src/compiler/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    virtual ir_visitor_status visit_enter(ir_assignment *ir);
  62
  63    ir_rvalue *handle_expression(ir_expression *ir);
  64    void handle_rvalue(ir_rvalue **rvalue);
  65    bool reassociate_constant(ir_expression *ir1,
  66                              int const_index,
  67                              ir_constant *constant,
  68                              ir_expression *ir2);
  69    void reassociate_operands(ir_expression *ir1,
  70                              int op1,
  71                              ir_expression *ir2,
  72                              int op2);
  73    ir_rvalue *swizzle_if_required(ir_expression *expr,
  74                                   ir_rvalue *operand);
  75
  76    const struct gl_shader_compiler_options *options;
  77    void *mem_ctx;
  78
  79    bool native_integers;
  80    bool progress;
  81 };
  82
  83 } /* unnamed namespace */
  84
  85 ir_visitor_status
  86 ir_algebraic_visitor::visit_enter(ir_assignment *ir)
  87 {
  88    ir_variable *var = ir->lhs->variable_referenced();
  89    if (var->data.invariant || var->data.precise) {
  90       /* If we're assigning to an invariant or precise variable, just bail.
  91        * Most of the algebraic optimizations aren't precision-safe.
  92        *
  93        * FINISHME: Find out which optimizations are precision-safe and enable
  94        * then only for invariant or precise trees.
  95        */
  96       return visit_continue_with_parent;
  97    } else {
  98       return visit_continue;
  99    }
 100 }
 101
 102 static inline bool
 103 is_vec_zero(ir_constant *ir)
 104 {
 105    return (ir == NULL) ? false : ir->is_zero();
 106 }
 107
 108 static inline bool
 109 is_vec_one(ir_constant *ir)
 110 {
 111    return (ir == NULL) ? false : ir->is_one();
 112 }
 113
 114 static inline bool
 115 is_vec_two(ir_constant *ir)
 116 {
 117    return (ir == NULL) ? false : ir->is_value(2.0, 2);
 118 }
 119
 120 static inline bool
 121 is_vec_four(ir_constant *ir)
 122 {
 123    return (ir == NULL) ? false : ir->is_value(4.0, 4);
 124 }
 125
 126 static inline bool
 127 is_vec_negative_one(ir_constant *ir)
 128 {
 129    return (ir == NULL) ? false : ir->is_negative_one();
 130 }
 131
 132 static inline bool
 133 is_valid_vec_const(ir_constant *ir)
 134 {
 135    if (ir == NULL)
 136       return false;
 137
 138    if (!ir->type->is_scalar() && !ir->type->is_vector())
 139       return false;
 140
 141    return true;
 142 }
 143
 144 static inline bool
 145 is_less_than_one(ir_constant *ir)
 146 {
 147    assert(ir->type->base_type == GLSL_TYPE_FLOAT);
 148
 149    if (!is_valid_vec_const(ir))
 150       return false;
 151
 152    unsigned component = 0;
 153    for (int c = 0; c < ir->type->vector_elements; c++) {
 154       if (ir->get_float_component(c) < 1.0f)
 155          component++;
 156    }
 157
 158    return (component == ir->type->vector_elements);
 159 }
 160
 161 static inline bool
 162 is_greater_than_zero(ir_constant *ir)
 163 {
 164    assert(ir->type->base_type == GLSL_TYPE_FLOAT);
 165
 166    if (!is_valid_vec_const(ir))
 167       return false;
 168
 169    unsigned component = 0;
 170    for (int c = 0; c < ir->type->vector_elements; c++) {
 171       if (ir->get_float_component(c) > 0.0f)
 172          component++;
 173    }
 174
 175    return (component == ir->type->vector_elements);
 176 }
 177
 178 static void
 179 update_type(ir_expression *ir)
 180 {
 181    if (ir->operands[0]->type->is_vector())
 182       ir->type = ir->operands[0]->type;
 183    else
 184       ir->type = ir->operands[1]->type;
 185 }
 186
 187 /* Recognize (v.x + v.y) + (v.z + v.w) as dot(v, 1.0) */
 188 static ir_expression *
 189 try_replace_with_dot(ir_expression *expr0, ir_expression *expr1, void *mem_ctx)
 190 {
 191    if (expr0 && expr0->operation == ir_binop_add &&
 192        expr0->type->is_float() &&
 193        expr1 && expr1->operation == ir_binop_add &&
 194        expr1->type->is_float()) {
 195       ir_swizzle *x = expr0->operands[0]->as_swizzle();
 196       ir_swizzle *y = expr0->operands[1]->as_swizzle();
 197       ir_swizzle *z = expr1->operands[0]->as_swizzle();
 198       ir_swizzle *w = expr1->operands[1]->as_swizzle();
 199
 200       if (!x || x->mask.num_components != 1 ||
 201           !y || y->mask.num_components != 1 ||
 202           !z || z->mask.num_components != 1 ||
 203           !w || w->mask.num_components != 1) {
 204          return NULL;
 205       }
 206
 207       bool swiz_seen[4] = {false, false, false, false};
 208       swiz_seen[x->mask.x] = true;
 209       swiz_seen[y->mask.x] = true;
 210       swiz_seen[z->mask.x] = true;
 211       swiz_seen[w->mask.x] = true;
 212
 213       if (!swiz_seen[0] || !swiz_seen[1] ||
 214           !swiz_seen[2] || !swiz_seen[3]) {
 215          return NULL;
 216       }
 217
 218       if (x->val->equals(y->val) &&
 219           x->val->equals(z->val) &&
 220           x->val->equals(w->val)) {
 221          return dot(x->val, new(mem_ctx) ir_constant(1.0f, 4));
 222       }
 223    }
 224    return NULL;
 225 }
 226
 227 void
 228 ir_algebraic_visitor::reassociate_operands(ir_expression *ir1,
 229                                            int op1,
 230                                            ir_expression *ir2,
 231                                            int op2)
 232 {
 233    ir_rvalue *temp = ir2->operands[op2];
 234    ir2->operands[op2] = ir1->operands[op1];
 235    ir1->operands[op1] = temp;
 236
 237    /* Update the type of ir2.  The type of ir1 won't have changed --
 238     * base types matched, and at least one of the operands of the 2
 239     * binops is still a vector if any of them were.
 240     */
 241    update_type(ir2);
 242
 243    this->progress = true;
 244 }
 245
 246 /**
 247  * Reassociates a constant down a tree of adds or multiplies.
 248  *
 249  * Consider (2 * (a * (b * 0.5))).  We want to send up with a * b.
 250  */
 251 bool
 252 ir_algebraic_visitor::reassociate_constant(ir_expression *ir1, int const_index,
 253                                            ir_constant *constant,
 254                                            ir_expression *ir2)
 255 {
 256    if (!ir2 || ir1->operation != ir2->operation)
 257       return false;
 258
 259    /* Don't want to even think about matrices. */
 260    if (ir1->operands[0]->type->is_matrix() ||
 261        ir1->operands[1]->type->is_matrix() ||
 262        ir2->operands[0]->type->is_matrix() ||
 263        ir2->operands[1]->type->is_matrix())
 264       return false;
 265
 266    ir_constant *ir2_const[2];
 267    ir2_const[0] = ir2->operands[0]->constant_expression_value();
 268    ir2_const[1] = ir2->operands[1]->constant_expression_value();
 269
 270    if (ir2_const[0] && ir2_const[1])
 271       return false;
 272
 273    if (ir2_const[0]) {
 274       reassociate_operands(ir1, const_index, ir2, 1);
 275       return true;
 276    } else if (ir2_const[1]) {
 277       reassociate_operands(ir1, const_index, ir2, 0);
 278       return true;
 279    }
 280
 281    if (reassociate_constant(ir1, const_index, constant,
 282                             ir2->operands[0]->as_expression())) {
 283       update_type(ir2);
 284       return true;
 285    }
 286
 287    if (reassociate_constant(ir1, const_index, constant,
 288                             ir2->operands[1]->as_expression())) {
 289       update_type(ir2);
 290       return true;
 291    }
 292
 293    return false;
 294 }
 295
 296 /* When eliminating an expression and just returning one of its operands,
 297  * we may need to swizzle that operand out to a vector if the expression was
 298  * vector type.
 299  */
 300 ir_rvalue *
 301 ir_algebraic_visitor::swizzle_if_required(ir_expression *expr,
 302                                           ir_rvalue *operand)
 303 {
 304    if (expr->type->is_vector() && operand->type->is_scalar()) {
 305       return new(mem_ctx) ir_swizzle(operand, 0, 0, 0, 0,
 306                                      expr->type->vector_elements);
 307    } else
 308       return operand;
 309 }
 310
 311 ir_rvalue *
 312 ir_algebraic_visitor::handle_expression(ir_expression *ir)
 313 {
 314    ir_constant *op_const[4] = {NULL, NULL, NULL, NULL};
 315    ir_expression *op_expr[4] = {NULL, NULL, NULL, NULL};
 316    unsigned int i;
 317
 318    if (ir->operation == ir_binop_mul &&
 319        ir->operands[0]->type->is_matrix() &&
 320        ir->operands[1]->type->is_vector()) {
 321       ir_expression *matrix_mul = ir->operands[0]->as_expression();
 322
 323       if (matrix_mul && matrix_mul->operation == ir_binop_mul &&
 324          matrix_mul->operands[0]->type->is_matrix() &&
 325          matrix_mul->operands[1]->type->is_matrix()) {
 326
 327          return mul(matrix_mul->operands[0],
 328                     mul(matrix_mul->operands[1], ir->operands[1]));
 329       }
 330    }
 331
 332    assert(ir->get_num_operands() <= 4);
 333    for (i = 0; i < ir->get_num_operands(); i++) {
 334       if (ir->operands[i]->type->is_matrix())
 335          return ir;
 336
 337       op_const[i] = ir->operands[i]->constant_expression_value();
 338       op_expr[i] = ir->operands[i]->as_expression();
 339    }
 340
 341    if (this->mem_ctx == NULL)
 342       this->mem_ctx = ralloc_parent(ir);
 343
 344    switch (ir->operation) {
 345    case ir_unop_bit_not:
 346       if (op_expr[0] && op_expr[0]->operation == ir_unop_bit_not)
 347          return op_expr[0]->operands[0];
 348       break;
 349
 350    case ir_unop_abs:
 351       if (op_expr[0] == NULL)
 352          break;
 353
 354       switch (op_expr[0]->operation) {
 355       case ir_unop_abs:
 356       case ir_unop_neg:
 357          return abs(op_expr[0]->operands[0]);
 358       default:
 359          break;
 360       }
 361       break;
 362
 363    case ir_unop_neg:
 364       if (op_expr[0] == NULL)
 365          break;
 366
 367       if (op_expr[0]->operation == ir_unop_neg) {
 368          return op_expr[0]->operands[0];
 369       }
 370       break;
 371
 372    case ir_unop_exp:
 373       if (op_expr[0] == NULL)
 374          break;
 375
 376       if (op_expr[0]->operation == ir_unop_log) {
 377          return op_expr[0]->operands[0];
 378       }
 379       break;
 380
 381    case ir_unop_log:
 382       if (op_expr[0] == NULL)
 383          break;
 384
 385       if (op_expr[0]->operation == ir_unop_exp) {
 386          return op_expr[0]->operands[0];
 387       }
 388       break;
 389
 390    case ir_unop_exp2:
 391       if (op_expr[0] == NULL)
 392          break;
 393
 394       if (op_expr[0]->operation == ir_unop_log2) {
 395          return op_expr[0]->operands[0];
 396       }
 397
 398       if (!options->EmitNoPow && op_expr[0]->operation == ir_binop_mul) {
 399          for (int log2_pos = 0; log2_pos < 2; log2_pos++) {
 400             ir_expression *log2_expr =
 401                op_expr[0]->operands[log2_pos]->as_expression();
 402
 403             if (log2_expr && log2_expr->operation == ir_unop_log2) {
 404                return new(mem_ctx) ir_expression(ir_binop_pow,
 405                                                  ir->type,
 406                                                  log2_expr->operands[0],
 407                                                  op_expr[0]->operands[1 - log2_pos]);
 408             }
 409          }
 410       }
 411       break;
 412
 413    case ir_unop_log2:
 414       if (op_expr[0] == NULL)
 415          break;
 416
 417       if (op_expr[0]->operation == ir_unop_exp2) {
 418          return op_expr[0]->operands[0];
 419       }
 420       break;
 421
 422    case ir_unop_f2i:
 423    case ir_unop_f2u:
 424       if (op_expr[0] && op_expr[0]->operation == ir_unop_trunc) {
 425          return new(mem_ctx) ir_expression(ir->operation,
 426                                            ir->type,
 427                                            op_expr[0]->operands[0]);
 428       }
 429       break;
 430
 431    case ir_unop_logic_not: {
 432       enum ir_expression_operation new_op = ir_unop_logic_not;
 433
 434       if (op_expr[0] == NULL)
 435          break;
 436
 437       switch (op_expr[0]->operation) {
 438       case ir_binop_less:    new_op = ir_binop_gequal;  break;
 439       case ir_binop_greater: new_op = ir_binop_lequal;  break;
 440       case ir_binop_lequal:  new_op = ir_binop_greater; break;
 441       case ir_binop_gequal:  new_op = ir_binop_less;    break;
 442       case ir_binop_equal:   new_op = ir_binop_nequal;  break;
 443       case ir_binop_nequal:  new_op = ir_binop_equal;   break;
 444       case ir_binop_all_equal:   new_op = ir_binop_any_nequal;  break;
 445       case ir_binop_any_nequal:  new_op = ir_binop_all_equal;   break;
 446
 447       default:
 448          /* The default case handler is here to silence a warning from GCC.
 449           */
 450          break;
 451       }
 452
 453       if (new_op != ir_unop_logic_not) {
 454          return new(mem_ctx) ir_expression(new_op,
 455                                            ir->type,
 456                                            op_expr[0]->operands[0],
 457                                            op_expr[0]->operands[1]);
 458       }
 459
 460       break;
 461    }
 462
 463    case ir_unop_saturate:
 464       if (op_expr[0] && op_expr[0]->operation == ir_binop_add) {
 465          ir_expression *b2f_0 = op_expr[0]->operands[0]->as_expression();
 466          ir_expression *b2f_1 = op_expr[0]->operands[1]->as_expression();
 467
 468          if (b2f_0 && b2f_0->operation == ir_unop_b2f &&
 469              b2f_1 && b2f_1->operation == ir_unop_b2f) {
 470             return b2f(logic_or(b2f_0->operands[0], b2f_1->operands[0]));
 471          }
 472       }
 473       break;
 474
 475       /* This macro CANNOT use the do { } while(true) mechanism because
 476        * then the breaks apply to the loop instead of the switch!
 477        */
 478 #define HANDLE_PACK_UNPACK_INVERSE(inverse_operation)                   \
 479       {                                                                 \
 480          ir_expression *const op = ir->operands[0]->as_expression();    \
 481          if (op == NULL)                                                \
 482             break;                                                      \
 483          if (op->operation == (inverse_operation))                      \
 484             return op->operands[0];                                     \
 485          break;                                                         \
 486       }
 487
 488    case ir_unop_unpack_uint_2x32:
 489       HANDLE_PACK_UNPACK_INVERSE(ir_unop_pack_uint_2x32);
 490    case ir_unop_pack_uint_2x32:
 491       HANDLE_PACK_UNPACK_INVERSE(ir_unop_unpack_uint_2x32);
 492    case ir_unop_unpack_int_2x32:
 493       HANDLE_PACK_UNPACK_INVERSE(ir_unop_pack_int_2x32);
 494    case ir_unop_pack_int_2x32:
 495       HANDLE_PACK_UNPACK_INVERSE(ir_unop_unpack_int_2x32);
 496    case ir_unop_unpack_double_2x32:
 497       HANDLE_PACK_UNPACK_INVERSE(ir_unop_pack_double_2x32);
 498    case ir_unop_pack_double_2x32:
 499       HANDLE_PACK_UNPACK_INVERSE(ir_unop_unpack_double_2x32);
 500
 501 #undef HANDLE_PACK_UNPACK_INVERSE
 502
 503    case ir_binop_add:
 504       if (is_vec_zero(op_const[0]))
 505          return ir->operands[1];
 506       if (is_vec_zero(op_const[1]))
 507          return ir->operands[0];
 508
 509       /* Reassociate addition of constants so that we can do constant
 510        * 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       /* Recognize (v.x + v.y) + (v.z + v.w) as dot(v, 1.0) */
 518       if (options->OptimizeForAOS) {
 519          ir_expression *expr = try_replace_with_dot(op_expr[0], op_expr[1],
 520                                                     mem_ctx);
 521          if (expr)
 522             return expr;
 523       }
 524
 525       /* Replace (-x + y) * a + x and commutative variations with lrp(x, y, a).
 526        *
 527        * (-x + y) * a + x
 528        * (x * -a) + (y * a) + x
 529        * x + (x * -a) + (y * a)
 530        * x * (1 - a) + y * a
 531        * lrp(x, y, a)
 532        */
 533       for (int mul_pos = 0; mul_pos < 2; mul_pos++) {
 534          ir_expression *mul = op_expr[mul_pos];
 535
 536          if (!mul || mul->operation != ir_binop_mul)
 537             continue;
 538
 539          /* Multiply found on one of the operands. Now check for an
 540           * inner addition operation.
 541           */
 542          for (int inner_add_pos = 0; inner_add_pos < 2; inner_add_pos++) {
 543             ir_expression *inner_add =
 544                mul->operands[inner_add_pos]->as_expression();
 545
 546             if (!inner_add || inner_add->operation != ir_binop_add)
 547                continue;
 548
 549             /* Inner addition found on one of the operands. Now check for
 550              * one of the operands of the inner addition to be the negative
 551              * of x_operand.
 552              */
 553             for (int neg_pos = 0; neg_pos < 2; neg_pos++) {
 554                ir_expression *neg =
 555                   inner_add->operands[neg_pos]->as_expression();
 556
 557                if (!neg || neg->operation != ir_unop_neg)
 558                   continue;
 559
 560                ir_rvalue *x_operand = ir->operands[1 - mul_pos];
 561
 562                if (!neg->operands[0]->equals(x_operand))
 563                   continue;
 564
 565                ir_rvalue *y_operand = inner_add->operands[1 - neg_pos];
 566                ir_rvalue *a_operand = mul->operands[1 - inner_add_pos];
 567
 568                if (x_operand->type != y_operand->type ||
 569                    x_operand->type != a_operand->type)
 570                   continue;
 571
 572                return lrp(x_operand, y_operand, a_operand);
 573             }
 574          }
 575       }
 576
 577       break;
 578
 579    case ir_binop_sub:
 580       if (is_vec_zero(op_const[0]))
 581          return neg(ir->operands[1]);
 582       if (is_vec_zero(op_const[1]))
 583          return ir->operands[0];
 584       break;
 585
 586    case ir_binop_mul:
 587       if (is_vec_one(op_const[0]))
 588          return ir->operands[1];
 589       if (is_vec_one(op_const[1]))
 590          return ir->operands[0];
 591
 592       if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1]))
 593          return ir_constant::zero(ir, ir->type);
 594
 595       if (is_vec_negative_one(op_const[0]))
 596          return neg(ir->operands[1]);
 597       if (is_vec_negative_one(op_const[1]))
 598          return neg(ir->operands[0]);
 599
 600       if (op_expr[0] && op_expr[0]->operation == ir_unop_b2f &&
 601           op_expr[1] && op_expr[1]->operation == ir_unop_b2f) {
 602          return b2f(logic_and(op_expr[0]->operands[0], op_expr[1]->operands[0]));
 603       }
 604
 605       /* Reassociate multiplication of constants so that we can do
 606        * constant folding.
 607        */
 608       if (op_const[0] && !op_const[1])
 609          reassociate_constant(ir, 0, op_const[0], op_expr[1]);
 610       if (op_const[1] && !op_const[0])
 611          reassociate_constant(ir, 1, op_const[1], op_expr[0]);
 612
 613       /* Optimizes
 614        *
 615        *    (mul (floor (add (abs x) 0.5) (sign x)))
 616        *
 617        * into
 618        *
 619        *    (trunc (add x (mul (sign x) 0.5)))
 620        */
 621       for (int i = 0; i < 2; i++) {
 622          ir_expression *sign_expr = ir->operands[i]->as_expression();
 623          ir_expression *floor_expr = ir->operands[1 - i]->as_expression();
 624
 625          if (!sign_expr || sign_expr->operation != ir_unop_sign ||
 626              !floor_expr || floor_expr->operation != ir_unop_floor)
 627             continue;
 628
 629          ir_expression *add_expr = floor_expr->operands[0]->as_expression();
 630          if (!add_expr || add_expr->operation != ir_binop_add)
 631             continue;
 632
 633          for (int j = 0; j < 2; j++) {
 634             ir_expression *abs_expr = add_expr->operands[j]->as_expression();
 635             if (!abs_expr || abs_expr->operation != ir_unop_abs)
 636                continue;
 637
 638             ir_constant *point_five = add_expr->operands[1 - j]->as_constant();
 639             if (!point_five || !point_five->is_value(0.5, 0))
 640                continue;
 641
 642             if (abs_expr->operands[0]->equals(sign_expr->operands[0])) {
 643                return trunc(add(abs_expr->operands[0],
 644                                 mul(sign_expr, point_five)));
 645             }
 646          }
 647       }
 648       break;
 649
 650    case ir_binop_div:
 651       if (is_vec_one(op_const[0]) && (
 652                 ir->type->base_type == GLSL_TYPE_FLOAT ||
 653                 ir->type->is_double())) {
 654          return new(mem_ctx) ir_expression(ir_unop_rcp,
 655                                            ir->operands[1]->type,
 656                                            ir->operands[1],
 657                                            NULL);
 658       }
 659       if (is_vec_one(op_const[1]))
 660          return ir->operands[0];
 661       break;
 662
 663    case ir_binop_dot:
 664       if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1]))
 665          return ir_constant::zero(mem_ctx, ir->type);
 666
 667       for (int i = 0; i < 2; i++) {
 668          if (!op_const[i])
 669             continue;
 670
 671          unsigned components[4] = { 0 }, count = 0;
 672
 673          for (unsigned c = 0; c < op_const[i]->type->vector_elements; c++) {
 674             if (op_const[i]->is_zero())
 675                continue;
 676
 677             components[count] = c;
 678             count++;
 679          }
 680
 681          /* No channels had zero values; bail. */
 682          if (count >= op_const[i]->type->vector_elements)
 683             break;
 684
 685          ir_expression_operation op = count == 1 ?
 686             ir_binop_mul : ir_binop_dot;
 687
 688          /* Swizzle both operands to remove the channels that were zero. */
 689          return new(mem_ctx)
 690             ir_expression(op, ir->type,
 691                           new(mem_ctx) ir_swizzle(ir->operands[0],
 692                                                   components, count),
 693                           new(mem_ctx) ir_swizzle(ir->operands[1],
 694                                                   components, count));
 695       }
 696       break;
 697
 698    case ir_binop_less:
 699    case ir_binop_lequal:
 700    case ir_binop_greater:
 701    case ir_binop_gequal:
 702    case ir_binop_equal:
 703    case ir_binop_nequal:
 704       for (int add_pos = 0; add_pos < 2; add_pos++) {
 705          ir_expression *add = op_expr[add_pos];
 706
 707          if (!add || add->operation != ir_binop_add)
 708             continue;
 709
 710          ir_constant *zero = op_const[1 - add_pos];
 711          if (!is_vec_zero(zero))
 712             continue;
 713
 714          /* Depending of the zero position we want to optimize
 715           * (0 cmp x+y) into (-x cmp y) or (x+y cmp 0) into (x cmp -y)
 716           */
 717          if (add_pos == 1) {
 718             return new(mem_ctx) ir_expression(ir->operation,
 719                                               neg(add->operands[0]),
 720                                               add->operands[1]);
 721          } else {
 722             return new(mem_ctx) ir_expression(ir->operation,
 723                                               add->operands[0],
 724                                               neg(add->operands[1]));
 725          }
 726       }
 727       break;
 728
 729    case ir_binop_all_equal:
 730    case ir_binop_any_nequal:
 731       if (ir->operands[0]->type->is_scalar() &&
 732           ir->operands[1]->type->is_scalar())
 733          return new(mem_ctx) ir_expression(ir->operation == ir_binop_all_equal
 734                                            ? ir_binop_equal : ir_binop_nequal,
 735                                            ir->operands[0],
 736                                            ir->operands[1]);
 737       break;
 738
 739    case ir_binop_rshift:
 740    case ir_binop_lshift:
 741       /* 0 >> x == 0 */
 742       if (is_vec_zero(op_const[0]))
 743          return ir->operands[0];
 744       /* x >> 0 == x */
 745       if (is_vec_zero(op_const[1]))
 746          return ir->operands[0];
 747       break;
 748
 749    case ir_binop_logic_and:
 750       if (is_vec_one(op_const[0])) {
 751          return ir->operands[1];
 752       } else if (is_vec_one(op_const[1])) {
 753          return ir->operands[0];
 754       } else if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1])) {
 755          return ir_constant::zero(mem_ctx, ir->type);
 756       } else if (op_expr[0] && op_expr[0]->operation == ir_unop_logic_not &&
 757                  op_expr[1] && op_expr[1]->operation == ir_unop_logic_not) {
 758          /* De Morgan's Law:
 759           *    (not A) and (not B) === not (A or B)
 760           */
 761          return logic_not(logic_or(op_expr[0]->operands[0],
 762                                    op_expr[1]->operands[0]));
 763       } else if (ir->operands[0]->equals(ir->operands[1])) {
 764          /* (a && a) == a */
 765          return ir->operands[0];
 766       }
 767       break;
 768
 769    case ir_binop_logic_xor:
 770       if (is_vec_zero(op_const[0])) {
 771          return ir->operands[1];
 772       } else if (is_vec_zero(op_const[1])) {
 773          return ir->operands[0];
 774       } else if (is_vec_one(op_const[0])) {
 775          return logic_not(ir->operands[1]);
 776       } else if (is_vec_one(op_const[1])) {
 777          return logic_not(ir->operands[0]);
 778       } else if (ir->operands[0]->equals(ir->operands[1])) {
 779          /* (a ^^ a) == false */
 780          return ir_constant::zero(mem_ctx, ir->type);
 781       }
 782       break;
 783
 784    case ir_binop_logic_or:
 785       if (is_vec_zero(op_const[0])) {
 786          return ir->operands[1];
 787       } else if (is_vec_zero(op_const[1])) {
 788          return ir->operands[0];
 789       } else if (is_vec_one(op_const[0]) || is_vec_one(op_const[1])) {
 790          ir_constant_data data;
 791
 792          for (unsigned i = 0; i < 16; i++)
 793             data.b[i] = true;
 794
 795          return new(mem_ctx) ir_constant(ir->type, &data);
 796       } else if (op_expr[0] && op_expr[0]->operation == ir_unop_logic_not &&
 797                  op_expr[1] && op_expr[1]->operation == ir_unop_logic_not) {
 798          /* De Morgan's Law:
 799           *    (not A) or (not B) === not (A and B)
 800           */
 801          return logic_not(logic_and(op_expr[0]->operands[0],
 802                                     op_expr[1]->operands[0]));
 803       } else if (ir->operands[0]->equals(ir->operands[1])) {
 804          /* (a || a) == a */
 805          return ir->operands[0];
 806       }
 807       break;
 808
 809    case ir_binop_pow:
 810       /* 1^x == 1 */
 811       if (is_vec_one(op_const[0]))
 812          return op_const[0];
 813
 814       /* x^1 == x */
 815       if (is_vec_one(op_const[1]))
 816          return ir->operands[0];
 817
 818       /* pow(2,x) == exp2(x) */
 819       if (is_vec_two(op_const[0]))
 820          return expr(ir_unop_exp2, ir->operands[1]);
 821
 822       if (is_vec_two(op_const[1])) {
 823          ir_variable *x = new(ir) ir_variable(ir->operands[1]->type, "x",
 824                                               ir_var_temporary);
 825          base_ir->insert_before(x);
 826          base_ir->insert_before(assign(x, ir->operands[0]));
 827          return mul(x, x);
 828       }
 829
 830       if (is_vec_four(op_const[1])) {
 831          ir_variable *x = new(ir) ir_variable(ir->operands[1]->type, "x",
 832                                               ir_var_temporary);
 833          base_ir->insert_before(x);
 834          base_ir->insert_before(assign(x, ir->operands[0]));
 835
 836          ir_variable *squared = new(ir) ir_variable(ir->operands[1]->type,
 837                                                     "squared",
 838                                                     ir_var_temporary);
 839          base_ir->insert_before(squared);
 840          base_ir->insert_before(assign(squared, mul(x, x)));
 841          return mul(squared, squared);
 842       }
 843
 844       break;
 845
 846    case ir_binop_min:
 847    case ir_binop_max:
 848       if (ir->type->base_type != GLSL_TYPE_FLOAT || options->EmitNoSat)
 849          break;
 850
 851       /* Replace min(max) operations and its commutative combinations with
 852        * a saturate operation
 853        */
 854       for (int op = 0; op < 2; op++) {
 855          ir_expression *inner_expr = op_expr[op];
 856          ir_constant *outer_const = op_const[1 - op];
 857          ir_expression_operation op_cond = (ir->operation == ir_binop_max) ?
 858             ir_binop_min : ir_binop_max;
 859
 860          if (!inner_expr || !outer_const || (inner_expr->operation != op_cond))
 861             continue;
 862
 863          /* One of these has to be a constant */
 864          if (!inner_expr->operands[0]->as_constant() &&
 865              !inner_expr->operands[1]->as_constant())
 866             break;
 867
 868          /* Found a min(max) combination. Now try to see if its operands
 869           * meet our conditions that we can do just a single saturate operation
 870           */
 871          for (int minmax_op = 0; minmax_op < 2; minmax_op++) {
 872             ir_rvalue *x = inner_expr->operands[minmax_op];
 873             ir_rvalue *y = inner_expr->operands[1 - minmax_op];
 874
 875             ir_constant *inner_const = y->as_constant();
 876             if (!inner_const)
 877                continue;
 878
 879             /* min(max(x, 0.0), 1.0) is sat(x) */
 880             if (ir->operation == ir_binop_min &&
 881                 inner_const->is_zero() &&
 882                 outer_const->is_one())
 883                return saturate(x);
 884
 885             /* max(min(x, 1.0), 0.0) is sat(x) */
 886             if (ir->operation == ir_binop_max &&
 887                 inner_const->is_one() &&
 888                 outer_const->is_zero())
 889                return saturate(x);
 890
 891             /* min(max(x, 0.0), b) where b < 1.0 is sat(min(x, b)) */
 892             if (ir->operation == ir_binop_min &&
 893                 inner_const->is_zero() &&
 894                 is_less_than_one(outer_const))
 895                return saturate(expr(ir_binop_min, x, outer_const));
 896
 897             /* max(min(x, b), 0.0) where b < 1.0 is sat(min(x, b)) */
 898             if (ir->operation == ir_binop_max &&
 899                 is_less_than_one(inner_const) &&
 900                 outer_const->is_zero())
 901                return saturate(expr(ir_binop_min, x, inner_const));
 902
 903             /* max(min(x, 1.0), b) where b > 0.0 is sat(max(x, b)) */
 904             if (ir->operation == ir_binop_max &&
 905                 inner_const->is_one() &&
 906                 is_greater_than_zero(outer_const))
 907                return saturate(expr(ir_binop_max, x, outer_const));
 908
 909             /* min(max(x, b), 1.0) where b > 0.0 is sat(max(x, b)) */
 910             if (ir->operation == ir_binop_min &&
 911                 is_greater_than_zero(inner_const) &&
 912                 outer_const->is_one())
 913                return saturate(expr(ir_binop_max, x, inner_const));
 914          }
 915       }
 916
 917       break;
 918
 919    case ir_unop_rcp:
 920       if (op_expr[0] && op_expr[0]->operation == ir_unop_rcp)
 921          return op_expr[0]->operands[0];
 922
 923       if (op_expr[0] && (op_expr[0]->operation == ir_unop_exp2 ||
 924                          op_expr[0]->operation == ir_unop_exp)) {
 925          return new(mem_ctx) ir_expression(op_expr[0]->operation, ir->type,
 926                                            neg(op_expr[0]->operands[0]));
 927       }
 928
 929       /* While ir_to_mesa.cpp will lower sqrt(x) to rcp(rsq(x)), it does so at
 930        * its IR level, so we can always apply this transformation.
 931        */
 932       if (op_expr[0] && op_expr[0]->operation == ir_unop_rsq)
 933          return sqrt(op_expr[0]->operands[0]);
 934
 935       /* As far as we know, all backends are OK with rsq. */
 936       if (op_expr[0] && op_expr[0]->operation == ir_unop_sqrt) {
 937          return rsq(op_expr[0]->operands[0]);
 938       }
 939
 940       break;
 941
 942    case ir_triop_fma:
 943       /* Operands are op0 * op1 + op2. */
 944       if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1])) {
 945          return ir->operands[2];
 946       } else if (is_vec_zero(op_const[2])) {
 947          return mul(ir->operands[0], ir->operands[1]);
 948       } else if (is_vec_one(op_const[0])) {
 949          return add(ir->operands[1], ir->operands[2]);
 950       } else if (is_vec_one(op_const[1])) {
 951          return add(ir->operands[0], ir->operands[2]);
 952       }
 953       break;
 954
 955    case ir_triop_lrp:
 956       /* Operands are (x, y, a). */
 957       if (is_vec_zero(op_const[2])) {
 958          return ir->operands[0];
 959       } else if (is_vec_one(op_const[2])) {
 960          return ir->operands[1];
 961       } else if (ir->operands[0]->equals(ir->operands[1])) {
 962          return ir->operands[0];
 963       } else if (is_vec_zero(op_const[0])) {
 964          return mul(ir->operands[1], ir->operands[2]);
 965       } else if (is_vec_zero(op_const[1])) {
 966          unsigned op2_components = ir->operands[2]->type->vector_elements;
 967          ir_constant *one;
 968
 969          switch (ir->type->base_type) {
 970          case GLSL_TYPE_FLOAT:
 971             one = new(mem_ctx) ir_constant(1.0f, op2_components);
 972             break;
 973          case GLSL_TYPE_DOUBLE:
 974             one = new(mem_ctx) ir_constant(1.0, op2_components);
 975             break;
 976          default:
 977             one = NULL;
 978             unreachable("unexpected type");
 979          }
 980
 981          return mul(ir->operands[0], add(one, neg(ir->operands[2])));
 982       }
 983       break;
 984
 985    case ir_triop_csel:
 986       if (is_vec_one(op_const[0]))
 987          return ir->operands[1];
 988       if (is_vec_zero(op_const[0]))
 989          return ir->operands[2];
 990       break;
 991
 992    /* Remove interpolateAt* instructions for demoted inputs. They are
 993     * assigned a constant expression to facilitate this.
 994     */
 995    case ir_unop_interpolate_at_centroid:
 996    case ir_binop_interpolate_at_offset:
 997    case ir_binop_interpolate_at_sample:
 998       if (op_const[0])
 999          return ir->operands[0];
1000       break;
1001
1002    default:
1003       break;
1004    }
1005
1006    return ir;
1007 }
1008
1009 void
1010 ir_algebraic_visitor::handle_rvalue(ir_rvalue **rvalue)
1011 {
1012    if (!*rvalue)
1013       return;
1014
1015    ir_expression *expr = (*rvalue)->as_expression();
1016    if (!expr || expr->operation == ir_quadop_vector)
1017       return;
1018
1019    ir_rvalue *new_rvalue = handle_expression(expr);
1020    if (new_rvalue == *rvalue)
1021       return;
1022
1023    /* If the expr used to be some vec OP scalar returning a vector, and the
1024     * optimization gave us back a scalar, we still need to turn it into a
1025     * vector.
1026     */
1027    *rvalue = swizzle_if_required(expr, new_rvalue);
1028
1029    this->progress = true;
1030 }
1031
1032 bool
1033 do_algebraic(exec_list *instructions, bool native_integers,
1034              const struct gl_shader_compiler_options *options)
1035 {
1036    ir_algebraic_visitor v(native_integers, options);
1037
1038    visit_list_elements(&v, instructions);
1039
1040    return v.progress;
1041 }