src/gallium/drivers/nvc0/nvc0_pc.h

   1 /*
   2  * Copyright 2010 Christoph Bumiller
   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  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
  18  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
  19  * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  20  * SOFTWARE.
  21  */
  22
  23 #ifndef __NVC0_COMPILER_H__
  24 #define __NVC0_COMPILER_H__
  25
  26 #include <stdio.h>
  27
  28 #ifndef NOUVEAU_DBG
  29 #ifdef NOUVEAU_DEBUG
  30 # define NOUVEAU_DBG(args...) debug_printf(args);
  31 #else
  32 # define NOUVEAU_DBG(args...)
  33 #endif
  34 #endif
  35
  36 #ifndef NOUVEAU_ERR
  37 #define NOUVEAU_ERR(fmt, args...) \
  38    fprintf(stderr, "%s:%d -  "fmt, __FUNCTION__, __LINE__, ##args);
  39 #endif
  40
  41 #include "pipe/p_defines.h"
  42 #include "util/u_inlines.h"
  43 #include "util/u_memory.h"
  44 #include "util/u_double_list.h"
  45
  46 /* pseudo opcodes */
  47 #define NV_OP_UNDEF      0
  48 #define NV_OP_BIND       1
  49 #define NV_OP_MERGE      2
  50 #define NV_OP_PHI        3
  51 #define NV_OP_SELECT     4
  52 #define NV_OP_NOP        5
  53
  54 /**
  55  * BIND forces source operand i into the same register as destination operand i
  56  * SELECT forces its multiple source operands and its destination operand into
  57  *  one and the same register.
  58  */
  59
  60 /* base opcodes */
  61 #define NV_OP_LD         6
  62 #define NV_OP_ST         7
  63 #define NV_OP_MOV        8
  64 #define NV_OP_AND        9
  65 #define NV_OP_OR        10
  66 #define NV_OP_XOR       11
  67 #define NV_OP_SHL       12
  68 #define NV_OP_SHR       13
  69 #define NV_OP_NOT       14
  70 #define NV_OP_SET       15
  71 #define NV_OP_ADD       16
  72 #define NV_OP_SUB       17
  73 #define NV_OP_MUL       18
  74 #define NV_OP_MAD       19
  75 #define NV_OP_ABS       20
  76 #define NV_OP_NEG       21
  77 #define NV_OP_MAX       22
  78 #define NV_OP_MIN       23
  79 #define NV_OP_CVT       24
  80 #define NV_OP_CEIL      25
  81 #define NV_OP_FLOOR     26
  82 #define NV_OP_TRUNC     27
  83 #define NV_OP_SAD       28
  84
  85 /* shader opcodes */
  86 #define NV_OP_VFETCH    29
  87 #define NV_OP_PFETCH    30
  88 #define NV_OP_EXPORT    31
  89 #define NV_OP_LINTERP   32
  90 #define NV_OP_PINTERP   33
  91 #define NV_OP_EMIT      34
  92 #define NV_OP_RESTART   35
  93 #define NV_OP_TEX       36
  94 #define NV_OP_TXB       37
  95 #define NV_OP_TXL       38
  96 #define NV_OP_TXF       39
  97 #define NV_OP_TXQ       40
  98 #define NV_OP_QUADOP    41
  99 #define NV_OP_DFDX      42
 100 #define NV_OP_DFDY      43
 101 #define NV_OP_KIL       44
 102
 103 /* control flow opcodes */
 104 #define NV_OP_BRA       45
 105 #define NV_OP_CALL      46
 106 #define NV_OP_RET       47
 107 #define NV_OP_EXIT      48
 108 #define NV_OP_BREAK     49
 109 #define NV_OP_BREAKADDR 50
 110 #define NV_OP_JOINAT    51
 111 #define NV_OP_JOIN      52
 112
 113 /* typed opcodes */
 114 #define NV_OP_ADD_F32   NV_OP_ADD
 115 #define NV_OP_ADD_B32   53
 116 #define NV_OP_MUL_F32   NV_OP_MUL
 117 #define NV_OP_MUL_B32   54
 118 #define NV_OP_ABS_F32   NV_OP_ABS
 119 #define NV_OP_ABS_S32   55
 120 #define NV_OP_NEG_F32   NV_OP_NEG
 121 #define NV_OP_NEG_S32   56
 122 #define NV_OP_MAX_F32   NV_OP_MAX
 123 #define NV_OP_MAX_S32   57
 124 #define NV_OP_MAX_U32   58
 125 #define NV_OP_MIN_F32   NV_OP_MIN
 126 #define NV_OP_MIN_S32   59
 127 #define NV_OP_MIN_U32   60
 128 #define NV_OP_SET_F32   61
 129 #define NV_OP_SET_S32   62
 130 #define NV_OP_SET_U32   63
 131 #define NV_OP_SAR       64
 132 #define NV_OP_RCP       65
 133 #define NV_OP_RSQ       66
 134 #define NV_OP_LG2       67
 135 #define NV_OP_SIN       68
 136 #define NV_OP_COS       69
 137 #define NV_OP_EX2       70
 138 #define NV_OP_PRESIN    71
 139 #define NV_OP_PREEX2    72
 140 #define NV_OP_SAT       73
 141
 142 /* newly added opcodes */
 143 #define NV_OP_SET_F32_AND 74
 144 #define NV_OP_SET_F32_OR  75
 145 #define NV_OP_SET_F32_XOR 76
 146 #define NV_OP_SELP        77
 147 #define NV_OP_SLCT        78
 148 #define NV_OP_SLCT_F32    NV_OP_SLCT
 149 #define NV_OP_SLCT_S32    79
 150 #define NV_OP_SLCT_U32    80
 151 #define NV_OP_SUB_F32     NV_OP_SUB
 152 #define NV_OP_SUB_S32     81
 153 #define NV_OP_MAD_F32     NV_OP_MAD
 154 #define NV_OP_FSET_F32    82
 155
 156 #define NV_OP_COUNT     83
 157
 158 /* nv50 files omitted */
 159 #define NV_FILE_GPR      0
 160 #define NV_FILE_COND     1
 161 #define NV_FILE_PRED     2
 162 #define NV_FILE_IMM      16
 163 #define NV_FILE_MEM_S    32
 164 #define NV_FILE_MEM_V    34
 165 #define NV_FILE_MEM_A    35
 166 #define NV_FILE_MEM_L    48
 167 #define NV_FILE_MEM_G    64
 168 #define NV_FILE_MEM_C(i) (80 + i)
 169
 170 #define NV_IS_MEMORY_FILE(f) ((f) >= NV_FILE_MEM_S)
 171
 172 #define NV_MOD_NEG 1
 173 #define NV_MOD_ABS 2
 174 #define NV_MOD_NOT 4
 175 #define NV_MOD_SAT 8
 176
 177 #define NV_TYPE_U8  0x00
 178 #define NV_TYPE_S8  0x01
 179 #define NV_TYPE_U16 0x02
 180 #define NV_TYPE_S16 0x03
 181 #define NV_TYPE_U32 0x04
 182 #define NV_TYPE_S32 0x05
 183 #define NV_TYPE_P32 0x07
 184 #define NV_TYPE_F32 0x09
 185 #define NV_TYPE_F64 0x0b
 186 #define NV_TYPE_VEC(x, n) (NV_TYPE_##x | (n << 4))
 187 #define NV_TYPE_ANY 0xff
 188
 189 #define NV_TYPE_ISINT(t) ((t) < 7)
 190 #define NV_TYPE_ISSGD(t) ((t) & 1)
 191
 192 #define NV_CC_FL 0x0
 193 #define NV_CC_LT 0x1
 194 #define NV_CC_EQ 0x2
 195 #define NV_CC_LE 0x3
 196 #define NV_CC_GT 0x4
 197 #define NV_CC_NE 0x5
 198 #define NV_CC_GE 0x6
 199 #define NV_CC_U  0x8
 200 #define NV_CC_TR 0xf
 201 #define NV_CC_O  0x10
 202 #define NV_CC_C  0x11
 203 #define NV_CC_A  0x12
 204 #define NV_CC_S  0x13
 205
 206 #define NV_PC_MAX_INSTRUCTIONS 2048
 207 #define NV_PC_MAX_VALUES (NV_PC_MAX_INSTRUCTIONS * 4)
 208
 209 #define NV_PC_MAX_BASIC_BLOCKS 1024
 210
 211 struct nv_op_info {
 212    uint base;                /* e.g. ADD_S32 -> ADD */
 213    char name[12];
 214    uint8_t type;
 215    uint8_t mods;
 216    unsigned flow        : 1;
 217    unsigned commutative : 1;
 218    unsigned vector      : 1;
 219    unsigned predicate   : 1;
 220    unsigned pseudo      : 1;
 221    unsigned immediate   : 3;
 222    unsigned memory      : 3;
 223 };
 224
 225 extern struct nv_op_info nvc0_op_info_table[];
 226
 227 #define NV_BASEOP(op) (nvc0_op_info_table[op].base)
 228 #define NV_OPTYPE(op) (nvc0_op_info_table[op].type)
 229
 230 static INLINE uint
 231 nv_op_base(uint opcode)
 232 {
 233    return nvc0_op_info_table[opcode].base;
 234 }
 235
 236 static INLINE boolean
 237 nv_is_texture_op(uint opcode)
 238 {
 239    return (opcode >= NV_OP_TEX && opcode <= NV_OP_TXQ);
 240 }
 241
 242 static INLINE boolean
 243 nv_is_vector_op(uint opcode)
 244 {
 245    return nvc0_op_info_table[opcode].vector ? TRUE : FALSE;
 246 }
 247
 248 static INLINE boolean
 249 nv_op_commutative(uint opcode)
 250 {
 251    return nvc0_op_info_table[opcode].commutative ? TRUE : FALSE;
 252 }
 253
 254 static INLINE uint8_t
 255 nv_op_supported_src_mods(uint opcode)
 256 {
 257    return nvc0_op_info_table[opcode].mods;
 258 }
 259
 260 static INLINE boolean
 261 nv_op_predicateable(uint opcode)
 262 {
 263    return nvc0_op_info_table[opcode].predicate ? TRUE : FALSE;
 264 }
 265
 266 static INLINE uint
 267 nv_type_order(ubyte type)
 268 {
 269    switch (type & 0xf) {
 270    case NV_TYPE_U8:
 271    case NV_TYPE_S8:
 272       return 0;
 273    case NV_TYPE_U16:
 274    case NV_TYPE_S16:
 275       return 1;
 276    case NV_TYPE_U32:
 277    case NV_TYPE_F32:
 278    case NV_TYPE_S32:
 279    case NV_TYPE_P32:
 280       return 2;
 281    case NV_TYPE_F64:
 282       return 3;
 283    }
 284    assert(0);
 285    return 0;
 286 }
 287
 288 static INLINE uint
 289 nv_type_sizeof(ubyte type)
 290 {
 291    if (type & 0xf0)
 292       return (1 << nv_type_order(type)) * (type >> 4);
 293    return 1 << nv_type_order(type);
 294 }
 295
 296 static INLINE uint
 297 nv_type_sizeof_base(ubyte type)
 298 {
 299    return 1 << nv_type_order(type);
 300 }
 301
 302 struct nv_reg {
 303    uint32_t address; /* for memory locations */
 304    int id; /* for registers */
 305    ubyte file;
 306    ubyte size;
 307    union {
 308       int32_t s32;
 309       int64_t s64;
 310       uint64_t u64;
 311       uint32_t u32;
 312       float f32;
 313       double f64;
 314    } imm;
 315 };
 316
 317 struct nv_range {
 318    struct nv_range *next;
 319    int bgn;
 320    int end;
 321 };
 322
 323 struct nv_ref;
 324
 325 struct nv_value {
 326    struct nv_reg reg;
 327    struct nv_instruction *insn;
 328    struct nv_value *join;
 329    struct nv_ref *last_use;
 330    int n;
 331    struct nv_range *livei;
 332    int refc;
 333    struct nv_value *next;
 334    struct nv_value *prev;
 335 };
 336
 337 struct nv_ref {
 338    struct nv_value *value;
 339    struct nv_instruction *insn;
 340    struct list_head list; /* connects uses of the same value */
 341    uint8_t mod;
 342    uint8_t flags;
 343 };
 344
 345 struct nv_basic_block;
 346
 347 struct nv_instruction {
 348    struct nv_instruction *next;
 349    struct nv_instruction *prev;
 350    uint opcode;
 351    uint serial;
 352
 353    struct nv_value *def[5];
 354    struct nv_ref *src[6];
 355
 356    int8_t predicate; /* index of predicate src */
 357    int8_t indirect;  /* index of pointer src */
 358
 359    union {
 360       struct {
 361          uint8_t t; /* TIC binding */
 362          uint8_t s; /* TSC binding */
 363       } tex;
 364       struct {
 365          uint8_t d; /* output type */
 366          uint8_t s; /* input type */
 367       } cvt;
 368    } ext;
 369
 370    struct nv_basic_block *bb;
 371    struct nv_basic_block *target; /* target block of control flow insn */
 372
 373    unsigned cc         : 5; /* condition code */
 374    unsigned fixed      : 1; /* don't optimize away (prematurely) */
 375    unsigned terminator : 1;
 376    unsigned join       : 1;
 377    unsigned set_cond   : 4; /* 2nd byte */
 378    unsigned saturate   : 1;
 379    unsigned centroid   : 1;
 380    unsigned flat       : 1;
 381    unsigned patch      : 1;
 382    unsigned lanes      : 4; /* 3rd byte */
 383    unsigned tex_argc   : 3;
 384    unsigned tex_live   : 1;
 385    unsigned tex_cube   : 1; /* 4th byte */
 386    unsigned tex_mask   : 4;
 387
 388    uint8_t quadop;
 389 };
 390
 391 static INLINE int
 392 nvi_vector_size(struct nv_instruction *nvi)
 393 {
 394    int i;
 395    assert(nvi);
 396    for (i = 0; i < 5 && nvi->def[i]; ++i);
 397    return i;
 398 }
 399
 400 #define CFG_EDGE_FORWARD     0
 401 #define CFG_EDGE_BACK        1
 402 #define CFG_EDGE_LOOP_ENTER  2
 403 #define CFG_EDGE_LOOP_LEAVE  4
 404 #define CFG_EDGE_FAKE        8
 405
 406 /* 'WALL' edge means where reachability check doesn't follow */
 407 /* 'LOOP' edge means just having to do with loops */
 408 #define IS_LOOP_EDGE(k) ((k) & 7)
 409 #define IS_WALL_EDGE(k) ((k) & 9)
 410
 411 struct nv_basic_block {
 412    struct nv_instruction *entry; /* first non-phi instruction */
 413    struct nv_instruction *exit;
 414    struct nv_instruction *phi; /* very first instruction */
 415    int num_instructions;
 416
 417    struct nv_basic_block *out[2]; /* no indirect branches -> 2 */
 418    struct nv_basic_block *in[8]; /* hope that suffices */
 419    uint num_in;
 420    ubyte out_kind[2];
 421    ubyte in_kind[8];
 422
 423    int id;
 424    int subroutine;
 425    uint priv; /* reset to 0 after you're done */
 426    uint pass_seq;
 427
 428    uint32_t emit_pos; /* position, size in emitted code */
 429    uint32_t emit_size;
 430
 431    uint32_t live_set[NV_PC_MAX_VALUES / 32];
 432 };
 433
 434 struct nvc0_translation_info;
 435
 436 struct nv_pc {
 437    struct nv_basic_block **root;
 438    struct nv_basic_block *current_block;
 439    struct nv_basic_block *parent_block;
 440
 441    int loop_nesting_bound;
 442    uint pass_seq;
 443
 444    struct nv_value values[NV_PC_MAX_VALUES];
 445    struct nv_instruction instructions[NV_PC_MAX_INSTRUCTIONS];
 446    struct nv_ref **refs;
 447    struct nv_basic_block *bb_list[NV_PC_MAX_BASIC_BLOCKS];
 448    int num_values;
 449    int num_instructions;
 450    int num_refs;
 451    int num_blocks;
 452    int num_subroutines;
 453
 454    int max_reg[4];
 455
 456    uint32_t *immd_buf; /* populated on emit */
 457    unsigned immd_count;
 458
 459    uint32_t *emit;
 460    unsigned emit_size;
 461    unsigned emit_pos;
 462
 463    void *reloc_entries;
 464    unsigned num_relocs;
 465
 466    /* optimization enables */
 467    boolean opt_reload_elim;
 468    boolean is_fragprog;
 469 };
 470
 471 void nvc0_insn_append(struct nv_basic_block *, struct nv_instruction *);
 472 void nvc0_insn_insert_after(struct nv_instruction *, struct nv_instruction *);
 473
 474 static INLINE struct nv_instruction *
 475 nv_alloc_instruction(struct nv_pc *pc, uint opcode)
 476 {
 477    struct nv_instruction *insn;
 478
 479    insn = &pc->instructions[pc->num_instructions++];
 480    assert(pc->num_instructions < NV_PC_MAX_INSTRUCTIONS);
 481
 482    insn->opcode = opcode;
 483    insn->cc = 0;
 484    insn->indirect = -1;
 485    insn->predicate = -1;
 486
 487    return insn;
 488 }
 489
 490 static INLINE struct nv_instruction *
 491 new_instruction(struct nv_pc *pc, uint opcode)
 492 {
 493    struct nv_instruction *insn = nv_alloc_instruction(pc, opcode);
 494
 495    nvc0_insn_append(pc->current_block, insn);
 496    return insn;
 497 }
 498
 499 static INLINE struct nv_instruction *
 500 new_instruction_at(struct nv_pc *pc, struct nv_instruction *at, uint opcode)
 501 {
 502    struct nv_instruction *insn = nv_alloc_instruction(pc, opcode);
 503
 504    nvc0_insn_insert_after(at, insn);
 505    return insn;
 506 }
 507
 508 static INLINE struct nv_value *
 509 new_value(struct nv_pc *pc, ubyte file, ubyte size)
 510 {
 511    struct nv_value *value = &pc->values[pc->num_values];
 512
 513    assert(pc->num_values < NV_PC_MAX_VALUES - 1);
 514
 515    value->n = pc->num_values++;
 516    value->join = value;
 517    value->reg.id = -1;
 518    value->reg.file = file;
 519    value->reg.size = size;
 520    return value;
 521 }
 522
 523 static INLINE struct nv_value *
 524 new_value_like(struct nv_pc *pc, struct nv_value *like)
 525 {
 526    return new_value(pc, like->reg.file, like->reg.size);
 527 }
 528
 529 static INLINE struct nv_ref *
 530 new_ref(struct nv_pc *pc, struct nv_value *val)
 531 {
 532    int i;
 533    struct nv_ref *ref;
 534
 535    if ((pc->num_refs % 64) == 0) {
 536       const unsigned old_size = pc->num_refs * sizeof(struct nv_ref *);
 537       const unsigned new_size = (pc->num_refs + 64) * sizeof(struct nv_ref *);
 538
 539       pc->refs = REALLOC(pc->refs, old_size, new_size);
 540
 541       ref = CALLOC(64, sizeof(struct nv_ref));
 542       for (i = 0; i < 64; ++i)
 543          pc->refs[pc->num_refs + i] = &ref[i];
 544    }
 545
 546    ref = pc->refs[pc->num_refs++];
 547    ref->value = val;
 548
 549    LIST_INITHEAD(&ref->list);
 550
 551    ++val->refc;
 552    return ref;
 553 }
 554
 555 static INLINE struct nv_basic_block *
 556 new_basic_block(struct nv_pc *pc)
 557 {
 558    struct nv_basic_block *bb;
 559
 560    if (pc->num_blocks >= NV_PC_MAX_BASIC_BLOCKS)
 561       return NULL;
 562
 563    bb = CALLOC_STRUCT(nv_basic_block);
 564
 565    bb->id = pc->num_blocks;
 566    pc->bb_list[pc->num_blocks++] = bb;
 567    return bb;
 568 }
 569
 570 static INLINE void
 571 nv_reference(struct nv_pc *pc,
 572              struct nv_instruction *nvi, int c, struct nv_value *s)
 573 {
 574    struct nv_ref **d = &nvi->src[c];
 575    assert(c < 6);
 576
 577    if (*d) {
 578       --(*d)->value->refc;
 579       LIST_DEL(&(*d)->list);
 580    }
 581
 582    if (s) {
 583       if (!*d) {
 584          *d = new_ref(pc, s);
 585          (*d)->insn = nvi;
 586       } else {
 587          LIST_DEL(&(*d)->list);
 588          (*d)->value = s;
 589          ++(s->refc);
 590       }
 591       if (!s->last_use)
 592          s->last_use = *d;
 593       else
 594          LIST_ADDTAIL(&s->last_use->list, &(*d)->list);
 595
 596       s->last_use = *d;
 597       (*d)->insn = nvi;
 598    } else {
 599       *d = NULL;
 600    }
 601 }
 602
 603 /* nvc0_emit.c */
 604 void nvc0_emit_instruction(struct nv_pc *, struct nv_instruction *);
 605
 606 /* nvc0_print.c */
 607 const char *nvc0_opcode_name(uint opcode);
 608 void nvc0_print_instruction(struct nv_instruction *);
 609
 610 /* nvc0_pc.c */
 611 void nvc0_print_function(struct nv_basic_block *root);
 612 void nvc0_print_program(struct nv_pc *);
 613
 614 boolean nvc0_insn_can_load(struct nv_instruction *, int s,
 615                            struct nv_instruction *);
 616 boolean nvc0_insn_is_predicateable(struct nv_instruction *);
 617
 618 int nvc0_insn_refcount(struct nv_instruction *);
 619 void nvc0_insn_delete(struct nv_instruction *);
 620 void nvc0_insns_permute(struct nv_instruction *prev, struct nv_instruction *);
 621
 622 void nvc0_bblock_attach(struct nv_basic_block *parent,
 623                         struct nv_basic_block *child, ubyte edge_kind);
 624 boolean nvc0_bblock_dominated_by(struct nv_basic_block *,
 625                                  struct nv_basic_block *);
 626 boolean nvc0_bblock_reachable_by(struct nv_basic_block *future,
 627                                  struct nv_basic_block *past,
 628                                  struct nv_basic_block *final);
 629 struct nv_basic_block *nvc0_bblock_dom_frontier(struct nv_basic_block *);
 630
 631 int nvc0_pc_replace_value(struct nv_pc *pc,
 632                           struct nv_value *old_val,
 633                           struct nv_value *new_val);
 634
 635 struct nv_value *nvc0_pc_find_immediate(struct nv_ref *);
 636 struct nv_value *nvc0_pc_find_constant(struct nv_ref *);
 637
 638 typedef void (*nv_pc_pass_func)(void *priv, struct nv_basic_block *b);
 639
 640 void nvc0_pc_pass_in_order(struct nv_basic_block *, nv_pc_pass_func, void *);
 641
 642 int nvc0_pc_exec_pass0(struct nv_pc *pc);
 643 int nvc0_pc_exec_pass1(struct nv_pc *pc);
 644 int nvc0_pc_exec_pass2(struct nv_pc *pc);
 645
 646 int nvc0_tgsi_to_nc(struct nv_pc *, struct nvc0_translation_info *);
 647
 648 #endif // NV50_COMPILER_H