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 && st->File <= 16);
 486    src->File = st->File;
 487    src->Index = st->Index;
 488    assert(st->File != PROGRAM_UNDEFINED);
 489    assert(st->Size >= 1);
 490    assert(st->Size <= 4);
 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    return inst;
 524 }
 525
 526
 527 #if 0
 528 /**
 529  * Return pointer to last instruction in program.
 530  */
 531 static struct prog_instruction *
 532 prev_instruction(struct gl_program *prog)
 533 {
 534    if (prog->NumInstructions == 0)
 535       return NULL;
 536    else
 537       return prog->Instructions + prog->NumInstructions - 1;
 538 }
 539 #endif
 540
 541
 542 static struct prog_instruction *
 543 emit(slang_emit_info *emitInfo, slang_ir_node *n);
 544
 545
 546 /**
 547  * Return an annotation string for given node's storage.
 548  */
 549 static char *
 550 storage_annotation(const slang_ir_node *n, const struct gl_program *prog)
 551 {
 552 #if ANNOTATE
 553    const slang_ir_storage *st = n->Store;
 554    static char s[100] = "";
 555
 556    if (!st)
 557       return _mesa_strdup("");
 558
 559    switch (st->File) {
 560    case PROGRAM_CONSTANT:
 561       if (st->Index >= 0) {
 562          const GLfloat *val = prog->Parameters->ParameterValues[st->Index];
 563          if (st->Swizzle == SWIZZLE_NOOP)
 564             sprintf(s, "{%g, %g, %g, %g}", val[0], val[1], val[2], val[3]);
 565          else {
 566             sprintf(s, "%g", val[GET_SWZ(st->Swizzle, 0)]);
 567          }
 568       }
 569       break;
 570    case PROGRAM_TEMPORARY:
 571       if (n->Var)
 572          sprintf(s, "%s", (char *) n->Var->a_name);
 573       else
 574          sprintf(s, "t[%d]", st->Index);
 575       break;
 576    case PROGRAM_STATE_VAR:
 577    case PROGRAM_UNIFORM:
 578       sprintf(s, "%s", prog->Parameters->Parameters[st->Index].Name);
 579       break;
 580    case PROGRAM_VARYING:
 581       sprintf(s, "%s", prog->Varying->Parameters[st->Index].Name);
 582       break;
 583    case PROGRAM_INPUT:
 584       sprintf(s, "input[%d]", st->Index);
 585       break;
 586    case PROGRAM_OUTPUT:
 587       sprintf(s, "output[%d]", st->Index);
 588       break;
 589    default:
 590       s[0] = 0;
 591    }
 592    return _mesa_strdup(s);
 593 #else
 594    return NULL;
 595 #endif
 596 }
 597
 598
 599 /**
 600  * Return an annotation string for an instruction.
 601  */
 602 static char *
 603 instruction_annotation(gl_inst_opcode opcode, char *dstAnnot,
 604                        char *srcAnnot0, char *srcAnnot1, char *srcAnnot2)
 605 {
 606 #if ANNOTATE
 607    const char *operator;
 608    char *s;
 609    int len = 50;
 610
 611    if (dstAnnot)
 612       len += strlen(dstAnnot);
 613    else
 614       dstAnnot = _mesa_strdup("");
 615
 616    if (srcAnnot0)
 617       len += strlen(srcAnnot0);
 618    else
 619       srcAnnot0 = _mesa_strdup("");
 620
 621    if (srcAnnot1)
 622       len += strlen(srcAnnot1);
 623    else
 624       srcAnnot1 = _mesa_strdup("");
 625
 626    if (srcAnnot2)
 627       len += strlen(srcAnnot2);
 628    else
 629       srcAnnot2 = _mesa_strdup("");
 630
 631    switch (opcode) {
 632    case OPCODE_ADD:
 633       operator = "+";
 634       break;
 635    case OPCODE_SUB:
 636       operator = "-";
 637       break;
 638    case OPCODE_MUL:
 639       operator = "*";
 640       break;
 641    case OPCODE_DP3:
 642       operator = "DP3";
 643       break;
 644    case OPCODE_DP4:
 645       operator = "DP4";
 646       break;
 647    case OPCODE_XPD:
 648       operator = "XPD";
 649       break;
 650    case OPCODE_RSQ:
 651       operator = "RSQ";
 652       break;
 653    case OPCODE_SGT:
 654       operator = ">";
 655       break;
 656    default:
 657       operator = ",";
 658    }
 659
 660    s = (char *) malloc(len);
 661    sprintf(s, "%s = %s %s %s %s", dstAnnot,
 662            srcAnnot0, operator, srcAnnot1, srcAnnot2);
 663    assert(_mesa_strlen(s) < len);
 664
 665    free(dstAnnot);
 666    free(srcAnnot0);
 667    free(srcAnnot1);
 668    free(srcAnnot2);
 669
 670    return s;
 671 #else
 672    return NULL;
 673 #endif
 674 }
 675
 676
 677
 678 /**
 679  * Generate code for a simple arithmetic instruction.
 680  * Either 1, 2 or 3 operands.
 681  */
 682 static struct prog_instruction *
 683 emit_arith(slang_emit_info *emitInfo, slang_ir_node *n)
 684 {
 685    struct prog_instruction *inst;
 686    const slang_ir_info *info = slang_find_ir_info(n->Opcode);
 687    char *srcAnnot[3], *dstAnnot;
 688    GLuint i;
 689
 690    assert(info);
 691    assert(info->InstOpcode != OPCODE_NOP);
 692
 693    srcAnnot[0] = srcAnnot[1] = srcAnnot[2] = dstAnnot = NULL;
 694
 695 #if PEEPHOLE_OPTIMIZATIONS
 696    /* Look for MAD opportunity */
 697    if (info->NumParams == 2 &&
 698        n->Opcode == IR_ADD && n->Children[0]->Opcode == IR_MUL) {
 699       /* found pattern IR_ADD(IR_MUL(A, B), C) */
 700       emit(emitInfo, n->Children[0]->Children[0]);  /* A */
 701       emit(emitInfo, n->Children[0]->Children[1]);  /* B */
 702       emit(emitInfo, n->Children[1]);  /* C */
 703       /* generate MAD instruction */
 704       inst = new_instruction(emitInfo, OPCODE_MAD);
 705       /* operands: A, B, C: */
 706       storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Children[0]->Store);
 707       storage_to_src_reg(&inst->SrcReg[1], n->Children[0]->Children[1]->Store);
 708       storage_to_src_reg(&inst->SrcReg[2], n->Children[1]->Store);
 709       free_temp_storage(emitInfo->vt, n->Children[0]->Children[0]);
 710       free_temp_storage(emitInfo->vt, n->Children[0]->Children[1]);
 711       free_temp_storage(emitInfo->vt, n->Children[1]);
 712    }
 713    else if (info->NumParams == 2 &&
 714             n->Opcode == IR_ADD && n->Children[1]->Opcode == IR_MUL) {
 715       /* found pattern IR_ADD(A, IR_MUL(B, C)) */
 716       emit(emitInfo, n->Children[0]);  /* A */
 717       emit(emitInfo, n->Children[1]->Children[0]);  /* B */
 718       emit(emitInfo, n->Children[1]->Children[1]);  /* C */
 719       /* generate MAD instruction */
 720       inst = new_instruction(emitInfo, OPCODE_MAD);
 721       /* operands: B, C, A */
 722       storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Children[0]->Store);
 723       storage_to_src_reg(&inst->SrcReg[1], n->Children[1]->Children[1]->Store);
 724       storage_to_src_reg(&inst->SrcReg[2], n->Children[0]->Store);
 725       free_temp_storage(emitInfo->vt, n->Children[1]->Children[0]);
 726       free_temp_storage(emitInfo->vt, n->Children[1]->Children[1]);
 727       free_temp_storage(emitInfo->vt, n->Children[0]);
 728    }
 729    else
 730 #endif
 731    {
 732       /* normal case */
 733
 734       /* gen code for children */
 735       for (i = 0; i < info->NumParams; i++)
 736          emit(emitInfo, n->Children[i]);
 737
 738       /* gen this instruction and src registers */
 739       inst = new_instruction(emitInfo, info->InstOpcode);
 740       for (i = 0; i < info->NumParams; i++)
 741          storage_to_src_reg(&inst->SrcReg[i], n->Children[i]->Store);
 742
 743       /* annotation */
 744       for (i = 0; i < info->NumParams; i++)
 745          srcAnnot[i] = storage_annotation(n->Children[i], emitInfo->prog);
 746
 747       /* free temps */
 748       for (i = 0; i < info->NumParams; i++)
 749          free_temp_storage(emitInfo->vt, n->Children[i]);
 750    }
 751
 752    /* result storage */
 753    if (!n->Store) {
 754       if (!alloc_temp_storage(emitInfo, n, info->ResultSize))
 755          return NULL;
 756    }
 757    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 758
 759    dstAnnot = storage_annotation(n, emitInfo->prog);
 760
 761    inst->Comment = instruction_annotation(inst->Opcode, dstAnnot, srcAnnot[0],
 762                                           srcAnnot[1], srcAnnot[2]);
 763
 764    /*_mesa_print_instruction(inst);*/
 765    return inst;
 766 }
 767
 768
 769 /**
 770  * Generate code for an IR_CLAMP instruction.
 771  */
 772 static struct prog_instruction *
 773 emit_clamp(slang_emit_info *emitInfo, slang_ir_node *n)
 774 {
 775    struct prog_instruction *inst;
 776
 777    assert(n->Opcode == IR_CLAMP);
 778    /* ch[0] = value
 779     * ch[1] = min limit
 780     * ch[2] = max limit
 781     */
 782
 783    inst = emit(emitInfo, n->Children[0]);
 784
 785    /* If lower limit == 0.0 and upper limit == 1.0,
 786     *    set prev instruction's SaturateMode field to SATURATE_ZERO_ONE.
 787     * Else,
 788     *    emit OPCODE_MIN, OPCODE_MAX sequence.
 789     */
 790 #if 0
 791    /* XXX this isn't quite finished yet */
 792    if (n->Children[1]->Opcode == IR_FLOAT &&
 793        n->Children[1]->Value[0] == 0.0 &&
 794        n->Children[1]->Value[1] == 0.0 &&
 795        n->Children[1]->Value[2] == 0.0 &&
 796        n->Children[1]->Value[3] == 0.0 &&
 797        n->Children[2]->Opcode == IR_FLOAT &&
 798        n->Children[2]->Value[0] == 1.0 &&
 799        n->Children[2]->Value[1] == 1.0 &&
 800        n->Children[2]->Value[2] == 1.0 &&
 801        n->Children[2]->Value[3] == 1.0) {
 802       if (!inst) {
 803          inst = prev_instruction(prog);
 804       }
 805       if (inst && inst->Opcode != OPCODE_NOP) {
 806          /* and prev instruction's DstReg matches n->Children[0]->Store */
 807          inst->SaturateMode = SATURATE_ZERO_ONE;
 808          n->Store = n->Children[0]->Store;
 809          return inst;
 810       }
 811    }
 812 #endif
 813
 814    if (!n->Store)
 815       if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size))
 816          return NULL;
 817
 818    emit(emitInfo, n->Children[1]);
 819    emit(emitInfo, n->Children[2]);
 820
 821    /* tmp = max(ch[0], ch[1]) */
 822    inst = new_instruction(emitInfo, OPCODE_MAX);
 823    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 824    storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
 825    storage_to_src_reg(&inst->SrcReg[1], n->Children[1]->Store);
 826
 827    /* tmp = min(tmp, ch[2]) */
 828    inst = new_instruction(emitInfo, OPCODE_MIN);
 829    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 830    storage_to_src_reg(&inst->SrcReg[0], n->Store);
 831    storage_to_src_reg(&inst->SrcReg[1], n->Children[2]->Store);
 832
 833    return inst;
 834 }
 835
 836
 837 static struct prog_instruction *
 838 emit_negation(slang_emit_info *emitInfo, slang_ir_node *n)
 839 {
 840    /* Implement as MOV dst, -src; */
 841    /* XXX we could look at the previous instruction and in some circumstances
 842     * modify it to accomplish the negation.
 843     */
 844    struct prog_instruction *inst;
 845
 846    emit(emitInfo, n->Children[0]);
 847
 848    if (!n->Store)
 849       if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size))
 850          return NULL;
 851
 852    inst = new_instruction(emitInfo, OPCODE_MOV);
 853    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 854    storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
 855    inst->SrcReg[0].NegateBase = NEGATE_XYZW;
 856    return inst;
 857 }
 858
 859
 860 static struct prog_instruction *
 861 emit_label(slang_emit_info *emitInfo, const slang_ir_node *n)
 862 {
 863    assert(n->Label);
 864    assert(_slang_label_get_location(n->Label) < 0);
 865    _slang_label_set_location(n->Label, emitInfo->prog->NumInstructions,
 866                              emitInfo->prog);
 867    return NULL;
 868 }
 869
 870
 871 static struct prog_instruction *
 872 emit_jump(slang_emit_info *emitInfo, slang_ir_node *n)
 873 {
 874    struct prog_instruction *inst;
 875    assert(n);
 876    assert(n->Label);
 877    inst = new_instruction(emitInfo, OPCODE_BRA);
 878    inst->DstReg.CondMask = COND_TR;  /* always branch */
 879    inst->BranchTarget = _slang_label_get_location(n->Label);
 880    if (inst->BranchTarget < 0) {
 881       _slang_label_add_reference(n->Label, emitInfo->prog->NumInstructions - 1);
 882    }
 883    return inst;
 884 }
 885
 886
 887 static struct prog_instruction *
 888 emit_kill(slang_emit_info *emitInfo)
 889 {
 890    struct prog_instruction *inst;
 891    /* NV-KILL - discard fragment depending on condition code.
 892     * Note that ARB-KILL depends on sign of vector operand.
 893     */
 894    inst = new_instruction(emitInfo, OPCODE_KIL_NV);
 895    inst->DstReg.CondMask = COND_TR;  /* always branch */
 896    return inst;
 897 }
 898
 899
 900 static struct prog_instruction *
 901 emit_tex(slang_emit_info *emitInfo, slang_ir_node *n)
 902 {
 903    struct prog_instruction *inst;
 904    if (n->Opcode == IR_TEX) {
 905       inst = new_instruction(emitInfo, OPCODE_TEX);
 906    }
 907    else if (n->Opcode == IR_TEXB) {
 908       inst = new_instruction(emitInfo, OPCODE_TXB);
 909    }
 910    else {
 911       assert(n->Opcode == IR_TEXP);
 912       inst = new_instruction(emitInfo, OPCODE_TXP);
 913    }
 914
 915    if (!n->Store)
 916       if (!alloc_temp_storage(emitInfo, n, 4))
 917          return NULL;
 918
 919    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 920
 921    (void) emit(emitInfo, n->Children[1]);
 922
 923    /* Child[1] is the coord */
 924    storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Store);
 925
 926    /* Child[0] is the sampler (a uniform which'll indicate the texture unit) */
 927    assert(n->Children[0]->Store);
 928    assert(n->Children[0]->Store->Size >= TEXTURE_1D_INDEX);
 929
 930    inst->Sampler = n->Children[0]->Store->Index; /* i.e. uniform's index */
 931    inst->TexSrcTarget = n->Children[0]->Store->Size;
 932    inst->TexSrcUnit = 27; /* Dummy value; the TexSrcUnit will be computed at
 933                            * link time, using the sampler uniform's value.
 934                            */
 935    return inst;
 936 }
 937
 938
 939 static struct prog_instruction *
 940 emit_move(slang_emit_info *emitInfo, slang_ir_node *n)
 941 {
 942    struct prog_instruction *inst;
 943
 944    /* rhs */
 945    assert(n->Children[1]);
 946    inst = emit(emitInfo, n->Children[1]);
 947
 948    assert(n->Children[1]->Store->Index >= 0);
 949
 950    /* lhs */
 951    emit(emitInfo, n->Children[0]);
 952
 953    assert(!n->Store);
 954    n->Store = n->Children[0]->Store;
 955
 956 #if PEEPHOLE_OPTIMIZATIONS
 957    if (inst && _slang_is_temp(emitInfo->vt, n->Children[1]->Store)) {
 958       /* Peephole optimization:
 959        * Just modify the RHS to put its result into the dest of this
 960        * MOVE operation.  Then, this MOVE is a no-op.
 961        */
 962       _slang_free_temp(emitInfo->vt, n->Children[1]->Store);
 963       *n->Children[1]->Store = *n->Children[0]->Store;
 964       /* fixup the prev (RHS) instruction */
 965       assert(n->Children[0]->Store->Index >= 0);
 966       assert(n->Children[0]->Store->Index < 16);
 967       storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store, n->Writemask);
 968       return inst;
 969    }
 970    else
 971 #endif
 972    {
 973       if (n->Children[0]->Store->Size > 4) {
 974          /* move matrix/struct etc (block of registers) */
 975          slang_ir_storage dstStore = *n->Children[0]->Store;
 976          slang_ir_storage srcStore = *n->Children[1]->Store;
 977          GLint size = srcStore.Size;
 978          ASSERT(n->Children[0]->Writemask == WRITEMASK_XYZW);
 979          ASSERT(n->Children[1]->Store->Swizzle == SWIZZLE_NOOP);
 980          dstStore.Size = 4;
 981          srcStore.Size = 4;
 982          while (size >= 4) {
 983             inst = new_instruction(emitInfo, OPCODE_MOV);
 984             inst->Comment = _mesa_strdup("IR_MOVE block");
 985             storage_to_dst_reg(&inst->DstReg, &dstStore, n->Writemask);
 986             storage_to_src_reg(&inst->SrcReg[0], &srcStore);
 987             srcStore.Index++;
 988             dstStore.Index++;
 989             size -= 4;
 990          }
 991       }
 992       else {
 993          /* single register move */
 994          char *srcAnnot, *dstAnnot;
 995          inst = new_instruction(emitInfo, OPCODE_MOV);
 996          assert(n->Children[0]->Store->Index >= 0);
 997          storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store, n->Writemask);
 998          storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Store);
 999          dstAnnot = storage_annotation(n->Children[0], emitInfo->prog);
1000          srcAnnot = storage_annotation(n->Children[1], emitInfo->prog);
1001          inst->Comment = instruction_annotation(inst->Opcode, dstAnnot,
1002                                                 srcAnnot, NULL, NULL);
1003       }
1004       free_temp_storage(emitInfo->vt, n->Children[1]);
1005       return inst;
1006    }
1007 }
1008
1009
1010 static struct prog_instruction *
1011 emit_cond(slang_emit_info *emitInfo, slang_ir_node *n)
1012 {
1013    /* Conditional expression (in if/while/for stmts).
1014     * Need to update condition code register.
1015     * Next instruction is typically an IR_IF.
1016     */
1017    struct prog_instruction *inst;
1018
1019    if (!n->Children[0])
1020       return NULL;
1021
1022    inst = emit(emitInfo, n->Children[0]);
1023    if (inst) {
1024       /* set inst's CondUpdate flag */
1025       inst->CondUpdate = GL_TRUE;
1026       return inst; /* XXX or null? */
1027    }
1028    else {
1029       /* This'll happen for things like "if (i) ..." where no code
1030        * is normally generated for the expression "i".
1031        * Generate a move instruction just to set condition codes.
1032        * Note: must use full 4-component vector since all four
1033        * condition codes must be set identically.
1034        */
1035       if (!alloc_temp_storage(emitInfo, n, 4))
1036          return NULL;
1037       inst = new_instruction(emitInfo, OPCODE_MOV);
1038       inst->CondUpdate = GL_TRUE;
1039       storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
1040       storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
1041       _slang_free_temp(emitInfo->vt, n->Store);
1042       inst->Comment = _mesa_strdup("COND expr");
1043       return inst; /* XXX or null? */
1044    }
1045 }
1046
1047
1048 /**
1049  * Logical-NOT
1050  */
1051 static struct prog_instruction *
1052 emit_not(slang_emit_info *emitInfo, slang_ir_node *n)
1053 {
1054    GLfloat zero = 0.0;
1055    slang_ir_storage st;
1056    struct prog_instruction *inst;
1057
1058    /* need zero constant */
1059    st.File = PROGRAM_CONSTANT;
1060    st.Size = 1;
1061    st.Index = _mesa_add_unnamed_constant(emitInfo->prog->Parameters, &zero,
1062                                          1, &st.Swizzle);
1063
1064    /* child expr */
1065    (void) emit(emitInfo, n->Children[0]);
1066    /* XXXX if child instr is SGT convert to SLE, if SEQ, SNE, etc */
1067
1068    if (!n->Store)
1069       if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size))
1070          return NULL;
1071
1072    inst = new_instruction(emitInfo, OPCODE_SEQ);
1073    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
1074    storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
1075    storage_to_src_reg(&inst->SrcReg[1], &st);
1076
1077    free_temp_storage(emitInfo->vt, n->Children[0]);
1078
1079    inst->Comment = _mesa_strdup("NOT");
1080    return inst;
1081 }
1082
1083
1084 static struct prog_instruction *
1085 emit_if(slang_emit_info *emitInfo, slang_ir_node *n)
1086 {
1087    struct gl_program *prog = emitInfo->prog;
1088    struct prog_instruction *ifInst;
1089    GLuint ifInstLoc, elseInstLoc = 0;
1090
1091    emit(emitInfo, n->Children[0]);  /* the condition */
1092    ifInstLoc = prog->NumInstructions;
1093    if (emitInfo->EmitHighLevelInstructions) {
1094       ifInst = new_instruction(emitInfo, OPCODE_IF);
1095       ifInst->DstReg.CondMask = COND_NE;  /* if cond is non-zero */
1096       ifInst->DstReg.CondSwizzle = SWIZZLE_X;
1097    }
1098    else {
1099       /* conditional jump to else, or endif */
1100       ifInst = new_instruction(emitInfo, OPCODE_BRA);
1101       ifInst->DstReg.CondMask = COND_EQ;  /* BRA if cond is zero */
1102       ifInst->DstReg.CondSwizzle = SWIZZLE_X;
1103       ifInst->Comment = _mesa_strdup("if zero");
1104    }
1105
1106    /* if body */
1107    emit(emitInfo, n->Children[1]);
1108
1109    if (n->Children[2]) {
1110       /* have else body */
1111       elseInstLoc = prog->NumInstructions;
1112       if (emitInfo->EmitHighLevelInstructions) {
1113          (void) new_instruction(emitInfo, OPCODE_ELSE);
1114       }
1115       else {
1116          /* jump to endif instruction */
1117          struct prog_instruction *inst;
1118          inst = new_instruction(emitInfo, OPCODE_BRA);
1119          inst->Comment = _mesa_strdup("else");
1120          inst->DstReg.CondMask = COND_TR;  /* always branch */
1121       }
1122       ifInst = prog->Instructions + ifInstLoc;
1123       ifInst->BranchTarget = prog->NumInstructions;
1124
1125       emit(emitInfo, n->Children[2]);
1126    }
1127    else {
1128       /* no else body */
1129       ifInst = prog->Instructions + ifInstLoc;
1130       ifInst->BranchTarget = prog->NumInstructions + 1;
1131    }
1132
1133    if (emitInfo->EmitHighLevelInstructions) {
1134       (void) new_instruction(emitInfo, OPCODE_ENDIF);
1135    }
1136
1137    if (n->Children[2]) {
1138       struct prog_instruction *elseInst;
1139       elseInst = prog->Instructions + elseInstLoc;
1140       elseInst->BranchTarget = prog->NumInstructions;
1141    }
1142    return NULL;
1143 }
1144
1145
1146 static struct prog_instruction *
1147 emit_loop(slang_emit_info *emitInfo, slang_ir_node *n)
1148 {
1149    struct gl_program *prog = emitInfo->prog;
1150    struct prog_instruction *beginInst, *endInst;
1151    GLuint beginInstLoc, endInstLoc;
1152    slang_ir_node *ir;
1153
1154    /* emit OPCODE_BGNLOOP */
1155    beginInstLoc = prog->NumInstructions;
1156    if (emitInfo->EmitHighLevelInstructions) {
1157       (void) new_instruction(emitInfo, OPCODE_BGNLOOP);
1158    }
1159
1160    /* body */
1161    emit(emitInfo, n->Children[0]);
1162
1163    endInstLoc = prog->NumInstructions;
1164    if (emitInfo->EmitHighLevelInstructions) {
1165       /* emit OPCODE_ENDLOOP */
1166       endInst = new_instruction(emitInfo, OPCODE_ENDLOOP);
1167    }
1168    else {
1169       /* emit unconditional BRA-nch */
1170       endInst = new_instruction(emitInfo, OPCODE_BRA);
1171       endInst->DstReg.CondMask = COND_TR;  /* always true */
1172    }
1173    /* end instruction's BranchTarget points to top of loop */
1174    endInst->BranchTarget = beginInstLoc;
1175
1176    if (emitInfo->EmitHighLevelInstructions) {
1177       /* BGNLOOP's BranchTarget points to the ENDLOOP inst */
1178       beginInst = prog->Instructions + beginInstLoc;
1179       beginInst->BranchTarget = prog->NumInstructions - 1;
1180    }
1181
1182    /* Done emitting loop code.  Now walk over the loop's linked list of
1183     * BREAK and CONT nodes, filling in their BranchTarget fields (which
1184     * will point to the ENDLOOP+1 or BGNLOOP instructions, respectively).
1185     */
1186    for (ir = n->BranchNode; ir; ir = ir->BranchNode) {
1187       struct prog_instruction *inst = prog->Instructions + ir->InstLocation;
1188       assert(inst->BranchTarget < 0);
1189       if (ir->Opcode == IR_BREAK ||
1190           ir->Opcode == IR_BREAK_IF_FALSE ||
1191           ir->Opcode == IR_BREAK_IF_TRUE) {
1192          assert(inst->Opcode == OPCODE_BRK ||
1193                 inst->Opcode == OPCODE_BRA);
1194          /* go to instruction after end of loop */
1195          inst->BranchTarget = endInstLoc + 1;
1196       }
1197       else {
1198          assert(ir->Opcode == IR_CONT ||
1199                 ir->Opcode == IR_CONT_IF_FALSE ||
1200                 ir->Opcode == IR_CONT_IF_TRUE);
1201          assert(inst->Opcode == OPCODE_CONT ||
1202                 inst->Opcode == OPCODE_BRA);
1203          /* to go instruction at top of loop */
1204          inst->BranchTarget = beginInstLoc;
1205       }
1206    }
1207    return NULL;
1208 }
1209
1210
1211 /**
1212  * "Continue" or "break" statement.
1213  * Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted.
1214  */
1215 static struct prog_instruction *
1216 emit_cont_break(slang_emit_info *emitInfo, slang_ir_node *n)
1217 {
1218    gl_inst_opcode opcode;
1219    struct prog_instruction *inst;
1220    n->InstLocation = emitInfo->prog->NumInstructions;
1221    if (emitInfo->EmitHighLevelInstructions) {
1222       opcode = (n->Opcode == IR_CONT) ? OPCODE_CONT : OPCODE_BRK;
1223    }
1224    else {
1225       opcode = OPCODE_BRA;
1226    }
1227    inst = new_instruction(emitInfo, opcode);
1228    inst->DstReg.CondMask = COND_TR;  /* always true */
1229    return inst;
1230 }
1231
1232
1233 /**
1234  * Conditional "continue" or "break" statement.
1235  * Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted.
1236  */
1237 static struct prog_instruction *
1238 emit_cont_break_if(slang_emit_info *emitInfo, slang_ir_node *n,
1239                    GLboolean breakTrue)
1240 {
1241    gl_inst_opcode opcode;
1242    struct prog_instruction *inst;
1243
1244    /* evaluate condition expr, setting cond codes */
1245    inst = emit(emitInfo, n->Children[0]);
1246    assert(inst);
1247    inst->CondUpdate = GL_TRUE;
1248
1249    n->InstLocation = emitInfo->prog->NumInstructions;
1250    if (emitInfo->EmitHighLevelInstructions) {
1251       if (n->Opcode == IR_CONT_IF_TRUE ||
1252           n->Opcode == IR_CONT_IF_FALSE)
1253          opcode = OPCODE_CONT;
1254       else
1255          opcode = OPCODE_BRK;
1256    }
1257    else {
1258       opcode = OPCODE_BRA;
1259    }
1260    inst = new_instruction(emitInfo, opcode);
1261    inst->DstReg.CondMask = breakTrue ? COND_NE : COND_EQ;
1262    return inst;
1263 }
1264
1265
1266
1267 /**
1268  * Remove any SWIZZLE_NIL terms from given swizzle mask (smear prev term).
1269  * Ex: fix_swizzle("zyNN") -> "zyyy"
1270  */
1271 static GLuint
1272 fix_swizzle(GLuint swizzle)
1273 {
1274    GLuint swz[4], i;
1275    for (i = 0; i < 4; i++) {
1276       swz[i] = GET_SWZ(swizzle, i);
1277       if (swz[i] == SWIZZLE_NIL) {
1278          swz[i] = swz[i - 1];
1279       }
1280    }
1281    return MAKE_SWIZZLE4(swz[0], swz[1], swz[2], swz[3]);
1282 }
1283
1284
1285 static struct prog_instruction *
1286 emit_swizzle(slang_emit_info *emitInfo, slang_ir_node *n)
1287 {
1288    GLuint swizzle;
1289
1290    /* swizzled storage access */
1291    (void) emit(emitInfo, n->Children[0]);
1292
1293    /* "pull-up" the child's storage info, applying our swizzle info */
1294    n->Store->File  = n->Children[0]->Store->File;
1295    n->Store->Index = n->Children[0]->Store->Index;
1296    n->Store->Size  = n->Children[0]->Store->Size;
1297    /*n->Var = n->Children[0]->Var; XXX for debug */
1298    assert(n->Store->Index >= 0);
1299
1300    swizzle = fix_swizzle(n->Store->Swizzle);
1301 #ifdef DEBUG
1302    {
1303       GLuint s = n->Children[0]->Store->Swizzle;
1304       assert(GET_SWZ(s, 0) != SWIZZLE_NIL);
1305       assert(GET_SWZ(s, 1) != SWIZZLE_NIL);
1306       assert(GET_SWZ(s, 2) != SWIZZLE_NIL);
1307       assert(GET_SWZ(s, 3) != SWIZZLE_NIL);
1308    }
1309 #endif
1310
1311    /* apply this swizzle to child's swizzle to get composed swizzle */
1312    n->Store->Swizzle = swizzle_swizzle(n->Children[0]->Store->Swizzle,
1313                                        swizzle);
1314    return NULL;
1315 }
1316
1317
1318 /**
1319  * Dereference array element.  Just resolve storage for the array
1320  * element represented by this node.
1321  */
1322 static struct prog_instruction *
1323 emit_array_element(slang_emit_info *emitInfo, slang_ir_node *n)
1324 {
1325    assert(n->Store);
1326    assert(n->Store->File != PROGRAM_UNDEFINED);
1327    assert(n->Store->Size > 0);
1328
1329    if (n->Store->File == PROGRAM_STATE_VAR) {
1330       n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
1331       return NULL;
1332    }
1333
1334
1335    if (n->Children[1]->Opcode == IR_FLOAT) {
1336       /* Constant index */
1337       const GLint arrayAddr = n->Children[0]->Store->Index;
1338       const GLint index = (GLint) n->Children[1]->Value[0];
1339       n->Store->Index = arrayAddr + index;
1340    }
1341    else {
1342       /* Variable index - PROBLEM */
1343       const GLint arrayAddr = n->Children[0]->Store->Index;
1344       const GLint index = 0;
1345       _mesa_problem(NULL, "variable array indexes not supported yet!");
1346       n->Store->Index = arrayAddr + index;
1347    }
1348    return NULL; /* no instruction */
1349 }
1350
1351
1352 /**
1353  * Resolve storage for accessing a structure field.
1354  */
1355 static struct prog_instruction *
1356 emit_struct_field(slang_emit_info *emitInfo, slang_ir_node *n)
1357 {
1358    if (n->Store->File == PROGRAM_STATE_VAR) {
1359       n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
1360    }
1361    else {
1362       _mesa_problem(NULL, "structs/fields not supported yet");
1363    }
1364    return NULL; /* no instruction */
1365 }
1366
1367
1368 static struct prog_instruction *
1369 emit(slang_emit_info *emitInfo, slang_ir_node *n)
1370 {
1371    struct prog_instruction *inst;
1372    if (!n)
1373       return NULL;
1374
1375    switch (n->Opcode) {
1376    case IR_SEQ:
1377       /* sequence of two sub-trees */
1378       assert(n->Children[0]);
1379       assert(n->Children[1]);
1380       emit(emitInfo, n->Children[0]);
1381       inst = emit(emitInfo, n->Children[1]);
1382       assert(!n->Store);
1383       n->Store = n->Children[1]->Store;
1384       return inst;
1385
1386    case IR_SCOPE:
1387       /* new variable scope */
1388       _slang_push_var_table(emitInfo->vt);
1389       inst = emit(emitInfo, n->Children[0]);
1390       _slang_pop_var_table(emitInfo->vt);
1391       return inst;
1392
1393    case IR_VAR_DECL:
1394       /* Variable declaration - allocate a register for it */
1395       assert(n->Store);
1396       assert(n->Store->File != PROGRAM_UNDEFINED);
1397       assert(n->Store->Size > 0);
1398       assert(n->Store->Index < 0);
1399       if (!n->Var || n->Var->isTemp) {
1400          /* a nameless/temporary variable, will be freed after first use */
1401          if (!_slang_alloc_temp(emitInfo->vt, n->Store)) {
1402             slang_info_log_error(emitInfo->log,
1403                                  "Ran out of registers, too many temporaries");
1404             return NULL;
1405          }
1406       }
1407       else {
1408          /* a regular variable */
1409          _slang_add_variable(emitInfo->vt, n->Var);
1410          if (!_slang_alloc_var(emitInfo->vt, n->Store)) {
1411             slang_info_log_error(emitInfo->log,
1412                                  "Ran out of registers, too many variables");
1413             return NULL;
1414          }
1415          /*
1416          printf("IR_VAR_DECL %s %d store %p\n",
1417                 (char*) n->Var->a_name, n->Store->Index, (void*) n->Store);
1418          */
1419          assert(n->Var->aux == n->Store);
1420       }
1421       if (emitInfo->EmitComments) {
1422          /* emit NOP with comment describing the variable's storage location */
1423          char s[1000];
1424          sprintf(s, "TEMP[%d]%s = %s (size %d)",
1425                  n->Store->Index,
1426                  _mesa_swizzle_string(n->Store->Swizzle, 0, GL_FALSE),
1427                  (char *) n->Var->a_name,
1428                  n->Store->Size);
1429          inst = new_instruction(emitInfo, OPCODE_NOP);
1430          inst->Comment = _mesa_strdup(s);
1431          return inst;
1432       }
1433       return NULL;
1434
1435    case IR_VAR:
1436       /* Reference to a variable
1437        * Storage should have already been resolved/allocated.
1438        */
1439       assert(n->Store);
1440       assert(n->Store->File != PROGRAM_UNDEFINED);
1441
1442       if (n->Store->File == PROGRAM_STATE_VAR &&
1443           n->Store->Index < 0) {
1444          n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
1445       }
1446
1447       if (n->Store->Index < 0) {
1448          printf("#### VAR %s not allocated!\n", (char*)n->Var->a_name);
1449       }
1450       assert(n->Store->Index >= 0);
1451       assert(n->Store->Size > 0);
1452       break;
1453
1454    case IR_ELEMENT:
1455       return emit_array_element(emitInfo, n);
1456    case IR_FIELD:
1457       return emit_struct_field(emitInfo, n);
1458    case IR_SWIZZLE:
1459       return emit_swizzle(emitInfo, n);
1460
1461    case IR_I_TO_F:
1462       /* just move */
1463       emit(emitInfo, n->Children[0]);
1464       inst = new_instruction(emitInfo, OPCODE_MOV);
1465       if (!n->Store) {
1466          if (!alloc_temp_storage(emitInfo, n, 1))
1467             return NULL;
1468       }
1469       storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
1470       storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
1471       if (emitInfo->EmitComments)
1472          inst->Comment = _mesa_strdup("int to float");
1473       return NULL;
1474
1475    /* Simple arithmetic */
1476    /* unary */
1477    case IR_RSQ:
1478    case IR_RCP:
1479    case IR_FLOOR:
1480    case IR_FRAC:
1481    case IR_F_TO_I:
1482    case IR_ABS:
1483    case IR_SIN:
1484    case IR_COS:
1485    case IR_DDX:
1486    case IR_DDY:
1487    case IR_NOISE1:
1488    case IR_NOISE2:
1489    case IR_NOISE3:
1490    case IR_NOISE4:
1491    /* binary */
1492    case IR_ADD:
1493    case IR_SUB:
1494    case IR_MUL:
1495    case IR_DOT4:
1496    case IR_DOT3:
1497    case IR_CROSS:
1498    case IR_MIN:
1499    case IR_MAX:
1500    case IR_SEQUAL:
1501    case IR_SNEQUAL:
1502    case IR_SGE:
1503    case IR_SGT:
1504    case IR_SLE:
1505    case IR_SLT:
1506    case IR_POW:
1507    case IR_EXP:
1508    case IR_EXP2:
1509    /* trinary operators */
1510    case IR_LRP:
1511       return emit_arith(emitInfo, n);
1512    case IR_CLAMP:
1513       return emit_clamp(emitInfo, n);
1514    case IR_TEX:
1515    case IR_TEXB:
1516    case IR_TEXP:
1517       return emit_tex(emitInfo, n);
1518    case IR_NEG:
1519       return emit_negation(emitInfo, n);
1520    case IR_FLOAT:
1521       /* find storage location for this float constant */
1522       n->Store->Index = _mesa_add_unnamed_constant(emitInfo->prog->Parameters, n->Value,
1523                                                    n->Store->Size,
1524                                                    &n->Store->Swizzle);
1525       if (n->Store->Index < 0) {
1526          slang_info_log_error(emitInfo->log, "Ran out of space for constants");
1527          return NULL;
1528       }
1529       return NULL;
1530
1531    case IR_MOVE:
1532       return emit_move(emitInfo, n);
1533
1534    case IR_COND:
1535       return emit_cond(emitInfo, n);
1536
1537    case IR_NOT:
1538       return emit_not(emitInfo, n);
1539
1540    case IR_LABEL:
1541       return emit_label(emitInfo, n);
1542    case IR_JUMP:
1543       assert(n);
1544       assert(n->Label);
1545       return emit_jump(emitInfo, n);
1546    case IR_KILL:
1547       return emit_kill(emitInfo);
1548
1549    case IR_IF:
1550       return emit_if(emitInfo, n);
1551
1552    case IR_LOOP:
1553       return emit_loop(emitInfo, n);
1554    case IR_BREAK_IF_FALSE:
1555    case IR_CONT_IF_FALSE:
1556       return emit_cont_break_if(emitInfo, n, GL_FALSE);
1557    case IR_BREAK_IF_TRUE:
1558    case IR_CONT_IF_TRUE:
1559       return emit_cont_break_if(emitInfo, n, GL_TRUE);
1560    case IR_BREAK:
1561       /* fall-through */
1562    case IR_CONT:
1563       return emit_cont_break(emitInfo, n);
1564
1565    case IR_BEGIN_SUB:
1566       return new_instruction(emitInfo, OPCODE_BGNSUB);
1567    case IR_END_SUB:
1568       return new_instruction(emitInfo, OPCODE_ENDSUB);
1569    case IR_RETURN:
1570       return new_instruction(emitInfo, OPCODE_RET);
1571
1572    case IR_NOP:
1573       return NULL;
1574
1575    default:
1576       _mesa_problem(NULL, "Unexpected IR opcode in emit()\n");
1577       abort();
1578    }
1579    return NULL;
1580 }
1581
1582
1583 GLboolean
1584 _slang_emit_code(slang_ir_node *n, slang_var_table *vt,
1585                  struct gl_program *prog, GLboolean withEnd,
1586                  slang_info_log *log)
1587 {
1588    GET_CURRENT_CONTEXT(ctx);
1589    GLboolean success;
1590    slang_emit_info emitInfo;
1591
1592    emitInfo.log = log;
1593    emitInfo.vt = vt;
1594    emitInfo.prog = prog;
1595
1596    emitInfo.EmitHighLevelInstructions = ctx->Shader.EmitHighLevelInstructions;
1597    emitInfo.EmitComments = ctx->Shader.EmitComments;
1598
1599    (void) emit(&emitInfo, n);
1600
1601    /* finish up by adding the END opcode to program */
1602    if (withEnd) {
1603       struct prog_instruction *inst;
1604       inst = new_instruction(&emitInfo, OPCODE_END);
1605    }
1606    success = GL_TRUE;
1607
1608    printf("*********** End generate code (%u inst):\n", prog->NumInstructions);
1609 #if 0
1610    _mesa_print_program(prog);
1611    _mesa_print_program_parameters(ctx,prog);
1612 #endif
1613
1614    return success;
1615 }