| 1 |
/* |
| 2 |
* Copyright © 2024 Imagination Technologies Ltd. |
| 3 |
* |
| 4 |
* based in part on asahi driver which is: |
| 5 |
* Copyright 2022 Alyssa Rosenzweig |
| 6 |
* |
| 7 |
* SPDX-License-Identifier: MIT |
| 8 |
*/ |
| 9 |
|
| 10 |
/** |
| 11 |
* \file pco_ra.c |
| 12 |
* |
| 13 |
* \brief PCO register allocator. |
| 14 |
*/ |
| 15 |
|
| 16 |
#include "hwdef/rogue_hw_utils.h" |
| 17 |
#include "pco.h" |
| 18 |
#include "pco_builder.h" |
| 19 |
#include "pco_internal.h" |
| 20 |
#include "util/bitset.h" |
| 21 |
#include "util/hash_table.h" |
| 22 |
#include "util/macros.h" |
| 23 |
#include "util/register_allocate.h" |
| 24 |
#include "util/sparse_array.h" |
| 25 |
#include "util/u_dynarray.h" |
| 26 |
|
| 27 |
#include <assert.h> |
| 28 |
#include <inttypes.h> |
| 29 |
#include <stdbool.h> |
| 30 |
#include <stdint.h> |
| 31 |
#include <stdio.h> |
| 32 |
#include <stdlib.h> |
| 33 |
|
| 34 |
/** Live range of an SSA variable. */ |
| 35 |
struct live_range { |
| 36 |
unsigned start; |
| 37 |
unsigned end; |
| 38 |
}; |
| 39 |
|
| 40 |
/** Vector override information. */ |
| 41 |
struct vec_override { |
| 42 |
pco_ref ref; |
| 43 |
unsigned offset; |
| 44 |
}; |
| 45 |
|
| 46 |
typedef struct _pco_ra_ctx { |
| 47 |
unsigned allocable_temps; |
| 48 |
unsigned allocable_vtxins; |
| 49 |
unsigned allocable_interns; |
| 50 |
|
| 51 |
unsigned temp_alloc_offset; |
| 52 |
|
| 53 |
bool spilling_setup; |
| 54 |
pco_ref spill_inst_addr_comps[2]; |
| 55 |
pco_ref spill_addr_comps[2]; |
| 56 |
pco_ref spill_data; |
| 57 |
pco_ref spill_addr; |
| 58 |
pco_ref spill_addr_data; |
| 59 |
unsigned spilled_temps; |
| 60 |
|
| 61 |
bool done; |
| 62 |
} pco_ra_ctx; |
| 63 |
|
| 64 |
/** |
| 65 |
* \brief Checks if a vec has ssa sources that are referenced more than once. |
| 66 |
* |
| 67 |
* \param[in] vec Vec instruction. |
| 68 |
* \return True if the vec has ssa sources that are referenced more than once. |
| 69 |
*/ |
| 70 |
static bool vec_has_repeated_ssas(pco_instr *vec) |
| 71 |
{ |
| 72 |
assert(vec->op == PCO_OP_VEC); |
| 73 |
|
| 74 |
pco_foreach_instr_src_ssa (psrc, vec) { |
| 75 |
pco_foreach_instr_src_ssa_from (psrc_inner, vec, psrc) { |
| 76 |
if (psrc_inner->val == psrc->val) |
| 77 |
return true; |
| 78 |
} |
| 79 |
} |
| 80 |
|
| 81 |
return false; |
| 82 |
} |
| 83 |
|
| 84 |
static void pco_extend_live_range(pco_ref origin, |
| 85 |
pco_ref current_ref, |
| 86 |
pco_instr *current_instr, |
| 87 |
struct hash_table_u64 *overrides, |
| 88 |
struct live_range *live_ranges) |
| 89 |
{ |
| 90 |
pco_foreach_instr_in_func_from (instr, current_instr) { |
| 91 |
pco_foreach_instr_src_ssa (psrc, instr) { |
| 92 |
if (current_ref.val != psrc->val) |
| 93 |
continue; |
| 94 |
|
| 95 |
pco_foreach_instr_dest_ssa (pdest, instr) { |
| 96 |
struct vec_override *override = |
| 97 |
_mesa_hash_table_u64_search(overrides, pdest->val); |
| 98 |
|
| 99 |
if (override) { |
| 100 |
live_ranges[origin.val].end = |
| 101 |
MAX2(live_ranges[origin.val].end, |
| 102 |
live_ranges[override->ref.val].end); |
| 103 |
break; |
| 104 |
} |
| 105 |
|
| 106 |
live_ranges[origin.val].end = |
| 107 |
MAX2(live_ranges[origin.val].end, instr->index); |
| 108 |
} |
| 109 |
|
| 110 |
break; |
| 111 |
} |
| 112 |
} |
| 113 |
} |
| 114 |
|
| 115 |
typedef struct _pco_use { |
| 116 |
pco_instr *instr; |
| 117 |
pco_ref *ref; |
| 118 |
} pco_use; |
| 119 |
|
| 120 |
static void preproc_vecs(pco_func *func) |
| 121 |
{ |
| 122 |
unsigned num_ssas = func->next_ssa; |
| 123 |
|
| 124 |
void *mem_ctx = ralloc_context(NULL); |
| 125 |
BITSET_WORD *multi_use_elems = rzalloc_array_size(mem_ctx, |
| 126 |
sizeof(*multi_use_elems), |
| 127 |
BITSET_WORDS(num_ssas)); |
| 128 |
struct hash_table_u64 *elem_uses = _mesa_hash_table_u64_create(mem_ctx); |
| 129 |
|
| 130 |
bool needs_reindex = false; |
| 131 |
pco_foreach_instr_in_func (instr, func) { |
| 132 |
if (instr->op != PCO_OP_VEC) |
| 133 |
continue; |
| 134 |
|
| 135 |
pco_foreach_instr_src_ssa (psrc, instr) { |
| 136 |
struct util_dynarray *uses = |
| 137 |
_mesa_hash_table_u64_search(elem_uses, psrc->val); |
| 138 |
if (!uses) { |
| 139 |
uses = rzalloc_size(elem_uses, sizeof(*uses)); |
| 140 |
util_dynarray_init(uses, uses); |
| 141 |
_mesa_hash_table_u64_insert(elem_uses, psrc->val, uses); |
| 142 |
} |
| 143 |
|
| 144 |
if (uses->size > 0) |
| 145 |
BITSET_SET(multi_use_elems, psrc->val); |
| 146 |
|
| 147 |
pco_use use = { |
| 148 |
.instr = instr, |
| 149 |
.ref = psrc, |
| 150 |
}; |
| 151 |
util_dynarray_append(uses, use); |
| 152 |
} |
| 153 |
} |
| 154 |
|
| 155 |
unsigned b; |
| 156 |
BITSET_FOREACH_SET (b, multi_use_elems, num_ssas) { |
| 157 |
struct util_dynarray *uses = _mesa_hash_table_u64_search(elem_uses, b); |
| 158 |
|
| 159 |
pco_instr *producer = NULL; |
| 160 |
pco_ref var; |
| 161 |
pco_foreach_instr_in_func (instr, func) { |
| 162 |
pco_foreach_instr_dest_ssa (pdest, instr) { |
| 163 |
if (pdest->val == b) { |
| 164 |
producer = instr; |
| 165 |
var = *pdest; |
| 166 |
break; |
| 167 |
} |
| 168 |
} |
| 169 |
|
| 170 |
if (producer) |
| 171 |
break; |
| 172 |
} |
| 173 |
assert(producer); |
| 174 |
|
| 175 |
pco_builder b = |
| 176 |
pco_builder_create(func, pco_cursor_after_instr(producer)); |
| 177 |
|
| 178 |
util_dynarray_foreach (uses, pco_use, use) { |
| 179 |
b.cursor = pco_cursor_before_instr(use->instr); |
| 180 |
if (pco_ref_get_chans(var) <= 4) { |
| 181 |
pco_ref dest = pco_ref_new_ssa_clone(func, var); |
| 182 |
pco_mbyp(&b, |
| 183 |
dest, |
| 184 |
var, |
| 185 |
.exec_cnd = pco_instr_has_exec_cnd(producer) |
| 186 |
? pco_instr_get_exec_cnd(producer) |
| 187 |
: PCO_EXEC_CND_E1_ZX, |
| 188 |
.rpt = pco_ref_get_chans(var)); |
| 189 |
*use->ref = dest; |
| 190 |
} else { |
| 191 |
assert(use->instr->op == PCO_OP_VEC && producer->op == PCO_OP_VEC); |
| 192 |
|
| 193 |
pco_instr *instr = |
| 194 |
pco_instr_create(func, |
| 195 |
1, |
| 196 |
use->instr->num_srcs + producer->num_srcs - 1); |
| 197 |
instr->op = PCO_OP_VEC; |
| 198 |
|
| 199 |
instr->dest[0] = use->instr->dest[0]; |
| 200 |
|
| 201 |
unsigned num_srcs = 0; |
| 202 |
for (unsigned s = 0; s < use->instr->num_srcs; ++s) { |
| 203 |
if (&use->instr->src[s] != use->ref) { |
| 204 |
instr->src[num_srcs++] = use->instr->src[s]; |
| 205 |
} else { |
| 206 |
for (unsigned t = 0; t < producer->num_srcs; ++t) |
| 207 |
instr->src[num_srcs++] = producer->src[t]; |
| 208 |
} |
| 209 |
} |
| 210 |
|
| 211 |
pco_instr_set_exec_cnd(instr, pco_instr_get_exec_cnd(use->instr)); |
| 212 |
pco_builder_insert_instr(&b, instr); |
| 213 |
|
| 214 |
pco_instr_delete(use->instr); |
| 215 |
needs_reindex = true; |
| 216 |
} |
| 217 |
} |
| 218 |
} |
| 219 |
|
| 220 |
ralloc_free(mem_ctx); |
| 221 |
|
| 222 |
if (needs_reindex) |
| 223 |
pco_index(func->parent_shader, false); |
| 224 |
} |
| 225 |
|
| 226 |
typedef struct _pco_copy { |
| 227 |
pco_ref src; |
| 228 |
pco_ref dest; |
| 229 |
bool s1; |
| 230 |
|
| 231 |
bool done; |
| 232 |
} pco_copy; |
| 233 |
|
| 234 |
static inline bool |
| 235 |
copy_blocked(pco_copy *copy, unsigned *temp_use_counts, unsigned lowest_temp) |
| 236 |
{ |
| 237 |
return temp_use_counts[pco_ref_get_temp(copy->dest) - lowest_temp] > 0; |
| 238 |
} |
| 239 |
|
| 240 |
static inline void |
| 241 |
do_copy(pco_builder *b, enum pco_exec_cnd exec_cnd, pco_copy *copy) |
| 242 |
{ |
| 243 |
if (copy->s1) |
| 244 |
pco_movs1(b, copy->dest, copy->src, .exec_cnd = exec_cnd); |
| 245 |
else |
| 246 |
pco_mbyp(b, copy->dest, copy->src, .exec_cnd = exec_cnd); |
| 247 |
} |
| 248 |
|
| 249 |
static inline void |
| 250 |
do_swap(pco_builder *b, enum pco_exec_cnd exec_cnd, pco_copy *copy) |
| 251 |
{ |
| 252 |
assert(!copy->s1); |
| 253 |
|
| 254 |
pco_mbyp2(b, |
| 255 |
copy->dest, |
| 256 |
pco_ref_reset_mods(copy->src), |
| 257 |
copy->src, |
| 258 |
copy->dest, |
| 259 |
.exec_cnd = exec_cnd); |
| 260 |
} |
| 261 |
|
| 262 |
static void emit_copies(pco_builder *b, |
| 263 |
struct util_dynarray *copies, |
| 264 |
enum pco_exec_cnd exec_cnd, |
| 265 |
unsigned highest_temp, |
| 266 |
unsigned lowest_temp) |
| 267 |
{ |
| 268 |
unsigned temp_range = highest_temp - lowest_temp + 1; |
| 269 |
unsigned *temp_use_counts = |
| 270 |
rzalloc_array_size(NULL, sizeof(*temp_use_counts), temp_range); |
| 271 |
pco_copy **temp_writes = |
| 272 |
rzalloc_array_size(NULL, sizeof(*temp_writes), temp_range); |
| 273 |
|
| 274 |
util_dynarray_foreach (copies, pco_copy, copy) { |
| 275 |
if (pco_ref_is_temp(copy->src)) |
| 276 |
++temp_use_counts[pco_ref_get_temp(copy->src) - lowest_temp]; |
| 277 |
|
| 278 |
temp_writes[pco_ref_get_temp(copy->dest) - lowest_temp] = copy; |
| 279 |
} |
| 280 |
|
| 281 |
bool progress = true; |
| 282 |
while (progress) { |
| 283 |
progress = false; |
| 284 |
|
| 285 |
util_dynarray_foreach (copies, pco_copy, copy) { |
| 286 |
if (!copy->done && !copy_blocked(copy, temp_use_counts, lowest_temp)) { |
| 287 |
copy->done = true; |
| 288 |
progress = true; |
| 289 |
do_copy(b, exec_cnd, copy); |
| 290 |
|
| 291 |
if (pco_ref_is_temp(copy->src)) |
| 292 |
--temp_use_counts[pco_ref_get_temp(copy->src) - lowest_temp]; |
| 293 |
|
| 294 |
temp_writes[pco_ref_get_temp(copy->dest) - lowest_temp] = NULL; |
| 295 |
} |
| 296 |
} |
| 297 |
|
| 298 |
if (progress) |
| 299 |
continue; |
| 300 |
|
| 301 |
util_dynarray_foreach (copies, pco_copy, copy) { |
| 302 |
if (copy->done) |
| 303 |
continue; |
| 304 |
|
| 305 |
if (pco_refs_are_equal(copy->src, copy->dest, true)) { |
| 306 |
copy->done = true; |
| 307 |
continue; |
| 308 |
} |
| 309 |
|
| 310 |
do_swap(b, exec_cnd, copy); |
| 311 |
copy->src = pco_ref_reset_mods(copy->src); |
| 312 |
|
| 313 |
util_dynarray_foreach (copies, pco_copy, blocking) { |
| 314 |
if (pco_ref_get_temp(blocking->src) >= |
| 315 |
pco_ref_get_temp(copy->dest) && |
| 316 |
pco_ref_get_temp(blocking->src) < |
| 317 |
(pco_ref_get_temp(copy->dest) + 1)) { |
| 318 |
blocking->src = pco_ref_offset(blocking->src, |
| 319 |
pco_ref_get_temp(copy->src) - |
| 320 |
pco_ref_get_temp(copy->dest)); |
| 321 |
} |
| 322 |
} |
| 323 |
|
| 324 |
copy->done = true; |
| 325 |
} |
| 326 |
} |
| 327 |
|
| 328 |
util_dynarray_foreach (copies, pco_copy, copy) { |
| 329 |
assert(copy->done); |
| 330 |
} |
| 331 |
|
| 332 |
ralloc_free(temp_writes); |
| 333 |
ralloc_free(temp_use_counts); |
| 334 |
} |
| 335 |
|
| 336 |
static void setup_spill_base(pco_shader *shader, |
| 337 |
pco_ref spill_inst_addr_comps[2]) |
| 338 |
{ |
| 339 |
pco_func *entry = pco_entrypoint(shader); |
| 340 |
pco_block *first_block = pco_func_first_block(entry); |
| 341 |
pco_builder b = |
| 342 |
pco_builder_create(entry, pco_cursor_before_block(first_block)); |
| 343 |
|
| 344 |
assert(shader->data.common.spill_info.count > 0); |
| 345 |
unsigned base_addr_lo_idx = shader->data.common.spill_info.start; |
| 346 |
unsigned base_addr_hi_idx = shader->data.common.spill_info.start + 1; |
| 347 |
unsigned block_size_idx = shader->data.common.spill_info.start + 2; |
| 348 |
|
| 349 |
pco_ref base_addr_lo = pco_ref_hwreg(base_addr_lo_idx, PCO_REG_CLASS_SHARED); |
| 350 |
pco_ref base_addr_hi = pco_ref_hwreg(base_addr_hi_idx, PCO_REG_CLASS_SHARED); |
| 351 |
pco_ref block_size = pco_ref_hwreg(block_size_idx, PCO_REG_CLASS_SHARED); |
| 352 |
pco_ref local_addr_inst_num = |
| 353 |
pco_ref_hwreg(PCO_SR_LOCAL_ADDR_INST_NUM, PCO_REG_CLASS_SPEC); |
| 354 |
|
| 355 |
pco_imadd64(&b, |
| 356 |
spill_inst_addr_comps[0], |
| 357 |
spill_inst_addr_comps[1], |
| 358 |
block_size, |
| 359 |
local_addr_inst_num, |
| 360 |
base_addr_lo, |
| 361 |
base_addr_hi, |
| 362 |
pco_ref_null()); |
| 363 |
} |
| 364 |
|
| 365 |
static void spill(unsigned spill_index, pco_func *func, pco_ra_ctx *ctx) |
| 366 |
{ |
| 367 |
unsigned spill_offset = ctx->spilled_temps++; |
| 368 |
|
| 369 |
pco_foreach_instr_in_func (instr, func) { |
| 370 |
pco_builder b = pco_builder_create(func, pco_cursor_before_instr(instr)); |
| 371 |
pco_foreach_instr_dest_ssa (pdest, instr) { |
| 372 |
if (pdest->val != spill_index) |
| 373 |
continue; |
| 374 |
|
| 375 |
pco_ref imm_off = pco_ref_imm32(spill_offset); |
| 376 |
pco_movi32(&b, ctx->spill_data, imm_off); |
| 377 |
pco_imadd64(&b, |
| 378 |
ctx->spill_addr_comps[0], |
| 379 |
ctx->spill_addr_comps[1], |
| 380 |
ctx->spill_data, |
| 381 |
pco_4, |
| 382 |
ctx->spill_inst_addr_comps[0], |
| 383 |
ctx->spill_inst_addr_comps[1], |
| 384 |
pco_ref_null()); |
| 385 |
|
| 386 |
/**/ |
| 387 |
|
| 388 |
*pdest = ctx->spill_data; |
| 389 |
|
| 390 |
pco_instr *next_instr = pco_next_instr(instr); |
| 391 |
if (next_instr && next_instr->op == PCO_OP_WDF) |
| 392 |
b.cursor = pco_cursor_after_instr(next_instr); |
| 393 |
else |
| 394 |
b.cursor = pco_cursor_after_instr(instr); |
| 395 |
|
| 396 |
pco_st32(&b, |
| 397 |
ctx->spill_data, |
| 398 |
pco_ref_drc(PCO_DRC_0), |
| 399 |
pco_ref_imm8(1), |
| 400 |
ctx->spill_addr_data, |
| 401 |
pco_ref_null()); |
| 402 |
|
| 403 |
pco_wdf(&b, pco_ref_drc(PCO_DRC_0)); |
| 404 |
|
| 405 |
break; |
| 406 |
} |
| 407 |
|
| 408 |
b.cursor = pco_cursor_before_instr(instr); |
| 409 |
bool load_done = false; |
| 410 |
pco_foreach_instr_src_ssa (pdest, instr) { |
| 411 |
if (pdest->val != spill_index) |
| 412 |
continue; |
| 413 |
|
| 414 |
if (!load_done) { |
| 415 |
pco_ref imm_off = pco_ref_imm32(spill_offset); |
| 416 |
pco_movi32(&b, ctx->spill_data, imm_off); |
| 417 |
pco_imadd64(&b, |
| 418 |
ctx->spill_addr_comps[0], |
| 419 |
ctx->spill_addr_comps[1], |
| 420 |
ctx->spill_data, |
| 421 |
pco_4, |
| 422 |
ctx->spill_inst_addr_comps[0], |
| 423 |
ctx->spill_inst_addr_comps[1], |
| 424 |
pco_ref_null()); |
| 425 |
|
| 426 |
pco_ld(&b, |
| 427 |
ctx->spill_data, |
| 428 |
pco_ref_drc(PCO_DRC_0), |
| 429 |
pco_ref_imm8(1), |
| 430 |
ctx->spill_addr); |
| 431 |
|
| 432 |
pco_wdf(&b, pco_ref_drc(PCO_DRC_0)); |
| 433 |
|
| 434 |
load_done = true; |
| 435 |
} |
| 436 |
|
| 437 |
*pdest = ctx->spill_data; |
| 438 |
} |
| 439 |
} |
| 440 |
|
| 441 |
pco_index(func->parent_shader, false); |
| 442 |
} |
| 443 |
|
| 444 |
/** |
| 445 |
* \brief Performs register allocation on a function. |
| 446 |
* |
| 447 |
* \param[in,out] func PCO shader. |
| 448 |
* \param[in] allocable_temps Number of allocatable temp registers. |
| 449 |
* \param[in] allocable_vtxins Number of allocatable vertex input registers. |
| 450 |
* \param[in] allocable_interns Number of allocatable internal registers. |
| 451 |
* \return True if registers were allocated. |
| 452 |
*/ |
| 453 |
static bool pco_ra_func(pco_func *func, pco_ra_ctx *ctx) |
| 454 |
{ |
| 455 |
/* TODO: support multiple functions and calls. */ |
| 456 |
assert(func->type == PCO_FUNC_TYPE_ENTRYPOINT); |
| 457 |
|
| 458 |
/* TODO: loop lifetime extension. |
| 459 |
* TODO: track successors/predecessors. |
| 460 |
*/ |
| 461 |
|
| 462 |
preproc_vecs(func); |
| 463 |
|
| 464 |
unsigned num_rsvd_vtxins = func->parent_shader->data.common.vtxins; |
| 465 |
unsigned num_ssas = func->next_ssa; |
| 466 |
unsigned num_vregs = func->next_vreg; |
| 467 |
unsigned num_vars = num_ssas + num_vregs + num_rsvd_vtxins; |
| 468 |
|
| 469 |
/* Collect used bit sizes. */ |
| 470 |
uint8_t used_bits = 0; |
| 471 |
pco_foreach_instr_in_func (instr, func) { |
| 472 |
pco_foreach_instr_dest_ssa (pdest, instr) { |
| 473 |
used_bits |= (1 << pdest->bits); |
| 474 |
} |
| 475 |
} |
| 476 |
|
| 477 |
/* vregs are always 32x1. */ |
| 478 |
if (num_vregs > 0) { |
| 479 |
used_bits |= (1 << PCO_BITS_32); |
| 480 |
} |
| 481 |
|
| 482 |
/* No registers to allocate. */ |
| 483 |
if (!used_bits) { |
| 484 |
ctx->done = true; |
| 485 |
return false; |
| 486 |
} |
| 487 |
|
| 488 |
/* 64-bit vars should've been lowered by now. */ |
| 489 |
assert(!(used_bits & (1 << PCO_BITS_64))); |
| 490 |
|
| 491 |
/* TODO: support multiple bit sizes. */ |
| 492 |
bool only_32bit = used_bits == (1 << PCO_BITS_32); |
| 493 |
assert(only_32bit); |
| 494 |
|
| 495 |
struct ra_regs *ra_regs = |
| 496 |
ra_alloc_reg_set(func, |
| 497 |
ctx->allocable_temps + ctx->allocable_vtxins, |
| 498 |
!only_32bit); |
| 499 |
|
| 500 |
BITSET_WORD *comps = |
| 501 |
rzalloc_array_size(ra_regs, sizeof(*comps), BITSET_WORDS(num_ssas)); |
| 502 |
|
| 503 |
/* Overrides for vector coalescing. */ |
| 504 |
struct hash_table_u64 *overrides = _mesa_hash_table_u64_create(ra_regs); |
| 505 |
pco_foreach_instr_in_func_rev (instr, func) { |
| 506 |
if (instr->op != PCO_OP_VEC) |
| 507 |
continue; |
| 508 |
|
| 509 |
/* Can't override vec ssa sources if they're referenced more than once. */ |
| 510 |
if (vec_has_repeated_ssas(instr)) |
| 511 |
continue; |
| 512 |
|
| 513 |
pco_ref dest = instr->dest[0]; |
| 514 |
unsigned offset = 0; |
| 515 |
|
| 516 |
struct vec_override *src_override = |
| 517 |
_mesa_hash_table_u64_search(overrides, dest.val); |
| 518 |
|
| 519 |
if (src_override) { |
| 520 |
dest = src_override->ref; |
| 521 |
offset += src_override->offset; |
| 522 |
} |
| 523 |
|
| 524 |
pco_foreach_instr_src (psrc, instr) { |
| 525 |
/* TODO: skip if vector producer is used by multiple things in a way |
| 526 |
* that doesn't allow coalescing. */ |
| 527 |
/* TODO: can NIR scalarise things so that the only remaining vectors |
| 528 |
* can be used in this way? */ |
| 529 |
|
| 530 |
if (pco_ref_is_ssa(*psrc)) { |
| 531 |
/* Make sure this hasn't already been overridden somewhere else! */ |
| 532 |
#if 1 |
| 533 |
if (_mesa_hash_table_u64_search(overrides, psrc->val)) { |
| 534 |
BITSET_SET(comps, psrc->val); |
| 535 |
continue; |
| 536 |
} |
| 537 |
#else |
| 538 |
assert(!_mesa_hash_table_u64_search(overrides, psrc->val)); |
| 539 |
#endif |
| 540 |
|
| 541 |
struct vec_override *src_override = |
| 542 |
rzalloc_size(overrides, sizeof(*src_override)); |
| 543 |
src_override->ref = dest; |
| 544 |
src_override->offset = offset; |
| 545 |
|
| 546 |
_mesa_hash_table_u64_insert(overrides, psrc->val, src_override); |
| 547 |
} |
| 548 |
|
| 549 |
offset += pco_ref_get_chans(*psrc); |
| 550 |
} |
| 551 |
} |
| 552 |
|
| 553 |
/* Overrides for vector component uses. */ |
| 554 |
pco_foreach_instr_in_func (instr, func) { |
| 555 |
if (instr->op != PCO_OP_COMP) |
| 556 |
continue; |
| 557 |
|
| 558 |
pco_ref dest = instr->dest[0]; |
| 559 |
pco_ref src = instr->src[0]; |
| 560 |
unsigned offset = pco_ref_get_imm(instr->src[1]); |
| 561 |
|
| 562 |
BITSET_SET(comps, dest.val); |
| 563 |
|
| 564 |
assert(pco_ref_is_ssa(src)); |
| 565 |
assert(pco_ref_is_ssa(dest)); |
| 566 |
|
| 567 |
struct vec_override *vec_override = |
| 568 |
_mesa_hash_table_u64_search(overrides, src.val); |
| 569 |
|
| 570 |
if (vec_override) { |
| 571 |
src = vec_override->ref; |
| 572 |
offset += vec_override->offset; |
| 573 |
} |
| 574 |
|
| 575 |
struct vec_override *src_override = |
| 576 |
rzalloc_size(overrides, sizeof(*src_override)); |
| 577 |
src_override->ref = src; |
| 578 |
src_override->offset = offset; |
| 579 |
_mesa_hash_table_u64_insert(overrides, dest.val, src_override); |
| 580 |
} |
| 581 |
|
| 582 |
/* Allocate classes. */ |
| 583 |
struct hash_table_u64 *ra_classes = _mesa_hash_table_u64_create(ra_regs); |
| 584 |
pco_foreach_instr_in_func (instr, func) { |
| 585 |
pco_foreach_instr_dest_ssa (pdest, instr) { |
| 586 |
unsigned chans = pco_ref_get_chans(*pdest); |
| 587 |
/* TODO: bitset instead of search? */ |
| 588 |
if (_mesa_hash_table_u64_search(ra_classes, chans)) |
| 589 |
continue; |
| 590 |
|
| 591 |
/* Skip if collated. */ |
| 592 |
if (_mesa_hash_table_u64_search(overrides, pdest->val)) |
| 593 |
continue; |
| 594 |
|
| 595 |
struct ra_class *ra_class = ra_alloc_contig_reg_class(ra_regs, chans); |
| 596 |
_mesa_hash_table_u64_insert(ra_classes, chans, ra_class); |
| 597 |
} |
| 598 |
} |
| 599 |
|
| 600 |
/* vregs are always 32x1. */ |
| 601 |
if (num_vregs > 0) { |
| 602 |
if (!_mesa_hash_table_u64_search(ra_classes, 1)) { |
| 603 |
struct ra_class *ra_class = ra_alloc_contig_reg_class(ra_regs, 1); |
| 604 |
_mesa_hash_table_u64_insert(ra_classes, 1, ra_class); |
| 605 |
} |
| 606 |
} |
| 607 |
|
| 608 |
/* Assign registers to classes. */ |
| 609 |
hash_table_u64_foreach (ra_classes, entry) { |
| 610 |
const unsigned stride = entry.key; |
| 611 |
struct ra_class *ra_class = entry.data; |
| 612 |
|
| 613 |
for (unsigned t = 0; t < ctx->allocable_temps - (stride - 1); ++t) |
| 614 |
ra_class_add_reg(ra_class, t); |
| 615 |
|
| 616 |
for (unsigned t = ctx->allocable_temps; |
| 617 |
t < ctx->allocable_temps + ctx->allocable_vtxins - (stride - 1); |
| 618 |
++t) |
| 619 |
ra_class_add_reg(ra_class, t); |
| 620 |
} |
| 621 |
|
| 622 |
ra_set_finalize(ra_regs, NULL); |
| 623 |
|
| 624 |
struct ra_graph *ra_graph = ra_alloc_interference_graph(ra_regs, num_vars); |
| 625 |
ralloc_steal(ra_regs, ra_graph); |
| 626 |
|
| 627 |
/* Allocate and calculate live ranges. */ |
| 628 |
struct live_range *live_ranges = |
| 629 |
rzalloc_array_size(ra_regs, sizeof(*live_ranges), num_vars); |
| 630 |
|
| 631 |
for (unsigned u = 0; u < num_vars; ++u) |
| 632 |
live_ranges[u].start = ~0U; |
| 633 |
|
| 634 |
pco_foreach_instr_in_func (instr, func) { |
| 635 |
pco_foreach_instr_dest_ssa (pdest, instr) { |
| 636 |
pco_ref dest = *pdest; |
| 637 |
struct vec_override *override = |
| 638 |
_mesa_hash_table_u64_search(overrides, dest.val); |
| 639 |
|
| 640 |
if (override) |
| 641 |
dest = override->ref; |
| 642 |
|
| 643 |
live_ranges[dest.val].start = |
| 644 |
MIN2(live_ranges[dest.val].start, instr->index); |
| 645 |
|
| 646 |
if (override) |
| 647 |
continue; |
| 648 |
|
| 649 |
/* Set class if it hasn't already been set up in an override. */ |
| 650 |
unsigned chans = pco_ref_get_chans(dest); |
| 651 |
struct ra_class *ra_class = |
| 652 |
_mesa_hash_table_u64_search(ra_classes, chans); |
| 653 |
assert(ra_class); |
| 654 |
|
| 655 |
ra_set_node_class(ra_graph, dest.val, ra_class); |
| 656 |
} |
| 657 |
|
| 658 |
pco_foreach_instr_src_ssa (psrc, instr) { |
| 659 |
pco_ref src = *psrc; |
| 660 |
struct vec_override *override = |
| 661 |
_mesa_hash_table_u64_search(overrides, src.val); |
| 662 |
|
| 663 |
if (override) |
| 664 |
src = override->ref; |
| 665 |
|
| 666 |
live_ranges[src.val].end = |
| 667 |
MAX2(live_ranges[src.val].end, instr->index); |
| 668 |
} |
| 669 |
|
| 670 |
pco_foreach_instr_dest_vreg (pdest, instr) { |
| 671 |
pco_ref dest = *pdest; |
| 672 |
|
| 673 |
/* Place vregs after ssa vars. */ |
| 674 |
dest.val += num_ssas; |
| 675 |
|
| 676 |
live_ranges[dest.val].start = |
| 677 |
MIN2(live_ranges[dest.val].start, instr->index); |
| 678 |
|
| 679 |
/* Set class if it hasn't already been set up in an override. */ |
| 680 |
unsigned chans = pco_ref_get_chans(dest); |
| 681 |
struct ra_class *ra_class = |
| 682 |
_mesa_hash_table_u64_search(ra_classes, chans); |
| 683 |
assert(ra_class); |
| 684 |
|
| 685 |
ra_set_node_class(ra_graph, dest.val, ra_class); |
| 686 |
} |
| 687 |
|
| 688 |
pco_foreach_instr_src_vreg (psrc, instr) { |
| 689 |
pco_ref src = *psrc; |
| 690 |
|
| 691 |
/* Place vregs after ssa vars. */ |
| 692 |
src.val += num_ssas; |
| 693 |
|
| 694 |
live_ranges[src.val].end = |
| 695 |
MAX2(live_ranges[src.val].end, instr->index); |
| 696 |
} |
| 697 |
|
| 698 |
/* Ensure that vertex input registers with pre-initialised data are not |
| 699 |
* clobbered too early. |
| 700 |
*/ |
| 701 |
if (ctx->allocable_vtxins > 0) { |
| 702 |
pco_foreach_instr_src_vtxin_reg (psrc, instr) { |
| 703 |
pco_ref src = *psrc; |
| 704 |
|
| 705 |
/* Place vtxin regs after ssa vars and vregs. */ |
| 706 |
src.val += num_ssas + num_vregs; |
| 707 |
|
| 708 |
live_ranges[src.val].end = |
| 709 |
MAX2(live_ranges[src.val].end, instr->index); |
| 710 |
live_ranges[src.val].start = 0; |
| 711 |
|
| 712 |
ra_set_node_reg(ra_graph, |
| 713 |
src.val, |
| 714 |
psrc->val + ctx->allocable_temps); |
| 715 |
} |
| 716 |
} |
| 717 |
} |
| 718 |
|
| 719 |
/* Extend lifetimes of non-overriden vecs that have comp instructions. */ |
| 720 |
pco_foreach_instr_in_func (instr, func) { |
| 721 |
if (instr->op != PCO_OP_COMP) |
| 722 |
continue; |
| 723 |
|
| 724 |
pco_ref dest = instr->dest[0]; |
| 725 |
pco_ref src_vec = instr->src[0]; |
| 726 |
|
| 727 |
struct vec_override *vec_override = |
| 728 |
_mesa_hash_table_u64_search(overrides, src_vec.val); |
| 729 |
|
| 730 |
/* Already taken care of. */ |
| 731 |
if (vec_override) { |
| 732 |
assert(live_ranges[src_vec.val].start == ~0U && |
| 733 |
live_ranges[src_vec.val].end == 0); |
| 734 |
continue; |
| 735 |
} |
| 736 |
|
| 737 |
pco_extend_live_range(src_vec, dest, instr, overrides, live_ranges); |
| 738 |
} |
| 739 |
|
| 740 |
/* Extend lifetimes of vars in loops. */ |
| 741 |
pco_foreach_loop_in_func (loop, func) { |
| 742 |
pco_block *prologue_block = |
| 743 |
pco_cf_node_as_block(pco_cf_node_head(&loop->prologue)); |
| 744 |
|
| 745 |
pco_block *epilogue_block = |
| 746 |
pco_cf_node_as_block(pco_cf_node_tail(&loop->epilogue)); |
| 747 |
|
| 748 |
unsigned loop_start_index = pco_first_instr(prologue_block)->index; |
| 749 |
unsigned loop_end_index = pco_last_instr(epilogue_block)->index; |
| 750 |
|
| 751 |
/* If a var is defined before a loop and stops being used during it, |
| 752 |
* extend its lifetime to the end of the loop. |
| 753 |
*/ |
| 754 |
|
| 755 |
for (unsigned var = 0; var < num_vars; ++var) { |
| 756 |
if (live_ranges[var].start < loop_start_index && |
| 757 |
live_ranges[var].end > loop_start_index && |
| 758 |
live_ranges[var].end < loop_end_index) { |
| 759 |
live_ranges[var].end = loop_end_index; |
| 760 |
} |
| 761 |
} |
| 762 |
} |
| 763 |
|
| 764 |
/* If there are instructions left with any unused dests that aren't/couldn't |
| 765 |
* be DCEd (e.g. because of side effects), ensure their range ends are setup |
| 766 |
* to avoid missing overlaps and clobbering regs. |
| 767 |
*/ |
| 768 |
for (unsigned var = 0; var < num_vars; ++var) |
| 769 |
if (live_ranges[var].start != ~0U && live_ranges[var].end == 0) |
| 770 |
live_ranges[var].end = live_ranges[var].start; |
| 771 |
|
| 772 |
/* Build interference graph from overlapping live ranges. */ |
| 773 |
for (unsigned var0 = 0; var0 < num_vars; ++var0) { |
| 774 |
for (unsigned var1 = var0 + 1; var1 < num_vars; ++var1) { |
| 775 |
/* If the live ranges overlap, the register nodes interfere. */ |
| 776 |
if ((live_ranges[var0].start != ~0U && live_ranges[var1].end != ~0U) && |
| 777 |
!(live_ranges[var0].start >= live_ranges[var1].end || |
| 778 |
live_ranges[var1].start >= live_ranges[var0].end)) { |
| 779 |
ra_add_node_interference(ra_graph, var0, var1); |
| 780 |
} |
| 781 |
} |
| 782 |
} |
| 783 |
|
| 784 |
pco_foreach_instr_in_func_rev (vec, func) { |
| 785 |
if (vec->op != PCO_OP_VEC) |
| 786 |
continue; |
| 787 |
|
| 788 |
pco_foreach_instr_src_ssa (psrc, vec) { |
| 789 |
ra_add_node_interference(ra_graph, vec->dest[0].val, psrc->val); |
| 790 |
} |
| 791 |
} |
| 792 |
|
| 793 |
/* Make srcs and dests interfere for instructions with repeat > 1. */ |
| 794 |
pco_foreach_instr_in_func_rev (instr, func) { |
| 795 |
if (!pco_instr_has_rpt(instr)) |
| 796 |
continue; |
| 797 |
|
| 798 |
if (pco_instr_get_rpt(instr) < 2) |
| 799 |
continue; |
| 800 |
|
| 801 |
pco_foreach_instr_dest_ssa (pdest, instr) { |
| 802 |
pco_foreach_instr_src_ssa (psrc, instr) { |
| 803 |
ra_add_node_interference(ra_graph, pdest->val, psrc->val); |
| 804 |
} |
| 805 |
} |
| 806 |
} |
| 807 |
|
| 808 |
bool allocated = ra_allocate(ra_graph); |
| 809 |
bool force_spill = false; |
| 810 |
if (!allocated || force_spill) { |
| 811 |
if (!ctx->spilling_setup) { |
| 812 |
ctx->spill_inst_addr_comps[0] = pco_ref_hwreg(0, PCO_REG_CLASS_TEMP); |
| 813 |
ctx->spill_inst_addr_comps[1] = pco_ref_hwreg(1, PCO_REG_CLASS_TEMP); |
| 814 |
|
| 815 |
ctx->spill_addr_comps[0] = pco_ref_hwreg(2, PCO_REG_CLASS_TEMP); |
| 816 |
ctx->spill_addr_comps[1] = pco_ref_hwreg(3, PCO_REG_CLASS_TEMP); |
| 817 |
|
| 818 |
ctx->spill_data = pco_ref_hwreg(4, PCO_REG_CLASS_TEMP); |
| 819 |
|
| 820 |
ctx->spill_addr = pco_ref_hwreg_vec(2, PCO_REG_CLASS_TEMP, 2); |
| 821 |
ctx->spill_addr_data = pco_ref_hwreg_vec(2, PCO_REG_CLASS_TEMP, 3); |
| 822 |
|
| 823 |
ctx->allocable_temps -= 5; |
| 824 |
ctx->temp_alloc_offset = 5; |
| 825 |
|
| 826 |
setup_spill_base(func->parent_shader, ctx->spill_inst_addr_comps); |
| 827 |
ctx->spilling_setup = true; |
| 828 |
} |
| 829 |
|
| 830 |
unsigned *uses = rzalloc_array_size(ra_regs, sizeof(*uses), num_ssas); |
| 831 |
pco_foreach_instr_in_func (instr, func) { |
| 832 |
pco_foreach_instr_src_ssa (psrc, instr) { |
| 833 |
if (pco_ref_get_chans(*psrc) > 1) |
| 834 |
continue; |
| 835 |
|
| 836 |
++uses[psrc->val]; |
| 837 |
} |
| 838 |
} |
| 839 |
|
| 840 |
for (unsigned u = 0; u < num_ssas; ++u) |
| 841 |
ra_set_node_spill_cost(ra_graph, u, (float)uses[u]); |
| 842 |
|
| 843 |
unsigned spill_index = ra_get_best_spill_node(ra_graph); |
| 844 |
if (spill_index == ~0) { |
| 845 |
fprintf(stderr, "FATAL: Failed to get best spill node.\n"); |
| 846 |
abort(); |
| 847 |
} |
| 848 |
|
| 849 |
spill(spill_index, func, ctx); |
| 850 |
|
| 851 |
ralloc_free(ra_regs); |
| 852 |
return false; |
| 853 |
} |
| 854 |
|
| 855 |
if (pco_should_print_shader(func->parent_shader) && PCO_DEBUG_PRINT(RA)) { |
| 856 |
printf("RA live ranges:\n"); |
| 857 |
for (unsigned u = 0; u < num_vars; ++u) |
| 858 |
printf(" %c%u: %u, %u\n", |
| 859 |
u >= num_ssas ? '$' : '%', |
| 860 |
u >= num_ssas ? u - num_ssas : u, |
| 861 |
live_ranges[u].start, |
| 862 |
live_ranges[u].end); |
| 863 |
|
| 864 |
if (_mesa_hash_table_u64_num_entries(overrides)) { |
| 865 |
printf("RA overrides:\n"); |
| 866 |
hash_table_u64_foreach (overrides, entry) { |
| 867 |
struct vec_override *override = entry.data; |
| 868 |
printf(" %%%" PRIu64 ": ref = ", entry.key); |
| 869 |
pco_print_ref(func->parent_shader, override->ref); |
| 870 |
printf(", offset = %u\n", override->offset); |
| 871 |
} |
| 872 |
printf("\n"); |
| 873 |
} |
| 874 |
} |
| 875 |
|
| 876 |
/* Replace vars with allocated registers. */ |
| 877 |
unsigned temps = 0; |
| 878 |
unsigned vtxins = 0; |
| 879 |
unsigned interns = 0; |
| 880 |
pco_foreach_instr_in_func_safe (instr, func) { |
| 881 |
#if 1 |
| 882 |
if (pco_should_print_shader(func->parent_shader) && PCO_DEBUG_PRINT(RA)) |
| 883 |
pco_print_shader(func->parent_shader, stdout, "ra debug"); |
| 884 |
#endif |
| 885 |
|
| 886 |
/* Insert movs for scalar components of super vecs. */ |
| 887 |
if (instr->op == PCO_OP_VEC) { |
| 888 |
pco_builder b = |
| 889 |
pco_builder_create(func, pco_cursor_before_instr(instr)); |
| 890 |
|
| 891 |
struct vec_override *override = |
| 892 |
_mesa_hash_table_u64_search(overrides, instr->dest[0].val); |
| 893 |
|
| 894 |
unsigned offset = override ? override->offset : 0; |
| 895 |
|
| 896 |
unsigned temp_dest_base = |
| 897 |
override ? ra_get_node_reg(ra_graph, override->ref.val) |
| 898 |
: ra_get_node_reg(ra_graph, instr->dest[0].val); |
| 899 |
|
| 900 |
struct util_dynarray copies = UTIL_DYNARRAY_INIT; |
| 901 |
|
| 902 |
unsigned highest_temp = 0; |
| 903 |
unsigned lowest_temp = ~0; |
| 904 |
|
| 905 |
enum pco_exec_cnd exec_cnd = pco_instr_get_exec_cnd(instr); |
| 906 |
pco_foreach_instr_src (psrc, instr) { |
| 907 |
if (!pco_ref_is_ssa(*psrc) || |
| 908 |
!_mesa_hash_table_u64_search(overrides, psrc->val) || |
| 909 |
BITSET_TEST(comps, psrc->val)) { |
| 910 |
unsigned chans = pco_ref_get_chans(*psrc); |
| 911 |
|
| 912 |
unsigned temp_src_base = ~0U; |
| 913 |
if (pco_ref_is_ssa(*psrc)) { |
| 914 |
temp_src_base = ra_get_node_reg(ra_graph, psrc->val); |
| 915 |
|
| 916 |
struct vec_override *src_override = |
| 917 |
_mesa_hash_table_u64_search(overrides, psrc->val); |
| 918 |
if (src_override) { |
| 919 |
temp_src_base = |
| 920 |
ra_get_node_reg(ra_graph, src_override->ref.val); |
| 921 |
temp_src_base += src_override->offset; |
| 922 |
} |
| 923 |
} else if (pco_ref_is_vreg(*psrc)) { |
| 924 |
temp_src_base = |
| 925 |
ra_get_node_reg(ra_graph, psrc->val + num_ssas); |
| 926 |
} |
| 927 |
|
| 928 |
for (unsigned u = 0; u < chans; ++u) { |
| 929 |
pco_ref dest; |
| 930 |
if (temp_dest_base + offset >= ctx->allocable_temps) { |
| 931 |
dest = pco_ref_hwreg(temp_dest_base + offset - |
| 932 |
ctx->allocable_temps, |
| 933 |
PCO_REG_CLASS_VTXIN); |
| 934 |
} else { |
| 935 |
dest = pco_ref_hwreg(temp_dest_base + offset, |
| 936 |
PCO_REG_CLASS_TEMP); |
| 937 |
} |
| 938 |
|
| 939 |
dest = pco_ref_offset(dest, u); |
| 940 |
dest = pco_ref_offset(dest, ctx->temp_alloc_offset); |
| 941 |
|
| 942 |
pco_ref src; |
| 943 |
if (pco_ref_is_ssa(*psrc) || pco_ref_is_vreg(*psrc)) { |
| 944 |
if (temp_src_base >= ctx->allocable_temps) { |
| 945 |
src = |
| 946 |
pco_ref_hwreg(temp_src_base - ctx->allocable_temps, |
| 947 |
PCO_REG_CLASS_VTXIN); |
| 948 |
} else { |
| 949 |
src = pco_ref_hwreg(temp_src_base, PCO_REG_CLASS_TEMP); |
| 950 |
} |
| 951 |
} else { |
| 952 |
src = pco_ref_chans(*psrc, 1); |
| 953 |
} |
| 954 |
|
| 955 |
src = pco_ref_offset(src, u); |
| 956 |
src = pco_ref_offset(src, ctx->temp_alloc_offset); |
| 957 |
|
| 958 |
pco_ref_xfer_mods(&src, psrc, false); |
| 959 |
|
| 960 |
/* if (!pco_refs_are_equal(src, dest, true)) */ { |
| 961 |
highest_temp = |
| 962 |
MAX3(highest_temp, |
| 963 |
pco_ref_is_temp(src) ? pco_ref_get_temp(src) |
| 964 |
: highest_temp, |
| 965 |
pco_ref_is_temp(dest) ? pco_ref_get_temp(dest) |
| 966 |
: highest_temp); |
| 967 |
|
| 968 |
lowest_temp = |
| 969 |
MIN3(lowest_temp, |
| 970 |
pco_ref_is_temp(src) ? pco_ref_get_temp(src) |
| 971 |
: lowest_temp, |
| 972 |
pco_ref_is_temp(dest) ? pco_ref_get_temp(dest) |
| 973 |
: lowest_temp); |
| 974 |
pco_copy copy = { |
| 975 |
.src = src, |
| 976 |
.dest = dest, |
| 977 |
.s1 = pco_ref_is_reg(src) && |
| 978 |
pco_ref_get_reg_class(src) == PCO_REG_CLASS_SPEC, |
| 979 |
}; |
| 980 |
|
| 981 |
/* |
| 982 |
if (pco_ref_is_reg(src) && |
| 983 |
pco_ref_get_reg_class(src) == PCO_REG_CLASS_SPEC) { |
| 984 |
pco_movs1(&b, dest, src, .exec_cnd = exec_cnd); |
| 985 |
} else { |
| 986 |
pco_mbyp(&b, dest, src, .exec_cnd = exec_cnd); |
| 987 |
} |
| 988 |
*/ |
| 989 |
|
| 990 |
util_dynarray_append(&copies, copy); |
| 991 |
} |
| 992 |
} |
| 993 |
|
| 994 |
temps = MAX2(temps, temp_dest_base + offset + chans); |
| 995 |
} |
| 996 |
|
| 997 |
offset += pco_ref_get_chans(*psrc); |
| 998 |
} |
| 999 |
|
| 1000 |
/* Emit copies. */ |
| 1001 |
emit_copies(&b, &copies, exec_cnd, highest_temp, lowest_temp); |
| 1002 |
|
| 1003 |
util_dynarray_fini(&copies); |
| 1004 |
|
| 1005 |
pco_instr_delete(instr); |
| 1006 |
continue; |
| 1007 |
} else if (instr->op == PCO_OP_COMP) { |
| 1008 |
pco_instr_delete(instr); |
| 1009 |
continue; |
| 1010 |
} |
| 1011 |
|
| 1012 |
pco_foreach_instr_dest_ssa (pdest, instr) { |
| 1013 |
struct vec_override *override = |
| 1014 |
_mesa_hash_table_u64_search(overrides, pdest->val); |
| 1015 |
|
| 1016 |
unsigned val = ra_get_node_reg(ra_graph, pdest->val); |
| 1017 |
unsigned dest_temps = val + pco_ref_get_chans(*pdest); |
| 1018 |
if (override) { |
| 1019 |
val = ra_get_node_reg(ra_graph, override->ref.val); |
| 1020 |
dest_temps = val + pco_ref_get_chans(override->ref); |
| 1021 |
val += override->offset; |
| 1022 |
} |
| 1023 |
|
| 1024 |
pdest->type = PCO_REF_TYPE_REG; |
| 1025 |
pdest->reg_class = PCO_REG_CLASS_TEMP; |
| 1026 |
pdest->val = val + ctx->temp_alloc_offset; |
| 1027 |
|
| 1028 |
/* Got a vertex input register. */ |
| 1029 |
if (val >= ctx->allocable_temps) { |
| 1030 |
pdest->reg_class = PCO_REG_CLASS_VTXIN; |
| 1031 |
pdest->val = val - ctx->allocable_temps; |
| 1032 |
vtxins = MAX2(vtxins, dest_temps - ctx->allocable_temps); |
| 1033 |
} else { |
| 1034 |
temps = MAX2(temps, dest_temps + ctx->temp_alloc_offset); |
| 1035 |
} |
| 1036 |
} |
| 1037 |
|
| 1038 |
pco_foreach_instr_src_ssa (psrc, instr) { |
| 1039 |
struct vec_override *override = |
| 1040 |
_mesa_hash_table_u64_search(overrides, psrc->val); |
| 1041 |
|
| 1042 |
unsigned val = |
| 1043 |
override |
| 1044 |
? ra_get_node_reg(ra_graph, override->ref.val) + override->offset |
| 1045 |
: ra_get_node_reg(ra_graph, psrc->val); |
| 1046 |
|
| 1047 |
psrc->type = PCO_REF_TYPE_REG; |
| 1048 |
psrc->reg_class = PCO_REG_CLASS_TEMP; |
| 1049 |
psrc->val = val + ctx->temp_alloc_offset; |
| 1050 |
|
| 1051 |
/* Got a vertex input register. */ |
| 1052 |
if (val >= ctx->allocable_temps) { |
| 1053 |
psrc->reg_class = PCO_REG_CLASS_VTXIN; |
| 1054 |
psrc->val = val - ctx->allocable_temps; |
| 1055 |
} |
| 1056 |
} |
| 1057 |
|
| 1058 |
pco_foreach_instr_dest_vreg (pdest, instr) { |
| 1059 |
unsigned val = ra_get_node_reg(ra_graph, pdest->val + num_ssas); |
| 1060 |
unsigned dest_temps = val + 1; |
| 1061 |
|
| 1062 |
pdest->type = PCO_REF_TYPE_REG; |
| 1063 |
pdest->reg_class = PCO_REG_CLASS_TEMP; |
| 1064 |
pdest->val = val + ctx->temp_alloc_offset; |
| 1065 |
|
| 1066 |
/* Got a vertex input register. */ |
| 1067 |
if (val >= ctx->allocable_temps) { |
| 1068 |
pdest->reg_class = PCO_REG_CLASS_VTXIN; |
| 1069 |
pdest->val = val - ctx->allocable_temps; |
| 1070 |
vtxins = MAX2(vtxins, dest_temps - ctx->allocable_temps); |
| 1071 |
} else { |
| 1072 |
temps = MAX2(temps, dest_temps); |
| 1073 |
} |
| 1074 |
} |
| 1075 |
|
| 1076 |
pco_foreach_instr_src_vreg (psrc, instr) { |
| 1077 |
unsigned val = ra_get_node_reg(ra_graph, psrc->val + num_ssas); |
| 1078 |
|
| 1079 |
psrc->type = PCO_REF_TYPE_REG; |
| 1080 |
psrc->reg_class = PCO_REG_CLASS_TEMP; |
| 1081 |
psrc->val = val + ctx->temp_alloc_offset; |
| 1082 |
|
| 1083 |
/* Got a vertex input register. */ |
| 1084 |
if (val >= ctx->allocable_temps) { |
| 1085 |
psrc->reg_class = PCO_REG_CLASS_VTXIN; |
| 1086 |
psrc->val = val - ctx->allocable_temps; |
| 1087 |
} |
| 1088 |
} |
| 1089 |
|
| 1090 |
/* Drop no-ops. */ |
| 1091 |
if (instr->op == PCO_OP_MBYP && |
| 1092 |
(pco_ref_is_ssa(instr->src[0]) || pco_ref_is_vreg(instr->src[0])) && |
| 1093 |
pco_refs_are_equal(instr->src[0], instr->dest[0], true)) { |
| 1094 |
pco_instr_delete(instr); |
| 1095 |
} |
| 1096 |
} |
| 1097 |
|
| 1098 |
ralloc_free(ra_regs); |
| 1099 |
|
| 1100 |
func->temps = temps; |
| 1101 |
func->vtxins = vtxins; |
| 1102 |
|
| 1103 |
if (pco_should_print_shader(func->parent_shader) && PCO_DEBUG_PRINT(RA)) { |
| 1104 |
printf( |
| 1105 |
"RA allocated %u temps, %u vtxins, %u interns from %u SSA vars, %u vregs.\n", |
| 1106 |
temps, |
| 1107 |
vtxins, |
| 1108 |
interns, |
| 1109 |
num_ssas, |
| 1110 |
num_vregs); |
| 1111 |
} |
| 1112 |
|
| 1113 |
ctx->done = true; |
| 1114 |
return true; |
| 1115 |
} |
| 1116 |
|
| 1117 |
/** |
| 1118 |
* \brief Register allocation pass. |
| 1119 |
* |
| 1120 |
* \param[in,out] shader PCO shader. |
| 1121 |
* \return True if the pass made progress. |
| 1122 |
*/ |
| 1123 |
bool pco_ra(pco_shader *shader) |
| 1124 |
{ |
| 1125 |
assert(!shader->is_grouped); |
| 1126 |
|
| 1127 |
/* Instruction indices need to be ordered for live ranges. */ |
| 1128 |
pco_index(shader, false); |
| 1129 |
|
| 1130 |
unsigned hw_temps = rogue_get_temps(shader->ctx->dev_info); |
| 1131 |
/* TODO: |
| 1132 |
* unsigned opt_temps = rogue_get_optimal_temps(shader->ctx->dev_info); |
| 1133 |
*/ |
| 1134 |
|
| 1135 |
/* If any vertex input registers are already used, round up to the nearest |
| 1136 |
* multiple of 4 as vertex input registers are allocated in blocks of 4. |
| 1137 |
* |
| 1138 |
* This number is used by default as the maximum safe number of vertex input |
| 1139 |
* registers that can be used is not currently known. |
| 1140 |
*/ |
| 1141 |
unsigned hw_vtxins = ALIGN_POT(shader->data.common.vtxins, |
| 1142 |
ROGUE_USRM_GRANULARITY_IN_REGISTERS); |
| 1143 |
|
| 1144 |
if (shader->stage != MESA_SHADER_FRAGMENT && !shader->is_internal) { |
| 1145 |
if (PCO_DEBUG(ALLOC_EXTRA_VTXINS)) { |
| 1146 |
hw_vtxins = MAX2(hw_vtxins, rogue_get_vtxins()); |
| 1147 |
} |
| 1148 |
} |
| 1149 |
|
| 1150 |
/* TODO: different number of temps available if preamble/phase change. */ |
| 1151 |
/* TODO: different number of temps available if barriers are in use. */ |
| 1152 |
/* TODO: support for internal registers. */ |
| 1153 |
pco_ra_ctx ctx = { |
| 1154 |
.allocable_temps = hw_temps, |
| 1155 |
.allocable_vtxins = hw_vtxins, |
| 1156 |
.allocable_interns = 0, |
| 1157 |
}; |
| 1158 |
|
| 1159 |
if (shader->stage == MESA_SHADER_COMPUTE) { |
| 1160 |
unsigned wg_size = shader->data.cs.workgroup_size[0] * |
| 1161 |
shader->data.cs.workgroup_size[1] * |
| 1162 |
shader->data.cs.workgroup_size[2]; |
| 1163 |
ctx.allocable_temps = |
| 1164 |
rogue_max_wg_temps(shader->ctx->dev_info, |
| 1165 |
ctx.allocable_temps, |
| 1166 |
wg_size, |
| 1167 |
shader->data.common.uses.barriers); |
| 1168 |
} |
| 1169 |
|
| 1170 |
/* Perform register allocation for each function. */ |
| 1171 |
bool progress = false; |
| 1172 |
pco_foreach_func_in_shader (func, shader) { |
| 1173 |
ctx.done = false; |
| 1174 |
while (!ctx.done) |
| 1175 |
progress |= pco_ra_func(func, &ctx); |
| 1176 |
|
| 1177 |
shader->data.common.temps = MAX2(shader->data.common.temps, func->temps); |
| 1178 |
shader->data.common.vtxins = |
| 1179 |
MAX2(shader->data.common.vtxins, func->vtxins); |
| 1180 |
} |
| 1181 |
|
| 1182 |
shader->data.common.spilled_temps = ctx.spilled_temps; |
| 1183 |
return progress; |
| 1184 |
} |