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  *  and the operands will be assigned consecutive registers (needed for TEX)
  57  * SELECT forces its multiple source operands and its destination operand into
  58  *  one and the same register.
  59  */
  60
  61 /* base opcodes */
  62 #define NV_OP_LD         6
  63 #define NV_OP_ST         7
  64 #define NV_OP_MOV        8
  65 #define NV_OP_AND        9
  66 #define NV_OP_OR        10
  67 #define NV_OP_XOR       11
  68 #define NV_OP_SHL       12
  69 #define NV_OP_SHR       13
  70 #define NV_OP_NOT       14
  71 #define NV_OP_SET       15
  72 #define NV_OP_ADD       16
  73 #define NV_OP_SUB       17
  74 #define NV_OP_MUL       18
  75 #define NV_OP_MAD       19
  76 #define NV_OP_ABS       20
  77 #define NV_OP_NEG       21
  78 #define NV_OP_MAX       22
  79 #define NV_OP_MIN       23
  80 #define NV_OP_CVT       24
  81 #define NV_OP_CEIL      25
  82 #define NV_OP_FLOOR     26
  83 #define NV_OP_TRUNC     27
  84 #define NV_OP_SAD       28
  85
  86 /* shader opcodes */
  87 #define NV_OP_VFETCH    29
  88 #define NV_OP_PFETCH    30
  89 #define NV_OP_EXPORT    31
  90 #define NV_OP_LINTERP   32
  91 #define NV_OP_PINTERP   33
  92 #define NV_OP_EMIT      34
  93 #define NV_OP_RESTART   35
  94 #define NV_OP_TEX       36
  95 #define NV_OP_TXB       37
  96 #define NV_OP_TXL       38
  97 #define NV_OP_TXF       39
  98 #define NV_OP_TXQ       40
  99 #define NV_OP_QUADOP    41
 100 #define NV_OP_DFDX      42
 101 #define NV_OP_DFDY      43
 102 #define NV_OP_KIL       44
 103
 104 /* control flow opcodes */
 105 #define NV_OP_BRA       45
 106 #define NV_OP_CALL      46
 107 #define NV_OP_RET       47
 108 #define NV_OP_EXIT      48
 109 #define NV_OP_BREAK     49
 110 #define NV_OP_BREAKADDR 50
 111 #define NV_OP_JOINAT    51
 112 #define NV_OP_JOIN      52
 113
 114 /* typed opcodes */
 115 #define NV_OP_ADD_F32   NV_OP_ADD
 116 #define NV_OP_ADD_B32   53
 117 #define NV_OP_MUL_F32   NV_OP_MUL
 118 #define NV_OP_MUL_B32   54
 119 #define NV_OP_ABS_F32   NV_OP_ABS
 120 #define NV_OP_ABS_S32   55
 121 #define NV_OP_NEG_F32   NV_OP_NEG
 122 #define NV_OP_NEG_S32   56
 123 #define NV_OP_MAX_F32   NV_OP_MAX
 124 #define NV_OP_MAX_S32   57
 125 #define NV_OP_MAX_U32   58
 126 #define NV_OP_MIN_F32   NV_OP_MIN
 127 #define NV_OP_MIN_S32   59
 128 #define NV_OP_MIN_U32   60
 129 #define NV_OP_SET_F32   61
 130 #define NV_OP_SET_S32   62
 131 #define NV_OP_SET_U32   63
 132 #define NV_OP_SAR       64
 133 #define NV_OP_RCP       65
 134 #define NV_OP_RSQ       66
 135 #define NV_OP_LG2       67
 136 #define NV_OP_SIN       68
 137 #define NV_OP_COS       69
 138 #define NV_OP_EX2       70
 139 #define NV_OP_PRESIN    71
 140 #define NV_OP_PREEX2    72
 141 #define NV_OP_SAT       73
 142
 143 /* newly added opcodes */
 144 #define NV_OP_SET_F32_AND 74
 145 #define NV_OP_SET_F32_OR  75
 146 #define NV_OP_SET_F32_XOR 76
 147 #define NV_OP_SELP        77
 148 #define NV_OP_SLCT        78
 149 #define NV_OP_SLCT_F32    NV_OP_SLCT
 150 #define NV_OP_SLCT_S32    79
 151 #define NV_OP_SLCT_U32    80
 152 #define NV_OP_SUB_F32     NV_OP_SUB
 153 #define NV_OP_SUB_S32     81
 154 #define NV_OP_MAD_F32     NV_OP_MAD
 155 #define NV_OP_FSET_F32    82
 156 #define NV_OP_TXG         83
 157
 158 #define NV_OP_COUNT     84
 159
 160 /* nv50 files omitted */
 161 #define NV_FILE_GPR      0
 162 #define NV_FILE_COND     1
 163 #define NV_FILE_PRED     2
 164 #define NV_FILE_IMM      16
 165 #define NV_FILE_MEM_S    32
 166 #define NV_FILE_MEM_V    34
 167 #define NV_FILE_MEM_A    35
 168 #define NV_FILE_MEM_L    48
 169 #define NV_FILE_MEM_G    64
 170 #define NV_FILE_MEM_C(i) (80 + i)
 171
 172 #define NV_IS_MEMORY_FILE(f) ((f) >= NV_FILE_MEM_S)
 173
 174 #define NV_MOD_NEG 1
 175 #define NV_MOD_ABS 2
 176 #define NV_MOD_NOT 4
 177 #define NV_MOD_SAT 8
 178
 179 #define NV_TYPE_U8  0x00
 180 #define NV_TYPE_S8  0x01
 181 #define NV_TYPE_U16 0x02
 182 #define NV_TYPE_S16 0x03
 183 #define NV_TYPE_U32 0x04
 184 #define NV_TYPE_S32 0x05
 185 #define NV_TYPE_P32 0x07
 186 #define NV_TYPE_F32 0x09
 187 #define NV_TYPE_F64 0x0b
 188 #define NV_TYPE_VEC(x, n) (NV_TYPE_##x | (n << 4))
 189 #define NV_TYPE_ANY 0xff
 190
 191 #define NV_TYPE_ISINT(t) ((t) < 7)
 192 #define NV_TYPE_ISSGD(t) ((t) & 1)
 193
 194 #define NV_CC_FL 0x0
 195 #define NV_CC_LT 0x1
 196 #define NV_CC_EQ 0x2
 197 #define NV_CC_LE 0x3
 198 #define NV_CC_GT 0x4
 199 #define NV_CC_NE 0x5
 200 #define NV_CC_GE 0x6
 201 #define NV_CC_U  0x8
 202 #define NV_CC_TR 0xf
 203 #define NV_CC_O  0x10
 204 #define NV_CC_C  0x11
 205 #define NV_CC_A  0x12
 206 #define NV_CC_S  0x13
 207
 208 #define NV_PC_MAX_INSTRUCTIONS 2048
 209 #define NV_PC_MAX_VALUES (NV_PC_MAX_INSTRUCTIONS * 4)
 210
 211 #define NV_PC_MAX_BASIC_BLOCKS 1024
 212
 213 struct nv_op_info {
 214    uint base;                /* e.g. ADD_S32 -> ADD */
 215    char name[12];
 216    uint8_t type;
 217    uint8_t mods;
 218    unsigned flow        : 1;
 219    unsigned commutative : 1;
 220    unsigned vector      : 1;
 221    unsigned predicate   : 1;
 222    unsigned pseudo      : 1;
 223    unsigned immediate   : 3;
 224    unsigned memory      : 3;
 225 };
 226
 227 extern struct nv_op_info nvc0_op_info_table[];
 228
 229 #define NV_BASEOP(op) (nvc0_op_info_table[op].base)
 230 #define NV_OPTYPE(op) (nvc0_op_info_table[op].type)
 231
 232 static INLINE uint
 233 nv_op_base(uint opcode)
 234 {
 235    return nvc0_op_info_table[opcode].base;
 236 }
 237
 238 static INLINE boolean
 239 nv_is_texture_op(uint opcode)
 240 {
 241    return (opcode >= NV_OP_TEX && opcode <= NV_OP_TXQ);
 242 }
 243
 244 static INLINE boolean
 245 nv_is_vector_op(uint opcode)
 246 {
 247    return nvc0_op_info_table[opcode].vector ? TRUE : FALSE;
 248 }
 249
 250 static INLINE boolean
 251 nv_op_commutative(uint opcode)
 252 {
 253    return nvc0_op_info_table[opcode].commutative ? TRUE : FALSE;
 254 }
 255
 256 static INLINE uint8_t
 257 nv_op_supported_src_mods(uint opcode)
 258 {
 259    return nvc0_op_info_table[opcode].mods;
 260 }
 261
 262 static INLINE boolean
 263 nv_op_predicateable(uint opcode)
 264 {
 265    return nvc0_op_info_table[opcode].predicate ? TRUE : FALSE;
 266 }
 267
 268 static INLINE uint
 269 nv_type_order(ubyte type)
 270 {
 271    switch (type & 0xf) {
 272    case NV_TYPE_U8:
 273    case NV_TYPE_S8:
 274       return 0;
 275    case NV_TYPE_U16:
 276    case NV_TYPE_S16:
 277       return 1;
 278    case NV_TYPE_U32:
 279    case NV_TYPE_F32:
 280    case NV_TYPE_S32:
 281    case NV_TYPE_P32:
 282       return 2;
 283    case NV_TYPE_F64:
 284       return 3;
 285    }
 286    assert(0);
 287    return 0;
 288 }
 289
 290 static INLINE uint
 291 nv_type_sizeof(ubyte type)
 292 {
 293    if (type & 0xf0)
 294       return (1 << nv_type_order(type)) * (type >> 4);
 295    return 1 << nv_type_order(type);
 296 }
 297
 298 static INLINE uint
 299 nv_type_sizeof_base(ubyte type)
 300 {
 301    return 1 << nv_type_order(type);
 302 }
 303
 304 struct nv_reg {
 305    uint32_t address; /* for memory locations */
 306    int id; /* for registers */
 307    ubyte file;
 308    ubyte size;
 309    union {
 310       int32_t s32;
 311       int64_t s64;
 312       uint64_t u64;
 313       uint32_t u32;
 314       float f32;
 315       double f64;
 316    } imm;
 317 };
 318
 319 struct nv_range {
 320    struct nv_range *next;
 321    int bgn;
 322    int end;
 323 };
 324
 325 struct nv_ref;
 326
 327 struct nv_value {
 328    struct nv_reg reg;
 329    struct nv_instruction *insn;
 330    struct nv_value *join;
 331    struct nv_ref *last_use;
 332    int n;
 333    struct nv_range *livei;
 334    int refc;
 335    struct nv_value *next;
 336    struct nv_value *prev;
 337 };
 338
 339 struct nv_ref {
 340    struct nv_value *value;
 341    struct nv_instruction *insn;
 342    struct list_head list; /* connects uses of the same value */
 343    uint8_t mod;
 344    uint8_t flags;
 345 };
 346
 347 struct nv_basic_block;
 348
 349 struct nv_instruction {
 350    struct nv_instruction *next;
 351    struct nv_instruction *prev;
 352    uint opcode;
 353    uint serial;
 354
 355    struct nv_value *def[5];
 356    struct nv_ref *src[6];
 357
 358    int8_t predicate; /* index of predicate src */
 359    int8_t indirect;  /* index of pointer src */
 360
 361    union {
 362       struct {
 363          uint8_t t; /* TIC binding */
 364          uint8_t s; /* TSC binding */
 365       } tex;
 366       struct {
 367          uint8_t d; /* output type */
 368          uint8_t s; /* input type */
 369       } cvt;
 370    } ext;
 371
 372    struct nv_basic_block *bb;
 373    struct nv_basic_block *target; /* target block of control flow insn */
 374
 375    unsigned cc         : 5; /* condition code */
 376    unsigned fixed      : 1; /* don't optimize away (prematurely) */
 377    unsigned terminator : 1;
 378    unsigned join       : 1;
 379    unsigned set_cond   : 4; /* 2nd byte */
 380    unsigned saturate   : 1;
 381    unsigned centroid   : 1;
 382    unsigned flat       : 1;
 383    unsigned patch      : 1;
 384    unsigned lanes      : 4; /* 3rd byte */
 385    unsigned tex_dim    : 2;
 386    unsigned tex_array  : 1;
 387    unsigned tex_cube   : 1;
 388    unsigned tex_shadow : 1; /* 4th byte */
 389    unsigned tex_live   : 1;
 390    unsigned tex_mask   : 4;
 391
 392    uint8_t quadop;
 393 };
 394
 395 static INLINE int
 396 nvi_vector_size(struct nv_instruction *nvi)
 397 {
 398    int i;
 399    assert(nvi);
 400    for (i = 0; i < 5 && nvi->def[i]; ++i);
 401    return i;
 402 }
 403
 404 #define CFG_EDGE_FORWARD     0
 405 #define CFG_EDGE_BACK        1
 406 #define CFG_EDGE_LOOP_ENTER  2
 407 #define CFG_EDGE_LOOP_LEAVE  4
 408 #define CFG_EDGE_FAKE        8
 409
 410 /* 'WALL' edge means where reachability check doesn't follow */
 411 /* 'LOOP' edge means just having to do with loops */
 412 #define IS_LOOP_EDGE(k) ((k) & 7)
 413 #define IS_WALL_EDGE(k) ((k) & 9)
 414
 415 struct nv_basic_block {
 416    struct nv_instruction *entry; /* first non-phi instruction */
 417    struct nv_instruction *exit;
 418    struct nv_instruction *phi; /* very first instruction */
 419    int num_instructions;
 420
 421    struct nv_basic_block *out[2]; /* no indirect branches -> 2 */
 422    struct nv_basic_block *in[8]; /* hope that suffices */
 423    uint num_in;
 424    ubyte out_kind[2];
 425    ubyte in_kind[8];
 426
 427    int id;
 428    int subroutine;
 429    uint priv; /* reset to 0 after you're done */
 430    uint pass_seq;
 431
 432    uint32_t emit_pos; /* position, size in emitted code (in bytes) */
 433    uint32_t emit_size;
 434
 435    uint32_t live_set[NV_PC_MAX_VALUES / 32];
 436 };
 437
 438 struct nvc0_translation_info;
 439
 440 struct nv_pc {
 441    struct nv_basic_block **root;
 442    struct nv_basic_block *current_block;
 443    struct nv_basic_block *parent_block;
 444
 445    int loop_nesting_bound;
 446    uint pass_seq;
 447
 448    struct nv_value values[NV_PC_MAX_VALUES];
 449    struct nv_instruction instructions[NV_PC_MAX_INSTRUCTIONS];
 450    struct nv_ref **refs;
 451    struct nv_basic_block *bb_list[NV_PC_MAX_BASIC_BLOCKS];
 452    int num_values;
 453    int num_instructions;
 454    int num_refs;
 455    int num_blocks;
 456    int num_subroutines;
 457
 458    int max_reg[4];
 459
 460    uint32_t *immd_buf; /* populated on emit */
 461    unsigned immd_count;
 462
 463    uint32_t *emit;
 464    uint32_t emit_size;
 465    uint32_t emit_pos;
 466
 467    void *reloc_entries;
 468    unsigned num_relocs;
 469
 470    /* optimization enables */
 471    boolean opt_reload_elim;
 472    boolean is_fragprog;
 473 };
 474
 475 void nvc0_insn_append(struct nv_basic_block *, struct nv_instruction *);
 476 void nvc0_insn_insert_after(struct nv_instruction *, struct nv_instruction *);
 477
 478 static INLINE struct nv_instruction *
 479 nv_alloc_instruction(struct nv_pc *pc, uint opcode)
 480 {
 481    struct nv_instruction *insn;
 482
 483    insn = &pc->instructions[pc->num_instructions++];
 484    assert(pc->num_instructions < NV_PC_MAX_INSTRUCTIONS);
 485
 486    insn->opcode = opcode;
 487    insn->cc = 0;
 488    insn->indirect = -1;
 489    insn->predicate = -1;
 490
 491    return insn;
 492 }
 493
 494 static INLINE struct nv_instruction *
 495 new_instruction(struct nv_pc *pc, uint opcode)
 496 {
 497    struct nv_instruction *insn = nv_alloc_instruction(pc, opcode);
 498
 499    nvc0_insn_append(pc->current_block, insn);
 500    return insn;
 501 }
 502
 503 static INLINE struct nv_instruction *
 504 new_instruction_at(struct nv_pc *pc, struct nv_instruction *at, uint opcode)
 505 {
 506    struct nv_instruction *insn = nv_alloc_instruction(pc, opcode);
 507
 508    nvc0_insn_insert_after(at, insn);
 509    return insn;
 510 }
 511
 512 static INLINE struct nv_value *
 513 new_value(struct nv_pc *pc, ubyte file, ubyte size)
 514 {
 515    struct nv_value *value = &pc->values[pc->num_values];
 516
 517    assert(pc->num_values < NV_PC_MAX_VALUES - 1);
 518
 519    value->n = pc->num_values++;
 520    value->join = value;
 521    value->reg.id = -1;
 522    value->reg.file = file;
 523    value->reg.size = size;
 524    return value;
 525 }
 526
 527 static INLINE struct nv_value *
 528 new_value_like(struct nv_pc *pc, struct nv_value *like)
 529 {
 530    return new_value(pc, like->reg.file, like->reg.size);
 531 }
 532
 533 static INLINE struct nv_ref *
 534 new_ref(struct nv_pc *pc, struct nv_value *val)
 535 {
 536    int i;
 537    struct nv_ref *ref;
 538
 539    if ((pc->num_refs % 64) == 0) {
 540       const unsigned old_size = pc->num_refs * sizeof(struct nv_ref *);
 541       const unsigned new_size = (pc->num_refs + 64) * sizeof(struct nv_ref *);
 542
 543       pc->refs = REALLOC(pc->refs, old_size, new_size);
 544
 545       ref = CALLOC(64, sizeof(struct nv_ref));
 546       for (i = 0; i < 64; ++i)
 547          pc->refs[pc->num_refs + i] = &ref[i];
 548    }
 549
 550    ref = pc->refs[pc->num_refs++];
 551    ref->value = val;
 552
 553    LIST_INITHEAD(&ref->list);
 554
 555    ++val->refc;
 556    return ref;
 557 }
 558
 559 static INLINE struct nv_basic_block *
 560 new_basic_block(struct nv_pc *pc)
 561 {
 562    struct nv_basic_block *bb;
 563
 564    if (pc->num_blocks >= NV_PC_MAX_BASIC_BLOCKS)
 565       return NULL;
 566
 567    bb = CALLOC_STRUCT(nv_basic_block);
 568
 569    bb->id = pc->num_blocks;
 570    pc->bb_list[pc->num_blocks++] = bb;
 571    return bb;
 572 }
 573
 574 static INLINE void
 575 nv_reference(struct nv_pc *pc,
 576              struct nv_instruction *nvi, int c, struct nv_value *s)
 577 {
 578    struct nv_ref **d = &nvi->src[c];
 579    assert(c < 6);
 580
 581    if (*d) {
 582       --(*d)->value->refc;
 583       LIST_DEL(&(*d)->list);
 584    }
 585
 586    if (s) {
 587       if (!*d) {
 588          *d = new_ref(pc, s);
 589          (*d)->insn = nvi;
 590       } else {
 591          LIST_DEL(&(*d)->list);
 592          (*d)->value = s;
 593          ++(s->refc);
 594       }
 595       if (!s->last_use)
 596          s->last_use = *d;
 597       else
 598          LIST_ADDTAIL(&s->last_use->list, &(*d)->list);
 599
 600       s->last_use = *d;
 601       (*d)->insn = nvi;
 602    } else {
 603       *d = NULL;
 604    }
 605 }
 606
 607 /* nvc0_emit.c */
 608 void nvc0_emit_instruction(struct nv_pc *, struct nv_instruction *);
 609
 610 /* nvc0_print.c */
 611 const char *nvc0_opcode_name(uint opcode);
 612 void nvc0_print_instruction(struct nv_instruction *);
 613
 614 /* nvc0_pc.c */
 615 void nvc0_print_function(struct nv_basic_block *root);
 616 void nvc0_print_program(struct nv_pc *);
 617
 618 boolean nvc0_insn_can_load(struct nv_instruction *, int s,
 619                            struct nv_instruction *);
 620 boolean nvc0_insn_is_predicateable(struct nv_instruction *);
 621
 622 int nvc0_insn_refcount(struct nv_instruction *);
 623 void nvc0_insn_delete(struct nv_instruction *);
 624 void nvc0_insns_permute(struct nv_instruction *prev, struct nv_instruction *);
 625
 626 void nvc0_bblock_attach(struct nv_basic_block *parent,
 627                         struct nv_basic_block *child, ubyte edge_kind);
 628 boolean nvc0_bblock_dominated_by(struct nv_basic_block *,
 629                                  struct nv_basic_block *);
 630 boolean nvc0_bblock_reachable_by(struct nv_basic_block *future,
 631                                  struct nv_basic_block *past,
 632                                  struct nv_basic_block *final);
 633 struct nv_basic_block *nvc0_bblock_dom_frontier(struct nv_basic_block *);
 634
 635 int nvc0_pc_replace_value(struct nv_pc *pc,
 636                           struct nv_value *old_val,
 637                           struct nv_value *new_val);
 638
 639 struct nv_value *nvc0_pc_find_immediate(struct nv_ref *);
 640 struct nv_value *nvc0_pc_find_constant(struct nv_ref *);
 641
 642 typedef void (*nv_pc_pass_func)(void *priv, struct nv_basic_block *b);
 643
 644 void nvc0_pc_pass_in_order(struct nv_basic_block *, nv_pc_pass_func, void *);
 645
 646 int nvc0_pc_exec_pass0(struct nv_pc *pc);
 647 int nvc0_pc_exec_pass1(struct nv_pc *pc);
 648 int nvc0_pc_exec_pass2(struct nv_pc *pc);
 649
 650 int nvc0_tgsi_to_nc(struct nv_pc *, struct nvc0_translation_info *);
 651
 652 #endif // NV50_COMPILER_H