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 "nv50/nv50_debug.h"
  27
  28 #include "pipe/p_defines.h"
  29 #include "util/u_inlines.h"
  30 #include "util/u_memory.h"
  31 #include "util/u_double_list.h"
  32
  33 /* pseudo opcodes */
  34 #define NV_OP_UNDEF      0
  35 #define NV_OP_BIND       1
  36 #define NV_OP_MERGE      2
  37 #define NV_OP_PHI        3
  38 #define NV_OP_SELECT     4
  39 #define NV_OP_NOP        5
  40
  41 /**
  42  * BIND forces source operand i into the same register as destination operand i,
  43  *  and the operands will be assigned consecutive registers (needed for TEX).
  44  *  Beware conflicts !
  45  * SELECT forces its multiple source operands and its destination operand into
  46  *  one and the same register.
  47  */
  48
  49 /* base opcodes */
  50 #define NV_OP_LD         6
  51 #define NV_OP_ST         7
  52 #define NV_OP_MOV        8
  53 #define NV_OP_AND        9
  54 #define NV_OP_OR        10
  55 #define NV_OP_XOR       11
  56 #define NV_OP_SHL       12
  57 #define NV_OP_SHR       13
  58 #define NV_OP_NOT       14
  59 #define NV_OP_SET       15
  60 #define NV_OP_ADD       16
  61 #define NV_OP_SUB       17
  62 #define NV_OP_MUL       18
  63 #define NV_OP_MAD       19
  64 #define NV_OP_ABS       20
  65 #define NV_OP_NEG       21
  66 #define NV_OP_MAX       22
  67 #define NV_OP_MIN       23
  68 #define NV_OP_CVT       24
  69 #define NV_OP_CEIL      25
  70 #define NV_OP_FLOOR     26
  71 #define NV_OP_TRUNC     27
  72 #define NV_OP_SAD       28
  73
  74 /* shader opcodes */
  75 #define NV_OP_VFETCH    29
  76 #define NV_OP_PFETCH    30
  77 #define NV_OP_EXPORT    31
  78 #define NV_OP_LINTERP   32
  79 #define NV_OP_PINTERP   33
  80 #define NV_OP_EMIT      34
  81 #define NV_OP_RESTART   35
  82 #define NV_OP_TEX       36
  83 #define NV_OP_TXB       37
  84 #define NV_OP_TXL       38
  85 #define NV_OP_TXF       39
  86 #define NV_OP_TXQ       40
  87 #define NV_OP_QUADOP    41
  88 #define NV_OP_DFDX      42
  89 #define NV_OP_DFDY      43
  90 #define NV_OP_KIL       44
  91
  92 /* control flow opcodes */
  93 #define NV_OP_BRA       45
  94 #define NV_OP_CALL      46
  95 #define NV_OP_RET       47
  96 #define NV_OP_EXIT      48
  97 #define NV_OP_BREAK     49
  98 #define NV_OP_BREAKADDR 50
  99 #define NV_OP_JOINAT    51
 100 #define NV_OP_JOIN      52
 101
 102 /* typed opcodes */
 103 #define NV_OP_ADD_F32   NV_OP_ADD
 104 #define NV_OP_ADD_B32   53
 105 #define NV_OP_MUL_F32   NV_OP_MUL
 106 #define NV_OP_MUL_B32   54
 107 #define NV_OP_ABS_F32   NV_OP_ABS
 108 #define NV_OP_ABS_S32   55
 109 #define NV_OP_NEG_F32   NV_OP_NEG
 110 #define NV_OP_NEG_S32   56
 111 #define NV_OP_MAX_F32   NV_OP_MAX
 112 #define NV_OP_MAX_S32   57
 113 #define NV_OP_MAX_U32   58
 114 #define NV_OP_MIN_F32   NV_OP_MIN
 115 #define NV_OP_MIN_S32   59
 116 #define NV_OP_MIN_U32   60
 117 #define NV_OP_SET_F32   61
 118 #define NV_OP_SET_S32   62
 119 #define NV_OP_SET_U32   63
 120 #define NV_OP_SAR       64
 121 #define NV_OP_RCP       65
 122 #define NV_OP_RSQ       66
 123 #define NV_OP_LG2       67
 124 #define NV_OP_SIN       68
 125 #define NV_OP_COS       69
 126 #define NV_OP_EX2       70
 127 #define NV_OP_PRESIN    71
 128 #define NV_OP_PREEX2    72
 129 #define NV_OP_SAT       73
 130
 131 /* newly added opcodes */
 132 #define NV_OP_SET_F32_AND 74
 133 #define NV_OP_SET_F32_OR  75
 134 #define NV_OP_SET_F32_XOR 76
 135 #define NV_OP_SELP        77
 136 #define NV_OP_SLCT        78
 137 #define NV_OP_SLCT_F32    NV_OP_SLCT
 138 #define NV_OP_SLCT_S32    79
 139 #define NV_OP_SLCT_U32    80
 140 #define NV_OP_SUB_F32     NV_OP_SUB
 141 #define NV_OP_SUB_S32     81
 142 #define NV_OP_MAD_F32     NV_OP_MAD
 143 #define NV_OP_FSET_F32    82
 144 #define NV_OP_TXG         83
 145
 146 #define NV_OP_COUNT     84
 147
 148 /* nv50 files omitted */
 149 #define NV_FILE_GPR      0
 150 #define NV_FILE_COND     1
 151 #define NV_FILE_PRED     2
 152 #define NV_FILE_IMM      16
 153 #define NV_FILE_MEM_S    32
 154 #define NV_FILE_MEM_V    34
 155 #define NV_FILE_MEM_A    35
 156 #define NV_FILE_MEM_L    48
 157 #define NV_FILE_MEM_G    64
 158 #define NV_FILE_MEM_C(i) (80 + i)
 159
 160 #define NV_IS_MEMORY_FILE(f) ((f) >= NV_FILE_MEM_S)
 161
 162 #define NV_MOD_NEG 1
 163 #define NV_MOD_ABS 2
 164 #define NV_MOD_NOT 4
 165 #define NV_MOD_SAT 8
 166
 167 #define NV_TYPE_U8  0x00
 168 #define NV_TYPE_S8  0x01
 169 #define NV_TYPE_U16 0x02
 170 #define NV_TYPE_S16 0x03
 171 #define NV_TYPE_U32 0x04
 172 #define NV_TYPE_S32 0x05
 173 #define NV_TYPE_P32 0x07
 174 #define NV_TYPE_F32 0x09
 175 #define NV_TYPE_F64 0x0b
 176 #define NV_TYPE_VEC(x, n) (NV_TYPE_##x | (n << 4))
 177 #define NV_TYPE_ANY 0xff
 178
 179 #define NV_TYPE_ISINT(t) ((t) < 7)
 180 #define NV_TYPE_ISSGD(t) ((t) & 1)
 181
 182 #define NV_CC_FL 0x0
 183 #define NV_CC_LT 0x1
 184 #define NV_CC_EQ 0x2
 185 #define NV_CC_LE 0x3
 186 #define NV_CC_GT 0x4
 187 #define NV_CC_NE 0x5
 188 #define NV_CC_GE 0x6
 189 #define NV_CC_U  0x8
 190 #define NV_CC_TR 0xf
 191 #define NV_CC_O  0x10
 192 #define NV_CC_C  0x11
 193 #define NV_CC_A  0x12
 194 #define NV_CC_S  0x13
 195 #define NV_CC_INVERSE(cc) ((cc) ^ 0x7)
 196 /* for 1 bit predicates: */
 197 #define NV_CC_P     0
 198 #define NV_CC_NOT_P 1
 199
 200 uint8_t nvc0_ir_reverse_cc(uint8_t cc);
 201
 202 #define NV_PC_MAX_INSTRUCTIONS 2048
 203 #define NV_PC_MAX_VALUES (NV_PC_MAX_INSTRUCTIONS * 4)
 204
 205 #define NV_PC_MAX_BASIC_BLOCKS 1024
 206
 207 struct nv_op_info {
 208    uint base;                /* e.g. ADD_S32 -> ADD */
 209    char name[12];
 210    uint8_t type;
 211    uint16_t mods;
 212    unsigned flow        : 1;
 213    unsigned commutative : 1;
 214    unsigned vector      : 1;
 215    unsigned predicate   : 1;
 216    unsigned pseudo      : 1;
 217    unsigned immediate   : 3;
 218    unsigned memory      : 3;
 219 };
 220
 221 extern struct nv_op_info nvc0_op_info_table[];
 222
 223 #define NV_BASEOP(op) (nvc0_op_info_table[op].base)
 224 #define NV_OPTYPE(op) (nvc0_op_info_table[op].type)
 225
 226 static INLINE boolean
 227 nv_is_texture_op(uint opcode)
 228 {
 229    return (opcode >= NV_OP_TEX && opcode <= NV_OP_TXQ);
 230 }
 231
 232 static INLINE boolean
 233 nv_is_vector_op(uint opcode)
 234 {
 235    return nvc0_op_info_table[opcode].vector ? TRUE : FALSE;
 236 }
 237
 238 static INLINE boolean
 239 nv_op_commutative(uint opcode)
 240 {
 241    return nvc0_op_info_table[opcode].commutative ? TRUE : FALSE;
 242 }
 243
 244 static INLINE uint8_t
 245 nv_op_supported_src_mods(uint opcode, int s)
 246 {
 247    return (nvc0_op_info_table[opcode].mods >> (s * 4)) & 0xf;
 248 }
 249
 250 static INLINE uint
 251 nv_type_order(ubyte type)
 252 {
 253    switch (type & 0xf) {
 254    case NV_TYPE_U8:
 255    case NV_TYPE_S8:
 256       return 0;
 257    case NV_TYPE_U16:
 258    case NV_TYPE_S16:
 259       return 1;
 260    case NV_TYPE_U32:
 261    case NV_TYPE_F32:
 262    case NV_TYPE_S32:
 263    case NV_TYPE_P32:
 264       return 2;
 265    case NV_TYPE_F64:
 266       return 3;
 267    }
 268    assert(0);
 269    return 0;
 270 }
 271
 272 static INLINE uint
 273 nv_type_sizeof(ubyte type)
 274 {
 275    if (type & 0xf0)
 276       return (1 << nv_type_order(type)) * (type >> 4);
 277    return 1 << nv_type_order(type);
 278 }
 279
 280 static INLINE uint
 281 nv_type_sizeof_base(ubyte type)
 282 {
 283    return 1 << nv_type_order(type);
 284 }
 285
 286 struct nv_reg {
 287    uint32_t address; /* for memory locations */
 288    int id; /* for registers */
 289    ubyte file;
 290    ubyte size;
 291    union {
 292       int32_t s32;
 293       int64_t s64;
 294       uint64_t u64;
 295       uint32_t u32; /* expected to be 0 for $r63 */
 296       float f32;
 297       double f64;
 298    } imm;
 299 };
 300
 301 struct nv_range {
 302    struct nv_range *next;
 303    int bgn;
 304    int end;
 305 };
 306
 307 struct nv_ref;
 308
 309 struct nv_value {
 310    struct nv_reg reg;
 311    struct nv_instruction *insn;
 312    struct nv_value *join;
 313    struct nv_ref *last_use;
 314    int n;
 315    struct nv_range *livei;
 316    int refc;
 317    struct nv_value *next;
 318    struct nv_value *prev;
 319 };
 320
 321 struct nv_ref {
 322    struct nv_value *value;
 323    struct nv_instruction *insn;
 324    struct list_head list; /* connects uses of the same value */
 325    uint8_t mod;
 326    uint8_t flags;
 327 };
 328
 329 #define NV_REF_FLAG_REGALLOC_PRIV (1 << 0)
 330
 331 struct nv_basic_block;
 332
 333 struct nv_instruction {
 334    struct nv_instruction *next;
 335    struct nv_instruction *prev;
 336    uint opcode;
 337    uint serial;
 338
 339    struct nv_value *def[5];
 340    struct nv_ref *src[6];
 341
 342    int8_t predicate; /* index of predicate src */
 343    int8_t indirect;  /* index of pointer src */
 344
 345    union {
 346       struct {
 347          uint8_t t; /* TIC binding */
 348          uint8_t s; /* TSC binding */
 349       } tex;
 350       struct {
 351          uint8_t d; /* output type */
 352          uint8_t s; /* input type */
 353       } cvt;
 354    } ext;
 355
 356    struct nv_basic_block *bb;
 357    struct nv_basic_block *target; /* target block of control flow insn */
 358
 359    unsigned cc         : 5; /* condition code */
 360    unsigned fixed      : 1; /* don't optimize away (prematurely) */
 361    unsigned terminator : 1;
 362    unsigned join       : 1;
 363    unsigned set_cond   : 4; /* 2nd byte */
 364    unsigned saturate   : 1;
 365    unsigned centroid   : 1;
 366    unsigned flat       : 1;
 367    unsigned patch      : 1;
 368    unsigned lanes      : 4; /* 3rd byte */
 369    unsigned tex_dim    : 2;
 370    unsigned tex_array  : 1;
 371    unsigned tex_cube   : 1;
 372    unsigned tex_shadow : 1; /* 4th byte */
 373    unsigned tex_live   : 1;
 374    unsigned tex_mask   : 4;
 375
 376    uint8_t quadop;
 377 };
 378
 379 static INLINE int
 380 nvi_vector_size(struct nv_instruction *nvi)
 381 {
 382    int i;
 383    assert(nvi);
 384    for (i = 0; i < 5 && nvi->def[i]; ++i);
 385    return i;
 386 }
 387
 388 #define CFG_EDGE_FORWARD     0
 389 #define CFG_EDGE_BACK        1
 390 #define CFG_EDGE_LOOP_ENTER  2
 391 #define CFG_EDGE_LOOP_LEAVE  4
 392 #define CFG_EDGE_FAKE        8
 393
 394 /* 'WALL' edge means where reachability check doesn't follow */
 395 /* 'LOOP' edge means just having to do with loops */
 396 #define IS_LOOP_EDGE(k) ((k) & 7)
 397 #define IS_WALL_EDGE(k) ((k) & 9)
 398
 399 struct nv_basic_block {
 400    struct nv_instruction *entry; /* first non-phi instruction */
 401    struct nv_instruction *exit;
 402    struct nv_instruction *phi; /* very first instruction */
 403    int num_instructions;
 404
 405    struct nv_basic_block *out[2]; /* no indirect branches -> 2 */
 406    struct nv_basic_block *in[8]; /* hope that suffices */
 407    uint num_in;
 408    ubyte out_kind[2];
 409    ubyte in_kind[8];
 410
 411    int id;
 412    int subroutine;
 413    uint priv; /* reset to 0 after you're done */
 414    uint pass_seq;
 415
 416    uint32_t emit_pos; /* position, size in emitted code (in bytes) */
 417    uint32_t emit_size;
 418
 419    uint32_t live_set[NV_PC_MAX_VALUES / 32];
 420 };
 421
 422 struct nvc0_translation_info;
 423
 424 struct nv_pc {
 425    struct nv_basic_block **root;
 426    struct nv_basic_block *current_block;
 427    struct nv_basic_block *parent_block;
 428
 429    int loop_nesting_bound;
 430    uint pass_seq;
 431
 432    struct nv_value values[NV_PC_MAX_VALUES];
 433    struct nv_instruction instructions[NV_PC_MAX_INSTRUCTIONS];
 434    struct nv_ref **refs;
 435    struct nv_basic_block *bb_list[NV_PC_MAX_BASIC_BLOCKS];
 436    int num_values;
 437    int num_instructions;
 438    int num_refs;
 439    int num_blocks;
 440    int num_subroutines;
 441
 442    int max_reg[4];
 443
 444    uint32_t *immd_buf; /* populated on emit */
 445    unsigned immd_count;
 446
 447    uint32_t *emit;
 448    uint32_t emit_size;
 449    uint32_t emit_pos;
 450
 451    void *reloc_entries;
 452    unsigned num_relocs;
 453
 454    /* optimization enables */
 455    boolean opt_reload_elim;
 456    boolean is_fragprog;
 457 };
 458
 459 void nvc0_insn_append(struct nv_basic_block *, struct nv_instruction *);
 460 void nvc0_insn_insert_before(struct nv_instruction *, struct nv_instruction *);
 461 void nvc0_insn_insert_after(struct nv_instruction *, struct nv_instruction *);
 462
 463 static INLINE struct nv_instruction *
 464 nv_alloc_instruction(struct nv_pc *pc, uint opcode)
 465 {
 466    struct nv_instruction *insn;
 467
 468    insn = &pc->instructions[pc->num_instructions++];
 469    assert(pc->num_instructions < NV_PC_MAX_INSTRUCTIONS);
 470
 471    insn->opcode = opcode;
 472    insn->cc = NV_CC_P;
 473    insn->indirect = -1;
 474    insn->predicate = -1;
 475
 476    return insn;
 477 }
 478
 479 static INLINE struct nv_instruction *
 480 new_instruction(struct nv_pc *pc, uint opcode)
 481 {
 482    struct nv_instruction *insn = nv_alloc_instruction(pc, opcode);
 483
 484    nvc0_insn_append(pc->current_block, insn);
 485    return insn;
 486 }
 487
 488 static INLINE struct nv_instruction *
 489 new_instruction_at(struct nv_pc *pc, struct nv_instruction *at, uint opcode)
 490 {
 491    struct nv_instruction *insn = nv_alloc_instruction(pc, opcode);
 492
 493    nvc0_insn_insert_after(at, insn);
 494    return insn;
 495 }
 496
 497 static INLINE struct nv_value *
 498 new_value(struct nv_pc *pc, ubyte file, ubyte size)
 499 {
 500    struct nv_value *value = &pc->values[pc->num_values];
 501
 502    assert(pc->num_values < NV_PC_MAX_VALUES - 1);
 503
 504    value->n = pc->num_values++;
 505    value->join = value;
 506    value->reg.id = -1;
 507    value->reg.file = file;
 508    value->reg.size = size;
 509    return value;
 510 }
 511
 512 static INLINE struct nv_value *
 513 new_value_like(struct nv_pc *pc, struct nv_value *like)
 514 {
 515    return new_value(pc, like->reg.file, like->reg.size);
 516 }
 517
 518 static INLINE struct nv_ref *
 519 new_ref(struct nv_pc *pc, struct nv_value *val)
 520 {
 521    int i;
 522    struct nv_ref *ref;
 523
 524    if ((pc->num_refs % 64) == 0) {
 525       const unsigned old_size = pc->num_refs * sizeof(struct nv_ref *);
 526       const unsigned new_size = (pc->num_refs + 64) * sizeof(struct nv_ref *);
 527
 528       pc->refs = REALLOC(pc->refs, old_size, new_size);
 529
 530       ref = CALLOC(64, sizeof(struct nv_ref));
 531       for (i = 0; i < 64; ++i)
 532          pc->refs[pc->num_refs + i] = &ref[i];
 533    }
 534
 535    ref = pc->refs[pc->num_refs++];
 536    ref->value = val;
 537
 538    LIST_INITHEAD(&ref->list);
 539
 540    ++val->refc;
 541    return ref;
 542 }
 543
 544 static INLINE struct nv_basic_block *
 545 new_basic_block(struct nv_pc *pc)
 546 {
 547    struct nv_basic_block *bb;
 548
 549    if (pc->num_blocks >= NV_PC_MAX_BASIC_BLOCKS)
 550       return NULL;
 551
 552    bb = CALLOC_STRUCT(nv_basic_block);
 553
 554    bb->id = pc->num_blocks;
 555    pc->bb_list[pc->num_blocks++] = bb;
 556    return bb;
 557 }
 558
 559 static INLINE void
 560 nv_reference(struct nv_pc *pc,
 561              struct nv_instruction *nvi, int c, struct nv_value *s)
 562 {
 563    struct nv_ref **d = &nvi->src[c];
 564    assert(c < 6);
 565
 566    if (*d) {
 567       --(*d)->value->refc;
 568       LIST_DEL(&(*d)->list);
 569    }
 570
 571    if (s) {
 572       if (!*d) {
 573          *d = new_ref(pc, s);
 574          (*d)->insn = nvi;
 575       } else {
 576          LIST_DEL(&(*d)->list);
 577          (*d)->value = s;
 578          ++(s->refc);
 579       }
 580       if (!s->last_use)
 581          s->last_use = *d;
 582       else
 583          LIST_ADDTAIL(&s->last_use->list, &(*d)->list);
 584
 585       s->last_use = *d;
 586       (*d)->insn = nvi;
 587    } else {
 588       *d = NULL;
 589    }
 590 }
 591
 592 /* nvc0_emit.c */
 593 void nvc0_emit_instruction(struct nv_pc *, struct nv_instruction *);
 594
 595 /* nvc0_print.c */
 596 const char *nvc0_opcode_name(uint opcode);
 597 void nvc0_print_instruction(struct nv_instruction *);
 598
 599 /* nvc0_pc.c */
 600 void nvc0_print_function(struct nv_basic_block *root);
 601 void nvc0_print_program(struct nv_pc *);
 602
 603 boolean nvc0_insn_can_load(struct nv_instruction *, int s,
 604                            struct nv_instruction *);
 605 boolean nvc0_insn_is_predicateable(struct nv_instruction *);
 606
 607 int nvc0_insn_refcount(struct nv_instruction *);
 608 void nvc0_insn_delete(struct nv_instruction *);
 609 void nvc0_insns_permute(struct nv_instruction *prev, struct nv_instruction *);
 610
 611 void nvc0_bblock_attach(struct nv_basic_block *parent,
 612                         struct nv_basic_block *child, ubyte edge_kind);
 613 boolean nvc0_bblock_dominated_by(struct nv_basic_block *,
 614                                  struct nv_basic_block *);
 615 boolean nvc0_bblock_reachable_by(struct nv_basic_block *future,
 616                                  struct nv_basic_block *past,
 617                                  struct nv_basic_block *final);
 618 struct nv_basic_block *nvc0_bblock_dom_frontier(struct nv_basic_block *);
 619
 620 int nvc0_pc_replace_value(struct nv_pc *pc,
 621                           struct nv_value *old_val,
 622                           struct nv_value *new_val);
 623
 624 struct nv_value *nvc0_pc_find_immediate(struct nv_ref *);
 625 struct nv_value *nvc0_pc_find_constant(struct nv_ref *);
 626
 627 typedef void (*nv_pc_pass_func)(void *priv, struct nv_basic_block *b);
 628
 629 void nvc0_pc_pass_in_order(struct nv_basic_block *, nv_pc_pass_func, void *);
 630
 631 int nvc0_pc_exec_pass0(struct nv_pc *pc);
 632 int nvc0_pc_exec_pass1(struct nv_pc *pc);
 633 int nvc0_pc_exec_pass2(struct nv_pc *pc);
 634
 635 int nvc0_tgsi_to_nc(struct nv_pc *, struct nvc0_translation_info *);
 636
 637 #endif // NV50_COMPILER_H