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  * Generate code for an IR_CLAMP instruction.
 775  */
 776 static struct prog_instruction *
 777 emit_clamp(slang_emit_info *emitInfo, slang_ir_node *n)
 778 {
 779    struct prog_instruction *inst;
 780
 781    assert(n->Opcode == IR_CLAMP);
 782    /* ch[0] = value
 783     * ch[1] = min limit
 784     * ch[2] = max limit
 785     */
 786
 787    inst = emit(emitInfo, n->Children[0]);
 788
 789    /* If lower limit == 0.0 and upper limit == 1.0,
 790     *    set prev instruction's SaturateMode field to SATURATE_ZERO_ONE.
 791     * Else,
 792     *    emit OPCODE_MIN, OPCODE_MAX sequence.
 793     */
 794 #if 0
 795    /* XXX this isn't quite finished yet */
 796    if (n->Children[1]->Opcode == IR_FLOAT &&
 797        n->Children[1]->Value[0] == 0.0 &&
 798        n->Children[1]->Value[1] == 0.0 &&
 799        n->Children[1]->Value[2] == 0.0 &&
 800        n->Children[1]->Value[3] == 0.0 &&
 801        n->Children[2]->Opcode == IR_FLOAT &&
 802        n->Children[2]->Value[0] == 1.0 &&
 803        n->Children[2]->Value[1] == 1.0 &&
 804        n->Children[2]->Value[2] == 1.0 &&
 805        n->Children[2]->Value[3] == 1.0) {
 806       if (!inst) {
 807          inst = prev_instruction(prog);
 808       }
 809       if (inst && inst->Opcode != OPCODE_NOP) {
 810          /* and prev instruction's DstReg matches n->Children[0]->Store */
 811          inst->SaturateMode = SATURATE_ZERO_ONE;
 812          n->Store = n->Children[0]->Store;
 813          return inst;
 814       }
 815    }
 816 #endif
 817
 818    if (!n->Store)
 819       if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size))
 820          return NULL;
 821
 822    emit(emitInfo, n->Children[1]);
 823    emit(emitInfo, n->Children[2]);
 824
 825    /* tmp = max(ch[0], ch[1]) */
 826    inst = new_instruction(emitInfo, OPCODE_MAX);
 827    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 828    storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
 829    storage_to_src_reg(&inst->SrcReg[1], n->Children[1]->Store);
 830
 831    /* tmp = min(tmp, ch[2]) */
 832    inst = new_instruction(emitInfo, OPCODE_MIN);
 833    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 834    storage_to_src_reg(&inst->SrcReg[0], n->Store);
 835    storage_to_src_reg(&inst->SrcReg[1], n->Children[2]->Store);
 836
 837    return inst;
 838 }
 839
 840
 841 static struct prog_instruction *
 842 emit_negation(slang_emit_info *emitInfo, slang_ir_node *n)
 843 {
 844    /* Implement as MOV dst, -src; */
 845    /* XXX we could look at the previous instruction and in some circumstances
 846     * modify it to accomplish the negation.
 847     */
 848    struct prog_instruction *inst;
 849
 850    emit(emitInfo, n->Children[0]);
 851
 852    if (!n->Store)
 853       if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size))
 854          return NULL;
 855
 856    inst = new_instruction(emitInfo, OPCODE_MOV);
 857    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 858    storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
 859    inst->SrcReg[0].NegateBase = NEGATE_XYZW;
 860    return inst;
 861 }
 862
 863
 864 static struct prog_instruction *
 865 emit_label(slang_emit_info *emitInfo, const slang_ir_node *n)
 866 {
 867    assert(n->Label);
 868    assert(_slang_label_get_location(n->Label) < 0);
 869    _slang_label_set_location(n->Label, emitInfo->prog->NumInstructions,
 870                              emitInfo->prog);
 871    return NULL;
 872 }
 873
 874
 875 static struct prog_instruction *
 876 emit_jump(slang_emit_info *emitInfo, slang_ir_node *n)
 877 {
 878    struct prog_instruction *inst;
 879    assert(n);
 880    assert(n->Label);
 881    inst = new_instruction(emitInfo, OPCODE_BRA);
 882    inst->DstReg.CondMask = COND_TR;  /* always branch */
 883    inst->BranchTarget = _slang_label_get_location(n->Label);
 884    if (inst->BranchTarget < 0) {
 885       _slang_label_add_reference(n->Label, emitInfo->prog->NumInstructions - 1);
 886    }
 887    return inst;
 888 }
 889
 890
 891 static struct prog_instruction *
 892 emit_kill(slang_emit_info *emitInfo)
 893 {
 894    struct prog_instruction *inst;
 895    /* NV-KILL - discard fragment depending on condition code.
 896     * Note that ARB-KILL depends on sign of vector operand.
 897     */
 898    inst = new_instruction(emitInfo, OPCODE_KIL_NV);
 899    inst->DstReg.CondMask = COND_TR;  /* always branch */
 900    return inst;
 901 }
 902
 903
 904 static struct prog_instruction *
 905 emit_tex(slang_emit_info *emitInfo, slang_ir_node *n)
 906 {
 907    struct prog_instruction *inst;
 908
 909    (void) emit(emitInfo, n->Children[1]);
 910
 911    if (n->Opcode == IR_TEX) {
 912       inst = new_instruction(emitInfo, OPCODE_TEX);
 913    }
 914    else if (n->Opcode == IR_TEXB) {
 915       inst = new_instruction(emitInfo, OPCODE_TXB);
 916    }
 917    else {
 918       assert(n->Opcode == IR_TEXP);
 919       inst = new_instruction(emitInfo, OPCODE_TXP);
 920    }
 921
 922    if (!n->Store)
 923       if (!alloc_temp_storage(emitInfo, n, 4))
 924          return NULL;
 925
 926    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 927
 928    /* Child[1] is the coord */
 929    assert(n->Children[1]->Store->File != PROGRAM_UNDEFINED);
 930    assert(n->Children[1]->Store->Index >= 0);
 931    storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Store);
 932
 933    /* Child[0] is the sampler (a uniform which'll indicate the texture unit) */
 934    assert(n->Children[0]->Store);
 935    assert(n->Children[0]->Store->Size >= TEXTURE_1D_INDEX);
 936
 937    inst->Sampler = n->Children[0]->Store->Index; /* i.e. uniform's index */
 938    inst->TexSrcTarget = n->Children[0]->Store->Size;
 939    inst->TexSrcUnit = 27; /* Dummy value; the TexSrcUnit will be computed at
 940                            * link time, using the sampler uniform's value.
 941                            */
 942    return inst;
 943 }
 944
 945
 946 static struct prog_instruction *
 947 emit_move(slang_emit_info *emitInfo, slang_ir_node *n)
 948 {
 949    struct prog_instruction *inst;
 950
 951    /* lhs */
 952    emit(emitInfo, n->Children[0]);
 953
 954    /* rhs */
 955    assert(n->Children[1]);
 956    inst = emit(emitInfo, n->Children[1]);
 957
 958    assert(n->Children[1]->Store->Index >= 0);
 959
 960    assert(!n->Store);
 961    n->Store = n->Children[0]->Store;
 962
 963 #if PEEPHOLE_OPTIMIZATIONS
 964    if (inst && _slang_is_temp(emitInfo->vt, n->Children[1]->Store)) {
 965       /* Peephole optimization:
 966        * Just modify the RHS to put its result into the dest of this
 967        * MOVE operation.  Then, this MOVE is a no-op.
 968        */
 969       _slang_free_temp(emitInfo->vt, n->Children[1]->Store);
 970       *n->Children[1]->Store = *n->Children[0]->Store;
 971       /* fixup the prev (RHS) instruction */
 972       assert(n->Children[0]->Store->Index >= 0);
 973       storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store, n->Writemask);
 974       return inst;
 975    }
 976    else
 977 #endif
 978    {
 979       if (n->Children[0]->Store->Size > 4) {
 980          /* move matrix/struct etc (block of registers) */
 981          slang_ir_storage dstStore = *n->Children[0]->Store;
 982          slang_ir_storage srcStore = *n->Children[1]->Store;
 983          GLint size = srcStore.Size;
 984          ASSERT(n->Children[0]->Writemask == WRITEMASK_XYZW);
 985          ASSERT(n->Children[1]->Store->Swizzle == SWIZZLE_NOOP);
 986          dstStore.Size = 4;
 987          srcStore.Size = 4;
 988          while (size >= 4) {
 989             inst = new_instruction(emitInfo, OPCODE_MOV);
 990             inst->Comment = _mesa_strdup("IR_MOVE block");
 991             storage_to_dst_reg(&inst->DstReg, &dstStore, n->Writemask);
 992             storage_to_src_reg(&inst->SrcReg[0], &srcStore);
 993             srcStore.Index++;
 994             dstStore.Index++;
 995             size -= 4;
 996          }
 997       }
 998       else {
 999          /* single register move */
1000          char *srcAnnot, *dstAnnot;
1001          inst = new_instruction(emitInfo, OPCODE_MOV);
1002          assert(n->Children[0]->Store->Index >= 0);
1003          storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store, n->Writemask);
1004          storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Store);
1005          dstAnnot = storage_annotation(n->Children[0], emitInfo->prog);
1006          srcAnnot = storage_annotation(n->Children[1], emitInfo->prog);
1007          inst->Comment = instruction_annotation(inst->Opcode, dstAnnot,
1008                                                 srcAnnot, NULL, NULL);
1009       }
1010       free_temp_storage(emitInfo->vt, n->Children[1]);
1011       return inst;
1012    }
1013 }
1014
1015
1016 static struct prog_instruction *
1017 emit_cond(slang_emit_info *emitInfo, slang_ir_node *n)
1018 {
1019    /* Conditional expression (in if/while/for stmts).
1020     * Need to update condition code register.
1021     * Next instruction is typically an IR_IF.
1022     */
1023    struct prog_instruction *inst;
1024
1025    if (!n->Children[0])
1026       return NULL;
1027
1028    inst = emit(emitInfo, n->Children[0]);
1029    if (inst) {
1030       /* set inst's CondUpdate flag */
1031       inst->CondUpdate = GL_TRUE;
1032       return inst; /* XXX or null? */
1033    }
1034    else {
1035       /* This'll happen for things like "if (i) ..." where no code
1036        * is normally generated for the expression "i".
1037        * Generate a move instruction just to set condition codes.
1038        * Note: must use full 4-component vector since all four
1039        * condition codes must be set identically.
1040        */
1041       if (!alloc_temp_storage(emitInfo, n, 4))
1042          return NULL;
1043       inst = new_instruction(emitInfo, OPCODE_MOV);
1044       inst->CondUpdate = GL_TRUE;
1045       storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
1046       storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
1047       _slang_free_temp(emitInfo->vt, n->Store);
1048       inst->Comment = _mesa_strdup("COND expr");
1049       return inst; /* XXX or null? */
1050    }
1051 }
1052
1053
1054 /**
1055  * Logical-NOT
1056  */
1057 static struct prog_instruction *
1058 emit_not(slang_emit_info *emitInfo, slang_ir_node *n)
1059 {
1060    GLfloat zero = 0.0;
1061    slang_ir_storage st;
1062    struct prog_instruction *inst;
1063
1064    /* need zero constant */
1065    st.File = PROGRAM_CONSTANT;
1066    st.Size = 1;
1067    st.Index = _mesa_add_unnamed_constant(emitInfo->prog->Parameters, &zero,
1068                                          1, &st.Swizzle);
1069
1070    /* child expr */
1071    (void) emit(emitInfo, n->Children[0]);
1072    /* XXXX if child instr is SGT convert to SLE, if SEQ, SNE, etc */
1073
1074    if (!n->Store)
1075       if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size))
1076          return NULL;
1077
1078    inst = new_instruction(emitInfo, OPCODE_SEQ);
1079    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
1080    storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
1081    storage_to_src_reg(&inst->SrcReg[1], &st);
1082
1083    free_temp_storage(emitInfo->vt, n->Children[0]);
1084
1085    inst->Comment = _mesa_strdup("NOT");
1086    return inst;
1087 }
1088
1089
1090 static struct prog_instruction *
1091 emit_if(slang_emit_info *emitInfo, slang_ir_node *n)
1092 {
1093    struct gl_program *prog = emitInfo->prog;
1094    struct prog_instruction *ifInst;
1095    GLuint ifInstLoc, elseInstLoc = 0;
1096
1097    emit(emitInfo, n->Children[0]);  /* the condition */
1098    ifInstLoc = prog->NumInstructions;
1099    if (emitInfo->EmitHighLevelInstructions) {
1100       ifInst = new_instruction(emitInfo, OPCODE_IF);
1101       ifInst->DstReg.CondMask = COND_NE;  /* if cond is non-zero */
1102       ifInst->DstReg.CondSwizzle = SWIZZLE_X;
1103    }
1104    else {
1105       /* conditional jump to else, or endif */
1106       ifInst = new_instruction(emitInfo, OPCODE_BRA);
1107       ifInst->DstReg.CondMask = COND_EQ;  /* BRA if cond is zero */
1108       ifInst->DstReg.CondSwizzle = SWIZZLE_X;
1109       ifInst->Comment = _mesa_strdup("if zero");
1110    }
1111
1112    /* if body */
1113    emit(emitInfo, n->Children[1]);
1114
1115    if (n->Children[2]) {
1116       /* have else body */
1117       elseInstLoc = prog->NumInstructions;
1118       if (emitInfo->EmitHighLevelInstructions) {
1119          (void) new_instruction(emitInfo, OPCODE_ELSE);
1120       }
1121       else {
1122          /* jump to endif instruction */
1123          struct prog_instruction *inst;
1124          inst = new_instruction(emitInfo, OPCODE_BRA);
1125          inst->Comment = _mesa_strdup("else");
1126          inst->DstReg.CondMask = COND_TR;  /* always branch */
1127       }
1128       ifInst = prog->Instructions + ifInstLoc;
1129       ifInst->BranchTarget = prog->NumInstructions;
1130
1131       emit(emitInfo, n->Children[2]);
1132    }
1133    else {
1134       /* no else body */
1135       ifInst = prog->Instructions + ifInstLoc;
1136       ifInst->BranchTarget = prog->NumInstructions + 1;
1137    }
1138
1139    if (emitInfo->EmitHighLevelInstructions) {
1140       (void) new_instruction(emitInfo, OPCODE_ENDIF);
1141    }
1142
1143    if (n->Children[2]) {
1144       struct prog_instruction *elseInst;
1145       elseInst = prog->Instructions + elseInstLoc;
1146       elseInst->BranchTarget = prog->NumInstructions;
1147    }
1148    return NULL;
1149 }
1150
1151
1152 static struct prog_instruction *
1153 emit_loop(slang_emit_info *emitInfo, slang_ir_node *n)
1154 {
1155    struct gl_program *prog = emitInfo->prog;
1156    struct prog_instruction *beginInst, *endInst;
1157    GLuint beginInstLoc, endInstLoc;
1158    slang_ir_node *ir;
1159
1160    /* emit OPCODE_BGNLOOP */
1161    beginInstLoc = prog->NumInstructions;
1162    if (emitInfo->EmitHighLevelInstructions) {
1163       (void) new_instruction(emitInfo, OPCODE_BGNLOOP);
1164    }
1165
1166    /* body */
1167    emit(emitInfo, n->Children[0]);
1168
1169    endInstLoc = prog->NumInstructions;
1170    if (emitInfo->EmitHighLevelInstructions) {
1171       /* emit OPCODE_ENDLOOP */
1172       endInst = new_instruction(emitInfo, OPCODE_ENDLOOP);
1173    }
1174    else {
1175       /* emit unconditional BRA-nch */
1176       endInst = new_instruction(emitInfo, OPCODE_BRA);
1177       endInst->DstReg.CondMask = COND_TR;  /* always true */
1178    }
1179    /* end instruction's BranchTarget points to top of loop */
1180    endInst->BranchTarget = beginInstLoc;
1181
1182    if (emitInfo->EmitHighLevelInstructions) {
1183       /* BGNLOOP's BranchTarget points to the ENDLOOP inst */
1184       beginInst = prog->Instructions + beginInstLoc;
1185       beginInst->BranchTarget = prog->NumInstructions - 1;
1186    }
1187
1188    /* Done emitting loop code.  Now walk over the loop's linked list of
1189     * BREAK and CONT nodes, filling in their BranchTarget fields (which
1190     * will point to the ENDLOOP+1 or BGNLOOP instructions, respectively).
1191     */
1192    for (ir = n->BranchNode; ir; ir = ir->BranchNode) {
1193       struct prog_instruction *inst = prog->Instructions + ir->InstLocation;
1194       assert(inst->BranchTarget < 0);
1195       if (ir->Opcode == IR_BREAK ||
1196           ir->Opcode == IR_BREAK_IF_FALSE ||
1197           ir->Opcode == IR_BREAK_IF_TRUE) {
1198          assert(inst->Opcode == OPCODE_BRK ||
1199                 inst->Opcode == OPCODE_BRA);
1200          /* go to instruction after end of loop */
1201          inst->BranchTarget = endInstLoc + 1;
1202       }
1203       else {
1204          assert(ir->Opcode == IR_CONT ||
1205                 ir->Opcode == IR_CONT_IF_FALSE ||
1206                 ir->Opcode == IR_CONT_IF_TRUE);
1207          assert(inst->Opcode == OPCODE_CONT ||
1208                 inst->Opcode == OPCODE_BRA);
1209          /* to go instruction at top of loop */
1210          inst->BranchTarget = beginInstLoc;
1211       }
1212    }
1213    return NULL;
1214 }
1215
1216
1217 /**
1218  * "Continue" or "break" statement.
1219  * Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted.
1220  */
1221 static struct prog_instruction *
1222 emit_cont_break(slang_emit_info *emitInfo, slang_ir_node *n)
1223 {
1224    gl_inst_opcode opcode;
1225    struct prog_instruction *inst;
1226    n->InstLocation = emitInfo->prog->NumInstructions;
1227    if (emitInfo->EmitHighLevelInstructions) {
1228       opcode = (n->Opcode == IR_CONT) ? OPCODE_CONT : OPCODE_BRK;
1229    }
1230    else {
1231       opcode = OPCODE_BRA;
1232    }
1233    inst = new_instruction(emitInfo, opcode);
1234    inst->DstReg.CondMask = COND_TR;  /* always true */
1235    return inst;
1236 }
1237
1238
1239 /**
1240  * Conditional "continue" or "break" statement.
1241  * Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted.
1242  */
1243 static struct prog_instruction *
1244 emit_cont_break_if(slang_emit_info *emitInfo, slang_ir_node *n,
1245                    GLboolean breakTrue)
1246 {
1247    gl_inst_opcode opcode;
1248    struct prog_instruction *inst;
1249
1250    /* evaluate condition expr, setting cond codes */
1251    inst = emit(emitInfo, n->Children[0]);
1252    assert(inst);
1253    inst->CondUpdate = GL_TRUE;
1254
1255    n->InstLocation = emitInfo->prog->NumInstructions;
1256    if (emitInfo->EmitHighLevelInstructions) {
1257       if (n->Opcode == IR_CONT_IF_TRUE ||
1258           n->Opcode == IR_CONT_IF_FALSE)
1259          opcode = OPCODE_CONT;
1260       else
1261          opcode = OPCODE_BRK;
1262    }
1263    else {
1264       opcode = OPCODE_BRA;
1265    }
1266    inst = new_instruction(emitInfo, opcode);
1267    inst->DstReg.CondMask = breakTrue ? COND_NE : COND_EQ;
1268    return inst;
1269 }
1270
1271
1272
1273 /**
1274  * Remove any SWIZZLE_NIL terms from given swizzle mask (smear prev term).
1275  * Ex: fix_swizzle("zyNN") -> "zyyy"
1276  */
1277 static GLuint
1278 fix_swizzle(GLuint swizzle)
1279 {
1280    GLuint swz[4], i;
1281    for (i = 0; i < 4; i++) {
1282       swz[i] = GET_SWZ(swizzle, i);
1283       if (swz[i] == SWIZZLE_NIL) {
1284          swz[i] = swz[i - 1];
1285       }
1286    }
1287    return MAKE_SWIZZLE4(swz[0], swz[1], swz[2], swz[3]);
1288 }
1289
1290
1291 static struct prog_instruction *
1292 emit_swizzle(slang_emit_info *emitInfo, slang_ir_node *n)
1293 {
1294    GLuint swizzle;
1295
1296    /* swizzled storage access */
1297    (void) emit(emitInfo, n->Children[0]);
1298
1299    /* "pull-up" the child's storage info, applying our swizzle info */
1300    n->Store->File  = n->Children[0]->Store->File;
1301    n->Store->Index = n->Children[0]->Store->Index;
1302    n->Store->Size  = n->Children[0]->Store->Size;
1303    /*n->Var = n->Children[0]->Var; XXX for debug */
1304    assert(n->Store->Index >= 0);
1305
1306    swizzle = fix_swizzle(n->Store->Swizzle);
1307 #ifdef DEBUG
1308    {
1309       GLuint s = n->Children[0]->Store->Swizzle;
1310       assert(GET_SWZ(s, 0) != SWIZZLE_NIL);
1311       assert(GET_SWZ(s, 1) != SWIZZLE_NIL);
1312       assert(GET_SWZ(s, 2) != SWIZZLE_NIL);
1313       assert(GET_SWZ(s, 3) != SWIZZLE_NIL);
1314    }
1315 #endif
1316
1317    /* apply this swizzle to child's swizzle to get composed swizzle */
1318    n->Store->Swizzle = swizzle_swizzle(n->Children[0]->Store->Swizzle,
1319                                        swizzle);
1320    return NULL;
1321 }
1322
1323
1324 /**
1325  * Dereference array element.  Just resolve storage for the array
1326  * element represented by this node.
1327  */
1328 static struct prog_instruction *
1329 emit_array_element(slang_emit_info *emitInfo, slang_ir_node *n)
1330 {
1331    assert(n->Store);
1332    assert(n->Store->File != PROGRAM_UNDEFINED);
1333    assert(n->Store->Size > 0);
1334
1335    if (n->Store->File == PROGRAM_STATE_VAR) {
1336       n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
1337       return NULL;
1338    }
1339
1340
1341    if (n->Children[1]->Opcode == IR_FLOAT) {
1342       /* Constant index */
1343       const GLint arrayAddr = n->Children[0]->Store->Index;
1344       const GLint index = (GLint) n->Children[1]->Value[0];
1345       n->Store->Index = arrayAddr + index;
1346    }
1347    else {
1348       /* Variable index - PROBLEM */
1349       const GLint arrayAddr = n->Children[0]->Store->Index;
1350       const GLint index = 0;
1351       _mesa_problem(NULL, "variable array indexes not supported yet!");
1352       n->Store->Index = arrayAddr + index;
1353    }
1354    return NULL; /* no instruction */
1355 }
1356
1357
1358 /**
1359  * Resolve storage for accessing a structure field.
1360  */
1361 static struct prog_instruction *
1362 emit_struct_field(slang_emit_info *emitInfo, slang_ir_node *n)
1363 {
1364    if (n->Store->File == PROGRAM_STATE_VAR) {
1365       n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
1366    }
1367    else {
1368       _mesa_problem(NULL, "structs/fields not supported yet");
1369    }
1370    return NULL; /* no instruction */
1371 }
1372
1373
1374 static struct prog_instruction *
1375 emit(slang_emit_info *emitInfo, slang_ir_node *n)
1376 {
1377    struct prog_instruction *inst;
1378    if (!n)
1379       return NULL;
1380
1381    switch (n->Opcode) {
1382    case IR_SEQ:
1383       /* sequence of two sub-trees */
1384       assert(n->Children[0]);
1385       assert(n->Children[1]);
1386       emit(emitInfo, n->Children[0]);
1387       inst = emit(emitInfo, n->Children[1]);
1388       assert(!n->Store);
1389       n->Store = n->Children[1]->Store;
1390       return inst;
1391
1392    case IR_SCOPE:
1393       /* new variable scope */
1394       _slang_push_var_table(emitInfo->vt);
1395       inst = emit(emitInfo, n->Children[0]);
1396       _slang_pop_var_table(emitInfo->vt);
1397       return inst;
1398
1399    case IR_VAR_DECL:
1400       /* Variable declaration - allocate a register for it */
1401       assert(n->Store);
1402       assert(n->Store->File != PROGRAM_UNDEFINED);
1403       assert(n->Store->Size > 0);
1404       assert(n->Store->Index < 0);
1405       if (!n->Var || n->Var->isTemp) {
1406          /* a nameless/temporary variable, will be freed after first use */
1407          if (!_slang_alloc_temp(emitInfo->vt, n->Store)) {
1408             slang_info_log_error(emitInfo->log,
1409                                  "Ran out of registers, too many temporaries");
1410             return NULL;
1411          }
1412       }
1413       else {
1414          /* a regular variable */
1415          _slang_add_variable(emitInfo->vt, n->Var);
1416          if (!_slang_alloc_var(emitInfo->vt, n->Store)) {
1417             slang_info_log_error(emitInfo->log,
1418                                  "Ran out of registers, too many variables");
1419             return NULL;
1420          }
1421          /*
1422          printf("IR_VAR_DECL %s %d store %p\n",
1423                 (char*) n->Var->a_name, n->Store->Index, (void*) n->Store);
1424          */
1425          assert(n->Var->aux == n->Store);
1426       }
1427       if (emitInfo->EmitComments) {
1428          /* emit NOP with comment describing the variable's storage location */
1429          char s[1000];
1430          sprintf(s, "TEMP[%d]%s = %s (size %d)",
1431                  n->Store->Index,
1432                  _mesa_swizzle_string(n->Store->Swizzle, 0, GL_FALSE),
1433                  (char *) n->Var->a_name,
1434                  n->Store->Size);
1435          inst = new_instruction(emitInfo, OPCODE_NOP);
1436          inst->Comment = _mesa_strdup(s);
1437          return inst;
1438       }
1439       return NULL;
1440
1441    case IR_VAR:
1442       /* Reference to a variable
1443        * Storage should have already been resolved/allocated.
1444        */
1445       assert(n->Store);
1446       assert(n->Store->File != PROGRAM_UNDEFINED);
1447
1448       if (n->Store->File == PROGRAM_STATE_VAR &&
1449           n->Store->Index < 0) {
1450          n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
1451       }
1452
1453       if (n->Store->Index < 0) {
1454          printf("#### VAR %s not allocated!\n", (char*)n->Var->a_name);
1455       }
1456       assert(n->Store->Index >= 0);
1457       assert(n->Store->Size > 0);
1458       break;
1459
1460    case IR_ELEMENT:
1461       return emit_array_element(emitInfo, n);
1462    case IR_FIELD:
1463       return emit_struct_field(emitInfo, n);
1464    case IR_SWIZZLE:
1465       return emit_swizzle(emitInfo, n);
1466
1467    case IR_I_TO_F:
1468       /* just move */
1469       emit(emitInfo, n->Children[0]);
1470       inst = new_instruction(emitInfo, OPCODE_MOV);
1471       if (!n->Store) {
1472          if (!alloc_temp_storage(emitInfo, n, 1))
1473             return NULL;
1474       }
1475       storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
1476       storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
1477       if (emitInfo->EmitComments)
1478          inst->Comment = _mesa_strdup("int to float");
1479       return NULL;
1480
1481    /* Simple arithmetic */
1482    /* unary */
1483    case IR_RSQ:
1484    case IR_RCP:
1485    case IR_FLOOR:
1486    case IR_FRAC:
1487    case IR_F_TO_I:
1488    case IR_ABS:
1489    case IR_SIN:
1490    case IR_COS:
1491    case IR_DDX:
1492    case IR_DDY:
1493    case IR_NOISE1:
1494    case IR_NOISE2:
1495    case IR_NOISE3:
1496    case IR_NOISE4:
1497    /* binary */
1498    case IR_ADD:
1499    case IR_SUB:
1500    case IR_MUL:
1501    case IR_DOT4:
1502    case IR_DOT3:
1503    case IR_CROSS:
1504    case IR_MIN:
1505    case IR_MAX:
1506    case IR_SEQUAL:
1507    case IR_SNEQUAL:
1508    case IR_SGE:
1509    case IR_SGT:
1510    case IR_SLE:
1511    case IR_SLT:
1512    case IR_POW:
1513    case IR_EXP:
1514    case IR_EXP2:
1515    /* trinary operators */
1516    case IR_LRP:
1517       return emit_arith(emitInfo, n);
1518    case IR_CLAMP:
1519       return emit_clamp(emitInfo, n);
1520    case IR_TEX:
1521    case IR_TEXB:
1522    case IR_TEXP:
1523       return emit_tex(emitInfo, n);
1524    case IR_NEG:
1525       return emit_negation(emitInfo, n);
1526    case IR_FLOAT:
1527       /* find storage location for this float constant */
1528       n->Store->Index = _mesa_add_unnamed_constant(emitInfo->prog->Parameters, n->Value,
1529                                                    n->Store->Size,
1530                                                    &n->Store->Swizzle);
1531       if (n->Store->Index < 0) {
1532          slang_info_log_error(emitInfo->log, "Ran out of space for constants");
1533          return NULL;
1534       }
1535       return NULL;
1536
1537    case IR_MOVE:
1538       return emit_move(emitInfo, n);
1539
1540    case IR_COND:
1541       return emit_cond(emitInfo, n);
1542
1543    case IR_NOT:
1544       return emit_not(emitInfo, n);
1545
1546    case IR_LABEL:
1547       return emit_label(emitInfo, n);
1548    case IR_JUMP:
1549       assert(n);
1550       assert(n->Label);
1551       return emit_jump(emitInfo, n);
1552    case IR_KILL:
1553       return emit_kill(emitInfo);
1554
1555    case IR_IF:
1556       return emit_if(emitInfo, n);
1557
1558    case IR_LOOP:
1559       return emit_loop(emitInfo, n);
1560    case IR_BREAK_IF_FALSE:
1561    case IR_CONT_IF_FALSE:
1562       return emit_cont_break_if(emitInfo, n, GL_FALSE);
1563    case IR_BREAK_IF_TRUE:
1564    case IR_CONT_IF_TRUE:
1565       return emit_cont_break_if(emitInfo, n, GL_TRUE);
1566    case IR_BREAK:
1567       /* fall-through */
1568    case IR_CONT:
1569       return emit_cont_break(emitInfo, n);
1570
1571    case IR_BEGIN_SUB:
1572       return new_instruction(emitInfo, OPCODE_BGNSUB);
1573    case IR_END_SUB:
1574       return new_instruction(emitInfo, OPCODE_ENDSUB);
1575    case IR_RETURN:
1576       return new_instruction(emitInfo, OPCODE_RET);
1577
1578    case IR_NOP:
1579       return NULL;
1580
1581    default:
1582       _mesa_problem(NULL, "Unexpected IR opcode in emit()\n");
1583       abort();
1584    }
1585    return NULL;
1586 }
1587
1588
1589 GLboolean
1590 _slang_emit_code(slang_ir_node *n, slang_var_table *vt,
1591                  struct gl_program *prog, GLboolean withEnd,
1592                  slang_info_log *log)
1593 {
1594    GET_CURRENT_CONTEXT(ctx);
1595    GLboolean success;
1596    slang_emit_info emitInfo;
1597
1598    emitInfo.log = log;
1599    emitInfo.vt = vt;
1600    emitInfo.prog = prog;
1601
1602    emitInfo.EmitHighLevelInstructions = ctx->Shader.EmitHighLevelInstructions;
1603    emitInfo.EmitComments = ctx->Shader.EmitComments;
1604
1605    (void) emit(&emitInfo, n);
1606
1607    /* finish up by adding the END opcode to program */
1608    if (withEnd) {
1609       struct prog_instruction *inst;
1610       inst = new_instruction(&emitInfo, OPCODE_END);
1611    }
1612    success = GL_TRUE;
1613
1614 #if 0
1615    printf("*********** End emit code (%u inst):\n", prog->NumInstructions);
1616    _mesa_print_program(prog);
1617    _mesa_print_program_parameters(ctx,prog);
1618 #endif
1619
1620    return success;
1621 }