src/mesa/shader/slang/slang_emit.c

   1 /*
   2  * Mesa 3-D graphics library
   3  * Version:  6.5.3
   4  *
   5  * Copyright (C) 2005-2007  Brian Paul   All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the "Software"),
   9  * to deal in the Software without restriction, including without limitation
  10  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  11  * and/or sell copies of the Software, and to permit persons to whom the
  12  * Software is furnished to do so, subject to the following conditions:
  13  *
  14  * The above copyright notice and this permission notice shall be included
  15  * in all copies or substantial portions of the Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  18  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
  21  * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  22  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  23  */
  24
  25 /**
  26  * \file slang_emit.c
  27  * Emit program instructions (PI code) from IR trees.
  28  * \author Brian Paul
  29  */
  30
  31 /***
  32  *** NOTES
  33  ***
  34  *** To emit GPU instructions, we basically just do an in-order traversal
  35  *** of the IR tree.
  36  ***/
  37
  38
  39 #include "imports.h"
  40 #include "context.h"
  41 #include "macros.h"
  42 #include "program.h"
  43 #include "prog_instruction.h"
  44 #include "prog_parameter.h"
  45 #include "prog_print.h"
  46 #include "slang_builtin.h"
  47 #include "slang_emit.h"
  48
  49
  50 #define PEEPHOLE_OPTIMIZATIONS 1
  51 #define ANNOTATE 0
  52
  53
  54 /* XXX temporarily here */
  55
  56
  57 typedef struct
  58 {
  59    slang_info_log *log;
  60    slang_var_table *vt;
  61    struct gl_program *prog;
  62    /* code-gen options */
  63    GLboolean EmitHighLevelInstructions;
  64    GLboolean EmitComments;
  65 } slang_emit_info;
  66
  67
  68 /**
  69  * Assembly and IR info
  70  */
  71 typedef struct
  72 {
  73    slang_ir_opcode IrOpcode;
  74    const char *IrName;
  75    gl_inst_opcode InstOpcode;
  76    GLuint ResultSize, NumParams;
  77 } slang_ir_info;
  78
  79
  80
  81 static const slang_ir_info IrInfo[] = {
  82    /* binary ops */
  83    { IR_ADD, "IR_ADD", OPCODE_ADD, 4, 2 },
  84    { IR_SUB, "IR_SUB", OPCODE_SUB, 4, 2 },
  85    { IR_MUL, "IR_MUL", OPCODE_MUL, 4, 2 },
  86    { IR_DIV, "IR_DIV", OPCODE_NOP, 0, 2 }, /* XXX broke */
  87    { IR_DOT4, "IR_DOT_4", OPCODE_DP4, 1, 2 },
  88    { IR_DOT3, "IR_DOT_3", OPCODE_DP3, 1, 2 },
  89    { IR_CROSS, "IR_CROSS", OPCODE_XPD, 3, 2 },
  90    { IR_LRP, "IR_LRP", OPCODE_LRP, 4, 3 },
  91    { IR_MIN, "IR_MIN", OPCODE_MIN, 4, 2 },
  92    { IR_MAX, "IR_MAX", OPCODE_MAX, 4, 2 },
  93    { IR_CLAMP, "IR_CLAMP", OPCODE_NOP, 4, 3 }, /* special case: emit_clamp() */
  94    { IR_SEQUAL, "IR_SEQUAL", OPCODE_SEQ, 4, 2 },
  95    { IR_SNEQUAL, "IR_SNEQUAL", OPCODE_SNE, 4, 2 },
  96    { IR_SGE, "IR_SGE", OPCODE_SGE, 4, 2 },
  97    { IR_SGT, "IR_SGT", OPCODE_SGT, 4, 2 },
  98    { IR_SLE, "IR_SLE", OPCODE_SLE, 4, 2 },
  99    { IR_SLT, "IR_SLT", OPCODE_SLT, 4, 2 },
 100    { IR_POW, "IR_POW", OPCODE_POW, 1, 2 },
 101    /* unary ops */
 102    { IR_I_TO_F, "IR_I_TO_F", OPCODE_NOP, 1, 1 },
 103    { IR_F_TO_I, "IR_F_TO_I", OPCODE_INT, 4, 1 }, /* 4 floats to 4 ints */
 104    { IR_EXP, "IR_EXP", OPCODE_EXP, 1, 1 },
 105    { IR_EXP2, "IR_EXP2", OPCODE_EX2, 1, 1 },
 106    { IR_LOG2, "IR_LOG2", OPCODE_LG2, 1, 1 },
 107    { IR_RSQ, "IR_RSQ", OPCODE_RSQ, 1, 1 },
 108    { IR_RCP, "IR_RCP", OPCODE_RCP, 1, 1 },
 109    { IR_FLOOR, "IR_FLOOR", OPCODE_FLR, 4, 1 },
 110    { IR_FRAC, "IR_FRAC", OPCODE_FRC, 4, 1 },
 111    { IR_ABS, "IR_ABS", OPCODE_ABS, 4, 1 },
 112    { IR_NEG, "IR_NEG", OPCODE_NOP, 4, 1 }, /* special case: emit_negation() */
 113    { IR_DDX, "IR_DDX", OPCODE_DDX, 4, 1 },
 114    { IR_DDX, "IR_DDY", OPCODE_DDX, 4, 1 },
 115    { IR_SIN, "IR_SIN", OPCODE_SIN, 1, 1 },
 116    { IR_COS, "IR_COS", OPCODE_COS, 1, 1 },
 117    { IR_NOISE1, "IR_NOISE1", OPCODE_NOISE1, 1, 1 },
 118    { IR_NOISE2, "IR_NOISE2", OPCODE_NOISE2, 1, 1 },
 119    { IR_NOISE3, "IR_NOISE3", OPCODE_NOISE3, 1, 1 },
 120    { IR_NOISE4, "IR_NOISE4", OPCODE_NOISE4, 1, 1 },
 121
 122    /* other */
 123    { IR_SEQ, "IR_SEQ", OPCODE_NOP, 0, 0 },
 124    { IR_SCOPE, "IR_SCOPE", OPCODE_NOP, 0, 0 },
 125    { IR_LABEL, "IR_LABEL", OPCODE_NOP, 0, 0 },
 126    { IR_JUMP, "IR_JUMP", OPCODE_NOP, 0, 0 },
 127    { IR_IF, "IR_IF", OPCODE_NOP, 0, 0 },
 128    { IR_KILL, "IR_KILL", OPCODE_NOP, 0, 0 },
 129    { IR_COND, "IR_COND", OPCODE_NOP, 0, 0 },
 130    { IR_CALL, "IR_CALL", OPCODE_NOP, 0, 0 },
 131    { IR_MOVE, "IR_MOVE", OPCODE_NOP, 0, 1 },
 132    { IR_NOT, "IR_NOT", OPCODE_NOP, 1, 1 },
 133    { IR_VAR, "IR_VAR", OPCODE_NOP, 0, 0 },
 134    { IR_VAR_DECL, "IR_VAR_DECL", OPCODE_NOP, 0, 0 },
 135    { IR_TEX, "IR_TEX", OPCODE_TEX, 4, 1 },
 136    { IR_TEXB, "IR_TEXB", OPCODE_TXB, 4, 1 },
 137    { IR_TEXP, "IR_TEXP", OPCODE_TXP, 4, 1 },
 138    { IR_FLOAT, "IR_FLOAT", OPCODE_NOP, 0, 0 }, /* float literal */
 139    { IR_FIELD, "IR_FIELD", OPCODE_NOP, 0, 0 },
 140    { IR_ELEMENT, "IR_ELEMENT", OPCODE_NOP, 0, 0 },
 141    { IR_SWIZZLE, "IR_SWIZZLE", OPCODE_NOP, 0, 0 },
 142    { IR_NOP, NULL, OPCODE_NOP, 0, 0 }
 143 };
 144
 145
 146 static const slang_ir_info *
 147 slang_find_ir_info(slang_ir_opcode opcode)
 148 {
 149    GLuint i;
 150    for (i = 0; IrInfo[i].IrName; i++) {
 151       if (IrInfo[i].IrOpcode == opcode) {
 152          return IrInfo + i;
 153       }
 154    }
 155    return NULL;
 156 }
 157
 158 static const char *
 159 slang_ir_name(slang_ir_opcode opcode)
 160 {
 161    return slang_find_ir_info(opcode)->IrName;
 162 }
 163
 164
 165 /**
 166  * Swizzle a swizzle.  That is, return swz2(swz1)
 167  */
 168 static GLuint
 169 swizzle_swizzle(GLuint swz1, GLuint swz2)
 170 {
 171    GLuint i, swz, s[4];
 172    for (i = 0; i < 4; i++) {
 173       GLuint c = GET_SWZ(swz2, i);
 174       s[i] = GET_SWZ(swz1, c);
 175    }
 176    swz = MAKE_SWIZZLE4(s[0], s[1], s[2], s[3]);
 177    return swz;
 178 }
 179
 180
 181 slang_ir_storage *
 182 _slang_new_ir_storage(enum register_file file, GLint index, GLint size)
 183 {
 184    slang_ir_storage *st;
 185    st = (slang_ir_storage *) _mesa_calloc(sizeof(slang_ir_storage));
 186    if (st) {
 187       st->File = file;
 188       st->Index = index;
 189       st->Size = size;
 190       st->Swizzle = SWIZZLE_NOOP;
 191    }
 192    return st;
 193 }
 194
 195
 196 static const char *
 197 swizzle_string(GLuint swizzle)
 198 {
 199    static char s[6];
 200    GLuint i;
 201    s[0] = '.';
 202    for (i = 1; i < 5; i++) {
 203       s[i] = "xyzw"[GET_SWZ(swizzle, i-1)];
 204    }
 205    s[i] = 0;
 206    return s;
 207 }
 208
 209 static const char *
 210 writemask_string(GLuint writemask)
 211 {
 212    static char s[6];
 213    GLuint i, j = 0;
 214    s[j++] = '.';
 215    for (i = 0; i < 4; i++) {
 216       if (writemask & (1 << i))
 217          s[j++] = "xyzw"[i];
 218    }
 219    s[j] = 0;
 220    return s;
 221 }
 222
 223 static const char *
 224 storage_string(const slang_ir_storage *st)
 225 {
 226    static const char *files[] = {
 227       "TEMP",
 228       "LOCAL_PARAM",
 229       "ENV_PARAM",
 230       "STATE",
 231       "INPUT",
 232       "OUTPUT",
 233       "NAMED_PARAM",
 234       "CONSTANT",
 235       "UNIFORM",
 236       "WRITE_ONLY",
 237       "ADDRESS",
 238       "SAMPLER",
 239       "UNDEFINED"
 240    };
 241    static char s[100];
 242 #if 0
 243    if (st->Size == 1)
 244       sprintf(s, "%s[%d]", files[st->File], st->Index);
 245    else
 246       sprintf(s, "%s[%d..%d]", files[st->File], st->Index,
 247               st->Index + st->Size - 1);
 248 #endif
 249    assert(st->File < (GLint) (sizeof(files) / sizeof(files[0])));
 250    sprintf(s, "%s[%d]", files[st->File], st->Index);
 251    return s;
 252 }
 253
 254
 255 static void
 256 spaces(int n)
 257 {
 258    while (n-- > 0) {
 259       printf(" ");
 260    }
 261 }
 262
 263 #define IND 0
 264 void
 265 slang_print_ir(const slang_ir_node *n, int indent)
 266 {
 267    if (!n)
 268       return;
 269 #if !IND
 270    if (n->Opcode != IR_SEQ)
 271 #else
 272       printf("%3d:", indent);
 273 #endif
 274       spaces(indent);
 275
 276    switch (n->Opcode) {
 277    case IR_SEQ:
 278 #if IND
 279       printf("SEQ  at %p\n", (void*) n);
 280 #endif
 281       assert(n->Children[0]);
 282       assert(n->Children[1]);
 283       slang_print_ir(n->Children[0], indent + IND);
 284       slang_print_ir(n->Children[1], indent + IND);
 285       break;
 286    case IR_SCOPE:
 287       printf("NEW SCOPE\n");
 288       assert(!n->Children[1]);
 289       slang_print_ir(n->Children[0], indent + 3);
 290       break;
 291    case IR_MOVE:
 292       printf("MOVE (writemask = %s)\n", writemask_string(n->Writemask));
 293       slang_print_ir(n->Children[0], indent+3);
 294       slang_print_ir(n->Children[1], indent+3);
 295       break;
 296    case IR_LABEL:
 297       printf("LABEL: %s\n", n->Label->Name);
 298       break;
 299    case IR_COND:
 300       printf("COND\n");
 301       slang_print_ir(n->Children[0], indent + 3);
 302       break;
 303    case IR_JUMP:
 304       printf("JUMP %s\n", n->Label->Name);
 305       break;
 306
 307    case IR_IF:
 308       printf("IF \n");
 309       slang_print_ir(n->Children[0], indent+3);
 310       spaces(indent);
 311       printf("THEN\n");
 312       slang_print_ir(n->Children[1], indent+3);
 313       if (n->Children[2]) {
 314          spaces(indent);
 315          printf("ELSE\n");
 316          slang_print_ir(n->Children[2], indent+3);
 317       }
 318       printf("ENDIF\n");
 319       break;
 320
 321    case IR_BEGIN_SUB:
 322       printf("BEGIN_SUB\n");
 323       break;
 324    case IR_END_SUB:
 325       printf("END_SUB\n");
 326       break;
 327    case IR_RETURN:
 328       printf("RETURN\n");
 329       break;
 330    case IR_CALL:
 331       printf("CALL\n");
 332       break;
 333
 334    case IR_LOOP:
 335       printf("LOOP\n");
 336       slang_print_ir(n->Children[0], indent+3);
 337       spaces(indent);
 338       printf("ENDLOOP\n");
 339       break;
 340    case IR_CONT:
 341       printf("CONT\n");
 342       break;
 343    case IR_BREAK:
 344       printf("BREAK\n");
 345       break;
 346    case IR_BREAK_IF_FALSE:
 347       printf("BREAK_IF_FALSE\n");
 348       slang_print_ir(n->Children[0], indent+3);
 349       break;
 350    case IR_BREAK_IF_TRUE:
 351       printf("BREAK_IF_TRUE\n");
 352       slang_print_ir(n->Children[0], indent+3);
 353       break;
 354    case IR_CONT_IF_FALSE:
 355       printf("CONT_IF_FALSE\n");
 356       slang_print_ir(n->Children[0], indent+3);
 357       break;
 358    case IR_CONT_IF_TRUE:
 359       printf("CONT_IF_TRUE\n");
 360       slang_print_ir(n->Children[0], indent+3);
 361       break;
 362
 363    case IR_VAR:
 364       printf("VAR %s%s at %s  store %p\n",
 365              (n->Var ? (char *) n->Var->a_name : "TEMP"),
 366              swizzle_string(n->Store->Swizzle),
 367              storage_string(n->Store), (void*) n->Store);
 368       break;
 369    case IR_VAR_DECL:
 370       printf("VAR_DECL %s (%p) at %s  store %p\n",
 371              (n->Var ? (char *) n->Var->a_name : "TEMP"),
 372              (void*) n->Var, storage_string(n->Store),
 373              (void*) n->Store);
 374       break;
 375    case IR_FIELD:
 376       printf("FIELD %s of\n", n->Field);
 377       slang_print_ir(n->Children[0], indent+3);
 378       break;
 379    case IR_FLOAT:
 380       printf("FLOAT %g %g %g %g\n",
 381              n->Value[0], n->Value[1], n->Value[2], n->Value[3]);
 382       break;
 383    case IR_I_TO_F:
 384       printf("INT_TO_FLOAT\n");
 385       slang_print_ir(n->Children[0], indent+3);
 386       break;
 387    case IR_F_TO_I:
 388       printf("FLOAT_TO_INT\n");
 389       slang_print_ir(n->Children[0], indent+3);
 390       break;
 391    case IR_SWIZZLE:
 392       printf("SWIZZLE %s of  (store %p) \n",
 393              swizzle_string(n->Store->Swizzle), (void*) n->Store);
 394       slang_print_ir(n->Children[0], indent + 3);
 395       break;
 396    default:
 397       printf("%s (%p, %p)  (store %p)\n", slang_ir_name(n->Opcode),
 398              (void*) n->Children[0], (void*) n->Children[1], (void*) n->Store);
 399       slang_print_ir(n->Children[0], indent+3);
 400       slang_print_ir(n->Children[1], indent+3);
 401    }
 402 }
 403
 404
 405 /**
 406  * Allocate temporary storage for an intermediate result (such as for
 407  * a multiply or add, etc.
 408  */
 409 static GLboolean
 410 alloc_temp_storage(slang_emit_info *emitInfo, slang_ir_node *n, GLint size)
 411 {
 412    assert(!n->Var);
 413    assert(!n->Store);
 414    assert(size > 0);
 415    n->Store = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, size);
 416    if (!_slang_alloc_temp(emitInfo->vt, n->Store)) {
 417       slang_info_log_error(emitInfo->log,
 418                            "Ran out of registers, too many temporaries");
 419       return GL_FALSE;
 420    }
 421    return GL_TRUE;
 422 }
 423
 424
 425 /**
 426  * Free temporary storage, if n->Store is, in fact, temp storage.
 427  * Otherwise, no-op.
 428  */
 429 static void
 430 free_temp_storage(slang_var_table *vt, slang_ir_node *n)
 431 {
 432    if (n->Store->File == PROGRAM_TEMPORARY && n->Store->Index >= 0) {
 433       if (_slang_is_temp(vt, n->Store)) {
 434          _slang_free_temp(vt, n->Store);
 435          n->Store->Index = -1;
 436          n->Store->Size = -1;
 437       }
 438    }
 439 }
 440
 441
 442 /**
 443  * Convert IR storage to an instruction dst register.
 444  */
 445 static void
 446 storage_to_dst_reg(struct prog_dst_register *dst, const slang_ir_storage *st,
 447                    GLuint writemask)
 448 {
 449    static const GLuint defaultWritemask[4] = {
 450       WRITEMASK_X,
 451       WRITEMASK_X | WRITEMASK_Y,
 452       WRITEMASK_X | WRITEMASK_Y | WRITEMASK_Z,
 453       WRITEMASK_X | WRITEMASK_Y | WRITEMASK_Z | WRITEMASK_W
 454    };
 455    assert(st->Index >= 0);
 456    dst->File = st->File;
 457    dst->Index = st->Index;
 458    assert(st->File != PROGRAM_UNDEFINED);
 459    assert(st->Size >= 1);
 460    assert(st->Size <= 4);
 461    if (st->Size == 1) {
 462       GLuint comp = GET_SWZ(st->Swizzle, 0);
 463       assert(comp < 4);
 464       assert(writemask & WRITEMASK_X);
 465       dst->WriteMask = WRITEMASK_X << comp;
 466    }
 467    else {
 468       dst->WriteMask = defaultWritemask[st->Size - 1] & writemask;
 469    }
 470 }
 471
 472
 473 /**
 474  * Convert IR storage to an instruction src register.
 475  */
 476 static void
 477 storage_to_src_reg(struct prog_src_register *src, const slang_ir_storage *st)
 478 {
 479    static const GLuint defaultSwizzle[4] = {
 480       MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_X, SWIZZLE_X, SWIZZLE_X),
 481       MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W),
 482       MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W),
 483       MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W)
 484    };
 485    assert(st->File >= 0);
 486    assert(st->File < PROGRAM_UNDEFINED);
 487    assert(st->Size >= 1);
 488    assert(st->Size <= 4);
 489    src->File = st->File;
 490    src->Index = st->Index;
 491    if (st->Swizzle != SWIZZLE_NOOP)
 492       src->Swizzle = st->Swizzle;
 493    else
 494       src->Swizzle = defaultSwizzle[st->Size - 1]; /*XXX really need this?*/
 495
 496    assert(GET_SWZ(src->Swizzle, 0) != SWIZZLE_NIL);
 497    assert(GET_SWZ(src->Swizzle, 1) != SWIZZLE_NIL);
 498    assert(GET_SWZ(src->Swizzle, 2) != SWIZZLE_NIL);
 499    assert(GET_SWZ(src->Swizzle, 3) != SWIZZLE_NIL);
 500 }
 501
 502
 503
 504 /**
 505  * Add new instruction at end of given program.
 506  * \param prog  the program to append instruction onto
 507  * \param opcode  opcode for the new instruction
 508  * \return pointer to the new instruction
 509  */
 510 static struct prog_instruction *
 511 new_instruction(slang_emit_info *emitInfo, gl_inst_opcode opcode)
 512 {
 513    struct gl_program *prog = emitInfo->prog;
 514    struct prog_instruction *inst;
 515    prog->Instructions = _mesa_realloc_instructions(prog->Instructions,
 516                                                    prog->NumInstructions,
 517                                                    prog->NumInstructions + 1);
 518    inst = prog->Instructions + prog->NumInstructions;
 519    prog->NumInstructions++;
 520    _mesa_init_instructions(inst, 1);
 521    inst->Opcode = opcode;
 522    inst->BranchTarget = -1; /* invalid */
 523    /*
 524    printf("New inst %d: %p %s\n", prog->NumInstructions-1,(void*)inst,
 525           _mesa_opcode_string(inst->Opcode));
 526    */
 527    return inst;
 528 }
 529
 530
 531 #if 0
 532 /**
 533  * Return pointer to last instruction in program.
 534  */
 535 static struct prog_instruction *
 536 prev_instruction(struct gl_program *prog)
 537 {
 538    if (prog->NumInstructions == 0)
 539       return NULL;
 540    else
 541       return prog->Instructions + prog->NumInstructions - 1;
 542 }
 543 #endif
 544
 545
 546 static struct prog_instruction *
 547 emit(slang_emit_info *emitInfo, slang_ir_node *n);
 548
 549
 550 /**
 551  * Return an annotation string for given node's storage.
 552  */
 553 static char *
 554 storage_annotation(const slang_ir_node *n, const struct gl_program *prog)
 555 {
 556 #if ANNOTATE
 557    const slang_ir_storage *st = n->Store;
 558    static char s[100] = "";
 559
 560    if (!st)
 561       return _mesa_strdup("");
 562
 563    switch (st->File) {
 564    case PROGRAM_CONSTANT:
 565       if (st->Index >= 0) {
 566          const GLfloat *val = prog->Parameters->ParameterValues[st->Index];
 567          if (st->Swizzle == SWIZZLE_NOOP)
 568             sprintf(s, "{%g, %g, %g, %g}", val[0], val[1], val[2], val[3]);
 569          else {
 570             sprintf(s, "%g", val[GET_SWZ(st->Swizzle, 0)]);
 571          }
 572       }
 573       break;
 574    case PROGRAM_TEMPORARY:
 575       if (n->Var)
 576          sprintf(s, "%s", (char *) n->Var->a_name);
 577       else
 578          sprintf(s, "t[%d]", st->Index);
 579       break;
 580    case PROGRAM_STATE_VAR:
 581    case PROGRAM_UNIFORM:
 582       sprintf(s, "%s", prog->Parameters->Parameters[st->Index].Name);
 583       break;
 584    case PROGRAM_VARYING:
 585       sprintf(s, "%s", prog->Varying->Parameters[st->Index].Name);
 586       break;
 587    case PROGRAM_INPUT:
 588       sprintf(s, "input[%d]", st->Index);
 589       break;
 590    case PROGRAM_OUTPUT:
 591       sprintf(s, "output[%d]", st->Index);
 592       break;
 593    default:
 594       s[0] = 0;
 595    }
 596    return _mesa_strdup(s);
 597 #else
 598    return NULL;
 599 #endif
 600 }
 601
 602
 603 /**
 604  * Return an annotation string for an instruction.
 605  */
 606 static char *
 607 instruction_annotation(gl_inst_opcode opcode, char *dstAnnot,
 608                        char *srcAnnot0, char *srcAnnot1, char *srcAnnot2)
 609 {
 610 #if ANNOTATE
 611    const char *operator;
 612    char *s;
 613    int len = 50;
 614
 615    if (dstAnnot)
 616       len += strlen(dstAnnot);
 617    else
 618       dstAnnot = _mesa_strdup("");
 619
 620    if (srcAnnot0)
 621       len += strlen(srcAnnot0);
 622    else
 623       srcAnnot0 = _mesa_strdup("");
 624
 625    if (srcAnnot1)
 626       len += strlen(srcAnnot1);
 627    else
 628       srcAnnot1 = _mesa_strdup("");
 629
 630    if (srcAnnot2)
 631       len += strlen(srcAnnot2);
 632    else
 633       srcAnnot2 = _mesa_strdup("");
 634
 635    switch (opcode) {
 636    case OPCODE_ADD:
 637       operator = "+";
 638       break;
 639    case OPCODE_SUB:
 640       operator = "-";
 641       break;
 642    case OPCODE_MUL:
 643       operator = "*";
 644       break;
 645    case OPCODE_DP3:
 646       operator = "DP3";
 647       break;
 648    case OPCODE_DP4:
 649       operator = "DP4";
 650       break;
 651    case OPCODE_XPD:
 652       operator = "XPD";
 653       break;
 654    case OPCODE_RSQ:
 655       operator = "RSQ";
 656       break;
 657    case OPCODE_SGT:
 658       operator = ">";
 659       break;
 660    default:
 661       operator = ",";
 662    }
 663
 664    s = (char *) malloc(len);
 665    sprintf(s, "%s = %s %s %s %s", dstAnnot,
 666            srcAnnot0, operator, srcAnnot1, srcAnnot2);
 667    assert(_mesa_strlen(s) < len);
 668
 669    free(dstAnnot);
 670    free(srcAnnot0);
 671    free(srcAnnot1);
 672    free(srcAnnot2);
 673
 674    return s;
 675 #else
 676    return NULL;
 677 #endif
 678 }
 679
 680
 681
 682 /**
 683  * Generate code for a simple arithmetic instruction.
 684  * Either 1, 2 or 3 operands.
 685  */
 686 static struct prog_instruction *
 687 emit_arith(slang_emit_info *emitInfo, slang_ir_node *n)
 688 {
 689    struct prog_instruction *inst;
 690    const slang_ir_info *info = slang_find_ir_info(n->Opcode);
 691    char *srcAnnot[3], *dstAnnot;
 692    GLuint i;
 693
 694    assert(info);
 695    assert(info->InstOpcode != OPCODE_NOP);
 696
 697    srcAnnot[0] = srcAnnot[1] = srcAnnot[2] = dstAnnot = NULL;
 698
 699 #if PEEPHOLE_OPTIMIZATIONS
 700    /* Look for MAD opportunity */
 701    if (info->NumParams == 2 &&
 702        n->Opcode == IR_ADD && n->Children[0]->Opcode == IR_MUL) {
 703       /* found pattern IR_ADD(IR_MUL(A, B), C) */
 704       emit(emitInfo, n->Children[0]->Children[0]);  /* A */
 705       emit(emitInfo, n->Children[0]->Children[1]);  /* B */
 706       emit(emitInfo, n->Children[1]);  /* C */
 707       /* generate MAD instruction */
 708       inst = new_instruction(emitInfo, OPCODE_MAD);
 709       /* operands: A, B, C: */
 710       storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Children[0]->Store);
 711       storage_to_src_reg(&inst->SrcReg[1], n->Children[0]->Children[1]->Store);
 712       storage_to_src_reg(&inst->SrcReg[2], n->Children[1]->Store);
 713       free_temp_storage(emitInfo->vt, n->Children[0]->Children[0]);
 714       free_temp_storage(emitInfo->vt, n->Children[0]->Children[1]);
 715       free_temp_storage(emitInfo->vt, n->Children[1]);
 716    }
 717    else if (info->NumParams == 2 &&
 718             n->Opcode == IR_ADD && n->Children[1]->Opcode == IR_MUL) {
 719       /* found pattern IR_ADD(A, IR_MUL(B, C)) */
 720       emit(emitInfo, n->Children[0]);  /* A */
 721       emit(emitInfo, n->Children[1]->Children[0]);  /* B */
 722       emit(emitInfo, n->Children[1]->Children[1]);  /* C */
 723       /* generate MAD instruction */
 724       inst = new_instruction(emitInfo, OPCODE_MAD);
 725       /* operands: B, C, A */
 726       storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Children[0]->Store);
 727       storage_to_src_reg(&inst->SrcReg[1], n->Children[1]->Children[1]->Store);
 728       storage_to_src_reg(&inst->SrcReg[2], n->Children[0]->Store);
 729       free_temp_storage(emitInfo->vt, n->Children[1]->Children[0]);
 730       free_temp_storage(emitInfo->vt, n->Children[1]->Children[1]);
 731       free_temp_storage(emitInfo->vt, n->Children[0]);
 732    }
 733    else
 734 #endif
 735    {
 736       /* normal case */
 737
 738       /* gen code for children */
 739       for (i = 0; i < info->NumParams; i++)
 740          emit(emitInfo, n->Children[i]);
 741
 742       /* gen this instruction and src registers */
 743       inst = new_instruction(emitInfo, info->InstOpcode);
 744       for (i = 0; i < info->NumParams; i++)
 745          storage_to_src_reg(&inst->SrcReg[i], n->Children[i]->Store);
 746
 747       /* annotation */
 748       for (i = 0; i < info->NumParams; i++)
 749          srcAnnot[i] = storage_annotation(n->Children[i], emitInfo->prog);
 750
 751       /* free temps */
 752       for (i = 0; i < info->NumParams; i++)
 753          free_temp_storage(emitInfo->vt, n->Children[i]);
 754    }
 755
 756    /* result storage */
 757    if (!n->Store) {
 758       if (!alloc_temp_storage(emitInfo, n, info->ResultSize))
 759          return NULL;
 760    }
 761    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 762
 763    dstAnnot = storage_annotation(n, emitInfo->prog);
 764
 765    inst->Comment = instruction_annotation(inst->Opcode, dstAnnot, srcAnnot[0],
 766                                           srcAnnot[1], srcAnnot[2]);
 767
 768    /*_mesa_print_instruction(inst);*/
 769    return inst;
 770 }
 771
 772
 773 /**
 774  * Emit code for == and != operators.  These could normally be handled
 775  * by emit_arith() except we need to be able to handle structure comparisons.
 776  */
 777 static struct prog_instruction *
 778 emit_compare(slang_emit_info *emitInfo, slang_ir_node *n)
 779 {
 780    struct prog_instruction *inst;
 781
 782    assert(n->Opcode == IR_SEQUAL || n->Opcode == IR_SNEQUAL);
 783
 784    /* gen code for children */
 785    emit(emitInfo, n->Children[0]);
 786    emit(emitInfo, n->Children[1]);
 787
 788    assert(n->Children[0]->Store->Size == n->Children[1]->Store->Size);
 789
 790    /* gen this instruction and src registers */
 791    inst = new_instruction(emitInfo,
 792                           (n->Opcode == IR_SEQUAL) ? OPCODE_SEQ : OPCODE_SNE);
 793    if (n->Children[0]->Store->Size > 4) {
 794       /* struct compare */
 795       _mesa_problem(NULL, "struct compare not implemented!");
 796       return NULL;
 797    }
 798    else {
 799       /* small/simple types */
 800       storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
 801       storage_to_src_reg(&inst->SrcReg[1], n->Children[1]->Store);
 802    }
 803
 804    /* free temps */
 805    free_temp_storage(emitInfo->vt, n->Children[0]);
 806    free_temp_storage(emitInfo->vt, n->Children[1]);
 807
 808    /* result storage */
 809    if (!n->Store) {
 810       if (!alloc_temp_storage(emitInfo, n, 1))  /* 1 bool */
 811          return NULL;
 812    }
 813    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 814
 815    return inst;
 816 }
 817
 818
 819
 820 /**
 821  * Generate code for an IR_CLAMP instruction.
 822  */
 823 static struct prog_instruction *
 824 emit_clamp(slang_emit_info *emitInfo, slang_ir_node *n)
 825 {
 826    struct prog_instruction *inst;
 827
 828    assert(n->Opcode == IR_CLAMP);
 829    /* ch[0] = value
 830     * ch[1] = min limit
 831     * ch[2] = max limit
 832     */
 833
 834    inst = emit(emitInfo, n->Children[0]);
 835
 836    /* If lower limit == 0.0 and upper limit == 1.0,
 837     *    set prev instruction's SaturateMode field to SATURATE_ZERO_ONE.
 838     * Else,
 839     *    emit OPCODE_MIN, OPCODE_MAX sequence.
 840     */
 841 #if 0
 842    /* XXX this isn't quite finished yet */
 843    if (n->Children[1]->Opcode == IR_FLOAT &&
 844        n->Children[1]->Value[0] == 0.0 &&
 845        n->Children[1]->Value[1] == 0.0 &&
 846        n->Children[1]->Value[2] == 0.0 &&
 847        n->Children[1]->Value[3] == 0.0 &&
 848        n->Children[2]->Opcode == IR_FLOAT &&
 849        n->Children[2]->Value[0] == 1.0 &&
 850        n->Children[2]->Value[1] == 1.0 &&
 851        n->Children[2]->Value[2] == 1.0 &&
 852        n->Children[2]->Value[3] == 1.0) {
 853       if (!inst) {
 854          inst = prev_instruction(prog);
 855       }
 856       if (inst && inst->Opcode != OPCODE_NOP) {
 857          /* and prev instruction's DstReg matches n->Children[0]->Store */
 858          inst->SaturateMode = SATURATE_ZERO_ONE;
 859          n->Store = n->Children[0]->Store;
 860          return inst;
 861       }
 862    }
 863 #endif
 864
 865    if (!n->Store)
 866       if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size))
 867          return NULL;
 868
 869    emit(emitInfo, n->Children[1]);
 870    emit(emitInfo, n->Children[2]);
 871
 872    /* tmp = max(ch[0], ch[1]) */
 873    inst = new_instruction(emitInfo, OPCODE_MAX);
 874    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 875    storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
 876    storage_to_src_reg(&inst->SrcReg[1], n->Children[1]->Store);
 877
 878    /* tmp = min(tmp, ch[2]) */
 879    inst = new_instruction(emitInfo, OPCODE_MIN);
 880    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 881    storage_to_src_reg(&inst->SrcReg[0], n->Store);
 882    storage_to_src_reg(&inst->SrcReg[1], n->Children[2]->Store);
 883
 884    return inst;
 885 }
 886
 887
 888 static struct prog_instruction *
 889 emit_negation(slang_emit_info *emitInfo, slang_ir_node *n)
 890 {
 891    /* Implement as MOV dst, -src; */
 892    /* XXX we could look at the previous instruction and in some circumstances
 893     * modify it to accomplish the negation.
 894     */
 895    struct prog_instruction *inst;
 896
 897    emit(emitInfo, n->Children[0]);
 898
 899    if (!n->Store)
 900       if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size))
 901          return NULL;
 902
 903    inst = new_instruction(emitInfo, OPCODE_MOV);
 904    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 905    storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
 906    inst->SrcReg[0].NegateBase = NEGATE_XYZW;
 907    return inst;
 908 }
 909
 910
 911 static struct prog_instruction *
 912 emit_label(slang_emit_info *emitInfo, const slang_ir_node *n)
 913 {
 914    assert(n->Label);
 915    assert(_slang_label_get_location(n->Label) < 0);
 916    _slang_label_set_location(n->Label, emitInfo->prog->NumInstructions,
 917                              emitInfo->prog);
 918    return NULL;
 919 }
 920
 921
 922 static struct prog_instruction *
 923 emit_jump(slang_emit_info *emitInfo, slang_ir_node *n)
 924 {
 925    struct prog_instruction *inst;
 926    assert(n);
 927    assert(n->Label);
 928    inst = new_instruction(emitInfo, OPCODE_BRA);
 929    inst->DstReg.CondMask = COND_TR;  /* always branch */
 930    inst->BranchTarget = _slang_label_get_location(n->Label);
 931    if (inst->BranchTarget < 0) {
 932       _slang_label_add_reference(n->Label, emitInfo->prog->NumInstructions - 1);
 933    }
 934    return inst;
 935 }
 936
 937
 938 static struct prog_instruction *
 939 emit_kill(slang_emit_info *emitInfo)
 940 {
 941    struct prog_instruction *inst;
 942    /* NV-KILL - discard fragment depending on condition code.
 943     * Note that ARB-KILL depends on sign of vector operand.
 944     */
 945    inst = new_instruction(emitInfo, OPCODE_KIL_NV);
 946    inst->DstReg.CondMask = COND_TR;  /* always branch */
 947    return inst;
 948 }
 949
 950
 951 static struct prog_instruction *
 952 emit_tex(slang_emit_info *emitInfo, slang_ir_node *n)
 953 {
 954    struct prog_instruction *inst;
 955
 956    (void) emit(emitInfo, n->Children[1]);
 957
 958    if (n->Opcode == IR_TEX) {
 959       inst = new_instruction(emitInfo, OPCODE_TEX);
 960    }
 961    else if (n->Opcode == IR_TEXB) {
 962       inst = new_instruction(emitInfo, OPCODE_TXB);
 963    }
 964    else {
 965       assert(n->Opcode == IR_TEXP);
 966       inst = new_instruction(emitInfo, OPCODE_TXP);
 967    }
 968
 969    if (!n->Store)
 970       if (!alloc_temp_storage(emitInfo, n, 4))
 971          return NULL;
 972
 973    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 974
 975    /* Child[1] is the coord */
 976    assert(n->Children[1]->Store->File != PROGRAM_UNDEFINED);
 977    assert(n->Children[1]->Store->Index >= 0);
 978    storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Store);
 979
 980    /* Child[0] is the sampler (a uniform which'll indicate the texture unit) */
 981    assert(n->Children[0]->Store);
 982    assert(n->Children[0]->Store->Size >= TEXTURE_1D_INDEX);
 983
 984    inst->Sampler = n->Children[0]->Store->Index; /* i.e. uniform's index */
 985    inst->TexSrcTarget = n->Children[0]->Store->Size;
 986    inst->TexSrcUnit = 27; /* Dummy value; the TexSrcUnit will be computed at
 987                            * link time, using the sampler uniform's value.
 988                            */
 989    return inst;
 990 }
 991
 992
 993 static struct prog_instruction *
 994 emit_move(slang_emit_info *emitInfo, slang_ir_node *n)
 995 {
 996    struct prog_instruction *inst;
 997
 998    /* lhs */
 999    emit(emitInfo, n->Children[0]);
1000
1001    /* rhs */
1002    assert(n->Children[1]);
1003    inst = emit(emitInfo, n->Children[1]);
1004
1005    assert(n->Children[1]->Store->Index >= 0);
1006
1007    assert(!n->Store);
1008    n->Store = n->Children[0]->Store;
1009
1010 #if PEEPHOLE_OPTIMIZATIONS
1011    if (inst && _slang_is_temp(emitInfo->vt, n->Children[1]->Store)) {
1012       /* Peephole optimization:
1013        * Just modify the RHS to put its result into the dest of this
1014        * MOVE operation.  Then, this MOVE is a no-op.
1015        */
1016       _slang_free_temp(emitInfo->vt, n->Children[1]->Store);
1017       *n->Children[1]->Store = *n->Children[0]->Store;
1018       /* fixup the prev (RHS) instruction */
1019       assert(n->Children[0]->Store->Index >= 0);
1020       storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store, n->Writemask);
1021       return inst;
1022    }
1023    else
1024 #endif
1025    {
1026       if (n->Children[0]->Store->Size > 4) {
1027          /* move matrix/struct etc (block of registers) */
1028          slang_ir_storage dstStore = *n->Children[0]->Store;
1029          slang_ir_storage srcStore = *n->Children[1]->Store;
1030          GLint size = srcStore.Size;
1031          ASSERT(n->Children[0]->Writemask == WRITEMASK_XYZW);
1032          ASSERT(n->Children[1]->Store->Swizzle == SWIZZLE_NOOP);
1033          dstStore.Size = 4;
1034          srcStore.Size = 4;
1035          while (size >= 4) {
1036             inst = new_instruction(emitInfo, OPCODE_MOV);
1037             inst->Comment = _mesa_strdup("IR_MOVE block");
1038             storage_to_dst_reg(&inst->DstReg, &dstStore, n->Writemask);
1039             storage_to_src_reg(&inst->SrcReg[0], &srcStore);
1040             srcStore.Index++;
1041             dstStore.Index++;
1042             size -= 4;
1043          }
1044       }
1045       else {
1046          /* single register move */
1047          char *srcAnnot, *dstAnnot;
1048          inst = new_instruction(emitInfo, OPCODE_MOV);
1049          assert(n->Children[0]->Store->Index >= 0);
1050          storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store, n->Writemask);
1051          storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Store);
1052          dstAnnot = storage_annotation(n->Children[0], emitInfo->prog);
1053          srcAnnot = storage_annotation(n->Children[1], emitInfo->prog);
1054          inst->Comment = instruction_annotation(inst->Opcode, dstAnnot,
1055                                                 srcAnnot, NULL, NULL);
1056       }
1057       free_temp_storage(emitInfo->vt, n->Children[1]);
1058       return inst;
1059    }
1060 }
1061
1062
1063 static struct prog_instruction *
1064 emit_cond(slang_emit_info *emitInfo, slang_ir_node *n)
1065 {
1066    /* Conditional expression (in if/while/for stmts).
1067     * Need to update condition code register.
1068     * Next instruction is typically an IR_IF.
1069     */
1070    struct prog_instruction *inst;
1071
1072    if (!n->Children[0])
1073       return NULL;
1074
1075    inst = emit(emitInfo, n->Children[0]);
1076    if (inst) {
1077       /* set inst's CondUpdate flag */
1078       inst->CondUpdate = GL_TRUE;
1079       return inst; /* XXX or null? */
1080    }
1081    else {
1082       /* This'll happen for things like "if (i) ..." where no code
1083        * is normally generated for the expression "i".
1084        * Generate a move instruction just to set condition codes.
1085        * Note: must use full 4-component vector since all four
1086        * condition codes must be set identically.
1087        */
1088       if (!alloc_temp_storage(emitInfo, n, 4))
1089          return NULL;
1090       inst = new_instruction(emitInfo, OPCODE_MOV);
1091       inst->CondUpdate = GL_TRUE;
1092       storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
1093       storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
1094       _slang_free_temp(emitInfo->vt, n->Store);
1095       inst->Comment = _mesa_strdup("COND expr");
1096       return inst; /* XXX or null? */
1097    }
1098 }
1099
1100
1101 /**
1102  * Logical-NOT
1103  */
1104 static struct prog_instruction *
1105 emit_not(slang_emit_info *emitInfo, slang_ir_node *n)
1106 {
1107    GLfloat zero = 0.0;
1108    slang_ir_storage st;
1109    struct prog_instruction *inst;
1110
1111    /* need zero constant */
1112    st.File = PROGRAM_CONSTANT;
1113    st.Size = 1;
1114    st.Index = _mesa_add_unnamed_constant(emitInfo->prog->Parameters, &zero,
1115                                          1, &st.Swizzle);
1116
1117    /* child expr */
1118    (void) emit(emitInfo, n->Children[0]);
1119    /* XXXX if child instr is SGT convert to SLE, if SEQ, SNE, etc */
1120
1121    if (!n->Store)
1122       if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size))
1123          return NULL;
1124
1125    inst = new_instruction(emitInfo, OPCODE_SEQ);
1126    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
1127    storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
1128    storage_to_src_reg(&inst->SrcReg[1], &st);
1129
1130    free_temp_storage(emitInfo->vt, n->Children[0]);
1131
1132    inst->Comment = _mesa_strdup("NOT");
1133    return inst;
1134 }
1135
1136
1137 static struct prog_instruction *
1138 emit_if(slang_emit_info *emitInfo, slang_ir_node *n)
1139 {
1140    struct gl_program *prog = emitInfo->prog;
1141    struct prog_instruction *ifInst;
1142    GLuint ifInstLoc, elseInstLoc = 0;
1143
1144    emit(emitInfo, n->Children[0]);  /* the condition */
1145    ifInstLoc = prog->NumInstructions;
1146    if (emitInfo->EmitHighLevelInstructions) {
1147       ifInst = new_instruction(emitInfo, OPCODE_IF);
1148       ifInst->DstReg.CondMask = COND_NE;  /* if cond is non-zero */
1149       ifInst->DstReg.CondSwizzle = SWIZZLE_X;
1150    }
1151    else {
1152       /* conditional jump to else, or endif */
1153       ifInst = new_instruction(emitInfo, OPCODE_BRA);
1154       ifInst->DstReg.CondMask = COND_EQ;  /* BRA if cond is zero */
1155       ifInst->DstReg.CondSwizzle = SWIZZLE_X;
1156       ifInst->Comment = _mesa_strdup("if zero");
1157    }
1158
1159    /* if body */
1160    emit(emitInfo, n->Children[1]);
1161
1162    if (n->Children[2]) {
1163       /* have else body */
1164       elseInstLoc = prog->NumInstructions;
1165       if (emitInfo->EmitHighLevelInstructions) {
1166          (void) new_instruction(emitInfo, OPCODE_ELSE);
1167       }
1168       else {
1169          /* jump to endif instruction */
1170          struct prog_instruction *inst;
1171          inst = new_instruction(emitInfo, OPCODE_BRA);
1172          inst->Comment = _mesa_strdup("else");
1173          inst->DstReg.CondMask = COND_TR;  /* always branch */
1174       }
1175       ifInst = prog->Instructions + ifInstLoc;
1176       ifInst->BranchTarget = prog->NumInstructions;
1177
1178       emit(emitInfo, n->Children[2]);
1179    }
1180    else {
1181       /* no else body */
1182       ifInst = prog->Instructions + ifInstLoc;
1183       ifInst->BranchTarget = prog->NumInstructions + 1;
1184    }
1185
1186    if (emitInfo->EmitHighLevelInstructions) {
1187       (void) new_instruction(emitInfo, OPCODE_ENDIF);
1188    }
1189
1190    if (n->Children[2]) {
1191       struct prog_instruction *elseInst;
1192       elseInst = prog->Instructions + elseInstLoc;
1193       elseInst->BranchTarget = prog->NumInstructions;
1194    }
1195    return NULL;
1196 }
1197
1198
1199 static struct prog_instruction *
1200 emit_loop(slang_emit_info *emitInfo, slang_ir_node *n)
1201 {
1202    struct gl_program *prog = emitInfo->prog;
1203    struct prog_instruction *beginInst, *endInst;
1204    GLuint beginInstLoc, endInstLoc;
1205    slang_ir_node *ir;
1206
1207    /* emit OPCODE_BGNLOOP */
1208    beginInstLoc = prog->NumInstructions;
1209    if (emitInfo->EmitHighLevelInstructions) {
1210       (void) new_instruction(emitInfo, OPCODE_BGNLOOP);
1211    }
1212
1213    /* body */
1214    emit(emitInfo, n->Children[0]);
1215
1216    endInstLoc = prog->NumInstructions;
1217    if (emitInfo->EmitHighLevelInstructions) {
1218       /* emit OPCODE_ENDLOOP */
1219       endInst = new_instruction(emitInfo, OPCODE_ENDLOOP);
1220    }
1221    else {
1222       /* emit unconditional BRA-nch */
1223       endInst = new_instruction(emitInfo, OPCODE_BRA);
1224       endInst->DstReg.CondMask = COND_TR;  /* always true */
1225    }
1226    /* end instruction's BranchTarget points to top of loop */
1227    endInst->BranchTarget = beginInstLoc;
1228
1229    if (emitInfo->EmitHighLevelInstructions) {
1230       /* BGNLOOP's BranchTarget points to the ENDLOOP inst */
1231       beginInst = prog->Instructions + beginInstLoc;
1232       beginInst->BranchTarget = prog->NumInstructions - 1;
1233    }
1234
1235    /* Done emitting loop code.  Now walk over the loop's linked list of
1236     * BREAK and CONT nodes, filling in their BranchTarget fields (which
1237     * will point to the ENDLOOP+1 or BGNLOOP instructions, respectively).
1238     */
1239    for (ir = n->BranchNode; ir; ir = ir->BranchNode) {
1240       struct prog_instruction *inst = prog->Instructions + ir->InstLocation;
1241       assert(inst->BranchTarget < 0);
1242       if (ir->Opcode == IR_BREAK ||
1243           ir->Opcode == IR_BREAK_IF_FALSE ||
1244           ir->Opcode == IR_BREAK_IF_TRUE) {
1245          assert(inst->Opcode == OPCODE_BRK ||
1246                 inst->Opcode == OPCODE_BRA);
1247          /* go to instruction after end of loop */
1248          inst->BranchTarget = endInstLoc + 1;
1249       }
1250       else {
1251          assert(ir->Opcode == IR_CONT ||
1252                 ir->Opcode == IR_CONT_IF_FALSE ||
1253                 ir->Opcode == IR_CONT_IF_TRUE);
1254          assert(inst->Opcode == OPCODE_CONT ||
1255                 inst->Opcode == OPCODE_BRA);
1256          /* to go instruction at top of loop */
1257          inst->BranchTarget = beginInstLoc;
1258       }
1259    }
1260    return NULL;
1261 }
1262
1263
1264 /**
1265  * "Continue" or "break" statement.
1266  * Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted.
1267  */
1268 static struct prog_instruction *
1269 emit_cont_break(slang_emit_info *emitInfo, slang_ir_node *n)
1270 {
1271    gl_inst_opcode opcode;
1272    struct prog_instruction *inst;
1273    n->InstLocation = emitInfo->prog->NumInstructions;
1274    if (emitInfo->EmitHighLevelInstructions) {
1275       opcode = (n->Opcode == IR_CONT) ? OPCODE_CONT : OPCODE_BRK;
1276    }
1277    else {
1278       opcode = OPCODE_BRA;
1279    }
1280    inst = new_instruction(emitInfo, opcode);
1281    inst->DstReg.CondMask = COND_TR;  /* always true */
1282    return inst;
1283 }
1284
1285
1286 /**
1287  * Conditional "continue" or "break" statement.
1288  * Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted.
1289  */
1290 static struct prog_instruction *
1291 emit_cont_break_if(slang_emit_info *emitInfo, slang_ir_node *n,
1292                    GLboolean breakTrue)
1293 {
1294    gl_inst_opcode opcode;
1295    struct prog_instruction *inst;
1296
1297    /* evaluate condition expr, setting cond codes */
1298    inst = emit(emitInfo, n->Children[0]);
1299    assert(inst);
1300    inst->CondUpdate = GL_TRUE;
1301
1302    n->InstLocation = emitInfo->prog->NumInstructions;
1303    if (emitInfo->EmitHighLevelInstructions) {
1304       if (n->Opcode == IR_CONT_IF_TRUE ||
1305           n->Opcode == IR_CONT_IF_FALSE)
1306          opcode = OPCODE_CONT;
1307       else
1308          opcode = OPCODE_BRK;
1309    }
1310    else {
1311       opcode = OPCODE_BRA;
1312    }
1313    inst = new_instruction(emitInfo, opcode);
1314    inst->DstReg.CondMask = breakTrue ? COND_NE : COND_EQ;
1315    return inst;
1316 }
1317
1318
1319
1320 /**
1321  * Remove any SWIZZLE_NIL terms from given swizzle mask (smear prev term).
1322  * Ex: fix_swizzle("zyNN") -> "zyyy"
1323  */
1324 static GLuint
1325 fix_swizzle(GLuint swizzle)
1326 {
1327    GLuint swz[4], i;
1328    for (i = 0; i < 4; i++) {
1329       swz[i] = GET_SWZ(swizzle, i);
1330       if (swz[i] == SWIZZLE_NIL) {
1331          swz[i] = swz[i - 1];
1332       }
1333    }
1334    return MAKE_SWIZZLE4(swz[0], swz[1], swz[2], swz[3]);
1335 }
1336
1337
1338 static struct prog_instruction *
1339 emit_swizzle(slang_emit_info *emitInfo, slang_ir_node *n)
1340 {
1341    GLuint swizzle;
1342
1343    /* swizzled storage access */
1344    (void) emit(emitInfo, n->Children[0]);
1345
1346    /* "pull-up" the child's storage info, applying our swizzle info */
1347    n->Store->File  = n->Children[0]->Store->File;
1348    n->Store->Index = n->Children[0]->Store->Index;
1349    n->Store->Size  = n->Children[0]->Store->Size;
1350    /*n->Var = n->Children[0]->Var; XXX for debug */
1351    assert(n->Store->Index >= 0);
1352
1353    swizzle = fix_swizzle(n->Store->Swizzle);
1354 #ifdef DEBUG
1355    {
1356       GLuint s = n->Children[0]->Store->Swizzle;
1357       assert(GET_SWZ(s, 0) != SWIZZLE_NIL);
1358       assert(GET_SWZ(s, 1) != SWIZZLE_NIL);
1359       assert(GET_SWZ(s, 2) != SWIZZLE_NIL);
1360       assert(GET_SWZ(s, 3) != SWIZZLE_NIL);
1361    }
1362 #endif
1363
1364    /* apply this swizzle to child's swizzle to get composed swizzle */
1365    n->Store->Swizzle = swizzle_swizzle(n->Children[0]->Store->Swizzle,
1366                                        swizzle);
1367    return NULL;
1368 }
1369
1370
1371 /**
1372  * Dereference array element.  Just resolve storage for the array
1373  * element represented by this node.
1374  */
1375 static struct prog_instruction *
1376 emit_array_element(slang_emit_info *emitInfo, slang_ir_node *n)
1377 {
1378    assert(n->Store);
1379    assert(n->Store->File != PROGRAM_UNDEFINED);
1380    assert(n->Store->Size > 0);
1381
1382    if (n->Store->File == PROGRAM_STATE_VAR) {
1383       n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
1384       return NULL;
1385    }
1386
1387    if (n->Children[1]->Opcode == IR_FLOAT) {
1388       /* Constant index */
1389       const GLint arrayAddr = n->Children[0]->Store->Index;
1390       const GLint index = (GLint) n->Children[1]->Value[0];
1391       n->Store->Index = arrayAddr + index;
1392    }
1393    else {
1394       /* Variable index - PROBLEM */
1395       const GLint arrayAddr = n->Children[0]->Store->Index;
1396       const GLint index = 0;
1397       _mesa_problem(NULL, "variable array indexes not supported yet!");
1398       n->Store->Index = arrayAddr + index;
1399    }
1400    return NULL; /* no instruction */
1401 }
1402
1403
1404 /**
1405  * Resolve storage for accessing a structure field.
1406  */
1407 static struct prog_instruction *
1408 emit_struct_field(slang_emit_info *emitInfo, slang_ir_node *n)
1409 {
1410    if (n->Store->File == PROGRAM_STATE_VAR) {
1411       n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
1412    }
1413    else {
1414       GLint offset = n->FieldOffset / 4;
1415       assert(n->Children[0]->Store->Index >= 0);
1416       n->Store->Index = n->Children[0]->Store->Index + offset;
1417       if (n->Store->Size == 1) {
1418          GLint swz = n->FieldOffset % 4;
1419          n->Store->Swizzle = MAKE_SWIZZLE4(swz, swz, swz, swz);
1420       }
1421       else {
1422          n->Store->Swizzle = SWIZZLE_XYZW;
1423       }
1424    }
1425    return NULL; /* no instruction */
1426 }
1427
1428
1429 static struct prog_instruction *
1430 emit(slang_emit_info *emitInfo, slang_ir_node *n)
1431 {
1432    struct prog_instruction *inst;
1433    if (!n)
1434       return NULL;
1435
1436    switch (n->Opcode) {
1437    case IR_SEQ:
1438       /* sequence of two sub-trees */
1439       assert(n->Children[0]);
1440       assert(n->Children[1]);
1441       emit(emitInfo, n->Children[0]);
1442       inst = emit(emitInfo, n->Children[1]);
1443       assert(!n->Store);
1444       n->Store = n->Children[1]->Store;
1445       return inst;
1446
1447    case IR_SCOPE:
1448       /* new variable scope */
1449       _slang_push_var_table(emitInfo->vt);
1450       inst = emit(emitInfo, n->Children[0]);
1451       _slang_pop_var_table(emitInfo->vt);
1452       return inst;
1453
1454    case IR_VAR_DECL:
1455       /* Variable declaration - allocate a register for it */
1456       assert(n->Store);
1457       assert(n->Store->File != PROGRAM_UNDEFINED);
1458       assert(n->Store->Size > 0);
1459       assert(n->Store->Index < 0);
1460       if (!n->Var || n->Var->isTemp) {
1461          /* a nameless/temporary variable, will be freed after first use */
1462          if (!_slang_alloc_temp(emitInfo->vt, n->Store)) {
1463             slang_info_log_error(emitInfo->log,
1464                                  "Ran out of registers, too many temporaries");
1465             return NULL;
1466          }
1467       }
1468       else {
1469          /* a regular variable */
1470          _slang_add_variable(emitInfo->vt, n->Var);
1471          if (!_slang_alloc_var(emitInfo->vt, n->Store)) {
1472             slang_info_log_error(emitInfo->log,
1473                                  "Ran out of registers, too many variables");
1474             return NULL;
1475          }
1476          /*
1477          printf("IR_VAR_DECL %s %d store %p\n",
1478                 (char*) n->Var->a_name, n->Store->Index, (void*) n->Store);
1479          */
1480          assert(n->Var->aux == n->Store);
1481       }
1482       if (emitInfo->EmitComments) {
1483          /* emit NOP with comment describing the variable's storage location */
1484          char s[1000];
1485          sprintf(s, "TEMP[%d]%s = variable %s (size %d)",
1486                  n->Store->Index,
1487                  _mesa_swizzle_string(n->Store->Swizzle, 0, GL_FALSE),
1488                  (n->Var ? (char *) n->Var->a_name : "anonymous"),
1489                  n->Store->Size);
1490          inst = new_instruction(emitInfo, OPCODE_NOP);
1491          inst->Comment = _mesa_strdup(s);
1492          return inst;
1493       }
1494       return NULL;
1495
1496    case IR_VAR:
1497       /* Reference to a variable
1498        * Storage should have already been resolved/allocated.
1499        */
1500       assert(n->Store);
1501       assert(n->Store->File != PROGRAM_UNDEFINED);
1502
1503       if (n->Store->File == PROGRAM_STATE_VAR &&
1504           n->Store->Index < 0) {
1505          n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
1506       }
1507
1508       if (n->Store->Index < 0) {
1509          printf("#### VAR %s not allocated!\n", (char*)n->Var->a_name);
1510       }
1511       assert(n->Store->Index >= 0);
1512       assert(n->Store->Size > 0);
1513       break;
1514
1515    case IR_ELEMENT:
1516       return emit_array_element(emitInfo, n);
1517    case IR_FIELD:
1518       return emit_struct_field(emitInfo, n);
1519    case IR_SWIZZLE:
1520       return emit_swizzle(emitInfo, n);
1521
1522    case IR_I_TO_F:
1523       /* just move */
1524       emit(emitInfo, n->Children[0]);
1525       inst = new_instruction(emitInfo, OPCODE_MOV);
1526       if (!n->Store) {
1527          if (!alloc_temp_storage(emitInfo, n, 1))
1528             return NULL;
1529       }
1530       storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
1531       storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
1532       if (emitInfo->EmitComments)
1533          inst->Comment = _mesa_strdup("int to float");
1534       return NULL;
1535
1536    /* Simple arithmetic */
1537    /* unary */
1538    case IR_RSQ:
1539    case IR_RCP:
1540    case IR_FLOOR:
1541    case IR_FRAC:
1542    case IR_F_TO_I:
1543    case IR_ABS:
1544    case IR_SIN:
1545    case IR_COS:
1546    case IR_DDX:
1547    case IR_DDY:
1548    case IR_NOISE1:
1549    case IR_NOISE2:
1550    case IR_NOISE3:
1551    case IR_NOISE4:
1552    /* binary */
1553    case IR_ADD:
1554    case IR_SUB:
1555    case IR_MUL:
1556    case IR_DOT4:
1557    case IR_DOT3:
1558    case IR_CROSS:
1559    case IR_MIN:
1560    case IR_MAX:
1561    case IR_SGE:
1562    case IR_SGT:
1563    case IR_SLE:
1564    case IR_SLT:
1565    case IR_POW:
1566    case IR_EXP:
1567    case IR_EXP2:
1568    /* trinary operators */
1569    case IR_LRP:
1570       return emit_arith(emitInfo, n);
1571
1572    case IR_SEQUAL:
1573    case IR_SNEQUAL:
1574       return emit_compare(emitInfo, n);
1575
1576    case IR_CLAMP:
1577       return emit_clamp(emitInfo, n);
1578    case IR_TEX:
1579    case IR_TEXB:
1580    case IR_TEXP:
1581       return emit_tex(emitInfo, n);
1582    case IR_NEG:
1583       return emit_negation(emitInfo, n);
1584    case IR_FLOAT:
1585       /* find storage location for this float constant */
1586       n->Store->Index = _mesa_add_unnamed_constant(emitInfo->prog->Parameters, n->Value,
1587                                                    n->Store->Size,
1588                                                    &n->Store->Swizzle);
1589       if (n->Store->Index < 0) {
1590          slang_info_log_error(emitInfo->log, "Ran out of space for constants");
1591          return NULL;
1592       }
1593       return NULL;
1594
1595    case IR_MOVE:
1596       return emit_move(emitInfo, n);
1597
1598    case IR_COND:
1599       return emit_cond(emitInfo, n);
1600
1601    case IR_NOT:
1602       return emit_not(emitInfo, n);
1603
1604    case IR_LABEL:
1605       return emit_label(emitInfo, n);
1606    case IR_JUMP:
1607       assert(n);
1608       assert(n->Label);
1609       return emit_jump(emitInfo, n);
1610    case IR_KILL:
1611       return emit_kill(emitInfo);
1612
1613    case IR_IF:
1614       return emit_if(emitInfo, n);
1615
1616    case IR_LOOP:
1617       return emit_loop(emitInfo, n);
1618    case IR_BREAK_IF_FALSE:
1619    case IR_CONT_IF_FALSE:
1620       return emit_cont_break_if(emitInfo, n, GL_FALSE);
1621    case IR_BREAK_IF_TRUE:
1622    case IR_CONT_IF_TRUE:
1623       return emit_cont_break_if(emitInfo, n, GL_TRUE);
1624    case IR_BREAK:
1625       /* fall-through */
1626    case IR_CONT:
1627       return emit_cont_break(emitInfo, n);
1628
1629    case IR_BEGIN_SUB:
1630       return new_instruction(emitInfo, OPCODE_BGNSUB);
1631    case IR_END_SUB:
1632       return new_instruction(emitInfo, OPCODE_ENDSUB);
1633    case IR_RETURN:
1634       return new_instruction(emitInfo, OPCODE_RET);
1635
1636    case IR_NOP:
1637       return NULL;
1638
1639    default:
1640       _mesa_problem(NULL, "Unexpected IR opcode in emit()\n");
1641       abort();
1642    }
1643    return NULL;
1644 }
1645
1646
1647 GLboolean
1648 _slang_emit_code(slang_ir_node *n, slang_var_table *vt,
1649                  struct gl_program *prog, GLboolean withEnd,
1650                  slang_info_log *log)
1651 {
1652    GET_CURRENT_CONTEXT(ctx);
1653    GLboolean success;
1654    slang_emit_info emitInfo;
1655
1656    emitInfo.log = log;
1657    emitInfo.vt = vt;
1658    emitInfo.prog = prog;
1659
1660    emitInfo.EmitHighLevelInstructions = ctx->Shader.EmitHighLevelInstructions;
1661    emitInfo.EmitComments = 1+ctx->Shader.EmitComments;
1662
1663    (void) emit(&emitInfo, n);
1664
1665    /* finish up by adding the END opcode to program */
1666    if (withEnd) {
1667       struct prog_instruction *inst;
1668       inst = new_instruction(&emitInfo, OPCODE_END);
1669    }
1670    success = GL_TRUE;
1671
1672 #if 0
1673    printf("*********** End emit code (%u inst):\n", prog->NumInstructions);
1674    _mesa_print_program(prog);
1675    _mesa_print_program_parameters(ctx,prog);
1676 #endif
1677
1678    return success;
1679 }