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_POW, "IR_POW", OPCODE_POW, 1, 2 },
  99    /* unary ops */
 100    { IR_I_TO_F, "IR_I_TO_F", OPCODE_NOP, 1, 1 },
 101    { IR_F_TO_I, "IR_F_TO_I", OPCODE_INT, 4, 1 }, /* 4 floats to 4 ints */
 102    { IR_EXP, "IR_EXP", OPCODE_EXP, 1, 1 },
 103    { IR_EXP2, "IR_EXP2", OPCODE_EX2, 1, 1 },
 104    { IR_LOG2, "IR_LOG2", OPCODE_LG2, 1, 1 },
 105    { IR_RSQ, "IR_RSQ", OPCODE_RSQ, 1, 1 },
 106    { IR_RCP, "IR_RCP", OPCODE_RCP, 1, 1 },
 107    { IR_FLOOR, "IR_FLOOR", OPCODE_FLR, 4, 1 },
 108    { IR_FRAC, "IR_FRAC", OPCODE_FRC, 4, 1 },
 109    { IR_ABS, "IR_ABS", OPCODE_ABS, 4, 1 },
 110    { IR_NEG, "IR_NEG", OPCODE_NOP, 4, 1 }, /* special case: emit_negation() */
 111    { IR_DDX, "IR_DDX", OPCODE_DDX, 4, 1 },
 112    { IR_DDX, "IR_DDY", OPCODE_DDX, 4, 1 },
 113    { IR_SIN, "IR_SIN", OPCODE_SIN, 1, 1 },
 114    { IR_COS, "IR_COS", OPCODE_COS, 1, 1 },
 115    { IR_NOISE1, "IR_NOISE1", OPCODE_NOISE1, 1, 1 },
 116    { IR_NOISE2, "IR_NOISE2", OPCODE_NOISE2, 1, 1 },
 117    { IR_NOISE3, "IR_NOISE3", OPCODE_NOISE3, 1, 1 },
 118    { IR_NOISE4, "IR_NOISE4", OPCODE_NOISE4, 1, 1 },
 119
 120    /* other */
 121    { IR_SEQ, "IR_SEQ", OPCODE_NOP, 0, 0 },
 122    { IR_SCOPE, "IR_SCOPE", OPCODE_NOP, 0, 0 },
 123    { IR_LABEL, "IR_LABEL", OPCODE_NOP, 0, 0 },
 124    { IR_JUMP, "IR_JUMP", OPCODE_NOP, 0, 0 },
 125    { IR_CJUMP0, "IR_CJUMP0", OPCODE_NOP, 0, 0 },
 126    { IR_CJUMP1, "IR_CJUMP1", 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    case IR_CJUMP0:
 307       printf("CJUMP0 %s\n", n->Label->Name);
 308       slang_print_ir(n->Children[0], indent+3);
 309       break;
 310    case IR_CJUMP1:
 311       printf("CJUMP1 %s\n", n->Label->Name);
 312       slang_print_ir(n->Children[0], indent+3);
 313       break;
 314
 315    case IR_IF:
 316       printf("IF \n");
 317       slang_print_ir(n->Children[0], indent+3);
 318       spaces(indent);
 319       printf("THEN\n");
 320       slang_print_ir(n->Children[1], indent+3);
 321       if (n->Children[2]) {
 322          spaces(indent);
 323          printf("ELSE\n");
 324          slang_print_ir(n->Children[2], indent+3);
 325       }
 326       printf("ENDIF\n");
 327       break;
 328
 329    case IR_BEGIN_SUB:
 330       printf("BEGIN_SUB\n");
 331       break;
 332    case IR_END_SUB:
 333       printf("END_SUB\n");
 334       break;
 335    case IR_RETURN:
 336       printf("RETURN\n");
 337       break;
 338    case IR_CALL:
 339       printf("CALL\n");
 340       break;
 341
 342    case IR_LOOP:
 343       printf("LOOP\n");
 344       slang_print_ir(n->Children[0], indent+3);
 345       spaces(indent);
 346       printf("ENDLOOP\n");
 347       break;
 348    case IR_CONT:
 349       printf("CONT\n");
 350       break;
 351    case IR_BREAK:
 352       printf("BREAK\n");
 353       break;
 354    case IR_BREAK_IF_FALSE:
 355       printf("BREAK_IF_FALSE\n");
 356       slang_print_ir(n->Children[0], indent+3);
 357       break;
 358    case IR_BREAK_IF_TRUE:
 359       printf("BREAK_IF_TRUE\n");
 360       slang_print_ir(n->Children[0], indent+3);
 361       break;
 362    case IR_CONT_IF_FALSE:
 363       printf("CONT_IF_FALSE\n");
 364       slang_print_ir(n->Children[0], indent+3);
 365       break;
 366    case IR_CONT_IF_TRUE:
 367       printf("CONT_IF_TRUE\n");
 368       slang_print_ir(n->Children[0], indent+3);
 369       break;
 370
 371    case IR_VAR:
 372       printf("VAR %s%s at %s  store %p\n",
 373              (n->Var ? (char *) n->Var->a_name : "TEMP"),
 374              swizzle_string(n->Store->Swizzle),
 375              storage_string(n->Store), (void*) n->Store);
 376       break;
 377    case IR_VAR_DECL:
 378       printf("VAR_DECL %s (%p) at %s  store %p\n",
 379              (n->Var ? (char *) n->Var->a_name : "TEMP"),
 380              (void*) n->Var, storage_string(n->Store),
 381              (void*) n->Store);
 382       break;
 383    case IR_FIELD:
 384       printf("FIELD %s of\n", n->Field);
 385       slang_print_ir(n->Children[0], indent+3);
 386       break;
 387    case IR_FLOAT:
 388       printf("FLOAT %g %g %g %g\n",
 389              n->Value[0], n->Value[1], n->Value[2], n->Value[3]);
 390       break;
 391    case IR_I_TO_F:
 392       printf("INT_TO_FLOAT\n");
 393       slang_print_ir(n->Children[0], indent+3);
 394       break;
 395    case IR_F_TO_I:
 396       printf("FLOAT_TO_INT\n");
 397       slang_print_ir(n->Children[0], indent+3);
 398       break;
 399    case IR_SWIZZLE:
 400       printf("SWIZZLE %s of  (store %p) \n",
 401              swizzle_string(n->Store->Swizzle), (void*) n->Store);
 402       slang_print_ir(n->Children[0], indent + 3);
 403       break;
 404    default:
 405       printf("%s (%p, %p)  (store %p)\n", slang_ir_name(n->Opcode),
 406              (void*) n->Children[0], (void*) n->Children[1], (void*) n->Store);
 407       slang_print_ir(n->Children[0], indent+3);
 408       slang_print_ir(n->Children[1], indent+3);
 409    }
 410 }
 411
 412
 413 /**
 414  * Allocate temporary storage for an intermediate result (such as for
 415  * a multiply or add, etc.
 416  */
 417 static GLboolean
 418 alloc_temp_storage(slang_emit_info *emitInfo, slang_ir_node *n, GLint size)
 419 {
 420    assert(!n->Var);
 421    assert(!n->Store);
 422    assert(size > 0);
 423    n->Store = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, size);
 424    if (!_slang_alloc_temp(emitInfo->vt, n->Store)) {
 425       slang_info_log_error(emitInfo->log,
 426                            "Ran out of registers, too many temporaries");
 427       return GL_FALSE;
 428    }
 429    return GL_TRUE;
 430 }
 431
 432
 433 /**
 434  * Free temporary storage, if n->Store is, in fact, temp storage.
 435  * Otherwise, no-op.
 436  */
 437 static void
 438 free_temp_storage(slang_var_table *vt, slang_ir_node *n)
 439 {
 440    if (n->Store->File == PROGRAM_TEMPORARY && n->Store->Index >= 0) {
 441       if (_slang_is_temp(vt, n->Store)) {
 442          _slang_free_temp(vt, n->Store);
 443          n->Store->Index = -1;
 444          n->Store->Size = -1;
 445       }
 446    }
 447 }
 448
 449
 450 /**
 451  * Convert IR storage to an instruction dst register.
 452  */
 453 static void
 454 storage_to_dst_reg(struct prog_dst_register *dst, const slang_ir_storage *st,
 455                    GLuint writemask)
 456 {
 457    static const GLuint defaultWritemask[4] = {
 458       WRITEMASK_X,
 459       WRITEMASK_X | WRITEMASK_Y,
 460       WRITEMASK_X | WRITEMASK_Y | WRITEMASK_Z,
 461       WRITEMASK_X | WRITEMASK_Y | WRITEMASK_Z | WRITEMASK_W
 462    };
 463    assert(st->Index >= 0 && st->Index <= 16);
 464    dst->File = st->File;
 465    dst->Index = st->Index;
 466    assert(st->File != PROGRAM_UNDEFINED);
 467    assert(st->Size >= 1);
 468    assert(st->Size <= 4);
 469    if (st->Size == 1) {
 470       GLuint comp = GET_SWZ(st->Swizzle, 0);
 471       assert(comp < 4);
 472       assert(writemask & WRITEMASK_X);
 473       dst->WriteMask = WRITEMASK_X << comp;
 474    }
 475    else {
 476       dst->WriteMask = defaultWritemask[st->Size - 1] & writemask;
 477    }
 478 }
 479
 480
 481 /**
 482  * Convert IR storage to an instruction src register.
 483  */
 484 static void
 485 storage_to_src_reg(struct prog_src_register *src, const slang_ir_storage *st)
 486 {
 487    static const GLuint defaultSwizzle[4] = {
 488       MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_X, SWIZZLE_X, SWIZZLE_X),
 489       MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W),
 490       MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W),
 491       MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W)
 492    };
 493    assert(st->File >= 0 && st->File <= 16);
 494    src->File = st->File;
 495    src->Index = st->Index;
 496    assert(st->File != PROGRAM_UNDEFINED);
 497    assert(st->Size >= 1);
 498    assert(st->Size <= 4);
 499    if (st->Swizzle != SWIZZLE_NOOP)
 500       src->Swizzle = st->Swizzle;
 501    else
 502       src->Swizzle = defaultSwizzle[st->Size - 1]; /*XXX really need this?*/
 503
 504    assert(GET_SWZ(src->Swizzle, 0) != SWIZZLE_NIL);
 505    assert(GET_SWZ(src->Swizzle, 1) != SWIZZLE_NIL);
 506    assert(GET_SWZ(src->Swizzle, 2) != SWIZZLE_NIL);
 507    assert(GET_SWZ(src->Swizzle, 3) != SWIZZLE_NIL);
 508 }
 509
 510
 511
 512 /**
 513  * Add new instruction at end of given program.
 514  * \param prog  the program to append instruction onto
 515  * \param opcode  opcode for the new instruction
 516  * \return pointer to the new instruction
 517  */
 518 static struct prog_instruction *
 519 new_instruction(slang_emit_info *emitInfo, gl_inst_opcode opcode)
 520 {
 521    struct gl_program *prog = emitInfo->prog;
 522    struct prog_instruction *inst;
 523    prog->Instructions = _mesa_realloc_instructions(prog->Instructions,
 524                                                    prog->NumInstructions,
 525                                                    prog->NumInstructions + 1);
 526    inst = prog->Instructions + prog->NumInstructions;
 527    prog->NumInstructions++;
 528    _mesa_init_instructions(inst, 1);
 529    inst->Opcode = opcode;
 530    inst->BranchTarget = -1; /* invalid */
 531    return inst;
 532 }
 533
 534
 535 #if 0
 536 /**
 537  * Return pointer to last instruction in program.
 538  */
 539 static struct prog_instruction *
 540 prev_instruction(struct gl_program *prog)
 541 {
 542    if (prog->NumInstructions == 0)
 543       return NULL;
 544    else
 545       return prog->Instructions + prog->NumInstructions - 1;
 546 }
 547 #endif
 548
 549
 550 static struct prog_instruction *
 551 emit(slang_emit_info *emitInfo, slang_ir_node *n);
 552
 553
 554 /**
 555  * Return an annotation string for given node's storage.
 556  */
 557 static char *
 558 storage_annotation(const slang_ir_node *n, const struct gl_program *prog)
 559 {
 560 #if ANNOTATE
 561    const slang_ir_storage *st = n->Store;
 562    static char s[100] = "";
 563
 564    if (!st)
 565       return _mesa_strdup("");
 566
 567    switch (st->File) {
 568    case PROGRAM_CONSTANT:
 569       if (st->Index >= 0) {
 570          const GLfloat *val = prog->Parameters->ParameterValues[st->Index];
 571          if (st->Swizzle == SWIZZLE_NOOP)
 572             sprintf(s, "{%g, %g, %g, %g}", val[0], val[1], val[2], val[3]);
 573          else {
 574             sprintf(s, "%g", val[GET_SWZ(st->Swizzle, 0)]);
 575          }
 576       }
 577       break;
 578    case PROGRAM_TEMPORARY:
 579       if (n->Var)
 580          sprintf(s, "%s", (char *) n->Var->a_name);
 581       else
 582          sprintf(s, "t[%d]", st->Index);
 583       break;
 584    case PROGRAM_STATE_VAR:
 585    case PROGRAM_UNIFORM:
 586       sprintf(s, "%s", prog->Parameters->Parameters[st->Index].Name);
 587       break;
 588    case PROGRAM_VARYING:
 589       sprintf(s, "%s", prog->Varying->Parameters[st->Index].Name);
 590       break;
 591    case PROGRAM_INPUT:
 592       sprintf(s, "input[%d]", st->Index);
 593       break;
 594    case PROGRAM_OUTPUT:
 595       sprintf(s, "output[%d]", st->Index);
 596       break;
 597    default:
 598       s[0] = 0;
 599    }
 600    return _mesa_strdup(s);
 601 #else
 602    return NULL;
 603 #endif
 604 }
 605
 606
 607 /**
 608  * Return an annotation string for an instruction.
 609  */
 610 static char *
 611 instruction_annotation(gl_inst_opcode opcode, char *dstAnnot,
 612                        char *srcAnnot0, char *srcAnnot1, char *srcAnnot2)
 613 {
 614 #if ANNOTATE
 615    const char *operator;
 616    char *s;
 617    int len = 50;
 618
 619    if (dstAnnot)
 620       len += strlen(dstAnnot);
 621    else
 622       dstAnnot = _mesa_strdup("");
 623
 624    if (srcAnnot0)
 625       len += strlen(srcAnnot0);
 626    else
 627       srcAnnot0 = _mesa_strdup("");
 628
 629    if (srcAnnot1)
 630       len += strlen(srcAnnot1);
 631    else
 632       srcAnnot1 = _mesa_strdup("");
 633
 634    if (srcAnnot2)
 635       len += strlen(srcAnnot2);
 636    else
 637       srcAnnot2 = _mesa_strdup("");
 638
 639    switch (opcode) {
 640    case OPCODE_ADD:
 641       operator = "+";
 642       break;
 643    case OPCODE_SUB:
 644       operator = "-";
 645       break;
 646    case OPCODE_MUL:
 647       operator = "*";
 648       break;
 649    case OPCODE_DP3:
 650       operator = "DP3";
 651       break;
 652    case OPCODE_DP4:
 653       operator = "DP4";
 654       break;
 655    case OPCODE_XPD:
 656       operator = "XPD";
 657       break;
 658    case OPCODE_RSQ:
 659       operator = "RSQ";
 660       break;
 661    case OPCODE_SGT:
 662       operator = ">";
 663       break;
 664    default:
 665       operator = ",";
 666    }
 667
 668    s = (char *) malloc(len);
 669    sprintf(s, "%s = %s %s %s %s", dstAnnot,
 670            srcAnnot0, operator, srcAnnot1, srcAnnot2);
 671    assert(_mesa_strlen(s) < len);
 672
 673    free(dstAnnot);
 674    free(srcAnnot0);
 675    free(srcAnnot1);
 676    free(srcAnnot2);
 677
 678    return s;
 679 #else
 680    return NULL;
 681 #endif
 682 }
 683
 684
 685
 686 /**
 687  * Generate code for a simple arithmetic instruction.
 688  * Either 1, 2 or 3 operands.
 689  */
 690 static struct prog_instruction *
 691 emit_arith(slang_emit_info *emitInfo, slang_ir_node *n)
 692 {
 693    struct prog_instruction *inst;
 694    const slang_ir_info *info = slang_find_ir_info(n->Opcode);
 695    char *srcAnnot[3], *dstAnnot;
 696    GLuint i;
 697
 698    assert(info);
 699    assert(info->InstOpcode != OPCODE_NOP);
 700
 701    srcAnnot[0] = srcAnnot[1] = srcAnnot[2] = dstAnnot = NULL;
 702
 703 #if PEEPHOLE_OPTIMIZATIONS
 704    /* Look for MAD opportunity */
 705    if (info->NumParams == 2 &&
 706        n->Opcode == IR_ADD && n->Children[0]->Opcode == IR_MUL) {
 707       /* found pattern IR_ADD(IR_MUL(A, B), C) */
 708       emit(emitInfo, n->Children[0]->Children[0]);  /* A */
 709       emit(emitInfo, n->Children[0]->Children[1]);  /* B */
 710       emit(emitInfo, n->Children[1]);  /* C */
 711       /* generate MAD instruction */
 712       inst = new_instruction(emitInfo, OPCODE_MAD);
 713       /* operands: A, B, C: */
 714       storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Children[0]->Store);
 715       storage_to_src_reg(&inst->SrcReg[1], n->Children[0]->Children[1]->Store);
 716       storage_to_src_reg(&inst->SrcReg[2], n->Children[1]->Store);
 717       free_temp_storage(emitInfo->vt, n->Children[0]->Children[0]);
 718       free_temp_storage(emitInfo->vt, n->Children[0]->Children[1]);
 719       free_temp_storage(emitInfo->vt, n->Children[1]);
 720    }
 721    else if (info->NumParams == 2 &&
 722             n->Opcode == IR_ADD && n->Children[1]->Opcode == IR_MUL) {
 723       /* found pattern IR_ADD(A, IR_MUL(B, C)) */
 724       emit(emitInfo, n->Children[0]);  /* A */
 725       emit(emitInfo, n->Children[1]->Children[0]);  /* B */
 726       emit(emitInfo, n->Children[1]->Children[1]);  /* C */
 727       /* generate MAD instruction */
 728       inst = new_instruction(emitInfo, OPCODE_MAD);
 729       /* operands: B, C, A */
 730       storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Children[0]->Store);
 731       storage_to_src_reg(&inst->SrcReg[1], n->Children[1]->Children[1]->Store);
 732       storage_to_src_reg(&inst->SrcReg[2], n->Children[0]->Store);
 733       free_temp_storage(emitInfo->vt, n->Children[1]->Children[0]);
 734       free_temp_storage(emitInfo->vt, n->Children[1]->Children[1]);
 735       free_temp_storage(emitInfo->vt, n->Children[0]);
 736    }
 737    else
 738 #endif
 739    {
 740       /* normal case */
 741
 742       /* gen code for children */
 743       for (i = 0; i < info->NumParams; i++)
 744          emit(emitInfo, n->Children[i]);
 745
 746       /* gen this instruction and src registers */
 747       inst = new_instruction(emitInfo, info->InstOpcode);
 748       for (i = 0; i < info->NumParams; i++)
 749          storage_to_src_reg(&inst->SrcReg[i], n->Children[i]->Store);
 750
 751       /* annotation */
 752       for (i = 0; i < info->NumParams; i++)
 753          srcAnnot[i] = storage_annotation(n->Children[i], emitInfo->prog);
 754
 755       /* free temps */
 756       for (i = 0; i < info->NumParams; i++)
 757          free_temp_storage(emitInfo->vt, n->Children[i]);
 758    }
 759
 760    /* result storage */
 761    if (!n->Store) {
 762       if (!alloc_temp_storage(emitInfo, n, info->ResultSize))
 763          return NULL;
 764    }
 765    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 766
 767    dstAnnot = storage_annotation(n, emitInfo->prog);
 768
 769    inst->Comment = instruction_annotation(inst->Opcode, dstAnnot, srcAnnot[0],
 770                                           srcAnnot[1], srcAnnot[2]);
 771
 772    /*_mesa_print_instruction(inst);*/
 773    return inst;
 774 }
 775
 776
 777 /**
 778  * Generate code for an IR_CLAMP instruction.
 779  */
 780 static struct prog_instruction *
 781 emit_clamp(slang_emit_info *emitInfo, slang_ir_node *n)
 782 {
 783    struct prog_instruction *inst;
 784
 785    assert(n->Opcode == IR_CLAMP);
 786    /* ch[0] = value
 787     * ch[1] = min limit
 788     * ch[2] = max limit
 789     */
 790
 791    inst = emit(emitInfo, n->Children[0]);
 792
 793    /* If lower limit == 0.0 and upper limit == 1.0,
 794     *    set prev instruction's SaturateMode field to SATURATE_ZERO_ONE.
 795     * Else,
 796     *    emit OPCODE_MIN, OPCODE_MAX sequence.
 797     */
 798 #if 0
 799    /* XXX this isn't quite finished yet */
 800    if (n->Children[1]->Opcode == IR_FLOAT &&
 801        n->Children[1]->Value[0] == 0.0 &&
 802        n->Children[1]->Value[1] == 0.0 &&
 803        n->Children[1]->Value[2] == 0.0 &&
 804        n->Children[1]->Value[3] == 0.0 &&
 805        n->Children[2]->Opcode == IR_FLOAT &&
 806        n->Children[2]->Value[0] == 1.0 &&
 807        n->Children[2]->Value[1] == 1.0 &&
 808        n->Children[2]->Value[2] == 1.0 &&
 809        n->Children[2]->Value[3] == 1.0) {
 810       if (!inst) {
 811          inst = prev_instruction(prog);
 812       }
 813       if (inst && inst->Opcode != OPCODE_NOP) {
 814          /* and prev instruction's DstReg matches n->Children[0]->Store */
 815          inst->SaturateMode = SATURATE_ZERO_ONE;
 816          n->Store = n->Children[0]->Store;
 817          return inst;
 818       }
 819    }
 820 #endif
 821
 822    if (!n->Store)
 823       if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size))
 824          return NULL;
 825
 826    emit(emitInfo, n->Children[1]);
 827    emit(emitInfo, n->Children[2]);
 828
 829    /* tmp = max(ch[0], ch[1]) */
 830    inst = new_instruction(emitInfo, OPCODE_MAX);
 831    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 832    storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
 833    storage_to_src_reg(&inst->SrcReg[1], n->Children[1]->Store);
 834
 835    /* tmp = min(tmp, ch[2]) */
 836    inst = new_instruction(emitInfo, OPCODE_MIN);
 837    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 838    storage_to_src_reg(&inst->SrcReg[0], n->Store);
 839    storage_to_src_reg(&inst->SrcReg[1], n->Children[2]->Store);
 840
 841    return inst;
 842 }
 843
 844
 845 static struct prog_instruction *
 846 emit_negation(slang_emit_info *emitInfo, slang_ir_node *n)
 847 {
 848    /* Implement as MOV dst, -src; */
 849    /* XXX we could look at the previous instruction and in some circumstances
 850     * modify it to accomplish the negation.
 851     */
 852    struct prog_instruction *inst;
 853
 854    emit(emitInfo, n->Children[0]);
 855
 856    if (!n->Store)
 857       if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size))
 858          return NULL;
 859
 860    inst = new_instruction(emitInfo, OPCODE_MOV);
 861    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 862    storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
 863    inst->SrcReg[0].NegateBase = NEGATE_XYZW;
 864    return inst;
 865 }
 866
 867
 868 static struct prog_instruction *
 869 emit_label(slang_emit_info *emitInfo, const slang_ir_node *n)
 870 {
 871    assert(n->Label);
 872    assert(_slang_label_get_location(n->Label) < 0);
 873    _slang_label_set_location(n->Label, emitInfo->prog->NumInstructions,
 874                              emitInfo->prog);
 875    return NULL;
 876 }
 877
 878
 879 static struct prog_instruction *
 880 emit_cjump(slang_emit_info *emitInfo, slang_ir_node *n, GLuint zeroOrOne)
 881 {
 882    struct prog_instruction *inst;
 883    assert(n->Opcode == IR_CJUMP0 || n->Opcode == IR_CJUMP1);
 884    inst = new_instruction(emitInfo, OPCODE_BRA);
 885    if (zeroOrOne)
 886       inst->DstReg.CondMask = COND_NE;  /* branch if non-zero */
 887    else
 888       inst->DstReg.CondMask = COND_EQ;  /* branch if equal to zero */
 889    inst->DstReg.CondSwizzle = SWIZZLE_X;
 890    inst->BranchTarget = _slang_label_get_location(n->Label);
 891    if (inst->BranchTarget < 0) {
 892       _slang_label_add_reference(n->Label, emitInfo->prog->NumInstructions - 1);
 893    }
 894    return inst;
 895 }
 896
 897
 898 static struct prog_instruction *
 899 emit_jump(slang_emit_info *emitInfo, slang_ir_node *n)
 900 {
 901    struct prog_instruction *inst;
 902    inst = new_instruction(emitInfo, OPCODE_BRA);
 903    inst->DstReg.CondMask = COND_TR;  /* always branch */
 904    inst->BranchTarget = _slang_label_get_location(n->Label);
 905    if (inst->BranchTarget < 0) {
 906       _slang_label_add_reference(n->Label, emitInfo->prog->NumInstructions - 1);
 907    }
 908    return inst;
 909 }
 910
 911
 912 static struct prog_instruction *
 913 emit_kill(slang_emit_info *emitInfo)
 914 {
 915    struct prog_instruction *inst;
 916    /* NV-KILL - discard fragment depending on condition code.
 917     * Note that ARB-KILL depends on sign of vector operand.
 918     */
 919    inst = new_instruction(emitInfo, OPCODE_KIL_NV);
 920    inst->DstReg.CondMask = COND_TR;  /* always branch */
 921    return inst;
 922 }
 923
 924
 925 static struct prog_instruction *
 926 emit_tex(slang_emit_info *emitInfo, slang_ir_node *n)
 927 {
 928    struct prog_instruction *inst;
 929    if (n->Opcode == IR_TEX) {
 930       inst = new_instruction(emitInfo, OPCODE_TEX);
 931    }
 932    else if (n->Opcode == IR_TEXB) {
 933       inst = new_instruction(emitInfo, OPCODE_TXB);
 934    }
 935    else {
 936       assert(n->Opcode == IR_TEXP);
 937       inst = new_instruction(emitInfo, OPCODE_TXP);
 938    }
 939
 940    if (!n->Store)
 941       if (!alloc_temp_storage(emitInfo, n, 4))
 942          return NULL;
 943
 944    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
 945
 946    (void) emit(emitInfo, n->Children[1]);
 947
 948    /* Child[1] is the coord */
 949    storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Store);
 950
 951    /* Child[0] is the sampler (a uniform which'll indicate the texture unit) */
 952    assert(n->Children[0]->Store);
 953    assert(n->Children[0]->Store->Size >= TEXTURE_1D_INDEX);
 954
 955    inst->Sampler = n->Children[0]->Store->Index; /* i.e. uniform's index */
 956    inst->TexSrcTarget = n->Children[0]->Store->Size;
 957    inst->TexSrcUnit = 27; /* Dummy value; the TexSrcUnit will be computed at
 958                            * link time, using the sampler uniform's value.
 959                            */
 960    return inst;
 961 }
 962
 963
 964 static struct prog_instruction *
 965 emit_move(slang_emit_info *emitInfo, slang_ir_node *n)
 966 {
 967    struct prog_instruction *inst;
 968
 969    /* rhs */
 970    assert(n->Children[1]);
 971    inst = emit(emitInfo, n->Children[1]);
 972
 973    assert(n->Children[1]->Store->Index >= 0);
 974
 975    /* lhs */
 976    emit(emitInfo, n->Children[0]);
 977
 978    assert(!n->Store);
 979    n->Store = n->Children[0]->Store;
 980
 981 #if PEEPHOLE_OPTIMIZATIONS
 982    if (inst && _slang_is_temp(emitInfo->vt, n->Children[1]->Store)) {
 983       /* Peephole optimization:
 984        * Just modify the RHS to put its result into the dest of this
 985        * MOVE operation.  Then, this MOVE is a no-op.
 986        */
 987       _slang_free_temp(emitInfo->vt, n->Children[1]->Store);
 988       *n->Children[1]->Store = *n->Children[0]->Store;
 989       /* fixup the prev (RHS) instruction */
 990       assert(n->Children[0]->Store->Index >= 0);
 991       assert(n->Children[0]->Store->Index < 16);
 992       storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store, n->Writemask);
 993       return inst;
 994    }
 995    else
 996 #endif
 997    {
 998       if (n->Children[0]->Store->Size > 4) {
 999          /* move matrix/struct etc (block of registers) */
1000          slang_ir_storage dstStore = *n->Children[0]->Store;
1001          slang_ir_storage srcStore = *n->Children[1]->Store;
1002          GLint size = srcStore.Size;
1003          ASSERT(n->Children[0]->Writemask == WRITEMASK_XYZW);
1004          ASSERT(n->Children[1]->Store->Swizzle == SWIZZLE_NOOP);
1005          dstStore.Size = 4;
1006          srcStore.Size = 4;
1007          while (size >= 4) {
1008             inst = new_instruction(emitInfo, OPCODE_MOV);
1009             inst->Comment = _mesa_strdup("IR_MOVE block");
1010             storage_to_dst_reg(&inst->DstReg, &dstStore, n->Writemask);
1011             storage_to_src_reg(&inst->SrcReg[0], &srcStore);
1012             srcStore.Index++;
1013             dstStore.Index++;
1014             size -= 4;
1015          }
1016       }
1017       else {
1018          /* single register move */
1019          char *srcAnnot, *dstAnnot;
1020          inst = new_instruction(emitInfo, OPCODE_MOV);
1021          assert(n->Children[0]->Store->Index >= 0);
1022          assert(n->Children[0]->Store->Index < 16);
1023          storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store, n->Writemask);
1024          storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Store);
1025          dstAnnot = storage_annotation(n->Children[0], emitInfo->prog);
1026          srcAnnot = storage_annotation(n->Children[1], emitInfo->prog);
1027          inst->Comment = instruction_annotation(inst->Opcode, dstAnnot,
1028                                                 srcAnnot, NULL, NULL);
1029       }
1030       free_temp_storage(emitInfo->vt, n->Children[1]);
1031       return inst;
1032    }
1033 }
1034
1035
1036 static struct prog_instruction *
1037 emit_cond(slang_emit_info *emitInfo, slang_ir_node *n)
1038 {
1039    /* Conditional expression (in if/while/for stmts).
1040     * Need to update condition code register.
1041     * Next instruction is typically an IR_CJUMP0/1.
1042     */
1043    /* last child expr instruction: */
1044    struct prog_instruction *inst = emit(emitInfo, n->Children[0]);
1045    if (inst) {
1046       /* set inst's CondUpdate flag */
1047       inst->CondUpdate = GL_TRUE;
1048       return inst; /* XXX or null? */
1049    }
1050    else {
1051       /* This'll happen for things like "if (i) ..." where no code
1052        * is normally generated for the expression "i".
1053        * Generate a move instruction just to set condition codes.
1054        * Note: must use full 4-component vector since all four
1055        * condition codes must be set identically.
1056        */
1057       if (!alloc_temp_storage(emitInfo, n, 4))
1058          return NULL;
1059       inst = new_instruction(emitInfo, OPCODE_MOV);
1060       inst->CondUpdate = GL_TRUE;
1061       storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
1062       storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
1063       _slang_free_temp(emitInfo->vt, n->Store);
1064       inst->Comment = _mesa_strdup("COND expr");
1065       return inst; /* XXX or null? */
1066    }
1067 }
1068
1069
1070 /**
1071  * Logical-NOT
1072  */
1073 static struct prog_instruction *
1074 emit_not(slang_emit_info *emitInfo, slang_ir_node *n)
1075 {
1076    GLfloat zero = 0.0;
1077    slang_ir_storage st;
1078    struct prog_instruction *inst;
1079
1080    /* need zero constant */
1081    st.File = PROGRAM_CONSTANT;
1082    st.Size = 1;
1083    st.Index = _mesa_add_unnamed_constant(emitInfo->prog->Parameters, &zero,
1084                                          1, &st.Swizzle);
1085
1086    /* child expr */
1087    (void) emit(emitInfo, n->Children[0]);
1088    /* XXXX if child instr is SGT convert to SLE, if SEQ, SNE, etc */
1089
1090    if (!n->Store)
1091       if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size))
1092          return NULL;
1093
1094    inst = new_instruction(emitInfo, OPCODE_SEQ);
1095    storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
1096    storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
1097    storage_to_src_reg(&inst->SrcReg[1], &st);
1098
1099    free_temp_storage(emitInfo->vt, n->Children[0]);
1100
1101    inst->Comment = _mesa_strdup("NOT");
1102    return inst;
1103 }
1104
1105
1106 static struct prog_instruction *
1107 emit_if(slang_emit_info *emitInfo, slang_ir_node *n)
1108 {
1109    struct gl_program *prog = emitInfo->prog;
1110    struct prog_instruction *ifInst;
1111    GLuint ifInstLoc, elseInstLoc = 0;
1112
1113    emit(emitInfo, n->Children[0]);  /* the condition */
1114    ifInstLoc = prog->NumInstructions;
1115    if (emitInfo->EmitHighLevelInstructions) {
1116       ifInst = new_instruction(emitInfo, OPCODE_IF);
1117       ifInst->DstReg.CondMask = COND_NE;  /* if cond is non-zero */
1118       ifInst->DstReg.CondSwizzle = SWIZZLE_X;
1119    }
1120    else {
1121       /* conditional jump to else, or endif */
1122       ifInst = new_instruction(emitInfo, OPCODE_BRA);
1123       ifInst->DstReg.CondMask = COND_EQ;  /* BRA if cond is zero */
1124       ifInst->DstReg.CondSwizzle = SWIZZLE_X;
1125       ifInst->Comment = _mesa_strdup("if zero");
1126    }
1127
1128    /* if body */
1129    emit(emitInfo, n->Children[1]);
1130
1131    if (n->Children[2]) {
1132       /* have else body */
1133       elseInstLoc = prog->NumInstructions;
1134       if (emitInfo->EmitHighLevelInstructions) {
1135          (void) new_instruction(emitInfo, OPCODE_ELSE);
1136       }
1137       else {
1138          /* jump to endif instruction */
1139          struct prog_instruction *inst;
1140          inst = new_instruction(emitInfo, OPCODE_BRA);
1141          inst->Comment = _mesa_strdup("else");
1142          inst->DstReg.CondMask = COND_TR;  /* always branch */
1143       }
1144       ifInst = prog->Instructions + ifInstLoc;
1145       ifInst->BranchTarget = prog->NumInstructions;
1146
1147       emit(emitInfo, n->Children[2]);
1148    }
1149    else {
1150       /* no else body */
1151       ifInst = prog->Instructions + ifInstLoc;
1152       ifInst->BranchTarget = prog->NumInstructions + 1;
1153    }
1154
1155    if (emitInfo->EmitHighLevelInstructions) {
1156       (void) new_instruction(emitInfo, OPCODE_ENDIF);
1157    }
1158
1159    if (n->Children[2]) {
1160       struct prog_instruction *elseInst;
1161       elseInst = prog->Instructions + elseInstLoc;
1162       elseInst->BranchTarget = prog->NumInstructions;
1163    }
1164    return NULL;
1165 }
1166
1167
1168 static struct prog_instruction *
1169 emit_loop(slang_emit_info *emitInfo, slang_ir_node *n)
1170 {
1171    struct gl_program *prog = emitInfo->prog;
1172    struct prog_instruction *beginInst, *endInst;
1173    GLuint beginInstLoc, endInstLoc;
1174    slang_ir_node *ir;
1175
1176    /* emit OPCODE_BGNLOOP */
1177    beginInstLoc = prog->NumInstructions;
1178    if (emitInfo->EmitHighLevelInstructions) {
1179       (void) new_instruction(emitInfo, OPCODE_BGNLOOP);
1180    }
1181
1182    /* body */
1183    emit(emitInfo, n->Children[0]);
1184
1185    endInstLoc = prog->NumInstructions;
1186    if (emitInfo->EmitHighLevelInstructions) {
1187       /* emit OPCODE_ENDLOOP */
1188       endInst = new_instruction(emitInfo, OPCODE_ENDLOOP);
1189    }
1190    else {
1191       /* emit unconditional BRA-nch */
1192       endInst = new_instruction(emitInfo, OPCODE_BRA);
1193       endInst->DstReg.CondMask = COND_TR;  /* always true */
1194    }
1195    /* end instruction's BranchTarget points to top of loop */
1196    endInst->BranchTarget = beginInstLoc;
1197
1198    if (emitInfo->EmitHighLevelInstructions) {
1199       /* BGNLOOP's BranchTarget points to the ENDLOOP inst */
1200       beginInst = prog->Instructions + beginInstLoc;
1201       beginInst->BranchTarget = prog->NumInstructions - 1;
1202    }
1203
1204    /* Done emitting loop code.  Now walk over the loop's linked list of
1205     * BREAK and CONT nodes, filling in their BranchTarget fields (which
1206     * will point to the ENDLOOP+1 or BGNLOOP instructions, respectively).
1207     */
1208    for (ir = n->BranchNode; ir; ir = ir->BranchNode) {
1209       struct prog_instruction *inst = prog->Instructions + ir->InstLocation;
1210       assert(inst->BranchTarget < 0);
1211       if (ir->Opcode == IR_BREAK ||
1212           ir->Opcode == IR_BREAK_IF_FALSE ||
1213           ir->Opcode == IR_BREAK_IF_TRUE) {
1214          assert(inst->Opcode == OPCODE_BRK ||
1215                 inst->Opcode == OPCODE_BRA);
1216          /* go to instruction after end of loop */
1217          inst->BranchTarget = endInstLoc + 1;
1218       }
1219       else {
1220          assert(ir->Opcode == IR_CONT ||
1221                 ir->Opcode == IR_CONT_IF_FALSE ||
1222                 ir->Opcode == IR_CONT_IF_TRUE);
1223          assert(inst->Opcode == OPCODE_CONT ||
1224                 inst->Opcode == OPCODE_BRA);
1225          /* to go instruction at top of loop */
1226          inst->BranchTarget = beginInstLoc;
1227       }
1228    }
1229    return NULL;
1230 }
1231
1232
1233 /**
1234  * "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(slang_emit_info *emitInfo, slang_ir_node *n)
1239 {
1240    gl_inst_opcode opcode;
1241    struct prog_instruction *inst;
1242    n->InstLocation = emitInfo->prog->NumInstructions;
1243    if (emitInfo->EmitHighLevelInstructions) {
1244       opcode = (n->Opcode == IR_CONT) ? OPCODE_CONT : OPCODE_BRK;
1245    }
1246    else {
1247       opcode = OPCODE_BRA;
1248    }
1249    inst = new_instruction(emitInfo, opcode);
1250    inst->DstReg.CondMask = COND_TR;  /* always true */
1251    return inst;
1252 }
1253
1254
1255 /**
1256  * Conditional "continue" or "break" statement.
1257  * Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted.
1258  */
1259 static struct prog_instruction *
1260 emit_cont_break_if(slang_emit_info *emitInfo, slang_ir_node *n,
1261                    GLboolean breakTrue)
1262 {
1263    gl_inst_opcode opcode;
1264    struct prog_instruction *inst;
1265
1266    /* evaluate condition expr, setting cond codes */
1267    inst = emit(emitInfo, n->Children[0]);
1268    assert(inst);
1269    inst->CondUpdate = GL_TRUE;
1270
1271    n->InstLocation = emitInfo->prog->NumInstructions;
1272    if (emitInfo->EmitHighLevelInstructions) {
1273       if (n->Opcode == IR_CONT_IF_TRUE ||
1274           n->Opcode == IR_CONT_IF_FALSE)
1275          opcode = OPCODE_CONT;
1276       else
1277          opcode = OPCODE_BRK;
1278    }
1279    else {
1280       opcode = OPCODE_BRA;
1281    }
1282    inst = new_instruction(emitInfo, opcode);
1283    inst->DstReg.CondMask = breakTrue ? COND_NE : COND_EQ;
1284    return inst;
1285 }
1286
1287
1288
1289 /**
1290  * Remove any SWIZZLE_NIL terms from given swizzle mask (smear prev term).
1291  * Ex: fix_swizzle("zyNN") -> "zyyy"
1292  */
1293 static GLuint
1294 fix_swizzle(GLuint swizzle)
1295 {
1296    GLuint swz[4], i;
1297    for (i = 0; i < 4; i++) {
1298       swz[i] = GET_SWZ(swizzle, i);
1299       if (swz[i] == SWIZZLE_NIL) {
1300          swz[i] = swz[i - 1];
1301       }
1302    }
1303    return MAKE_SWIZZLE4(swz[0], swz[1], swz[2], swz[3]);
1304 }
1305
1306
1307 static struct prog_instruction *
1308 emit_swizzle(slang_emit_info *emitInfo, slang_ir_node *n)
1309 {
1310    GLuint swizzle;
1311
1312    /* swizzled storage access */
1313    (void) emit(emitInfo, n->Children[0]);
1314
1315    /* "pull-up" the child's storage info, applying our swizzle info */
1316    n->Store->File  = n->Children[0]->Store->File;
1317    n->Store->Index = n->Children[0]->Store->Index;
1318    n->Store->Size  = n->Children[0]->Store->Size;
1319    /*n->Var = n->Children[0]->Var; XXX for debug */
1320    assert(n->Store->Index >= 0);
1321
1322    swizzle = fix_swizzle(n->Store->Swizzle);
1323 #ifdef DEBUG
1324    {
1325       GLuint s = n->Children[0]->Store->Swizzle;
1326       assert(GET_SWZ(s, 0) != SWIZZLE_NIL);
1327       assert(GET_SWZ(s, 1) != SWIZZLE_NIL);
1328       assert(GET_SWZ(s, 2) != SWIZZLE_NIL);
1329       assert(GET_SWZ(s, 3) != SWIZZLE_NIL);
1330    }
1331 #endif
1332
1333    /* apply this swizzle to child's swizzle to get composed swizzle */
1334    n->Store->Swizzle = swizzle_swizzle(n->Children[0]->Store->Swizzle,
1335                                        swizzle);
1336    return NULL;
1337 }
1338
1339
1340 /**
1341  * Dereference array element.  Just resolve storage for the array
1342  * element represented by this node.
1343  */
1344 static struct prog_instruction *
1345 emit_array_element(slang_emit_info *emitInfo, slang_ir_node *n)
1346 {
1347    assert(n->Store);
1348    assert(n->Store->File != PROGRAM_UNDEFINED);
1349    assert(n->Store->Size > 0);
1350
1351    if (n->Store->File == PROGRAM_STATE_VAR) {
1352       n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
1353       return NULL;
1354    }
1355
1356
1357    if (n->Children[1]->Opcode == IR_FLOAT) {
1358       /* Constant index */
1359       const GLint arrayAddr = n->Children[0]->Store->Index;
1360       const GLint index = (GLint) n->Children[1]->Value[0];
1361       n->Store->Index = arrayAddr + index;
1362    }
1363    else {
1364       /* Variable index - PROBLEM */
1365       const GLint arrayAddr = n->Children[0]->Store->Index;
1366       const GLint index = 0;
1367       _mesa_problem(NULL, "variable array indexes not supported yet!");
1368       n->Store->Index = arrayAddr + index;
1369    }
1370    return NULL; /* no instruction */
1371 }
1372
1373
1374 /**
1375  * Resolve storage for accessing a structure field.
1376  */
1377 static struct prog_instruction *
1378 emit_struct_field(slang_emit_info *emitInfo, slang_ir_node *n)
1379 {
1380    if (n->Store->File == PROGRAM_STATE_VAR) {
1381       n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
1382       return NULL;
1383    }
1384    else {
1385       _mesa_problem(NULL, "structs/fields not supported yet");
1386    }
1387    return NULL; /* no instruction */
1388 }
1389
1390
1391 static struct prog_instruction *
1392 emit(slang_emit_info *emitInfo, slang_ir_node *n)
1393 {
1394    struct prog_instruction *inst;
1395    if (!n)
1396       return NULL;
1397
1398    switch (n->Opcode) {
1399    case IR_SEQ:
1400       /* sequence of two sub-trees */
1401       assert(n->Children[0]);
1402       assert(n->Children[1]);
1403       emit(emitInfo, n->Children[0]);
1404       inst = emit(emitInfo, n->Children[1]);
1405       assert(!n->Store);
1406       n->Store = n->Children[1]->Store;
1407       return inst;
1408
1409    case IR_SCOPE:
1410       /* new variable scope */
1411       _slang_push_var_table(emitInfo->vt);
1412       inst = emit(emitInfo, n->Children[0]);
1413       _slang_pop_var_table(emitInfo->vt);
1414       return inst;
1415
1416    case IR_VAR_DECL:
1417       /* Variable declaration - allocate a register for it */
1418       assert(n->Store);
1419       assert(n->Store->File != PROGRAM_UNDEFINED);
1420       assert(n->Store->Size > 0);
1421       assert(n->Store->Index < 0);
1422       if (!n->Var || n->Var->isTemp) {
1423          /* a nameless/temporary variable, will be freed after first use */
1424          if (!_slang_alloc_temp(emitInfo->vt, n->Store)) {
1425             slang_info_log_error(emitInfo->log,
1426                                  "Ran out of registers, too many temporaries");
1427             return NULL;
1428          }
1429       }
1430       else {
1431          /* a regular variable */
1432          _slang_add_variable(emitInfo->vt, n->Var);
1433          if (!_slang_alloc_var(emitInfo->vt, n->Store)) {
1434             slang_info_log_error(emitInfo->log,
1435                                  "Ran out of registers, too many variables");
1436             return NULL;
1437          }
1438          /*
1439          printf("IR_VAR_DECL %s %d store %p\n",
1440                 (char*) n->Var->a_name, n->Store->Index, (void*) n->Store);
1441          */
1442          assert(n->Var->aux == n->Store);
1443       }
1444       if (emitInfo->EmitComments) {
1445          /* emit NOP with comment describing the variable's storage location */
1446          char s[1000];
1447          sprintf(s, "TEMP[%d]%s = %s (size %d)",
1448                  n->Store->Index,
1449                  _mesa_swizzle_string(n->Store->Swizzle, 0, GL_FALSE),
1450                  (char *) n->Var->a_name,
1451                  n->Store->Size);
1452          inst = new_instruction(emitInfo, OPCODE_NOP);
1453          inst->Comment = _mesa_strdup(s);
1454          return inst;
1455       }
1456       return NULL;
1457
1458    case IR_VAR:
1459       /* Reference to a variable
1460        * Storage should have already been resolved/allocated.
1461        */
1462       assert(n->Store);
1463       assert(n->Store->File != PROGRAM_UNDEFINED);
1464
1465       if (n->Store->File == PROGRAM_STATE_VAR &&
1466           n->Store->Index < 0) {
1467          n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
1468       }
1469
1470       if (n->Store->Index < 0) {
1471          printf("#### VAR %s not allocated!\n", (char*)n->Var->a_name);
1472       }
1473       assert(n->Store->Index >= 0);
1474       assert(n->Store->Size > 0);
1475       break;
1476
1477    case IR_ELEMENT:
1478       return emit_array_element(emitInfo, n);
1479    case IR_FIELD:
1480       return emit_struct_field(emitInfo, n);
1481    case IR_SWIZZLE:
1482       return emit_swizzle(emitInfo, n);
1483
1484    case IR_I_TO_F:
1485       {
1486          n->Store = n->Children[0]->Store;
1487       }
1488       return NULL;
1489
1490    /* Simple arithmetic */
1491    /* unary */
1492    case IR_RSQ:
1493    case IR_RCP:
1494    case IR_FLOOR:
1495    case IR_FRAC:
1496    case IR_F_TO_I:
1497    case IR_ABS:
1498    case IR_SIN:
1499    case IR_COS:
1500    case IR_DDX:
1501    case IR_DDY:
1502    case IR_NOISE1:
1503    case IR_NOISE2:
1504    case IR_NOISE3:
1505    case IR_NOISE4:
1506    /* binary */
1507    case IR_ADD:
1508    case IR_SUB:
1509    case IR_MUL:
1510    case IR_DOT4:
1511    case IR_DOT3:
1512    case IR_CROSS:
1513    case IR_MIN:
1514    case IR_MAX:
1515    case IR_SEQUAL:
1516    case IR_SNEQUAL:
1517    case IR_SGE:
1518    case IR_SGT:
1519    case IR_POW:
1520    case IR_EXP:
1521    case IR_EXP2:
1522    /* trinary operators */
1523    case IR_LRP:
1524       return emit_arith(emitInfo, n);
1525    case IR_CLAMP:
1526       return emit_clamp(emitInfo, n);
1527    case IR_TEX:
1528    case IR_TEXB:
1529    case IR_TEXP:
1530       return emit_tex(emitInfo, n);
1531    case IR_NEG:
1532       return emit_negation(emitInfo, n);
1533    case IR_FLOAT:
1534       /* find storage location for this float constant */
1535       n->Store->Index = _mesa_add_unnamed_constant(emitInfo->prog->Parameters, n->Value,
1536                                                    n->Store->Size,
1537                                                    &n->Store->Swizzle);
1538       if (n->Store->Index < 0) {
1539          slang_info_log_error(emitInfo->log, "Ran out of space for constants");
1540          return NULL;
1541       }
1542       return NULL;
1543
1544    case IR_MOVE:
1545       return emit_move(emitInfo, n);
1546
1547    case IR_COND:
1548       return emit_cond(emitInfo, n);
1549
1550    case IR_NOT:
1551       return emit_not(emitInfo, n);
1552
1553    case IR_LABEL:
1554       return emit_label(emitInfo, n);
1555    case IR_JUMP:
1556       return emit_jump(emitInfo, n);
1557    case IR_CJUMP0:
1558       return emit_cjump(emitInfo, n, 0);
1559    case IR_CJUMP1:
1560       return emit_cjump(emitInfo, n, 1);
1561    case IR_KILL:
1562       return emit_kill(emitInfo);
1563
1564    case IR_IF:
1565       return emit_if(emitInfo, n);
1566
1567    case IR_LOOP:
1568       return emit_loop(emitInfo, n);
1569    case IR_BREAK_IF_FALSE:
1570    case IR_CONT_IF_FALSE:
1571       return emit_cont_break_if(emitInfo, n, GL_FALSE);
1572    case IR_BREAK_IF_TRUE:
1573    case IR_CONT_IF_TRUE:
1574       return emit_cont_break_if(emitInfo, n, GL_TRUE);
1575    case IR_BREAK:
1576       /* fall-through */
1577    case IR_CONT:
1578       return emit_cont_break(emitInfo, n);
1579
1580    case IR_BEGIN_SUB:
1581       return new_instruction(emitInfo, OPCODE_BGNSUB);
1582    case IR_END_SUB:
1583       return new_instruction(emitInfo, OPCODE_ENDSUB);
1584    case IR_RETURN:
1585       return new_instruction(emitInfo, OPCODE_RET);
1586
1587    case IR_NOP:
1588       return NULL;
1589
1590    default:
1591       _mesa_problem(NULL, "Unexpected IR opcode in emit()\n");
1592       abort();
1593    }
1594    return NULL;
1595 }
1596
1597
1598 GLboolean
1599 _slang_emit_code(slang_ir_node *n, slang_var_table *vt,
1600                  struct gl_program *prog, GLboolean withEnd,
1601                  slang_info_log *log)
1602 {
1603    GET_CURRENT_CONTEXT(ctx);
1604    GLboolean success;
1605    slang_emit_info emitInfo;
1606
1607    emitInfo.log = log;
1608    emitInfo.vt = vt;
1609    emitInfo.prog = prog;
1610
1611    emitInfo.EmitHighLevelInstructions = ctx->Shader.EmitHighLevelInstructions;
1612    emitInfo.EmitComments = ctx->Shader.EmitComments;
1613
1614    (void) emit(&emitInfo, n);
1615
1616    /* finish up by adding the END opcode to program */
1617    if (withEnd) {
1618       struct prog_instruction *inst;
1619       inst = new_instruction(&emitInfo, OPCODE_END);
1620    }
1621    success = GL_TRUE;
1622
1623    printf("*********** End generate code (%u inst):\n", prog->NumInstructions);
1624 #if 0
1625    _mesa_print_program(prog);
1626    _mesa_print_program_parameters(ctx,prog);
1627 #endif
1628
1629    return success;
1630 }