1/* 2 * Copyright 2013 Vadim Girlin <vadimgirlin@gmail.com> 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * on the rights to use, copy, modify, merge, publish, distribute, sub 8 * license, and/or sell copies of the Software, and to permit persons to whom 9 * the Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM, 19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 21 * USE OR OTHER DEALINGS IN THE SOFTWARE. 22 * 23 * Authors: 24 * Vadim Girlin 25 */ 26 27#ifndef SB_BC_H_ 28#define SB_BC_H_ 29 30#include <stdint.h> 31#include "r600_isa.h" 32 33#include <cstdio> 34#include <string> 35#include <vector> 36#include <stack> 37 38struct r600_bytecode; 39struct r600_shader; 40 41namespace r600_sb { 42 43class hw_encoding_format; 44class node; 45class alu_node; 46class cf_node; 47class fetch_node; 48class alu_group_node; 49class region_node; 50class shader; 51class value; 52 53class sb_ostream { 54public: 55 sb_ostream() {} 56 57 virtual void write(const char *s) = 0; 58 59 sb_ostream& operator <<(const char *s) { 60 write(s); 61 return *this; 62 } 63 64 sb_ostream& operator <<(const std::string& s) { 65 return *this << s.c_str(); 66 } 67 68 sb_ostream& operator <<(void *p) { 69 char b[32]; 70 sprintf(b, "%p", p); 71 return *this << b; 72 } 73 74 sb_ostream& operator <<(char c) { 75 char b[2]; 76 sprintf(b, "%c", c); 77 return *this << b; 78 } 79 80 sb_ostream& operator <<(int n) { 81 char b[32]; 82 sprintf(b, "%d", n); 83 return *this << b; 84 } 85 86 sb_ostream& operator <<(unsigned n) { 87 char b[32]; 88 sprintf(b, "%u", n); 89 return *this << b; 90 } 91 92 sb_ostream& operator <<(double d) { 93 char b[32]; 94 snprintf(b, 32, "%g", d); 95 return *this << b; 96 } 97 98 // print as field of specified width, right aligned 99 void print_w(int n, int width) { 100 char b[256],f[8]; 101 sprintf(f, "%%%dd", width); 102 snprintf(b, 256, f, n); 103 write(b); 104 } 105 106 // print as field of specified width, left aligned 107 void print_wl(int n, int width) { 108 char b[256],f[8]; 109 sprintf(f, "%%-%dd", width); 110 snprintf(b, 256, f, n); 111 write(b); 112 } 113 114 // print as field of specified width, left aligned 115 void print_wl(const std::string &s, int width) { 116 write(s.c_str()); 117 int l = s.length(); 118 while (l++ < width) { 119 write(" "); 120 } 121 } 122 123 // print int as field of specified width, right aligned, zero-padded 124 void print_zw(int n, int width) { 125 char b[256],f[8]; 126 sprintf(f, "%%0%dd", width); 127 snprintf(b, 256, f, n); 128 write(b); 129 } 130 131 // print int as field of specified width, right aligned, zero-padded, hex 132 void print_zw_hex(int n, int width) { 133 char b[256],f[8]; 134 sprintf(f, "%%0%dx", width); 135 snprintf(b, 256, f, n); 136 write(b); 137 } 138}; 139 140class sb_ostringstream : public sb_ostream { 141 std::string data; 142public: 143 sb_ostringstream() : data() {} 144 145 virtual void write(const char *s) { 146 data += s; 147 } 148 149 void clear() { data.clear(); } 150 151 const char* c_str() { return data.c_str(); } 152 std::string& str() { return data; } 153}; 154 155class sb_log : public sb_ostream { 156 FILE *o; 157public: 158 sb_log() : o(stderr) {} 159 160 virtual void write(const char *s) { 161 fputs(s, o); 162 } 163}; 164 165extern sb_log sblog; 166 167enum shader_target 168{ 169 TARGET_UNKNOWN, 170 TARGET_VS, 171 TARGET_ES, 172 TARGET_PS, 173 TARGET_GS, 174 TARGET_GS_COPY, 175 TARGET_COMPUTE, 176 TARGET_FETCH, 177 TARGET_HS, 178 TARGET_LS, 179 180 TARGET_NUM 181}; 182 183enum sb_hw_class_bits 184{ 185 HB_R6 = (1<<0), 186 HB_R7 = (1<<1), 187 HB_EG = (1<<2), 188 HB_CM = (1<<3), 189 190 HB_R6R7 = (HB_R6 | HB_R7), 191 HB_EGCM = (HB_EG | HB_CM), 192 HB_R6R7EG = (HB_R6 | HB_R7 | HB_EG), 193 HB_R7EGCM = (HB_R7 | HB_EG | HB_CM), 194 195 HB_ALL = (HB_R6 | HB_R7 | HB_EG | HB_CM) 196}; 197 198enum sb_hw_chip 199{ 200 HW_CHIP_UNKNOWN, 201 HW_CHIP_R600, 202 HW_CHIP_RV610, 203 HW_CHIP_RV630, 204 HW_CHIP_RV670, 205 HW_CHIP_RV620, 206 HW_CHIP_RV635, 207 HW_CHIP_RS780, 208 HW_CHIP_RS880, 209 HW_CHIP_RV770, 210 HW_CHIP_RV730, 211 HW_CHIP_RV710, 212 HW_CHIP_RV740, 213 HW_CHIP_CEDAR, 214 HW_CHIP_REDWOOD, 215 HW_CHIP_JUNIPER, 216 HW_CHIP_CYPRESS, 217 HW_CHIP_HEMLOCK, 218 HW_CHIP_PALM, 219 HW_CHIP_SUMO, 220 HW_CHIP_SUMO2, 221 HW_CHIP_BARTS, 222 HW_CHIP_TURKS, 223 HW_CHIP_CAICOS, 224 HW_CHIP_CAYMAN, 225 HW_CHIP_ARUBA 226}; 227 228enum sb_hw_class 229{ 230 HW_CLASS_UNKNOWN, 231 HW_CLASS_R600, 232 HW_CLASS_R700, 233 HW_CLASS_EVERGREEN, 234 HW_CLASS_CAYMAN 235}; 236 237enum alu_slots { 238 SLOT_X = 0, 239 SLOT_Y = 1, 240 SLOT_Z = 2, 241 SLOT_W = 3, 242 SLOT_TRANS = 4 243}; 244 245enum misc_consts { 246 MAX_ALU_LITERALS = 4, 247 MAX_ALU_SLOTS = 128, 248 MAX_GPR = 128, 249 MAX_CHAN = 4 250 251}; 252 253enum alu_src_sel { 254 255 ALU_SRC_LDS_OQ_A = 219, 256 ALU_SRC_LDS_OQ_B = 220, 257 ALU_SRC_LDS_OQ_A_POP = 221, 258 ALU_SRC_LDS_OQ_B_POP = 222, 259 ALU_SRC_LDS_DIRECT_A = 223, 260 ALU_SRC_LDS_DIRECT_B = 224, 261 ALU_SRC_TIME_HI = 227, 262 ALU_SRC_TIME_LO = 228, 263 ALU_SRC_MASK_HI = 229, 264 ALU_SRC_MASK_LO = 230, 265 ALU_SRC_HW_WAVE_ID = 231, 266 ALU_SRC_SIMD_ID = 232, 267 ALU_SRC_SE_ID = 233, 268 ALU_SRC_HW_THREADGRP_ID = 234, 269 ALU_SRC_WAVE_ID_IN_GRP = 235, 270 ALU_SRC_NUM_THREADGRP_WAVES = 236, 271 ALU_SRC_HW_ALU_ODD = 237, 272 ALU_SRC_LOOP_IDX = 238, 273 ALU_SRC_PARAM_BASE_ADDR = 240, 274 ALU_SRC_NEW_PRIM_MASK = 241, 275 ALU_SRC_PRIM_MASK_HI = 242, 276 ALU_SRC_PRIM_MASK_LO = 243, 277 ALU_SRC_1_DBL_L = 244, 278 ALU_SRC_1_DBL_M = 245, 279 ALU_SRC_0_5_DBL_L = 246, 280 ALU_SRC_0_5_DBL_M = 247, 281 ALU_SRC_0 = 248, 282 ALU_SRC_1 = 249, 283 ALU_SRC_1_INT = 250, 284 ALU_SRC_M_1_INT = 251, 285 ALU_SRC_0_5 = 252, 286 ALU_SRC_LITERAL = 253, 287 ALU_SRC_PV = 254, 288 ALU_SRC_PS = 255, 289 290 ALU_SRC_PARAM_OFFSET = 448 291}; 292 293enum alu_predicate_select 294{ 295 PRED_SEL_OFF = 0, 296// RESERVED = 1, 297 PRED_SEL_0 = 2, 298 PRED_SEL_1 = 3 299}; 300 301 302enum alu_omod { 303 OMOD_OFF = 0, 304 OMOD_M2 = 1, 305 OMOD_M4 = 2, 306 OMOD_D2 = 3 307}; 308 309enum alu_index_mode { 310 INDEX_AR_X = 0, 311 INDEX_AR_Y_R600 = 1, 312 INDEX_AR_Z_R600 = 2, 313 INDEX_AR_W_R600 = 3, 314 315 INDEX_LOOP = 4, 316 INDEX_GLOBAL = 5, 317 INDEX_GLOBAL_AR_X = 6 318}; 319 320enum alu_cayman_mova_dst { 321 CM_MOVADST_AR_X, 322 CM_MOVADST_PC, 323 CM_MOVADST_IDX0, 324 CM_MOVADST_IDX1, 325 CM_MOVADST_CG0, // clause-global byte 0 326 CM_MOVADST_CG1, 327 CM_MOVADST_CG2, 328 CM_MOVADST_CG3 329}; 330 331enum alu_cayman_exec_mask_op { 332 CM_EMO_DEACTIVATE, 333 CM_EMO_BREAK, 334 CM_EMO_CONTINUE, 335 CM_EMO_KILL 336}; 337 338 339enum cf_exp_type { 340 EXP_PIXEL, 341 EXP_POS, 342 EXP_PARAM, 343 344 EXP_TYPE_COUNT 345}; 346 347enum cf_mem_type { 348 MEM_WRITE, 349 MEM_WRITE_IND, 350 MEM_WRITE_ACK, 351 MEM_WRITE_IND_ACK 352}; 353 354 355enum alu_kcache_mode { 356 KC_LOCK_NONE, 357 KC_LOCK_1, 358 KC_LOCK_2, 359 KC_LOCK_LOOP 360}; 361 362enum alu_kcache_index_mode { 363 KC_INDEX_NONE, 364 KC_INDEX_0, 365 KC_INDEX_1, 366 KC_INDEX_INVALID 367}; 368 369enum chan_select { 370 SEL_X = 0, 371 SEL_Y = 1, 372 SEL_Z = 2, 373 SEL_W = 3, 374 SEL_0 = 4, 375 SEL_1 = 5, 376// RESERVED = 6, 377 SEL_MASK = 7 378}; 379 380enum bank_swizzle { 381 VEC_012 = 0, 382 VEC_021 = 1, 383 VEC_120 = 2, 384 VEC_102 = 3, 385 VEC_201 = 4, 386 VEC_210 = 5, 387 388 VEC_NUM = 6, 389 390 SCL_210 = 0, 391 SCL_122 = 1, 392 SCL_212 = 2, 393 SCL_221 = 3, 394 395 SCL_NUM = 4 396 397}; 398 399enum sched_queue_id { 400 SQ_CF, 401 SQ_ALU, 402 SQ_TEX, 403 SQ_VTX, 404 SQ_GDS, 405 406 SQ_NUM 407}; 408 409struct literal { 410 union { 411 int32_t i; 412 uint32_t u; 413 float f; 414 }; 415 416 literal(int32_t i = 0) : i(i) {} 417 literal(uint32_t u) : u(u) {} 418 literal(float f) : f(f) {} 419 literal(double f) : f(f) {} 420 operator uint32_t() const { return u; } 421 bool operator ==(literal l) { return u == l.u; } 422 bool operator ==(int v_int) { return i == v_int; } 423 bool operator ==(unsigned v_uns) { return u == v_uns; } 424}; 425 426struct bc_kcache { 427 unsigned mode; 428 unsigned bank; 429 unsigned addr; 430 unsigned index_mode; 431} ; 432 433// TODO optimize bc structures 434 435struct bc_cf { 436 437 bc_kcache kc[4]; 438 439 unsigned id; 440 441 442 const cf_op_info * op_ptr; 443 unsigned op; 444 445 unsigned addr:32; 446 447 unsigned alt_const:1; 448 unsigned uses_waterfall:1; 449 450 unsigned barrier:1; 451 unsigned count:7; 452 unsigned pop_count:3; 453 unsigned call_count:6; 454 unsigned whole_quad_mode:1; 455 unsigned valid_pixel_mode:1; 456 457 unsigned jumptable_sel:3; 458 unsigned cf_const:5; 459 unsigned cond:2; 460 unsigned end_of_program:1; 461 462 unsigned array_base:13; 463 unsigned elem_size:2; 464 unsigned index_gpr:7; 465 unsigned rw_gpr:7; 466 unsigned rw_rel:1; 467 unsigned type:2; 468 469 unsigned burst_count:4; 470 unsigned mark:1; 471 unsigned sel[4]; 472 473 unsigned array_size:12; 474 unsigned comp_mask:4; 475 476 unsigned rat_id:4; 477 unsigned rat_inst:6; 478 unsigned rat_index_mode:2; 479 480 void set_op(unsigned op) { this->op = op; op_ptr = r600_isa_cf(op); } 481 482 bool is_alu_extended() { 483 assert(op_ptr->flags & CF_ALU); 484 return kc[2].mode != KC_LOCK_NONE || kc[3].mode != KC_LOCK_NONE || 485 kc[0].index_mode != KC_INDEX_NONE || kc[1].index_mode != KC_INDEX_NONE || 486 kc[2].index_mode != KC_INDEX_NONE || kc[3].index_mode != KC_INDEX_NONE; 487 } 488 489}; 490 491struct bc_alu_src { 492 unsigned sel:9; 493 unsigned chan:2; 494 unsigned neg:1; 495 unsigned abs:1; 496 unsigned rel:1; 497 literal value; 498}; 499 500struct bc_alu { 501 const alu_op_info * op_ptr; 502 unsigned op; 503 504 bc_alu_src src[3]; 505 506 unsigned dst_gpr:7; 507 unsigned dst_chan:2; 508 unsigned dst_rel:1; 509 unsigned clamp:1; 510 unsigned omod:2; 511 unsigned bank_swizzle:3; 512 513 unsigned index_mode:3; 514 unsigned last:1; 515 unsigned pred_sel:2; 516 517 unsigned fog_merge:1; 518 unsigned write_mask:1; 519 unsigned update_exec_mask:1; 520 unsigned update_pred:1; 521 522 unsigned slot:3; 523 524 unsigned lds_idx_offset:6; 525 526 alu_op_flags slot_flags; 527 528 void set_op(unsigned op) { 529 this->op = op; 530 op_ptr = r600_isa_alu(op); 531 } 532}; 533 534struct bc_fetch { 535 const fetch_op_info * op_ptr; 536 unsigned op; 537 538 unsigned bc_frac_mode:1; 539 unsigned fetch_whole_quad:1; 540 unsigned resource_id:8; 541 542 unsigned src_gpr:7; 543 unsigned src_rel:1; 544 unsigned src_rel_global:1; /* for GDS ops */ 545 unsigned src_sel[4]; 546 547 unsigned dst_gpr:7; 548 unsigned dst_rel:1; 549 unsigned dst_rel_global:1; /* for GDS ops */ 550 unsigned dst_sel[4]; 551 552 unsigned alt_const:1; 553 554 unsigned inst_mod:2; 555 unsigned resource_index_mode:2; 556 unsigned sampler_index_mode:2; 557 558 unsigned coord_type[4]; 559 unsigned lod_bias:7; 560 561 unsigned offset[3]; 562 563 unsigned sampler_id:5; 564 565 566 unsigned fetch_type:2; 567 unsigned mega_fetch_count:6; 568 unsigned coalesced_read:1; 569 unsigned structured_read:2; 570 unsigned lds_req:1; 571 572 unsigned data_format:6; 573 unsigned format_comp_all:1; 574 unsigned num_format_all:2; 575 unsigned semantic_id:8; 576 unsigned srf_mode_all:1; 577 unsigned use_const_fields:1; 578 579 unsigned const_buf_no_stride:1; 580 unsigned endian_swap:2; 581 unsigned mega_fetch:1; 582 583 unsigned src2_gpr:7; /* for GDS */ 584 unsigned alloc_consume:1; 585 unsigned uav_id:4; 586 unsigned uav_index_mode:2; 587 unsigned bcast_first_req:1; 588 589 /* for MEM ops */ 590 unsigned elem_size:2; 591 unsigned uncached:1; 592 unsigned indexed:1; 593 unsigned burst_count:4; 594 unsigned array_base:13; 595 unsigned array_size:12; 596 597 void set_op(unsigned op) { this->op = op; op_ptr = r600_isa_fetch(op); } 598}; 599 600struct shader_stats { 601 unsigned ndw; 602 unsigned ngpr; 603 unsigned nstack; 604 605 unsigned cf; // clause instructions not included 606 unsigned alu; 607 unsigned alu_clauses; 608 unsigned fetch_clauses; 609 unsigned fetch; 610 unsigned alu_groups; 611 612 unsigned shaders; // number of shaders (for accumulated stats) 613 614 shader_stats() : ndw(), ngpr(), nstack(), cf(), alu(), alu_clauses(), 615 fetch_clauses(), fetch(), alu_groups(), shaders() {} 616 617 void collect(node *n); 618 void accumulate(shader_stats &s); 619 void dump(); 620 void dump_diff(shader_stats &s); 621}; 622 623class sb_context { 624 625public: 626 627 shader_stats src_stats, opt_stats; 628 629 r600_isa *isa; 630 631 sb_hw_chip hw_chip; 632 sb_hw_class hw_class; 633 634 unsigned alu_temp_gprs; 635 unsigned max_fetch; 636 bool has_trans; 637 unsigned vtx_src_num; 638 unsigned num_slots; 639 bool uses_mova_gpr; 640 641 bool r6xx_gpr_index_workaround; 642 643 bool stack_workaround_8xx; 644 bool stack_workaround_9xx; 645 646 unsigned wavefront_size; 647 unsigned stack_entry_size; 648 649 static unsigned dump_pass; 650 static unsigned dump_stat; 651 652 static unsigned dry_run; 653 static unsigned no_fallback; 654 static unsigned safe_math; 655 656 static unsigned dskip_start; 657 static unsigned dskip_end; 658 static unsigned dskip_mode; 659 660 sb_context() : src_stats(), opt_stats(), isa(0), 661 hw_chip(HW_CHIP_UNKNOWN), hw_class(HW_CLASS_UNKNOWN) {} 662 663 int init(r600_isa *isa, sb_hw_chip chip, sb_hw_class cclass); 664 665 bool is_r600() {return hw_class == HW_CLASS_R600;} 666 bool is_r700() {return hw_class == HW_CLASS_R700;} 667 bool is_evergreen() {return hw_class == HW_CLASS_EVERGREEN;} 668 bool is_cayman() {return hw_class == HW_CLASS_CAYMAN;} 669 bool is_egcm() {return hw_class >= HW_CLASS_EVERGREEN;} 670 671 bool needs_8xx_stack_workaround() { 672 if (!is_evergreen()) 673 return false; 674 675 switch (hw_chip) { 676 case HW_CHIP_HEMLOCK: 677 case HW_CHIP_CYPRESS: 678 case HW_CHIP_JUNIPER: 679 return false; 680 default: 681 return true; 682 } 683 } 684 685 bool needs_9xx_stack_workaround() { 686 return is_cayman(); 687 } 688 689 sb_hw_class_bits hw_class_bit() { 690 switch (hw_class) { 691 case HW_CLASS_R600:return HB_R6; 692 case HW_CLASS_R700:return HB_R7; 693 case HW_CLASS_EVERGREEN:return HB_EG; 694 case HW_CLASS_CAYMAN:return HB_CM; 695 default: assert(!"unknown hw class"); return (sb_hw_class_bits)0; 696 697 } 698 } 699 700 unsigned cf_opcode(unsigned op) { 701 return r600_isa_cf_opcode(isa->hw_class, op); 702 } 703 704 unsigned alu_opcode(unsigned op) { 705 return r600_isa_alu_opcode(isa->hw_class, op); 706 } 707 708 unsigned alu_slots(unsigned op) { 709 return r600_isa_alu_slots(isa->hw_class, op); 710 } 711 712 unsigned alu_slots(const alu_op_info * op_ptr) { 713 return op_ptr->slots[isa->hw_class]; 714 } 715 716 unsigned alu_slots_mask(const alu_op_info * op_ptr) { 717 unsigned mask = 0; 718 unsigned slot_flags = alu_slots(op_ptr); 719 if (slot_flags & AF_V) 720 mask = 0x0F; 721 if (!is_cayman() && (slot_flags & AF_S)) 722 mask |= 0x10; 723 /* Force LDS_IDX ops into SLOT_X */ 724 if (op_ptr->opcode[0] == -1 && ((op_ptr->opcode[1] & 0xFF) == 0x11)) 725 mask = 0x01; 726 return mask; 727 } 728 729 unsigned fetch_opcode(unsigned op) { 730 return r600_isa_fetch_opcode(isa->hw_class, op); 731 } 732 733 bool is_kcache_sel(unsigned sel) { 734 return ((sel >= 128 && sel < 192) || (sel >= 256 && sel < 320)); 735 } 736 737 bool is_lds_oq(unsigned sel) { 738 return (sel >= 0xdb && sel <= 0xde); 739 } 740 741 const char * get_hw_class_name(); 742 const char * get_hw_chip_name(); 743 744}; 745 746#define SB_DUMP_STAT(a) do { if (sb_context::dump_stat) { a } } while (0) 747#define SB_DUMP_PASS(a) do { if (sb_context::dump_pass) { a } } while (0) 748 749class bc_decoder { 750 751 sb_context &ctx; 752 753 uint32_t* dw; 754 unsigned ndw; 755 756public: 757 758 bc_decoder(sb_context &sctx, uint32_t *data, unsigned size) 759 : ctx(sctx), dw(data), ndw(size) {} 760 761 int decode_cf(unsigned &i, bc_cf &bc); 762 int decode_alu(unsigned &i, bc_alu &bc); 763 int decode_fetch(unsigned &i, bc_fetch &bc); 764 765private: 766 int decode_cf_alu(unsigned &i, bc_cf &bc); 767 int decode_cf_exp(unsigned &i, bc_cf &bc); 768 int decode_cf_mem(unsigned &i, bc_cf &bc); 769 770 int decode_fetch_vtx(unsigned &i, bc_fetch &bc); 771 int decode_fetch_gds(unsigned &i, bc_fetch &bc); 772 int decode_fetch_mem(unsigned &i, bc_fetch &bc); 773}; 774 775// bytecode format definition 776 777class hw_encoding_format { 778 const sb_hw_class_bits hw_target; //FIXME: debug - remove after testing 779 hw_encoding_format(); 780protected: 781 uint32_t value; 782public: 783 hw_encoding_format(sb_hw_class_bits hw) 784 : hw_target(hw), value(0) {} 785 hw_encoding_format(uint32_t v, sb_hw_class_bits hw) 786 : hw_target(hw), value(v) {} 787 uint32_t get_value(sb_hw_class_bits hw) const { 788 assert((hw & hw_target) == hw); 789 return value; 790 } 791}; 792 793#define BC_FORMAT_BEGIN_HW(fmt, hwset) \ 794class fmt##_##hwset : public hw_encoding_format {\ 795 typedef fmt##_##hwset thistype; \ 796public: \ 797 fmt##_##hwset() : hw_encoding_format(HB_##hwset) {}; \ 798 fmt##_##hwset(uint32_t v) : hw_encoding_format(v, HB_##hwset) {}; 799 800#define BC_FORMAT_BEGIN(fmt) BC_FORMAT_BEGIN_HW(fmt, ALL) 801 802#define BC_FORMAT_END(fmt) }; 803 804// bytecode format field definition 805 806#define BC_FIELD(fmt, name, shortname, last_bit, first_bit) \ 807 thistype & name(unsigned v) { \ 808 value |= ((v&((1ull<<((last_bit)-(first_bit)+1))-1))<<(first_bit)); \ 809 return *this; \ 810 } \ 811 unsigned get_##name() const { \ 812 return (value>>(first_bit))&((1ull<<((last_bit)-(first_bit)+1))-1); \ 813 } 814 815#define BC_RSRVD(fmt, last_bit, first_bit) 816 817// CLAMP macro defined elsewhere interferes with bytecode field name 818#undef CLAMP 819#include "sb_bc_fmt_def.inc" 820 821#undef BC_FORMAT_BEGIN 822#undef BC_FORMAT_END 823#undef BC_FIELD 824#undef BC_RSRVD 825 826class bc_parser { 827 sb_context & ctx; 828 829 bc_decoder *dec; 830 831 r600_bytecode *bc; 832 r600_shader *pshader; 833 834 uint32_t *dw; 835 unsigned bc_ndw; 836 837 unsigned max_cf; 838 839 shader *sh; 840 841 int error; 842 843 alu_node *slots[2][5]; 844 unsigned cgroup; 845 846 typedef std::vector<cf_node*> id_cf_map; 847 id_cf_map cf_map; 848 849 typedef std::stack<region_node*> region_stack; 850 region_stack loop_stack; 851 852 bool gpr_reladdr; 853 854 // Note: currently relies on input emitting SET_CF in same basic block as uses 855 value *cf_index_value[2]; 856 alu_node *mova; 857public: 858 859 bc_parser(sb_context &sctx, r600_bytecode *bc, r600_shader* pshader) : 860 ctx(sctx), dec(), bc(bc), pshader(pshader), 861 dw(), bc_ndw(), max_cf(), 862 sh(), error(), slots(), cgroup(), 863 cf_map(), loop_stack(), gpr_reladdr(), cf_index_value(), mova() { } 864 865 int decode(); 866 int prepare(); 867 868 shader* get_shader() { assert(!error); return sh; } 869 870private: 871 872 int decode_shader(); 873 874 int parse_decls(); 875 876 int decode_cf(unsigned &i, bool &eop); 877 878 int decode_alu_clause(cf_node *cf); 879 int decode_alu_group(cf_node* cf, unsigned &i, unsigned &gcnt); 880 881 int decode_fetch_clause(cf_node *cf); 882 883 int prepare_ir(); 884 int prepare_alu_clause(cf_node *cf); 885 int prepare_alu_group(cf_node* cf, alu_group_node *g); 886 int prepare_fetch_clause(cf_node *cf); 887 888 int prepare_loop(cf_node *c); 889 int prepare_if(cf_node *c); 890 891 void save_set_cf_index(value *val, unsigned idx); 892 value *get_cf_index_value(unsigned idx); 893 void save_mova(alu_node *mova); 894 alu_node *get_mova(); 895}; 896 897 898 899 900class bytecode { 901 typedef std::vector<uint32_t> bc_vector; 902 sb_hw_class_bits hw_class_bit; 903 904 bc_vector bc; 905 906 unsigned pos; 907 908public: 909 910 bytecode(sb_hw_class_bits hw, unsigned rdw = 256) 911 : hw_class_bit(hw), pos(0) { bc.reserve(rdw); } 912 913 unsigned ndw() { return bc.size(); } 914 915 void write_data(uint32_t* dst) { 916 std::copy(bc.begin(), bc.end(), dst); 917 } 918 919 void align(unsigned a) { 920 unsigned size = bc.size(); 921 size = (size + a - 1) & ~(a-1); 922 bc.resize(size); 923 } 924 925 void set_size(unsigned sz) { 926 assert(sz >= bc.size()); 927 bc.resize(sz); 928 } 929 930 void seek(unsigned p) { 931 if (p != pos) { 932 if (p > bc.size()) { 933 bc.resize(p); 934 } 935 pos = p; 936 } 937 } 938 939 unsigned get_pos() { return pos; } 940 uint32_t *data() { return &bc[0]; } 941 942 bytecode & operator <<(uint32_t v) { 943 if (pos == ndw()) { 944 bc.push_back(v); 945 } else 946 bc.at(pos) = v; 947 ++pos; 948 return *this; 949 } 950 951 bytecode & operator <<(const hw_encoding_format &e) { 952 *this << e.get_value(hw_class_bit); 953 return *this; 954 } 955 956 bytecode & operator <<(const bytecode &b) { 957 bc.insert(bc.end(), b.bc.begin(), b.bc.end()); 958 return *this; 959 } 960 961 uint32_t at(unsigned dw_id) { return bc.at(dw_id); } 962}; 963 964 965class bc_builder { 966 shader &sh; 967 sb_context &ctx; 968 bytecode bb; 969 int error; 970 971public: 972 973 bc_builder(shader &s); 974 int build(); 975 bytecode& get_bytecode() { assert(!error); return bb; } 976 977private: 978 979 int build_cf(cf_node *n); 980 981 int build_cf_alu(cf_node *n); 982 int build_cf_mem(cf_node *n); 983 int build_cf_exp(cf_node *n); 984 985 int build_alu_clause(cf_node *n); 986 int build_alu_group(alu_group_node *n); 987 int build_alu(alu_node *n); 988 989 int build_fetch_clause(cf_node *n); 990 int build_fetch_tex(fetch_node *n); 991 int build_fetch_vtx(fetch_node *n); 992 int build_fetch_gds(fetch_node *n); 993 int build_fetch_mem(fetch_node* n); 994}; 995 996} // namespace r600_sb 997 998#endif /* SB_BC_H_ */ 999