class ir_algebraic_visitor : public ir_rvalue_visitor {
public:
- ir_algebraic_visitor()
+ ir_algebraic_visitor(bool native_integers,
+ const struct gl_shader_compiler_options *options)
+ : options(options)
{
this->progress = false;
this->mem_ctx = NULL;
+ this->native_integers = native_integers;
}
virtual ~ir_algebraic_visitor()
ir_rvalue *swizzle_if_required(ir_expression *expr,
ir_rvalue *operand);
+ const struct gl_shader_compiler_options *options;
void *mem_ctx;
+ bool native_integers;
bool progress;
};
return (ir == NULL) ? false : ir->is_one();
}
+static inline bool
+is_vec_two(ir_constant *ir)
+{
+ return (ir == NULL) ? false : ir->is_value(2.0, 2);
+}
+
static inline bool
is_vec_negative_one(ir_constant *ir)
{
ir->type = ir->operands[1]->type;
}
+/* Recognize (v.x + v.y) + (v.z + v.w) as dot(v, 1.0) */
+static ir_expression *
+try_replace_with_dot(ir_expression *expr0, ir_expression *expr1, void *mem_ctx)
+{
+ if (expr0 && expr0->operation == ir_binop_add &&
+ expr0->type->is_float() &&
+ expr1 && expr1->operation == ir_binop_add &&
+ expr1->type->is_float()) {
+ ir_swizzle *x = expr0->operands[0]->as_swizzle();
+ ir_swizzle *y = expr0->operands[1]->as_swizzle();
+ ir_swizzle *z = expr1->operands[0]->as_swizzle();
+ ir_swizzle *w = expr1->operands[1]->as_swizzle();
+
+ if (!x || x->mask.num_components != 1 ||
+ !y || y->mask.num_components != 1 ||
+ !z || z->mask.num_components != 1 ||
+ !w || w->mask.num_components != 1) {
+ return NULL;
+ }
+
+ bool swiz_seen[4] = {false, false, false, false};
+ swiz_seen[x->mask.x] = true;
+ swiz_seen[y->mask.x] = true;
+ swiz_seen[z->mask.x] = true;
+ swiz_seen[w->mask.x] = true;
+
+ if (!swiz_seen[0] || !swiz_seen[1] ||
+ !swiz_seen[2] || !swiz_seen[3]) {
+ return NULL;
+ }
+
+ if (x->val->equals(y->val) &&
+ x->val->equals(z->val) &&
+ x->val->equals(w->val)) {
+ return dot(x->val, new(mem_ctx) ir_constant(1.0f, 4));
+ }
+ }
+ return NULL;
+}
+
void
ir_algebraic_visitor::reassociate_operands(ir_expression *ir1,
int op1,
this->mem_ctx = ralloc_parent(ir);
switch (ir->operation) {
+ case ir_unop_bit_not:
+ if (op_expr[0] && op_expr[0]->operation == ir_unop_bit_not)
+ return op_expr[0]->operands[0];
+ break;
+
case ir_unop_abs:
if (op_expr[0] == NULL)
break;
}
break;
+ case ir_unop_exp:
+ if (op_expr[0] == NULL)
+ break;
+
+ if (op_expr[0]->operation == ir_unop_log) {
+ return op_expr[0]->operands[0];
+ }
+ break;
+
+ case ir_unop_log:
+ if (op_expr[0] == NULL)
+ break;
+
+ if (op_expr[0]->operation == ir_unop_exp) {
+ return op_expr[0]->operands[0];
+ }
+ break;
+
+ case ir_unop_exp2:
+ if (op_expr[0] == NULL)
+ break;
+
+ if (op_expr[0]->operation == ir_unop_log2) {
+ return op_expr[0]->operands[0];
+ }
+ break;
+
+ case ir_unop_log2:
+ if (op_expr[0] == NULL)
+ break;
+
+ if (op_expr[0]->operation == ir_unop_exp2) {
+ return op_expr[0]->operands[0];
+ }
+ break;
+
case ir_unop_logic_not: {
enum ir_expression_operation new_op = ir_unop_logic_not;
reassociate_constant(ir, 0, op_const[0], op_expr[1]);
if (op_const[1] && !op_const[0])
reassociate_constant(ir, 1, op_const[1], op_expr[0]);
+
+ /* Recognize (v.x + v.y) + (v.z + v.w) as dot(v, 1.0) */
+ if (options->OptimizeForAOS) {
+ ir_expression *expr = try_replace_with_dot(op_expr[0], op_expr[1],
+ mem_ctx);
+ if (expr)
+ return expr;
+ }
+
+ /* Replace (-x + y) * a + x and commutative variations with lrp(x, y, a).
+ *
+ * (-x + y) * a + x
+ * (x * -a) + (y * a) + x
+ * x + (x * -a) + (y * a)
+ * x * (1 - a) + y * a
+ * lrp(x, y, a)
+ */
+ for (int mul_pos = 0; mul_pos < 2; mul_pos++) {
+ ir_expression *mul = op_expr[mul_pos];
+
+ if (!mul || mul->operation != ir_binop_mul)
+ continue;
+
+ /* Multiply found on one of the operands. Now check for an
+ * inner addition operation.
+ */
+ for (int inner_add_pos = 0; inner_add_pos < 2; inner_add_pos++) {
+ ir_expression *inner_add =
+ mul->operands[inner_add_pos]->as_expression();
+
+ if (!inner_add || inner_add->operation != ir_binop_add)
+ continue;
+
+ /* Inner addition found on one of the operands. Now check for
+ * one of the operands of the inner addition to be the negative
+ * of x_operand.
+ */
+ for (int neg_pos = 0; neg_pos < 2; neg_pos++) {
+ ir_expression *neg =
+ inner_add->operands[neg_pos]->as_expression();
+
+ if (!neg || neg->operation != ir_unop_neg)
+ continue;
+
+ ir_rvalue *x_operand = ir->operands[1 - mul_pos];
+
+ if (!neg->operands[0]->equals(x_operand))
+ continue;
+
+ ir_rvalue *y_operand = inner_add->operands[1 - neg_pos];
+ ir_rvalue *a_operand = mul->operands[1 - inner_add_pos];
+
+ if (x_operand->type != y_operand->type ||
+ x_operand->type != a_operand->type)
+ continue;
+
+ return lrp(x_operand, y_operand, a_operand);
+ }
+ }
+ }
+
break;
case ir_binop_sub:
}
break;
+ case ir_binop_less:
+ case ir_binop_lequal:
+ case ir_binop_greater:
+ case ir_binop_gequal:
+ case ir_binop_equal:
+ case ir_binop_nequal:
+ for (int add_pos = 0; add_pos < 2; add_pos++) {
+ ir_expression *add = op_expr[add_pos];
+
+ if (!add || add->operation != ir_binop_add)
+ continue;
+
+ ir_constant *zero = op_const[1 - add_pos];
+ if (!is_vec_zero(zero))
+ continue;
+
+ return new(mem_ctx) ir_expression(ir->operation,
+ add->operands[0],
+ neg(add->operands[1]));
+ }
+ break;
+
case ir_binop_rshift:
case ir_binop_lshift:
/* 0 >> x == 0 */
break;
case ir_binop_logic_and:
- /* FINISHME: Also simplify (a && a) to (a). */
if (is_vec_one(op_const[0])) {
return ir->operands[1];
} else if (is_vec_one(op_const[1])) {
*/
return logic_not(logic_or(op_expr[0]->operands[0],
op_expr[1]->operands[0]));
+ } else if (ir->operands[0]->equals(ir->operands[1])) {
+ /* (a && a) == a */
+ return ir->operands[0];
}
break;
case ir_binop_logic_xor:
- /* FINISHME: Also simplify (a ^^ a) to (false). */
if (is_vec_zero(op_const[0])) {
return ir->operands[1];
} else if (is_vec_zero(op_const[1])) {
return logic_not(ir->operands[1]);
} else if (is_vec_one(op_const[1])) {
return logic_not(ir->operands[0]);
+ } else if (ir->operands[0]->equals(ir->operands[1])) {
+ /* (a ^^ a) == false */
+ return ir_constant::zero(mem_ctx, ir->type);
}
break;
}
break;
+ case ir_binop_pow:
+ /* 1^x == 1 */
+ if (is_vec_one(op_const[0]))
+ return op_const[0];
+
+ /* x^1 == x */
+ if (is_vec_one(op_const[1]))
+ return ir->operands[0];
+
+ /* pow(2,x) == exp2(x) */
+ if (is_vec_two(op_const[0]))
+ return expr(ir_unop_exp2, ir->operands[1]);
+
+ if (is_vec_two(op_const[1])) {
+ ir_variable *x = new(ir) ir_variable(ir->operands[1]->type, "x",
+ ir_var_temporary);
+ base_ir->insert_before(x);
+ base_ir->insert_before(assign(x, ir->operands[0]));
+ return mul(x, x);
+ }
+
+ break;
+
case ir_unop_rcp:
if (op_expr[0] && op_expr[0]->operation == ir_unop_rcp)
return op_expr[0]->operands[0];
- /* FINISHME: We should do rcp(rsq(x)) -> sqrt(x) for some
- * backends, except that some backends will have done sqrt ->
- * rcp(rsq(x)) and we don't want to undo it for them.
+ /* While ir_to_mesa.cpp will lower sqrt(x) to rcp(rsq(x)), it does so at
+ * its IR level, so we can always apply this transformation.
*/
+ if (op_expr[0] && op_expr[0]->operation == ir_unop_rsq)
+ return sqrt(op_expr[0]->operands[0]);
/* As far as we know, all backends are OK with rsq. */
if (op_expr[0] && op_expr[0]->operation == ir_unop_sqrt) {
break;
+ case ir_triop_fma:
+ /* Operands are op0 * op1 + op2. */
+ if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1])) {
+ return ir->operands[2];
+ } else if (is_vec_zero(op_const[2])) {
+ return mul(ir->operands[0], ir->operands[1]);
+ } else if (is_vec_one(op_const[0])) {
+ return add(ir->operands[1], ir->operands[2]);
+ } else if (is_vec_one(op_const[1])) {
+ return add(ir->operands[0], ir->operands[2]);
+ }
+ break;
+
case ir_triop_lrp:
/* Operands are (x, y, a). */
if (is_vec_zero(op_const[2])) {
return ir->operands[0];
} else if (is_vec_one(op_const[2])) {
return ir->operands[1];
+ } else if (ir->operands[0]->equals(ir->operands[1])) {
+ return ir->operands[0];
+ } else if (is_vec_zero(op_const[0])) {
+ return mul(ir->operands[1], ir->operands[2]);
+ } else if (is_vec_zero(op_const[1])) {
+ unsigned op2_components = ir->operands[2]->type->vector_elements;
+ ir_constant *one = new(mem_ctx) ir_constant(1.0f, op2_components);
+ return mul(ir->operands[0], add(one, neg(ir->operands[2])));
}
break;
+ case ir_triop_csel:
+ if (is_vec_one(op_const[0]))
+ return ir->operands[1];
+ if (is_vec_zero(op_const[0]))
+ return ir->operands[2];
+ break;
+
default:
break;
}
}
bool
-do_algebraic(exec_list *instructions)
+do_algebraic(exec_list *instructions, bool native_integers,
+ const struct gl_shader_compiler_options *options)
{
- ir_algebraic_visitor v;
+ ir_algebraic_visitor v(native_integers, options);
visit_list_elements(&v, instructions);
projects / mesa.git / blobdiff