mov r, a" cases */
|
| 524 |
switch (op) {
|
| 525 |
CASE_OP_32_64(or): |
| 526 |
CASE_OP_32_64(and): |
| 527 |
if (args[1] == args[2]) { |
| 528 |
if (temps_are_copies(args[0], args[1])) { |
| 529 |
gen_opc_buf[op_index] = INDEX_op_nop; |
| 530 |
} else {
|
| 531 |
gen_opc_buf[op_index] = op_to_mov(op); |
| 532 |
tcg_opt_gen_mov(s, gen_args, args[0], args[1]); |
| 533 |
gen_args += 2;
|
| 534 |
} |
| 535 |
args += 3;
|
| 536 |
continue;
|
| 537 |
} |
| 538 |
break;
|
| 539 |
default:
|
| 540 |
break;
|
| 541 |
} |
| 542 |
|
| 543 |
/* Propagate constants through copy operations and do constant
|
| 544 |
folding. Constants will be substituted to arguments by register
|
| 545 |
allocator where needed and possible. Also detect copies. */
|
| 546 |
switch (op) {
|
| 547 |
CASE_OP_32_64(mov): |
| 548 |
if (temps_are_copies(args[0], args[1])) { |
| 549 |
args += 2;
|
| 550 |
gen_opc_buf[op_index] = INDEX_op_nop; |
| 551 |
break;
|
| 552 |
} |
| 553 |
if (temps[args[1]].state != TCG_TEMP_CONST) { |
| 554 |
tcg_opt_gen_mov(s, gen_args, args[0], args[1]); |
| 555 |
gen_args += 2;
|
| 556 |
args += 2;
|
| 557 |
break;
|
| 558 |
} |
| 559 |
/* Source argument is constant. Rewrite the operation and
|
| 560 |
let movi case handle it. */
|
| 561 |
op = op_to_movi(op); |
| 562 |
gen_opc_buf[op_index] = op; |
| 563 |
args[1] = temps[args[1]].val; |
| 564 |
/* fallthrough */
|
| 565 |
CASE_OP_32_64(movi): |
| 566 |
tcg_opt_gen_movi(gen_args, args[0], args[1]); |
| 567 |
gen_args += 2;
|
| 568 |
args += 2;
|
| 569 |
break;
|
| 570 |
CASE_OP_32_64(not): |
| 571 |
CASE_OP_32_64(neg): |
| 572 |
CASE_OP_32_64(ext8s): |
| 573 |
CASE_OP_32_64(ext8u): |
| 574 |
CASE_OP_32_64(ext16s): |
| 575 |
CASE_OP_32_64(ext16u): |
| 576 |
case INDEX_op_ext32s_i64:
|
| 577 |
case INDEX_op_ext32u_i64:
|
| 578 |
if (temps[args[1]].state == TCG_TEMP_CONST) { |
| 579 |
gen_opc_buf[op_index] = op_to_movi(op); |
| 580 |
tmp = do_constant_folding(op, temps[args[1]].val, 0); |
| 581 |
tcg_opt_gen_movi(gen_args, args[0], tmp);
|
| 582 |
} else {
|
| 583 |
reset_temp(args[0]);
|
| 584 |
gen_args[0] = args[0]; |
| 585 |
gen_args[1] = args[1]; |
| 586 |
} |
| 587 |
gen_args += 2;
|
| 588 |
args += 2;
|
| 589 |
break;
|
| 590 |
CASE_OP_32_64(add): |
| 591 |
CASE_OP_32_64(sub): |
| 592 |
CASE_OP_32_64(mul): |
| 593 |
CASE_OP_32_64(or): |
| 594 |
CASE_OP_32_64(and): |
| 595 |
CASE_OP_32_64(xor): |
| 596 |
CASE_OP_32_64(shl): |
| 597 |
CASE_OP_32_64(shr): |
| 598 |
CASE_OP_32_64(sar): |
| 599 |
CASE_OP_32_64(rotl): |
| 600 |
CASE_OP_32_64(rotr): |
| 601 |
CASE_OP_32_64(andc): |
| 602 |
CASE_OP_32_64(orc): |
| 603 |
CASE_OP_32_64(eqv): |
| 604 |
CASE_OP_32_64(nand): |
| 605 |
CASE_OP_32_64(nor): |
| 606 |
if (temps[args[1]].state == TCG_TEMP_CONST |
| 607 |
&& temps[args[2]].state == TCG_TEMP_CONST) {
|
| 608 |
gen_opc_buf[op_index] = op_to_movi(op); |
| 609 |
tmp = do_constant_folding(op, temps[args[1]].val,
|
| 610 |
temps[args[2]].val);
|
| 611 |
tcg_opt_gen_movi(gen_args, args[0], tmp);
|
| 612 |
gen_args += 2;
|
| 613 |
} else {
|
| 614 |
reset_temp(args[0]);
|
| 615 |
gen_args[0] = args[0]; |
| 616 |
gen_args[1] = args[1]; |
| 617 |
gen_args[2] = args[2]; |
| 618 |
gen_args += 3;
|
| 619 |
} |
| 620 |
args += 3;
|
| 621 |
break;
|
| 622 |
CASE_OP_32_64(setcond): |
| 623 |
if (temps[args[1]].state == TCG_TEMP_CONST |
| 624 |
&& temps[args[2]].state == TCG_TEMP_CONST) {
|
| 625 |
gen_opc_buf[op_index] = op_to_movi(op); |
| 626 |
tmp = do_constant_folding_cond(op, temps[args[1]].val,
|
| 627 |
temps[args[2]].val, args[3]); |
| 628 |
tcg_opt_gen_movi(gen_args, args[0], tmp);
|
| 629 |
gen_args += 2;
|
| 630 |
} else {
|
| 631 |
reset_temp(args[0]);
|
| 632 |
gen_args[0] = args[0]; |
| 633 |
gen_args[1] = args[1]; |
| 634 |
gen_args[2] = args[2]; |
| 635 |
gen_args[3] = args[3]; |
| 636 |
gen_args += 4;
|
| 637 |
} |
| 638 |
args += 4;
|
| 639 |
break;
|
| 640 |
CASE_OP_32_64(brcond): |
| 641 |
if (temps[args[0]].state == TCG_TEMP_CONST |
| 642 |
&& temps[args[1]].state == TCG_TEMP_CONST) {
|
| 643 |
if (do_constant_folding_cond(op, temps[args[0]].val, |
| 644 |
temps[args[1]].val, args[2])) { |
| 645 |
memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info)); |
| 646 |
gen_opc_buf[op_index] = INDEX_op_br; |
| 647 |
gen_args[0] = args[3]; |
| 648 |
gen_args += 1;
|
| 649 |
} else {
|
| 650 |
gen_opc_buf[op_index] = INDEX_op_nop; |
| 651 |
} |
| 652 |
} else {
|
| 653 |
memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info)); |
| 654 |
reset_temp(args[0]);
|
| 655 |
gen_args[0] = args[0]; |
| 656 |
gen_args[1] = args[1]; |
| 657 |
gen_args[2] = args[2]; |
| 658 |
gen_args[3] = args[3]; |
| 659 |
gen_args += 4;
|
| 660 |
} |
| 661 |
args += 4;
|
| 662 |
break;
|
| 663 |
CASE_OP_32_64(movcond): |
| 664 |
if (temps[args[1]].state == TCG_TEMP_CONST |
| 665 |
&& temps[args[2]].state == TCG_TEMP_CONST) {
|
| 666 |
tmp = do_constant_folding_cond(op, temps[args[1]].val,
|
| 667 |
temps[args[2]].val, args[5]); |
| 668 |
if (temps_are_copies(args[0], args[4-tmp])) { |
| 669 |
gen_opc_buf[op_index] = INDEX_op_nop; |
| 670 |
} else if (temps[args[4-tmp]].state == TCG_TEMP_CONST) { |
| 671 |
gen_opc_buf[op_index] = op_to_movi(op); |
| 672 |
tcg_opt_gen_movi(gen_args, args[0], temps[args[4-tmp]].val); |
| 673 |
gen_args += 2;
|
| 674 |
} else {
|
| 675 |
gen_opc_buf[op_index] = op_to_mov(op); |
| 676 |
tcg_opt_gen_mov(s, gen_args, args[0], args[4-tmp]); |
| 677 |
gen_args += 2;
|
| 678 |
} |
| 679 |
} else {
|
| 680 |
reset_temp(args[0]);
|
| 681 |
gen_args[0] = args[0]; |
| 682 |
gen_args[1] = args[1]; |
| 683 |
gen_args[2] = args[2]; |
| 684 |
gen_args[3] = args[3]; |
| 685 |
gen_args[4] = args[4]; |
| 686 |
gen_args[5] = args[5]; |
| 687 |
gen_args += 6;
|
| 688 |
} |
| 689 |
args += 6;
|
| 690 |
break;
|
| 691 |
case INDEX_op_call:
|
| 692 |
nb_call_args = (args[0] >> 16) + (args[0] & 0xffff); |
| 693 |
if (!(args[nb_call_args + 1] & (TCG_CALL_CONST | TCG_CALL_PURE))) { |
| 694 |
for (i = 0; i < nb_globals; i++) { |
| 695 |
reset_temp(i); |
| 696 |
} |
| 697 |
} |
| 698 |
for (i = 0; i < (args[0] >> 16); i++) { |
| 699 |
reset_temp(args[i + 1]);
|
| 700 |
} |
| 701 |
i = nb_call_args + 3;
|
| 702 |
while (i) {
|
| 703 |
*gen_args = *args; |
| 704 |
args++; |
| 705 |
gen_args++; |
| 706 |
i--; |
| 707 |
} |
| 708 |
break;
|
| 709 |
default:
|
| 710 |
/* Default case: we do know nothing about operation so no
|
| 711 |
propagation is done. We trash everything if the operation
|
| 712 |
is the end of a basic block, otherwise we only trash the
|
| 713 |
output args. */
|
| 714 |
if (def->flags & TCG_OPF_BB_END) {
|
| 715 |
memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info)); |
| 716 |
} else {
|
| 717 |
for (i = 0; i < def->nb_oargs; i++) { |
| 718 |
reset_temp(args[i]); |
| 719 |
} |
| 720 |
} |
| 721 |
for (i = 0; i < def->nb_args; i++) { |
| 722 |
gen_args[i] = args[i]; |
| 723 |
} |
| 724 |
args += def->nb_args; |
| 725 |
gen_args += def->nb_args; |
| 726 |
break;
|
| 727 |
} |
| 728 |
} |
| 729 |
|
| 730 |
return gen_args;
|
| 731 |
} |
| 732 |
|
| 733 |
TCGArg *tcg_optimize(TCGContext *s, uint16_t *tcg_opc_ptr, |
| 734 |
TCGArg *args, TCGOpDef *tcg_op_defs) |
| 735 |
{
|
| 736 |
TCGArg *res; |
| 737 |
res = tcg_constant_folding(s, tcg_opc_ptr, args, tcg_op_defs); |
| 738 |
return res;
|
| 739 |
} |