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 "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_negative_one(ir_constant *ir)
 103 {
 104    return (ir == NULL) ? false : ir->is_negative_one();
 105 }
 106
 107 static inline bool
 108 is_valid_vec_const(ir_constant *ir)
 109 {
 110    if (ir == NULL)
 111       return false;
 112
 113    if (!ir->type->is_scalar() && !ir->type->is_vector())
 114       return false;
 115
 116    return true;
 117 }
 118
 119 static inline bool
 120 is_less_than_one(ir_constant *ir)
 121 {
 122    if (!is_valid_vec_const(ir))
 123       return false;
 124
 125    unsigned component = 0;
 126    for (int c = 0; c < ir->type->vector_elements; c++) {
 127       if (ir->get_float_component(c) < 1.0f)
 128          component++;
 129    }
 130
 131    return (component == ir->type->vector_elements);
 132 }
 133
 134 static inline bool
 135 is_greater_than_zero(ir_constant *ir)
 136 {
 137    if (!is_valid_vec_const(ir))
 138       return false;
 139
 140    unsigned component = 0;
 141    for (int c = 0; c < ir->type->vector_elements; c++) {
 142       if (ir->get_float_component(c) > 0.0f)
 143          component++;
 144    }
 145
 146    return (component == ir->type->vector_elements);
 147 }
 148
 149 static void
 150 update_type(ir_expression *ir)
 151 {
 152    if (ir->operands[0]->type->is_vector())
 153       ir->type = ir->operands[0]->type;
 154    else
 155       ir->type = ir->operands[1]->type;
 156 }
 157
 158 /* Recognize (v.x + v.y) + (v.z + v.w) as dot(v, 1.0) */
 159 static ir_expression *
 160 try_replace_with_dot(ir_expression *expr0, ir_expression *expr1, void *mem_ctx)
 161 {
 162    if (expr0 && expr0->operation == ir_binop_add &&
 163        expr0->type->is_float() &&
 164        expr1 && expr1->operation == ir_binop_add &&
 165        expr1->type->is_float()) {
 166       ir_swizzle *x = expr0->operands[0]->as_swizzle();
 167       ir_swizzle *y = expr0->operands[1]->as_swizzle();
 168       ir_swizzle *z = expr1->operands[0]->as_swizzle();
 169       ir_swizzle *w = expr1->operands[1]->as_swizzle();
 170
 171       if (!x || x->mask.num_components != 1 ||
 172           !y || y->mask.num_components != 1 ||
 173           !z || z->mask.num_components != 1 ||
 174           !w || w->mask.num_components != 1) {
 175          return NULL;
 176       }
 177
 178       bool swiz_seen[4] = {false, false, false, false};
 179       swiz_seen[x->mask.x] = true;
 180       swiz_seen[y->mask.x] = true;
 181       swiz_seen[z->mask.x] = true;
 182       swiz_seen[w->mask.x] = true;
 183
 184       if (!swiz_seen[0] || !swiz_seen[1] ||
 185           !swiz_seen[2] || !swiz_seen[3]) {
 186          return NULL;
 187       }
 188
 189       if (x->val->equals(y->val) &&
 190           x->val->equals(z->val) &&
 191           x->val->equals(w->val)) {
 192          return dot(x->val, new(mem_ctx) ir_constant(1.0f, 4));
 193       }
 194    }
 195    return NULL;
 196 }
 197
 198 void
 199 ir_algebraic_visitor::reassociate_operands(ir_expression *ir1,
 200                                            int op1,
 201                                            ir_expression *ir2,
 202                                            int op2)
 203 {
 204    ir_rvalue *temp = ir2->operands[op2];
 205    ir2->operands[op2] = ir1->operands[op1];
 206    ir1->operands[op1] = temp;
 207
 208    /* Update the type of ir2.  The type of ir1 won't have changed --
 209     * base types matched, and at least one of the operands of the 2
 210     * binops is still a vector if any of them were.
 211     */
 212    update_type(ir2);
 213
 214    this->progress = true;
 215 }
 216
 217 /**
 218  * Reassociates a constant down a tree of adds or multiplies.
 219  *
 220  * Consider (2 * (a * (b * 0.5))).  We want to send up with a * b.
 221  */
 222 bool
 223 ir_algebraic_visitor::reassociate_constant(ir_expression *ir1, int const_index,
 224                                            ir_constant *constant,
 225                                            ir_expression *ir2)
 226 {
 227    if (!ir2 || ir1->operation != ir2->operation)
 228       return false;
 229
 230    /* Don't want to even think about matrices. */
 231    if (ir1->operands[0]->type->is_matrix() ||
 232        ir1->operands[1]->type->is_matrix() ||
 233        ir2->operands[0]->type->is_matrix() ||
 234        ir2->operands[1]->type->is_matrix())
 235       return false;
 236
 237    ir_constant *ir2_const[2];
 238    ir2_const[0] = ir2->operands[0]->constant_expression_value();
 239    ir2_const[1] = ir2->operands[1]->constant_expression_value();
 240
 241    if (ir2_const[0] && ir2_const[1])
 242       return false;
 243
 244    if (ir2_const[0]) {
 245       reassociate_operands(ir1, const_index, ir2, 1);
 246       return true;
 247    } else if (ir2_const[1]) {
 248       reassociate_operands(ir1, const_index, ir2, 0);
 249       return true;
 250    }
 251
 252    if (reassociate_constant(ir1, const_index, constant,
 253                             ir2->operands[0]->as_expression())) {
 254       update_type(ir2);
 255       return true;
 256    }
 257
 258    if (reassociate_constant(ir1, const_index, constant,
 259                             ir2->operands[1]->as_expression())) {
 260       update_type(ir2);
 261       return true;
 262    }
 263
 264    return false;
 265 }
 266
 267 /* When eliminating an expression and just returning one of its operands,
 268  * we may need to swizzle that operand out to a vector if the expression was
 269  * vector type.
 270  */
 271 ir_rvalue *
 272 ir_algebraic_visitor::swizzle_if_required(ir_expression *expr,
 273                                           ir_rvalue *operand)
 274 {
 275    if (expr->type->is_vector() && operand->type->is_scalar()) {
 276       return new(mem_ctx) ir_swizzle(operand, 0, 0, 0, 0,
 277                                      expr->type->vector_elements);
 278    } else
 279       return operand;
 280 }
 281
 282 ir_rvalue *
 283 ir_algebraic_visitor::handle_expression(ir_expression *ir)
 284 {
 285    ir_constant *op_const[4] = {NULL, NULL, NULL, NULL};
 286    ir_expression *op_expr[4] = {NULL, NULL, NULL, NULL};
 287    unsigned int i;
 288
 289    assert(ir->get_num_operands() <= 4);
 290    for (i = 0; i < ir->get_num_operands(); i++) {
 291       if (ir->operands[i]->type->is_matrix())
 292          return ir;
 293
 294       op_const[i] = ir->operands[i]->constant_expression_value();
 295       op_expr[i] = ir->operands[i]->as_expression();
 296    }
 297
 298    if (this->mem_ctx == NULL)
 299       this->mem_ctx = ralloc_parent(ir);
 300
 301    switch (ir->operation) {
 302    case ir_unop_bit_not:
 303       if (op_expr[0] && op_expr[0]->operation == ir_unop_bit_not)
 304          return op_expr[0]->operands[0];
 305       break;
 306
 307    case ir_unop_abs:
 308       if (op_expr[0] == NULL)
 309          break;
 310
 311       switch (op_expr[0]->operation) {
 312       case ir_unop_abs:
 313       case ir_unop_neg:
 314          return abs(op_expr[0]->operands[0]);
 315       default:
 316          break;
 317       }
 318       break;
 319
 320    case ir_unop_neg:
 321       if (op_expr[0] == NULL)
 322          break;
 323
 324       if (op_expr[0]->operation == ir_unop_neg) {
 325          return op_expr[0]->operands[0];
 326       }
 327       break;
 328
 329    case ir_unop_exp:
 330       if (op_expr[0] == NULL)
 331          break;
 332
 333       if (op_expr[0]->operation == ir_unop_log) {
 334          return op_expr[0]->operands[0];
 335       }
 336       break;
 337
 338    case ir_unop_log:
 339       if (op_expr[0] == NULL)
 340          break;
 341
 342       if (op_expr[0]->operation == ir_unop_exp) {
 343          return op_expr[0]->operands[0];
 344       }
 345       break;
 346
 347    case ir_unop_exp2:
 348       if (op_expr[0] == NULL)
 349          break;
 350
 351       if (op_expr[0]->operation == ir_unop_log2) {
 352          return op_expr[0]->operands[0];
 353       }
 354
 355       if (!options->EmitNoPow && op_expr[0]->operation == ir_binop_mul) {
 356          for (int log2_pos = 0; log2_pos < 2; log2_pos++) {
 357             ir_expression *log2_expr =
 358                op_expr[0]->operands[log2_pos]->as_expression();
 359
 360             if (log2_expr && log2_expr->operation == ir_unop_log2) {
 361                return new(mem_ctx) ir_expression(ir_binop_pow,
 362                                                  ir->type,
 363                                                  log2_expr->operands[0],
 364                                                  op_expr[0]->operands[1 - log2_pos]);
 365             }
 366          }
 367       }
 368       break;
 369
 370    case ir_unop_log2:
 371       if (op_expr[0] == NULL)
 372          break;
 373
 374       if (op_expr[0]->operation == ir_unop_exp2) {
 375          return op_expr[0]->operands[0];
 376       }
 377       break;
 378
 379    case ir_unop_logic_not: {
 380       enum ir_expression_operation new_op = ir_unop_logic_not;
 381
 382       if (op_expr[0] == NULL)
 383          break;
 384
 385       switch (op_expr[0]->operation) {
 386       case ir_binop_less:    new_op = ir_binop_gequal;  break;
 387       case ir_binop_greater: new_op = ir_binop_lequal;  break;
 388       case ir_binop_lequal:  new_op = ir_binop_greater; break;
 389       case ir_binop_gequal:  new_op = ir_binop_less;    break;
 390       case ir_binop_equal:   new_op = ir_binop_nequal;  break;
 391       case ir_binop_nequal:  new_op = ir_binop_equal;   break;
 392       case ir_binop_all_equal:   new_op = ir_binop_any_nequal;  break;
 393       case ir_binop_any_nequal:  new_op = ir_binop_all_equal;   break;
 394
 395       default:
 396          /* The default case handler is here to silence a warning from GCC.
 397           */
 398          break;
 399       }
 400
 401       if (new_op != ir_unop_logic_not) {
 402          return new(mem_ctx) ir_expression(new_op,
 403                                            ir->type,
 404                                            op_expr[0]->operands[0],
 405                                            op_expr[0]->operands[1]);
 406       }
 407
 408       break;
 409    }
 410
 411    case ir_binop_add:
 412       if (is_vec_zero(op_const[0]))
 413          return ir->operands[1];
 414       if (is_vec_zero(op_const[1]))
 415          return ir->operands[0];
 416
 417       /* Reassociate addition of constants so that we can do constant
 418        * folding.
 419        */
 420       if (op_const[0] && !op_const[1])
 421          reassociate_constant(ir, 0, op_const[0], op_expr[1]);
 422       if (op_const[1] && !op_const[0])
 423          reassociate_constant(ir, 1, op_const[1], op_expr[0]);
 424
 425       /* Recognize (v.x + v.y) + (v.z + v.w) as dot(v, 1.0) */
 426       if (options->OptimizeForAOS) {
 427          ir_expression *expr = try_replace_with_dot(op_expr[0], op_expr[1],
 428                                                     mem_ctx);
 429          if (expr)
 430             return expr;
 431       }
 432
 433       /* Replace (-x + y) * a + x and commutative variations with lrp(x, y, a).
 434        *
 435        * (-x + y) * a + x
 436        * (x * -a) + (y * a) + x
 437        * x + (x * -a) + (y * a)
 438        * x * (1 - a) + y * a
 439        * lrp(x, y, a)
 440        */
 441       for (int mul_pos = 0; mul_pos < 2; mul_pos++) {
 442          ir_expression *mul = op_expr[mul_pos];
 443
 444          if (!mul || mul->operation != ir_binop_mul)
 445             continue;
 446
 447          /* Multiply found on one of the operands. Now check for an
 448           * inner addition operation.
 449           */
 450          for (int inner_add_pos = 0; inner_add_pos < 2; inner_add_pos++) {
 451             ir_expression *inner_add =
 452                mul->operands[inner_add_pos]->as_expression();
 453
 454             if (!inner_add || inner_add->operation != ir_binop_add)
 455                continue;
 456
 457             /* Inner addition found on one of the operands. Now check for
 458              * one of the operands of the inner addition to be the negative
 459              * of x_operand.
 460              */
 461             for (int neg_pos = 0; neg_pos < 2; neg_pos++) {
 462                ir_expression *neg =
 463                   inner_add->operands[neg_pos]->as_expression();
 464
 465                if (!neg || neg->operation != ir_unop_neg)
 466                   continue;
 467
 468                ir_rvalue *x_operand = ir->operands[1 - mul_pos];
 469
 470                if (!neg->operands[0]->equals(x_operand))
 471                   continue;
 472
 473                ir_rvalue *y_operand = inner_add->operands[1 - neg_pos];
 474                ir_rvalue *a_operand = mul->operands[1 - inner_add_pos];
 475
 476                if (x_operand->type != y_operand->type ||
 477                    x_operand->type != a_operand->type)
 478                   continue;
 479
 480                return lrp(x_operand, y_operand, a_operand);
 481             }
 482          }
 483       }
 484
 485       break;
 486
 487    case ir_binop_sub:
 488       if (is_vec_zero(op_const[0]))
 489          return neg(ir->operands[1]);
 490       if (is_vec_zero(op_const[1]))
 491          return ir->operands[0];
 492       break;
 493
 494    case ir_binop_mul:
 495       if (is_vec_one(op_const[0]))
 496          return ir->operands[1];
 497       if (is_vec_one(op_const[1]))
 498          return ir->operands[0];
 499
 500       if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1]))
 501          return ir_constant::zero(ir, ir->type);
 502
 503       if (is_vec_negative_one(op_const[0]))
 504          return neg(ir->operands[1]);
 505       if (is_vec_negative_one(op_const[1]))
 506          return neg(ir->operands[0]);
 507
 508
 509       /* Reassociate multiplication of constants so that we can do
 510        * constant folding.
 511        */
 512       if (op_const[0] && !op_const[1])
 513          reassociate_constant(ir, 0, op_const[0], op_expr[1]);
 514       if (op_const[1] && !op_const[0])
 515          reassociate_constant(ir, 1, op_const[1], op_expr[0]);
 516
 517       break;
 518
 519    case ir_binop_div:
 520       if (is_vec_one(op_const[0]) && ir->type->base_type == GLSL_TYPE_FLOAT) {
 521          return new(mem_ctx) ir_expression(ir_unop_rcp,
 522                                            ir->operands[1]->type,
 523                                            ir->operands[1],
 524                                            NULL);
 525       }
 526       if (is_vec_one(op_const[1]))
 527          return ir->operands[0];
 528       break;
 529
 530    case ir_binop_dot:
 531       if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1]))
 532          return ir_constant::zero(mem_ctx, ir->type);
 533
 534       for (int i = 0; i < 2; i++) {
 535          if (!op_const[i])
 536             continue;
 537
 538          unsigned components[4] = { 0 }, count = 0;
 539
 540          for (unsigned c = 0; c < op_const[i]->type->vector_elements; c++) {
 541             if (op_const[i]->value.f[c] == 0.0)
 542                continue;
 543
 544             components[count] = c;
 545             count++;
 546          }
 547
 548          /* No channels had zero values; bail. */
 549          if (count >= op_const[i]->type->vector_elements)
 550             break;
 551
 552          ir_expression_operation op = count == 1 ?
 553             ir_binop_mul : ir_binop_dot;
 554
 555          /* Swizzle both operands to remove the channels that were zero. */
 556          return new(mem_ctx)
 557             ir_expression(op, glsl_type::float_type,
 558                           new(mem_ctx) ir_swizzle(ir->operands[0],
 559                                                   components, count),
 560                           new(mem_ctx) ir_swizzle(ir->operands[1],
 561                                                   components, count));
 562       }
 563       break;
 564
 565    case ir_binop_less:
 566    case ir_binop_lequal:
 567    case ir_binop_greater:
 568    case ir_binop_gequal:
 569    case ir_binop_equal:
 570    case ir_binop_nequal:
 571       for (int add_pos = 0; add_pos < 2; add_pos++) {
 572          ir_expression *add = op_expr[add_pos];
 573
 574          if (!add || add->operation != ir_binop_add)
 575             continue;
 576
 577          ir_constant *zero = op_const[1 - add_pos];
 578          if (!is_vec_zero(zero))
 579             continue;
 580
 581          return new(mem_ctx) ir_expression(ir->operation,
 582                                            add->operands[0],
 583                                            neg(add->operands[1]));
 584       }
 585       break;
 586
 587    case ir_binop_all_equal:
 588    case ir_binop_any_nequal:
 589       if (ir->operands[0]->type->is_scalar() &&
 590           ir->operands[1]->type->is_scalar())
 591          return new(mem_ctx) ir_expression(ir->operation == ir_binop_all_equal
 592                                            ? ir_binop_equal : ir_binop_nequal,
 593                                            ir->operands[0],
 594                                            ir->operands[1]);
 595       break;
 596
 597    case ir_binop_rshift:
 598    case ir_binop_lshift:
 599       /* 0 >> x == 0 */
 600       if (is_vec_zero(op_const[0]))
 601          return ir->operands[0];
 602       /* x >> 0 == x */
 603       if (is_vec_zero(op_const[1]))
 604          return ir->operands[0];
 605       break;
 606
 607    case ir_binop_logic_and:
 608       if (is_vec_one(op_const[0])) {
 609          return ir->operands[1];
 610       } else if (is_vec_one(op_const[1])) {
 611          return ir->operands[0];
 612       } else if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1])) {
 613          return ir_constant::zero(mem_ctx, ir->type);
 614       } else if (op_expr[0] && op_expr[0]->operation == ir_unop_logic_not &&
 615                  op_expr[1] && op_expr[1]->operation == ir_unop_logic_not) {
 616          /* De Morgan's Law:
 617           *    (not A) and (not B) === not (A or B)
 618           */
 619          return logic_not(logic_or(op_expr[0]->operands[0],
 620                                    op_expr[1]->operands[0]));
 621       } else if (ir->operands[0]->equals(ir->operands[1])) {
 622          /* (a && a) == a */
 623          return ir->operands[0];
 624       }
 625       break;
 626
 627    case ir_binop_logic_xor:
 628       if (is_vec_zero(op_const[0])) {
 629          return ir->operands[1];
 630       } else if (is_vec_zero(op_const[1])) {
 631          return ir->operands[0];
 632       } else if (is_vec_one(op_const[0])) {
 633          return logic_not(ir->operands[1]);
 634       } else if (is_vec_one(op_const[1])) {
 635          return logic_not(ir->operands[0]);
 636       } else if (ir->operands[0]->equals(ir->operands[1])) {
 637          /* (a ^^ a) == false */
 638          return ir_constant::zero(mem_ctx, ir->type);
 639       }
 640       break;
 641
 642    case ir_binop_logic_or:
 643       if (is_vec_zero(op_const[0])) {
 644          return ir->operands[1];
 645       } else if (is_vec_zero(op_const[1])) {
 646          return ir->operands[0];
 647       } else if (is_vec_one(op_const[0]) || is_vec_one(op_const[1])) {
 648          ir_constant_data data;
 649
 650          for (unsigned i = 0; i < 16; i++)
 651             data.b[i] = true;
 652
 653          return new(mem_ctx) ir_constant(ir->type, &data);
 654       } else if (op_expr[0] && op_expr[0]->operation == ir_unop_logic_not &&
 655                  op_expr[1] && op_expr[1]->operation == ir_unop_logic_not) {
 656          /* De Morgan's Law:
 657           *    (not A) or (not B) === not (A and B)
 658           */
 659          return logic_not(logic_and(op_expr[0]->operands[0],
 660                                     op_expr[1]->operands[0]));
 661       } else if (ir->operands[0]->equals(ir->operands[1])) {
 662          /* (a || a) == a */
 663          return ir->operands[0];
 664       }
 665       break;
 666
 667    case ir_binop_pow:
 668       /* 1^x == 1 */
 669       if (is_vec_one(op_const[0]))
 670          return op_const[0];
 671
 672       /* x^1 == x */
 673       if (is_vec_one(op_const[1]))
 674          return ir->operands[0];
 675
 676       /* pow(2,x) == exp2(x) */
 677       if (is_vec_two(op_const[0]))
 678          return expr(ir_unop_exp2, ir->operands[1]);
 679
 680       if (is_vec_two(op_const[1])) {
 681          ir_variable *x = new(ir) ir_variable(ir->operands[1]->type, "x",
 682                                               ir_var_temporary);
 683          base_ir->insert_before(x);
 684          base_ir->insert_before(assign(x, ir->operands[0]));
 685          return mul(x, x);
 686       }
 687
 688       break;
 689
 690    case ir_binop_min:
 691    case ir_binop_max:
 692       if (ir->type->base_type != GLSL_TYPE_FLOAT || options->EmitNoSat)
 693          break;
 694
 695       /* Replace min(max) operations and its commutative combinations with
 696        * a saturate operation
 697        */
 698       for (int op = 0; op < 2; op++) {
 699          ir_expression *minmax = op_expr[op];
 700          ir_constant *outer_const = op_const[1 - op];
 701          ir_expression_operation op_cond = (ir->operation == ir_binop_max) ?
 702             ir_binop_min : ir_binop_max;
 703
 704          if (!minmax || !outer_const || (minmax->operation != op_cond))
 705             continue;
 706
 707          /* Found a min(max) combination. Now try to see if its operands
 708           * meet our conditions that we can do just a single saturate operation
 709           */
 710          for (int minmax_op = 0; minmax_op < 2; minmax_op++) {
 711             ir_rvalue *inner_val_a = minmax->operands[minmax_op];
 712             ir_rvalue *inner_val_b = minmax->operands[1 - minmax_op];
 713
 714             if (!inner_val_a || !inner_val_b)
 715                continue;
 716
 717             /* Found a {min|max} ({max|min} (x, 0.0), 1.0) operation and its variations */
 718             if ((outer_const->is_one() && inner_val_a->is_zero()) ||
 719                 (inner_val_a->is_one() && outer_const->is_zero()))
 720                return saturate(inner_val_b);
 721
 722             /* Found a {min|max} ({max|min} (x, 0.0), b) where b < 1.0
 723              * and its variations
 724              */
 725             if (is_less_than_one(outer_const) && inner_val_b->is_zero())
 726                return expr(ir_binop_min, saturate(inner_val_a), outer_const);
 727
 728             if (!inner_val_b->as_constant())
 729                continue;
 730
 731             if (is_less_than_one(inner_val_b->as_constant()) && outer_const->is_zero())
 732                return expr(ir_binop_min, saturate(inner_val_a), inner_val_b);
 733
 734             /* Found a {min|max} ({max|min} (x, b), 1.0), where b > 0.0
 735              * and its variations
 736              */
 737             if (outer_const->is_one() && is_greater_than_zero(inner_val_b->as_constant()))
 738                return expr(ir_binop_max, saturate(inner_val_a), inner_val_b);
 739             if (inner_val_b->as_constant()->is_one() && is_greater_than_zero(outer_const))
 740                return expr(ir_binop_max, saturate(inner_val_a), outer_const);
 741          }
 742       }
 743
 744       break;
 745
 746    case ir_unop_rcp:
 747       if (op_expr[0] && op_expr[0]->operation == ir_unop_rcp)
 748          return op_expr[0]->operands[0];
 749
 750       /* While ir_to_mesa.cpp will lower sqrt(x) to rcp(rsq(x)), it does so at
 751        * its IR level, so we can always apply this transformation.
 752        */
 753       if (op_expr[0] && op_expr[0]->operation == ir_unop_rsq)
 754          return sqrt(op_expr[0]->operands[0]);
 755
 756       /* As far as we know, all backends are OK with rsq. */
 757       if (op_expr[0] && op_expr[0]->operation == ir_unop_sqrt) {
 758          return rsq(op_expr[0]->operands[0]);
 759       }
 760
 761       break;
 762
 763    case ir_triop_fma:
 764       /* Operands are op0 * op1 + op2. */
 765       if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1])) {
 766          return ir->operands[2];
 767       } else if (is_vec_zero(op_const[2])) {
 768          return mul(ir->operands[0], ir->operands[1]);
 769       } else if (is_vec_one(op_const[0])) {
 770          return add(ir->operands[1], ir->operands[2]);
 771       } else if (is_vec_one(op_const[1])) {
 772          return add(ir->operands[0], ir->operands[2]);
 773       }
 774       break;
 775
 776    case ir_triop_lrp:
 777       /* Operands are (x, y, a). */
 778       if (is_vec_zero(op_const[2])) {
 779          return ir->operands[0];
 780       } else if (is_vec_one(op_const[2])) {
 781          return ir->operands[1];
 782       } else if (ir->operands[0]->equals(ir->operands[1])) {
 783          return ir->operands[0];
 784       } else if (is_vec_zero(op_const[0])) {
 785          return mul(ir->operands[1], ir->operands[2]);
 786       } else if (is_vec_zero(op_const[1])) {
 787          unsigned op2_components = ir->operands[2]->type->vector_elements;
 788          ir_constant *one = new(mem_ctx) ir_constant(1.0f, op2_components);
 789          return mul(ir->operands[0], add(one, neg(ir->operands[2])));
 790       }
 791       break;
 792
 793    case ir_triop_csel:
 794       if (is_vec_one(op_const[0]))
 795          return ir->operands[1];
 796       if (is_vec_zero(op_const[0]))
 797          return ir->operands[2];
 798       break;
 799
 800    default:
 801       break;
 802    }
 803
 804    return ir;
 805 }
 806
 807 void
 808 ir_algebraic_visitor::handle_rvalue(ir_rvalue **rvalue)
 809 {
 810    if (!*rvalue)
 811       return;
 812
 813    ir_expression *expr = (*rvalue)->as_expression();
 814    if (!expr || expr->operation == ir_quadop_vector)
 815       return;
 816
 817    ir_rvalue *new_rvalue = handle_expression(expr);
 818    if (new_rvalue == *rvalue)
 819       return;
 820
 821    /* If the expr used to be some vec OP scalar returning a vector, and the
 822     * optimization gave us back a scalar, we still need to turn it into a
 823     * vector.
 824     */
 825    *rvalue = swizzle_if_required(expr, new_rvalue);
 826
 827    this->progress = true;
 828 }
 829
 830 bool
 831 do_algebraic(exec_list *instructions, bool native_integers,
 832              const struct gl_shader_compiler_options *options)
 833 {
 834    ir_algebraic_visitor v(native_integers, options);
 835
 836    visit_list_elements(&v, instructions);
 837
 838    return v.progress;
 839 }