src/gallium/drivers/nv50/nv50_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 __NV50_COMPILER_H__
  24 #define __NV50_COMPILER_H__
  25
  26 #ifdef NV50PC_DEBUG
  27 # define NV50_DBGMSG(args...) debug_printf(args)
  28 #else
  29 # define NV50_DBGMSG(args...)
  30 #endif
  31
  32 #include "pipe/p_defines.h"
  33 #include "util/u_inlines.h"
  34 #include "util/u_memory.h"
  35
  36 #define NV_OP_PHI       0
  37 #define NV_OP_EXTRACT   1
  38 #define NV_OP_COMBINE   2
  39 #define NV_OP_LDA       3
  40 #define NV_OP_STA       4
  41 #define NV_OP_MOV       5
  42 #define NV_OP_ADD       6
  43 #define NV_OP_SUB       7
  44 #define NV_OP_NEG       8
  45 #define NV_OP_MUL       9
  46 #define NV_OP_MAD       10
  47 #define NV_OP_CVT       11
  48 #define NV_OP_SAT       12
  49 #define NV_OP_NOT       13
  50 #define NV_OP_AND       14
  51 #define NV_OP_OR        15
  52 #define NV_OP_XOR       16
  53 #define NV_OP_SHL       17
  54 #define NV_OP_SHR       18
  55 #define NV_OP_RCP       19
  56 #define NV_OP_UNDEF     20
  57 #define NV_OP_RSQ       21
  58 #define NV_OP_LG2       22
  59 #define NV_OP_SIN       23
  60 #define NV_OP_COS       24
  61 #define NV_OP_EX2       25
  62 #define NV_OP_PRESIN    26
  63 #define NV_OP_PREEX2    27
  64 #define NV_OP_MIN       28
  65 #define NV_OP_MAX       29
  66 #define NV_OP_SET       30
  67 #define NV_OP_SAD       31
  68 #define NV_OP_KIL       32
  69 #define NV_OP_BRA       33
  70 #define NV_OP_CALL      34
  71 #define NV_OP_RET       35
  72 #define NV_OP_BREAK     36
  73 #define NV_OP_BREAKADDR 37
  74 #define NV_OP_JOINAT    38
  75 #define NV_OP_TEX       39
  76 #define NV_OP_TXB       40
  77 #define NV_OP_TXL       41
  78 #define NV_OP_TXF       42
  79 #define NV_OP_TXQ       43
  80 #define NV_OP_DFDX      44
  81 #define NV_OP_DFDY      45
  82 #define NV_OP_QUADOP    46
  83 #define NV_OP_LINTERP   47
  84 #define NV_OP_PINTERP   48
  85 #define NV_OP_ABS       49
  86 #define NV_OP_CEIL      50
  87 #define NV_OP_FLOOR     51
  88 #define NV_OP_TRUNC     52
  89 #define NV_OP_NOP       53
  90 #define NV_OP_SELECT    54
  91 #define NV_OP_EXPORT    55
  92 #define NV_OP_JOIN      56
  93 #define NV_OP_COUNT     57
  94
  95 #define NV_FILE_GPR      0
  96 #define NV_FILE_OUT      1
  97 #define NV_FILE_ADDR     2
  98 #define NV_FILE_FLAGS    3
  99 #define NV_FILE_IMM      16
 100 #define NV_FILE_MEM_S    32
 101 #define NV_FILE_MEM_P    33
 102 #define NV_FILE_MEM_V    34
 103 #define NV_FILE_MEM_L    48
 104 #define NV_FILE_MEM_G(i) (64 + i)
 105 #define NV_FILE_MEM_C(i) (80 + i)
 106
 107 #define NV_MOD_NEG 1
 108 #define NV_MOD_ABS 2
 109 #define NV_MOD_NOT 4
 110 #define NV_MOD_SAT 8
 111
 112 #define NV_TYPE_U8  0x00
 113 #define NV_TYPE_S8  0x01
 114 #define NV_TYPE_U16 0x02
 115 #define NV_TYPE_S16 0x03
 116 #define NV_TYPE_U32 0x04
 117 #define NV_TYPE_S32 0x05
 118 #define NV_TYPE_P32 0x07
 119 #define NV_TYPE_F32 0x09
 120 #define NV_TYPE_F64 0x0b
 121 #define NV_TYPE_VEC(x, n) (NV_TYPE_##x | (n << 4))
 122 #define NV_TYPE_LO  0x00
 123 #define NV_TYPE_HI  0x80
 124 #define NV_TYPE_ANY 0xff
 125
 126 #define NV_TYPE_ISINT(t) ((t) <= 5)
 127 #define NV_TYPE_ISFLT(t) ((t) & 0x08)
 128
 129 /* $cX registers contain 4 bits: OCSZ (Z is bit 0) */
 130 #define NV_CC_FL 0x0
 131 #define NV_CC_LT 0x1
 132 #define NV_CC_EQ 0x2
 133 #define NV_CC_LE 0x3
 134 #define NV_CC_GT 0x4
 135 #define NV_CC_NE 0x5
 136 #define NV_CC_GE 0x6
 137 #define NV_CC_U  0x8
 138 #define NV_CC_TR 0xf
 139 #define NV_CC_O  0x10
 140 #define NV_CC_C  0x11
 141 #define NV_CC_A  0x12
 142 #define NV_CC_S  0x13
 143
 144 #define NV_PC_MAX_INSTRUCTIONS 2048
 145 #define NV_PC_MAX_VALUES (NV_PC_MAX_INSTRUCTIONS * 4)
 146
 147 #define NV_PC_MAX_BASIC_BLOCKS 1024
 148
 149 static INLINE boolean
 150 nv_is_vector_op(uint opcode)
 151 {
 152    return (opcode >= NV_OP_TEX) && (opcode <= NV_OP_TXQ);
 153 }
 154
 155 static INLINE uint
 156 nv_type_order(ubyte type)
 157 {
 158    switch (type & 0xf) {
 159    case NV_TYPE_U8:
 160    case NV_TYPE_S8:
 161       return 0;
 162    case NV_TYPE_U16:
 163    case NV_TYPE_S16:
 164       return 1;
 165    case NV_TYPE_U32:
 166    case NV_TYPE_F32:
 167    case NV_TYPE_S32:
 168    case NV_TYPE_P32:
 169       return 2;
 170    case NV_TYPE_F64:
 171       return 3;
 172    }
 173    assert(0);
 174    return 0;
 175 }
 176
 177 static INLINE uint
 178 nv_type_sizeof(ubyte type)
 179 {
 180    if (type & 0xf0)
 181       return (1 << nv_type_order(type)) * (type >> 4);
 182    return 1 << nv_type_order(type);
 183 }
 184
 185 static INLINE uint
 186 nv_type_sizeof_base(ubyte type)
 187 {
 188    return 1 << nv_type_order(type);
 189 }
 190
 191 struct nv_reg {
 192    int id;
 193    ubyte file;
 194    ubyte type; /* type of generating instruction's result */
 195    ubyte as_type; /* default type for new references to this value */
 196    union {
 197       float f32;
 198       double f64;
 199       int32_t s32;
 200       uint32_t u32;
 201    } imm;
 202 };
 203
 204 struct nv_range {
 205    struct nv_range *next;
 206    int bgn;
 207    int end;
 208 };
 209
 210 struct nv_value {
 211    struct nv_reg reg;
 212    struct nv_instruction *insn;
 213    struct nv_value *join;
 214    int n;
 215    struct nv_range *livei;
 216    int refc;
 217
 218    struct nv_value *next;
 219    struct nv_value *prev;
 220 };
 221
 222 struct nv_ref {
 223    struct nv_value *value;
 224    ubyte mod;
 225    ubyte typecast;
 226    ubyte flags; /* not used yet */
 227 };
 228
 229 struct nv_basic_block;
 230
 231 struct nv_instruction {
 232    struct nv_instruction *next;
 233    struct nv_instruction *prev;
 234    uint opcode;
 235    int serial;
 236    struct nv_value *def[4];
 237    struct nv_value *flags_def;
 238    struct nv_ref *src[5];
 239    struct nv_ref *flags_src;
 240    struct nv_basic_block *bb;
 241    struct nv_basic_block *target; /* target block of control flow insn */
 242    ubyte cc;
 243    unsigned set_cond      : 4;
 244    unsigned fixed         : 1; /* don't optimize away */
 245    unsigned is_terminator : 1;
 246    unsigned is_join       : 1;
 247    unsigned is_long       : 1; /* for emission */
 248    /* */
 249    unsigned saturate : 1;
 250    unsigned centroid : 1;
 251    unsigned flat     : 1;
 252    unsigned lanes    : 4;
 253    unsigned tex_live : 1;
 254    /* */
 255    ubyte tex_t; /* TIC binding */
 256    ubyte tex_s; /* TSC binding */
 257    unsigned tex_argc : 3;
 258    unsigned tex_cube : 1;
 259    unsigned tex_mask : 4;
 260    /* */
 261    ubyte quadop;
 262 };
 263
 264 #define CFG_EDGE_FORWARD     0
 265 #define CFG_EDGE_BACK        1
 266 #define CFG_EDGE_LOOP_ENTER  2
 267 #define CFG_EDGE_LOOP_LEAVE  4
 268 #define CFG_EDGE_FAKE        8
 269
 270 /* 'WALL' edge means where reachability check doesn't follow */
 271 /* 'LOOP' edge means just having to do with loops */
 272 #define IS_LOOP_EDGE(k) ((k) & 7)
 273 #define IS_WALL_EDGE(k) ((k) & 9)
 274
 275 struct nv_basic_block {
 276    struct nv_instruction *entry; /* first non-phi instruction */
 277    struct nv_instruction *exit;
 278    struct nv_instruction *phi; /* very first instruction */
 279    int num_instructions;
 280
 281    struct nv_basic_block *out[2]; /* no indirect branches -> 2 */
 282    struct nv_basic_block *in[8]; /* hope that suffices */
 283    uint num_in;
 284    ubyte out_kind[2];
 285    ubyte in_kind[8];
 286
 287    int id;
 288    int subroutine;
 289    uint priv; /* reset to 0 after you're done */
 290    uint pass_seq;
 291
 292    uint32_t bin_pos; /* position, size in emitted code */
 293    uint32_t bin_size;
 294
 295    uint32_t live_set[NV_PC_MAX_VALUES / 32];
 296 };
 297
 298 struct nv50_translation_info;
 299
 300 struct nv_pc {
 301    struct nv_basic_block **root;
 302    struct nv_basic_block *current_block;
 303    struct nv_basic_block *parent_block;
 304
 305    int loop_nesting_bound;
 306    uint pass_seq;
 307
 308    struct nv_value values[NV_PC_MAX_VALUES];
 309    struct nv_instruction instructions[NV_PC_MAX_INSTRUCTIONS];
 310    struct nv_ref **refs;
 311    struct nv_basic_block *bb_list[NV_PC_MAX_BASIC_BLOCKS];
 312    int num_values;
 313    int num_instructions;
 314    int num_refs;
 315    int num_blocks;
 316    int num_subroutines;
 317
 318    int max_reg[4];
 319
 320    uint32_t *immd_buf; /* populated on emit */
 321    unsigned immd_count;
 322
 323    uint32_t *emit;
 324    unsigned bin_size;
 325    unsigned bin_pos;
 326
 327    void *fixups;
 328    unsigned num_fixups;
 329
 330    /* optimization enables */
 331    boolean opt_reload_elim;
 332 };
 333
 334 void nvbb_insert_tail(struct nv_basic_block *, struct nv_instruction *);
 335 void nvi_insert_after(struct nv_instruction *, struct nv_instruction *);
 336
 337 static INLINE struct nv_instruction *
 338 nv_alloc_instruction(struct nv_pc *pc, uint opcode)
 339 {
 340    struct nv_instruction *insn;
 341
 342    insn = &pc->instructions[pc->num_instructions++];
 343    assert(pc->num_instructions < NV_PC_MAX_INSTRUCTIONS);
 344
 345    insn->cc = NV_CC_TR;
 346    insn->opcode = opcode;
 347
 348    return insn;
 349 }
 350
 351 static INLINE struct nv_instruction *
 352 new_instruction(struct nv_pc *pc, uint opcode)
 353 {
 354    struct nv_instruction *insn = nv_alloc_instruction(pc, opcode);
 355
 356    nvbb_insert_tail(pc->current_block, insn);
 357    return insn;
 358 }
 359
 360 static INLINE struct nv_instruction *
 361 new_instruction_at(struct nv_pc *pc, struct nv_instruction *at, uint opcode)
 362 {
 363    struct nv_instruction *insn = nv_alloc_instruction(pc, opcode);
 364
 365    nvi_insert_after(at, insn);
 366    return insn;
 367 }
 368
 369 static INLINE struct nv_value *
 370 new_value(struct nv_pc *pc, ubyte file, ubyte type)
 371 {
 372    struct nv_value *value = &pc->values[pc->num_values];
 373
 374    assert(pc->num_values < NV_PC_MAX_VALUES - 1);
 375
 376    value->n = pc->num_values++;
 377    value->join = value;
 378    value->reg.id = -1;
 379    value->reg.file = file;
 380    value->reg.type = value->reg.as_type = type;
 381    return value;
 382 }
 383
 384 static INLINE struct nv_value *
 385 new_value_like(struct nv_pc *pc, struct nv_value *like)
 386 {
 387    struct nv_value *val = new_value(pc, like->reg.file, like->reg.type);
 388    val->reg.as_type = like->reg.as_type;
 389    return val;
 390 }
 391
 392 static INLINE struct nv_ref *
 393 new_ref(struct nv_pc *pc, struct nv_value *val)
 394 {
 395    int i;
 396    struct nv_ref *ref;
 397
 398    if ((pc->num_refs % 64) == 0) {
 399       const unsigned old_size = pc->num_refs * sizeof(struct nv_ref *);
 400       const unsigned new_size = (pc->num_refs + 64) * sizeof(struct nv_ref *);
 401
 402       pc->refs = REALLOC(pc->refs, old_size, new_size);
 403
 404       ref = CALLOC(64, sizeof(struct nv_ref));
 405       for (i = 0; i < 64; ++i)
 406          pc->refs[pc->num_refs + i] = &ref[i];
 407    }
 408
 409    ref = pc->refs[pc->num_refs++];
 410    ref->value = val;
 411    ref->typecast = val->reg.as_type;
 412
 413    ++val->refc;
 414    return ref;
 415 }
 416
 417 static INLINE struct nv_basic_block *
 418 new_basic_block(struct nv_pc *pc)
 419 {
 420    struct nv_basic_block *bb;
 421
 422    if (pc->num_blocks >= NV_PC_MAX_BASIC_BLOCKS)
 423       return NULL;
 424
 425    bb = CALLOC_STRUCT(nv_basic_block);
 426
 427    bb->id = pc->num_blocks;
 428    pc->bb_list[pc->num_blocks++] = bb;
 429    return bb;
 430 }
 431
 432 static INLINE void
 433 nv_reference(struct nv_pc *pc, struct nv_ref **d, struct nv_value *s)
 434 {
 435    if (*d)
 436       --(*d)->value->refc;
 437
 438    if (s) {
 439       if (!*d)
 440          *d = new_ref(pc, s);
 441       else {
 442          (*d)->value = s;
 443          ++(s->refc);
 444       }
 445    } else {
 446       *d = NULL;
 447    }
 448 }
 449
 450 /* nv50_emit.c */
 451 void nv50_emit_instruction(struct nv_pc *, struct nv_instruction *);
 452 unsigned nv50_inst_min_size(struct nv_instruction *);
 453
 454 /* nv50_print.c */
 455 const char *nv_opcode_name(uint opcode);
 456 void nv_print_instruction(struct nv_instruction *);
 457
 458 /* nv50_pc.c */
 459
 460 void nv_print_function(struct nv_basic_block *root);
 461 void nv_print_program(struct nv_pc *);
 462
 463 boolean nv_op_commutative(uint opcode);
 464 int nv50_indirect_opnd(struct nv_instruction *);
 465 boolean nv50_nvi_can_use_imm(struct nv_instruction *, int s);
 466 boolean nv50_nvi_can_predicate(struct nv_instruction *);
 467 boolean nv50_nvi_can_load(struct nv_instruction *, int s, struct nv_value *);
 468 boolean nv50_op_can_write_flags(uint opcode);
 469 ubyte nv50_supported_src_mods(uint opcode, int s);
 470 int nv_nvi_refcount(struct nv_instruction *);
 471 void nv_nvi_delete(struct nv_instruction *);
 472 void nv_nvi_permute(struct nv_instruction *, struct nv_instruction *);
 473 void nvbb_attach_block(struct nv_basic_block *parent,
 474                        struct nv_basic_block *, ubyte edge_kind);
 475 boolean nvbb_dominated_by(struct nv_basic_block *, struct nv_basic_block *);
 476 boolean nvbb_reachable_by(struct nv_basic_block *, struct nv_basic_block *,
 477                           struct nv_basic_block *);
 478 struct nv_basic_block *nvbb_dom_frontier(struct nv_basic_block *);
 479 int nvcg_replace_value(struct nv_pc *pc, struct nv_value *old_val,
 480                        struct nv_value *new_val);
 481 struct nv_value *nvcg_find_immediate(struct nv_ref *);
 482 struct nv_value *nvcg_find_constant(struct nv_ref *);
 483
 484 typedef void (*nv_pc_pass_func)(void *priv, struct nv_basic_block *b);
 485
 486 void nv_pc_pass_in_order(struct nv_basic_block *, nv_pc_pass_func, void *);
 487
 488 int nv_pc_exec_pass0(struct nv_pc *pc);
 489 int nv_pc_exec_pass1(struct nv_pc *pc);
 490 int nv_pc_exec_pass2(struct nv_pc *pc);
 491
 492 int nv50_tgsi_to_nc(struct nv_pc *, struct nv50_translation_info *);
 493
 494 #endif // NV50_COMPILER_H