src/glsl/opt_algebraic.cpp

   1 /*
   2  * Copyright © 2010 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  21  * DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 /**
  25  * \file opt_algebraic.cpp
  26  *
  27  * Takes advantage of association, commutivity, and other algebraic
  28  * properties to simplify expressions.
  29  */
  30
  31 #include "ir.h"
  32 #include "ir_visitor.h"
  33 #include "ir_rvalue_visitor.h"
  34 #include "ir_optimization.h"
  35 #include "ir_builder.h"
  36 #include "compiler/glsl_types.h"
  37
  38 using namespace ir_builder;
  39
  40 namespace {
  41
  42 /**
  43  * Visitor class for replacing expressions with ir_constant values.
  44  */
  45
  46 class ir_algebraic_visitor : public ir_rvalue_visitor {
  47 public:
  48    ir_algebraic_visitor(bool native_integers,
  49                         const struct gl_shader_compiler_options *options)
  50       : options(options)
  51    {
  52       this->progress = false;
  53       this->mem_ctx = NULL;
  54       this->native_integers = native_integers;
  55    }
  56
  57    virtual ~ir_algebraic_visitor()
  58    {
  59    }
  60
  61    ir_rvalue *handle_expression(ir_expression *ir);
  62    void handle_rvalue(ir_rvalue **rvalue);
  63    bool reassociate_constant(ir_expression *ir1,
  64                              int const_index,
  65                              ir_constant *constant,
  66                              ir_expression *ir2);
  67    void reassociate_operands(ir_expression *ir1,
  68                              int op1,
  69                              ir_expression *ir2,
  70                              int op2);
  71    ir_rvalue *swizzle_if_required(ir_expression *expr,
  72                                   ir_rvalue *operand);
  73
  74    const struct gl_shader_compiler_options *options;
  75    void *mem_ctx;
  76
  77    bool native_integers;
  78    bool progress;
  79 };
  80
  81 } /* unnamed namespace */
  82
  83 static inline bool
  84 is_vec_zero(ir_constant *ir)
  85 {
  86    return (ir == NULL) ? false : ir->is_zero();
  87 }
  88
  89 static inline bool
  90 is_vec_one(ir_constant *ir)
  91 {
  92    return (ir == NULL) ? false : ir->is_one();
  93 }
  94
  95 static inline bool
  96 is_vec_two(ir_constant *ir)
  97 {
  98    return (ir == NULL) ? false : ir->is_value(2.0, 2);
  99 }
 100
 101 static inline bool
 102 is_vec_four(ir_constant *ir)
 103 {
 104    return (ir == NULL) ? false : ir->is_value(4.0, 4);
 105 }
 106
 107 static inline bool
 108 is_vec_negative_one(ir_constant *ir)
 109 {
 110    return (ir == NULL) ? false : ir->is_negative_one();
 111 }
 112
 113 static inline bool
 114 is_valid_vec_const(ir_constant *ir)
 115 {
 116    if (ir == NULL)
 117       return false;
 118
 119    if (!ir->type->is_scalar() && !ir->type->is_vector())
 120       return false;
 121
 122    return true;
 123 }
 124
 125 static inline bool
 126 is_less_than_one(ir_constant *ir)
 127 {
 128    assert(ir->type->base_type == GLSL_TYPE_FLOAT);
 129
 130    if (!is_valid_vec_const(ir))
 131       return false;
 132
 133    unsigned component = 0;
 134    for (int c = 0; c < ir->type->vector_elements; c++) {
 135       if (ir->get_float_component(c) < 1.0f)
 136          component++;
 137    }
 138
 139    return (component == ir->type->vector_elements);
 140 }
 141
 142 static inline bool
 143 is_greater_than_zero(ir_constant *ir)
 144 {
 145    assert(ir->type->base_type == GLSL_TYPE_FLOAT);
 146
 147    if (!is_valid_vec_const(ir))
 148       return false;
 149
 150    unsigned component = 0;
 151    for (int c = 0; c < ir->type->vector_elements; c++) {
 152       if (ir->get_float_component(c) > 0.0f)
 153          component++;
 154    }
 155
 156    return (component == ir->type->vector_elements);
 157 }
 158
 159 static void
 160 update_type(ir_expression *ir)
 161 {
 162    if (ir->operands[0]->type->is_vector())
 163       ir->type = ir->operands[0]->type;
 164    else
 165       ir->type = ir->operands[1]->type;
 166 }
 167
 168 /* Recognize (v.x + v.y) + (v.z + v.w) as dot(v, 1.0) */
 169 static ir_expression *
 170 try_replace_with_dot(ir_expression *expr0, ir_expression *expr1, void *mem_ctx)
 171 {
 172    if (expr0 && expr0->operation == ir_binop_add &&
 173        expr0->type->is_float() &&
 174        expr1 && expr1->operation == ir_binop_add &&
 175        expr1->type->is_float()) {
 176       ir_swizzle *x = expr0->operands[0]->as_swizzle();
 177       ir_swizzle *y = expr0->operands[1]->as_swizzle();
 178       ir_swizzle *z = expr1->operands[0]->as_swizzle();
 179       ir_swizzle *w = expr1->operands[1]->as_swizzle();
 180
 181       if (!x || x->mask.num_components != 1 ||
 182           !y || y->mask.num_components != 1 ||
 183           !z || z->mask.num_components != 1 ||
 184           !w || w->mask.num_components != 1) {
 185          return NULL;
 186       }
 187
 188       bool swiz_seen[4] = {false, false, false, false};
 189       swiz_seen[x->mask.x] = true;
 190       swiz_seen[y->mask.x] = true;
 191       swiz_seen[z->mask.x] = true;
 192       swiz_seen[w->mask.x] = true;
 193
 194       if (!swiz_seen[0] || !swiz_seen[1] ||
 195           !swiz_seen[2] || !swiz_seen[3]) {
 196          return NULL;
 197       }
 198
 199       if (x->val->equals(y->val) &&
 200           x->val->equals(z->val) &&
 201           x->val->equals(w->val)) {
 202          return dot(x->val, new(mem_ctx) ir_constant(1.0f, 4));
 203       }
 204    }
 205    return NULL;
 206 }
 207
 208 void
 209 ir_algebraic_visitor::reassociate_operands(ir_expression *ir1,
 210                                            int op1,
 211                                            ir_expression *ir2,
 212                                            int op2)
 213 {
 214    ir_rvalue *temp = ir2->operands[op2];
 215    ir2->operands[op2] = ir1->operands[op1];
 216    ir1->operands[op1] = temp;
 217
 218    /* Update the type of ir2.  The type of ir1 won't have changed --
 219     * base types matched, and at least one of the operands of the 2
 220     * binops is still a vector if any of them were.
 221     */
 222    update_type(ir2);
 223
 224    this->progress = true;
 225 }
 226
 227 /**
 228  * Reassociates a constant down a tree of adds or multiplies.
 229  *
 230  * Consider (2 * (a * (b * 0.5))).  We want to send up with a * b.
 231  */
 232 bool
 233 ir_algebraic_visitor::reassociate_constant(ir_expression *ir1, int const_index,
 234                                            ir_constant *constant,
 235                                            ir_expression *ir2)
 236 {
 237    if (!ir2 || ir1->operation != ir2->operation)
 238       return false;
 239
 240    /* Don't want to even think about matrices. */
 241    if (ir1->operands[0]->type->is_matrix() ||
 242        ir1->operands[1]->type->is_matrix() ||
 243        ir2->operands[0]->type->is_matrix() ||
 244        ir2->operands[1]->type->is_matrix())
 245       return false;
 246
 247    ir_constant *ir2_const[2];
 248    ir2_const[0] = ir2->operands[0]->constant_expression_value();
 249    ir2_const[1] = ir2->operands[1]->constant_expression_value();
 250
 251    if (ir2_const[0] && ir2_const[1])
 252       return false;
 253
 254    if (ir2_const[0]) {
 255       reassociate_operands(ir1, const_index, ir2, 1);
 256       return true;
 257    } else if (ir2_const[1]) {
 258       reassociate_operands(ir1, const_index, ir2, 0);
 259       return true;
 260    }
 261
 262    if (reassociate_constant(ir1, const_index, constant,
 263                             ir2->operands[0]->as_expression())) {
 264       update_type(ir2);
 265       return true;
 266    }
 267
 268    if (reassociate_constant(ir1, const_index, constant,
 269                             ir2->operands[1]->as_expression())) {
 270       update_type(ir2);
 271       return true;
 272    }
 273
 274    return false;
 275 }
 276
 277 /* When eliminating an expression and just returning one of its operands,
 278  * we may need to swizzle that operand out to a vector if the expression was
 279  * vector type.
 280  */
 281 ir_rvalue *
 282 ir_algebraic_visitor::swizzle_if_required(ir_expression *expr,
 283                                           ir_rvalue *operand)
 284 {
 285    if (expr->type->is_vector() && operand->type->is_scalar()) {
 286       return new(mem_ctx) ir_swizzle(operand, 0, 0, 0, 0,
 287                                      expr->type->vector_elements);
 288    } else
 289       return operand;
 290 }
 291
 292 ir_rvalue *
 293 ir_algebraic_visitor::handle_expression(ir_expression *ir)
 294 {
 295    ir_constant *op_const[4] = {NULL, NULL, NULL, NULL};
 296    ir_expression *op_expr[4] = {NULL, NULL, NULL, NULL};
 297    unsigned int i;
 298
 299    if (ir->operation == ir_binop_mul &&
 300        ir->operands[0]->type->is_matrix() &&
 301        ir->operands[1]->type->is_vector()) {
 302       ir_expression *matrix_mul = ir->operands[0]->as_expression();
 303
 304       if (matrix_mul && matrix_mul->operation == ir_binop_mul &&
 305          matrix_mul->operands[0]->type->is_matrix() &&
 306          matrix_mul->operands[1]->type->is_matrix()) {
 307
 308          return mul(matrix_mul->operands[0],
 309                     mul(matrix_mul->operands[1], ir->operands[1]));
 310       }
 311    }
 312
 313    assert(ir->get_num_operands() <= 4);
 314    for (i = 0; i < ir->get_num_operands(); i++) {
 315       if (ir->operands[i]->type->is_matrix())
 316          return ir;
 317
 318       op_const[i] = ir->operands[i]->constant_expression_value();
 319       op_expr[i] = ir->operands[i]->as_expression();
 320    }
 321
 322    if (this->mem_ctx == NULL)
 323       this->mem_ctx = ralloc_parent(ir);
 324
 325    switch (ir->operation) {
 326    case ir_unop_bit_not:
 327       if (op_expr[0] && op_expr[0]->operation == ir_unop_bit_not)
 328          return op_expr[0]->operands[0];
 329       break;
 330
 331    case ir_unop_abs:
 332       if (op_expr[0] == NULL)
 333          break;
 334
 335       switch (op_expr[0]->operation) {
 336       case ir_unop_abs:
 337       case ir_unop_neg:
 338          return abs(op_expr[0]->operands[0]);
 339       default:
 340          break;
 341       }
 342       break;
 343
 344    case ir_unop_neg:
 345       if (op_expr[0] == NULL)
 346          break;
 347
 348       if (op_expr[0]->operation == ir_unop_neg) {
 349          return op_expr[0]->operands[0];
 350       }
 351       break;
 352
 353    case ir_unop_exp:
 354       if (op_expr[0] == NULL)
 355          break;
 356
 357       if (op_expr[0]->operation == ir_unop_log) {
 358          return op_expr[0]->operands[0];
 359       }
 360       break;
 361
 362    case ir_unop_log:
 363       if (op_expr[0] == NULL)
 364          break;
 365
 366       if (op_expr[0]->operation == ir_unop_exp) {
 367          return op_expr[0]->operands[0];
 368       }
 369       break;
 370
 371    case ir_unop_exp2:
 372       if (op_expr[0] == NULL)
 373          break;
 374
 375       if (op_expr[0]->operation == ir_unop_log2) {
 376          return op_expr[0]->operands[0];
 377       }
 378
 379       if (!options->EmitNoPow && op_expr[0]->operation == ir_binop_mul) {
 380          for (int log2_pos = 0; log2_pos < 2; log2_pos++) {
 381             ir_expression *log2_expr =
 382                op_expr[0]->operands[log2_pos]->as_expression();
 383
 384             if (log2_expr && log2_expr->operation == ir_unop_log2) {
 385                return new(mem_ctx) ir_expression(ir_binop_pow,
 386                                                  ir->type,
 387                                                  log2_expr->operands[0],
 388                                                  op_expr[0]->operands[1 - log2_pos]);
 389             }
 390          }
 391       }
 392       break;
 393
 394    case ir_unop_log2:
 395       if (op_expr[0] == NULL)
 396          break;
 397
 398       if (op_expr[0]->operation == ir_unop_exp2) {
 399          return op_expr[0]->operands[0];
 400       }
 401       break;
 402
 403    case ir_unop_f2i:
 404    case ir_unop_f2u:
 405       if (op_expr[0] && op_expr[0]->operation == ir_unop_trunc) {
 406          return new(mem_ctx) ir_expression(ir->operation,
 407                                            ir->type,
 408                                            op_expr[0]->operands[0]);
 409       }
 410       break;
 411
 412    case ir_unop_logic_not: {
 413       enum ir_expression_operation new_op = ir_unop_logic_not;
 414
 415       if (op_expr[0] == NULL)
 416          break;
 417
 418       switch (op_expr[0]->operation) {
 419       case ir_binop_less:    new_op = ir_binop_gequal;  break;
 420       case ir_binop_greater: new_op = ir_binop_lequal;  break;
 421       case ir_binop_lequal:  new_op = ir_binop_greater; break;
 422       case ir_binop_gequal:  new_op = ir_binop_less;    break;
 423       case ir_binop_equal:   new_op = ir_binop_nequal;  break;
 424       case ir_binop_nequal:  new_op = ir_binop_equal;   break;
 425       case ir_binop_all_equal:   new_op = ir_binop_any_nequal;  break;
 426       case ir_binop_any_nequal:  new_op = ir_binop_all_equal;   break;
 427
 428       default:
 429          /* The default case handler is here to silence a warning from GCC.
 430           */
 431          break;
 432       }
 433
 434       if (new_op != ir_unop_logic_not) {
 435          return new(mem_ctx) ir_expression(new_op,
 436                                            ir->type,
 437                                            op_expr[0]->operands[0],
 438                                            op_expr[0]->operands[1]);
 439       }
 440
 441       break;
 442    }
 443
 444    case ir_unop_saturate:
 445       if (op_expr[0] && op_expr[0]->operation == ir_binop_add) {
 446          ir_expression *b2f_0 = op_expr[0]->operands[0]->as_expression();
 447          ir_expression *b2f_1 = op_expr[0]->operands[1]->as_expression();
 448
 449          if (b2f_0 && b2f_0->operation == ir_unop_b2f &&
 450              b2f_1 && b2f_1->operation == ir_unop_b2f) {
 451             return b2f(logic_or(b2f_0->operands[0], b2f_1->operands[0]));
 452          }
 453       }
 454       break;
 455
 456    case ir_binop_add:
 457       if (is_vec_zero(op_const[0]))
 458          return ir->operands[1];
 459       if (is_vec_zero(op_const[1]))
 460          return ir->operands[0];
 461
 462       /* Reassociate addition of constants so that we can do constant
 463        * folding.
 464        */
 465       if (op_const[0] && !op_const[1])
 466          reassociate_constant(ir, 0, op_const[0], op_expr[1]);
 467       if (op_const[1] && !op_const[0])
 468          reassociate_constant(ir, 1, op_const[1], op_expr[0]);
 469
 470       /* Recognize (v.x + v.y) + (v.z + v.w) as dot(v, 1.0) */
 471       if (options->OptimizeForAOS) {
 472          ir_expression *expr = try_replace_with_dot(op_expr[0], op_expr[1],
 473                                                     mem_ctx);
 474          if (expr)
 475             return expr;
 476       }
 477
 478       /* Replace (-x + y) * a + x and commutative variations with lrp(x, y, a).
 479        *
 480        * (-x + y) * a + x
 481        * (x * -a) + (y * a) + x
 482        * x + (x * -a) + (y * a)
 483        * x * (1 - a) + y * a
 484        * lrp(x, y, a)
 485        */
 486       for (int mul_pos = 0; mul_pos < 2; mul_pos++) {
 487          ir_expression *mul = op_expr[mul_pos];
 488
 489          if (!mul || mul->operation != ir_binop_mul)
 490             continue;
 491
 492          /* Multiply found on one of the operands. Now check for an
 493           * inner addition operation.
 494           */
 495          for (int inner_add_pos = 0; inner_add_pos < 2; inner_add_pos++) {
 496             ir_expression *inner_add =
 497                mul->operands[inner_add_pos]->as_expression();
 498
 499             if (!inner_add || inner_add->operation != ir_binop_add)
 500                continue;
 501
 502             /* Inner addition found on one of the operands. Now check for
 503              * one of the operands of the inner addition to be the negative
 504              * of x_operand.
 505              */
 506             for (int neg_pos = 0; neg_pos < 2; neg_pos++) {
 507                ir_expression *neg =
 508                   inner_add->operands[neg_pos]->as_expression();
 509
 510                if (!neg || neg->operation != ir_unop_neg)
 511                   continue;
 512
 513                ir_rvalue *x_operand = ir->operands[1 - mul_pos];
 514
 515                if (!neg->operands[0]->equals(x_operand))
 516                   continue;
 517
 518                ir_rvalue *y_operand = inner_add->operands[1 - neg_pos];
 519                ir_rvalue *a_operand = mul->operands[1 - inner_add_pos];
 520
 521                if (x_operand->type != y_operand->type ||
 522                    x_operand->type != a_operand->type)
 523                   continue;
 524
 525                return lrp(x_operand, y_operand, a_operand);
 526             }
 527          }
 528       }
 529
 530       break;
 531
 532    case ir_binop_sub:
 533       if (is_vec_zero(op_const[0]))
 534          return neg(ir->operands[1]);
 535       if (is_vec_zero(op_const[1]))
 536          return ir->operands[0];
 537       break;
 538
 539    case ir_binop_mul:
 540       if (is_vec_one(op_const[0]))
 541          return ir->operands[1];
 542       if (is_vec_one(op_const[1]))
 543          return ir->operands[0];
 544
 545       if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1]))
 546          return ir_constant::zero(ir, ir->type);
 547
 548       if (is_vec_negative_one(op_const[0]))
 549          return neg(ir->operands[1]);
 550       if (is_vec_negative_one(op_const[1]))
 551          return neg(ir->operands[0]);
 552
 553       if (op_expr[0] && op_expr[0]->operation == ir_unop_b2f &&
 554           op_expr[1] && op_expr[1]->operation == ir_unop_b2f) {
 555          return b2f(logic_and(op_expr[0]->operands[0], op_expr[1]->operands[0]));
 556       }
 557
 558       /* Reassociate multiplication of constants so that we can do
 559        * constant folding.
 560        */
 561       if (op_const[0] && !op_const[1])
 562          reassociate_constant(ir, 0, op_const[0], op_expr[1]);
 563       if (op_const[1] && !op_const[0])
 564          reassociate_constant(ir, 1, op_const[1], op_expr[0]);
 565
 566       /* Optimizes
 567        *
 568        *    (mul (floor (add (abs x) 0.5) (sign x)))
 569        *
 570        * into
 571        *
 572        *    (trunc (add x (mul (sign x) 0.5)))
 573        */
 574       for (int i = 0; i < 2; i++) {
 575          ir_expression *sign_expr = ir->operands[i]->as_expression();
 576          ir_expression *floor_expr = ir->operands[1 - i]->as_expression();
 577
 578          if (!sign_expr || sign_expr->operation != ir_unop_sign ||
 579              !floor_expr || floor_expr->operation != ir_unop_floor)
 580             continue;
 581
 582          ir_expression *add_expr = floor_expr->operands[0]->as_expression();
 583          if (!add_expr || add_expr->operation != ir_binop_add)
 584             continue;
 585
 586          for (int j = 0; j < 2; j++) {
 587             ir_expression *abs_expr = add_expr->operands[j]->as_expression();
 588             if (!abs_expr || abs_expr->operation != ir_unop_abs)
 589                continue;
 590
 591             ir_constant *point_five = add_expr->operands[1 - j]->as_constant();
 592             if (!point_five || !point_five->is_value(0.5, 0))
 593                continue;
 594
 595             if (abs_expr->operands[0]->equals(sign_expr->operands[0])) {
 596                return trunc(add(abs_expr->operands[0],
 597                                 mul(sign_expr, point_five)));
 598             }
 599          }
 600       }
 601       break;
 602
 603    case ir_binop_div:
 604       if (is_vec_one(op_const[0]) && (
 605                 ir->type->base_type == GLSL_TYPE_FLOAT ||
 606                 ir->type->base_type == GLSL_TYPE_DOUBLE)) {
 607          return new(mem_ctx) ir_expression(ir_unop_rcp,
 608                                            ir->operands[1]->type,
 609                                            ir->operands[1],
 610                                            NULL);
 611       }
 612       if (is_vec_one(op_const[1]))
 613          return ir->operands[0];
 614       break;
 615
 616    case ir_binop_dot:
 617       if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1]))
 618          return ir_constant::zero(mem_ctx, ir->type);
 619
 620       for (int i = 0; i < 2; i++) {
 621          if (!op_const[i])
 622             continue;
 623
 624          unsigned components[4] = { 0 }, count = 0;
 625
 626          for (unsigned c = 0; c < op_const[i]->type->vector_elements; c++) {
 627             if (op_const[i]->is_zero())
 628                continue;
 629
 630             components[count] = c;
 631             count++;
 632          }
 633
 634          /* No channels had zero values; bail. */
 635          if (count >= op_const[i]->type->vector_elements)
 636             break;
 637
 638          ir_expression_operation op = count == 1 ?
 639             ir_binop_mul : ir_binop_dot;
 640
 641          /* Swizzle both operands to remove the channels that were zero. */
 642          return new(mem_ctx)
 643             ir_expression(op, ir->type,
 644                           new(mem_ctx) ir_swizzle(ir->operands[0],
 645                                                   components, count),
 646                           new(mem_ctx) ir_swizzle(ir->operands[1],
 647                                                   components, count));
 648       }
 649       break;
 650
 651    case ir_binop_less:
 652    case ir_binop_lequal:
 653    case ir_binop_greater:
 654    case ir_binop_gequal:
 655    case ir_binop_equal:
 656    case ir_binop_nequal:
 657       for (int add_pos = 0; add_pos < 2; add_pos++) {
 658          ir_expression *add = op_expr[add_pos];
 659
 660          if (!add || add->operation != ir_binop_add)
 661             continue;
 662
 663          ir_constant *zero = op_const[1 - add_pos];
 664          if (!is_vec_zero(zero))
 665             continue;
 666
 667          /* Depending of the zero position we want to optimize
 668           * (0 cmp x+y) into (-x cmp y) or (x+y cmp 0) into (x cmp -y)
 669           */
 670          if (add_pos == 1) {
 671             return new(mem_ctx) ir_expression(ir->operation,
 672                                               neg(add->operands[0]),
 673                                               add->operands[1]);
 674          } else {
 675             return new(mem_ctx) ir_expression(ir->operation,
 676                                               add->operands[0],
 677                                               neg(add->operands[1]));
 678          }
 679       }
 680       break;
 681
 682    case ir_binop_all_equal:
 683    case ir_binop_any_nequal:
 684       if (ir->operands[0]->type->is_scalar() &&
 685           ir->operands[1]->type->is_scalar())
 686          return new(mem_ctx) ir_expression(ir->operation == ir_binop_all_equal
 687                                            ? ir_binop_equal : ir_binop_nequal,
 688                                            ir->operands[0],
 689                                            ir->operands[1]);
 690       break;
 691
 692    case ir_binop_rshift:
 693    case ir_binop_lshift:
 694       /* 0 >> x == 0 */
 695       if (is_vec_zero(op_const[0]))
 696          return ir->operands[0];
 697       /* x >> 0 == x */
 698       if (is_vec_zero(op_const[1]))
 699          return ir->operands[0];
 700       break;
 701
 702    case ir_binop_logic_and:
 703       if (is_vec_one(op_const[0])) {
 704          return ir->operands[1];
 705       } else if (is_vec_one(op_const[1])) {
 706          return ir->operands[0];
 707       } else if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1])) {
 708          return ir_constant::zero(mem_ctx, ir->type);
 709       } else if (op_expr[0] && op_expr[0]->operation == ir_unop_logic_not &&
 710                  op_expr[1] && op_expr[1]->operation == ir_unop_logic_not) {
 711          /* De Morgan's Law:
 712           *    (not A) and (not B) === not (A or B)
 713           */
 714          return logic_not(logic_or(op_expr[0]->operands[0],
 715                                    op_expr[1]->operands[0]));
 716       } else if (ir->operands[0]->equals(ir->operands[1])) {
 717          /* (a && a) == a */
 718          return ir->operands[0];
 719       }
 720       break;
 721
 722    case ir_binop_logic_xor:
 723       if (is_vec_zero(op_const[0])) {
 724          return ir->operands[1];
 725       } else if (is_vec_zero(op_const[1])) {
 726          return ir->operands[0];
 727       } else if (is_vec_one(op_const[0])) {
 728          return logic_not(ir->operands[1]);
 729       } else if (is_vec_one(op_const[1])) {
 730          return logic_not(ir->operands[0]);
 731       } else if (ir->operands[0]->equals(ir->operands[1])) {
 732          /* (a ^^ a) == false */
 733          return ir_constant::zero(mem_ctx, ir->type);
 734       }
 735       break;
 736
 737    case ir_binop_logic_or:
 738       if (is_vec_zero(op_const[0])) {
 739          return ir->operands[1];
 740       } else if (is_vec_zero(op_const[1])) {
 741          return ir->operands[0];
 742       } else if (is_vec_one(op_const[0]) || is_vec_one(op_const[1])) {
 743          ir_constant_data data;
 744
 745          for (unsigned i = 0; i < 16; i++)
 746             data.b[i] = true;
 747
 748          return new(mem_ctx) ir_constant(ir->type, &data);
 749       } else if (op_expr[0] && op_expr[0]->operation == ir_unop_logic_not &&
 750                  op_expr[1] && op_expr[1]->operation == ir_unop_logic_not) {
 751          /* De Morgan's Law:
 752           *    (not A) or (not B) === not (A and B)
 753           */
 754          return logic_not(logic_and(op_expr[0]->operands[0],
 755                                     op_expr[1]->operands[0]));
 756       } else if (ir->operands[0]->equals(ir->operands[1])) {
 757          /* (a || a) == a */
 758          return ir->operands[0];
 759       }
 760       break;
 761
 762    case ir_binop_pow:
 763       /* 1^x == 1 */
 764       if (is_vec_one(op_const[0]))
 765          return op_const[0];
 766
 767       /* x^1 == x */
 768       if (is_vec_one(op_const[1]))
 769          return ir->operands[0];
 770
 771       /* pow(2,x) == exp2(x) */
 772       if (is_vec_two(op_const[0]))
 773          return expr(ir_unop_exp2, ir->operands[1]);
 774
 775       if (is_vec_two(op_const[1])) {
 776          ir_variable *x = new(ir) ir_variable(ir->operands[1]->type, "x",
 777                                               ir_var_temporary);
 778          base_ir->insert_before(x);
 779          base_ir->insert_before(assign(x, ir->operands[0]));
 780          return mul(x, x);
 781       }
 782
 783       if (is_vec_four(op_const[1])) {
 784          ir_variable *x = new(ir) ir_variable(ir->operands[1]->type, "x",
 785                                               ir_var_temporary);
 786          base_ir->insert_before(x);
 787          base_ir->insert_before(assign(x, ir->operands[0]));
 788
 789          ir_variable *squared = new(ir) ir_variable(ir->operands[1]->type,
 790                                                     "squared",
 791                                                     ir_var_temporary);
 792          base_ir->insert_before(squared);
 793          base_ir->insert_before(assign(squared, mul(x, x)));
 794          return mul(squared, squared);
 795       }
 796
 797       break;
 798
 799    case ir_binop_min:
 800    case ir_binop_max:
 801       if (ir->type->base_type != GLSL_TYPE_FLOAT || options->EmitNoSat)
 802          break;
 803
 804       /* Replace min(max) operations and its commutative combinations with
 805        * a saturate operation
 806        */
 807       for (int op = 0; op < 2; op++) {
 808          ir_expression *inner_expr = op_expr[op];
 809          ir_constant *outer_const = op_const[1 - op];
 810          ir_expression_operation op_cond = (ir->operation == ir_binop_max) ?
 811             ir_binop_min : ir_binop_max;
 812
 813          if (!inner_expr || !outer_const || (inner_expr->operation != op_cond))
 814             continue;
 815
 816          /* One of these has to be a constant */
 817          if (!inner_expr->operands[0]->as_constant() &&
 818              !inner_expr->operands[1]->as_constant())
 819             break;
 820
 821          /* Found a min(max) combination. Now try to see if its operands
 822           * meet our conditions that we can do just a single saturate operation
 823           */
 824          for (int minmax_op = 0; minmax_op < 2; minmax_op++) {
 825             ir_rvalue *x = inner_expr->operands[minmax_op];
 826             ir_rvalue *y = inner_expr->operands[1 - minmax_op];
 827
 828             ir_constant *inner_const = y->as_constant();
 829             if (!inner_const)
 830                continue;
 831
 832             /* min(max(x, 0.0), 1.0) is sat(x) */
 833             if (ir->operation == ir_binop_min &&
 834                 inner_const->is_zero() &&
 835                 outer_const->is_one())
 836                return saturate(x);
 837
 838             /* max(min(x, 1.0), 0.0) is sat(x) */
 839             if (ir->operation == ir_binop_max &&
 840                 inner_const->is_one() &&
 841                 outer_const->is_zero())
 842                return saturate(x);
 843
 844             /* min(max(x, 0.0), b) where b < 1.0 is sat(min(x, b)) */
 845             if (ir->operation == ir_binop_min &&
 846                 inner_const->is_zero() &&
 847                 is_less_than_one(outer_const))
 848                return saturate(expr(ir_binop_min, x, outer_const));
 849
 850             /* max(min(x, b), 0.0) where b < 1.0 is sat(min(x, b)) */
 851             if (ir->operation == ir_binop_max &&
 852                 is_less_than_one(inner_const) &&
 853                 outer_const->is_zero())
 854                return saturate(expr(ir_binop_min, x, inner_const));
 855
 856             /* max(min(x, 1.0), b) where b > 0.0 is sat(max(x, b)) */
 857             if (ir->operation == ir_binop_max &&
 858                 inner_const->is_one() &&
 859                 is_greater_than_zero(outer_const))
 860                return saturate(expr(ir_binop_max, x, outer_const));
 861
 862             /* min(max(x, b), 1.0) where b > 0.0 is sat(max(x, b)) */
 863             if (ir->operation == ir_binop_min &&
 864                 is_greater_than_zero(inner_const) &&
 865                 outer_const->is_one())
 866                return saturate(expr(ir_binop_max, x, inner_const));
 867          }
 868       }
 869
 870       break;
 871
 872    case ir_unop_rcp:
 873       if (op_expr[0] && op_expr[0]->operation == ir_unop_rcp)
 874          return op_expr[0]->operands[0];
 875
 876       if (op_expr[0] && (op_expr[0]->operation == ir_unop_exp2 ||
 877                          op_expr[0]->operation == ir_unop_exp)) {
 878          return new(mem_ctx) ir_expression(op_expr[0]->operation, ir->type,
 879                                            neg(op_expr[0]->operands[0]));
 880       }
 881
 882       /* While ir_to_mesa.cpp will lower sqrt(x) to rcp(rsq(x)), it does so at
 883        * its IR level, so we can always apply this transformation.
 884        */
 885       if (op_expr[0] && op_expr[0]->operation == ir_unop_rsq)
 886          return sqrt(op_expr[0]->operands[0]);
 887
 888       /* As far as we know, all backends are OK with rsq. */
 889       if (op_expr[0] && op_expr[0]->operation == ir_unop_sqrt) {
 890          return rsq(op_expr[0]->operands[0]);
 891       }
 892
 893       break;
 894
 895    case ir_triop_fma:
 896       /* Operands are op0 * op1 + op2. */
 897       if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1])) {
 898          return ir->operands[2];
 899       } else if (is_vec_zero(op_const[2])) {
 900          return mul(ir->operands[0], ir->operands[1]);
 901       } else if (is_vec_one(op_const[0])) {
 902          return add(ir->operands[1], ir->operands[2]);
 903       } else if (is_vec_one(op_const[1])) {
 904          return add(ir->operands[0], ir->operands[2]);
 905       }
 906       break;
 907
 908    case ir_triop_lrp:
 909       /* Operands are (x, y, a). */
 910       if (is_vec_zero(op_const[2])) {
 911          return ir->operands[0];
 912       } else if (is_vec_one(op_const[2])) {
 913          return ir->operands[1];
 914       } else if (ir->operands[0]->equals(ir->operands[1])) {
 915          return ir->operands[0];
 916       } else if (is_vec_zero(op_const[0])) {
 917          return mul(ir->operands[1], ir->operands[2]);
 918       } else if (is_vec_zero(op_const[1])) {
 919          unsigned op2_components = ir->operands[2]->type->vector_elements;
 920          ir_constant *one;
 921
 922          switch (ir->type->base_type) {
 923          case GLSL_TYPE_FLOAT:
 924             one = new(mem_ctx) ir_constant(1.0f, op2_components);
 925             break;
 926          case GLSL_TYPE_DOUBLE:
 927             one = new(mem_ctx) ir_constant(1.0, op2_components);
 928             break;
 929          default:
 930             one = NULL;
 931             unreachable("unexpected type");
 932          }
 933
 934          return mul(ir->operands[0], add(one, neg(ir->operands[2])));
 935       }
 936       break;
 937
 938    case ir_triop_csel:
 939       if (is_vec_one(op_const[0]))
 940          return ir->operands[1];
 941       if (is_vec_zero(op_const[0]))
 942          return ir->operands[2];
 943       break;
 944
 945    default:
 946       break;
 947    }
 948
 949    return ir;
 950 }
 951
 952 void
 953 ir_algebraic_visitor::handle_rvalue(ir_rvalue **rvalue)
 954 {
 955    if (!*rvalue)
 956       return;
 957
 958    ir_expression *expr = (*rvalue)->as_expression();
 959    if (!expr || expr->operation == ir_quadop_vector)
 960       return;
 961
 962    ir_rvalue *new_rvalue = handle_expression(expr);
 963    if (new_rvalue == *rvalue)
 964       return;
 965
 966    /* If the expr used to be some vec OP scalar returning a vector, and the
 967     * optimization gave us back a scalar, we still need to turn it into a
 968     * vector.
 969     */
 970    *rvalue = swizzle_if_required(expr, new_rvalue);
 971
 972    this->progress = true;
 973 }
 974
 975 bool
 976 do_algebraic(exec_list *instructions, bool native_integers,
 977              const struct gl_shader_compiler_options *options)
 978 {
 979    ir_algebraic_visitor v(native_integers, options);
 980
 981    visit_list_elements(&v, instructions);
 982
 983    return v.progress;
 984 }