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 static INLINE boolean
 148 nv_is_vector_op(uint opcode)
 149 {
 150    return (opcode >= NV_OP_TEX) && (opcode <= NV_OP_TXQ);
 151 }
 152
 153 static INLINE uint
 154 nv_type_order(ubyte type)
 155 {
 156    switch (type & 0xf) {
 157    case NV_TYPE_U8:
 158    case NV_TYPE_S8:
 159       return 0;
 160    case NV_TYPE_U16:
 161    case NV_TYPE_S16:
 162       return 1;
 163    case NV_TYPE_U32:
 164    case NV_TYPE_F32:
 165    case NV_TYPE_S32:
 166    case NV_TYPE_P32:
 167       return 2;
 168    case NV_TYPE_F64:
 169       return 3;
 170    }
 171    assert(0);
 172 }
 173
 174 static INLINE uint
 175 nv_type_sizeof(ubyte type)
 176 {
 177    if (type & 0xf0)
 178       return (1 << nv_type_order(type)) * (type >> 4);
 179    return 1 << nv_type_order(type);
 180 }
 181
 182 static INLINE uint
 183 nv_type_sizeof_base(ubyte type)
 184 {
 185    return 1 << nv_type_order(type);
 186 }
 187
 188 struct nv_reg {
 189    int id;
 190    ubyte file;
 191    ubyte type; /* type of generating instruction's result */
 192    union {
 193       float f32;
 194       double f64;
 195       int32_t s32;
 196       uint32_t u32;
 197    } imm;
 198 };
 199
 200 struct nv_range {
 201    struct nv_range *next;
 202    int bgn;
 203    int end;
 204 };
 205
 206 struct nv_value {
 207    struct nv_reg reg;
 208    struct nv_instruction *insn;
 209    struct nv_value *join;
 210    int n;
 211    struct nv_range *livei;
 212    int refc;
 213
 214    struct nv_value *next;
 215    struct nv_value *prev;
 216 };
 217
 218 struct nv_ref {
 219    struct nv_value *value;
 220    struct nv_instruction *insn;
 221    ubyte mod;
 222    ubyte typecast;
 223    ubyte flags; /* not used yet */
 224 };
 225
 226 struct nv_basic_block;
 227
 228 struct nv_instruction {
 229    struct nv_instruction *next;
 230    struct nv_instruction *prev;
 231    uint opcode;
 232    int serial;
 233    struct nv_value *def[4];
 234    struct nv_value *flags_def;
 235    struct nv_ref *src[5];
 236    struct nv_ref *flags_src;
 237    struct nv_basic_block *bb;
 238    struct nv_basic_block *target; /* target block of control flow insn */
 239    ubyte cc;
 240    ubyte set_cond      : 4;
 241    ubyte fixed         : 1; /* don't optimize away */
 242    ubyte is_terminator : 1;
 243    ubyte is_join       : 1;
 244    ubyte is_long       : 1; /* for emission */
 245    /* */
 246    ubyte saturate : 1;
 247    ubyte centroid : 1;
 248    ubyte flat     : 1;
 249    ubyte lanes    : 4;
 250    ubyte tex_live : 1;
 251    /* */
 252    ubyte tex_t; /* TIC binding */
 253    ubyte tex_s; /* TSC binding */
 254    ubyte tex_argc : 3;
 255    ubyte tex_cube : 1;
 256    ubyte tex_mask : 4;
 257    /* */
 258    ubyte quadop;
 259 };
 260
 261 #define CFG_EDGE_FORWARD     0
 262 #define CFG_EDGE_BACK        1
 263 #define CFG_EDGE_LOOP_ENTER  2
 264 #define CFG_EDGE_LOOP_LEAVE  4
 265 #define CFG_EDGE_FAKE        8
 266
 267 /* 'WALL' edge means where reachability check doesn't follow */
 268 /* 'LOOP' edge means just having to do with loops */
 269 #define IS_LOOP_EDGE(k) ((k) & 7)
 270 #define IS_WALL_EDGE(k) ((k) & 9)
 271
 272 struct nv_basic_block {
 273    struct nv_instruction *entry; /* first non-phi instruction */
 274    struct nv_instruction *exit;
 275    struct nv_instruction *phi; /* very first instruction */
 276    int num_instructions;
 277
 278    struct nv_basic_block *out[2]; /* no indirect branches -> 2 */
 279    struct nv_basic_block *in[8]; /* hope that suffices */
 280    uint num_in;
 281    ubyte out_kind[2];
 282    ubyte in_kind[8];
 283
 284    int id;
 285    struct nv_basic_block *last_visitor;
 286    uint priv;
 287    uint pass_seq;
 288
 289    uint32_t bin_pos; /* position, size in emitted code */
 290    uint32_t bin_size;
 291
 292    uint32_t live_set[NV_PC_MAX_VALUES / 32];
 293 };
 294
 295 #define NV_FIXUP_CFLOW_RELOC 0
 296 #define NV_FIXUP_PARAM_RELOC 1
 297
 298 struct nv_fixup {
 299    ubyte type;
 300    ubyte shift;
 301    uint32_t mask;
 302    uint32_t data;
 303    uint32_t offset;
 304 };
 305
 306 static INLINE void
 307 nv_fixup_apply(uint32_t *bin, struct nv_fixup *fixup, uint32_t data)
 308 {
 309    uint32_t val;
 310
 311    val = bin[fixup->offset / 4] & ~fixup->mask;
 312    data = (fixup->shift < 0) ? (data >> fixup->shift) : (data << fixup->shift);
 313    val |= (fixup->data + data) & fixup->mask;
 314    bin[fixup->offset / 4] = val;
 315 }
 316
 317 struct nv_pc {
 318    struct nv50_translation_info *ti;
 319
 320    struct nv_basic_block *root;
 321    struct nv_basic_block *current_block;
 322    struct nv_basic_block *parent_block;
 323
 324    int loop_nesting_bound;
 325    uint pass_seq;
 326
 327    struct nv_value values[NV_PC_MAX_VALUES];
 328    struct nv_instruction instructions[NV_PC_MAX_INSTRUCTIONS];
 329    struct nv_ref **refs;
 330    struct nv_basic_block **bb_list;
 331    int num_values;
 332    int num_instructions;
 333    int num_refs;
 334    int num_blocks;
 335
 336    int max_reg[4];
 337
 338    uint32_t *immd_buf; /* populated on emit */
 339    unsigned immd_count;
 340
 341    uint32_t *emit;
 342    unsigned bin_size;
 343    unsigned bin_pos;
 344
 345    struct nv_fixup *fixups;
 346    int num_fixups;
 347 };
 348
 349 void nvbb_insert_tail(struct nv_basic_block *, struct nv_instruction *);
 350 void nvi_insert_after(struct nv_instruction *, struct nv_instruction *);
 351
 352 static INLINE struct nv_instruction *
 353 nv_alloc_instruction(struct nv_pc *pc, uint opcode)
 354 {
 355    struct nv_instruction *insn;
 356
 357    insn = &pc->instructions[pc->num_instructions++];
 358    assert(pc->num_instructions < NV_PC_MAX_INSTRUCTIONS);
 359
 360    insn->cc = NV_CC_TR;
 361    insn->opcode = opcode;
 362
 363    return insn;
 364 }
 365
 366 static INLINE struct nv_instruction *
 367 new_instruction(struct nv_pc *pc, uint opcode)
 368 {
 369    struct nv_instruction *insn = nv_alloc_instruction(pc, opcode);
 370
 371    nvbb_insert_tail(pc->current_block, insn);
 372    return insn;
 373 }
 374
 375 static INLINE struct nv_instruction *
 376 new_instruction_at(struct nv_pc *pc, struct nv_instruction *at, uint opcode)
 377 {
 378    struct nv_instruction *insn = nv_alloc_instruction(pc, opcode);
 379
 380    nvi_insert_after(at, insn);
 381    return insn;
 382 }
 383
 384 static INLINE struct nv_value *
 385 new_value(struct nv_pc *pc, ubyte file, ubyte type)
 386 {
 387    struct nv_value *value = &pc->values[pc->num_values];
 388
 389    assert(pc->num_values < NV_PC_MAX_VALUES - 1);
 390
 391    value->n = pc->num_values++;
 392    value->join = value;
 393    value->reg.id = -1;
 394    value->reg.file = file;
 395    value->reg.type = type;
 396    return value;
 397 }
 398
 399 static INLINE struct nv_value *
 400 new_value_like(struct nv_pc *pc, struct nv_value *like)
 401 {
 402    return new_value(pc, like->reg.file, like->reg.type);
 403 }
 404
 405 static INLINE struct nv_ref *
 406 new_ref(struct nv_pc *pc, struct nv_value *val)
 407 {
 408    int i;
 409    struct nv_ref *ref;
 410
 411    if ((pc->num_refs % 64) == 0) {
 412       const unsigned old_size = pc->num_refs * sizeof(struct nv_ref *);
 413       const unsigned new_size = (pc->num_refs + 64) * sizeof(struct nv_ref *);
 414
 415       pc->refs = REALLOC(pc->refs, old_size, new_size);
 416
 417       ref = CALLOC(64, sizeof(struct nv_ref));
 418       for (i = 0; i < 64; ++i)
 419          pc->refs[pc->num_refs + i] = &ref[i];
 420    }
 421
 422    ref = pc->refs[pc->num_refs++];
 423    ref->value = val;
 424    ref->typecast = val->reg.type;
 425
 426    ++val->refc;
 427    return ref;
 428 }
 429
 430 static INLINE struct nv_basic_block *
 431 new_basic_block(struct nv_pc *pc)
 432 {
 433    struct nv_basic_block *bb = CALLOC_STRUCT(nv_basic_block);
 434
 435    bb->id = pc->num_blocks++;
 436    return bb;
 437 }
 438
 439 static INLINE void
 440 nv_reference(struct nv_pc *pc, struct nv_ref **d, struct nv_value *s)
 441 {
 442    if (*d)
 443       --(*d)->value->refc;
 444
 445    if (s) {
 446       if (!*d)
 447          *d = new_ref(pc, s);
 448       else {
 449          (*d)->value = s;
 450          ++(s->refc);
 451       }
 452    } else {
 453       *d = NULL;
 454    }
 455 }
 456
 457 /* nv50_emit.c */
 458 void nv50_emit_instruction(struct nv_pc *, struct nv_instruction *);
 459
 460 /* nv50_print.c */
 461 const char *nv_opcode_name(uint opcode);
 462 void nv_print_instruction(struct nv_instruction *);
 463
 464 /* nv50_pc.c */
 465
 466 void nv_print_program(struct nv_basic_block *b);
 467
 468 boolean nv_op_commutative(uint opcode);
 469 int nv50_indirect_opnd(struct nv_instruction *);
 470 boolean nv50_nvi_can_use_imm(struct nv_instruction *, int s);
 471 boolean nv50_nvi_can_predicate(struct nv_instruction *);
 472 boolean nv50_nvi_can_load(struct nv_instruction *, int s, struct nv_value *);
 473 ubyte nv50_supported_src_mods(uint opcode, int s);
 474 int nv_nvi_refcount(struct nv_instruction *);
 475 void nv_nvi_delete(struct nv_instruction *);
 476 void nv_nvi_permute(struct nv_instruction *, struct nv_instruction *);
 477 void nvbb_attach_block(struct nv_basic_block *parent,
 478                        struct nv_basic_block *, ubyte edge_kind);
 479 boolean nvbb_dominated_by(struct nv_basic_block *, struct nv_basic_block *);
 480 boolean nvbb_reachable_by(struct nv_basic_block *, struct nv_basic_block *,
 481                           struct nv_basic_block *);
 482 struct nv_basic_block *nvbb_dom_frontier(struct nv_basic_block *);
 483 int nvcg_replace_value(struct nv_pc *pc, struct nv_value *old_val,
 484                        struct nv_value *new_val);
 485 struct nv_value *nvcg_find_immediate(struct nv_ref *);
 486 struct nv_value *nvcg_find_constant(struct nv_ref *);
 487
 488 typedef void (*nv_pc_pass_func)(void *priv, struct nv_basic_block *b);
 489
 490 void nv_pc_pass_in_order(struct nv_basic_block *, nv_pc_pass_func, void *);
 491
 492 int nv_pc_exec_pass0(struct nv_pc *pc);
 493 int nv_pc_exec_pass1(struct nv_pc *pc);
 494 int nv_pc_exec_pass2(struct nv_pc *pc);
 495
 496 int nv50_tgsi_to_nc(struct nv_pc *, struct nv50_translation_info *);
 497
 498 #endif // NV50_COMPILER_H