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->is_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->is_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 end 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    void *mem_ctx = ralloc_parent(ir2);
 267
 268    ir_constant *ir2_const[2];
 269    ir2_const[0] = ir2->operands[0]->constant_expression_value(mem_ctx);
 270    ir2_const[1] = ir2->operands[1]->constant_expression_value(mem_ctx);
 271
 272    if (ir2_const[0] && ir2_const[1])
 273       return false;
 274
 275    if (ir2_const[0]) {
 276       reassociate_operands(ir1, const_index, ir2, 1);
 277       return true;
 278    } else if (ir2_const[1]) {
 279       reassociate_operands(ir1, const_index, ir2, 0);
 280       return true;
 281    }
 282
 283    if (reassociate_constant(ir1, const_index, constant,
 284                             ir2->operands[0]->as_expression())) {
 285       update_type(ir2);
 286       return true;
 287    }
 288
 289    if (reassociate_constant(ir1, const_index, constant,
 290                             ir2->operands[1]->as_expression())) {
 291       update_type(ir2);
 292       return true;
 293    }
 294
 295    return false;
 296 }
 297
 298 /* When eliminating an expression and just returning one of its operands,
 299  * we may need to swizzle that operand out to a vector if the expression was
 300  * vector type.
 301  */
 302 ir_rvalue *
 303 ir_algebraic_visitor::swizzle_if_required(ir_expression *expr,
 304                                           ir_rvalue *operand)
 305 {
 306    if (expr->type->is_vector() && operand->type->is_scalar()) {
 307       return new(mem_ctx) ir_swizzle(operand, 0, 0, 0, 0,
 308                                      expr->type->vector_elements);
 309    } else
 310       return operand;
 311 }
 312
 313 ir_rvalue *
 314 ir_algebraic_visitor::handle_expression(ir_expression *ir)
 315 {
 316    ir_constant *op_const[4] = {NULL, NULL, NULL, NULL};
 317    ir_expression *op_expr[4] = {NULL, NULL, NULL, NULL};
 318
 319    if (ir->operation == ir_binop_mul &&
 320        ir->operands[0]->type->is_matrix() &&
 321        ir->operands[1]->type->is_vector()) {
 322       ir_expression *matrix_mul = ir->operands[0]->as_expression();
 323
 324       if (matrix_mul && matrix_mul->operation == ir_binop_mul &&
 325          matrix_mul->operands[0]->type->is_matrix() &&
 326          matrix_mul->operands[1]->type->is_matrix()) {
 327
 328          return mul(matrix_mul->operands[0],
 329                     mul(matrix_mul->operands[1], ir->operands[1]));
 330       }
 331    }
 332
 333    assert(ir->num_operands <= 4);
 334    for (unsigned i = 0; i < ir->num_operands; i++) {
 335       if (ir->operands[i]->type->is_matrix())
 336          return ir;
 337
 338       op_const[i] =
 339          ir->operands[i]->constant_expression_value(ralloc_parent(ir));
 340       op_expr[i] = ir->operands[i]->as_expression();
 341    }
 342
 343    if (this->mem_ctx == NULL)
 344       this->mem_ctx = ralloc_parent(ir);
 345
 346    switch (ir->operation) {
 347    case ir_unop_bit_not:
 348       if (op_expr[0] && op_expr[0]->operation == ir_unop_bit_not)
 349          return op_expr[0]->operands[0];
 350       break;
 351
 352    case ir_unop_abs:
 353       if (op_expr[0] == NULL)
 354          break;
 355
 356       switch (op_expr[0]->operation) {
 357       case ir_unop_abs:
 358       case ir_unop_neg:
 359          return abs(op_expr[0]->operands[0]);
 360       default:
 361          break;
 362       }
 363       break;
 364
 365    case ir_unop_neg:
 366       if (op_expr[0] == NULL)
 367          break;
 368
 369       if (op_expr[0]->operation == ir_unop_neg) {
 370          return op_expr[0]->operands[0];
 371       }
 372       break;
 373
 374    case ir_unop_exp:
 375       if (op_expr[0] == NULL)
 376          break;
 377
 378       if (op_expr[0]->operation == ir_unop_log) {
 379          return op_expr[0]->operands[0];
 380       }
 381       break;
 382
 383    case ir_unop_log:
 384       if (op_expr[0] == NULL)
 385          break;
 386
 387       if (op_expr[0]->operation == ir_unop_exp) {
 388          return op_expr[0]->operands[0];
 389       }
 390       break;
 391
 392    case ir_unop_exp2:
 393       if (op_expr[0] == NULL)
 394          break;
 395
 396       if (op_expr[0]->operation == ir_unop_log2) {
 397          return op_expr[0]->operands[0];
 398       }
 399
 400       if (!options->EmitNoPow && op_expr[0]->operation == ir_binop_mul) {
 401          for (int log2_pos = 0; log2_pos < 2; log2_pos++) {
 402             ir_expression *log2_expr =
 403                op_expr[0]->operands[log2_pos]->as_expression();
 404
 405             if (log2_expr && log2_expr->operation == ir_unop_log2) {
 406                return new(mem_ctx) ir_expression(ir_binop_pow,
 407                                                  ir->type,
 408                                                  log2_expr->operands[0],
 409                                                  op_expr[0]->operands[1 - log2_pos]);
 410             }
 411          }
 412       }
 413       break;
 414
 415    case ir_unop_log2:
 416       if (op_expr[0] == NULL)
 417          break;
 418
 419       if (op_expr[0]->operation == ir_unop_exp2) {
 420          return op_expr[0]->operands[0];
 421       }
 422       break;
 423
 424    case ir_unop_f2i:
 425    case ir_unop_f2u:
 426       if (op_expr[0] && op_expr[0]->operation == ir_unop_trunc) {
 427          return new(mem_ctx) ir_expression(ir->operation,
 428                                            ir->type,
 429                                            op_expr[0]->operands[0]);
 430       }
 431       break;
 432
 433    case ir_unop_logic_not: {
 434       enum ir_expression_operation new_op = ir_unop_logic_not;
 435
 436       if (op_expr[0] == NULL)
 437          break;
 438
 439       switch (op_expr[0]->operation) {
 440       case ir_binop_less:    new_op = ir_binop_gequal;  break;
 441       case ir_binop_greater: new_op = ir_binop_lequal;  break;
 442       case ir_binop_lequal:  new_op = ir_binop_greater; break;
 443       case ir_binop_gequal:  new_op = ir_binop_less;    break;
 444       case ir_binop_equal:   new_op = ir_binop_nequal;  break;
 445       case ir_binop_nequal:  new_op = ir_binop_equal;   break;
 446       case ir_binop_all_equal:   new_op = ir_binop_any_nequal;  break;
 447       case ir_binop_any_nequal:  new_op = ir_binop_all_equal;   break;
 448
 449       default:
 450          /* The default case handler is here to silence a warning from GCC.
 451           */
 452          break;
 453       }
 454
 455       if (new_op != ir_unop_logic_not) {
 456          return new(mem_ctx) ir_expression(new_op,
 457                                            ir->type,
 458                                            op_expr[0]->operands[0],
 459                                            op_expr[0]->operands[1]);
 460       }
 461
 462       break;
 463    }
 464
 465    case ir_unop_saturate:
 466       if (op_expr[0] && op_expr[0]->operation == ir_binop_add) {
 467          ir_expression *b2f_0 = op_expr[0]->operands[0]->as_expression();
 468          ir_expression *b2f_1 = op_expr[0]->operands[1]->as_expression();
 469
 470          if (b2f_0 && b2f_0->operation == ir_unop_b2f &&
 471              b2f_1 && b2f_1->operation == ir_unop_b2f) {
 472             return b2f(logic_or(b2f_0->operands[0], b2f_1->operands[0]));
 473          }
 474       }
 475       break;
 476
 477       /* This macro CANNOT use the do { } while(true) mechanism because
 478        * then the breaks apply to the loop instead of the switch!
 479        */
 480 #define HANDLE_PACK_UNPACK_INVERSE(inverse_operation)                   \
 481       {                                                                 \
 482          ir_expression *const op = ir->operands[0]->as_expression();    \
 483          if (op == NULL)                                                \
 484             break;                                                      \
 485          if (op->operation == (inverse_operation))                      \
 486             return op->operands[0];                                     \
 487          break;                                                         \
 488       }
 489
 490    case ir_unop_unpack_uint_2x32:
 491       HANDLE_PACK_UNPACK_INVERSE(ir_unop_pack_uint_2x32);
 492    case ir_unop_pack_uint_2x32:
 493       HANDLE_PACK_UNPACK_INVERSE(ir_unop_unpack_uint_2x32);
 494    case ir_unop_unpack_int_2x32:
 495       HANDLE_PACK_UNPACK_INVERSE(ir_unop_pack_int_2x32);
 496    case ir_unop_pack_int_2x32:
 497       HANDLE_PACK_UNPACK_INVERSE(ir_unop_unpack_int_2x32);
 498    case ir_unop_unpack_double_2x32:
 499       HANDLE_PACK_UNPACK_INVERSE(ir_unop_pack_double_2x32);
 500    case ir_unop_pack_double_2x32:
 501       HANDLE_PACK_UNPACK_INVERSE(ir_unop_unpack_double_2x32);
 502
 503 #undef HANDLE_PACK_UNPACK_INVERSE
 504
 505    case ir_binop_add:
 506       if (is_vec_zero(op_const[0]))
 507          return ir->operands[1];
 508       if (is_vec_zero(op_const[1]))
 509          return ir->operands[0];
 510
 511       /* Reassociate addition of constants so that we can do constant
 512        * folding.
 513        */
 514       if (op_const[0] && !op_const[1])
 515          reassociate_constant(ir, 0, op_const[0], op_expr[1]);
 516       if (op_const[1] && !op_const[0])
 517          reassociate_constant(ir, 1, op_const[1], op_expr[0]);
 518
 519       /* Recognize (v.x + v.y) + (v.z + v.w) as dot(v, 1.0) */
 520       if (options->OptimizeForAOS) {
 521          ir_expression *expr = try_replace_with_dot(op_expr[0], op_expr[1],
 522                                                     mem_ctx);
 523          if (expr)
 524             return expr;
 525       }
 526
 527       /* Replace (-x + y) * a + x and commutative variations with lrp(x, y, a).
 528        *
 529        * (-x + y) * a + x
 530        * (x * -a) + (y * a) + x
 531        * x + (x * -a) + (y * a)
 532        * x * (1 - a) + y * a
 533        * lrp(x, y, a)
 534        */
 535       for (int mul_pos = 0; mul_pos < 2; mul_pos++) {
 536          ir_expression *mul = op_expr[mul_pos];
 537
 538          if (!mul || mul->operation != ir_binop_mul)
 539             continue;
 540
 541          /* Multiply found on one of the operands. Now check for an
 542           * inner addition operation.
 543           */
 544          for (int inner_add_pos = 0; inner_add_pos < 2; inner_add_pos++) {
 545             ir_expression *inner_add =
 546                mul->operands[inner_add_pos]->as_expression();
 547
 548             if (!inner_add || inner_add->operation != ir_binop_add)
 549                continue;
 550
 551             /* Inner addition found on one of the operands. Now check for
 552              * one of the operands of the inner addition to be the negative
 553              * of x_operand.
 554              */
 555             for (int neg_pos = 0; neg_pos < 2; neg_pos++) {
 556                ir_expression *neg =
 557                   inner_add->operands[neg_pos]->as_expression();
 558
 559                if (!neg || neg->operation != ir_unop_neg)
 560                   continue;
 561
 562                ir_rvalue *x_operand = ir->operands[1 - mul_pos];
 563
 564                if (!neg->operands[0]->equals(x_operand))
 565                   continue;
 566
 567                ir_rvalue *y_operand = inner_add->operands[1 - neg_pos];
 568                ir_rvalue *a_operand = mul->operands[1 - inner_add_pos];
 569
 570                if (x_operand->type != y_operand->type ||
 571                    x_operand->type != a_operand->type)
 572                   continue;
 573
 574                return lrp(x_operand, y_operand, a_operand);
 575             }
 576          }
 577       }
 578
 579       break;
 580
 581    case ir_binop_sub:
 582       if (is_vec_zero(op_const[0]))
 583          return neg(ir->operands[1]);
 584       if (is_vec_zero(op_const[1]))
 585          return ir->operands[0];
 586       break;
 587
 588    case ir_binop_mul:
 589       if (is_vec_one(op_const[0]))
 590          return ir->operands[1];
 591       if (is_vec_one(op_const[1]))
 592          return ir->operands[0];
 593
 594       if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1]))
 595          return ir_constant::zero(ir, ir->type);
 596
 597       if (is_vec_negative_one(op_const[0]))
 598          return neg(ir->operands[1]);
 599       if (is_vec_negative_one(op_const[1]))
 600          return neg(ir->operands[0]);
 601
 602       if (op_expr[0] && op_expr[0]->operation == ir_unop_b2f &&
 603           op_expr[1] && op_expr[1]->operation == ir_unop_b2f) {
 604          return b2f(logic_and(op_expr[0]->operands[0], op_expr[1]->operands[0]));
 605       }
 606
 607       /* Reassociate multiplication of constants so that we can do
 608        * constant folding.
 609        */
 610       if (op_const[0] && !op_const[1])
 611          reassociate_constant(ir, 0, op_const[0], op_expr[1]);
 612       if (op_const[1] && !op_const[0])
 613          reassociate_constant(ir, 1, op_const[1], op_expr[0]);
 614
 615       /* Optimizes
 616        *
 617        *    (mul (floor (add (abs x) 0.5) (sign x)))
 618        *
 619        * into
 620        *
 621        *    (trunc (add x (mul (sign x) 0.5)))
 622        */
 623       for (int i = 0; i < 2; i++) {
 624          ir_expression *sign_expr = ir->operands[i]->as_expression();
 625          ir_expression *floor_expr = ir->operands[1 - i]->as_expression();
 626
 627          if (!sign_expr || sign_expr->operation != ir_unop_sign ||
 628              !floor_expr || floor_expr->operation != ir_unop_floor)
 629             continue;
 630
 631          ir_expression *add_expr = floor_expr->operands[0]->as_expression();
 632          if (!add_expr || add_expr->operation != ir_binop_add)
 633             continue;
 634
 635          for (int j = 0; j < 2; j++) {
 636             ir_expression *abs_expr = add_expr->operands[j]->as_expression();
 637             if (!abs_expr || abs_expr->operation != ir_unop_abs)
 638                continue;
 639
 640             ir_constant *point_five = add_expr->operands[1 - j]->as_constant();
 641             if (!point_five || !point_five->is_value(0.5, 0))
 642                continue;
 643
 644             if (abs_expr->operands[0]->equals(sign_expr->operands[0])) {
 645                return trunc(add(abs_expr->operands[0],
 646                                 mul(sign_expr, point_five)));
 647             }
 648          }
 649       }
 650       break;
 651
 652    case ir_binop_div:
 653       if (is_vec_one(op_const[0]) && (
 654                 ir->type->is_float() || ir->type->is_double())) {
 655          return new(mem_ctx) ir_expression(ir_unop_rcp,
 656                                            ir->operands[1]->type,
 657                                            ir->operands[1],
 658                                            NULL);
 659       }
 660       if (is_vec_one(op_const[1]))
 661          return ir->operands[0];
 662       break;
 663
 664    case ir_binop_dot:
 665       if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1]))
 666          return ir_constant::zero(mem_ctx, ir->type);
 667
 668       for (int i = 0; i < 2; i++) {
 669          if (!op_const[i])
 670             continue;
 671
 672          unsigned components[4] = { 0 }, count = 0;
 673
 674          for (unsigned c = 0; c < op_const[i]->type->vector_elements; c++) {
 675             if (op_const[i]->is_zero())
 676                continue;
 677
 678             components[count] = c;
 679             count++;
 680          }
 681
 682          /* No channels had zero values; bail. */
 683          if (count >= op_const[i]->type->vector_elements)
 684             break;
 685
 686          ir_expression_operation op = count == 1 ?
 687             ir_binop_mul : ir_binop_dot;
 688
 689          /* Swizzle both operands to remove the channels that were zero. */
 690          return new(mem_ctx)
 691             ir_expression(op, ir->type,
 692                           new(mem_ctx) ir_swizzle(ir->operands[0],
 693                                                   components, count),
 694                           new(mem_ctx) ir_swizzle(ir->operands[1],
 695                                                   components, count));
 696       }
 697       break;
 698
 699    case ir_binop_less:
 700    case ir_binop_lequal:
 701    case ir_binop_greater:
 702    case ir_binop_gequal:
 703    case ir_binop_equal:
 704    case ir_binop_nequal:
 705       for (int add_pos = 0; add_pos < 2; add_pos++) {
 706          ir_expression *add = op_expr[add_pos];
 707
 708          if (!add || add->operation != ir_binop_add)
 709             continue;
 710
 711          ir_constant *zero = op_const[1 - add_pos];
 712          if (!is_vec_zero(zero))
 713             continue;
 714
 715          /* Depending of the zero position we want to optimize
 716           * (0 cmp x+y) into (-x cmp y) or (x+y cmp 0) into (x cmp -y)
 717           */
 718          if (add_pos == 1) {
 719             return new(mem_ctx) ir_expression(ir->operation,
 720                                               neg(add->operands[0]),
 721                                               add->operands[1]);
 722          } else {
 723             return new(mem_ctx) ir_expression(ir->operation,
 724                                               add->operands[0],
 725                                               neg(add->operands[1]));
 726          }
 727       }
 728       break;
 729
 730    case ir_binop_all_equal:
 731    case ir_binop_any_nequal:
 732       if (ir->operands[0]->type->is_scalar() &&
 733           ir->operands[1]->type->is_scalar())
 734          return new(mem_ctx) ir_expression(ir->operation == ir_binop_all_equal
 735                                            ? ir_binop_equal : ir_binop_nequal,
 736                                            ir->operands[0],
 737                                            ir->operands[1]);
 738       break;
 739
 740    case ir_binop_rshift:
 741    case ir_binop_lshift:
 742       /* 0 >> x == 0 */
 743       if (is_vec_zero(op_const[0]))
 744          return ir->operands[0];
 745       /* x >> 0 == x */
 746       if (is_vec_zero(op_const[1]))
 747          return ir->operands[0];
 748       break;
 749
 750    case ir_binop_logic_and:
 751       if (is_vec_one(op_const[0])) {
 752          return ir->operands[1];
 753       } else if (is_vec_one(op_const[1])) {
 754          return ir->operands[0];
 755       } else if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1])) {
 756          return ir_constant::zero(mem_ctx, ir->type);
 757       } else if (op_expr[0] && op_expr[0]->operation == ir_unop_logic_not &&
 758                  op_expr[1] && op_expr[1]->operation == ir_unop_logic_not) {
 759          /* De Morgan's Law:
 760           *    (not A) and (not B) === not (A or B)
 761           */
 762          return logic_not(logic_or(op_expr[0]->operands[0],
 763                                    op_expr[1]->operands[0]));
 764       } else if (ir->operands[0]->equals(ir->operands[1])) {
 765          /* (a && a) == a */
 766          return ir->operands[0];
 767       }
 768       break;
 769
 770    case ir_binop_logic_xor:
 771       if (is_vec_zero(op_const[0])) {
 772          return ir->operands[1];
 773       } else if (is_vec_zero(op_const[1])) {
 774          return ir->operands[0];
 775       } else if (is_vec_one(op_const[0])) {
 776          return logic_not(ir->operands[1]);
 777       } else if (is_vec_one(op_const[1])) {
 778          return logic_not(ir->operands[0]);
 779       } else if (ir->operands[0]->equals(ir->operands[1])) {
 780          /* (a ^^ a) == false */
 781          return ir_constant::zero(mem_ctx, ir->type);
 782       }
 783       break;
 784
 785    case ir_binop_logic_or:
 786       if (is_vec_zero(op_const[0])) {
 787          return ir->operands[1];
 788       } else if (is_vec_zero(op_const[1])) {
 789          return ir->operands[0];
 790       } else if (is_vec_one(op_const[0]) || is_vec_one(op_const[1])) {
 791          ir_constant_data data;
 792
 793          for (unsigned i = 0; i < 16; i++)
 794             data.b[i] = true;
 795
 796          return new(mem_ctx) ir_constant(ir->type, &data);
 797       } else if (op_expr[0] && op_expr[0]->operation == ir_unop_logic_not &&
 798                  op_expr[1] && op_expr[1]->operation == ir_unop_logic_not) {
 799          /* De Morgan's Law:
 800           *    (not A) or (not B) === not (A and B)
 801           */
 802          return logic_not(logic_and(op_expr[0]->operands[0],
 803                                     op_expr[1]->operands[0]));
 804       } else if (ir->operands[0]->equals(ir->operands[1])) {
 805          /* (a || a) == a */
 806          return ir->operands[0];
 807       }
 808       break;
 809
 810    case ir_binop_pow:
 811       /* 1^x == 1 */
 812       if (is_vec_one(op_const[0]))
 813          return op_const[0];
 814
 815       /* x^1 == x */
 816       if (is_vec_one(op_const[1]))
 817          return ir->operands[0];
 818
 819       /* pow(2,x) == exp2(x) */
 820       if (is_vec_two(op_const[0]))
 821          return expr(ir_unop_exp2, ir->operands[1]);
 822
 823       if (is_vec_two(op_const[1])) {
 824          ir_variable *x = new(ir) ir_variable(ir->operands[1]->type, "x",
 825                                               ir_var_temporary);
 826          base_ir->insert_before(x);
 827          base_ir->insert_before(assign(x, ir->operands[0]));
 828          return mul(x, x);
 829       }
 830
 831       if (is_vec_four(op_const[1])) {
 832          ir_variable *x = new(ir) ir_variable(ir->operands[1]->type, "x",
 833                                               ir_var_temporary);
 834          base_ir->insert_before(x);
 835          base_ir->insert_before(assign(x, ir->operands[0]));
 836
 837          ir_variable *squared = new(ir) ir_variable(ir->operands[1]->type,
 838                                                     "squared",
 839                                                     ir_var_temporary);
 840          base_ir->insert_before(squared);
 841          base_ir->insert_before(assign(squared, mul(x, x)));
 842          return mul(squared, squared);
 843       }
 844
 845       break;
 846
 847    case ir_binop_min:
 848    case ir_binop_max:
 849       if (!ir->type->is_float() || options->EmitNoSat)
 850          break;
 851
 852       /* Replace min(max) operations and its commutative combinations with
 853        * a saturate operation
 854        */
 855       for (int op = 0; op < 2; op++) {
 856          ir_expression *inner_expr = op_expr[op];
 857          ir_constant *outer_const = op_const[1 - op];
 858          ir_expression_operation op_cond = (ir->operation == ir_binop_max) ?
 859             ir_binop_min : ir_binop_max;
 860
 861          if (!inner_expr || !outer_const || (inner_expr->operation != op_cond))
 862             continue;
 863
 864          /* One of these has to be a constant */
 865          if (!inner_expr->operands[0]->as_constant() &&
 866              !inner_expr->operands[1]->as_constant())
 867             break;
 868
 869          /* Found a min(max) combination. Now try to see if its operands
 870           * meet our conditions that we can do just a single saturate operation
 871           */
 872          for (int minmax_op = 0; minmax_op < 2; minmax_op++) {
 873             ir_rvalue *x = inner_expr->operands[minmax_op];
 874             ir_rvalue *y = inner_expr->operands[1 - minmax_op];
 875
 876             ir_constant *inner_const = y->as_constant();
 877             if (!inner_const)
 878                continue;
 879
 880             /* min(max(x, 0.0), 1.0) is sat(x) */
 881             if (ir->operation == ir_binop_min &&
 882                 inner_const->is_zero() &&
 883                 outer_const->is_one())
 884                return saturate(x);
 885
 886             /* max(min(x, 1.0), 0.0) is sat(x) */
 887             if (ir->operation == ir_binop_max &&
 888                 inner_const->is_one() &&
 889                 outer_const->is_zero())
 890                return saturate(x);
 891
 892             /* min(max(x, 0.0), b) where b < 1.0 is sat(min(x, b)) */
 893             if (ir->operation == ir_binop_min &&
 894                 inner_const->is_zero() &&
 895                 is_less_than_one(outer_const))
 896                return saturate(expr(ir_binop_min, x, outer_const));
 897
 898             /* max(min(x, b), 0.0) where b < 1.0 is sat(min(x, b)) */
 899             if (ir->operation == ir_binop_max &&
 900                 is_less_than_one(inner_const) &&
 901                 outer_const->is_zero())
 902                return saturate(expr(ir_binop_min, x, inner_const));
 903
 904             /* max(min(x, 1.0), b) where b > 0.0 is sat(max(x, b)) */
 905             if (ir->operation == ir_binop_max &&
 906                 inner_const->is_one() &&
 907                 is_greater_than_zero(outer_const))
 908                return saturate(expr(ir_binop_max, x, outer_const));
 909
 910             /* min(max(x, b), 1.0) where b > 0.0 is sat(max(x, b)) */
 911             if (ir->operation == ir_binop_min &&
 912                 is_greater_than_zero(inner_const) &&
 913                 outer_const->is_one())
 914                return saturate(expr(ir_binop_max, x, inner_const));
 915          }
 916       }
 917
 918       break;
 919
 920    case ir_unop_rcp:
 921       if (op_expr[0] && op_expr[0]->operation == ir_unop_rcp)
 922          return op_expr[0]->operands[0];
 923
 924       if (op_expr[0] && (op_expr[0]->operation == ir_unop_exp2 ||
 925                          op_expr[0]->operation == ir_unop_exp)) {
 926          return new(mem_ctx) ir_expression(op_expr[0]->operation, ir->type,
 927                                            neg(op_expr[0]->operands[0]));
 928       }
 929
 930       /* While ir_to_mesa.cpp will lower sqrt(x) to rcp(rsq(x)), it does so at
 931        * its IR level, so we can always apply this transformation.
 932        */
 933       if (op_expr[0] && op_expr[0]->operation == ir_unop_rsq)
 934          return sqrt(op_expr[0]->operands[0]);
 935
 936       /* As far as we know, all backends are OK with rsq. */
 937       if (op_expr[0] && op_expr[0]->operation == ir_unop_sqrt) {
 938          return rsq(op_expr[0]->operands[0]);
 939       }
 940
 941       break;
 942
 943    case ir_triop_fma:
 944       /* Operands are op0 * op1 + op2. */
 945       if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1])) {
 946          return ir->operands[2];
 947       } else if (is_vec_zero(op_const[2])) {
 948          return mul(ir->operands[0], ir->operands[1]);
 949       } else if (is_vec_one(op_const[0])) {
 950          return add(ir->operands[1], ir->operands[2]);
 951       } else if (is_vec_one(op_const[1])) {
 952          return add(ir->operands[0], ir->operands[2]);
 953       }
 954       break;
 955
 956    case ir_triop_lrp:
 957       /* Operands are (x, y, a). */
 958       if (is_vec_zero(op_const[2])) {
 959          return ir->operands[0];
 960       } else if (is_vec_one(op_const[2])) {
 961          return ir->operands[1];
 962       } else if (ir->operands[0]->equals(ir->operands[1])) {
 963          return ir->operands[0];
 964       } else if (is_vec_zero(op_const[0])) {
 965          return mul(ir->operands[1], ir->operands[2]);
 966       } else if (is_vec_zero(op_const[1])) {
 967          unsigned op2_components = ir->operands[2]->type->vector_elements;
 968          ir_constant *one;
 969
 970          switch (ir->type->base_type) {
 971          case GLSL_TYPE_FLOAT:
 972             one = new(mem_ctx) ir_constant(1.0f, op2_components);
 973             break;
 974          case GLSL_TYPE_DOUBLE:
 975             one = new(mem_ctx) ir_constant(1.0, op2_components);
 976             break;
 977          default:
 978             one = NULL;
 979             unreachable("unexpected type");
 980          }
 981
 982          return mul(ir->operands[0], add(one, neg(ir->operands[2])));
 983       }
 984       break;
 985
 986    case ir_triop_csel:
 987       if (is_vec_one(op_const[0]))
 988          return ir->operands[1];
 989       if (is_vec_zero(op_const[0]))
 990          return ir->operands[2];
 991       break;
 992
 993    /* Remove interpolateAt* instructions for demoted inputs. They are
 994     * assigned a constant expression to facilitate this.
 995     */
 996    case ir_unop_interpolate_at_centroid:
 997    case ir_binop_interpolate_at_offset:
 998    case ir_binop_interpolate_at_sample:
 999       if (op_const[0])
1000          return ir->operands[0];
1001       break;
1002
1003    default:
1004       break;
1005    }
1006
1007    return ir;
1008 }
1009
1010 void
1011 ir_algebraic_visitor::handle_rvalue(ir_rvalue **rvalue)
1012 {
1013    if (!*rvalue)
1014       return;
1015
1016    ir_expression *expr = (*rvalue)->as_expression();
1017    if (!expr || expr->operation == ir_quadop_vector)
1018       return;
1019
1020    ir_rvalue *new_rvalue = handle_expression(expr);
1021    if (new_rvalue == *rvalue)
1022       return;
1023
1024    /* If the expr used to be some vec OP scalar returning a vector, and the
1025     * optimization gave us back a scalar, we still need to turn it into a
1026     * vector.
1027     */
1028    *rvalue = swizzle_if_required(expr, new_rvalue);
1029
1030    this->progress = true;
1031 }
1032
1033 bool
1034 do_algebraic(exec_list *instructions, bool native_integers,
1035              const struct gl_shader_compiler_options *options)
1036 {
1037    ir_algebraic_visitor v(native_integers, options);
1038
1039    visit_list_elements(&v, instructions);
1040
1041    return v.progress;
1042 }