https://github.com/halide/Halide
Tip revision: a317ebe4aa0bbf3412163b528923158b977f4a73 authored by Steven Johnson on 28 March 2018, 00:45:23 UTC
Various test fixes
Various test fixes
Tip revision: a317ebe
Simplify.cpp
#include <iostream>
#include <algorithm>
#include <cmath>
#include <limits>
#include <stdio.h>
#include "Simplify.h"
#include "IROperator.h"
#include "IREquality.h"
#include "IRPrinter.h"
#include "IRMutator.h"
#include "Scope.h"
#include "Var.h"
#include "Debug.h"
#include "ModulusRemainder.h"
#include "Substitute.h"
#include "Bounds.h"
#include "Deinterleave.h"
#include "ExprUsesVar.h"
#ifdef _MSC_VER
#define snprintf _snprintf
#endif
namespace Halide {
namespace Internal {
using std::string;
using std::map;
using std::pair;
using std::ostringstream;
using std::vector;
#define LOG_EXPR_MUTATIONS 0
#define LOG_STMT_MUTATIONS 0
namespace {
// Things that we can constant fold: Immediates and broadcasts of immediates.
bool is_simple_const(const Expr &e) {
if (e.as<IntImm>()) return true;
if (e.as<UIntImm>()) return true;
// Don't consider NaN to be a "simple const", since it doesn't obey equality rules assumed elsewere
const FloatImm *f = e.as<FloatImm>();
if (f && !std::isnan(f->value)) return true;
if (const Broadcast *b = e.as<Broadcast>()) {
return is_simple_const(b->value);
}
return false;
}
// If the Expr is (var relop const) or (const relop var),
// fill in the var name and return true.
template<typename RelOp>
bool is_var_relop_simple_const(const Expr &e, string* name) {
if (const RelOp *r = e.as<RelOp>()) {
if (is_simple_const(r->b)) {
const Variable *v = r->a.template as<Variable>();
if (v) {
*name = v->name;
return true;
}
} else if (is_simple_const(r->a)) {
const Variable *v = r->b.template as<Variable>();
if (v) {
*name = v->name;
return true;
}
}
}
return false;
}
bool is_var_simple_const_comparison(const Expr &e, string* name) {
// It's not clear if GT, LT, etc would be useful
// here; leaving them out until proven otherwise.
return is_var_relop_simple_const<EQ>(e, name) ||
is_var_relop_simple_const<NE>(e, name);
}
// Returns true iff t is an integral type where overflow is undefined
bool no_overflow_int(Type t) {
return t.is_int() && t.bits() >= 32;
}
bool no_overflow_scalar_int(Type t) {
return t.is_scalar() && no_overflow_int(t);
}
// Returns true iff t does not have a well defined overflow behavior.
bool no_overflow(Type t) {
return t.is_float() || no_overflow_int(t);
}
// Make a poison value used when overflow is detected during constant
// folding.
Expr signed_integer_overflow_error(Type t) {
// Mark each call with an atomic counter, so that the errors can't
// cancel against each other.
static std::atomic<int> counter;
return Call::make(t, Call::signed_integer_overflow, {counter++}, Call::Intrinsic);
}
// Make a poison value used when integer div/mod-by-zero is detected during constant folding.
Expr indeterminate_expression_error(Type t) {
// Mark each call with an atomic counter, so that the errors can't
// cancel against each other.
static std::atomic<int> counter;
return Call::make(t, Call::indeterminate_expression, {counter++}, Call::Intrinsic);
}
// If 'e' is indeterminate_expression of type t,
// set *expr to it and return true.
// If 'e' is indeterminate_expression of other type,
// make a new indeterminate_expression of the proper type, set *expr to it and return true.
// Otherwise, leave *expr untouched and return false.
bool propagate_indeterminate_expression(const Expr &e, Type t, Expr *expr) {
const Call *call = e.as<Call>();
if (call && call->is_intrinsic(Call::indeterminate_expression)) {
if (call->type != t) {
*expr = indeterminate_expression_error(t);
} else {
*expr = e;
}
return true;
}
return false;
}
bool propagate_indeterminate_expression(const Expr &e0, const Expr &e1, Type t, Expr *expr) {
return propagate_indeterminate_expression(e0, t, expr) ||
propagate_indeterminate_expression(e1, t, expr);
}
bool propagate_indeterminate_expression(const Expr &e0, const Expr &e1, const Expr &e2, Type t, Expr *expr) {
return propagate_indeterminate_expression(e0, t, expr) ||
propagate_indeterminate_expression(e1, t, expr) ||
propagate_indeterminate_expression(e2, t, expr);
}
#if LOG_EXPR_MUTATIONS || LOG_STMT_MUTATIONS
static int debug_indent = 0;
#endif
}
class Simplify : public IRMutator2 {
public:
Simplify(bool r, const Scope<Interval> *bi, const Scope<ModulusRemainder> *ai) :
remove_dead_lets(r) {
alignment_info.set_containing_scope(ai);
// Only respect the constant bounds from the containing scope.
for (Scope<Interval>::const_iterator iter = bi->cbegin(); iter != bi->cend(); ++iter) {
int64_t i_min, i_max;
if (const_int(iter.value().min, &i_min) &&
const_int(iter.value().max, &i_max)) {
bounds_info.push(iter.name(), { i_min, i_max });
}
}
}
#if LOG_EXPR_MUTATIONS
Expr mutate(const Expr &e) override {
const std::string spaces(debug_indent, ' ');
debug(1) << spaces << "Simplifying Expr: " << e << "\n";
debug_indent++;
Expr new_e = IRMutator2::mutate(e);
debug_indent--;
if (!new_e.same_as(e)) {
debug(1)
<< spaces << "Before: " << e << "\n"
<< spaces << "After: " << new_e << "\n";
}
internal_assert(e.type() == new_e.type());
return new_e;
}
#endif
#if LOG_STMT_MUTATIONS
Stmt mutate(const Stmt &s) override {
const std::string spaces(debug_indent, ' ');
debug(1) << spaces << "Simplifying Stmt: " << s << "\n";
debug_indent++;
Stmt new_s = IRMutator2::mutate(s);
debug_indent--;
if (!new_s.same_as(s)) {
debug(1)
<< spaces << "Before: " << s << "\n"
<< spaces << "After: " << new_s << "\n";
}
return new_s;
}
#endif
using IRMutator2::mutate;
private:
bool remove_dead_lets;
struct VarInfo {
Expr replacement;
int old_uses, new_uses;
};
Scope<VarInfo> var_info;
Scope<pair<int64_t, int64_t>> bounds_info;
Scope<ModulusRemainder> alignment_info;
// If we encounter a reference to a buffer (a Load, Store, Call,
// or Provide), there's an implicit dependence on some associated
// symbols.
void found_buffer_reference(const string &name, size_t dimensions = 0) {
for (size_t i = 0; i < dimensions; i++) {
string stride = name + ".stride." + std::to_string(i);
if (var_info.contains(stride)) {
var_info.ref(stride).old_uses++;
}
string min = name + ".min." + std::to_string(i);
if (var_info.contains(min)) {
var_info.ref(min).old_uses++;
}
}
if (var_info.contains(name)) {
var_info.ref(name).old_uses++;
}
}
using IRMutator2::visit;
// Wrappers for as_const_foo that are more convenient to use in
// the large chains of conditions in the visit methods
// below. Unlike the versions in IROperator, these only match
// scalars.
bool const_float(const Expr &e, double *f) {
if (e.type().is_vector()) {
return false;
} else if (const double *p = as_const_float(e)) {
*f = *p;
return true;
} else {
return false;
}
}
bool const_int(const Expr &e, int64_t *i) {
if (e.type().is_vector()) {
return false;
} else if (const int64_t *p = as_const_int(e)) {
*i = *p;
return true;
} else {
return false;
}
}
bool const_uint(const Expr &e, uint64_t *u) {
if (e.type().is_vector()) {
return false;
} else if (const uint64_t *p = as_const_uint(e)) {
*u = *p;
return true;
} else {
return false;
}
}
// Similar to bounds_of_expr_in_scope, but gives up immediately if
// anything isn't a constant. This stops rules from taking the
// bounds of something then having to simplify it to see whether
// it constant-folds. For some expressions the bounds of the
// expression is at least as complex as the expression, so
// recursively mutating the bounds causes havoc.
bool const_int_bounds(const Expr &e, int64_t *min_val, int64_t *max_val) {
Type t = e.type();
if (const int64_t *i = as_const_int(e)) {
*min_val = *max_val = *i;
return true;
} else if (const Variable *v = e.as<Variable>()) {
if (bounds_info.contains(v->name)) {
pair<int64_t, int64_t> b = bounds_info.get(v->name);
*min_val = b.first;
*max_val = b.second;
return true;
}
} else if (const Broadcast *b = e.as<Broadcast>()) {
return const_int_bounds(b->value, min_val, max_val);
} else if (const Max *max = e.as<Max>()) {
int64_t min_a, min_b, max_a, max_b;
// We only need to check the LHS for Min expr since simplify would
// canonicalize min/max to always be in the LHS.
if (const Min *min = max->a.as<Min>()) {
// Bound of max(min(x, a), b) : [min_b, max(max_a, max_b)].
// We need to check both LHS and RHS of the min, since if a is
// a min/max clamp instead of a constant, simplify would have
// reordered x and a.
if (const_int_bounds(max->b, &min_b, &max_b) &&
(const_int_bounds(min->b, &min_a, &max_a) ||
const_int_bounds(min->a, &min_a, &max_a))) {
*min_val = min_b;
*max_val = std::max(max_a, max_b);
return true;
}
} else if (const_int_bounds(max->a, &min_a, &max_a) &&
const_int_bounds(max->b, &min_b, &max_b)) {
*min_val = std::max(min_a, min_b);
*max_val = std::max(max_a, max_b);
return true;
}
} else if (const Min *min = e.as<Min>()) {
int64_t min_a, min_b, max_a, max_b;
// We only need to check the LHS for Max expr since simplify would
// canonicalize min/max to always be in the LHS.
if (const Max *max = min->a.as<Max>()) {
// Bound of min(max(x, a), b) : [min(min_a, min_b), max_b].
// We need to check both LHS and RHS of the max, since if a is
// a min/max clamp instead of a constant, simplify would have
// reordered x and a.
if (const_int_bounds(min->b, &min_b, &max_b) &&
(const_int_bounds(max->b, &min_a, &max_a) ||
const_int_bounds(max->a, &min_a, &max_a))) {
*min_val = std::min(min_a, min_b);
*max_val = max_b;
return true;
}
} else if (const_int_bounds(min->a, &min_a, &max_a) &&
const_int_bounds(min->b, &min_b, &max_b)) {
*min_val = std::min(min_a, min_b);
*max_val = std::min(max_a, max_b);
return true;
}
} else if (const Select *sel = e.as<Select>()) {
int64_t min_a, min_b, max_a, max_b;
if (const_int_bounds(sel->true_value, &min_a, &max_a) &&
const_int_bounds(sel->false_value, &min_b, &max_b)) {
*min_val = std::min(min_a, min_b);
*max_val = std::max(max_a, max_b);
return true;
}
} else if (const Add *add = e.as<Add>()) {
int64_t min_a, min_b, max_a, max_b;
if (const_int_bounds(add->a, &min_a, &max_a) &&
const_int_bounds(add->b, &min_b, &max_b)) {
*min_val = min_a + min_b;
*max_val = max_a + max_b;
return no_overflow_scalar_int(t.element_of()) ||
(t.can_represent(*min_val) && t.can_represent(*max_val));
}
} else if (const Sub *sub = e.as<Sub>()) {
int64_t min_a, min_b, max_a, max_b;
if (const_int_bounds(sub->a, &min_a, &max_a) &&
const_int_bounds(sub->b, &min_b, &max_b)) {
*min_val = min_a - max_b;
*max_val = max_a - min_b;
return no_overflow_scalar_int(t.element_of()) ||
(t.can_represent(*min_val) && t.can_represent(*max_val));
}
} else if (const Mul *mul = e.as<Mul>()) {
int64_t min_a, min_b, max_a, max_b;
if (const_int_bounds(mul->a, &min_a, &max_a) &&
const_int_bounds(mul->b, &min_b, &max_b)) {
int64_t
t0 = min_a*min_b,
t1 = min_a*max_b,
t2 = max_a*min_b,
t3 = max_a*max_b;
*min_val = std::min(std::min(t0, t1), std::min(t2, t3));
*max_val = std::max(std::max(t0, t1), std::max(t2, t3));
return no_overflow_scalar_int(t.element_of()) ||
(t.can_represent(*min_val) && t.can_represent(*max_val));
}
} else if (const Mod *mod = e.as<Mod>()) {
int64_t min_b, max_b;
if (const_int_bounds(mod->b, &min_b, &max_b) &&
(min_b > 0 || max_b < 0)) {
*min_val = 0;
*max_val = std::max(std::abs(min_b), std::abs(max_b)) - 1;
return no_overflow_scalar_int(t.element_of()) ||
(t.can_represent(*min_val) && t.can_represent(*max_val));
}
} else if (const Div *div = e.as<Div>()) {
int64_t min_a, min_b, max_a, max_b;
if (const_int_bounds(div->a, &min_a, &max_a) &&
const_int_bounds(div->b, &min_b, &max_b) &&
(min_b > 0 || max_b < 0)) {
int64_t
t0 = div_imp(min_a, min_b),
t1 = div_imp(min_a, max_b),
t2 = div_imp(max_a, min_b),
t3 = div_imp(max_a, max_b);
*min_val = std::min(std::min(t0, t1), std::min(t2, t3));
*max_val = std::max(std::max(t0, t1), std::max(t2, t3));
return no_overflow_scalar_int(t.element_of()) ||
(t.can_represent(*min_val) && t.can_represent(*max_val));
}
} else if (const Ramp *r = e.as<Ramp>()) {
int64_t min_base, max_base, min_stride, max_stride;
if (const_int_bounds(r->base, &min_base, &max_base) &&
const_int_bounds(r->stride, &min_stride, &max_stride)) {
int64_t min_last_lane = min_base + min_stride * (r->lanes - 1);
int64_t max_last_lane = max_base + max_stride * (r->lanes - 1);
*min_val = std::min(min_base, min_last_lane);
*max_val = std::max(max_base, max_last_lane);
return no_overflow_scalar_int(t.element_of()) ||
(t.can_represent(*min_val) && t.can_represent(*max_val));
}
}
return false;
}
// Check if an Expr is integer-division-rounding-up by the given
// factor. If so, return the core expression.
Expr is_round_up_div(const Expr &e, int64_t factor) {
if (!no_overflow(e.type())) return Expr();
const Div *div = e.as<Div>();
if (!div) return Expr();
if (!is_const(div->b, factor)) return Expr();
const Add *add = div->a.as<Add>();
if (!add) return Expr();
if (!is_const(add->b, factor-1)) return Expr();
return add->a;
}
// Check if an Expr is a rounding-up operation, and if so, return
// the factor.
Expr is_round_up(const Expr &e, int64_t *factor) {
if (!no_overflow(e.type())) return Expr();
const Mul *mul = e.as<Mul>();
if (!mul) return Expr();
if (!const_int(mul->b, factor)) return Expr();
return is_round_up_div(mul->a, *factor);
}
Expr visit(const Cast *op) override {
Expr value = mutate(op->value);
Expr expr;
if (propagate_indeterminate_expression(value, op->type, &expr)) {
return expr;
}
const Cast *cast = value.as<Cast>();
const Broadcast *broadcast_value = value.as<Broadcast>();
const Ramp *ramp_value = value.as<Ramp>();
const Add *add = value.as<Add>();
double f = 0.0;
int64_t i = 0;
uint64_t u = 0;
if (value.type() == op->type) {
return value;
} else if (op->type.is_int() &&
const_float(value, &f)) {
// float -> int
return IntImm::make(op->type, (int64_t)f);
} else if (op->type.is_uint() &&
const_float(value, &f)) {
// float -> uint
return UIntImm::make(op->type, (uint64_t)f);
} else if (op->type.is_float() &&
const_float(value, &f)) {
// float -> float
return FloatImm::make(op->type, f);
} else if (op->type.is_int() &&
const_int(value, &i)) {
// int -> int
return IntImm::make(op->type, i);
} else if (op->type.is_uint() &&
const_int(value, &i)) {
// int -> uint
return UIntImm::make(op->type, (uint64_t)i);
} else if (op->type.is_float() &&
const_int(value, &i)) {
// int -> float
return FloatImm::make(op->type, (double)i);
} else if (op->type.is_int() &&
const_uint(value, &u)) {
// uint -> int
return IntImm::make(op->type, (int64_t)u);
} else if (op->type.is_uint() &&
const_uint(value, &u)) {
// uint -> uint
return UIntImm::make(op->type, u);
} else if (op->type.is_float() &&
const_uint(value, &u)) {
// uint -> float
return FloatImm::make(op->type, (double)u);
} else if (cast &&
op->type.code() == cast->type.code() &&
op->type.bits() < cast->type.bits()) {
// If this is a cast of a cast of the same type, where the
// outer cast is narrower, the inner cast can be
// eliminated.
return mutate(Cast::make(op->type, cast->value));
} else if (cast &&
(op->type.is_int() || op->type.is_uint()) &&
(cast->type.is_int() || cast->type.is_uint()) &&
op->type.bits() <= cast->type.bits() &&
op->type.bits() <= op->value.type().bits()) {
// If this is a cast between integer types, where the
// outer cast is narrower than the inner cast and the
// inner cast's argument, the inner cast can be
// eliminated. The inner cast is either a sign extend
// or a zero extend, and the outer cast truncates the extended bits
return mutate(Cast::make(op->type, cast->value));
} else if (broadcast_value) {
// cast(broadcast(x)) -> broadcast(cast(x))
return mutate(Broadcast::make(Cast::make(op->type.element_of(), broadcast_value->value), broadcast_value->lanes));
} else if (ramp_value &&
op->type.element_of() == Int(64) &&
op->value.type().element_of() == Int(32)) {
// cast(ramp(a, b, w)) -> ramp(cast(a), cast(b), w)
return mutate(Ramp::make(Cast::make(op->type.element_of(), ramp_value->base),
Cast::make(op->type.element_of(), ramp_value->stride),
ramp_value->lanes));
} else if (add &&
op->type == Int(64) &&
op->value.type() == Int(32) &&
is_const(add->b)) {
// In the interest of moving constants outwards so they
// can cancel, pull the addition outside of the cast.
return mutate(Cast::make(op->type, add->a) + add->b);
} else if (value.same_as(op->value)) {
return op;
} else {
return Cast::make(op->type, value);
}
}
Expr visit(const Variable *op) override {
if (bounds_info.contains(op->name)) {
std::pair<int64_t, int64_t> bounds = bounds_info.get(op->name);
if (bounds.first == bounds.second) {
return make_const(op->type, bounds.first);
}
}
if (var_info.contains(op->name)) {
VarInfo &info = var_info.ref(op->name);
// if replacement is defined, we should substitute it in (unless
// it's a var that has been hidden by a nested scope).
if (info.replacement.defined()) {
internal_assert(info.replacement.type() == op->type) << "Cannot replace variable " << op->name
<< " of type " << op->type << " with expression of type " << info.replacement.type() << "\n";
info.new_uses++;
return info.replacement;
} else {
// This expression was not something deemed
// substitutable - no replacement is defined.
info.old_uses++;
return op;
}
} else {
// We never encountered a let that defines this var. Must
// be a uniform. Don't touch it.
return op;
}
}
Expr visit(const Add *op) override {
int64_t ia = 0, ib = 0, ic = 0;
uint64_t ua = 0, ub = 0;
double fa = 0.0f, fb = 0.0f;
Expr a = mutate(op->a);
Expr b = mutate(op->b);
Expr expr;
if (propagate_indeterminate_expression(a, b, op->type, &expr)) {
return expr;
}
// Rearrange a few patterns to cut down on the number of cases
// to check later.
if ((is_simple_const(a) && !is_simple_const(b)) ||
(b.as<Min>() && !a.as<Min>()) ||
(b.as<Max>() && !a.as<Max>())) {
std::swap(a, b);
}
if ((b.as<Min>() && a.as<Max>())) {
std::swap(a, b);
}
const Call *call_a = a.as<Call>();
const Call *call_b = b.as<Call>();
const Shuffle *shuffle_a = a.as<Shuffle>();
const Shuffle *shuffle_b = b.as<Shuffle>();
const Ramp *ramp_a = a.as<Ramp>();
const Ramp *ramp_b = b.as<Ramp>();
const Broadcast *broadcast_a = a.as<Broadcast>();
const Broadcast *broadcast_b = b.as<Broadcast>();
const Add *add_a = a.as<Add>();
const Add *add_b = b.as<Add>();
const Sub *sub_a = a.as<Sub>();
const Sub *sub_b = b.as<Sub>();
const Mul *mul_a = a.as<Mul>();
const Mul *mul_b = b.as<Mul>();
const Div *div_a = a.as<Div>();
const Div *div_b = b.as<Div>();
const Add *add_div_a_a = div_a ? div_a->a.as<Add>(): nullptr;
const Sub *sub_div_a_a = div_a ? div_a->a.as<Sub>(): nullptr;
const Add *add_div_b_a = div_b ? div_b->a.as<Add>(): nullptr;
const Sub *sub_div_b_a = div_b ? div_b->a.as<Sub>(): nullptr;
const Div *div_a_a = mul_a ? mul_a->a.as<Div>() : nullptr;
const Mod *mod_a = a.as<Mod>();
const Mod *mod_b = b.as<Mod>();
const Mul *mul_a_a = add_a ? add_a->a.as<Mul>(): nullptr;
const Mod *mod_a_a = add_a ? add_a->a.as<Mod>(): nullptr;
const Mul *mul_a_b = add_a ? add_a->b.as<Mul>(): nullptr;
const Mod *mod_a_b = add_a ? add_a->b.as<Mod>(): nullptr;
const Div *div_a_b = add_a ? add_a->b.as<Div>() : nullptr;
const Add *add_a_b_a = div_a_b ? div_a_b->a.as<Add>() : nullptr;
const Max *max_b = b.as<Max>();
const Min *min_a = a.as<Min>();
const Max *max_a = a.as<Max>();
const Sub *sub_a_a = min_a ? min_a->a.as<Sub>() : nullptr;
const Sub *sub_a_b = min_a ? min_a->b.as<Sub>() : nullptr;
const Add *add_a_a = min_a ? min_a->a.as<Add>() : nullptr;
const Add *add_a_b = min_a ? min_a->b.as<Add>() : nullptr;
sub_a_a = max_a ? max_a->a.as<Sub>() : sub_a_a;
sub_a_b = max_a ? max_a->b.as<Sub>() : sub_a_b;
add_a_a = max_a ? max_a->a.as<Add>() : add_a_a;
add_a_b = max_a ? max_a->b.as<Add>() : add_a_b;
add_a_a = mul_a ? mul_a->a.as<Add>() : add_a_a;
add_a_a = div_a ? div_a->a.as<Add>() : add_a_a;
const Div *div_a_a_a = add_a_a ? add_a_a->a.as<Div>() : nullptr;
const Div *div_a_a_b = add_a_a ? add_a_a->b.as<Div>() : nullptr;
const Select *select_a = a.as<Select>();
const Select *select_b = b.as<Select>();
if (const_int(a, &ia) &&
const_int(b, &ib)) {
if (no_overflow(a.type()) &&
add_would_overflow(a.type().bits(), ia, ib)) {
return signed_integer_overflow_error(a.type());
} else {
return IntImm::make(a.type(), ia + ib);
}
} else if (const_uint(a, &ua) &&
const_uint(b, &ub)) {
// const uint + const uint
return UIntImm::make(a.type(), ua + ub);
} else if (const_float(a, &fa) &&
const_float(b, &fb)) {
// const float + const float
return FloatImm::make(a.type(), fa + fb);
} else if (is_zero(b)) {
return a;
} else if (is_zero(a)) {
return b;
} else if (equal(a, b)) {
// x + x = x*2
return mutate(a * make_const(op->type, 2));
} else if (call_a &&
call_a->is_intrinsic(Call::signed_integer_overflow)) {
return a;
} else if (call_b &&
call_b->is_intrinsic(Call::signed_integer_overflow)) {
return b;
} else if (shuffle_a && shuffle_b &&
shuffle_a->is_slice() &&
shuffle_b->is_slice()) {
if (a.same_as(op->a) && b.same_as(op->b)) {
return hoist_slice_vector<Add>(op);
} else {
return hoist_slice_vector<Add>(Add::make(a, b));
}
} else if (ramp_a &&
ramp_b) {
// Ramp + Ramp
return mutate(Ramp::make(ramp_a->base + ramp_b->base,
ramp_a->stride + ramp_b->stride, ramp_a->lanes));
} else if (ramp_a &&
broadcast_b) {
// Ramp + Broadcast
return mutate(Ramp::make(ramp_a->base + broadcast_b->value,
ramp_a->stride, ramp_a->lanes));
} else if (broadcast_a &&
ramp_b) {
// Broadcast + Ramp
return mutate(Ramp::make(broadcast_a->value + ramp_b->base,
ramp_b->stride, ramp_b->lanes));
} else if (broadcast_a &&
broadcast_b) {
// Broadcast + Broadcast
return Broadcast::make(mutate(broadcast_a->value + broadcast_b->value),
broadcast_a->lanes);
} else if (select_a &&
select_b &&
equal(select_a->condition, select_b->condition)) {
// select(c, a, b) + select(c, d, e) -> select(c, a+d, b+e)
return mutate(Select::make(select_a->condition,
select_a->true_value + select_b->true_value,
select_a->false_value + select_b->false_value));
} else if (select_a &&
is_simple_const(b) &&
(is_simple_const(select_a->true_value) ||
is_simple_const(select_a->false_value))) {
// select(c, c1, c2) + c3 -> select(c, c1+c3, c2+c3)
return mutate(Select::make(select_a->condition,
select_a->true_value + b,
select_a->false_value + b));
} else if (add_a &&
is_simple_const(add_a->b)) {
// In ternary expressions, pull constants outside
if (is_simple_const(b)) {
return mutate(add_a->a + (add_a->b + b));
} else {
return mutate((add_a->a + b) + add_a->b);
}
} else if (add_b &&
is_simple_const(add_b->b)) {
return mutate((a + add_b->a) + add_b->b);
} else if (sub_a &&
is_simple_const(sub_a->a)) {
if (is_simple_const(b)) {
return mutate((sub_a->a + b) - sub_a->b);
} else {
return mutate((b - sub_a->b) + sub_a->a);
}
} else if (sub_a &&
equal(b, sub_a->b)) {
// Additions that cancel an inner term
// (a - b) + b
return sub_a->a;
} else if (sub_a &&
is_zero(sub_a->a)) {
return mutate(b - sub_a->b);
} else if (sub_b && equal(a, sub_b->b)) {
// a + (b - a)
return sub_b->a;
} else if (sub_b &&
is_simple_const(sub_b->a)) {
// a + (7 - b) -> (a - b) + 7
return mutate((a - sub_b->b) + sub_b->a);
} else if (sub_a &&
sub_b &&
equal(sub_a->b, sub_b->a)) {
// (a - b) + (b - c) -> a - c
return mutate(sub_a->a - sub_b->b);
} else if (sub_a &&
sub_b &&
equal(sub_a->a, sub_b->b)) {
// (a - b) + (c - a) -> c - b
return mutate(sub_b->a - sub_a->b);
} else if (mul_b &&
is_negative_negatable_const(mul_b->b)) {
// a + b*-x -> a - b*x
return mutate(a - mul_b->a * (-mul_b->b));
} else if (mul_a &&
!is_const(b) && // Leave constants on the right
is_negative_negatable_const(mul_a->b)) {
// a*-x + b -> b - a*x
return mutate(b - mul_a->a * (-mul_a->b));
} else if (mul_b &&
!is_const(a) &&
equal(a, mul_b->a) &&
no_overflow(op->type)) {
// a + a*b -> a*(1 + b)
return mutate(a * (make_one(op->type) + mul_b->b));
} else if (mul_b &&
!is_const(a) &&
equal(a, mul_b->b) &&
no_overflow(op->type)) {
// a + b*a -> (1 + b)*a
return mutate((make_one(op->type) + mul_b->a) * a);
} else if (mul_a &&
!is_const(b) &&
equal(mul_a->a, b) &&
no_overflow(op->type)) {
// a*b + a -> a*(b + 1)
return mutate(mul_a->a * (mul_a->b + make_one(op->type)));
} else if (mul_a &&
!is_const(b) &&
equal(mul_a->b, b) &&
no_overflow(op->type)) {
// a*b + b -> (a + 1)*b
return mutate((mul_a->a + make_one(op->type)) * b);
} else if (no_overflow(op->type) &&
div_a && add_div_a_a &&
is_simple_const(add_div_a_a->b) &&
is_simple_const(div_a->b) &&
is_simple_const(b)) {
// (y + c1)/c2 + c3 -> (y + (c1 + c2*c3))/c2
return mutate((add_div_a_a->a + (add_div_a_a->b + div_a->b*b))/div_a->b);
} else if (no_overflow(op->type) &&
add_a &&
div_a_b &&
add_a_b_a &&
is_simple_const(add_a_b_a->b) &&
is_simple_const(div_a_b->b) &&
is_simple_const(b)) {
// (x + (y + c1)/c2) + c3 -> x + (y + (c1 + c2*c3))/c2
return mutate(add_a->a + (add_a_b_a->a + (add_a_b_a->b + div_a_b->b*b))/div_a_b->b);
} else if (no_overflow(op->type) &&
div_a && sub_div_a_a &&
!is_zero(sub_div_a_a->a) &&
is_simple_const(sub_div_a_a->a) &&
is_simple_const(div_a->b) &&
is_simple_const(b)) {
// (c1 - y)/c2 + c3 + -> ((c1 + c2*c3) - y)/c2
// If c1 == 0, we shouldn't pull in c3 inside the division; otherwise,
// it will cause a cycle with the division simplification rule.
return mutate(((sub_div_a_a->a + div_a->b*b) - sub_div_a_a->b)/div_a->b);
} else if (no_overflow(op->type) &&
div_b && add_div_b_a &&
is_simple_const(div_b->b) &&
equal(a, add_div_b_a->a)) {
// x + (x + y)/c -> ((c + 1)*x + y)/c
return mutate(((div_b->b + 1)*a + add_div_b_a->b)/div_b->b);
} else if (no_overflow(op->type) &&
div_b && sub_div_b_a &&
is_simple_const(div_b->b) &&
equal(a, sub_div_b_a->a)) {
// x + (x - y)/c -> ((c + 1)*x - y)/c
return mutate(((div_b->b + 1)*a - sub_div_b_a->b)/div_b->b);
} else if (no_overflow(op->type) &&
div_b && add_div_b_a &&
is_simple_const(div_b->b) &&
equal(a, add_div_b_a->b)) {
// x + (y + x)/c -> ((c + 1)*x + y)/c
return mutate(((div_b->b + 1)*a + add_div_b_a->a)/div_b->b);
} else if (no_overflow(op->type) &&
div_b && sub_div_b_a &&
is_simple_const(div_b->b) &&
equal(a, sub_div_b_a->b)) {
// x + (y - x)/c -> ((c - 1)*x + y)/c
return mutate(((div_b->b - 1)*a + sub_div_b_a->a)/div_b->b);
} else if (no_overflow(op->type) &&
div_a && add_div_a_a &&
is_simple_const(div_a->b) &&
equal(b, add_div_a_a->a)) {
// (x + y)/c + x + -> ((c + 1)*x + y)/c
return mutate(((div_a->b + 1)*b + add_div_a_a->b)/div_a->b);
} else if (no_overflow(op->type) &&
div_a && sub_div_a_a &&
is_simple_const(div_a->b) &&
equal(b, sub_div_a_a->a)) {
// (x - y)/c + x + -> ((1 + c)*x - y)/c
return mutate(((1 + div_a->b)*b - sub_div_a_a->b)/div_a->b);
} else if (no_overflow(op->type) &&
div_a && add_div_a_a &&
is_simple_const(div_a->b) &&
equal(b, add_div_a_a->b)) {
// (y + x)/c + x -> (y + (1 + c)*x)/c
return mutate((add_div_a_a->a + (1 + div_a->b)*b)/div_a->b);
} else if (no_overflow(op->type) &&
div_a && sub_div_a_a &&
is_simple_const(div_a->b) &&
equal(b, sub_div_a_a->b)) {
// (y - x)/c + x -> (y + (-1 + c)*x)/c
return mutate((sub_div_a_a->a + (- 1 + div_a->b)*b)/div_a->b);
} else if (no_overflow(op->type) &&
min_a &&
sub_a_b &&
equal(sub_a_b->b, b)) {
// min(a, b-c) + c -> min(a+c, b)
return mutate(Min::make(Add::make(min_a->a, b), sub_a_b->a));
} else if (no_overflow(op->type) &&
min_a &&
sub_a_a &&
equal(sub_a_a->b, b)) {
// min(a-c, b) + c -> min(a, b+c)
return mutate(Min::make(sub_a_a->a, Add::make(min_a->b, b)));
} else if (no_overflow(op->type) &&
max_a &&
sub_a_b &&
equal(sub_a_b->b, b)) {
// max(a, b-c) + c -> max(a+c, b)
return mutate(Max::make(Add::make(max_a->a, b), sub_a_b->a));
} else if (no_overflow(op->type) &&
max_a &&
sub_a_a &&
equal(sub_a_a->b, b)) {
// max(a-c, b) + c -> max(a, b+c)
return mutate(Max::make(sub_a_a->a, Add::make(max_a->b, b)));
} else if (no_overflow(op->type) &&
min_a &&
add_a_b &&
const_int(add_a_b->b, &ia) &&
const_int(b, &ib) &&
ia + ib == 0) {
// min(a, b + (-2)) + 2 -> min(a + 2, b)
return mutate(Min::make(Add::make(min_a->a, b), add_a_b->a));
} else if (no_overflow(op->type) &&
min_a &&
add_a_a &&
const_int(add_a_a->b, &ia) &&
const_int(b, &ib) &&
ia + ib == 0) {
// min(a + (-2), b) + 2 -> min(a, b + 2)
return mutate(Min::make(add_a_a->a, Add::make(min_a->b, b)));
} else if (no_overflow(op->type) &&
max_a &&
add_a_b &&
const_int(add_a_b->b, &ia) &&
const_int(b, &ib) &&
ia + ib == 0) {
// max(a, b + (-2)) + 2 -> max(a + 2, b)
return mutate(Max::make(Add::make(max_a->a, b), add_a_b->a));
} else if (no_overflow(op->type) &&
max_a &&
add_a_a &&
const_int(add_a_a->b, &ia) &&
const_int(b, &ib) &&
ia + ib == 0) {
// max(a + (-2), b) + 2 -> max(a, b + 2)
return mutate(Max::make(add_a_a->a, Add::make(max_a->b, b)));
} else if (min_a &&
max_b &&
equal(min_a->a, max_b->a) &&
equal(min_a->b, max_b->b)) {
// min(x, y) + max(x, y) -> x + y
return mutate(min_a->a + min_a->b);
} else if (min_a &&
max_b &&
equal(min_a->a, max_b->b) &&
equal(min_a->b, max_b->a)) {
// min(x, y) + max(y, x) -> x + y
return mutate(min_a->a + min_a->b);
} else if (no_overflow(op->type) &&
div_a &&
add_a_a &&
const_int(add_a_a->b, &ia) &&
const_int(div_a->b, &ib) && ib &&
const_int(b, &ic)) {
// ((a + ia) / ib + ic) -> (a + (ia + ib*ic)) / ib
return mutate((add_a_a->a + IntImm::make(op->type, ia + ib*ic)) / div_a->b);
} else if (mul_a &&
mul_b &&
equal(mul_a->a, mul_b->a)) {
// Pull out common factors a*x + b*x
return mutate(mul_a->a * (mul_a->b + mul_b->b));
} else if (mul_a &&
mul_b &&
equal(mul_a->b, mul_b->a)) {
return mutate(mul_a->b * (mul_a->a + mul_b->b));
} else if (mul_a &&
mul_b &&
equal(mul_a->b, mul_b->b)) {
return mutate(mul_a->b * (mul_a->a + mul_b->a));
} else if (mul_a &&
mul_b &&
equal(mul_a->a, mul_b->b)) {
return mutate(mul_a->a * (mul_a->b + mul_b->a));
} else if (mul_a &&
mul_b &&
const_int(mul_a->b, &ia) &&
const_int(mul_b->b, &ib) &&
(ia % ib == 0 || ib % ia == 0)) {
if (ia % ib == 0) {
Expr factor = make_const(op->type, div_imp(ia, ib));
return mutate((mul_a->a * factor + mul_b->a) * mul_b->b);
} else {
Expr factor = make_const(op->type, div_imp(ib, ia));
return mutate((mul_a->a + mul_b->a * factor) * mul_a->b);
}
} else if (mod_a &&
mul_b &&
equal(mod_a->b, mul_b->b)) {
// (x%3) + y*3 -> y*3 + x%3
return mutate(b + a);
} else if (no_overflow_int(op->type) &&
mul_a &&
mod_b &&
div_a_a &&
equal(mul_a->b, div_a_a->b) &&
equal(mul_a->b, mod_b->b) &&
equal(div_a_a->a, mod_b->a)) {
// (x/3)*3 + x%3 -> x
return div_a_a->a;
} else if (no_overflow_int(op->type) &&
mul_a &&
mod_b &&
add_a_a &&
div_a_a_b &&
equal(mul_a->b, div_a_a_b->b) &&
equal(mul_a->b, mod_b->b) &&
equal(div_a_a_b->a, mod_b->a)) {
// (y + (x/3))*3 + x%3 -> y*3 + x
return mutate(add_a_a->a * mul_a->b + mod_b->a);
} else if (no_overflow_int(op->type) &&
mul_a &&
mod_b &&
add_a_a &&
div_a_a_a &&
equal(mul_a->b, div_a_a_a->b) &&
equal(mul_a->b, mod_b->b) &&
equal(div_a_a_a->a, mod_b->a)) {
// ((x/3) + y)*3 + x%3 -> x + y*3
return mutate(mod_b->a + add_a_a->b * mul_a->b);
} else if (no_overflow_int(op->type) &&
add_a &&
mul_a_a &&
mod_b &&
equal(mul_a_a->b, mod_b->b) &&
(!mod_a_b || !equal(mod_a_b->b, mod_b->b))) {
// ((x*3) + y) + z%3 -> (x*3 + z%3) + y
return mutate((add_a->a + b) + add_a->b);
} else if (no_overflow_int(op->type) &&
add_a &&
mod_a_a &&
mul_b &&
equal(mod_a_a->b, mul_b->b) &&
(!mod_a_b || !equal(mod_a_b->b, mul_b->b))) {
// ((x%3) + y) + z*3 -> (z*3 + x%3) + y
return mutate((b + add_a->a) + add_a->b);
} else if (no_overflow_int(op->type) &&
add_a &&
mul_a_b &&
mod_b &&
equal(mul_a_b->b, mod_b->b) &&
(!mod_a_a || !equal(mod_a_a->b, mod_b->b))) {
// (y + (x*3)) + z%3 -> y + (x*3 + z%3)
return mutate(add_a->a + (add_a->b + b));
} else if (no_overflow_int(op->type) &&
add_a &&
mod_a_b &&
mul_b &&
equal(mod_a_b->b, mul_b->b) &&
(!mod_a_a || !equal(mod_a_a->b, mul_b->b))) {
// (y + (x%3)) + z*3 -> y + (z*3 + x%3)
return mutate(add_a->a + (b + add_a->b));
} else if (mul_a && mul_b &&
const_int(mul_a->b, &ia) &&
const_int(mul_b->b, &ib) &&
ia % ib == 0) {
// x*4 + y*2 -> (x*2 + y)*2
Expr ratio = make_const(a.type(), div_imp(ia, ib));
return mutate((mul_a->a * ratio + mul_b->a) * mul_b->b);
} else if (no_overflow_int(op->type) &&
div_a &&
mod_b &&
is_two(div_a->b) &&
is_two(mod_b->b) &&
equal(div_a->a, mod_b->a)) {
// x / 2 + x % 2 -> (x + 1) / 2
return mutate((div_a->a + make_one(op->type)) / div_a->b);
} else if (no_overflow_int(op->type) &&
div_b &&
mod_a &&
is_two(div_b->b) &&
is_two(mod_a->b) &&
equal(div_b->a, mod_a->a)) {
// x % 2 + x / 2 -> (x + 1) / 2
return mutate((div_b->a + make_one(op->type)) / div_b->b);
} else if (a.same_as(op->a) && b.same_as(op->b)) {
// If we've made no changes, and can't find a rule to apply, return the operator unchanged.
return op;
} else {
return Add::make(a, b);
}
}
Expr visit(const Sub *op) override {
Expr a = mutate(op->a);
Expr b = mutate(op->b);
Expr expr;
if (propagate_indeterminate_expression(a, b, op->type, &expr)) {
return expr;
}
int64_t ia = 0, ib = 0;
uint64_t ua = 0, ub = 0;
double fa = 0.0f, fb = 0.0f;
const Call *call_a = a.as<Call>();
const Call *call_b = b.as<Call>();
const Ramp *ramp_a = a.as<Ramp>();
const Ramp *ramp_b = b.as<Ramp>();
const Broadcast *broadcast_a = a.as<Broadcast>();
const Broadcast *broadcast_b = b.as<Broadcast>();
const Add *add_a = a.as<Add>();
const Add *add_b = b.as<Add>();
const Sub *sub_a = a.as<Sub>();
const Sub *sub_b = b.as<Sub>();
const Mul *mul_a = a.as<Mul>();
const Mul *mul_b = b.as<Mul>();
const Div *div_a_a = mul_a ? mul_a->a.as<Div>() : nullptr;
const Div *div_b_a = mul_b ? mul_b->a.as<Div>() : nullptr;
const Div *div_a = a.as<Div>();
const Div *div_b = b.as<Div>();
const Add *add_div_a_a = div_a ? div_a->a.as<Add>(): nullptr;
const Sub *sub_div_a_a = div_a ? div_a->a.as<Sub>(): nullptr;
const Add *add_div_b_a = div_b ? div_b->a.as<Add>(): nullptr;
const Sub *sub_div_b_a = div_b ? div_b->a.as<Sub>(): nullptr;
const Min *min_b = b.as<Min>();
const Add *add_b_a = min_b ? min_b->a.as<Add>() : nullptr;
const Add *add_b_b = min_b ? min_b->b.as<Add>() : nullptr;
const Min *min_a = a.as<Min>();
const Add *add_a_a = min_a ? min_a->a.as<Add>() : nullptr;
const Add *add_a_b = min_a ? min_a->b.as<Add>() : nullptr;
if (add_a) {
add_a_a = add_a->a.as<Add>();
add_a_b = add_a->b.as<Add>();
}
if (div_a) {
add_a_a = div_a->a.as<Add>();
add_a_b = div_a->b.as<Add>();
}
if (div_b) {
add_b_a = div_b->a.as<Add>();
add_b_b = div_b->b.as<Add>();
}
const Max *max_a = a.as<Max>();
const Max *max_b = b.as<Max>();
const Sub *sub_a_a = div_a ? div_a->a.as<Sub>() : nullptr;
const Sub *sub_b_a = div_b ? div_b->a.as<Sub>() : nullptr;
const Select *select_a = a.as<Select>();
const Select *select_b = b.as<Select>();
int64_t a_round_up_factor = 0, b_round_up_factor = 0;
Expr a_round_up = is_round_up(a, &a_round_up_factor);
Expr b_round_up = is_round_up(b, &b_round_up_factor);
if (is_zero(b)) {
return a;
} else if (equal(a, b)) {
return make_zero(op->type);
} else if (const_int(a, &ia) && const_int(b, &ib)) {
if (no_overflow(a.type()) &&
sub_would_overflow(a.type().bits(), ia, ib)) {
return signed_integer_overflow_error(a.type());
} else {
return IntImm::make(a.type(), ia - ib);
}
} else if (const_uint(a, &ua) && const_uint(b, &ub)) {
return UIntImm::make(a.type(), ua - ub);
} else if (const_float(a, &fa) && const_float(b, &fb)) {
return FloatImm::make(a.type(), fa - fb);
} else if (const_int(b, &ib)) {
return mutate(a + IntImm::make(a.type(), (-ib)));
} else if (const_float(b, &fb)) {
return mutate(a + FloatImm::make(a.type(), (-fb)));
} else if (call_a &&
call_a->is_intrinsic(Call::signed_integer_overflow)) {
return a;
} else if (call_b &&
call_b->is_intrinsic(Call::signed_integer_overflow)) {
return b;
} else if (ramp_a && ramp_b) {
// Ramp - Ramp
return mutate(Ramp::make(ramp_a->base - ramp_b->base,
ramp_a->stride - ramp_b->stride, ramp_a->lanes));
} else if (ramp_a && broadcast_b) {
// Ramp - Broadcast
return mutate(Ramp::make(ramp_a->base - broadcast_b->value,
ramp_a->stride, ramp_a->lanes));
} else if (broadcast_a && ramp_b) {
// Broadcast - Ramp
return mutate(Ramp::make(broadcast_a->value - ramp_b->base,
make_zero(ramp_b->stride.type())- ramp_b->stride,
ramp_b->lanes));
} else if (broadcast_a && broadcast_b) {
// Broadcast + Broadcast
return Broadcast::make(mutate(broadcast_a->value - broadcast_b->value),
broadcast_a->lanes);
} else if (select_a && select_b &&
equal(select_a->condition, select_b->condition)) {
// select(c, a, b) - select(c, d, e) -> select(c, a+d, b+e)
return mutate(Select::make(select_a->condition,
select_a->true_value - select_b->true_value,
select_a->false_value - select_b->false_value));
} else if (select_a &&
equal(select_a->true_value, b)) {
// select(c, a, b) - a -> select(c, 0, b-a)
return mutate(Select::make(select_a->condition,
make_zero(op->type),
select_a->false_value - select_a->true_value));
} else if (select_a &&
equal(select_a->false_value, b)) {
// select(c, a, b) - b -> select(c, a-b, 0)
return mutate(Select::make(select_a->condition,
select_a->true_value - select_a->false_value,
make_zero(op->type)));
} else if (select_b &&
equal(select_b->true_value, a)) {
// a - select(c, a, b) -> select(c, 0, a-b)
return mutate(Select::make(select_b->condition,
make_zero(op->type),
select_b->true_value - select_b->false_value));
} else if (select_b &&
equal(select_b->false_value, a)) {
// b - select(c, a, b) -> select(c, b-a, 0)
return mutate(Select::make(select_b->condition,
select_b->false_value - select_b->true_value,
make_zero(op->type)));
} else if (add_a && equal(add_a->b, b)) {
// Ternary expressions where a term cancels
return add_a->a;
} else if (add_a &&
equal(add_a->a, b)) {
return add_a->b;
} else if (add_b &&
equal(add_b->b, a)) {
return mutate(make_zero(add_b->a.type()) - add_b->a);
} else if (add_b &&
equal(add_b->a, a)) {
return mutate(make_zero(add_b->a.type()) - add_b->b);
} else if (max_a &&
equal(max_a->a, b) &&
!is_const(b) &&
no_overflow(op->type)) {
// max(a, b) - a -> max(0, b-a)
return mutate(Max::make(make_zero(op->type), max_a->b - max_a->a));
} else if (min_a &&
equal(min_a->a, b) &&
!is_const(b) &&
no_overflow(op->type)) {
// min(a, b) - a -> min(0, b-a)
return mutate(Min::make(make_zero(op->type), min_a->b - min_a->a));
} else if (max_a &&
equal(max_a->b, b) &&
!is_const(b) &&
no_overflow(op->type)) {
// max(a, b) - b -> max(a-b, 0)
return mutate(Max::make(max_a->a - max_a->b, make_zero(op->type)));
} else if (min_a &&
equal(min_a->b, b) &&
!is_const(b) &&
no_overflow(op->type)) {
// min(a, b) - b -> min(a-b, 0)
return mutate(Min::make(min_a->a - min_a->b, make_zero(op->type)));
} else if (max_b &&
equal(max_b->a, a) &&
!is_const(a) &&
no_overflow(op->type)) {
// a - max(a, b) -> 0 - max(0, b-a) -> min(0, a-b)
return mutate(Min::make(make_zero(op->type), max_b->a - max_b->b));
} else if (min_b &&
equal(min_b->a, a) &&
!is_const(a) &&
no_overflow(op->type)) {
// a - min(a, b) -> 0 - min(0, b-a) -> max(0, a-b)
return mutate(Max::make(make_zero(op->type), min_b->a - min_b->b));
} else if (max_b &&
equal(max_b->b, a) &&
!is_const(a) &&
no_overflow(op->type)) {
// b - max(a, b) -> 0 - max(a-b, 0) -> min(b-a, 0)
return mutate(Min::make(max_b->b - max_b->a, make_zero(op->type)));
} else if (min_b &&
equal(min_b->b, a) &&
!is_const(a) &&
no_overflow(op->type)) {
// b - min(a, b) -> 0 - min(a-b, 0) -> max(b-a, 0)
return mutate(Max::make(min_b->b - min_b->a, make_zero(op->type)));
} else if (add_a &&
is_simple_const(add_a->b)) {
// In ternary expressions, pull constants outside
if (is_simple_const(b)) {
return mutate(add_a->a + (add_a->b - b));
} else {
return mutate((add_a->a - b) + add_a->b);
}
} else if (sub_a &&
sub_b &&
is_const(sub_a->a) &&
is_const(sub_b->a)) {
// (c1 - a) - (c2 - b) -> (b - a) + (c1 - c2)
return mutate((sub_b->b - sub_a->b) + (sub_a->a - sub_b->a));
} else if (sub_b) {
// a - (b - c) -> a + (c - b)
return mutate(a + (sub_b->b - sub_b->a));
} else if (mul_b &&
is_negative_negatable_const(mul_b->b)) {
// a - b*-x -> a + b*x
return mutate(a + mul_b->a * (-mul_b->b));
} else if (mul_b &&
!is_const(a) &&
equal(a, mul_b->a) &&
no_overflow(op->type)) {
// a - a*b -> a*(1 - b)
return mutate(a * (make_one(op->type) - mul_b->b));
} else if (mul_b &&
!is_const(a) &&
equal(a, mul_b->b) &&
no_overflow(op->type)) {
// a - b*a -> (1 - b)*a
return mutate((make_one(op->type) - mul_b->a) * a);
} else if (mul_a &&
!is_const(b) &&
equal(mul_a->a, b) &&
no_overflow(op->type)) {
// a*b - a -> a*(b - 1)
return mutate(mul_a->a * (mul_a->b - make_one(op->type)));
} else if (mul_a &&
!is_const(b) &&
equal(mul_a->b, b) &&
no_overflow(op->type)) {
// a*b - b -> (a - 1)*b
return mutate((mul_a->a - make_one(op->type)) * b);
} else if (add_b &&
is_simple_const(add_b->b)) {
return mutate((a - add_b->a) - add_b->b);
} else if (sub_a &&
is_simple_const(sub_a->a) &&
is_simple_const(b)) {
return mutate((sub_a->a - b) - sub_a->b);
} else if (mul_a &&
mul_b &&
equal(mul_a->a, mul_b->a)) {
// Pull out common factors a*x + b*x
return mutate(mul_a->a * (mul_a->b - mul_b->b));
} else if (mul_a &&
mul_b &&
equal(mul_a->b, mul_b->a)) {
return mutate(mul_a->b * (mul_a->a - mul_b->b));
} else if (mul_a &&
mul_b &&
equal(mul_a->b, mul_b->b)) {
return mutate(mul_a->b * (mul_a->a - mul_b->a));
} else if (mul_a &&
mul_b &&
equal(mul_a->a, mul_b->b)) {
return mutate(mul_a->a * (mul_a->b - mul_b->a));
} else if (add_a &&
add_b &&
equal(add_a->b, add_b->b)) {
// Quaternary expressions where a term cancels
// (a + b) - (c + b) -> a - c
return mutate(add_a->a - add_b->a);
} else if (add_a &&
add_b &&
equal(add_a->a, add_b->a)) {
// (a + b) - (a + c) -> b - c
return mutate(add_a->b - add_b->b);
} else if (add_a &&
add_b &&
equal(add_a->a, add_b->b)) {
// (a + b) - (c + a) -> b - c
return mutate(add_a->b - add_b->a);
} else if (add_a &&
add_b &&
equal(add_a->b, add_b->a)) {
// (b + a) - (a + c) -> b - c
return mutate(add_a->a - add_b->b);
} else if (add_a &&
add_a_a &&
equal(add_a_a->a, b)) {
// ((a + b) + c) - a -> b + c
return mutate(add_a_a->b + add_a->b);
} else if (add_a &&
add_a_a &&
equal(add_a_a->b, b)) {
// ((a + b) + c) - b -> a + c
return mutate(add_a_a->a + add_a->b);
} else if (add_a &&
add_a_b &&
equal(add_a_b->a, b)) {
// (a + (b + c)) - b -> a + c
return mutate(add_a->a + add_a_b->b);
} else if (add_a &&
add_a_b &&
equal(add_a_b->b, b)) {
// (a + (b + c)) - c -> a + b
return mutate(add_a->a + add_a_b->a);
} else if (no_overflow(op->type) &&
div_b && sub_div_b_a &&
is_simple_const(a) &&
is_simple_const(sub_div_b_a->a) &&
is_simple_const(div_b->b) &&
is_positive_const(div_b->b)) {
// c1 - (c2 - y)/c3 and c3 > 0-> ((c1*c3 - c2 + (c3 - 1)) + y)/c3
return mutate(((a*div_b->b - sub_div_b_a->a) + sub_div_b_a->b + (div_b->b - 1))/div_b->b);
} else if (no_overflow(op->type) &&
div_b && add_div_b_a &&
is_simple_const(a) &&
is_simple_const(add_div_b_a->b) &&
is_simple_const(div_b->b) &&
is_positive_const(div_b->b)) {
// c1 - (y + c2)/c3 and c3 > 0 -> ((c1*c3 - c2 + (c3 - 1)) - y)/c3
return mutate(((a*div_b->b - add_div_b_a->b) - add_div_b_a->a + (div_b->b - 1))/div_b->b);
} else if (no_overflow(op->type) &&
div_b && add_div_b_a &&
is_simple_const(div_b->b) &&
is_positive_const(div_b->b) &&
equal(a, add_div_b_a->a)) {
// x - (x + y)/c and c > 0 -> ((c - 1)*x - y + (c - 1))/c
return mutate(((div_b->b - 1)*a - add_div_b_a->b + (div_b->b - 1))/div_b->b);
} else if (no_overflow(op->type) &&
div_b && sub_div_b_a &&
is_simple_const(div_b->b) &&
is_positive_const(div_b->b) &&
equal(a, sub_div_b_a->a)) {
// x - (x - y)/c and c > 0 -> ((c - 1)*x + y + (c - 1))/c
return mutate(((div_b->b - 1)*a + sub_div_b_a->b + (div_b->b - 1))/div_b->b);
} else if (no_overflow(op->type) &&
div_b && add_div_b_a &&
is_simple_const(div_b->b) &&
is_positive_const(div_b->b) &&
equal(a, add_div_b_a->b)) {
// x - (y + x)/c and c > 0 -> ((c - 1)*x - y + (c - 1))/c
return mutate(((div_b->b - 1)*a - add_div_b_a->a + (div_b->b - 1))/div_b->b);
} else if (no_overflow(op->type) &&
div_b && sub_div_b_a &&
is_simple_const(div_b->b) &&
is_positive_const(div_b->b) &&
equal(a, sub_div_b_a->b)) {
// x - (y - x)/c and c > 0 -> ((c + 1)*x - y + (c - 1))/c
return mutate(((div_b->b + 1)*a - sub_div_b_a->a + (div_b->b - 1))/div_b->b);
} else if (no_overflow(op->type) &&
div_a && add_div_a_a &&
is_simple_const(div_a->b) &&
equal(b, add_div_a_a->a)) {
// (x + y)/c - x + -> ((1 - c)*x + y)/c
return mutate(((1 - div_a->b)*b + add_div_a_a->b)/div_a->b);
} else if (no_overflow(op->type) &&
div_a && sub_div_a_a &&
is_simple_const(div_a->b) &&
equal(b, sub_div_a_a->a)) {
// (x - y)/c - x + -> ((1 - c)*x - y)/c
return mutate(((1 - div_a->b)*b - sub_div_a_a->b)/div_a->b);
} else if (no_overflow(op->type) &&
div_a && add_div_a_a &&
is_simple_const(div_a->b) &&
equal(b, add_div_a_a->b)) {
// (y + x)/c - x -> (y + (1 - c)*x)/c
return mutate((add_div_a_a->a + (1 - div_a->b)*b)/div_a->b);
} else if (no_overflow(op->type) &&
div_a && sub_div_a_a &&
is_simple_const(div_a->b) &&
equal(b, sub_div_a_a->b)) {
// (y - x)/c - x -> (y - (c + 1)*x)/c
return mutate((sub_div_a_a->a - (div_a->b + 1)*b)/div_a->b);
} else if (no_overflow(op->type) &&
min_b &&
add_b_a &&
equal(a, add_b_a->a)) {
// Quaternary expressions involving mins where a term
// cancels. These are important for bounds inference
// simplifications.
// a - min(a + b, c) -> max(-b, a-c)
return mutate(max(0 - add_b_a->b, a - min_b->b));
} else if (no_overflow(op->type) &&
min_b &&
add_b_a &&
equal(a, add_b_a->b)) {
// a - min(b + a, c) -> max(-b, a-c)
return mutate(max(0 - add_b_a->a, a - min_b->b));
} else if (no_overflow(op->type) &&
min_b &&
add_b_b &&
equal(a, add_b_b->a)) {
// a - min(c, a + b) -> max(-b, a-c)
return mutate(max(0 - add_b_b->b, a - min_b->a));
} else if (no_overflow(op->type) &&
min_b &&
add_b_b &&
equal(a, add_b_b->b)) {
// a - min(c, b + a) -> max(-b, a-c)
return mutate(max(0 - add_b_b->a, a - min_b->a));
} else if (no_overflow(op->type) &&
min_a &&
add_a_a &&
equal(b, add_a_a->a)) {
// min(a + b, c) - a -> min(b, c-a)
return mutate(min(add_a_a->b, min_a->b - b));
} else if (no_overflow(op->type) &&
min_a &&
add_a_a &&
equal(b, add_a_a->b)) {
// min(b + a, c) - a -> min(b, c-a)
return mutate(min(add_a_a->a, min_a->b - b));
} else if (no_overflow(op->type) &&
min_a &&
add_a_b &&
equal(b, add_a_b->a)) {
// min(c, a + b) - a -> min(b, c-a)
return mutate(min(add_a_b->b, min_a->a - b));
} else if (no_overflow(op->type) &&
min_a &&
add_a_b &&
equal(b, add_a_b->b)) {
// min(c, b + a) - a -> min(b, c-a)
return mutate(min(add_a_b->a, min_a->a - b));
} else if (min_a &&
min_b &&
equal(min_a->a, min_b->b) &&
equal(min_a->b, min_b->a)) {
// min(a, b) - min(b, a) -> 0
return make_zero(op->type);
} else if (max_a &&
max_b &&
equal(max_a->a, max_b->b) &&
equal(max_a->b, max_b->a)) {
// max(a, b) - max(b, a) -> 0
return make_zero(op->type);
} else if (no_overflow(op->type) &&
min_a &&
min_b &&
is_zero(mutate((min_a->a + min_b->b) - (min_a->b + min_b->a)))) {
// min(a, b) - min(c, d) where a-b == c-d -> b - d
return mutate(min_a->b - min_b->b);
} else if (no_overflow(op->type) &&
max_a &&
max_b &&
is_zero(mutate((max_a->a + max_b->b) - (max_a->b + max_b->a)))) {
// max(a, b) - max(c, d) where a-b == c-d -> b - d
return mutate(max_a->b - max_b->b);
} else if (no_overflow(op->type) &&
min_a &&
min_b &&
is_zero(mutate((min_a->a + min_b->a) - (min_a->b + min_b->b)))) {
// min(a, b) - min(c, d) where a-b == d-c -> b - c
return mutate(min_a->b - min_b->a);
} else if (no_overflow(op->type) &&
max_a &&
max_b &&
is_zero(mutate((max_a->a + max_b->a) - (max_a->b + max_b->b)))) {
// max(a, b) - max(c, d) where a-b == d-c -> b - c
return mutate(max_a->b - max_b->a);
} else if (no_overflow(op->type) &&
(op->type.is_int() || op->type.is_uint()) &&
mul_a &&
div_a_a &&
is_positive_const(mul_a->b) &&
equal(mul_a->b, div_a_a->b) &&
equal(div_a_a->a, b)) {
// (x/4)*4 - x -> -(x%4)
return mutate(make_zero(a.type()) - (b % mul_a->b));
} else if (no_overflow(op->type) &&
(op->type.is_int() || op->type.is_uint()) &&
mul_b &&
div_b_a &&
is_positive_const(mul_b->b) &&
equal(mul_b->b, div_b_a->b) &&
equal(div_b_a->a, a)) {
// x - (x/4)*4 -> x%4
return mutate(a % mul_b->b);
} else if (no_overflow_int(op->type) &&
a_round_up.defined() &&
a_round_up_factor == 2 &&
equal(a_round_up, b)) {
// ((x + 1)/2)*2 - x -> x%2
return mutate(b % make_const(op->type, a_round_up_factor));
} else if (no_overflow_int(op->type) &&
b_round_up.defined() &&
b_round_up_factor == 2 &&
equal(b_round_up, a)) {
// x - ((x + 1)/2)*2 -> -(x%2)
return mutate(make_zero(op->type) - (a % make_const(op->type, b_round_up_factor)));
} else if (mul_a &&
mul_b &&
const_int(mul_a->b, &ia) &&
const_int(mul_b->b, &ib) &&
ib % ia == 0) {
// x * a - y * (a * b) -> (x - y * b) * a
Expr ratio = make_const(a.type(), div_imp(ib, ia));
return mutate((mul_a->a - mul_b->a * ratio) * mul_a->b);
} else if (mul_a &&
mul_b &&
const_int(mul_a->b, &ia) &&
const_int(mul_b->b, &ib) &&
ia % ib == 0) {
// x * (a * b) - y * a -> (x * b - y) * a
Expr ratio = make_const(a.type(), div_imp(ia, ib));
return mutate((mul_a->a * ratio - mul_b->a) * mul_b->b);
} else if (div_a &&
div_b &&
is_positive_const(div_a->b) &&
equal(div_a->b, div_b->b) &&
op->type.is_int() &&
no_overflow(op->type) &&
add_a_a &&
add_b_a &&
equal(add_a_a->a, add_b_a->a) &&
(is_simple_const(add_a_a->b) ||
is_simple_const(add_b_a->b))) {
// This pattern comes up in bounds inference on upsampling code:
// (x + a)/c - (x + b)/c ->
// ((c + a - 1 - b) - (x + a)%c)/c (duplicates a)
// or ((x + b)%c + (a - b))/c (duplicates b)
Expr x = add_a_a->a, a = add_a_a->b, b = add_b_a->b, c = div_a->b;
if (is_simple_const(b)) {
// Use the version that injects two copies of b
return mutate((((x + (b % c)) % c) + (a - b))/c);
} else {
// Use the version that injects two copies of a
return mutate((((c + a - 1) - b) - ((x + (a % c)) % c))/c);
}
} else if (div_a &&
div_b &&
is_positive_const(div_a->b) &&
equal(div_a->b, div_b->b) &&
op->type.is_int() &&
no_overflow(op->type) &&
add_b_a &&
equal(div_a->a, add_b_a->a)) {
// Same as above, where a == 0
Expr x = div_a->a, b = add_b_a->b, c = div_a->b;
return mutate(((c - 1 - b) - (x % c))/c);
} else if (div_a &&
div_b &&
is_positive_const(div_a->b) &&
equal(div_a->b, div_b->b) &&
op->type.is_int() &&
no_overflow(op->type) &&
add_a_a &&
equal(add_a_a->a, div_b->a)) {
// Same as above, where b == 0
Expr x = add_a_a->a, a = add_a_a->b, c = div_a->b;
return mutate(((x % c) + a)/c);
} else if (div_a &&
div_b &&
is_positive_const(div_a->b) &&
equal(div_a->b, div_b->b) &&
op->type.is_int() &&
no_overflow(op->type) &&
sub_b_a &&
equal(div_a->a, sub_b_a->a)) {
// Same as above, where a == 0 and b is subtracted
Expr x = div_a->a, b = sub_b_a->b, c = div_a->b;
return mutate(((c - 1 + b) - (x % c))/c);
} else if (div_a &&
div_b &&
is_positive_const(div_a->b) &&
equal(div_a->b, div_b->b) &&
op->type.is_int() &&
no_overflow(op->type) &&
sub_a_a &&
equal(sub_a_a->a, div_b->a)) {
// Same as above, where b == 0, and a is subtracted
Expr x = sub_a_a->a, a = sub_a_a->b, c = div_a->b;
return mutate(((x % c) - a)/c);
} else if (div_a &&
div_b &&
is_positive_const(div_a->b) &&
equal(div_a->b, div_b->b) &&
op->type.is_int() &&
no_overflow(op->type) &&
sub_a_a &&
add_b_a &&
equal(sub_a_a->a, add_b_a->a) &&
is_simple_const(add_b_a->b)) {
// Same as above, where a is subtracted and b is a constant
// (x - a)/c - (x + b)/c -> ((x + b)%c - a - b)/c
Expr x = sub_a_a->a, a = sub_a_a->b, b = add_b_a->b, c = div_a->b;
return mutate((((x + (b % c)) % c) - a - b)/c);
} else if (div_a &&
div_b &&
is_positive_const(div_a->b) &&
equal(div_a->b, div_b->b) &&
op->type.is_int() &&
no_overflow(op->type) &&
add_a_a &&
sub_b_a &&
equal(add_a_a->a, sub_b_a->a) &&
is_simple_const(add_a_a->b)) {
// Same as above, where b is subtracted and a is a constant
// (x + a)/c - (x - b)/c -> (b - (x + a)%c + (a + c - 1))/c
Expr x = add_a_a->a, a = add_a_a->b, b = sub_b_a->b, c = div_a->b;
return mutate((b - (x + (a % c))%c + (a + c - 1))/c);
} else if (no_overflow(op->type) &&
min_a &&
min_b &&
equal(min_a->a, min_b->a) &&
is_simple_const(min_a->b) &&
is_simple_const(min_b->b)) {
// min(x, c1) - min(x, c2) where c1 and c2 are constants
// if c1 >= c2 -> clamp(x, c2, c1) - c2
// else -> c1 - clamp(x, c1, c2)
if (is_one(mutate(min_a->b >= min_b->b))) {
return mutate(clamp(min_a->a, min_b->b, min_a->b) - min_b->b);
} else {
return mutate(min_a->b - clamp(min_a->a, min_a->b, min_b->b));
}
} else if (no_overflow(op->type) &&
max_a &&
max_b &&
equal(max_a->a, max_b->a) &&
is_simple_const(max_a->b) &&
is_simple_const(max_b->b)) {
// max(x, c1) - max(x, c2) where c1 and c2 are constants
// if c1 >= c2 -> c1 - clamp(x, c2, c1)
// else -> clamp(x, c1, c2) - c2
if (is_one(mutate(max_a->b >= max_b->b))) {
return mutate(max_a->b - clamp(max_a->a, max_b->b, max_a->b));
} else {
return mutate(clamp(max_a->a, max_a->b, max_b->b)- max_b->b);
}
} else if (no_overflow(op->type) &&
min_a &&
min_b) {
// min(a + c1, b + c2) - min(a + c3, b + c4)
// where delta_a = c1 - c3 and delta_b = c2 - c4 are constants
// if delta_b - delta_a <= 0 -> clamp((b + c2) - (a + c1), delta_b - delta_a, 0) + delta_a
// else -> delta_b - clamp((b + c2) - (a + c1), 0, delta_b - delta_a)
Expr delta_a = mutate(min_a->a - min_b->a);
Expr delta_b = mutate(min_a->b - min_b->b);
if (is_simple_const(delta_a) &&
is_simple_const(delta_b)) {
Expr diff = delta_b - delta_a;
if (is_one(mutate(diff <= make_zero(op->type)))) {
return mutate(clamp(min_a->b - min_a->a, diff, make_zero(op->type)) + delta_a);
} else {
return mutate(delta_b - clamp(min_a->b - min_a->a, make_zero(op->type), diff));
}
} else if (is_simple_const(mutate(min_a->a - min_b->b)) &&
is_simple_const(mutate(min_a->b - min_b->a))) {
// Canonicalize min(a + c1, b + c2) - min(b + c4, a + c3)
// where c1, c2, c3, and c4 are constants
// into min(a + c1, b + c2) - min(a + c3, b + c4)
// so that the previous rule can pick it up
return mutate(a - Min::make(min_b->b, min_b->a));
} else if (a.same_as(op->a) && b.same_as(op->b)) {
return op;
} else {
return Sub::make(a, b);
}
} else if (no_overflow(op->type) &&
max_a &&
max_b) {
// max(a + c1, b + c2) - max(a + c3, b + c4)
// where delta_a = c1 - c3 and delta_b = c2 - c4 are constants
// if delta_b - delta_a <= 0 -> delta_b - clamp((b + c2) - (a + c1), delta_b - delta_a, 0)
// else -> clamp((b + c2) - (a + c1), 0, delta_b - delta_a) + delta_a
Expr delta_a = mutate(max_a->a - max_b->a);
Expr delta_b = mutate(max_a->b - max_b->b);
if (is_simple_const(delta_a) &&
is_simple_const(delta_b)) {
Expr diff = delta_b - delta_a;
if (is_one(mutate(diff <= make_zero(op->type)))) {
return mutate(delta_b - clamp(max_a->b - max_a->a, diff, make_zero(op->type)));
} else {
return mutate(clamp(max_a->b - max_a->a, make_zero(op->type), diff) + delta_a);
}
} else if (is_simple_const(mutate(max_a->a - max_b->b)) &&
is_simple_const(mutate(max_a->b - max_b->a))) {
// Canonicalize max(a + c1, b + c2) - max(b + c4, a + c3)
// where c1, c2, c3, and c4 are constants
// into max(a + c1, b + c2) - max(a + c3, b + c4)
// so that the previous rule can pick it up
return mutate(a - Max::make(max_b->b, max_b->a));
} else if (a.same_as(op->a) && b.same_as(op->b)) {
return op;
} else {
return Sub::make(a, b);
}
} else if (a.same_as(op->a) && b.same_as(op->b)) {
return op;
} else {
return Sub::make(a, b);
}
}
Expr visit(const Mul *op) override {
Expr a = mutate(op->a);
Expr b = mutate(op->b);
Expr expr;
if (propagate_indeterminate_expression(a, b, op->type, &expr)) {
return expr;
}
if (is_simple_const(a) ||
(b.as<Min>() && a.as<Max>())) {
std::swap(a, b);
}
int64_t ia = 0, ib = 0;
uint64_t ua = 0, ub = 0;
double fa = 0.0f, fb = 0.0f;
const Call *call_a = a.as<Call>();
const Call *call_b = b.as<Call>();
const Shuffle *shuffle_a = a.as<Shuffle>();
const Shuffle *shuffle_b = b.as<Shuffle>();
const Ramp *ramp_a = a.as<Ramp>();
const Ramp *ramp_b = b.as<Ramp>();
const Broadcast *broadcast_a = a.as<Broadcast>();
const Broadcast *broadcast_b = b.as<Broadcast>();
const Add *add_a = a.as<Add>();
const Sub *sub_a = a.as<Sub>();
const Mul *mul_a = a.as<Mul>();
const Min *min_a = a.as<Min>();
const Mul *mul_b = b.as<Mul>();
const Max *max_b = b.as<Max>();
if (is_zero(a)) {
return a;
} else if (is_zero(b)) {
return b;
} else if (is_one(a)) {
return b;
} else if (is_one(b)) {
return a;
} else if (const_int(a, &ia) && const_int(b, &ib)) {
if (no_overflow(a.type()) &&
mul_would_overflow(a.type().bits(), ia, ib)) {
return signed_integer_overflow_error(a.type());
} else {
return IntImm::make(a.type(), ia * ib);
}
} else if (const_uint(a, &ua) && const_uint(b, &ub)) {
return UIntImm::make(a.type(), ua * ub);
} else if (const_float(a, &fa) && const_float(b, &fb)) {
return FloatImm::make(a.type(), fa * fb);
} else if (call_a &&
call_a->is_intrinsic(Call::signed_integer_overflow)) {
return a;
} else if (call_b &&
call_b->is_intrinsic(Call::signed_integer_overflow)) {
return b;
} else if (shuffle_a && shuffle_b &&
shuffle_a->is_slice() &&
shuffle_b->is_slice()) {
if (a.same_as(op->a) && b.same_as(op->b)) {
return hoist_slice_vector<Mul>(op);
} else {
return hoist_slice_vector<Mul>(Mul::make(a, b));
}
}else if (broadcast_a && broadcast_b) {
return Broadcast::make(mutate(broadcast_a->value * broadcast_b->value), broadcast_a->lanes);
} else if (ramp_a && broadcast_b) {
Expr m = broadcast_b->value;
return mutate(Ramp::make(ramp_a->base * m, ramp_a->stride * m, ramp_a->lanes));
} else if (broadcast_a && ramp_b) {
Expr m = broadcast_a->value;
return mutate(Ramp::make(m * ramp_b->base, m * ramp_b->stride, ramp_b->lanes));
} else if (add_a &&
!(add_a->b.as<Ramp>() && ramp_b) &&
is_simple_const(add_a->b) &&
is_simple_const(b)) {
return mutate(add_a->a * b + add_a->b * b);
} else if (sub_a && is_negative_negatable_const(b)) {
return mutate(Mul::make(Sub::make(sub_a->b, sub_a->a), -b));
} else if (mul_a && is_simple_const(mul_a->b) && is_simple_const(b)) {
return mutate(mul_a->a * (mul_a->b * b));
} else if (mul_b && is_simple_const(mul_b->b)) {
// Pull constants outside
return mutate((a * mul_b->a) * mul_b->b);
} else if (min_a &&
max_b &&
equal(min_a->a, max_b->a) &&
equal(min_a->b, max_b->b)) {
// min(x, y) * max(x, y) -> x*y
return mutate(min_a->a * min_a->b);
} else if (min_a &&
max_b &&
equal(min_a->a, max_b->b) &&
equal(min_a->b, max_b->a)) {
// min(x, y) * max(y, x) -> x*y
return mutate(min_a->a * min_a->b);
} else if (a.same_as(op->a) && b.same_as(op->b)) {
return op;
} else {
return Mul::make(a, b);
}
}
Expr visit(const Div *op) override {
Expr a = mutate(op->a);
Expr b = mutate(op->b);
Expr expr;
if (propagate_indeterminate_expression(a, b, op->type, &expr)) {
return expr;
}
int64_t ia = 0, ib = 0, ic = 0, id = 0;
uint64_t ua = 0, ub = 0;
double fa = 0.0f, fb = 0.0f;
const Mul *mul_a = a.as<Mul>();
const Add *add_a = a.as<Add>();
const Sub *sub_a = a.as<Sub>();
const Div *div_a = a.as<Div>();
const Div *div_a_a = nullptr;
const Mul *mul_a_a = nullptr;
const Mul *mul_a_b = nullptr;
const Add *add_a_a = nullptr;
const Add *add_a_b = nullptr;
const Sub *sub_a_a = nullptr;
const Sub *sub_a_b = nullptr;
const Mul *mul_a_a_a = nullptr;
const Mul *mul_a_a_b = nullptr;
const Mul *mul_a_b_a = nullptr;
const Mul *mul_a_b_b = nullptr;
const Mod *mod_a_a = nullptr;
const Broadcast *broadcast_a = a.as<Broadcast>();
const Ramp *ramp_a = a.as<Ramp>();
const Broadcast *broadcast_b = b.as<Broadcast>();
if (add_a) {
div_a_a = add_a->a.as<Div>();
mul_a_a = add_a->a.as<Mul>();
mul_a_b = add_a->b.as<Mul>();
add_a_a = add_a->a.as<Add>();
add_a_b = add_a->b.as<Add>();
sub_a_a = add_a->a.as<Sub>();
sub_a_b = add_a->b.as<Sub>();
mod_a_a = add_a->a.as<Mod>();
} else if (sub_a) {
mul_a_a = sub_a->a.as<Mul>();
mul_a_b = sub_a->b.as<Mul>();
add_a_a = sub_a->a.as<Add>();
add_a_b = sub_a->b.as<Add>();
sub_a_a = sub_a->a.as<Sub>();
sub_a_b = sub_a->b.as<Sub>();
}
if (add_a_a) {
mul_a_a_a = add_a_a->a.as<Mul>();
mul_a_a_b = add_a_a->b.as<Mul>();
} else if (sub_a_a) {
mul_a_a_a = sub_a_a->a.as<Mul>();
mul_a_a_b = sub_a_a->b.as<Mul>();
}
if (add_a_b) {
mul_a_b_a = add_a_b->a.as<Mul>();
mul_a_b_b = add_a_b->b.as<Mul>();
} else if (sub_a_b) {
mul_a_b_a = sub_a_b->a.as<Mul>();
mul_a_b_b = sub_a_b->b.as<Mul>();
}
if (ramp_a) {
mul_a_a = ramp_a->base.as<Mul>();
}
// Check for bounded numerators divided by constant
// denominators.
int64_t num_min, num_max;
if (const_int(b, &ib) && ib &&
const_int_bounds(a, &num_min, &num_max) &&
div_imp(num_max, ib) == div_imp(num_min, ib)) {
return make_const(op->type, div_imp(num_max, ib));
}
ModulusRemainder mod_rem(0, 1);
if (ramp_a && no_overflow_scalar_int(ramp_a->base.type())) {
// Do modulus remainder analysis on the base.
mod_rem = modulus_remainder(ramp_a->base, alignment_info);
}
if (is_zero(b) && !op->type.is_float()) {
return indeterminate_expression_error(op->type);
} else if (is_zero(a)) {
return a;
} else if (is_one(b)) {
return a;
} else if (equal(a, b)) {
return make_one(op->type);
} else if (const_int(a, &ia) &&
const_int(b, &ib)) {
return IntImm::make(op->type, div_imp(ia, ib));
} else if (const_uint(a, &ua) &&
const_uint(b, &ub)) {
return UIntImm::make(op->type, ua / ub);
} else if (const_float(a, &fa) &&
const_float(b, &fb) &&
fb != 0.0f) {
return FloatImm::make(op->type, fa / fb);
} else if (broadcast_a && broadcast_b) {
return mutate(Broadcast::make(Div::make(broadcast_a->value, broadcast_b->value), broadcast_a->lanes));
} else if (no_overflow_scalar_int(op->type) &&
is_const(a, -1)) {
// -1/x -> select(x < 0, 1, -1)
return mutate(select(b < make_zero(op->type),
make_one(op->type),
make_const(op->type, -1)));
} else if (ramp_a &&
no_overflow_scalar_int(ramp_a->base.type()) &&
const_int(ramp_a->stride, &ia) &&
broadcast_b &&
const_int(broadcast_b->value, &ib) &&
ib &&
ia % ib == 0) {
// ramp(x, 4, w) / broadcast(2, w) -> ramp(x / 2, 2, w)
Type t = op->type.element_of();
return mutate(Ramp::make(ramp_a->base / broadcast_b->value,
IntImm::make(t, div_imp(ia, ib)),
ramp_a->lanes));
} else if (ramp_a &&
no_overflow_scalar_int(ramp_a->base.type()) &&
const_int(ramp_a->stride, &ia) &&
broadcast_b &&
const_int(broadcast_b->value, &ib) &&
ib != 0 &&
(ic = gcd(mod_rem.modulus, ib)) > 1 &&
div_imp((int64_t)mod_rem.remainder, ic) == div_imp(mod_rem.remainder + (ramp_a->lanes-1)*ia, ic)) {
// ramp(k*(a*c) + x, y, w) / (b*c) = broadcast(k/b, w) if x/c == (x + (w-1)*y)/c
// The ramp lanes can't actually change the result, so we
// can just divide the base and broadcast it.
return mutate(Broadcast::make(ramp_a->base / broadcast_b->value, ramp_a->lanes));
} else if (no_overflow(op->type) &&
div_a &&
const_int(div_a->b, &ia) &&
ia >= 0 &&
const_int(b, &ib) &&
ib >= 0) {
// (x / 3) / 4 -> x / 12
return mutate(div_a->a / make_const(op->type, ia * ib));
} else if (no_overflow(op->type) &&
div_a_a &&
add_a &&
const_int(div_a_a->b, &ia) &&
ia >= 0 &&
const_int(add_a->b, &ib) &&
const_int(b, &ic) &&
ic >= 0) {
// (x / ia + ib) / ic -> (x + ia*ib) / (ia*ic)
return mutate((div_a_a->a + make_const(op->type, ia*ib)) / make_const(op->type, ia*ic));
} else if (no_overflow(op->type) &&
mul_a &&
const_int(mul_a->b, &ia) &&
const_int(b, &ib) &&
ia > 0 &&
ib > 0 &&
(ia % ib == 0 || ib % ia == 0)) {
if (ia % ib == 0) {
// (x * 4) / 2 -> x * 2
return mutate(mul_a->a * make_const(op->type, div_imp(ia, ib)));
} else {
// (x * 2) / 4 -> x / 2
return mutate(mul_a->a / make_const(op->type, div_imp(ib, ia)));
}
} else if (no_overflow(op->type) &&
add_a &&
mul_a_a &&
const_int(mul_a_a->b, &ia) &&
const_int(b, &ib) &&
ib > 0 &&
(ia % ib == 0)) {
// Pull terms that are a multiple of the divisor out
// (x*4 + y) / 2 -> x*2 + y/2
Expr ratio = make_const(op->type, div_imp(ia, ib));
return mutate((mul_a_a->a * ratio) + (add_a->b / b));
} else if (no_overflow(op->type) &&
add_a &&
mul_a_b &&
const_int(mul_a_b->b, &ia) &&
const_int(b, &ib) &&
ib > 0 &&
(ia % ib == 0)) {
// (y + x*4) / 2 -> y/2 + x*2
Expr ratio = make_const(op->type, div_imp(ia, ib));
return mutate((add_a->a / b) + (mul_a_b->a * ratio));
} else if (no_overflow(op->type) &&
sub_a &&
mul_a_a &&
const_int(mul_a_a->b, &ia) &&
const_int(b, &ib) &&
ib > 0 &&
(ia % ib == 0)) {
// Pull terms that are a multiple of the divisor out
// (x*4 - y) / 2 -> x*2 + (-y)/2
Expr ratio = make_const(op->type, div_imp(ia, ib));
return mutate((mul_a_a->a * ratio) + (-sub_a->b) / b);
} else if (no_overflow(op->type) &&
sub_a &&
mul_a_b &&
const_int(mul_a_b->b, &ia) &&
const_int(b, &ib) &&
ib > 0 &&
(ia % ib == 0)) {
// (y - x*4) / 2 -> y/2 - x*2
Expr ratio = make_const(op->type, div_imp(ia, ib));
return mutate((sub_a->a / b) - (mul_a_b->a * ratio));
} else if (no_overflow(op->type) &&
add_a &&
add_a_a &&
mul_a_a_a &&
const_int(mul_a_a_a->b, &ia) &&
const_int(b, &ib) &&
ib > 0 &&
(ia % ib == 0)) {
// Pull terms that are a multiple of the divisor out
// ((x*4 + y) + z) / 2 -> x*2 + (y + z)/2
Expr ratio = make_const(op->type, div_imp(ia, ib));
return mutate((mul_a_a_a->a * ratio) + (add_a_a->b + add_a->b) / b);
} else if (no_overflow(op->type) &&
add_a &&
add_a_a &&
mul_a_a_b &&
const_int(mul_a_a_b->b, &ia) &&
const_int(b, &ib) &&
ib > 0 &&
(ia % ib == 0)) {
// ((x + y*4) + z) / 2 -> y*2 + (x + z)/2
Expr ratio = make_const(op->type, div_imp(ia, ib));
return mutate((mul_a_a_b->a * ratio) + (add_a_a->a + add_a->b) / b);
} else if (no_overflow(op->type) &&
add_a &&
sub_a_a &&
mul_a_a_a &&
const_int(mul_a_a_a->b, &ia) &&
const_int(b, &ib) &&
ib > 0 &&
(ia % ib == 0)) {
// ((x*4 - y) + z) / 2 -> x*2 + (z - y)/2
Expr ratio = make_const(op->type, div_imp(ia, ib));
return mutate((mul_a_a_a->a * ratio) + (add_a->b - sub_a_a->b) / b);
} else if (no_overflow(op->type) &&
sub_a &&
add_a_a &&
mul_a_a_a &&
const_int(mul_a_a_a->b, &ia) &&
const_int(b, &ib) &&
ib > 0 &&
(ia % ib == 0)) {
// ((x*4 + y) - z) / 2 -> x*2 + (y - z)/2
Expr ratio = make_const(op->type, div_imp(ia, ib));
return mutate((mul_a_a_a->a * ratio) + (add_a_a->b - sub_a->b) / b);
} else if (no_overflow(op->type) &&
sub_a &&
sub_a_a &&
mul_a_a_a &&
const_int(mul_a_a_a->b, &ia) &&
const_int(b, &ib) &&
ib > 0 &&
(ia % ib == 0)) {
// ((x*4 - y) - z) / 2 -> x*2 + (0 - y - z)/2
Expr ratio = make_const(op->type, div_imp(ia, ib));
return mutate((mul_a_a_a->a * ratio) + (- sub_a_a->b - sub_a->b) / b);
} else if (no_overflow(op->type) &&
add_a &&
add_a_b &&
mul_a_b_a &&
const_int(mul_a_b_a->b, &ia) &&
const_int(b, &ib) &&
ib > 0 &&
(ia % ib == 0)) {
// (x + (y*4 + z)) / 2 -> y*2 + (x + z)/2
Expr ratio = make_const(op->type, div_imp(ia, ib));
return mutate((mul_a_b_a->a * ratio) + (add_a->a + add_a_b->b) / b);
} else if (no_overflow(op->type) &&
add_a &&
sub_a_b &&
mul_a_b_a &&
const_int(mul_a_b_a->b, &ia) &&
const_int(b, &ib) &&
ib > 0 &&
(ia % ib == 0)) {
// (x + (y*4 - z)) / 2 -> y*2 + (x - z)/2
Expr ratio = make_const(op->type, div_imp(ia, ib));
return mutate((mul_a_b_a->a * ratio) + (add_a->a - sub_a_b->b) / b);
} else if (no_overflow(op->type) &&
sub_a &&
add_a_b &&
mul_a_b_a &&
const_int(mul_a_b_a->b, &ia) &&
const_int(b, &ib) &&
ib > 0 &&
(ia % ib == 0)) {
// (x - (y*4 + z)) / 2 -> (x - z)/2 - y*2
Expr ratio = make_const(op->type, div_imp(ia, ib));
return mutate((sub_a->a - add_a_b->b) / b - (mul_a_b_a->a * ratio));
} else if (no_overflow(op->type) &&
add_a &&
sub_a_b &&
mul_a_b_b &&
const_int(mul_a_b_b->b, &ia) &&
const_int(b, &ib) &&
ib > 0 &&
(ia % ib == 0)) {
// (x - (z*4 - y)) / 2 -> (x + (y - z*4)) / 2 -- by a rule from Sub
// (x + (y - z*4)) / 2 -> (x + y)/2 - z*2 -- by this rule
Expr ratio = make_const(op->type, div_imp(ia, ib));
return mutate((add_a->a + sub_a_b->a) / b - (mul_a_b_b->a * ratio));
} else if (no_overflow(op->type) &&
add_a &&
const_int(add_a->b, &ia) &&
const_int(b, &ib) &&
ib > 0 &&
(ia % ib == 0)) {
// (y + 8) / 2 -> y/2 + 4
Expr ratio = make_const(op->type, div_imp(ia, ib));
return mutate((add_a->a / b) + ratio);
} else if (no_overflow(op->type) &&
add_a &&
const_int(add_a->b, &ib) &&
mul_a_a &&
const_int(mul_a_a->b, &ia) &&
const_int(b, &ic) &&
ic > 0 &&
(id = gcd(ia, ic)) != 1) {
// In expressions of the form (x*a + b)/c, we can divide all the constants by gcd(a, c)
// E.g. (y*12 + 5)/9 = (y*4 + 2)/3
ia = div_imp(ia, id);
ib = div_imp(ib, id);
ic = div_imp(ic, id);
return mutate((mul_a_a->a * make_const(op->type, ia) + make_const(op->type, ib)) / make_const(op->type, ic));
} else if (no_overflow(op->type) &&
add_a &&
equal(add_a->a, b)) {
// (x + y)/x -> y/x + 1
return mutate(add_a->b/b + make_one(op->type));
} else if (no_overflow(op->type) &&
add_a &&
equal(add_a->b, b)) {
// (y + x)/x -> y/x + 1
return mutate(add_a->a/b + make_one(op->type));
} else if (no_overflow(op->type) &&
sub_a &&
!is_zero(b) &&
equal(sub_a->a, b)) {
// (x - y)/x -> (-y)/x + 1
return mutate((make_zero(op->type) - sub_a->b)/b + make_one(op->type));
} else if (no_overflow(op->type) &&
sub_a &&
equal(sub_a->b, b)) {
// (y - x)/x -> y/x - 1
return mutate(sub_a->a/b + make_const(op->type, -1));
} else if (no_overflow(op->type) &&
add_a &&
add_a_a &&
equal(add_a_a->a, b)) {
// ((x + y) + z)/x -> ((y + z) + x)/x -> (y+z)/x + 1
return mutate((add_a_a->b + add_a->b)/b + make_one(op->type));
} else if (no_overflow(op->type) &&
add_a &&
add_a_a &&
equal(add_a_a->b, b)) {
// ((y + x) + z)/x -> ((y + z) + x)/x -> (y+z)/x + 1
return mutate((add_a_a->a + add_a->b)/b + make_one(op->type));
} else if (no_overflow(op->type) &&
add_a &&
add_a_b &&
equal(add_a_b->b, b)) {
// (y + (z + x))/x -> ((y + z) + x)/x -> (y+z)/x + 1
return mutate((add_a->a + add_a_b->a)/b + make_one(op->type));
} else if (no_overflow(op->type) &&
add_a &&
add_a_b &&
equal(add_a_b->a, b)) {
// (y + (x + z))/x -> ((y + z) + x)/x -> (y+z)/x + 1
return mutate((add_a->a + add_a_b->b)/b + make_one(op->type));
} else if (no_overflow(op->type) &&
mul_a &&
equal(mul_a->b, b)) {
// (x*y)/y
return mul_a->a;
} else if (no_overflow(op->type) &&
mul_a &&
equal(mul_a->a, b)) {
// (y*x)/y
return mul_a->b;
} else if (no_overflow(op->type) &&
add_a &&
mul_a_a &&
equal(mul_a_a->b, b)) {
// (x*a + y) / a -> x + y/a
return mutate(mul_a_a->a + (add_a->b / b));
} else if (no_overflow(op->type) &&
add_a &&
mul_a_a &&
equal(mul_a_a->a, b)) {
// (a*x + y) / a -> x + y/a
return mutate(mul_a_a->b + (add_a->b / b));
} else if (no_overflow(op->type) &&
add_a &&
mul_a_b &&
equal(mul_a_b->b, b)) {
// (y + x*a) / a -> y/a + x
return mutate((add_a->a / b) + mul_a_b->a);
} else if (no_overflow(op->type) &&
add_a &&
mul_a_b &&
equal(mul_a_b->a, b)) {
// (y + a*x) / a -> y/a + x
return mutate((add_a->a / b) + mul_a_b->b);
} else if (no_overflow(op->type) &&
sub_a &&
mul_a_a &&
equal(mul_a_a->b, b)) {
// (x*a - y) / a -> x + (-y)/a
return mutate(mul_a_a->a + ((make_zero(op->type) - sub_a->b) / b));
} else if (no_overflow(op->type) &&
sub_a &&
mul_a_a &&
equal(mul_a_a->a, b)) {
// (a*x - y) / a -> x + (-y)/a
return mutate(mul_a_a->b + ((make_zero(op->type) - sub_a->b) / b));
} else if (no_overflow(op->type) &&
sub_a &&
mul_a_b &&
equal(mul_a_b->b, b)) {
// (y - x*a) / a -> y/a - x
return mutate((sub_a->a / b) - mul_a_b->a);
} else if (no_overflow(op->type) &&
sub_a &&
mul_a_b &&
equal(mul_a_b->a, b)) {
// (y - a*x) / a -> y/a - x
return mutate((sub_a->a / b) - mul_a_b->b);
} else if (no_overflow_int(op->type) &&
is_two(b) &&
add_a &&
is_const(add_a->b) &&
mod_a_a &&
is_two(mod_a_a->b)) {
// A very specific pattern that comes up in bounds in upsampling code.
// ((x % 2) + c)/2 -> (x % 2) + c / 2 (for odd c)
// We know c is odd or a rule above would have triggered
return mutate(add_a->a + add_a->b / b);
} else if (b.type().is_float() && is_simple_const(b)) {
// Convert const float division to multiplication
// x / 2 -> x * 0.5
return mutate(a * (make_one(b.type()) / b));
} else if (a.same_as(op->a) && b.same_as(op->b)) {
return op;
} else {
return Div::make(a, b);
}
}
Expr visit(const Mod *op) override {
Expr a = mutate(op->a);
Expr b = mutate(op->b);
Expr expr;
if (propagate_indeterminate_expression(a, b, op->type, &expr)) {
return expr;
}
int64_t ia = 0, ib = 0;
uint64_t ua = 0, ub = 0;
double fa = 0.0f, fb = 0.0f;
const Broadcast *broadcast_a = a.as<Broadcast>();
const Broadcast *broadcast_b = b.as<Broadcast>();
const Mul *mul_a = a.as<Mul>();
const Add *add_a = a.as<Add>();
const Mul *mul_a_a = add_a ? add_a->a.as<Mul>() : nullptr;
const Mul *mul_a_b = add_a ? add_a->b.as<Mul>() : nullptr;
const Ramp *ramp_a = a.as<Ramp>();
// If the RHS is a constant, do modulus remainder analysis on the LHS
ModulusRemainder mod_rem(0, 1);
if (const_int(b, &ib) &&
ib &&
no_overflow_scalar_int(op->type)) {
// If the LHS is bounded, we can possibly bail out early
int64_t a_min, a_max;
if (const_int_bounds(a, &a_min, &a_max) &&
a_max < ib && a_min >= 0) {
return a;
}
mod_rem = modulus_remainder(a, alignment_info);
}
// If the RHS is a constant and the LHS is a ramp, do modulus
// remainder analysis on the base.
if (broadcast_b &&
const_int(broadcast_b->value, &ib) &&
ib &&
ramp_a &&
no_overflow_scalar_int(ramp_a->base.type())) {
mod_rem = modulus_remainder(ramp_a->base, alignment_info);
}
if (is_zero(b) && !op->type.is_float()) {
return indeterminate_expression_error(op->type);
} else if (is_one(b) && !op->type.is_float()) {
return make_zero(op->type);
} else if (is_zero(a)) {
return a;
} else if (const_int(a, &ia) && const_int(b, &ib)) {
return IntImm::make(op->type, mod_imp(ia, ib));
} else if (const_uint(a, &ua) && const_uint(b, &ub)) {
return UIntImm::make(op->type, ua % ub);
} else if (const_float(a, &fa) && const_float(b, &fb)) {
return FloatImm::make(op->type, mod_imp(fa, fb));
} else if (broadcast_a && broadcast_b) {
return mutate(Broadcast::make(Mod::make(broadcast_a->value, broadcast_b->value), broadcast_a->lanes));
} else if (no_overflow(op->type) &&
mul_a &&
const_int(b, &ib) &&
ib &&
const_int(mul_a->b, &ia) &&
(ia % ib == 0)) {
// (x * (b*a)) % b -> 0
return make_zero(op->type);
} else if (no_overflow(op->type) &&
mul_a &&
const_int(b, &ib) &&
ib &&
const_int(mul_a->b, &ia) &&
ia > 0 &&
(ib % ia == 0)) {
// (x * a) % (a * b) -> (x % b) * a
Expr ratio = make_const(a.type(), div_imp(ib, ia));
return mutate((mul_a->a % ratio) * mul_a->b);
} else if (no_overflow(op->type) &&
add_a &&
mul_a_a &&
const_int(mul_a_a->b, &ia) &&
const_int(b, &ib) &&
ib &&
(ia % ib == 0)) {
// (x * (b*a) + y) % b -> (y % b)
return mutate(add_a->b % b);
} else if (no_overflow(op->type) &&
add_a &&
const_int(add_a->b, &ia) &&
const_int(b, &ib) &&
ib &&
(ia % ib == 0)) {
// (y + (b*a)) % b -> (y % b)
return mutate(add_a->a % b);
} else if (no_overflow(op->type) &&
add_a &&
mul_a_b &&
const_int(mul_a_b->b, &ia) &&
const_int(b, &ib) &&
ib &&
(ia % ib == 0)) {
// (y + x * (b*a)) % b -> (y % b)
return mutate(add_a->a % b);
} else if (no_overflow_scalar_int(op->type) &&
const_int(b, &ib) &&
ib &&
mod_rem.modulus % ib == 0) {
// ((a*b)*x + c) % a -> c % a
return make_const(op->type, mod_imp((int64_t)mod_rem.remainder, ib));
} else if (no_overflow(op->type) &&
ramp_a &&
const_int(ramp_a->stride, &ia) &&
broadcast_b &&
const_int(broadcast_b->value, &ib) &&
ib &&
ia % ib == 0) {
// ramp(x, 4, w) % broadcast(2, w)
return mutate(Broadcast::make(ramp_a->base % broadcast_b->value, ramp_a->lanes));
} else if (ramp_a &&
no_overflow_scalar_int(ramp_a->base.type()) &&
const_int(ramp_a->stride, &ia) &&
broadcast_b &&
const_int(broadcast_b->value, &ib) &&
ib != 0 &&
mod_rem.modulus % ib == 0 &&
div_imp((int64_t)mod_rem.remainder, ib) == div_imp(mod_rem.remainder + (ramp_a->lanes-1)*ia, ib)) {
// ramp(k*z + x, y, w) % z = ramp(x, y, w) if x/z == (x + (w-1)*y)/z
Expr new_base = make_const(ramp_a->base.type(), mod_imp((int64_t)mod_rem.remainder, ib));
return mutate(Ramp::make(new_base, ramp_a->stride, ramp_a->lanes));
} else if (ramp_a &&
no_overflow_scalar_int(ramp_a->base.type()) &&
const_int(ramp_a->stride, &ia) &&
!is_const(ramp_a->base) &&
broadcast_b &&
const_int(broadcast_b->value, &ib) &&
ib != 0 &&
mod_rem.modulus % ib == 0) {
// ramp(k*z + x, y, w) % z = ramp(x, y, w) % z
Type t = ramp_a->base.type();
Expr new_base = make_const(t, mod_imp((int64_t)mod_rem.remainder, ib));
return mutate(Ramp::make(new_base, ramp_a->stride, ramp_a->lanes) % b);
} else if (a.same_as(op->a) && b.same_as(op->b)) {
return op;
} else {
return Mod::make(a, b);
}
}
Expr visit(const Min *op) override {
Expr a = mutate(op->a);
Expr b = mutate(op->b);
Expr expr;
if (propagate_indeterminate_expression(a, b, op->type, &expr)) {
return expr;
}
// Move constants to the right to cut down on number of cases to check
if (is_simple_const(a) && !is_simple_const(b)) {
std::swap(a, b);
} else if (a.as<Broadcast>() && !b.as<Broadcast>()) {
std::swap(a, b);
} else if (!a.as<Max>() && b.as<Max>()) {
std::swap(a, b);
}
int64_t ia = 0, ib = 0, ic = 0;
uint64_t ua = 0, ub = 0;
double fa = 0.0f, fb = 0.0f;
int64_t a_min, a_max, b_min, b_max;
const Broadcast *broadcast_a = a.as<Broadcast>();
const Broadcast *broadcast_b = b.as<Broadcast>();
const Ramp *ramp_a = a.as<Ramp>();
const Add *add_a = a.as<Add>();
const Add *add_a_a = add_a ? add_a->a.as<Add>() : nullptr;
const Add *add_a_b = add_a ? add_a->b.as<Add>() : nullptr;
const Add *add_b = b.as<Add>();
const Add *add_b_a = add_b ? add_b->a.as<Add>() : nullptr;
const Add *add_b_b = add_b ? add_b->b.as<Add>() : nullptr;
const Mul *mul_a = a.as<Mul>();
const Mul *mul_b = b.as<Mul>();
const Mul *mul_a_a = add_a ? add_a->a.as<Mul>() : nullptr;
const Mul *mul_b_a = add_b ? add_b->a.as<Mul>() : nullptr;
const Div *div_a = a.as<Div>();
const Div *div_b = b.as<Div>();
const Div *div_a_a = add_a ? add_a->a.as<Div>() : mul_a ? mul_a->a.as<Div>() : nullptr;
const Div *div_b_a = add_b ? add_b->a.as<Div>() : mul_b ? mul_b->a.as<Div>() : nullptr;
const Div *div_a_a_a = mul_a_a ? mul_a_a->a.as<Div>() : nullptr;
const Div *div_b_a_a = mul_b_a ? mul_b_a->a.as<Div>() : nullptr;
const Sub *sub_a = a.as<Sub>();
const Sub *sub_b = b.as<Sub>();
const Min *min_a = a.as<Min>();
const Min *min_b = b.as<Min>();
const Min *min_a_a = min_a ? min_a->a.as<Min>() : nullptr;
const Min *min_a_a_a = min_a_a ? min_a_a->a.as<Min>() : nullptr;
const Min *min_a_a_a_a = min_a_a_a ? min_a_a_a->a.as<Min>() : nullptr;
const Max *max_a = a.as<Max>();
const Max *max_b = b.as<Max>();
const Call *call_a = a.as<Call>();
const Call *call_b = b.as<Call>();
const Shuffle *shuffle_a = a.as<Shuffle>();
const Shuffle *shuffle_b = b.as<Shuffle>();
const Select *select_a = a.as<Select>();
const Select *select_b = b.as<Select>();
const Broadcast *broadcast_a_b = min_a ? min_a->b.as<Broadcast>() : nullptr;
min_a_a = max_a ? max_a->a.as<Min>() : min_a_a;
// Detect if the lhs or rhs is a rounding-up operation
int64_t a_round_up_factor = 0, b_round_up_factor = 0;
Expr a_round_up = is_round_up(a, &a_round_up_factor);
Expr b_round_up = is_round_up(b, &b_round_up_factor);
int64_t ramp_min, ramp_max;
if (equal(a, b)) {
return a;
} else if (const_int(a, &ia) &&
const_int(b, &ib)) {
return IntImm::make(op->type, std::min(ia, ib));
} else if (const_uint(a, &ua) &&
const_uint(b, &ub)) {
return UIntImm::make(op->type, std::min(ua, ub));
} else if (const_float(a, &fa) &&
const_float(b, &fb)) {
return FloatImm::make(op->type, std::min(fa, fb));
} else if (const_int(b, &ib) &&
b.type().is_max(ib)) {
// Compute minimum of expression of type and maximum of type --> expression
return a;
} else if (const_int(b, &ib) &&
b.type().is_min(ib)) {
// Compute minimum of expression of type and minimum of type --> min of type
return b;
} else if (const_uint(b, &ub) &&
b.type().is_max(ub)) {
// Compute minimum of expression of type and maximum of type --> expression
return a;
} else if (op->type.is_uint() &&
is_zero(b)) {
// Compute minimum of expression of type and minimum of type --> min of type
return b;
} else if (broadcast_a &&
broadcast_b) {
return mutate(Broadcast::make(Min::make(broadcast_a->value, broadcast_b->value), broadcast_a->lanes));
} else if (const_int_bounds(a, &a_min, &a_max) &&
const_int_bounds(b, &b_min, &b_max)) {
if (a_min >= b_max) {
return b;
} else if (b_min >= a_max) {
return a;
}
} else if (no_overflow(op->type) &&
ramp_a &&
broadcast_b &&
const_int_bounds(ramp_a, &ramp_min, &ramp_max) &&
const_int(broadcast_b->value, &ic)) {
// min(ramp(a, b, n), broadcast(c, n))
if (ramp_min <= ic && ramp_max <= ic) {
// ramp dominates
return a;
} if (ramp_min >= ic && ramp_max >= ic) {
// broadcast dominates
return b;
}
}
if (no_overflow(op->type) &&
add_a &&
const_int(add_a->b, &ia) &&
add_b &&
const_int(add_b->b, &ib) &&
equal(add_a->a, add_b->a)) {
// min(x + 3, x - 2) -> x - 2
if (ia > ib) {
return b;
} else {
return a;
}
} else if (no_overflow(op->type) &&
add_a &&
const_int(add_a->b, &ia) &&
equal(add_a->a, b)) {
// min(x + 5, x) -> x
if (ia > 0) {
return b;
} else {
return a;
}
} else if (no_overflow(op->type) &&
add_b &&
const_int(add_b->b, &ib) &&
equal(add_b->a, a)) {
// min(x, x + 5) -> x
if (ib > 0) {
return a;
} else {
return b;
}
} else if (no_overflow(op->type) &&
sub_a &&
sub_b &&
equal(sub_a->b, sub_b->b) &&
const_int(sub_a->a, &ia) &&
const_int(sub_b->a, &ib)) {
// min (100-x, 101-x) -> 100-x
if (ia < ib) {
return a;
} else {
return b;
}
} else if (a_round_up.defined() &&
equal(a_round_up, b)) {
// min(((a + 3)/4)*4, a) -> a
return b;
} else if (a_round_up.defined() &&
max_b &&
equal(a_round_up, max_b->a) &&
is_const(max_b->b, a_round_up_factor)) {
// min(((a + 3)/4)*4, max(a, 4)) -> max(a, 4)
return b;
} else if (b_round_up.defined() &&
equal(b_round_up, a)) {
// min(a, ((a + 3)/4)*4) -> a
return a;
} else if (b_round_up.defined() &&
max_a &&
equal(b_round_up, max_a->a) &&
is_const(max_a->b, b_round_up_factor)) {
// min(max(a, 4), ((a + 3)/4)*4) -> max(a, 4)
return a;
} else if (mul_a &&
div_a_a &&
is_positive_const(mul_a->b) &&
equal(mul_a->b, div_a_a->b) &&
equal(div_a_a->a, b)) {
// min((a/4)*4, a) -> (a/4)*4
return a;
} else if (mul_b &&
div_b_a &&
is_positive_const(mul_b->b) &&
equal(mul_b->b, div_b_a->b) &&
equal(div_b_a->a, a)) {
// min(a, (a/4)*4) -> (a/4)*4
return b;
} else if (add_a &&
const_int(add_a->b, &ia) &&
mul_a_a &&
div_a_a_a &&
const_int(mul_a_a->b, &ib) &&
const_int(div_a_a_a->b, &ic) &&
ib > 0 &&
ib == ic &&
ia >= ib - 1 &&
equal(div_a_a_a->a, b)) {
// min((b/4)*4 + c, b) -> b (where c >= 3)
return b;
} else if (add_b &&
const_int(add_b->b, &ia) &&
mul_b_a &&
div_b_a_a &&
const_int(mul_b_a->b, &ib) &&
const_int(div_b_a_a->b, &ic) &&
ib > 0 &&
ib == ic &&
ia >= ib - 1 &&
equal(div_b_a_a->a, b)) {
// min(a, (a/4)*4 + c) -> a (where c >= 3)
return a;
} else if (max_a &&
min_b &&
equal(max_a->a, min_b->a) &&
equal(max_a->b, min_b->b)) {
// min(max(x, y), min(x, y)) -> min(x, y)
return mutate(min(max_a->a, max_a->b));
} else if (max_a &&
min_b &&
equal(max_a->a, min_b->b) &&
equal(max_a->b, min_b->a)) {
// min(max(x, y), min(y, x)) -> min(x, y)
return mutate(min(max_a->a, max_a->b));
} else if (max_a &&
(equal(max_a->a, b) || equal(max_a->b, b))) {
// min(max(x, y), x) -> x
// min(max(x, y), y) -> y
return b;
} else if (min_a &&
(equal(min_a->b, b) || equal(min_a->a, b))) {
// min(min(x, y), y) -> min(x, y)
return a;
} else if (min_b &&
(equal(min_b->b, a) || equal(min_b->a, a))) {
// min(y, min(x, y)) -> min(x, y)
return b;
} else if (min_a &&
broadcast_a_b &&
broadcast_b ) {
// min(min(x, broadcast(y, n)), broadcast(z, n))) -> min(x, broadcast(min(y, z), n))
return mutate(Min::make(min_a->a, Broadcast::make(Min::make(broadcast_a_b->value, broadcast_b->value), broadcast_b->lanes)));
} else if (min_a &&
min_a_a &&
equal(min_a_a->b, b)) {
// min(min(min(x, y), z), y) -> min(min(x, y), z)
return a;
} else if (min_a &&
min_a_a_a &&
equal(min_a_a_a->b, b)) {
// min(min(min(min(x, y), z), w), y) -> min(min(min(x, y), z), w)
return a;
} else if (min_a &&
min_a_a_a_a &&
equal(min_a_a_a_a->b, b)) {
// min(min(min(min(min(x, y), z), w), l), y) -> min(min(min(min(x, y), z), w), l)
return a;
} else if (max_a &&
max_b &&
equal(max_a->a, max_b->a)) {
// Distributive law for min/max
// min(max(x, y), max(x, z)) -> max(min(y, z), x)
return mutate(Max::make(Min::make(max_a->b, max_b->b), max_a->a));
} else if (max_a &&
max_b &&
equal(max_a->a, max_b->b)) {
// min(max(x, y), max(z, x)) -> max(min(y, z), x)
return mutate(Max::make(Min::make(max_a->b, max_b->a), max_a->a));
} else if (max_a &&
max_b &&
equal(max_a->b, max_b->a)) {
// min(max(y, x), max(x, z)) -> max(min(y, z), x)
return mutate(Max::make(Min::make(max_a->a, max_b->b), max_a->b));
} else if (max_a &&
max_b &&
equal(max_a->b, max_b->b)) {
// min(max(y, x), max(z, x)) -> max(min(y, z), x)
return mutate(Max::make(Min::make(max_a->a, max_b->a), max_a->b));
} else if (min_a &&
min_b &&
equal(min_a->a, min_b->a)) {
// min(min(x, y), min(x, z)) -> min(min(y, z), x)
return mutate(Min::make(Min::make(min_a->b, min_b->b), min_a->a));
} else if (min_a &&
min_b &&
equal(min_a->a, min_b->b)) {
// min(min(x, y), min(z, x)) -> min(min(y, z), x)
return mutate(Min::make(Min::make(min_a->b, min_b->a), min_a->a));
} else if (min_a &&
min_b &&
equal(min_a->b, min_b->a)) {
// min(min(y, x), min(x, z)) -> min(min(y, z), x)
return mutate(Min::make(Min::make(min_a->a, min_b->b), min_a->b));
} else if (min_a &&
min_b &&
equal(min_a->b, min_b->b)) {
// min(min(y, x), min(z, x)) -> min(min(y, z), x)
return mutate(Min::make(Min::make(min_a->a, min_b->a), min_a->b));
} else if (max_a &&
min_a_a &&
equal(min_a_a->b, b)) {
// min(max(min(x, y), z), y) -> min(max(x, z), y)
return mutate(min(max(min_a_a->a, max_a->b), b));
} else if (max_a &&
min_a_a &&
equal(min_a_a->a, b)) {
// min(max(min(y, x), z), y) -> min(max(x, z), y)
return mutate(min(max(min_a_a->b, max_a->b), b));
} else if (no_overflow(op->type) &&
add_a &&
add_b &&
equal(add_a->b, add_b->b)) {
// Distributive law for addition
// min(a + b, c + b) -> min(a, c) + b
return mutate(min(add_a->a, add_b->a)) + add_a->b;
} else if (no_overflow(op->type) &&
add_a &&
add_b &&
equal(add_a->a, add_b->a)) {
// min(b + a, b + c) -> min(a, c) + b
return mutate(min(add_a->b, add_b->b)) + add_a->a;
} else if (no_overflow(op->type) &&
add_a &&
add_b &&
equal(add_a->a, add_b->b)) {
// min(b + a, c + b) -> min(a, c) + b
return mutate(min(add_a->b, add_b->a)) + add_a->a;
} else if (no_overflow(op->type) &&
add_a &&
add_b &&
equal(add_a->b, add_b->a)) {
// min(a + b, b + c) -> min(a, c) + b
return mutate(min(add_a->a, add_b->b)) + add_a->b;
} else if (no_overflow(op->type) &&
add_a_a &&
add_b &&
equal(add_a_a->a, add_b->a)) {
// min((a + b) + c, a + d) -> min(b + c, d) + a
return mutate(min(add_a_a->b + add_a->b, add_b->b)) + add_b->a;
} else if (no_overflow(op->type) &&
add_a_a &&
add_b &&
equal(add_a_a->b, add_b->a)) {
// min((b + a) + c, a + d) -> min(b + c, d) + a
return mutate(min(add_a_a->a + add_a->b, add_b->b)) + add_b->a;
} else if (no_overflow(op->type) &&
add_a &&
add_b_a &&
equal(add_a->a, add_b_a->a)) {
// min(a + d, (a + b) + c) -> min(d, b + c) + a
return mutate(min(add_a->b, add_b_a->b + add_b->b)) + add_a->a;
} else if (no_overflow(op->type) &&
add_a &&
add_b_a &&
equal(add_a->a, add_b_a->b)) {
// min(a + d, (b + a) + c) -> min(d, b + c) + a
return mutate(min(add_a->b, add_b_a->a + add_b->b)) + add_a->a;
} else if (no_overflow(op->type) &&
add_a_b &&
add_b &&
equal(add_a_b->a, add_b->a)) {
// min(a + (b + c), b + d) -> min(a + c, d) + b
return mutate(min(add_a->a + add_a_b->b, add_b->b)) + add_b->a;
} else if (no_overflow(op->type) &&
add_a_b &&
add_b &&
equal(add_a_b->b, add_b->a)) {
// min(a + (c + b), b + d) -> min(a + c, d) + b
return mutate(min(add_a->a + add_a_b->a, add_b->b)) + add_b->a;
} else if (no_overflow(op->type) &&
add_a &&
add_b_b &&
equal(add_a->a, add_b_b->a)) {
// min(b + d, a + (b + c)) -> min(d, a + c) + b
return mutate(min(add_a->b, add_b->a + add_b_b->b)) + add_a->a;
} else if (no_overflow(op->type) &&
add_a &&
add_b_b &&
equal(add_a->a, add_b_b->b)) {
// min(b + d, a + (c + b)) -> min(d, a + c) + b
return mutate(min(add_a->b, add_b->a + add_b_b->a)) + add_a->a;
} else if (min_a &&
is_simple_const(min_a->b)) {
if (is_simple_const(b)) {
// min(min(x, 4), 5) -> min(x, 4)
return Min::make(min_a->a, mutate(Min::make(b, min_a->b)));
} else {
// min(min(x, 4), y) -> min(min(x, y), 4)
return mutate(Min::make(Min::make(min_a->a, b), min_a->b));
}
} else if (no_overflow(op->type) &&
div_a &&
div_b &&
const_int(div_a->b, &ia) &&
ia &&
const_int(div_b->b, &ib) &&
(ia == ib)) {
// min(a / 4, b / 4) -> min(a, b) / 4
Expr factor = make_const(op->type, ia);
if (ia > 0) {
return mutate(min(div_a->a, div_b->a) / factor);
} else {
return mutate(max(div_a->a, div_b->a) / factor);
}
} else if (no_overflow(op->type) &&
add_a &&
is_simple_const(add_a->b) &&
div_a_a &&
div_b &&
const_int(div_a_a->b, &ia) &&
ia &&
const_int(div_b->b, &ib) &&
(ia == ib)) {
// min(a / 4 + c, b / 4) -> min(a + c*4, b) / 4
Expr factor = make_const(op->type, ia);
if (ia > 0) {
return mutate(min(div_a_a->a + (add_a->b * factor), div_b->a) / factor);
} else {
return mutate(max(div_a_a->a + (add_a->b * factor), div_b->a) / factor);
}
} else if (no_overflow(op->type) &&
add_b &&
is_simple_const(add_b->b) &&
div_b_a &&
div_a &&
const_int(div_b_a->b, &ia) &&
ia &&
const_int(div_a->b, &ib) &&
(ia == ib)) {
// min(a / 4, b / 4 + c) -> min(a, b + c*4) / 4
Expr factor = make_const(op->type, ia);
if (ia > 0) {
return mutate(min(div_a->a, div_b_a->a + (add_b->b * factor)) / factor);
} else {
return mutate(max(div_a->a, div_b_a->a + (add_b->b * factor)) / factor);
}
} else if (no_overflow(op->type) &&
mul_a &&
mul_b &&
const_int(mul_a->b, &ia) &&
const_int(mul_b->b, &ib) &&
(ia == ib)) {
Expr factor = make_const(op->type, ia);
if (ia > 0) {
return mutate(min(mul_a->a, mul_b->a) * factor);
} else {
return mutate(max(mul_a->a, mul_b->a) * factor);
}
} else if (no_overflow(op->type) &&
mul_a &&
const_int(mul_a->b, &ia) &&
const_int(b, &ib) &&
ia &&
(ib % ia == 0)) {
// min(x*8, 24) -> min(x, 3)*8
Expr ratio = make_const(op->type, ib / ia);
Expr factor = make_const(op->type, ia);
if (ia > 0) {
return mutate(min(mul_a->a, ratio) * factor);
} else {
return mutate(max(mul_a->a, ratio) * factor);
}
} else if (no_overflow(op->type) &&
add_a &&
const_int(add_a->b, &ic) &&
mul_a_a &&
mul_b &&
const_int(mul_a_a->b, &ia) &&
ia &&
const_int(mul_b->b, &ib) &&
(ia == ib) &&
(ic % ia == 0)) {
// min(a * 4 + c, b * 4) -> min(a + c / 4, b) * 4 when c % 4 == 0
Expr factor = make_const(op->type, ia);
if (ia > 0) {
return mutate(min(mul_a_a->a + (add_a->b / factor), mul_b->a) * factor);
} else {
return mutate(max(mul_a_a->a + (add_a->b / factor), mul_b->a) * factor);
}
} else if (no_overflow(op->type) &&
add_b &&
const_int(add_b->b, &ic) &&
mul_b_a &&
mul_a &&
const_int(mul_b_a->b, &ia) &&
ia &&
const_int(mul_a->b, &ib) &&
(ia == ib) &&
(ic % ia == 0)) {
// min(a * 4, b * 4 + c) -> min(a, b + c/4) * 4 when c % 4 == 0
Expr factor = make_const(op->type, ia);
if (ia > 0) {
return mutate(min(mul_a->a, mul_b_a->a + (add_b->b / factor)) * factor);
} else {
return mutate(max(mul_a->a, mul_b_a->a + (add_b->b / factor)) * factor);
}
} else if (call_a &&
call_a->is_intrinsic(Call::likely) &&
equal(call_a->args[0], b)) {
// min(likely(b), b) -> likely(b)
return a;
} else if (call_b &&
call_b->is_intrinsic(Call::likely) &&
equal(call_b->args[0], a)) {
// min(a, likely(a)) -> likely(a)
return b;
} else if (shuffle_a && shuffle_b &&
shuffle_a->is_slice() &&
shuffle_b->is_slice()) {
if (a.same_as(op->a) && b.same_as(op->b)) {
return hoist_slice_vector<Min>(op);
} else {
return hoist_slice_vector<Min>(min(a, b));
}
} else if (no_overflow(op->type) &&
sub_a &&
is_const(sub_a->a) &&
is_const(b)) {
// min(8 - x, 3) -> 8 - max(x, 5)
return mutate(sub_a->a - max(sub_a->b, sub_a->a - b));
} else if (select_a &&
select_b &&
equal(select_a->condition, select_b->condition)) {
return mutate(select(select_a->condition,
min(select_a->true_value, select_b->true_value),
min(select_a->false_value, select_b->false_value)));
} else if (a.same_as(op->a) && b.same_as(op->b)) {
return op;
} else {
return Min::make(a, b);
}
}
Expr visit(const Max *op) override {
Expr a = mutate(op->a), b = mutate(op->b);
Expr expr;
if (propagate_indeterminate_expression(a, b, op->type, &expr)) {
return expr;
}
// Move constants to the right to cut down on number of cases to check
if (is_simple_const(a) && !is_simple_const(b)) {
std::swap(a, b);
} else if (a.as<Broadcast>() && !b.as<Broadcast>()) {
std::swap(a, b);
} else if (!a.as<Min>() && b.as<Min>()) {
std::swap(a, b);
}
int64_t ia = 0, ib = 0, ic = 0;
uint64_t ua = 0, ub = 0;
double fa = 0.0f, fb = 0.0f;
int64_t a_min, a_max, b_min, b_max;
const Broadcast *broadcast_a = a.as<Broadcast>();
const Broadcast *broadcast_b = b.as<Broadcast>();
const Ramp *ramp_a = a.as<Ramp>();
const Add *add_a = a.as<Add>();
const Add *add_a_a = add_a ? add_a->a.as<Add>() : nullptr;
const Add *add_a_b = add_a ? add_a->b.as<Add>() : nullptr;
const Add *add_b = b.as<Add>();
const Add *add_b_a = add_b ? add_b->a.as<Add>() : nullptr;
const Add *add_b_b = add_b ? add_b->b.as<Add>() : nullptr;
const Mul *mul_a = a.as<Mul>();
const Mul *mul_b = b.as<Mul>();
const Mul *mul_a_a = add_a ? add_a->a.as<Mul>() : nullptr;
const Mul *mul_b_a = add_b ? add_b->a.as<Mul>() : nullptr;
const Div *div_a = a.as<Div>();
const Div *div_b = b.as<Div>();
const Div *div_a_a = add_a ? add_a->a.as<Div>() : mul_a ? mul_a->a.as<Div>() : nullptr;
const Div *div_b_a = add_b ? add_b->a.as<Div>() : mul_b ? mul_b->a.as<Div>() : nullptr;
const Div *div_a_a_a = mul_a_a ? mul_a_a->a.as<Div>() : nullptr;
const Div *div_b_a_a = mul_b_a ? mul_b_a->a.as<Div>() : nullptr;
const Sub *sub_a = a.as<Sub>();
const Sub *sub_b = b.as<Sub>();
const Max *max_a = a.as<Max>();
const Max *max_b = b.as<Max>();
const Max *max_a_a = max_a ? max_a->a.as<Max>() : nullptr;
const Max *max_a_a_a = max_a_a ? max_a_a->a.as<Max>() : nullptr;
const Max *max_a_a_a_a = max_a_a_a ? max_a_a_a->a.as<Max>() : nullptr;
const Min *min_a = a.as<Min>();
const Min *min_b = b.as<Min>();
const Call *call_a = a.as<Call>();
const Call *call_b = b.as<Call>();
const Shuffle *shuffle_a = a.as<Shuffle>();
const Shuffle *shuffle_b = b.as<Shuffle>();
const Select *select_a = a.as<Select>();
const Select *select_b = b.as<Select>();
const Broadcast *broadcast_a_b = max_a ? max_a->b.as<Broadcast>() : nullptr;
max_a_a = min_a ? min_a->a.as<Max>() : max_a_a;
// Detect if the lhs or rhs is a rounding-up operation
int64_t a_round_up_factor = 0, b_round_up_factor = 0;
Expr a_round_up = is_round_up(a, &a_round_up_factor);
Expr b_round_up = is_round_up(b, &b_round_up_factor);
int64_t ramp_min, ramp_max;
if (equal(a, b)) {
return a;
} else if (const_int(a, &ia) &&
const_int(b, &ib)) {
return IntImm::make(op->type, std::max(ia, ib));
} else if (const_uint(a, &ua) &&
const_uint(b, &ub)) {
return UIntImm::make(op->type, std::max(ua, ub));
} else if (const_float(a, &fa) &&
const_float(b, &fb)) {
return FloatImm::make(op->type, std::max(fa, fb));
} else if (const_int(b, &ib) &&
b.type().is_min(ib)) {
// Compute maximum of expression of type and minimum of type --> expression
return a;
} else if (const_int(b, &ib) &&
b.type().is_max(ib)) {
// Compute maximum of expression of type and maximum of type --> max of type
return b;
} else if (op->type.is_uint() &&
is_zero(b)) {
// Compute maximum of expression of type and minimum of type --> expression
return a;
} else if (const_uint(b, &ub) &&
b.type().is_max(ub)) {
// Compute maximum of expression of type and maximum of type --> max of type
return b;
} else if (broadcast_a && broadcast_b) {
return mutate(Broadcast::make(Max::make(broadcast_a->value, broadcast_b->value), broadcast_a->lanes));
} else if (const_int_bounds(a, &a_min, &a_max) &&
const_int_bounds(b, &b_min, &b_max)) {
if (a_min >= b_max) {
return a;
} else if (b_min >= a_max) {
return b;
}
} else if (no_overflow(op->type) &&
ramp_a &&
broadcast_b &&
const_int_bounds(ramp_a, &ramp_min, &ramp_max) &&
const_int(broadcast_b->value, &ic)) {
// max(ramp(a, b, n), broadcast(c, n))
if (ramp_min >= ic && ramp_max >= ic) {
// ramp dominates
return a;
}
if (ramp_min <= ic && ramp_max <= ic) {
// broadcast dominates
return b;
}
}
if (no_overflow(op->type) &&
add_a &&
const_int(add_a->b, &ia) &&
add_b &&
const_int(add_b->b, &ib) &&
equal(add_a->a, add_b->a)) {
// max(x + 3, x - 2) -> x - 2
if (ia > ib) {
return a;
} else {
return b;
}
} else if (no_overflow(op->type) &&
add_a &&
const_int(add_a->b, &ia) &&
equal(add_a->a, b)) {
// max(x + 5, x)
if (ia > 0) {
return a;
} else {
return b;
}
} else if (no_overflow(op->type) &&
add_b &&
const_int(add_b->b, &ib) &&
equal(add_b->a, a)) {
// max(x, x + 5)
if (ib > 0) {
return b;
} else {
return a;
}
} else if (no_overflow(op->type) &&
sub_a &&
sub_b &&
equal(sub_a->b, sub_b->b) &&
const_int(sub_a->a, &ia) &&
const_int(sub_b->a, &ib)) {
// max (100-x, 101-x) -> 101-x
if (ia > ib) {
return a;
} else {
return b;
}
} else if (a_round_up.defined() &&
equal(a_round_up, b)) {
// max(((a + 3)/4)*4, a) -> ((a + 3)/4)*4
return a;
} else if (b_round_up.defined() &&
equal(b_round_up, a)) {
// max(a, ((a + 3)/4)*4) -> ((a + 3)/4)*4
return b;
} else if (mul_a &&
div_a_a &&
is_positive_const(mul_a->b) &&
equal(mul_a->b, div_a_a->b) &&
equal(div_a_a->a, b)) {
// max((a/4)*4, a) -> a;
return b;
} else if (mul_b &&
div_b_a &&
is_positive_const(mul_b->b) &&
equal(mul_b->b, div_b_a->b) &&
equal(div_b_a->a, a)) {
// max(a, (a/4)*4) -> a
return a;
} else if (add_a &&
const_int(add_a->b, &ia) &&
mul_a_a &&
div_a_a_a &&
const_int(mul_a_a->b, &ib) &&
const_int(div_a_a_a->b, &ic) &&
ib > 0 &&
ib == ic &&
ia >= ib - 1 &&
equal(div_a_a_a->a, b)) {
// max((b/4)*4 + c, b) -> (b/4)*4 + c (where c >= 3)
return a;
} else if (add_b &&
const_int(add_b->b, &ia) &&
mul_b_a &&
div_b_a_a &&
const_int(mul_b_a->b, &ib) &&
const_int(div_b_a_a->b, &ic) &&
ib > 0 &&
ib == ic &&
ia >= ib - 1 &&
equal(div_b_a_a->a, b)) {
// max(a, (a/4)*4 + c) -> (a/4)*4 + c (where c >= 3)
return b;
} else if (min_a &&
max_b &&
equal(min_a->a, max_b->a) &&
equal(min_a->b, max_b->b)) {
// max(min(x, y), max(x, y)) -> max(x, y)
return mutate(max(min_a->a, min_a->b));
} else if (min_a &&
max_b &&
equal(min_a->a, max_b->b) &&
equal(min_a->b, max_b->a)) {
// max(min(x, y), max(y, x)) -> max(x, y)
return mutate(max(min_a->a, min_a->b));
} else if (min_a &&
(equal(min_a->a, b) || equal(min_a->b, b))) {
// max(min(x, y), x) -> x
// max(min(x, y), y) -> y
return b;
} else if (max_a &&
(equal(max_a->b, b) || equal(max_a->a, b))) {
// max(max(x, y), y) -> max(x, y)
return a;
} else if (max_b &&
(equal(max_b->b, a) || equal(max_b->a, a))) {
// max(y, max(x, y)) -> max(x, y)
return b;
} else if (max_a &&
broadcast_a_b &&
broadcast_b ) {
// max(max(x, broadcast(y, n)), broadcast(z, n))) -> max(x, broadcast(max(y, z), n))
return mutate(Max::make(max_a->a, Broadcast::make(Max::make(broadcast_a_b->value, broadcast_b->value), broadcast_b->lanes)));
} else if (max_a &&
max_a_a &&
equal(max_a_a->b, b)) {
// max(max(max(x, y), z), y) -> max(max(x, y), z)
return a;
} else if (max_a_a_a &&
equal(max_a_a_a->b, b)) {
// max(max(max(max(x, y), z), w), y) -> max(max(max(x, y), z), w)
return a;
} else if (max_a_a_a_a &&
equal(max_a_a_a_a->b, b)) {
// max(max(max(max(max(x, y), z), w), l), y) -> max(max(max(max(x, y), z), w), l)
return a;
} else if (max_a &&
max_b &&
equal(max_a->a, max_b->a)) {
// Distributive law for min/max
// max(max(x, y), max(x, z)) -> max(max(y, z), x)
return mutate(Max::make(Max::make(max_a->b, max_b->b), max_a->a));
} else if (max_a &&
max_b &&
equal(max_a->a, max_b->b)) {
// max(max(x, y), max(z, x)) -> max(max(y, z), x)
return mutate(Max::make(Max::make(max_a->b, max_b->a), max_a->a));
} else if (max_a &&
max_b &&
equal(max_a->b, max_b->a)) {
// max(max(y, x), max(x, z)) -> max(max(y, z), x)
return mutate(Max::make(Max::make(max_a->a, max_b->b), max_a->b));
} else if (max_a &&
max_b &&
equal(max_a->b, max_b->b)) {
// max(max(y, x), max(z, x)) -> max(max(y, z), x)
return mutate(Max::make(Max::make(max_a->a, max_b->a), max_a->b));
} else if (min_a &&
min_b &&
equal(min_a->a, min_b->a)) {
// max(min(x, y), min(x, z)) -> min(max(y, z), x)
return mutate(Min::make(Max::make(min_a->b, min_b->b), min_a->a));
} else if (min_a &&
min_b &&
equal(min_a->a, min_b->b)) {
// max(min(x, y), min(z, x)) -> min(max(y, z), x)
return mutate(Min::make(Max::make(min_a->b, min_b->a), min_a->a));
} else if (min_a &&
min_b &&
equal(min_a->b, min_b->a)) {
// max(min(y, x), min(x, z)) -> min(max(y, z), x)
return mutate(Min::make(Max::make(min_a->a, min_b->b), min_a->b));
} else if (min_a &&
min_b &&
equal(min_a->b, min_b->b)) {
// max(min(y, x), min(z, x)) -> min(max(y, z), x)
return mutate(Min::make(Max::make(min_a->a, min_b->a), min_a->b));
} else if (min_a &&
max_a_a &&
equal(max_a_a->b, b)) {
// max(min(max(x, y), z), y) -> max(min(x, z), y)
return mutate(max(min(max_a_a->a, min_a->b), b));
} else if (min_a &&
max_a_a &&
equal(max_a_a->a, b)) {
// max(min(max(y, x), z), y) -> max(min(x, z), y)
return mutate(max(min(max_a_a->b, min_a->b), b));
} else if (no_overflow(op->type) &&
add_a &&
add_b &&
equal(add_a->b, add_b->b)) {
// Distributive law for addition
// max(a + b, c + b) -> max(a, c) + b
return mutate(max(add_a->a, add_b->a)) + add_a->b;
} else if (no_overflow(op->type) &&
add_a &&
add_b &&
equal(add_a->a, add_b->a)) {
// max(b + a, b + c) -> max(a, c) + b
return mutate(max(add_a->b, add_b->b)) + add_a->a;
} else if (no_overflow(op->type) &&
add_a &&
add_b &&
equal(add_a->a, add_b->b)) {
// max(b + a, c + b) -> max(a, c) + b
return mutate(max(add_a->b, add_b->a)) + add_a->a;
} else if (no_overflow(op->type) &&
add_a &&
add_b &&
equal(add_a->b, add_b->a)) {
// max(a + b, b + c) -> max(a, c) + b
return mutate(max(add_a->a, add_b->b)) + add_a->b;
} else if (no_overflow(op->type) &&
add_a_a &&
add_b &&
equal(add_a_a->a, add_b->a)) {
// max((a + b) + c, a + d) -> max(b + c, d) + a
return mutate(max(add_a_a->b + add_a->b, add_b->b)) + add_b->a;
} else if (no_overflow(op->type) &&
add_a_a &&
add_b &&
equal(add_a_a->b, add_b->a)) {
// max((b + a) + c, a + d) -> max(b + c, d) + a
return mutate(max(add_a_a->a + add_a->b, add_b->b)) + add_b->a;
} else if (no_overflow(op->type) &&
add_a_b &&
add_b &&
equal(add_a_b->a, add_b->a)) {
// max(a + (b + c), b + d) -> max(a + c, d) + b
return mutate(max(add_a->a + add_a_b->b, add_b->b)) + add_b->a;
} else if (no_overflow(op->type) &&
add_a_b &&
add_b &&
equal(add_a_b->b, add_b->a)) {
// max(a + (c + b), b + d) -> max(a + c, d) + b
return mutate(max(add_a->a + add_a_b->a, add_b->b)) + add_b->a;
} else if (no_overflow(op->type) &&
add_a &&
add_b_a &&
equal(add_a->a, add_b_a->a)) {
// max(a + d, (a + b) + c) -> max(d, b + c) + a
return mutate(max(add_a->b, add_b_a->b + add_b->b)) + add_a->a;
} else if (no_overflow(op->type) &&
add_a &&
add_b_a &&
equal(add_a->a, add_b_a->b)) {
// max(a + d, (b + a) + c) -> max(d, b + c) + a
return mutate(max(add_a->b, add_b_a->a + add_b->b)) + add_a->a;
} else if (no_overflow(op->type) &&
add_a &&
add_b_b &&
equal(add_a->a, add_b_b->a)) {
// max(b + d, a + (b + c)) -> max(d, a + c) + b
return mutate(max(add_a->b, add_b->a + add_b_b->b)) + add_a->a;
} else if (no_overflow(op->type) &&
add_a &&
add_b_b &&
equal(add_a->a, add_b_b->b)) {
// max(b + d, a + (c + b)) -> max(d, a + c) + b
return mutate(max(add_a->b, add_b->a + add_b_b->a)) + add_a->a;
} else if (max_a && is_simple_const(max_a->b)) {
if (is_simple_const(b)) {
// max(max(x, 4), 5) -> max(x, 4)
return Max::make(max_a->a, mutate(Max::make(b, max_a->b)));
} else {
// max(max(x, 4), y) -> max(max(x, y), 4)
return mutate(Max::make(Max::make(max_a->a, b), max_a->b));
}
} else if (no_overflow(op->type) &&
div_a &&
div_b &&
const_int(div_a->b, &ia) &&
ia &&
const_int(div_b->b, &ib) &&
(ia == ib)) {
// max(a / 4, b / 4) -> max(a, b) / 4
Expr factor = make_const(op->type, ia);
if (ia > 0) {
return mutate(max(div_a->a, div_b->a) / factor);
} else {
return mutate(min(div_a->a, div_b->a) / factor);
}
} else if (no_overflow(op->type) &&
add_a &&
is_simple_const(add_a->b) &&
div_a_a &&
div_b &&
const_int(div_a_a->b, &ia) &&
ia &&
const_int(div_b->b, &ib) &&
(ia == ib)) {
// max(a / 4 + c, b / 4) -> max(a + c*4, b) / 4
Expr factor = make_const(op->type, ia);
if (ia > 0) {
return mutate(max(div_a_a->a + (add_a->b * factor), div_b->a) / factor);
} else {
return mutate(min(div_a_a->a + (add_a->b * factor), div_b->a) / factor);
}
} else if (no_overflow(op->type) &&
add_b &&
is_simple_const(add_b->b) &&
div_b_a &&
div_a &&
const_int(div_b_a->b, &ia) &&
ia &&
const_int(div_a->b, &ib) &&
(ia == ib)) {
// max(a / 4, b / 4 + c) -> max(a, b + c*4) / 4
Expr factor = make_const(op->type, ia);
if (ia > 0) {
return mutate(max(div_a->a, div_b_a->a + (add_b->b * factor)) / factor);
} else {
return mutate(min(div_a->a, div_b_a->a + (add_b->b * factor)) / factor);
}
} else if (no_overflow(op->type) &&
mul_a &&
mul_b &&
const_int(mul_a->b, &ia) &&
const_int(mul_b->b, &ib) &&
(ia == ib)) {
Expr factor = make_const(op->type, ia);
if (ia > 0) {
return mutate(max(mul_a->a, mul_b->a) * factor);
} else {
return mutate(min(mul_a->a, mul_b->a) * factor);
}
} else if (no_overflow(op->type) &&
mul_a &&
const_int(mul_a->b, &ia) &&
const_int(b, &ib) &&
ia &&
(ib % ia == 0)) {
// max(x*8, 24) -> max(x, 3)*8
Expr ratio = make_const(op->type, ib / ia);
Expr factor = make_const(op->type, ia);
if (ia > 0) {
return mutate(max(mul_a->a, ratio) * factor);
} else {
return mutate(min(mul_a->a, ratio) * factor);
}
} else if (no_overflow(op->type) &&
add_a &&
const_int(add_a->b, &ic) &&
mul_a_a &&
mul_b &&
const_int(mul_a_a->b, &ia) &&
ia &&
const_int(mul_b->b, &ib) &&
(ia == ib) &&
(ic % ia == 0)) {
// max(a * 4 + c, b * 4) -> max(a + c / 4, b) * 4 when c % 4 == 0
Expr factor = make_const(op->type, ia);
if (ia > 0) {
return mutate(max(mul_a_a->a + (add_a->b / factor), mul_b->a) * factor);
} else {
return mutate(min(mul_a_a->a + (add_a->b / factor), mul_b->a) * factor);
}
} else if (no_overflow(op->type) &&
add_b &&
const_int(add_b->b, &ic) &&
mul_b_a &&
mul_a &&
const_int(mul_b_a->b, &ia) &&
ia &&
const_int(mul_a->b, &ib) &&
(ia == ib) &&
(ic % ia == 0)) {
// max(a * 4, b * 4 + c) -> max(a, b + c/4) * 4 when c % 4 == 0
Expr factor = make_const(op->type, ia);
if (ia > 0) {
return mutate(max(mul_a->a, mul_b_a->a + (add_b->b / factor)) * factor);
} else {
return mutate(min(mul_a->a, mul_b_a->a + (add_b->b / factor)) * factor);
}
} else if (call_a &&
call_a->is_intrinsic(Call::likely) &&
equal(call_a->args[0], b)) {
// max(likely(b), b) -> likely(b)
return a;
} else if (call_b &&
call_b->is_intrinsic(Call::likely) &&
equal(call_b->args[0], a)) {
// max(a, likely(a)) -> likely(a)
return b;
} else if (shuffle_a && shuffle_b &&
shuffle_a->is_slice() &&
shuffle_b->is_slice()) {
if (a.same_as(op->a) && b.same_as(op->b)) {
return hoist_slice_vector<Max>(op);
} else {
return hoist_slice_vector<Max>(max(a, b));
}
} else if (no_overflow(op->type) &&
sub_a &&
is_simple_const(sub_a->a) &&
is_simple_const(b)) {
// max(8 - x, 3) -> 8 - min(x, 5)
return mutate(sub_a->a - min(sub_a->b, sub_a->a - b));
} else if (select_a &&
select_b &&
equal(select_a->condition, select_b->condition)) {
return mutate(select(select_a->condition,
max(select_a->true_value, select_b->true_value),
max(select_a->false_value, select_b->false_value)));
} else if (a.same_as(op->a) && b.same_as(op->b)) {
return op;
} else {
return Max::make(a, b);
}
}
Expr visit(const EQ *op) override {
Expr delta = mutate(op->a - op->b);
Expr expr;
if (propagate_indeterminate_expression(delta, op->type, &expr)) {
return expr;
}
const Broadcast *broadcast = delta.as<Broadcast>();
const Add *add = delta.as<Add>();
const Sub *sub = delta.as<Sub>();
const Mul *mul = delta.as<Mul>();
const Select *sel = delta.as<Select>();
Expr zero = make_zero(delta.type());
if (is_zero(delta)) {
return const_true(op->type.lanes());
} else if (is_const(delta)) {
bool t = true;
bool f = true;
for (int i = 0; i < delta.type().lanes(); i++) {
Expr deltai = extract_lane(delta, i);
if (is_zero(deltai)) {
f = false;
} else {
t = false;
}
}
if (t) {
return const_true(op->type.lanes());
} else if (f) {
return const_false(op->type.lanes());
}
} else if (no_overflow_scalar_int(delta.type())) {
// Attempt to disprove using modulus remainder analysis
ModulusRemainder mod_rem = modulus_remainder(delta, alignment_info);
if (mod_rem.remainder) {
return const_false();
}
// Attempt to disprove using bounds analysis
int64_t delta_min, delta_max;
if (const_int_bounds(delta, &delta_min, &delta_max) &&
(delta_min > 0 || delta_max < 0)) {
return const_false();
}
}
if (broadcast) {
// Push broadcasts outwards
return Broadcast::make(mutate(broadcast->value ==
make_zero(broadcast->value.type())),
broadcast->lanes);
} else if (add && is_const(add->b)) {
// x + const = 0 -> x = -const
return (add->a == mutate(make_zero(delta.type()) - add->b));
} else if (sub) {
if (is_const(sub->a)) {
// const - x == 0 -> x == const
return sub->b == sub->a;
} else if (sub->a.same_as(op->a) && sub->b.same_as(op->b)) {
return op;
} else {
// x - y == 0 -> x == y
return (sub->a == sub->b);
}
} else if (mul &&
no_overflow(mul->type)) {
// Restrict to int32 and greater, because, e.g. 64 * 4 == 0 as a uint8.
return mutate(mul->a == zero || mul->b == zero);
} else if (sel && is_zero(sel->true_value)) {
// select(c, 0, f) == 0 -> c || (f == 0)
return mutate(sel->condition || (sel->false_value == zero));
} else if (sel &&
(is_positive_const(sel->true_value) || is_negative_const(sel->true_value))) {
// select(c, 4, f) == 0 -> !c && (f == 0)
return mutate((!sel->condition) && (sel->false_value == zero));
} else if (sel && is_zero(sel->false_value)) {
// select(c, t, 0) == 0 -> !c || (t == 0)
return mutate((!sel->condition) || (sel->true_value == zero));
} else if (sel &&
(is_positive_const(sel->false_value) || is_negative_const(sel->false_value))) {
// select(c, t, 4) == 0 -> c && (t == 0)
return mutate((sel->condition) && (sel->true_value == zero));
} else {
return (delta == make_zero(delta.type()));
}
}
Expr visit(const NE *op) override {
return mutate(Not::make(op->a == op->b));
}
Expr visit(const LT *op) override {
Expr a = mutate(op->a), b = mutate(op->b);
Expr expr;
if (propagate_indeterminate_expression(a, b, op->type, &expr)) {
return expr;
}
int64_t a_min, a_max, b_min, b_max;
if (const_int_bounds(a, &a_min, &a_max) &&
const_int_bounds(b, &b_min, &b_max)) {
if (a_max < b_min) {
return const_true(op->type.lanes());
}
if (a_min >= b_max) {
return const_false(op->type.lanes());
}
}
Expr delta = mutate(a - b);
const Ramp *ramp_a = a.as<Ramp>();
const Ramp *ramp_b = b.as<Ramp>();
const Ramp *delta_ramp = delta.as<Ramp>();
const Broadcast *broadcast_a = a.as<Broadcast>();
const Broadcast *broadcast_b = b.as<Broadcast>();
const Add *add_a = a.as<Add>();
const Add *add_b = b.as<Add>();
const Sub *sub_a = a.as<Sub>();
const Sub *sub_b = b.as<Sub>();
const Mul *mul_a = a.as<Mul>();
const Mul *mul_b = b.as<Mul>();
const Div *div_a = a.as<Div>();
const Div *div_b = b.as<Div>();
const Min *min_a = a.as<Min>();
const Min *min_b = b.as<Min>();
const Max *max_a = a.as<Max>();
const Max *max_b = b.as<Max>();
const Div *div_a_a = mul_a ? mul_a->a.as<Div>() : nullptr;
const Add *add_a_a_a = div_a_a ? div_a_a->a.as<Add>() : nullptr;
int64_t ia = 0, ib = 0, ic = 0;
uint64_t ua = 0, ub = 0;
double fa, fb;
ModulusRemainder mod_rem(0, 1);
if (delta_ramp &&
no_overflow_scalar_int(delta_ramp->base.type())) {
// Do modulus remainder analysis on the base.
mod_rem = modulus_remainder(delta_ramp->base, alignment_info);
}
// Note that the computation of delta could be incorrect if
// ia and/or ib are large unsigned integer constants, especially when
// int is 32 bits on the machine.
// Explicit comparison is preferred.
if (const_int(a, &ia) &&
const_int(b, &ib)) {
return make_bool(ia < ib, op->type.lanes());
} else if (const_uint(a, &ua) &&
const_uint(b, &ub)) {
return make_bool(ua < ub, op->type.lanes());
} else if (const_float(a, &fa) &&
const_float(b, &fb)) {
return make_bool(fa < fb, op->type.lanes());
} else if (const_int(a, &ia) &&
a.type().is_max(ia)) {
// Comparing maximum of type < expression of type. This can never be true.
return const_false(op->type.lanes());
} else if (const_int(b, &ib) &&
b.type().is_min(ib)) {
// Comparing expression of type < minimum of type. This can never be true.
return const_false(op->type.lanes());
} else if (no_overflow(delta.type()) &&
const_int_bounds(delta, &ia, &ib) &&
(ia >= 0 || ib < 0)) {
return make_bool(ib < 0, op->type.lanes());
} else if (broadcast_a &&
broadcast_b) {
// Push broadcasts outwards
return mutate(Broadcast::make(broadcast_a->value < broadcast_b->value, broadcast_a->lanes));
} else if (no_overflow(delta.type())) {
if (ramp_a &&
ramp_b &&
equal(ramp_a->stride, ramp_b->stride)) {
// Ramps with matching stride
Expr bases_lt = (ramp_a->base < ramp_b->base);
return mutate(Broadcast::make(bases_lt, ramp_a->lanes));
} else if (add_a &&
add_b &&
equal(add_a->a, add_b->a)) {
// Subtract a term from both sides
return mutate(add_a->b < add_b->b);
} else if (add_a &&
add_b &&
equal(add_a->a, add_b->b)) {
return mutate(add_a->b < add_b->a);
} else if (add_a &&
add_b &&
equal(add_a->b, add_b->a)) {
return mutate(add_a->a < add_b->b);
} else if (add_a &&
add_b &&
equal(add_a->b, add_b->b)) {
return mutate(add_a->a < add_b->a);
} else if (sub_a &&
sub_b &&
equal(sub_a->a, sub_b->a)) {
// Add a term to both sides and negate.
return mutate(sub_b->b < sub_a->b);
} else if (sub_a &&
sub_b &&
equal(sub_a->b, sub_b->b)) {
return mutate(sub_a->a < sub_b->a);
} else if (add_a) {
// Rearrange so that all adds and subs are on the rhs to cut down on further cases
return mutate(add_a->a < (b - add_a->b));
} else if (sub_a) {
return mutate(sub_a->a < (b + sub_a->b));
} else if (add_b &&
equal(add_b->a, a)) {
// Subtract a term from both sides
return mutate(make_zero(add_b->b.type()) < add_b->b);
} else if (add_b &&
equal(add_b->b, a)) {
return mutate(make_zero(add_b->a.type()) < add_b->a);
} else if (add_b &&
is_simple_const(a) &&
is_simple_const(add_b->b)) {
// a < x + b -> (a - b) < x
return mutate((a - add_b->b) < add_b->a);
} else if (sub_b &&
equal(sub_b->a, a)) {
// Subtract a term from both sides
return mutate(sub_b->b < make_zero(sub_b->b.type()));
} else if (sub_b &&
is_const(a) &&
is_const(sub_b->a) &&
!is_const(sub_b->b)) {
// (c1 < c2 - x) -> (x < c2 - c1)
return mutate(sub_b->b < (sub_b->a - a));
} else if (mul_a &&
mul_b &&
is_positive_const(mul_a->b) &&
is_positive_const(mul_b->b) &&
equal(mul_a->b, mul_b->b)) {
// Divide both sides by a constant
return mutate(mul_a->a < mul_b->a);
} else if (mul_a &&
is_positive_const(mul_a->b) &&
is_const(b)) {
if (mul_a->type.is_int()) {
// (a * c1 < c2) <=> (a < (c2 - 1) / c1 + 1)
return mutate(mul_a->a < (((b - 1) / mul_a->b) + 1));
} else {
// (a * c1 < c2) <=> (a < c2 / c1)
return mutate(mul_a->a < (b / mul_a->b));
}
} else if (mul_b &&
is_positive_const(mul_b->b) &&
is_simple_const(mul_b->b) &&
is_simple_const(a)) {
// (c1 < b * c2) <=> ((c1 / c2) < b)
return mutate((a / mul_b->b) < mul_b->a);
} else if (a.type().is_int() &&
div_a &&
is_positive_const(div_a->b) &&
is_const(b)) {
// a / c1 < c2 <=> a < c1*c2
return mutate(div_a->a < (div_a->b * b));
} else if (a.type().is_int() &&
div_b &&
is_positive_const(div_b->b) &&
is_const(a)) {
// c1 < b / c2 <=> (c1+1)*c2-1 < b
Expr one = make_one(a.type());
return mutate((a + one)*div_b->b - one < div_b->a);
} else if (min_a) {
// (min(a, b) < c) <=> (a < c || b < c)
// See if that would simplify usefully:
Expr lt_a = mutate(min_a->a < b);
Expr lt_b = mutate(min_a->b < b);
if (is_const(lt_a) || is_const(lt_b)) {
return mutate(lt_a || lt_b);
} else if (a.same_as(op->a) && b.same_as(op->b)) {
return op;
} else {
return LT::make(a, b);
}
} else if (max_a) {
// (max(a, b) < c) <=> (a < c && b < c)
Expr lt_a = mutate(max_a->a < b);
Expr lt_b = mutate(max_a->b < b);
if (is_const(lt_a) || is_const(lt_b)) {
return mutate(lt_a && lt_b);
} else if (a.same_as(op->a) && b.same_as(op->b)) {
return op;
} else {
return LT::make(a, b);
}
} else if (min_b) {
// (a < min(b, c)) <=> (a < b && a < c)
Expr lt_a = mutate(a < min_b->a);
Expr lt_b = mutate(a < min_b->b);
if (is_const(lt_a) || is_const(lt_b)) {
return mutate(lt_a && lt_b);
} else if (a.same_as(op->a) && b.same_as(op->b)) {
return op;
} else {
return LT::make(a, b);
}
} else if (max_b) {
// (a < max(b, c)) <=> (a < b || a < c)
Expr lt_a = mutate(a < max_b->a);
Expr lt_b = mutate(a < max_b->b);
if (is_const(lt_a) || is_const(lt_b)) {
return mutate(lt_a || lt_b);
} else if (a.same_as(op->a) && b.same_as(op->b)) {
return op;
} else {
return LT::make(a, b);
}
} else if (mul_a &&
div_a_a &&
const_int(div_a_a->b, &ia) &&
const_int(mul_a->b, &ib) &&
ia > 0 &&
ia == ib &&
equal(div_a_a->a, b)) {
// subtract (x/c1)*c1 from both sides
// (x/c1)*c1 < x -> 0 < x % c1
return mutate(0 < b % make_const(a.type(), ia));
} else if (mul_a &&
div_a_a &&
add_b &&
const_int(div_a_a->b, &ia) &&
const_int(mul_a->b, &ib) &&
ia > 0 &&
ia == ib &&
equal(div_a_a->a, add_b->a)) {
// subtract (x/c1)*c1 from both sides
// (x/c1)*c1 < x + y -> 0 < x % c1 + y
return mutate(0 < add_b->a % div_a_a->b + add_b->b);
} else if (mul_a &&
div_a_a &&
sub_b &&
const_int(div_a_a->b, &ia) &&
const_int(mul_a->b, &ib) &&
ia > 0 &&
ia == ib &&
equal(div_a_a->a, sub_b->a)) {
// subtract (x/c1)*c1 from both sides
// (x/c1)*c1 < x - y -> y < x % c1
return mutate(sub_b->b < sub_b->a % div_a_a->b);
} else if (mul_a &&
div_a_a &&
add_a_a_a &&
const_int(div_a_a->b, &ia) &&
const_int(mul_a->b, &ib) &&
const_int(add_a_a_a->b, &ic) &&
ia > 0 &&
ia == ib &&
equal(add_a_a_a->a, b)) {
// subtract ((x+c2)/c1)*c1 from both sides
// ((x+c2)/c1)*c1 < x -> c2 < (x+c2) % c1
return mutate(add_a_a_a->b < div_a_a->a % div_a_a->b);
} else if (mul_a &&
div_a_a &&
add_b &&
add_a_a_a &&
const_int(div_a_a->b, &ia) &&
const_int(mul_a->b, &ib) &&
const_int(add_a_a_a->b, &ic) &&
ia > 0 &&
ia == ib &&
equal(add_a_a_a->a, add_b->a)) {
// subtract ((x+c2)/c1)*c1 from both sides
// ((x+c2)/c1)*c1 < x + y -> c2 < (x+c2) % c1 + y
return mutate(add_a_a_a->b < div_a_a->a % div_a_a->b + add_b->b);
} else if (mul_a &&
div_a_a &&
add_a_a_a &&
sub_b &&
const_int(div_a_a->b, &ia) &&
const_int(mul_a->b, &ib) &&
const_int(add_a_a_a->b, &ic) &&
ia > 0 &&
ia == ib &&
equal(add_a_a_a->a, sub_b->a)) {
// subtract ((x+c2)/c1)*c1 from both sides
// ((x+c2)/c1)*c1 < x - y -> y < (x+c2) % c1 + (-c2)
return mutate(sub_b->b < div_a_a->a % div_a_a->b + make_const(a.type(), -ic));
} else if (delta_ramp &&
is_positive_const(delta_ramp->stride) &&
is_one(mutate(delta_ramp->base + delta_ramp->stride*(delta_ramp->lanes - 1) < 0))) {
return const_true(delta_ramp->lanes);
} else if (delta_ramp &&
is_positive_const(delta_ramp->stride) &&
is_one(mutate(delta_ramp->base >= 0))) {
return const_false(delta_ramp->lanes);
} else if (delta_ramp &&
is_negative_const(delta_ramp->stride) &&
is_one(mutate(delta_ramp->base < 0))) {
return const_true(delta_ramp->lanes);
} else if (delta_ramp &&
is_negative_const(delta_ramp->stride) &&
is_one(mutate(delta_ramp->base + delta_ramp->stride*(delta_ramp->lanes - 1) >= 0))) {
return const_false(delta_ramp->lanes);
} else if (delta_ramp && mod_rem.modulus > 0 &&
const_int(delta_ramp->stride, &ia) &&
0 <= ia * (delta_ramp->lanes - 1) + mod_rem.remainder &&
ia * (delta_ramp->lanes - 1) + mod_rem.remainder < mod_rem.modulus) {
// ramp(x, a, b) < 0 -> broadcast(x < 0, b)
return Broadcast::make(mutate(LT::make(delta_ramp->base / mod_rem.modulus, 0)), delta_ramp->lanes);
} else if (a.same_as(op->a) && b.same_as(op->b)) {
return op;
} else {
return LT::make(a, b);
}
} else if (a.same_as(op->a) && b.same_as(op->b)) {
return op;
} else {
return LT::make(a, b);
}
}
Expr visit(const LE *op) override {
return mutate(!(op->b < op->a));
}
Expr visit(const GT *op) override {
return mutate(op->b < op->a);
}
Expr visit(const GE *op) override {
return mutate(!(op->a < op->b));
}
Expr visit(const And *op) override {
Expr a = mutate(op->a);
Expr b = mutate(op->b);
Expr expr;
if (propagate_indeterminate_expression(a, b, op->type, &expr)) {
return expr;
}
const Broadcast *broadcast_a = a.as<Broadcast>();
const Broadcast *broadcast_b = b.as<Broadcast>();
const LE *le_a = a.as<LE>();
const LE *le_b = b.as<LE>();
const LT *lt_a = a.as<LT>();
const LT *lt_b = b.as<LT>();
const EQ *eq_a = a.as<EQ>();
const EQ *eq_b = b.as<EQ>();
const NE *neq_a = a.as<NE>();
const NE *neq_b = b.as<NE>();
const Not *not_a = a.as<Not>();
const Not *not_b = b.as<Not>();
const Variable *var_a = a.as<Variable>();
const Variable *var_b = b.as<Variable>();
int64_t ia = 0, ib = 0;
if (is_one(a)) {
return b;
} else if (is_one(b)) {
return a;
} else if (is_zero(a)) {
return a;
} else if (is_zero(b)) {
return b;
} else if (equal(a, b)) {
// a && a -> a
return a;
} else if (le_a &&
le_b &&
equal(le_a->a, le_b->a)) {
// (x <= foo && x <= bar) -> x <= min(foo, bar)
return mutate(le_a->a <= min(le_a->b, le_b->b));
} else if (le_a &&
le_b &&
equal(le_a->b, le_b->b)) {
// (foo <= x && bar <= x) -> max(foo, bar) <= x
return mutate(max(le_a->a, le_b->a) <= le_a->b);
} else if (lt_a &&
lt_b &&
equal(lt_a->a, lt_b->a)) {
// (x < foo && x < bar) -> x < min(foo, bar)
return mutate(lt_a->a < min(lt_a->b, lt_b->b));
} else if (lt_a &&
lt_b &&
equal(lt_a->b, lt_b->b)) {
// (foo < x && bar < x) -> max(foo, bar) < x
return mutate(max(lt_a->a, lt_b->a) < lt_a->b);
} else if (eq_a &&
neq_b &&
((equal(eq_a->a, neq_b->a) && equal(eq_a->b, neq_b->b)) ||
(equal(eq_a->a, neq_b->b) && equal(eq_a->b, neq_b->a)))) {
// a == b && a != b
return const_false(op->type.lanes());
} else if (eq_b &&
neq_a &&
((equal(eq_b->a, neq_a->a) && equal(eq_b->b, neq_a->b)) ||
(equal(eq_b->a, neq_a->b) && equal(eq_b->b, neq_a->a)))) {
// a != b && a == b
return const_false(op->type.lanes());
} else if ((not_a && equal(not_a->a, b)) ||
(not_b && equal(not_b->a, a))) {
// a && !a
return const_false(op->type.lanes());
} else if (le_a &&
lt_b &&
equal(le_a->a, lt_b->b) &&
equal(le_a->b, lt_b->a)) {
// a <= b && b < a
return const_false(op->type.lanes());
} else if (lt_a &&
le_b &&
equal(lt_a->a, le_b->b) &&
equal(lt_a->b, le_b->a)) {
// a < b && b <= a
return const_false(op->type.lanes());
} else if (lt_a &&
lt_b &&
equal(lt_a->a, lt_b->b) &&
const_int(lt_a->b, &ia) &&
const_int(lt_b->a, &ib) &&
ib + 1 >= ia) {
// (a < ia && ib < a) where there is no integer a s.t. ib < a < ia
return const_false(op->type.lanes());
} else if (lt_a &&
lt_b &&
equal(lt_a->b, lt_b->a) &&
const_int(lt_b->b, &ia) &&
const_int(lt_a->a, &ib) &&
ib + 1 >= ia) {
// (ib < a && a < ia) where there is no integer a s.t. ib < a < ia
return const_false(op->type.lanes());
} else if (le_a &&
lt_b &&
equal(le_a->a, lt_b->b) &&
const_int(le_a->b, &ia) &&
const_int(lt_b->a, &ib) &&
ib >= ia) {
// (a <= ia && ib < a) where there is no integer a s.t. ib < a <= ia
return const_false(op->type.lanes());
} else if (le_a &&
lt_b &&
equal(le_a->b, lt_b->a) &&
const_int(lt_b->b, &ia) &&
const_int(le_a->a, &ib) &&
ib >= ia) {
// (ib <= a && a < ia) where there is no integer a s.t. ib < a <= ia
return const_false(op->type.lanes());
} else if (lt_a &&
le_b &&
equal(lt_a->a, le_b->b) &&
const_int(lt_a->b, &ia) &&
const_int(le_b->a, &ib) &&
ib >= ia) {
// (a < ia && ib <= a) where there is no integer a s.t. ib <= a < ia
return const_false(op->type.lanes());
} else if (lt_a &&
le_b &&
equal(lt_a->b, le_b->a) &&
const_int(le_b->b, &ia) &&
const_int(lt_a->a, &ib) &&
ib >= ia) {
// (ib < a && a <= ia) where there is no integer a s.t. ib <= a < ia
return const_false(op->type.lanes());
} else if (le_a &&
le_b &&
equal(le_a->a, le_b->b) &&
const_int(le_a->b, &ia) &&
const_int(le_b->a, &ib) &&
ib > ia) {
// (a <= ia && ib <= a) where there is no integer a s.t. ib <= a <= ia
return const_false(op->type.lanes());
} else if (le_a &&
le_b &&
equal(le_a->b, le_b->a) &&
const_int(le_b->b, &ia) &&
const_int(le_a->a, &ib) &&
ib > ia) {
// (ib <= a && a <= ia) where there is no integer a s.t. ib <= a <= ia
return const_false(op->type.lanes());
} else if (eq_a &&
neq_b &&
equal(eq_a->a, neq_b->a) &&
is_simple_const(eq_a->b) &&
is_simple_const(neq_b->b)) {
// (a == k1) && (a != k2) -> (a == k1) && (k1 != k2)
// (second term always folds away)
return mutate(And::make(a, NE::make(eq_a->b, neq_b->b)));
} else if (neq_a &&
eq_b &&
equal(neq_a->a, eq_b->a) &&
is_simple_const(neq_a->b) &&
is_simple_const(eq_b->b)) {
// (a != k1) && (a == k2) -> (a == k2) && (k1 != k2)
// (second term always folds away)
return mutate(And::make(b, NE::make(neq_a->b, eq_b->b)));
} else if (eq_a &&
eq_a->a.as<Variable>() &&
is_simple_const(eq_a->b) &&
expr_uses_var(b, eq_a->a.as<Variable>()->name)) {
// (somevar == k) && b -> (somevar == k) && substitute(somevar, k, b)
return mutate(And::make(a, substitute(eq_a->a.as<Variable>(), eq_a->b, b)));
} else if (eq_b &&
eq_b->a.as<Variable>() &&
is_simple_const(eq_b->b) &&
expr_uses_var(a, eq_b->a.as<Variable>()->name)) {
// a && (somevar == k) -> substitute(somevar, k1, a) && (somevar == k)
return mutate(And::make(substitute(eq_b->a.as<Variable>(), eq_b->b, a), b));
} else if (broadcast_a &&
broadcast_b &&
broadcast_a->lanes == broadcast_b->lanes) {
// x8(a) && x8(b) -> x8(a && b)
return Broadcast::make(mutate(And::make(broadcast_a->value, broadcast_b->value)), broadcast_a->lanes);
} else if (var_a && expr_uses_var(b, var_a->name)) {
return mutate(a && substitute(var_a->name, make_one(a.type()), b));
} else if (var_b && expr_uses_var(a, var_b->name)) {
return mutate(substitute(var_b->name, make_one(b.type()), a) && b);
} else if (a.same_as(op->a) &&
b.same_as(op->b)) {
return op;
} else {
return And::make(a, b);
}
}
Expr visit(const Or *op) override {
Expr a = mutate(op->a), b = mutate(op->b);
Expr expr;
if (propagate_indeterminate_expression(a, b, op->type, &expr)) {
return expr;
}
const Broadcast *broadcast_a = a.as<Broadcast>();
const Broadcast *broadcast_b = b.as<Broadcast>();
const EQ *eq_a = a.as<EQ>();
const EQ *eq_b = b.as<EQ>();
const NE *neq_a = a.as<NE>();
const NE *neq_b = b.as<NE>();
const Not *not_a = a.as<Not>();
const Not *not_b = b.as<Not>();
const LE *le_a = a.as<LE>();
const LE *le_b = b.as<LE>();
const LT *lt_a = a.as<LT>();
const LT *lt_b = b.as<LT>();
const Variable *var_a = a.as<Variable>();
const Variable *var_b = b.as<Variable>();
const And *and_a = a.as<And>();
const And *and_b = b.as<And>();
string name_a, name_b, name_c;
int64_t ia = 0, ib = 0;
if (is_one(a)) {
return a;
} else if (is_one(b)) {
return b;
} else if (is_zero(a)) {
return b;
} else if (is_zero(b)) {
return a;
} else if (equal(a, b)) {
return a;
} else if (eq_a &&
neq_b &&
((equal(eq_a->a, neq_b->a) && equal(eq_a->b, neq_b->b)) ||
(equal(eq_a->a, neq_b->b) && equal(eq_a->b, neq_b->a)))) {
// a == b || a != b
return const_true(op->type.lanes());
} else if (neq_a &&
eq_b &&
((equal(eq_b->a, neq_a->a) && equal(eq_b->b, neq_a->b)) ||
(equal(eq_b->a, neq_a->b) && equal(eq_b->b, neq_a->a)))) {
// a != b || a == b
return const_true(op->type.lanes());
} else if ((not_a && equal(not_a->a, b)) ||
(not_b && equal(not_b->a, a))) {
// a || !a
return const_true(op->type.lanes());
} else if (le_a &&
lt_b &&
equal(le_a->a, lt_b->b) &&
equal(le_a->b, lt_b->a)) {
// a <= b || b < a
return const_true(op->type.lanes());
} else if (lt_a &&
le_b &&
equal(lt_a->a, le_b->b) &&
equal(lt_a->b, le_b->a)) {
// a < b || b <= a
return const_true(op->type.lanes());
} else if (lt_a &&
lt_b &&
equal(lt_a->a, lt_b->b) &&
const_int(lt_a->b, &ia) &&
const_int(lt_b->a, &ib) &&
ib < ia) {
// (a < ia || ib < a) where ib < ia
return const_true(op->type.lanes());
} else if (lt_a &&
lt_b &&
equal(lt_a->b, lt_b->a) &&
const_int(lt_b->b, &ia) &&
const_int(lt_a->a, &ib) &&
ib < ia) {
// (ib < a || a < ia) where ib < ia
return const_true(op->type.lanes());
} else if (le_a &&
lt_b &&
equal(le_a->a, lt_b->b) &&
const_int(le_a->b, &ia) &&
const_int(lt_b->a, &ib) &&
ib <= ia) {
// (a <= ia || ib < a) where ib <= ia
return const_true(op->type.lanes());
} else if (le_a &&
lt_b &&
equal(le_a->b, lt_b->a) &&
const_int(lt_b->b, &ia) &&
const_int(le_a->a, &ib) &&
ib <= ia) {
// (ib <= a || a < ia) where ib <= ia
return const_true(op->type.lanes());
} else if (lt_a &&
le_b &&
equal(lt_a->a, le_b->b) &&
const_int(lt_a->b, &ia) &&
const_int(le_b->a, &ib) &&
ib <= ia) {
// (a < ia || ib <= a) where ib <= ia
return const_true(op->type.lanes());
} else if (lt_a &&
le_b &&
equal(lt_a->b, le_b->a) &&
const_int(le_b->b, &ia) &&
const_int(lt_a->a, &ib) &&
ib <= ia) {
// (ib < a || a <= ia) where ib <= ia
return const_true(op->type.lanes());
} else if (le_a &&
le_b &&
equal(le_a->a, le_b->b) &&
const_int(le_a->b, &ia) &&
const_int(le_b->a, &ib) &&
ib <= ia + 1) {
// (a <= ia || ib <= a) where ib <= ia + 1
return const_true(op->type.lanes());
} else if (le_a &&
le_b &&
equal(le_a->b, le_b->a) &&
const_int(le_b->b, &ia) &&
const_int(le_a->a, &ib) &&
ib <= ia + 1) {
// (ib <= a || a <= ia) where ib <= ia + 1
return const_true(op->type.lanes());
} else if (broadcast_a &&
broadcast_b &&
broadcast_a->lanes == broadcast_b->lanes) {
// x8(a) || x8(b) -> x8(a || b)
return Broadcast::make(mutate(Or::make(broadcast_a->value, broadcast_b->value)), broadcast_a->lanes);
} else if (eq_a &&
neq_b &&
equal(eq_a->a, neq_b->a) &&
is_simple_const(eq_a->b) &&
is_simple_const(neq_b->b)) {
// (a == k1) || (a != k2) -> (a != k2) || (k1 == k2)
// (second term always folds away)
return mutate(Or::make(b, EQ::make(eq_a->b, neq_b->b)));
} else if (neq_a &&
eq_b &&
equal(neq_a->a, eq_b->a) &&
is_simple_const(neq_a->b) &&
is_simple_const(eq_b->b)) {
// (a != k1) || (a == k2) -> (a != k1) || (k1 == k2)
// (second term always folds away)
return mutate(Or::make(a, EQ::make(neq_a->b, eq_b->b)));
} else if (var_a && expr_uses_var(b, var_a->name)) {
return mutate(a || substitute(var_a->name, make_zero(a.type()), b));
} else if (var_b && expr_uses_var(a, var_b->name)) {
return mutate(substitute(var_b->name, make_zero(b.type()), a) || b);
} else if (is_var_simple_const_comparison(b, &name_c) &&
and_a &&
((is_var_simple_const_comparison(and_a->a, &name_a) && name_a == name_c) ||
(is_var_simple_const_comparison(and_a->b, &name_b) && name_b == name_c))) {
// (a && b) || (c) -> (a || c) && (b || c)
// iff c and at least one of a or b is of the form
// (var == const) or (var != const)
// (and the vars are the same)
return mutate(And::make(Or::make(and_a->a, b), Or::make(and_a->b, b)));
} else if (is_var_simple_const_comparison(a, &name_c) &&
and_b &&
((is_var_simple_const_comparison(and_b->a, &name_a) && name_a == name_c) ||
(is_var_simple_const_comparison(and_b->b, &name_b) && name_b == name_c))) {
// (c) || (a && b) -> (a || c) && (b || c)
// iff c and at least one of a or b is of the form
// (var == const) or (var != const)
// (and the vars are the same)
return mutate(And::make(Or::make(and_b->a, a), Or::make(and_b->b, a)));
} else if (a.same_as(op->a) && b.same_as(op->b)) {
return op;
} else {
return Or::make(a, b);
}
}
Expr visit(const Not *op) override {
Expr a = mutate(op->a);
Expr expr;
if (propagate_indeterminate_expression(a, op->type, &expr)) {
return expr;
}
const Call *c;
if (is_one(a)) {
return make_zero(a.type());
} else if (is_zero(a)) {
return make_one(a.type());
} else if (const Not *n = a.as<Not>()) {
// Double negatives cancel
return n->a;
} else if (const LE *n = a.as<LE>()) {
return LT::make(n->b, n->a);
} else if (const GE *n = a.as<GE>()) {
return LT::make(n->a, n->b);
} else if (const LT *n = a.as<LT>()) {
return LE::make(n->b, n->a);
} else if (const GT *n = a.as<GT>()) {
return LE::make(n->a, n->b);
} else if (const NE *n = a.as<NE>()) {
return EQ::make(n->a, n->b);
} else if (const EQ *n = a.as<EQ>()) {
return NE::make(n->a, n->b);
} else if (const Broadcast *n = a.as<Broadcast>()) {
return mutate(Broadcast::make(!n->value, n->lanes));
} else if ((c = a.as<Call>()) != nullptr && c->is_intrinsic(Call::likely)) {
// !likely(e) -> likely(!e)
return likely(mutate(Not::make(c->args[0])));
} else if (a.same_as(op->a)) {
return op;
} else {
return Not::make(a);
}
}
Expr visit(const Select *op) override {
Expr condition = mutate(op->condition);
Expr true_value = mutate(op->true_value);
Expr false_value = mutate(op->false_value);
Expr expr;
if (propagate_indeterminate_expression(condition, true_value, false_value, op->type, &expr)) {
return expr;
}
const Call *ct = true_value.as<Call>();
const Call *cf = false_value.as<Call>();
const Select *sel_t = true_value.as<Select>();
const Select *sel_f = false_value.as<Select>();
const Add *add_t = true_value.as<Add>();
const Add *add_f = false_value.as<Add>();
const Sub *sub_t = true_value.as<Sub>();
const Sub *sub_f = false_value.as<Sub>();
const Mul *mul_t = true_value.as<Mul>();
const Mul *mul_f = false_value.as<Mul>();
if (is_zero(condition)) {
return false_value;
} else if (is_one(condition)) {
return true_value;
} else if (equal(true_value, false_value)) {
return true_value;
} else if (true_value.type().is_bool() &&
is_one(true_value) &&
is_zero(false_value)) {
if (true_value.type().is_vector() && condition.type().is_scalar()) {
return Broadcast::make(condition, true_value.type().lanes());
} else {
return condition;
}
} else if (true_value.type().is_bool() &&
is_zero(true_value) &&
is_one(false_value)) {
if (true_value.type().is_vector() && condition.type().is_scalar()) {
return Broadcast::make(mutate(!condition), true_value.type().lanes());
} else {
return mutate(!condition);
}
} else if (const Broadcast *b = condition.as<Broadcast>()) {
// Select of broadcast -> scalar select
return mutate(Select::make(b->value, true_value, false_value));
} else if (const NE *ne = condition.as<NE>()) {
// Normalize select(a != b, c, d) to select(a == b, d, c)
return mutate(Select::make(ne->a == ne->b, false_value, true_value));
} else if (const LE *le = condition.as<LE>()) {
// Normalize select(a <= b, c, d) to select(b < a, d, c)
return mutate(Select::make(le->b < le->a, false_value, true_value));
} else if (ct && ct->is_intrinsic(Call::likely) &&
equal(ct->args[0], false_value)) {
// select(cond, likely(a), a) -> likely(a)
return true_value;
} else if (cf &&
cf->is_intrinsic(Call::likely) &&
equal(cf->args[0], true_value)) {
// select(cond, a, likely(a)) -> likely(a)
return false_value;
} else if (sel_t &&
equal(sel_t->true_value, false_value)) {
// select(a, select(b, c, d), c) -> select(a && !b, d, c)
return mutate(Select::make(condition && !sel_t->condition, sel_t->false_value, false_value));
} else if (sel_t &&
equal(sel_t->false_value, false_value)) {
// select(a, select(b, c, d), d) -> select(a && b, c, d)
return mutate(Select::make(condition && sel_t->condition, sel_t->true_value, false_value));
} else if (sel_f &&
equal(sel_f->false_value, true_value)) {
// select(a, d, select(b, c, d)) -> select(a || !b, d, c)
return mutate(Select::make(condition || !sel_f->condition, true_value, sel_f->true_value));
} else if (sel_f &&
equal(sel_f->true_value, true_value)) {
// select(a, d, select(b, d, c)) -> select(a || b, d, c)
return mutate(Select::make(condition || sel_f->condition, true_value, sel_f->false_value));
} else if (sel_t &&
equal(sel_t->condition, condition)) {
// select(a, select(a, b, c), d) -> select(a, b, d)
return mutate(Select::make(condition, sel_t->true_value, false_value));
} else if (sel_f &&
equal(sel_f->condition, condition)) {
// select(a, b, select(a, c, d)) -> select(a, b, d)
return mutate(Select::make(condition, true_value, sel_f->false_value));
} else if (add_t &&
add_f &&
equal(add_t->a, add_f->a)) {
// select(c, a+b, a+d) -> a + select(x, b, d)
return mutate(add_t->a + Select::make(condition, add_t->b, add_f->b));
} else if (add_t &&
add_f &&
equal(add_t->a, add_f->b)) {
// select(c, a+b, d+a) -> a + select(x, b, d)
return mutate(add_t->a + Select::make(condition, add_t->b, add_f->a));
} else if (add_t &&
add_f &&
equal(add_t->b, add_f->a)) {
// select(c, b+a, a+d) -> a + select(x, b, d)
return mutate(add_t->b + Select::make(condition, add_t->a, add_f->b));
} else if (add_t &&
add_f &&
equal(add_t->b, add_f->b)) {
// select(c, b+a, d+a) -> select(x, b, d) + a
return mutate(Select::make(condition, add_t->a, add_f->a) + add_t->b);
} else if (sub_t &&
sub_f &&
equal(sub_t->a, sub_f->a)) {
// select(c, a-b, a-d) -> a - select(x, b, d)
return mutate(sub_t->a - Select::make(condition, sub_t->b, sub_f->b));
} else if (sub_t &&
sub_f &&
equal(sub_t->b, sub_f->b)) {
// select(c, b-a, d-a) -> select(x, b, d) - a
return mutate(Select::make(condition, sub_t->a, sub_f->a) - sub_t->b);\
} else if (add_t &&
sub_f &&
equal(add_t->a, sub_f->a)) {
// select(c, a+b, a-d) -> a + select(x, b, 0-d)
return mutate(add_t->a + Select::make(condition, add_t->b, make_zero(sub_f->b.type()) - sub_f->b));
} else if (add_t &&
sub_f &&
equal(add_t->b, sub_f->a)) {
// select(c, b+a, a-d) -> a + select(x, b, 0-d)
return mutate(add_t->b + Select::make(condition, add_t->a, make_zero(sub_f->b.type()) - sub_f->b));
} else if (sub_t &&
add_f &&
equal(sub_t->a, add_f->a)) {
// select(c, a-b, a+d) -> a + select(x, 0-b, d)
return mutate(sub_t->a + Select::make(condition, make_zero(sub_t->b.type()) - sub_t->b, add_f->b));
} else if (sub_t &&
add_f &&
equal(sub_t->a, add_f->b)) {
// select(c, a-b, d+a) -> a + select(x, 0-b, d)
return mutate(sub_t->a + Select::make(condition, make_zero(sub_t->b.type()) - sub_t->b, add_f->a));
} else if (mul_t &&
mul_f &&
equal(mul_t->a, mul_f->a)) {
// select(c, a*b, a*d) -> a * select(x, b, d)
return mutate(mul_t->a * Select::make(condition, mul_t->b, mul_f->b));
} else if (mul_t &&
mul_f &&
equal(mul_t->a, mul_f->b)) {
// select(c, a*b, d*a) -> a * select(x, b, d)
return mutate(mul_t->a * Select::make(condition, mul_t->b, mul_f->a));
} else if (mul_t &&
mul_f &&
equal(mul_t->b, mul_f->a)) {
// select(c, b*a, a*d) -> a * select(x, b, d)
return mutate(mul_t->b * Select::make(condition, mul_t->a, mul_f->b));
} else if (mul_t &&
mul_f &&
equal(mul_t->b, mul_f->b)) {
// select(c, b*a, d*a) -> select(x, b, d) * a
return mutate(Select::make(condition, mul_t->a, mul_f->a) * mul_t->b);
} else if (condition.same_as(op->condition) &&
true_value.same_as(op->true_value) &&
false_value.same_as(op->false_value)) {
return op;
} else {
return Select::make(condition, true_value, false_value);
}
}
Expr visit(const Ramp *op) override {
Expr base = mutate(op->base);
Expr stride = mutate(op->stride);
if (is_zero(stride)) {
return Broadcast::make(base, op->lanes);
} else if (base.same_as(op->base) &&
stride.same_as(op->stride)) {
return op;
} else {
return Ramp::make(base, stride, op->lanes);
}
}
Stmt visit(const IfThenElse *op) override {
Expr condition = mutate(op->condition);
// If (true) ...
if (is_one(condition)) {
return mutate(op->then_case);
}
// If (false) ...
if (is_zero(condition)) {
Stmt stmt = mutate(op->else_case);
if (!stmt.defined()) {
// Emit a noop
stmt = Evaluate::make(0);
}
return stmt;
}
Stmt then_case = mutate(op->then_case);
Stmt else_case = mutate(op->else_case);
// If both sides are no-ops, bail out.
if (is_no_op(then_case) && is_no_op(else_case)) {
return then_case;
}
// Remember the statements before substitution.
Stmt then_nosubs = then_case;
Stmt else_nosubs = else_case;
// Mine the condition for useful constraints to apply (eg var == value && bool_param).
vector<Expr> stack;
stack.push_back(condition);
bool and_chain = false, or_chain = false;
while (!stack.empty()) {
Expr next = stack.back();
stack.pop_back();
if (!or_chain) {
then_case = substitute(next, const_true(), then_case);
}
if (!and_chain) {
else_case = substitute(next, const_false(), else_case);
}
if (const And *a = next.as<And>()) {
if (!or_chain) {
stack.push_back(a->b);
stack.push_back(a->a);
and_chain = true;
}
} else if (const Or *o = next.as<Or>()) {
if (!and_chain) {
stack.push_back(o->b);
stack.push_back(o->a);
or_chain = true;
}
} else {
const EQ *eq = next.as<EQ>();
const NE *ne = next.as<NE>();
const Variable *var = eq ? eq->a.as<Variable>() : next.as<Variable>();
if (eq && var) {
if (!or_chain) {
then_case = substitute(var->name, eq->b, then_case);
}
if (!and_chain && eq->b.type().is_bool()) {
else_case = substitute(var->name, !eq->b, else_case);
}
} else if (var) {
if (!or_chain) {
then_case = substitute(var->name, const_true(), then_case);
}
if (!and_chain) {
else_case = substitute(var->name, const_false(), else_case);
}
} else if (eq && is_const(eq->b) && !or_chain) {
// some_expr = const
then_case = substitute(eq->a, eq->b, then_case);
} else if (ne && is_const(ne->b) && !and_chain) {
// some_expr != const
else_case = substitute(ne->a, ne->b, else_case);
}
}
}
// If substitutions have been made, simplify again.
if (!then_case.same_as(then_nosubs)) {
then_case = mutate(then_case);
}
if (!else_case.same_as(else_nosubs)) {
else_case = mutate(else_case);
}
if (condition.same_as(op->condition) &&
then_case.same_as(op->then_case) &&
else_case.same_as(op->else_case)) {
return op;
} else {
return IfThenElse::make(condition, then_case, else_case);
}
}
Expr visit(const Load *op) override {
found_buffer_reference(op->name);
Expr predicate = mutate(op->predicate);
Expr index = mutate(op->index);
const Broadcast *b_index = index.as<Broadcast>();
const Broadcast *b_pred = predicate.as<Broadcast>();
if (is_zero(predicate)) {
// Predicate is always false
return undef(op->type);
} else if (b_index && b_pred) {
// Load of a broadcast should be broadcast of the load
Expr load = Load::make(op->type.element_of(), op->name, b_index->value, op->image, op->param, b_pred->value);
return Broadcast::make(load, b_index->lanes);
} else if (predicate.same_as(op->predicate) && index.same_as(op->index)) {
return op;
} else {
return Load::make(op->type, op->name, index, op->image, op->param, predicate);
}
}
Expr visit(const Call *op) override {
// Calls implicitly depend on host, dev, mins, and strides of the buffer referenced
if (op->call_type == Call::Image || op->call_type == Call::Halide) {
found_buffer_reference(op->name, op->args.size());
}
if (op->is_intrinsic(Call::shift_left) ||
op->is_intrinsic(Call::shift_right)) {
Expr a = mutate(op->args[0]), b = mutate(op->args[1]);
Expr expr;
if (propagate_indeterminate_expression(a, b, op->type, &expr)) {
return expr;
}
int64_t ib = 0;
if (const_int(b, &ib) || const_uint(b, (uint64_t *)(&ib))) {
Type t = op->type;
bool shift_left = op->is_intrinsic(Call::shift_left);
if (t.is_int() && ib < 0) {
shift_left = !shift_left;
ib = -ib;
}
if (ib >= 0 && ib < std::min(t.bits(), 64) - 1) {
ib = 1LL << ib;
b = make_const(t, ib);
if (shift_left) {
return mutate(Mul::make(a, b));
} else {
return mutate(Div::make(a, b));
}
}
}
if (a.same_as(op->args[0]) && b.same_as(op->args[1])) {
return op;
} else if (op->is_intrinsic(Call::shift_left)) {
return a << b;
} else {
return a >> b;
}
} else if (op->is_intrinsic(Call::bitwise_and)) {
Expr a = mutate(op->args[0]), b = mutate(op->args[1]);
Expr expr;
if (propagate_indeterminate_expression(a, b, op->type, &expr)) {
return expr;
}
int64_t ia, ib = 0;
uint64_t ua, ub = 0;
int bits;
if (const_int(a, &ia) &&
const_int(b, &ib)) {
return make_const(op->type, ia & ib);
} else if (const_uint(a, &ua) &&
const_uint(b, &ub)) {
return make_const(op->type, ua & ub) ;
} else if (const_int(b, &ib) &&
!b.type().is_max(ib) &&
is_const_power_of_two_integer(make_const(a.type(), ib + 1), &bits)) {
return Mod::make(a, make_const(a.type(), ib + 1));
} else if (const_uint(b, &ub) &&
b.type().is_max(ub)) {
return a;
} else if (const_uint(b, &ub) &&
is_const_power_of_two_integer(make_const(a.type(), ub + 1), &bits)) {
return Mod::make(a, make_const(a.type(), ub + 1));
} else if (a.same_as(op->args[0]) && b.same_as(op->args[1])) {
return op;
} else {
return a & b;
}
} else if (op->is_intrinsic(Call::bitwise_or)) {
Expr a = mutate(op->args[0]), b = mutate(op->args[1]);
Expr expr;
if (propagate_indeterminate_expression(a, b, op->type, &expr)) {
return expr;
}
int64_t ia, ib;
uint64_t ua, ub;
if (const_int(a, &ia) &&
const_int(b, &ib)) {
return make_const(op->type, ia | ib);
} else if (const_uint(a, &ua) &&
const_uint(b, &ub)) {
return make_const(op->type, ua | ub);
} else if (a.same_as(op->args[0]) && b.same_as(op->args[1])) {
return op;
} else {
return a | b;
}
} else if (op->is_intrinsic(Call::bitwise_not)) {
Expr a = mutate(op->args[0]);
Expr expr;
if (propagate_indeterminate_expression(a, op->type, &expr)) {
return expr;
}
int64_t ia;
uint64_t ua;
if (const_int(a, &ia)) {
return make_const(op->type, ~ia);
} else if (const_uint(a, &ua)) {
return make_const(op->type, ~ua);
} else if (a.same_as(op->args[0])) {
return op;
} else {
return ~a;
}
} else if (op->is_intrinsic(Call::reinterpret)) {
Expr a = mutate(op->args[0]);
Expr expr;
if (propagate_indeterminate_expression(a, op->type, &expr)) {
return expr;
}
int64_t ia;
uint64_t ua;
bool vector = op->type.is_vector() || a.type().is_vector();
if (op->type == a.type()) {
return a;
} else if (const_int(a, &ia) && op->type.is_uint() && !vector) {
// int -> uint
return make_const(op->type, (uint64_t)ia);
} else if (const_uint(a, &ua) && op->type.is_int() && !vector) {
// uint -> int
return make_const(op->type, (int64_t)ua);
} else if (a.same_as(op->args[0])) {
return op;
} else {
return reinterpret(op->type, a);
}
} else if (op->is_intrinsic(Call::abs)) {
// Constant evaluate abs(x).
Expr a = mutate(op->args[0]);
Expr expr;
if (propagate_indeterminate_expression(a, op->type, &expr)) {
return expr;
}
Type ta = a.type();
int64_t ia = 0;
double fa = 0;
if (ta.is_int() && const_int(a, &ia)) {
if (ia < 0 && !(Int(64).is_min(ia))) {
ia = -ia;
}
return make_const(op->type, ia);
} else if (ta.is_uint()) {
// abs(uint) is a no-op.
return a;
} else if (const_float(a, &fa)) {
if (fa < 0) {
fa = -fa;
}
return make_const(a.type(), fa);
} else if (a.same_as(op->args[0])) {
return op;
} else {
return abs(a);
}
} else if (op->call_type == Call::PureExtern &&
op->name == "is_nan_f32") {
Expr arg = mutate(op->args[0]);
double f = 0.0;
if (const_float(arg, &f)) {
return std::isnan(f);
} else if (arg.same_as(op->args[0])) {
return op;
} else {
return Call::make(op->type, op->name, {arg}, op->call_type);
}
} else if (op->is_intrinsic(Call::stringify)) {
// Eagerly concat constant arguments to a stringify.
bool changed = false;
vector<Expr> new_args;
const StringImm *last = nullptr;
for (size_t i = 0; i < op->args.size(); i++) {
Expr arg = mutate(op->args[i]);
if (!arg.same_as(op->args[i])) {
changed = true;
}
const StringImm *string_imm = arg.as<StringImm>();
const IntImm *int_imm = arg.as<IntImm>();
const FloatImm *float_imm = arg.as<FloatImm>();
// We use snprintf here rather than stringstreams,
// because the runtime's float printing is guaranteed
// to match snprintf.
char buf[64]; // Large enough to hold the biggest float literal.
if (last && string_imm) {
new_args.back() = last->value + string_imm->value;
changed = true;
} else if (int_imm) {
snprintf(buf, sizeof(buf), "%lld", (long long)int_imm->value);
if (last) {
new_args.back() = last->value + buf;
} else {
new_args.push_back(string(buf));
}
changed = true;
} else if (last && float_imm) {
snprintf(buf, sizeof(buf), "%f", float_imm->value);
if (last) {
new_args.back() = last->value + buf;
} else {
new_args.push_back(string(buf));
}
changed = true;
} else {
new_args.push_back(arg);
}
last = new_args.back().as<StringImm>();
}
if (new_args.size() == 1 && new_args[0].as<StringImm>()) {
// stringify of a string constant is just the string constant
return new_args[0];
} else if (changed) {
return Call::make(op->type, op->name, new_args, op->call_type);
} else {
return op;
}
} else if (op->call_type == Call::PureExtern &&
op->name == "sqrt_f32") {
Expr arg = mutate(op->args[0]);
Expr expr;
if (propagate_indeterminate_expression(arg, op->type, &expr)) {
return expr;
}
if (const double *f = as_const_float(arg)) {
return FloatImm::make(arg.type(), std::sqrt(*f));
} else if (!arg.same_as(op->args[0])) {
return Call::make(op->type, op->name, {arg}, op->call_type);
} else {
return op;
}
} else if (op->call_type == Call::PureExtern &&
op->name == "log_f32") {
Expr arg = mutate(op->args[0]);
Expr expr;
if (propagate_indeterminate_expression(arg, op->type, &expr)) {
return expr;
}
if (const double *f = as_const_float(arg)) {
return FloatImm::make(arg.type(), std::log(*f));
} else if (!arg.same_as(op->args[0])) {
return Call::make(op->type, op->name, {arg}, op->call_type);
} else {
return op;
}
} else if (op->call_type == Call::PureExtern &&
op->name == "exp_f32") {
Expr arg = mutate(op->args[0]);
Expr expr;
if (propagate_indeterminate_expression(arg, op->type, &expr)) {
return expr;
}
if (const double *f = as_const_float(arg)) {
return FloatImm::make(arg.type(), std::exp(*f));
} else if (!arg.same_as(op->args[0])) {
return Call::make(op->type, op->name, {arg}, op->call_type);
} else {
return op;
}
} else if (op->call_type == Call::PureExtern &&
op->name == "pow_f32") {
Expr arg0 = mutate(op->args[0]);
Expr arg1 = mutate(op->args[1]);
Expr expr;
if (propagate_indeterminate_expression(arg0, arg1, op->type, &expr)) {
return expr;
}
const double *f0 = as_const_float(arg0);
const double *f1 = as_const_float(arg1);
if (f0 && f1) {
return FloatImm::make(arg0.type(), std::pow(*f0, *f1));
} else if (!arg0.same_as(op->args[0]) || !arg1.same_as(op->args[1])) {
return Call::make(op->type, op->name, {arg0, arg1}, op->call_type);
} else {
return op;
}
} else if (op->call_type == Call::PureExtern &&
(op->name == "floor_f32" || op->name == "ceil_f32" ||
op->name == "round_f32" || op->name == "trunc_f32")) {
internal_assert(op->args.size() == 1);
Expr arg = mutate(op->args[0]);
Expr expr;
if (propagate_indeterminate_expression(arg, op->type, &expr)) {
return expr;
}
const Call *call = arg.as<Call>();
if (const double *f = as_const_float(arg)) {
if (op->name == "floor_f32") {
return FloatImm::make(arg.type(), std::floor(*f));
} else if (op->name == "ceil_f32") {
return FloatImm::make(arg.type(), std::ceil(*f));
} else if (op->name == "round_f32") {
return FloatImm::make(arg.type(), std::nearbyint(*f));
} else if (op->name == "trunc_f32") {
return FloatImm::make(arg.type(), (*f < 0 ? std::ceil(*f) : std::floor(*f)));
} else {
return op;
}
} else if (call && call->call_type == Call::PureExtern &&
(call->name == "floor_f32" || call->name == "ceil_f32" ||
call->name == "round_f32" || call->name == "trunc_f32")) {
// For any combination of these integer-valued functions, we can
// discard the outer function. For example, floor(ceil(x)) == ceil(x).
return call;
} else if (!arg.same_as(op->args[0])) {
return Call::make(op->type, op->name, {arg}, op->call_type);
} else {
return op;
}
} else if (op->is_intrinsic(Call::prefetch)) {
// Collapse the prefetched region into lower dimension whenever is possible.
// TODO(psuriana): Deal with negative strides and overlaps.
internal_assert(op->args.size() % 2 == 0); // Format: {base, offset, extent0, min0, ...}
vector<Expr> args(op->args);
bool changed = false;
for (size_t i = 0; i < op->args.size(); ++i) {
args[i] = mutate(op->args[i]);
if (!args[i].same_as(op->args[i])) {
changed = true;
}
}
// The {extent, stride} args in the prefetch call are sorted
// based on the storage dimension in ascending order (i.e. innermost
// first and outermost last), so, it is enough to check for the upper
// triangular pairs to see if any contiguous addresses exist.
for (size_t i = 2; i < args.size(); i += 2) {
Expr extent_0 = args[i];
Expr stride_0 = args[i + 1];
for (size_t j = i + 2; j < args.size(); j += 2) {
Expr extent_1 = args[j];
Expr stride_1 = args[j + 1];
if (can_prove(extent_0 * stride_0 == stride_1)) {
Expr new_extent = mutate(extent_0 * extent_1);
Expr new_stride = stride_0;
args.erase(args.begin() + j, args.begin() + j + 2);
args[i] = new_extent;
args[i + 1] = new_stride;
i -= 2;
break;
}
}
}
internal_assert(args.size() <= op->args.size());
if (changed || (args.size() != op->args.size())) {
return Call::make(op->type, Call::prefetch, args, Call::Intrinsic);
} else {
return op;
}
} else if (op->is_intrinsic(Call::require)) {
Expr cond = mutate(op->args[0]);
// likely(const-bool) is deliberately not reduced
// by the simplify(), but for our purposes here, we want
// to ignore the likely() wrapper. (Note that this is
// equivalent to calling can_prove() without needing to
// create a new Simplifier instance.)
if (const Call *c = cond.as<Call>()) {
if (c->is_intrinsic(Call::likely)) {
cond = c->args[0];
}
}
if (is_one(cond)) {
return mutate(op->args[1]);
} else {
if (is_zero(cond)) {
// (We could simplify this to avoid evaluating the provably-false
// expression, but since this is a degenerate condition, don't bother.)
user_warning << "This pipeline is guaranteed to fail a require() expression at runtime: \n"
<< Expr(op) << "\n";
}
return IRMutator2::visit(op);
}
} else {
return IRMutator2::visit(op);
}
}
Expr visit(const Shuffle *op) override {
if (op->is_extract_element() &&
(op->vectors[0].as<Ramp>() ||
op->vectors[0].as<Broadcast>())) {
// Extracting a single lane of a ramp or broadcast
if (const Ramp *r = op->vectors[0].as<Ramp>()) {
return mutate(r->base + op->indices[0]*r->stride);
} else if (const Broadcast *b = op->vectors[0].as<Broadcast>()) {
return mutate(b->value);
} else {
internal_error << "Unreachable";
return Expr();
}
}
// Mutate the vectors
vector<Expr> new_vectors;
bool changed = false;
for (Expr vector : op->vectors) {
Expr new_vector = mutate(vector);
if (!vector.same_as(new_vector)) {
changed = true;
}
new_vectors.push_back(new_vector);
}
// Try to convert a load with shuffled indices into a
// shuffle of a dense load.
if (const Load *first_load = new_vectors[0].as<Load>()) {
vector<Expr> load_predicates;
vector<Expr> load_indices;
bool unpredicated = true;
for (Expr e : new_vectors) {
const Load *load = e.as<Load>();
if (load && load->name == first_load->name) {
load_predicates.push_back(load->predicate);
load_indices.push_back(load->index);
unpredicated = unpredicated && is_one(load->predicate);
} else {
break;
}
}
if (load_indices.size() == new_vectors.size()) {
Type t = load_indices[0].type().with_lanes(op->indices.size());
Expr shuffled_index = Shuffle::make(load_indices, op->indices);
shuffled_index = mutate(shuffled_index);
if (shuffled_index.as<Ramp>()) {
Expr shuffled_predicate;
if (unpredicated) {
shuffled_predicate = const_true(t.lanes());
} else {
shuffled_predicate = Shuffle::make(load_predicates, op->indices);
shuffled_predicate = mutate(shuffled_predicate);
}
t = first_load->type;
t = t.with_lanes(op->indices.size());
return Load::make(t, first_load->name, shuffled_index, first_load->image,
first_load->param, shuffled_predicate);
}
}
}
// Try to collapse a shuffle of broadcasts into a single
// broadcast. Note that it doesn't matter what the indices
// are.
const Broadcast *b1 = new_vectors[0].as<Broadcast>();
if (b1) {
bool can_collapse = true;
for (size_t i = 1; i < new_vectors.size() && can_collapse; i++) {
if (const Broadcast *b2 = new_vectors[i].as<Broadcast>()) {
Expr check = mutate(b1->value - b2->value);
can_collapse &= is_zero(check);
} else {
can_collapse = false;
}
}
if (can_collapse) {
if (op->indices.size() == 1) {
return b1->value;
} else {
return Broadcast::make(b1->value, op->indices.size());
}
}
}
if (op->is_interleave()) {
int terms = (int)new_vectors.size();
// Try to collapse an interleave of ramps into a single ramp.
const Ramp *r = new_vectors[0].as<Ramp>();
if (r) {
bool can_collapse = true;
for (size_t i = 1; i < new_vectors.size() && can_collapse; i++) {
// If we collapse these terms into a single ramp,
// the new stride is going to be the old stride
// divided by the number of terms, so the
// difference between two adjacent terms in the
// interleave needs to be a broadcast of the new
// stride.
Expr diff = mutate(new_vectors[i] - new_vectors[i-1]);
const Broadcast *b = diff.as<Broadcast>();
if (b) {
Expr check = mutate(b->value * terms - r->stride);
can_collapse &= is_zero(check);
} else {
can_collapse = false;
}
}
if (can_collapse) {
return Ramp::make(r->base, mutate(r->stride / terms), r->lanes * terms);
}
}
// Try to collapse an interleave of slices of vectors from
// the same vector into a single vector.
if (const Shuffle *first_shuffle = new_vectors[0].as<Shuffle>()) {
if (first_shuffle->is_slice()) {
bool can_collapse = true;
for (size_t i = 0; i < new_vectors.size() && can_collapse; i++) {
const Shuffle *i_shuffle = new_vectors[i].as<Shuffle>();
// Check that the current shuffle is a slice...
if (!i_shuffle || !i_shuffle->is_slice()) {
can_collapse = false;
break;
}
// ... and that it is a slice in the right place...
if (i_shuffle->slice_begin() != (int)i || i_shuffle->slice_stride() != terms) {
can_collapse = false;
break;
}
if (i > 0) {
// ... and that the vectors being sliced are the same.
if (first_shuffle->vectors.size() != i_shuffle->vectors.size()) {
can_collapse = false;
break;
}
for (size_t j = 0; j < first_shuffle->vectors.size() && can_collapse; j++) {
if (!equal(first_shuffle->vectors[j], i_shuffle->vectors[j])) {
can_collapse = false;
}
}
}
}
if (can_collapse) {
return Shuffle::make_concat(first_shuffle->vectors);
}
}
}
} else if (op->is_concat()) {
// Try to collapse a concat of ramps into a single ramp.
const Ramp *r = new_vectors[0].as<Ramp>();
if (r) {
bool can_collapse = true;
for (size_t i = 1; i < new_vectors.size() && can_collapse; i++) {
Expr diff;
if (new_vectors[i].type().lanes() == new_vectors[i-1].type().lanes()) {
diff = mutate(new_vectors[i] - new_vectors[i-1]);
}
const Broadcast *b = diff.as<Broadcast>();
if (b) {
Expr check = mutate(b->value - r->stride * new_vectors[i-1].type().lanes());
can_collapse &= is_zero(check);
} else {
can_collapse = false;
}
}
if (can_collapse) {
return Ramp::make(r->base, r->stride, op->indices.size());
}
}
// Try to collapse a concat of scalars into a ramp.
if (new_vectors[0].type().is_scalar() && new_vectors[1].type().is_scalar()) {
bool can_collapse = true;
Expr stride = mutate(new_vectors[1] - new_vectors[0]);
for (size_t i = 1; i < new_vectors.size() && can_collapse; i++) {
if (!new_vectors[i].type().is_scalar()) {
can_collapse = false;
break;
}
Expr check = mutate(new_vectors[i] - new_vectors[i - 1] - stride);
if (!is_zero(check)) {
can_collapse = false;
}
}
if (can_collapse) {
return Ramp::make(new_vectors[0], stride, op->indices.size());
}
}
}
if (!changed) {
return op;
} else {
return Shuffle::make(new_vectors, op->indices);
}
}
template <typename T>
Expr hoist_slice_vector(Expr e) {
const T *op = e.as<T>();
internal_assert(op);
const Shuffle *shuffle_a = op->a.template as<Shuffle>();
const Shuffle *shuffle_b = op->b.template as<Shuffle>();
internal_assert(shuffle_a && shuffle_b &&
shuffle_a->is_slice() &&
shuffle_b->is_slice());
if (shuffle_a->indices != shuffle_b->indices) {
return e;
}
const std::vector<Expr> &slices_a = shuffle_a->vectors;
const std::vector<Expr> &slices_b = shuffle_b->vectors;
if (slices_a.size() != slices_b.size()) {
return e;
}
for (size_t i = 0; i < slices_a.size(); i++) {
if (slices_a[i].type() != slices_b[i].type()) {
return e;
}
}
vector<Expr> new_slices;
for (size_t i = 0; i < slices_a.size(); i++) {
new_slices.push_back(T::make(slices_a[i], slices_b[i]));
}
return Shuffle::make(new_slices, shuffle_a->indices);
}
template<typename T, typename Body>
Body simplify_let(const T *op) {
internal_assert(!var_info.contains(op->name))
<< "Simplify only works on code where every name is unique. Repeated name: " << op->name << "\n";
// If the value is trivial, make a note of it in the scope so
// we can subs it in later
Expr value = mutate(op->value);
Body body = op->body;
// Iteratively peel off certain operations from the let value and push them inside.
Expr new_value = value;
string new_name = op->name + ".s";
Expr new_var = Variable::make(new_value.type(), new_name);
Expr replacement = new_var;
debug(4) << "simplify let " << op->name << " = " << value << " in ... " << op->name << " ...\n";
while (1) {
const Variable *var = new_value.as<Variable>();
const Add *add = new_value.as<Add>();
const Sub *sub = new_value.as<Sub>();
const Mul *mul = new_value.as<Mul>();
const Div *div = new_value.as<Div>();
const Mod *mod = new_value.as<Mod>();
const Min *min = new_value.as<Min>();
const Max *max = new_value.as<Max>();
const Ramp *ramp = new_value.as<Ramp>();
const Cast *cast = new_value.as<Cast>();
const Broadcast *broadcast = new_value.as<Broadcast>();
const Shuffle *shuffle = new_value.as<Shuffle>();
const Variable *var_b = nullptr;
const Variable *var_a = nullptr;
if (add) {
var_a = add->a.as<Variable>();
var_b = add->b.as<Variable>();
} else if (sub) {
var_b = sub->b.as<Variable>();
} else if (mul) {
var_b = mul->b.as<Variable>();
} else if (shuffle && shuffle->is_concat() && shuffle->vectors.size() == 2) {
var_a = shuffle->vectors[0].as<Variable>();
var_b = shuffle->vectors[1].as<Variable>();
}
if (is_const(new_value)) {
replacement = substitute(new_name, new_value, replacement);
new_value = Expr();
break;
} else if (var) {
replacement = substitute(new_name, var, replacement);
new_value = Expr();
break;
} else if (add && (is_const(add->b) || var_b)) {
replacement = substitute(new_name, Add::make(new_var, add->b), replacement);
new_value = add->a;
} else if (add && var_a) {
replacement = substitute(new_name, Add::make(add->a, new_var), replacement);
new_value = add->b;
} else if (mul && (is_const(mul->b) || var_b)) {
replacement = substitute(new_name, Mul::make(new_var, mul->b), replacement);
new_value = mul->a;
} else if (div && is_const(div->b)) {
replacement = substitute(new_name, Div::make(new_var, div->b), replacement);
new_value = div->a;
} else if (sub && (is_const(sub->b) || var_b)) {
replacement = substitute(new_name, Sub::make(new_var, sub->b), replacement);
new_value = sub->a;
} else if (mod && is_const(mod->b)) {
replacement = substitute(new_name, Mod::make(new_var, mod->b), replacement);
new_value = mod->a;
} else if (min && is_const(min->b)) {
replacement = substitute(new_name, Min::make(new_var, min->b), replacement);
new_value = min->a;
} else if (max && is_const(max->b)) {
replacement = substitute(new_name, Max::make(new_var, max->b), replacement);
new_value = max->a;
} else if (ramp && is_const(ramp->stride)) {
new_value = ramp->base;
new_var = Variable::make(new_value.type(), new_name);
replacement = substitute(new_name, Ramp::make(new_var, ramp->stride, ramp->lanes), replacement);
} else if (broadcast) {
new_value = broadcast->value;
new_var = Variable::make(new_value.type(), new_name);
replacement = substitute(new_name, Broadcast::make(new_var, broadcast->lanes), replacement);
} else if (cast && cast->type.bits() > cast->value.type().bits()) {
// Widening casts get pushed inwards, narrowing casts
// stay outside. This keeps the temporaries small, and
// helps with peephole optimizations in codegen that
// skip the widening entirely.
new_value = cast->value;
new_var = Variable::make(new_value.type(), new_name);
replacement = substitute(new_name, Cast::make(cast->type, new_var), replacement);
} else if (shuffle && shuffle->is_slice()) {
// Replacing new_value below might free the shuffle
// indices vector, so save them now.
std::vector<int> slice_indices = shuffle->indices;
new_value = Shuffle::make_concat(shuffle->vectors);
new_var = Variable::make(new_value.type(), new_name);
replacement = substitute(new_name, Shuffle::make({new_var}, slice_indices), replacement);
} else if (shuffle && shuffle->is_concat() &&
((var_a && !var_b) || (!var_a && var_b))) {
new_var = Variable::make(var_a ? shuffle->vectors[1].type() : shuffle->vectors[0].type(), new_name);
Expr op_a = var_a ? shuffle->vectors[0] : new_var;
Expr op_b = var_a ? new_var : shuffle->vectors[1];
replacement = substitute(new_name, Shuffle::make_concat({op_a, op_b}), replacement);
new_value = var_a ? shuffle->vectors[1] : shuffle->vectors[0];
} else {
break;
}
}
if (new_value.same_as(value)) {
// Nothing to substitute
new_value = Expr();
replacement = Expr();
} else {
debug(4) << "new let " << new_name << " = " << new_value << " in ... " << replacement << " ...\n";
}
VarInfo info;
info.old_uses = 0;
info.new_uses = 0;
info.replacement = replacement;
var_info.push(op->name, info);
// Before we enter the body, track the alignment info
bool new_value_alignment_tracked = false, new_value_bounds_tracked = false;
if (new_value.defined() && no_overflow_scalar_int(new_value.type())) {
ModulusRemainder mod_rem = modulus_remainder(new_value, alignment_info);
if (mod_rem.modulus > 1) {
alignment_info.push(new_name, mod_rem);
new_value_alignment_tracked = true;
}
int64_t val_min, val_max;
if (const_int_bounds(new_value, &val_min, &val_max)) {
bounds_info.push(new_name, { val_min, val_max });
new_value_bounds_tracked = true;
}
}
bool value_alignment_tracked = false, value_bounds_tracked = false;;
if (no_overflow_scalar_int(value.type())) {
ModulusRemainder mod_rem = modulus_remainder(value, alignment_info);
if (mod_rem.modulus > 1) {
alignment_info.push(op->name, mod_rem);
value_alignment_tracked = true;
}
int64_t val_min, val_max;
if (const_int_bounds(value, &val_min, &val_max)) {
bounds_info.push(op->name, { val_min, val_max });
value_bounds_tracked = true;
}
}
body = mutate(body);
if (value_alignment_tracked) {
alignment_info.pop(op->name);
}
if (value_bounds_tracked) {
bounds_info.pop(op->name);
}
if (new_value_alignment_tracked) {
alignment_info.pop(new_name);
}
if (new_value_bounds_tracked) {
bounds_info.pop(new_name);
}
info = var_info.get(op->name);
var_info.pop(op->name);
Body result = body;
if (new_value.defined() && info.new_uses > 0) {
// The new name/value may be used
result = T::make(new_name, new_value, result);
}
if (info.old_uses > 0 || !remove_dead_lets) {
// The old name is still in use. We'd better keep it as well.
result = T::make(op->name, value, result);
}
// Don't needlessly make a new Let/LetStmt node. (Here's a
// piece of template syntax I've never needed before).
const T *new_op = result.template as<T>();
if (new_op &&
new_op->name == op->name &&
new_op->body.same_as(op->body) &&
new_op->value.same_as(op->value)) {
return op;
}
return result;
}
Expr visit(const Let *op) override {
return simplify_let<Let, Expr>(op);
}
Stmt visit(const LetStmt *op) override {
return simplify_let<LetStmt, Stmt>(op);
}
Stmt visit(const AssertStmt *op) override {
Stmt stmt = IRMutator2::visit(op);
const AssertStmt *a = stmt.as<AssertStmt>();
if (a && is_zero(a->condition)) {
// Usually, assert(const-false) should generate a warning;
// in at least one case (specialize_fail()), we want to suppress
// the warning, because the assertion is generated internally
// by Halide and is expected to always fail.
const Call *call = a->message.as<Call>();
const bool const_false_conditions_expected =
call && call->name == "halide_error_specialize_fail";
if (!const_false_conditions_expected) {
user_warning << "This pipeline is guaranteed to fail an assertion at runtime: \n"
<< stmt << "\n";
}
} else if (a && is_one(a->condition)) {
stmt = Evaluate::make(0);
}
return stmt;
}
Stmt visit(const For *op) override {
Expr new_min = mutate(op->min);
Expr new_extent = mutate(op->extent);
int64_t new_min_int, new_extent_int;
bool bounds_tracked = false;
if (const_int(new_min, &new_min_int) &&
const_int(new_extent, &new_extent_int)) {
bounds_tracked = true;
int64_t new_max_int = new_min_int + new_extent_int - 1;
bounds_info.push(op->name, { new_min_int, new_max_int });
}
Stmt new_body = mutate(op->body);
if (bounds_tracked) {
bounds_info.pop(op->name);
}
if (is_no_op(new_body)) {
return new_body;
} else if (op->min.same_as(new_min) &&
op->extent.same_as(new_extent) &&
op->body.same_as(new_body)) {
return op;
} else {
return For::make(op->name, new_min, new_extent, op->for_type, op->device_api, new_body);
}
}
Stmt visit(const Provide *op) override {
found_buffer_reference(op->name, op->args.size());
return IRMutator2::visit(op);
}
Stmt visit(const Store *op) override {
found_buffer_reference(op->name);
Expr predicate = mutate(op->predicate);
Expr value = mutate(op->value);
Expr index = mutate(op->index);
const Load *load = value.as<Load>();
const Broadcast *scalar_pred = predicate.as<Broadcast>();
if (is_zero(predicate)) {
// Predicate is always false
return Evaluate::make(0);
} else if (scalar_pred && !is_one(scalar_pred->value)) {
return IfThenElse::make(scalar_pred->value,
Store::make(op->name, value, index, op->param, const_true(value.type().lanes())));
} else if (is_undef(value) || (load && load->name == op->name && equal(load->index, index))) {
// foo[x] = foo[x] or foo[x] = undef is a no-op
return Evaluate::make(0);
} else if (predicate.same_as(op->predicate) && value.same_as(op->value) && index.same_as(op->index)) {
return op;
} else {
return Store::make(op->name, value, index, op->param, predicate);
}
}
Stmt visit(const Allocate *op) override {
std::vector<Expr> new_extents;
bool all_extents_unmodified = true;
for (size_t i = 0; i < op->extents.size(); i++) {
new_extents.push_back(mutate(op->extents[i]));
all_extents_unmodified &= new_extents[i].same_as(op->extents[i]);
}
Stmt body = mutate(op->body);
Expr condition = mutate(op->condition);
Expr new_expr;
if (op->new_expr.defined()) {
new_expr = mutate(op->new_expr);
}
const IfThenElse *body_if = body.as<IfThenElse>();
if (body_if &&
op->condition.defined() &&
equal(op->condition, body_if->condition)) {
// We can move the allocation into the if body case. The
// else case must not use it.
Stmt stmt = Allocate::make(op->name, op->type, op->memory_type,
new_extents, condition, body_if->then_case,
new_expr, op->free_function);
return IfThenElse::make(body_if->condition, stmt, body_if->else_case);
} else if (all_extents_unmodified &&
body.same_as(op->body) &&
condition.same_as(op->condition) &&
new_expr.same_as(op->new_expr)) {
return op;
} else {
return Allocate::make(op->name, op->type, op->memory_type,
new_extents, condition, body,
new_expr, op->free_function);
}
}
Stmt visit(const Evaluate *op) override {
Expr value = mutate(op->value);
// Rewrite Lets inside an evaluate as LetStmts outside the Evaluate.
vector<pair<string, Expr>> lets;
while (const Let *let = value.as<Let>()) {
lets.push_back({let->name, let->value});
value = let->body;
}
if (value.same_as(op->value)) {
internal_assert(lets.empty());
return op;
} else {
// Rewrap the lets outside the evaluate node
Stmt stmt = Evaluate::make(value);
for (size_t i = lets.size(); i > 0; i--) {
stmt = LetStmt::make(lets[i-1].first, lets[i-1].second, stmt);
}
return stmt;
}
}
Stmt visit(const ProducerConsumer *op) override {
Stmt body = mutate(op->body);
if (is_no_op(body)) {
return Evaluate::make(0);
} else if (body.same_as(op->body)) {
return op;
} else {
return ProducerConsumer::make(op->name, op->is_producer, body);
}
}
Stmt visit(const Block *op) override {
Stmt first = mutate(op->first);
Stmt rest = mutate(op->rest);
// Check if both halves start with a let statement.
const LetStmt *let_first = first.as<LetStmt>();
const LetStmt *let_rest = rest.as<LetStmt>();
const IfThenElse *if_first = first.as<IfThenElse>();
const IfThenElse *if_rest = rest.as<IfThenElse>();
if (is_no_op(first) &&
is_no_op(rest)) {
return Evaluate::make(0);
} else if (is_no_op(first)) {
return rest;
} else if (is_no_op(rest)) {
return first;
} else if (let_first &&
let_rest &&
equal(let_first->value, let_rest->value) &&
is_pure(let_first->value)) {
// Do both first and rest start with the same let statement (occurs when unrolling).
Stmt new_block = mutate(Block::make(let_first->body, let_rest->body));
// We need to make a new name since we're pulling it out to a
// different scope.
string var_name = unique_name('t');
Expr new_var = Variable::make(let_first->value.type(), var_name);
new_block = substitute(let_first->name, new_var, new_block);
new_block = substitute(let_rest->name, new_var, new_block);
return LetStmt::make(var_name, let_first->value, new_block);
} else if (if_first &&
if_rest &&
equal(if_first->condition, if_rest->condition) &&
is_pure(if_first->condition)) {
// Two ifs with matching conditions
Stmt then_case = mutate(Block::make(if_first->then_case, if_rest->then_case));
Stmt else_case;
if (if_first->else_case.defined() && if_rest->else_case.defined()) {
else_case = mutate(Block::make(if_first->else_case, if_rest->else_case));
} else if (if_first->else_case.defined()) {
// We already simplified the body of the ifs.
else_case = if_first->else_case;
} else {
else_case = if_rest->else_case;
}
return IfThenElse::make(if_first->condition, then_case, else_case);
} else if (if_first &&
if_rest &&
!if_rest->else_case.defined() &&
is_pure(if_first->condition) &&
is_pure(if_rest->condition) &&
is_one(mutate((if_first->condition && if_rest->condition) == if_rest->condition))) {
// Two ifs where the second condition is tighter than
// the first condition. The second if can be nested
// inside the first one, because if it's true the
// first one must also be true.
Stmt then_case = mutate(Block::make(if_first->then_case, if_rest));
Stmt else_case = mutate(if_first->else_case);
return IfThenElse::make(if_first->condition, then_case, else_case);
} else if (op->first.same_as(first) &&
op->rest.same_as(rest)) {
return op;
} else {
return Block::make(first, rest);
}
}
};
Expr simplify(Expr e, bool remove_dead_lets,
const Scope<Interval> &bounds,
const Scope<ModulusRemainder> &alignment) {
return Simplify(remove_dead_lets, &bounds, &alignment).mutate(e);
}
Stmt simplify(Stmt s, bool remove_dead_lets,
const Scope<Interval> &bounds,
const Scope<ModulusRemainder> &alignment) {
return Simplify(remove_dead_lets, &bounds, &alignment).mutate(s);
}
class SimplifyExprs : public IRMutator2 {
public:
using IRMutator2::mutate;
Expr mutate(const Expr &e) override {
return simplify(e);
}
};
Stmt simplify_exprs(Stmt s) {
return SimplifyExprs().mutate(s);
}
bool can_prove(Expr e) {
internal_assert(e.type().is_bool())
<< "Argument to can_prove is not a boolean Expr: " << e << "\n";
e = simplify(e);
// likely(const-bool) is deliberately left unsimplified, because
// things like max(likely(1), x) are meaningful, but we do want to
// have can_prove(likely(1)) return true.
if (const Call *c = e.as<Call>()) {
if (c->is_intrinsic(Call::likely)) {
e = c->args[0];
}
}
return is_one(e);
}
namespace {
void check(const Expr &a, const Expr &b) {
//debug(0) << "Checking that " << a << " -> " << b << "\n";
Expr simpler = simplify(a);
if (!equal(simpler, b)) {
internal_error
<< "\nSimplification failure:\n"
<< "Input: " << a << '\n'
<< "Output: " << simpler << '\n'
<< "Expected output: " << b << '\n';
}
}
void check(const Stmt &a, const Stmt &b) {
//debug(0) << "Checking that " << a << " -> " << b << "\n";
Stmt simpler = simplify(a);
if (!equal(simpler, b)) {
internal_error
<< "\nSimplification failure:\n"
<< "Input: " << a << '\n'
<< "Output: " << simpler << '\n'
<< "Expected output: " << b << '\n';
}
}
void check_in_bounds(const Expr &a, const Expr &b, const Scope<Interval> &bi) {
//debug(0) << "Checking that " << a << " -> " << b << "\n";
Expr simpler = simplify(a, true, bi);
if (!equal(simpler, b)) {
internal_error
<< "\nSimplification failure:\n"
<< "Input: " << a << '\n'
<< "Output: " << simpler << '\n'
<< "Expected output: " << b << '\n';
}
}
// Helper functions to use in the tests below
Expr interleave_vectors(const vector<Expr> &e) {
return Shuffle::make_interleave(e);
}
Expr concat_vectors(const vector<Expr> &e) {
return Shuffle::make_concat(e);
}
Expr slice(const Expr &e, int begin, int stride, int w) {
return Shuffle::make_slice(e, begin, stride, w);
}
Expr ramp(const Expr &base, const Expr &stride, int w) {
return Ramp::make(base, stride, w);
}
Expr broadcast(const Expr &base, int w) {
return Broadcast::make(base, w);
}
void check_casts() {
Expr x = Var("x");
check(cast(Int(32), cast(Int(32), x)), x);
check(cast(Float(32), 3), 3.0f);
check(cast(Int(32), 5.0f), 5);
check(cast(Int(32), cast(Int(8), 3)), 3);
check(cast(Int(32), cast(Int(8), 1232)), -48);
// Check redundant casts
check(cast(Float(32), cast(Float(64), x)), cast(Float(32), x));
check(cast(Int(16), cast(Int(32), x)), cast(Int(16), x));
check(cast(Int(16), cast(UInt(32), x)), cast(Int(16), x));
check(cast(UInt(16), cast(Int(32), x)), cast(UInt(16), x));
check(cast(UInt(16), cast(UInt(32), x)), cast(UInt(16), x));
// Check evaluation of constant expressions involving casts
check(cast(UInt(16), 53) + cast(UInt(16), 87), make_const(UInt(16), 140));
check(cast(Int(8), 127) + cast(Int(8), 1), make_const(Int(8), -128));
check(cast(UInt(16), -1) - cast(UInt(16), 1), make_const(UInt(16), 65534));
check(cast(Int(16), 4) * cast(Int(16), -5), make_const(Int(16), -20));
check(cast(Int(16), 16) / cast(Int(16), 4), make_const(Int(16), 4));
check(cast(Int(16), 23) % cast(Int(16), 5), make_const(Int(16), 3));
check(min(cast(Int(16), 30000), cast(Int(16), -123)), make_const(Int(16), -123));
check(max(cast(Int(16), 30000), cast(Int(16), 65000)), make_const(Int(16), 30000));
check(cast(UInt(16), -1) == cast(UInt(16), 65535), const_true());
check(cast(UInt(16), 65) == cast(UInt(16), 66), const_false());
check(cast(UInt(16), -1) < cast(UInt(16), 65535), const_false());
check(cast(UInt(16), 65) < cast(UInt(16), 66), const_true());
check(cast(UInt(16), 123.4f), make_const(UInt(16), 123));
check(cast(Float(32), cast(UInt(16), 123456.0f)), 57920.0f);
// Specific checks for 32 bit unsigned expressions - ensure simplifications are actually unsigned.
// 4000000000 (4 billion) is less than 2^32 but more than 2^31. As an int, it is negative.
check(cast(UInt(32), (int) 4000000000UL) + cast(UInt(32), 5), make_const(UInt(32), (int) 4000000005UL));
check(cast(UInt(32), (int) 4000000000UL) - cast(UInt(32), 5), make_const(UInt(32), (int) 3999999995UL));
check(cast(UInt(32), (int) 4000000000UL) / cast(UInt(32), 5), make_const(UInt(32), 800000000));
check(cast(UInt(32), 800000000) * cast(UInt(32), 5), make_const(UInt(32), (int) 4000000000UL));
check(cast(UInt(32), (int) 4000000023UL) % cast(UInt(32), 100), make_const(UInt(32), 23));
check(min(cast(UInt(32), (int) 4000000023UL) , cast(UInt(32), 1000)), make_const(UInt(32), (int) 1000));
check(max(cast(UInt(32), (int) 4000000023UL) , cast(UInt(32), 1000)), make_const(UInt(32), (int) 4000000023UL));
check(cast(UInt(32), (int) 4000000023UL) < cast(UInt(32), 1000), const_false());
check(cast(UInt(32), (int) 4000000023UL) == cast(UInt(32), 1000), const_false());
check(cast(Float(64), 0.5f), Expr(0.5));
check((x - cast(Float(64), 0.5f)) * (x - cast(Float(64), 0.5f)),
(x + Expr(-0.5)) * (x + Expr(-0.5)));
check(cast(Int(64, 3), ramp(5.5f, 2.0f, 3)),
cast(Int(64, 3), ramp(5.5f, 2.0f, 3)));
check(cast(Int(64, 3), ramp(x, 2, 3)),
ramp(cast(Int(64), x), cast(Int(64), 2), 3));
// Check cancellations can occur through casts
check(cast(Int(64), x + 1) - cast(Int(64), x), cast(Int(64), 1));
check(cast(Int(64), 1 + x) - cast(Int(64), x), cast(Int(64), 1));
// But only when overflow is undefined for the type
check(cast(UInt(8), x + 1) - cast(UInt(8), x),
cast(UInt(8), x + 1) - cast(UInt(8), x));
}
void check_algebra() {
Expr x = Var("x"), y = Var("y"), z = Var("z"), w = Var("w"), v = Var("v");
Expr xf = cast<float>(x);
Expr yf = cast<float>(y);
Expr t = const_true(), f = const_false();
check(3 + x, x + 3);
check(x + 0, x);
check(0 + x, x);
check(Expr(ramp(x, 2, 3)) + Expr(ramp(y, 4, 3)), ramp(x+y, 6, 3));
check(Expr(broadcast(4.0f, 5)) + Expr(ramp(3.25f, 4.5f, 5)), ramp(7.25f, 4.5f, 5));
check(Expr(ramp(3.25f, 4.5f, 5)) + Expr(broadcast(4.0f, 5)), ramp(7.25f, 4.5f, 5));
check(Expr(broadcast(3, 3)) + Expr(broadcast(1, 3)), broadcast(4, 3));
check((x + 3) + 4, x + 7);
check(4 + (3 + x), x + 7);
check((x + 3) + y, (x + y) + 3);
check(y + (x + 3), (y + x) + 3);
check((3 - x) + x, 3);
check(x + (3 - x), 3);
check(x*y + x*z, x*(y+z));
check(x*y + z*x, x*(y+z));
check(y*x + x*z, x*(y+z));
check(y*x + z*x, x*(y+z));
check(x - 0, x);
check((x/y) - (x/y), 0);
check(x - 2, x + (-2));
check(Expr(ramp(x, 2, 3)) - Expr(ramp(y, 4, 3)), ramp(x-y, -2, 3));
check(Expr(broadcast(4.0f, 5)) - Expr(ramp(3.25f, 4.5f, 5)), ramp(0.75f, -4.5f, 5));
check(Expr(ramp(3.25f, 4.5f, 5)) - Expr(broadcast(4.0f, 5)), ramp(-0.75f, 4.5f, 5));
check(Expr(broadcast(3, 3)) - Expr(broadcast(1, 3)), broadcast(2, 3));
check((x + y) - x, y);
check((x + y) - y, x);
check(x - (x + y), 0 - y);
check(x - (y + x), 0 - y);
check((x + 3) - 2, x + 1);
check((x + 3) - y, (x - y) + 3);
check((x - 3) - y, (x - y) + (-3));
check(x - (y - 2), (x - y) + 2);
check(3 - (y - 2), 5 - y);
check(x - (0 - y), x + y);
check(x + (0 - y), x - y);
check((0 - x) + y, y - x);
check(x*y - x*z, x*(y-z));
check(x*y - z*x, x*(y-z));
check(y*x - x*z, x*(y-z));
check(y*x - z*x, x*(y-z));
check(x - y*-2, x + y*2);
check(x + y*-2, x - y*2);
check(x*-2 + y, y - x*2);
check(xf - yf*-2.0f, xf + y*2.0f);
check(xf + yf*-2.0f, xf - y*2.0f);
check(xf*-2.0f + yf, yf - x*2.0f);
check((x * 8) - (y * 4), (x * 2 - y) * 4);
check((x * 4) - (y * 8), (x - y * 2) * 4);
check((x * 2) % 6, (x % 3) * 2);
check(x - (x/8)*8, x % 8);
check((x/8)*8 - x, -(x % 8));
check((x/8)*8 < x + y, 0 < x%8 + y);
check((x/8)*8 < x - y, y < x%8);
check((x/8)*8 < x, 0 < x%8);
check(((x+3)/8)*8 < x + y, 3 < (x+3)%8 + y);
check(((x+3)/8)*8 < x - y, y < (x+3)%8 + (-3));
check(((x+3)/8)*8 < x, 3 < (x+3)%8);
check(x*0, 0);
check(0*x, 0);
check(x*1, x);
check(1*x, x);
check(Expr(2.0f)*4.0f, 8.0f);
check(Expr(2)*4, 8);
check((3*x)*4, x*12);
check(4*(3+x), x*4 + 12);
check(Expr(broadcast(4.0f, 5)) * Expr(ramp(3.0f, 4.0f, 5)), ramp(12.0f, 16.0f, 5));
check(Expr(ramp(3.0f, 4.0f, 5)) * Expr(broadcast(2.0f, 5)), ramp(6.0f, 8.0f, 5));
check(Expr(broadcast(3, 3)) * Expr(broadcast(2, 3)), broadcast(6, 3));
check(x*y + x, x*(y + 1));
check(x*y - x, x*(y + -1));
check(x + x*y, x*(y + 1));
check(x - x*y, x*(1 - y));
check(x*y + y, (x + 1)*y);
check(x*y - y, (x + -1)*y);
check(y + x*y, (x + 1)*y);
check(y - x*y, (1 - x)*y);
check(0/x, 0);
check(x/1, x);
check(x/x, 1);
check((-1)/x, select(x < 0, 1, -1));
check(Expr(7)/3, 2);
check(Expr(6.0f)/2.0f, 3.0f);
check((x / 3) / 4, x / 12);
check((x*4)/2, x*2);
check((x*2)/4, x/2);
check((x*4 + y)/2, x*2 + y/2);
check((y + x*4)/2, y/2 + x*2);
check((x*4 - y)/2, x*2 + (0 - y)/2);
check((y - x*4)/2, y/2 - x*2);
check((x + 3)/2 + 7, (x + 17)/2);
check((x/2 + 3)/5, (x + 6)/10);
check((x + (y + 3)/5) + 5, x + (y + 28)/5);
check((x + 8)/2, x/2 + 4);
check((x - y)*-2, (y - x)*2);
check((xf - yf)*-2.0f, (yf - xf)*2.0f);
check(x*3 + y*9, (x + y*3)*3);
check(x*9 + y*3, (x*3 + y)*3);
// Pull terms that are a multiple of the divisor out of a ternary expression
check(((x*4 + y) + z) / 2, x*2 + (y + z)/2);
check(((x*4 - y) + z) / 2, x*2 + (z - y)/2);
check(((x*4 + y) - z) / 2, x*2 + (y - z)/2);
check(((x*4 - y) - z) / 2, x*2 + (0 - y - z)/2);
check((x + (y*4 + z)) / 2, y*2 + (x + z)/2);
check(((x + y*4) + z) / 2, y*2 + (x + z)/2);
check((x + (y*4 - z)) / 2, y*2 + (x - z)/2);
check((x - (y*4 + z)) / 2, (x - z)/2 - y*2);
check((x - (y*4 - z)) / 2, (x + z)/2 - y*2);
// Pull out the gcd of the numerator and divisor
check((x * 12 + 5) / 9, (x * 4 + 1) / 3);
check((x * 12 + 19) / 9, (x * 4) / 3 + 2);
// Cancellations in non-const integer divisions
check((x*y)/x, y);
check((y*x)/x, y);
check((x*y + z)/x, y + z/x);
check((y*x + z)/x, y + z/x);
check((z + x*y)/x, z/x + y);
check((z + y*x)/x, z/x + y);
check((x*y - z)/x, y + (-z)/x);
check((y*x - z)/x, y + (-z)/x);
check((z - x*y)/x, z/x - y);
check((z - y*x)/x, z/x - y);
check((x + y)/x, y/x + 1);
check((y + x)/x, y/x + 1);
check((x - y)/x, (-y)/x + 1);
check((y - x)/x, y/x + (-1));
check(((x + y) + z)/x, (y + z)/x + 1);
check(((y + x) + z)/x, (y + z)/x + 1);
check((y + (x + z))/x, (y + z)/x + 1);
check((y + (z + x))/x, (y + z)/x + 1);
check(xf / 4.0f, xf * 0.25f);
// Some quaternary rules with cancellations
check((x + y) - (z + y), x - z);
check((x + y) - (y + z), x - z);
check((y + x) - (z + y), x - z);
check((y + x) - (y + z), x - z);
check((x - y) - (z - y), x - z);
check((y - z) - (y - x), x - z);
check(((x + y) + z) - x, y + z);
check(((x + y) + z) - y, x + z);
check((x + (y + z)) - y, x + z);
check((x + (y + z)) - z, x + y);
check((x*8) % 4, 0);
check((x*8 + y) % 4, y % 4);
check((y + 8) % 4, y % 4);
check((y + x*8) % 4, y % 4);
check((y*16 + 13) % 2, 1);
check((x*y) % 1, 0);
// Check an optimization important for fusing dimensions
check((x/3)*3 + x%3, x);
check(x%3 + (x/3)*3, x);
check(((x/3)*3 + y) + x%3, x + y);
check(((x/3) + y)*3 + x%3, x + y*3);
check((x%3 + y) + (x/3)*3, x + y);
check((y + x%3) + (x/3)*3, y + x);
check((y + (x/3*3)) + x%3, y + x);
check((y + (x/3))*3 + x%3, y*3 + x);
check(x/2 + x%2, (x + 1)/2);
check(x%2 + x/2, (x + 1)/2);
check(((x+1)/2)*2 - x, x%2);
check(x - ((x+1)/2)*2, -(x%2));
check((x%2 + 4)/2, 2);
check((x%2 + 5)/2, x%2 + 2);
// Almost-cancellations through integer divisions. These rules all
// deduplicate x and wrap it in a modulo operator, neutering it
// for the purposes of bounds inference. Patterns below look
// confusing, but were brute-force tested.
check((x + 17)/3 - (x + 7)/3, ((x+1)%3 + 10)/3);
check((x + 17)/3 - (x + y)/3, (19 - y - (x+2)%3)/3);
check((x + y )/3 - (x + 7)/3, ((x+1)%3 + y + -7)/3);
check( x /3 - (x + y)/3, (2 - y - x % 3)/3);
check((x + y )/3 - x /3, (x%3 + y)/3);
check( x /3 - (x + 7)/3, (-5 - x%3)/3);
check((x + 17)/3 - x /3, (x%3 + 17)/3);
check((x + 17)/3 - (x - y)/3, (y - (x+2)%3 + 19)/3);
check((x - y )/3 - (x + 7)/3, ((x+1)%3 - y + (-7))/3);
check( x /3 - (x - y)/3, (y - x%3 + 2)/3);
check((x - y )/3 - x /3, (x%3 - y)/3);
// Check some specific expressions involving div and mod
check(Expr(23) / 4, Expr(5));
check(Expr(-23) / 4, Expr(-6));
check(Expr(-23) / -4, Expr(6));
check(Expr(23) / -4, Expr(-5));
check(Expr(-2000000000) / 1000000001, Expr(-2));
check(Expr(23) % 4, Expr(3));
check(Expr(-23) % 4, Expr(1));
check(Expr(-23) % -4, Expr(1));
check(Expr(23) % -4, Expr(3));
check(Expr(-2000000000) % 1000000001, Expr(2));
check(Expr(3) + Expr(8), 11);
check(Expr(3.25f) + Expr(7.75f), 11.0f);
check(Expr(7) % 2, 1);
check(Expr(7.25f) % 2.0f, 1.25f);
check(Expr(-7.25f) % 2.0f, 0.75f);
check(Expr(-7.25f) % -2.0f, -1.25f);
check(Expr(7.25f) % -2.0f, -0.75f);
check(2*x + (2*x + y)/5, (x*12 + y)/5);
check(x + (x - y)/4, (x*5 - y)/4);
check((x + z) + (y + (x + z))/3, ((x + z)*4 + y)/3);
check(x + ((y + w) - x)/2, (x + (y + w))/2);
check((x + y)/3 + x, (x*4 + y)/3);
check((x - y)/4 + x, (x*5 - y)/4);
check((y + x)/3 + x, (y + x*4)/3);
check((y - x)/3 + x, (y + x*2)/3);
check(1 + (1 + y)/2, (y + 3)/2);
check((y + 1)/2 + 1, (y + 3)/2);
check((0 - y)/5 + 1, (0 - y)/5 + 1);
check(x - (x + y)/3, (x*2 - y + 2)/3);
check((w + x) - ((w + x) - y*z)/3, ((w + x)*2 + y*z + 2)/3);
check(x - (y + x)/2, (x - y + 1)/2);
check(x - (y - x)/6, (x*7 - y + 5)/6);
check(x - (x + y)/-3, x - (x + y)/-3);
check((w + x) - ((w + x) - y*z)/-3, (w + x) - ((w + x) - y*z)/-3);
check(x - (y + x)/-2, x - (y + x)/-2);
check(x - (y - x)/-6, x - (y - x)/-6);
check((x + y)/3 - x, (y - x*2)/3);
check((x*y - w)/4 - x*y, (x*y*(-3) - w)/4);
check((y + x)/5 - x, (y - x*4)/5);
check((y - x)/6 - x, (y - x*7)/6);
check(1 - (1 + y)/2 - 1, (0 - y)/2);
check(1 - (-y + 1)/2 - 1, y/2);
check(1 - (0 - y)/5, (y + 9)/5);
}
void check_vectors() {
Expr x = Var("x"), y = Var("y"), z = Var("z");
check(Expr(broadcast(y, 4)) / Expr(broadcast(x, 4)),
Expr(broadcast(y/x, 4)));
check(Expr(ramp(x, 4, 4)) / 2, ramp(x/2, 2, 4));
check(Expr(ramp(x, -4, 7)) / 2, ramp(x/2, -2, 7));
check(Expr(ramp(x, 4, 5)) / -2, ramp(x/-2, -2, 5));
check(Expr(ramp(x, -8, 5)) / -2, ramp(x/-2, 4, 5));
check(Expr(ramp(4*x, 1, 4)) / 4, broadcast(x, 4));
check(Expr(ramp(x*4, 1, 3)) / 4, broadcast(x, 3));
check(Expr(ramp(x*8, 2, 4)) / 8, broadcast(x, 4));
check(Expr(ramp(x*8, 3, 3)) / 8, broadcast(x, 3));
check(Expr(ramp(0, 1, 8)) % 16, Expr(ramp(0, 1, 8)));
check(Expr(ramp(8, 1, 8)) % 16, Expr(ramp(8, 1, 8)));
check(Expr(ramp(9, 1, 8)) % 16, Expr(ramp(9, 1, 8)) % 16);
check(Expr(ramp(16, 1, 8)) % 16, Expr(ramp(0, 1, 8)));
check(Expr(ramp(0, 1, 8)) % 8, Expr(ramp(0, 1, 8)));
check(Expr(ramp(x*8+17, 1, 4)) % 8, Expr(ramp(1, 1, 4)));
check(Expr(ramp(x*8+17, 1, 8)) % 8, Expr(ramp(1, 1, 8) % 8));
check(Expr(broadcast(x, 4)) % Expr(broadcast(y, 4)),
Expr(broadcast(x % y, 4)));
check(Expr(ramp(x, 2, 4)) % (broadcast(2, 4)),
broadcast(x % 2, 4));
check(Expr(ramp(2*x+1, 4, 4)) % (broadcast(2, 4)),
broadcast(1, 4));
check(max(broadcast(24, 2), broadcast(x, 2) % ramp(-8, -33, 2)),
max(broadcast(x, 2) % ramp(-8, -33, 2), broadcast(24, 2)));
check(max(broadcast(41, 2), broadcast(x, 2) % ramp(-8, -33, 2)),
broadcast(41, 2));
check(ramp(0, 1, 4) == broadcast(2, 4),
ramp(-2, 1, 4) == broadcast(0, 4));
{
Expr test = select(ramp(const_true(), const_true(), 2),
ramp(const_false(), const_true(), 2),
broadcast(const_false(), 2)) ==
broadcast(const_false(), 2);
Expr expected = !(ramp(const_true(), const_true(), 2)) ||
(ramp(const_false(), const_true(), 2) == broadcast(const_false(), 2));
check(test, expected);
}
{
Expr test = select(ramp(const_true(), const_true(), 2),
broadcast(const_true(), 2),
ramp(const_false(), const_true(), 2)) ==
broadcast(const_false(), 2);
Expr expected = (!ramp(const_true(), const_true(), 2)) &&
(ramp(const_false(), const_true(), 2) == broadcast(const_false(), 2));
check(test, expected);
}
}
void check_bounds() {
Expr x = Var("x"), y = Var("y"), z = Var("z"), w = Var("w");
check(min(Expr(7), 3), 3);
check(min(Expr(4.25f), 1.25f), 1.25f);
check(min(broadcast(x, 4), broadcast(y, 4)),
broadcast(min(x, y), 4));
check(min(x, x+3), x);
check(min(x+4, x), x);
check(min(x-1, x+2), x+(-1));
check(min(7, min(x, 3)), min(x, 3));
check(min(min(x, y), x), min(x, y));
check(min(min(x, y), y), min(x, y));
check(min(x, min(x, y)), min(x, y));
check(min(y, min(x, y)), min(x, y));
check(max(Expr(7), 3), 7);
check(max(Expr(4.25f), 1.25f), 4.25f);
check(max(broadcast(x, 4), broadcast(y, 4)),
broadcast(max(x, y), 4));
check(max(x, x+3), x+3);
check(max(x+4, x), x+4);
check(max(x-1, x+2), x+2);
check(max(7, max(x, 3)), max(x, 7));
check(max(max(x, y), x), max(x, y));
check(max(max(x, y), y), max(x, y));
check(max(x, max(x, y)), max(x, y));
check(max(y, max(x, y)), max(x, y));
// Check that simplifier can recognise instances where the extremes of the
// datatype appear as constants in comparisons, Min and Max expressions.
// The result of min/max with extreme is known to be either the extreme or
// the other expression. The result of < or > comparison is known to be true or false.
check(x <= Int(32).max(), const_true());
check(cast(Int(16), x) >= Int(16).min(), const_true());
check(x < Int(32).min(), const_false());
check(min(cast(UInt(16), x), cast(UInt(16), 65535)), cast(UInt(16), x));
check(min(x, Int(32).max()), x);
check(min(Int(32).min(), x), Int(32).min());
check(max(cast(Int(8), x), cast(Int(8), -128)), cast(Int(8), x));
check(max(x, Int(32).min()), x);
check(max(x, Int(32).max()), Int(32).max());
// Check that non-extremes do not lead to incorrect simplification
check(max(cast(Int(8), x), cast(Int(8), -127)), max(cast(Int(8), x), make_const(Int(8), -127)));
// Some quaternary rules with cancellations
check((x + y) - (z + y), x - z);
check((x + y) - (y + z), x - z);
check((y + x) - (z + y), x - z);
check((y + x) - (y + z), x - z);
check((x - y) - (z - y), x - z);
check((y - z) - (y - x), x - z);
check((x + 3) / 4 - (x + 2) / 4, ((x + 2) % 4 + 1)/4);
check(x - min(x + y, z), max(-y, x-z));
check(x - min(y + x, z), max(-y, x-z));
check(x - min(z, x + y), max(-y, x-z));
check(x - min(z, y + x), max(-y, x-z));
check(min(x + y, z) - x, min(y, z-x));
check(min(y + x, z) - x, min(y, z-x));
check(min(z, x + y) - x, min(y, z-x));
check(min(z, y + x) - x, min(y, z-x));
check(min(x + y, z + y), min(x, z) + y);
check(min(y + x, z + y), min(x, z) + y);
check(min(x + y, y + z), min(x, z) + y);
check(min(y + x, y + z), min(x, z) + y);
check(min(x, y) - min(y, x), 0);
check(max(x, y) - max(y, x), 0);
check(min(123 - x, 1 - x), 1 - x);
check(max(123 - x, 1 - x), 123 - x);
check(min(x*43, y*43), min(x, y)*43);
check(max(x*43, y*43), max(x, y)*43);
check(min(x*-43, y*-43), max(x, y)*-43);
check(max(x*-43, y*-43), min(x, y)*-43);
check(min(min(x, 4), y), min(min(x, y), 4));
check(max(max(x, 4), y), max(max(x, y), 4));
check(min(x*8, 24), min(x, 3)*8);
check(max(x*8, 24), max(x, 3)*8);
check(min(x*-8, 24), max(x, -3)*-8);
check(max(x*-8, 24), min(x, -3)*-8);
check(min(clamp(x, -10, 14), clamp(y, -10, 14)), clamp(min(x, y), -10, 14));
check(min(x/4, y/4), min(x, y)/4);
check(max(x/4, y/4), max(x, y)/4);
check(min(x/(-4), y/(-4)), max(x, y)/(-4));
check(max(x/(-4), y/(-4)), min(x, y)/(-4));
check(min(x/4 + 2, y/4), min(x + 8, y)/4);
check(max(x/4 + 2, y/4), max(x + 8, y)/4);
check(min(x/4, y/4 + 2), min(x, y + 8)/4);
check(max(x/4, y/4 + 2), max(x, y + 8)/4);
check(min(x/(-4) + 2, y/(-4)), max(x + -8, y)/(-4));
check(max(x/(-4) + 2, y/(-4)), min(x + -8, y)/(-4));
check(min(x/(-4), y/(-4) + 2), max(x, y + -8)/(-4));
check(max(x/(-4), y/(-4) + 2), min(x, y + -8)/(-4));
check(min(x*4 + 8, y*4), min(x + 2, y)*4);
check(max(x*4 + 8, y*4), max(x + 2, y)*4);
check(min(x*4, y*4 + 8), min(x, y + 2)*4);
check(max(x*4, y*4 + 8), max(x, y + 2)*4);
check(min(x*(-4) + 8, y*(-4)), max(x + -2, y)*(-4));
check(max(x*(-4) + 8, y*(-4)), min(x + -2, y)*(-4));
check(min(x*(-4), y*(-4) + 8), max(x, y + -2)*(-4));
check(max(x*(-4), y*(-4) + 8), min(x, y + -2)*(-4));
// Min and max of clamped expressions
check(min(clamp(x+1, y, z), clamp(x-1, y, z)), clamp(x+(-1), y, z));
check(max(clamp(x+1, y, z), clamp(x-1, y, z)), clamp(x+1, y, z));
// Additions that cancel a term inside a min or max
check(x + min(y - x, z), min(y, z + x));
check(x + max(y - x, z), max(y, z + x));
check(min(y + (-2), z) + 2, min(y, z + 2));
check(max(y + (-2), z) + 2, max(y, z + 2));
check(x + min(y - x, z), min(y, z + x));
check(x + max(y - x, z), max(y, z + x));
check(min(y + (-2), z) + 2, min(y, z + 2));
check(max(y + (-2), z) + 2, max(y, z + 2));
// Min/Max distributive law
check(max(max(x, y), max(x, z)), max(max(y, z), x));
check(min(max(x, y), max(x, z)), max(min(y, z), x));
check(min(min(x, y), min(x, z)), min(min(y, z), x));
check(max(min(x, y), min(x, z)), min(max(y, z), x));
// Mins of expressions and rounded up versions of them
check(min(((x+7)/8)*8, x), x);
check(min(x, ((x+7)/8)*8), x);
check(max(((x+7)/8)*8, x), ((x+7)/8)*8);
check(max(x, ((x+7)/8)*8), ((x+7)/8)*8);
// And rounded down...
check(max((x/8)*8, x), x);
check(max(x, (x/8)*8), x);
check(min((x/8)*8, x), (x/8)*8);
check(min(x, (x/8)*8), (x/8)*8);
check(min(((x+7)/8)*8, max(x, 8)), max(x, 8));
check(min(max(x, 8), ((x+7)/8)*8), max(x, 8));
check(min(x, likely(x)), likely(x));
check(min(likely(x), x), likely(x));
check(max(x, likely(x)), likely(x));
check(max(likely(x), x), likely(x));
check(select(x > y, likely(x), x), likely(x));
check(select(x > y, x, likely(x)), likely(x));
check(min(x + 1, y) - min(x, y - 1), 1);
check(max(x + 1, y) - max(x, y - 1), 1);
check(min(x + 1, y) - min(y - 1, x), 1);
check(max(x + 1, y) - max(y - 1, x), 1);
// min and max on constant ramp v broadcast
check(max(ramp(0, 1, 8), 0), ramp(0, 1, 8));
check(min(ramp(0, 1, 8), 7), ramp(0, 1, 8));
check(max(ramp(0, 1, 8), 7), broadcast(7, 8));
check(min(ramp(0, 1, 8), 0), broadcast(0, 8));
check(min(ramp(0, 1, 8), 4), min(ramp(0, 1, 8), 4));
check(max(ramp(7, -1, 8), 0), ramp(7, -1, 8));
check(min(ramp(7, -1, 8), 7), ramp(7, -1, 8));
check(max(ramp(7, -1, 8), 7), broadcast(7, 8));
check(min(ramp(7, -1, 8), 0), broadcast(0, 8));
check(min(ramp(7, -1, 8), 4), min(ramp(7, -1, 8), 4));
check(max(0, ramp(0, 1, 8)), ramp(0, 1, 8));
check(min(7, ramp(0, 1, 8)), ramp(0, 1, 8));
check(min(8 - x, 2), 8 - max(x, 6));
check(max(3, 77 - x), 77 - min(x, 74));
check(min(max(8-x, 0), 8), 8 - max(min(x, 8), 0));
check(x - min(x, 2), max(x + -2, 0));
check(x - max(x, 2), min(x + -2, 0));
check(min(x, 2) - x, 2 - max(x, 2));
check(max(x, 2) - x, 2 - min(x, 2));
check(x - min(2, x), max(x + -2, 0));
check(x - max(2, x), min(x + -2, 0));
check(min(2, x) - x, 2 - max(x, 2));
check(max(2, x) - x, 2 - min(x, 2));
check(max(min(x, y), x), x);
check(max(min(x, y), y), y);
check(min(max(x, y), x), x);
check(min(max(x, y), y), y);
check(max(min(x, y), x) + y, x + y);
check(max(min(max(x, y), z), y), max(min(x, z), y));
check(max(min(z, max(x, y)), y), max(min(x, z), y));
check(max(y, min(max(x, y), z)), max(min(x, z), y));
check(max(y, min(z, max(x, y))), max(min(x, z), y));
check(max(min(max(y, x), z), y), max(min(x, z), y));
check(max(min(z, max(y, x)), y), max(min(x, z), y));
check(max(y, min(max(y, x), z)), max(min(x, z), y));
check(max(y, min(z, max(y, x))), max(min(x, z), y));
check(min(max(min(x, y), z), y), min(max(x, z), y));
check(min(max(z, min(x, y)), y), min(max(x, z), y));
check(min(y, max(min(x, y), z)), min(max(x, z), y));
check(min(y, max(z, min(x, y))), min(max(x, z), y));
check(min(max(min(y, x), z), y), min(max(x, z), y));
check(min(max(z, min(y, x)), y), min(max(x, z), y));
check(min(y, max(min(y, x), z)), min(max(x, z), y));
check(min(y, max(z, min(y, x))), min(max(x, z), y));
{
Expr one = broadcast(cast(Int(16), 1), 64);
Expr three = broadcast(cast(Int(16), 3), 64);
Expr four = broadcast(cast(Int(16), 4), 64);
Expr five = broadcast(cast(Int(16), 5), 64);
Expr v1 = Variable::make(Int(16).with_lanes(64), "x");
Expr v2 = Variable::make(Int(16).with_lanes(64), "y");
// Bound: [-4, 4]
std::vector<Expr> clamped = {
max(min(v1, four), -four),
max(-four, min(v1, four)),
min(max(v1, -four), four),
min(four, max(v1, -four)),
clamp(v1, -four, four)
};
for (size_t i = 0; i < clamped.size(); ++i) {
// min(v, 4) where v=[-4, 4] -> v
check(min(clamped[i], four), simplify(clamped[i]));
// min(v, 5) where v=[-4, 4] -> v
check(min(clamped[i], five), simplify(clamped[i]));
// min(v, 3) where v=[-4, 4] -> min(v, 3)
check(min(clamped[i], three), simplify(min(clamped[i], three)));
// min(v, -5) where v=[-4, 4] -> -5
check(min(clamped[i], -five), simplify(-five));
}
for (size_t i = 0; i < clamped.size(); ++i) {
// max(v, 4) where v=[-4, 4] -> 4
check(max(clamped[i], four), simplify(four));
// max(v, 5) where v=[-4, 4] -> 5
check(max(clamped[i], five), simplify(five));
// max(v, 3) where v=[-4, 4] -> max(v, 3)
check(max(clamped[i], three), simplify(max(clamped[i], three)));
// max(v, -5) where v=[-4, 4] -> v
check(max(clamped[i], -five), simplify(clamped[i]));
}
for (size_t i = 0; i < clamped.size(); ++i) {
// max(min(v, 5), -5) where v=[-4, 4] -> v
check(max(min(clamped[i], five), -five), simplify(clamped[i]));
// max(min(v, 5), 5) where v=[-4, 4] -> 5
check(max(min(clamped[i], five), five), simplify(five));
// max(min(v, -5), -5) where v=[-4, 4] -> -5
check(max(min(clamped[i], -five), -five), simplify(-five));
// max(min(v, -5), 5) where v=[-4, 4] -> 5
check(max(min(clamped[i], -five), five), simplify(five));
// max(min(v, -5), 3) where v=[-4, 4] -> 3
check(max(min(clamped[2], -five), three), simplify(three));
}
// max(min(v, 5), 3) where v=[-4, 4] -> max(v, 3)
check(max(min(clamped[2], five), three), simplify(max(clamped[2], three)));
// max(min(v, 5), 3) where v=[-4, 4] -> max(v, 3) -> v=[3, 4]
// There is simplification rule that will simplify max(max(min(x, 4), -4), 3)
// further into max(min(x, 4), 3)
check(max(min(clamped[0], five), three), simplify(max(min(v1, four), three)));
for (size_t i = 0; i < clamped.size(); ++i) {
// min(v + 1, 4) where v=[-4, 4] -> min(v + 1, 4)
check(min(clamped[i] + one, four), simplify(min(clamped[i] + one, four)));
// min(v + 1, 5) where v=[-4, 4] -> v + 1
check(min(clamped[i] + one, five), simplify(clamped[i] + one));
// min(v + 1, -4) where v=[-4, 4] -> -4
check(min(clamped[i] + one, -four), simplify(-four));
// max(min(v + 1, 4), -4) where v=[-4, 4] -> min(v + 1, 4)
check(max(min(clamped[i] + one, four), -four), simplify(min(clamped[i] + one, four)));
}
for (size_t i = 0; i < clamped.size(); ++i) {
// max(v + 1, 4) where v=[-4, 4] -> max(v + 1, 4)
check(max(clamped[i] + one, four), simplify(max(clamped[i] + one, four)));
// max(v + 1, 5) where v=[-4, 4] -> 5
check(max(clamped[i] + one, five), simplify(five));
// max(v + 1, -4) where v=[-4, 4] -> -v + 1
check(max(clamped[i] + one, -four), simplify(clamped[i] + one));
// min(max(v + 1, -4), 4) where v=[-4, 4] -> min(v + 1, 4)
check(min(max(clamped[i] + one, -four), four), simplify(min(clamped[i] + one, four)));
}
Expr t1 = clamp(v1, one, four);
Expr t2 = clamp(v1, -five, -four);
check(min(max(min(v2, t1), t2), five), simplify(max(min(t1, v2), t2)));
}
{
Expr xv = Variable::make(Int(16).with_lanes(64), "x");
Expr yv = Variable::make(Int(16).with_lanes(64), "y");
Expr zv = Variable::make(Int(16).with_lanes(64), "z");
// min(min(x, broadcast(y, n)), broadcast(z, n))) -> min(x, broadcast(min(y, z), n))
check(min(min(xv, broadcast(y, 64)), broadcast(z, 64)), min(xv, broadcast(min(y, z), 64)));
// min(min(broadcast(x, n), y), broadcast(z, n))) -> min(y, broadcast(min(x, z), n))
check(min(min(broadcast(x, 64), yv), broadcast(z, 64)), min(yv, broadcast(min(x, z), 64)));
// min(broadcast(x, n), min(y, broadcast(z, n)))) -> min(y, broadcast(min(x, z), n))
check(min(broadcast(x, 64), min(yv, broadcast(z, 64))), min(yv, broadcast(min(z, x), 64)));
// min(broadcast(x, n), min(broadcast(y, n), z))) -> min(z, broadcast(min(x, y), n))
check(min(broadcast(x, 64), min(broadcast(y, 64), zv)), min(zv, broadcast(min(y, x), 64)));
// max(max(x, broadcast(y, n)), broadcast(z, n))) -> max(x, broadcast(max(y, z), n))
check(max(max(xv, broadcast(y, 64)), broadcast(z, 64)), max(xv, broadcast(max(y, z), 64)));
// max(max(broadcast(x, n), y), broadcast(z, n))) -> max(y, broadcast(max(x, z), n))
check(max(max(broadcast(x, 64), yv), broadcast(z, 64)), max(yv, broadcast(max(x, z), 64)));
// max(broadcast(x, n), max(y, broadcast(z, n)))) -> max(y, broadcast(max(x, z), n))
check(max(broadcast(x, 64), max(yv, broadcast(z, 64))), max(yv, broadcast(max(z, x), 64)));
// max(broadcast(x, n), max(broadcast(y, n), z))) -> max(z, broadcast(max(x, y), n))
check(max(broadcast(x, 64), max(broadcast(y, 64), zv)), max(zv, broadcast(max(y, x), 64)));
}
// Pull out common addition term inside min/max
check(min((x + y) + z, x + w), min(y + z, w) + x);
check(min((y + x) + z, x + w), min(y + z, w) + x);
check(min(x + y, (x + z) + w), min(y, z + w) + x);
check(min(x + y, (z + x) + w), min(y, z + w) + x);
check(min(x + (y + z), y + w), min(x + z, w) + y);
check(min(x + (z + y), y + w), min(x + z, w) + y);
check(min(x + y, z + (x + w)), min(y, z + w) + x);
check(min(x + y, z + (w + x)), min(y, z + w) + x);
check(min(x + y/2 + 13, x + (0 - y)/2), min(y + 26, 0 - y)/2 + x);
check(max((x + y) + z, x + w), max(y + z, w) + x);
check(max((y + x) + z, x + w), max(y + z, w) + x);
check(max(x + y, (x + z) + w), max(y, z + w) + x);
check(max(x + y, (z + x) + w), max(y, z + w) + x);
check(max(x + (y + z), y + w), max(x + z, w) + y);
check(max(x + (z + y), y + w), max(x + z, w) + y);
check(max(x + y, z + (x + w)), max(y, z + w) + x);
check(max(x + y, z + (w + x)), max(y, z + w) + x);
}
void check_boolean() {
Expr x = Var("x"), y = Var("y"), z = Var("z"), w = Var("w");
Expr xf = cast<float>(x);
Expr yf = cast<float>(y);
Expr t = const_true(), f = const_false();
Expr b1 = Variable::make(Bool(), "b1");
Expr b2 = Variable::make(Bool(), "b2");
check(x == x, t);
check(x == (x+1), f);
check(x-2 == y+3, (x-y) == 5);
check(x+y == y+z, x == z);
check(y+x == y+z, x == z);
check(x+y == z+y, x == z);
check(y+x == z+y, x == z);
check((y+x)*17 == (z+y)*17, x == z);
check(x*0 == y*0, t);
check(x == x+y, y == 0);
check(x+y == x, y == 0);
check(100 - x == 99 - y, (y-x) == -1);
check(x < x, f);
check(x < (x+1), t);
check(x-2 < y+3, x < y+5);
check(x+y < y+z, x < z);
check(y+x < y+z, x < z);
check(x+y < z+y, x < z);
check(y+x < z+y, x < z);
check((y+x)*17 < (z+y)*17, x < z);
check(x*0 < y*0, f);
check(x < x+y, 0 < y);
check(x+y < x, y < 0);
check(select(x < 3, 2, 2), 2);
check(select(x < (x+1), 9, 2), 9);
check(select(x > (x+1), 9, 2), 2);
// Selects of comparisons should always become selects of LT or selects of EQ
check(select(x != 5, 2, 3), select(x == 5, 3, 2));
check(select(x >= 5, 2, 3), select(x < 5, 3, 2));
check(select(x <= 5, 2, 3), select(5 < x, 3, 2));
check(select(x > 5, 2, 3), select(5 < x, 2, 3));
check(select(x > 5, 2, 3) + select(x > 5, 6, 2), select(5 < x, 8, 5));
check(select(x > 5, 8, 3) - select(x > 5, 6, 2), select(5 < x, 2, 1));
check(select(x < 5, select(x < 5, 0, 1), 2), select(x < 5, 0, 2));
check(select(x < 5, 0, select(x < 5, 1, 2)), select(x < 5, 0, 2));
check((1 - xf)*6 < 3, 0.5f < xf);
check(!f, t);
check(!t, f);
check(!(x < y), y <= x);
check(!(x > y), x <= y);
check(!(x >= y), x < y);
check(!(x <= y), y < x);
check(!(x == y), x != y);
check(!(x != y), x == y);
check(!(!(x == 0)), x == 0);
check(!Expr(broadcast(x > y, 4)),
broadcast(x <= y, 4));
check(b1 || !b1, t);
check(!b1 || b1, t);
check(b1 && !b1, f);
check(!b1 && b1, f);
check(b1 && b1, b1);
check(b1 || b1, b1);
check(broadcast(b1, 4) || broadcast(!b1, 4), broadcast(t, 4));
check(broadcast(!b1, 4) || broadcast(b1, 4), broadcast(t, 4));
check(broadcast(b1, 4) && broadcast(!b1, 4), broadcast(f, 4));
check(broadcast(!b1, 4) && broadcast(b1, 4), broadcast(f, 4));
check(broadcast(b1, 4) && broadcast(b1, 4), broadcast(b1, 4));
check(broadcast(b1, 4) || broadcast(b1, 4), broadcast(b1, 4));
check((x == 1) && (x != 2), (x == 1));
check((x != 1) && (x == 2), (x == 2));
check((x == 1) && (x != 1), f);
check((x != 1) && (x == 1), f);
check((x == 1) || (x != 2), (x != 2));
check((x != 1) || (x == 2), (x != 1));
check((x == 1) || (x != 1), t);
check((x != 1) || (x == 1), t);
check(x < 20 || x > 19, t);
check(x > 19 || x < 20, t);
check(x < 20 || x > 20, x < 20 || 20 < x);
check(x > 20 || x < 20, 20 < x || x < 20);
check(x < 20 && x > 19, f);
check(x > 19 && x < 20, f);
check(x < 20 && x > 18, x < 20 && 18 < x);
check(x > 18 && x < 20, 18 < x && x < 20);
check(x <= 20 || x > 19, t);
check(x > 19 || x <= 20, t);
check(x <= 18 || x > 20, x <= 18 || 20 < x);
check(x > 20 || x <= 18, 20 < x || x <= 18);
check(x <= 18 && x > 19, f);
check(x > 19 && x <= 18, f);
check(x <= 20 && x > 19, x <= 20 && 19 < x);
check(x > 19 && x <= 20, 19 < x && x <= 20);
check(x < 20 || x >= 19, t);
check(x >= 19 || x < 20, t);
check(x < 18 || x >= 20, x < 18 || 20 <= x);
check(x >= 20 || x < 18, 20 <= x || x < 18);
check(x < 18 && x >= 19, f);
check(x >= 19 && x < 18, f);
check(x < 20 && x >= 19, x < 20 && 19 <= x);
check(x >= 19 && x < 20, 19 <= x && x < 20);
check(x <= 20 || x >= 21, t);
check(x >= 21 || x <= 20, t);
check(x <= 18 || x >= 20, x <= 18 || 20 <= x);
check(x >= 20 || x <= 18, 20 <= x || x <= 18);
check(x <= 18 && x >= 19, f);
check(x >= 19 && x <= 18, f);
check(x <= 20 && x >= 20, x <= 20 && 20 <= x);
check(x >= 20 && x <= 20, 20 <= x && x <= 20);
// check for substitution patterns
check((b1 == t) && (b1 && b2), (b1 == t) && b2);
check((b1 && b2) && (b1 == t), b2 && (b1 == t));
{
Expr i = Variable::make(Int(32), "i");
check((i!=2 && (i!=4 && (i!=8 && i!=16))) || (i==16), (i!=2 && (i!=4 && (i!=8))));
check((i==16) || (i!=2 && (i!=4 && (i!=8 && i!=16))), (i!=2 && (i!=4 && (i!=8))));
}
check(t && (x < 0), x < 0);
check(f && (x < 0), f);
check(t || (x < 0), t);
check(f || (x < 0), x < 0);
check(x == y || y != x, t);
check(x == y || x != y, t);
check(x == y && x != y, f);
check(x == y && y != x, f);
check(x < y || x >= y, t);
check(x <= y || x > y, t);
check(x < y && x >= y, f);
check(x <= y && x > y, f);
check(x <= max(x, y), t);
check(x < min(x, y), f);
check(min(x, y) <= x, t);
check(max(x, y) < x, f);
check(max(x, y) <= y, x <= y);
check(min(x, y) >= y, y <= x);
check((1 < y) && (2 < y), 2 < y);
check(x*5 < 4, x < 1);
check(x*5 < 5, x < 1);
check(x*5 < 6, x < 2);
check(x*5 <= 4, x <= 0);
check(x*5 <= 5, x <= 1);
check(x*5 <= 6, x <= 1);
check(x*5 > 4, 0 < x);
check(x*5 > 5, 1 < x);
check(x*5 > 6, 1 < x);
check(x*5 >= 4, 1 <= x);
check(x*5 >= 5, 1 <= x);
check(x*5 >= 6, 2 <= x);
check(x/4 < 3, x < 12);
check(3 < x/4, 15 < x);
check(4 - x <= 0, 4 <= x);
check((x/8)*8 < x - 8, f);
check((x/8)*8 < x - 9, f);
check((x/8)*8 < x - 7, f);
check((x/8)*8 < x - 6, 6 < x % 8);
check(ramp(x*4, 1, 4) < broadcast(y*4, 4), broadcast(x < y, 4));
check(ramp(x*8, 1, 4) < broadcast(y*8, 4), broadcast(x < y, 4));
check(ramp(x*8 + 1, 1, 4) < broadcast(y*8, 4), broadcast(x < y, 4));
check(ramp(x*8 + 4, 1, 4) < broadcast(y*8, 4), broadcast(x < y, 4));
check(ramp(x*8 + 8, 1, 4) < broadcast(y*8, 4), broadcast(x < y + (-1), 4));
check(ramp(x*8 + 5, 1, 4) < broadcast(y*8, 4), ramp(x*8 + 5, 1, 4) < broadcast(y*8, 4));
check(ramp(x*8 - 1, 1, 4) < broadcast(y*8, 4), ramp(x*8 + (-1), 1, 4) < broadcast(y*8, 4));
check(ramp(x*8, 1, 4) < broadcast(y*4, 4), broadcast(x*2 < y, 4));
check(ramp(x*8, 2, 4) < broadcast(y*8, 4), broadcast(x < y, 4));
check(ramp(x*8 + 1, 2, 4) < broadcast(y*8, 4), broadcast(x < y, 4));
check(ramp(x*8 + 2, 2, 4) < broadcast(y*8, 4), ramp(x*8 + 2, 2, 4) < broadcast(y*8, 4));
check(ramp(x*8, 3, 4) < broadcast(y*8, 4), ramp(x*8, 3, 4) < broadcast(y*8, 4));
check(select(ramp((x/16)*16, 1, 8) < broadcast((y/8)*8, 8), broadcast(1, 8), broadcast(3, 8)),
select((x/16)*2 < y/8, broadcast(1, 8), broadcast(3, 8)));
check(ramp(x*8, -1, 4) < broadcast(y*8, 4), ramp(x*8, -1, 4) < broadcast(y*8, 4));
check(ramp(x*8 + 1, -1, 4) < broadcast(y*8, 4), ramp(x*8 + 1, -1, 4) < broadcast(y*8, 4));
check(ramp(x*8 + 4, -1, 4) < broadcast(y*8, 4), broadcast(x < y, 4));
check(ramp(x*8 + 8, -1, 4) < broadcast(y*8, 4), ramp(x*8 + 8, -1, 4) < broadcast(y*8, 4));
check(ramp(x*8 + 5, -1, 4) < broadcast(y*8, 4), broadcast(x < y, 4));
check(ramp(x*8 - 1, -1, 4) < broadcast(y*8, 4), broadcast(x < y + 1, 4));
// Check anded conditions apply to the then case only
check(IfThenElse::make(x == 4 && y == 5,
Evaluate::make(z + x + y),
Evaluate::make(z + x - y)),
IfThenElse::make(x == 4 && y == 5,
Evaluate::make(z + 9),
Evaluate::make(z + x - y)));
// Check ored conditions apply to the else case only
check(IfThenElse::make(b1 || b2,
Evaluate::make(select(b1, x+3, y+4) + select(b2, x+5, y+7)),
Evaluate::make(select(b1, x+3, y+8) - select(b2, x+5, y+7))),
IfThenElse::make(b1 || b2,
Evaluate::make(select(b1, x+3, y+4) + select(b2, x+5, y+7)),
Evaluate::make(1)));
// Check single conditions apply to both cases of an ifthenelse
check(IfThenElse::make(b1,
Evaluate::make(select(b1, x, y)),
Evaluate::make(select(b1, z, w))),
IfThenElse::make(b1,
Evaluate::make(x),
Evaluate::make(w)));
check(IfThenElse::make(x < y,
IfThenElse::make(x < y, Evaluate::make(y), Evaluate::make(x)),
Evaluate::make(x)),
IfThenElse::make(x < y,
Evaluate::make(y),
Evaluate::make(x)));
check(Block::make(IfThenElse::make(x < y, Evaluate::make(x+1), Evaluate::make(x+2)),
IfThenElse::make(x < y, Evaluate::make(x+3), Evaluate::make(x+4))),
IfThenElse::make(x < y,
Block::make(Evaluate::make(x+1), Evaluate::make(x+3)),
Block::make(Evaluate::make(x+2), Evaluate::make(x+4))));
check(Block::make(IfThenElse::make(x < y, Evaluate::make(x+1)),
IfThenElse::make(x < y, Evaluate::make(x+2))),
IfThenElse::make(x < y, Block::make(Evaluate::make(x+1), Evaluate::make(x+2))));
check(Block::make(IfThenElse::make(x < y, Evaluate::make(x+1), Evaluate::make(x+2)),
IfThenElse::make(x < y, Evaluate::make(x+3))),
IfThenElse::make(x < y,
Block::make(Evaluate::make(x+1), Evaluate::make(x+3)),
Evaluate::make(x+2)));
check(Block::make(IfThenElse::make(x < y, Evaluate::make(x+1)),
IfThenElse::make(x < y, Evaluate::make(x+2), Evaluate::make(x+3))),
IfThenElse::make(x < y,
Block::make(Evaluate::make(x+1), Evaluate::make(x+2)),
Evaluate::make(x+3)));
// Check conditions involving entire exprs
Expr foo = x + 3*y;
Expr foo_simple = x + y*3;
check(IfThenElse::make(foo == 17,
Evaluate::make(x+foo+1),
Evaluate::make(x+foo+2)),
IfThenElse::make(foo_simple == 17,
Evaluate::make(x+18),
Evaluate::make(x+foo_simple+2)));
check(IfThenElse::make(foo != 17,
Evaluate::make(x+foo+1),
Evaluate::make(x+foo+2)),
IfThenElse::make(foo_simple != 17,
Evaluate::make(x+foo_simple+1),
Evaluate::make(x+19)));
// The construct
// if (var == expr) then a else b;
// was being simplified incorrectly, but *only* if var was of type Bool.
Stmt then_clause = AssertStmt::make(b2, Expr(22));
Stmt else_clause = AssertStmt::make(b2, Expr(33));
check(IfThenElse::make(b1 == b2, then_clause, else_clause),
IfThenElse::make(b1 == b2, then_clause, else_clause));
// Simplifications of selects
check(select(x == 3, 5, 7) + 7, select(x == 3, 12, 14));
check(select(x == 3, 5, 7) - 7, select(x == 3, -2, 0));
check(select(x == 3, 5, y) - y, select(x == 3, 5 - y, 0));
check(select(x == 3, y, 5) - y, select(x == 3, 0, 5 - y));
check(y - select(x == 3, 5, y), select(x == 3, y + (-5), 0));
check(y - select(x == 3, y, 5), select(x == 3, 0, y + (-5)));
check(select(x == 3, 5, 7) == 7, x != 3);
check(select(x == 3, z, y) == z, (x == 3) || (y == z));
check(select(x == 3, 4, 2) == 0, const_false());
check(select(x == 3, y, 2) == 4, (x == 3) && (y == 4));
check(select(x == 3, 2, y) == 4, (x != 3) && (y == 4));
check(min(select(x == 2, y*3, 8), select(x == 2, y+8, y*7)),
select(x == 2, min(y*3, y+8), min(y*7, 8)));
check(max(select(x == 2, y*3, 8), select(x == 2, y+8, y*7)),
select(x == 2, max(y*3, y+8), max(y*7, 8)));
check(select(x == 2, x+1, x+5), x + select(x == 2, 1, 5));
check(select(x == 2, x+y, x+z), x + select(x == 2, y, z));
check(select(x == 2, y+x, x+z), x + select(x == 2, y, z));
check(select(x == 2, y+x, z+x), select(x == 2, y, z) + x);
check(select(x == 2, x+y, z+x), x + select(x == 2, y, z));
check(select(x == 2, x*2, x*5), x * select(x == 2, 2, 5));
check(select(x == 2, x*y, x*z), x * select(x == 2, y, z));
check(select(x == 2, y*x, x*z), x * select(x == 2, y, z));
check(select(x == 2, y*x, z*x), select(x == 2, y, z) * x);
check(select(x == 2, x*y, z*x), x * select(x == 2, y, z));
check(select(x == 2, x-y, x-z), x - select(x == 2, y, z));
check(select(x == 2, y-x, z-x), select(x == 2, y, z) - x);
check(select(x == 2, x+y, x-z), x + select(x == 2, y, 0-z));
check(select(x == 2, y+x, x-z), x + select(x == 2, y, 0-z));
check(select(x == 2, x-z, x+y), x + select(x == 2, 0-z, y));
check(select(x == 2, x-z, y+x), x + select(x == 2, 0-z, y));
{
Expr b[12];
for (int i = 0; i < 12; i++) {
b[i] = Variable::make(Bool(), unique_name('b'));
}
// Some rules that collapse selects
check(select(b[0], x, select(b[1], x, y)),
select(b[0] || b[1], x, y));
check(select(b[0], x, select(b[1], y, x)),
select(b[0] || !b[1], x, y));
check(select(b[0], select(b[1], x, y), x),
select(b[0] && !b[1], y, x));
check(select(b[0], select(b[1], y, x), x),
select(b[0] && b[1], y, x));
// Ternary boolean expressions in two variables
check(b[0] || (b[0] && b[1]), b[0]);
check((b[0] && b[1]) || b[0], b[0]);
check(b[0] && (b[0] || b[1]), b[0]);
check((b[0] || b[1]) && b[0], b[0]);
check(b[0] && (b[0] && b[1]), b[0] && b[1]);
check((b[0] && b[1]) && b[0], b[1] && b[0]);
check(b[0] || (b[0] || b[1]), b[0] || b[1]);
check((b[0] || b[1]) || b[0], b[1] || b[0]);
// A nasty unsimplified boolean Expr seen in the wild
Expr nasty = ((((((((((((((((((((((((((((((((((((((((((((b[0] && b[1]) || (b[2] && b[1])) || b[0]) || b[2]) || b[0]) || b[2]) && ((b[0] && b[6]) || (b[2] && b[6]))) || b[0]) || b[2]) || b[0]) || b[2]) && ((b[0] && b[3]) || (b[2] && b[3]))) || b[0]) || b[2]) || b[0]) || b[2]) && ((b[0] && b[7]) || (b[2] && b[7]))) || b[0]) || b[2]) || b[0]) || b[2]) && ((b[0] && b[4]) || (b[2] && b[4]))) || b[0]) || b[2]) || b[0]) || b[2]) && ((b[0] && b[8]) || (b[2] && b[8]))) || b[0]) || b[2]) || b[0]) || b[2]) && ((b[0] && b[5]) || (b[2] && b[5]))) || b[0]) || b[2]) || b[0]) || b[2]) && ((b[0] && b[10]) || (b[2] && b[10]))) || b[0]) || b[2]) || b[0]) || b[2]) && ((b[0] && b[9]) || (b[2] && b[9]))) || b[0]) || b[2]);
check(nasty, b[0] || b[2]);
}
{
// verify that likely(const-bool) is *not* simplified.
check(likely(t), likely(t));
check(likely(f), likely(f));
// verify that !likely(e) -> likely(!e)
check(!likely(t), likely(f));
check(!likely(f), likely(t));
check(!likely(x == 2), likely(x != 2));
// can_prove(likely(const-true)) = true
// can_prove(!likely(const-false)) = true
internal_assert(can_prove(likely(t)));
internal_assert(can_prove(!likely(f)));
// unprovable cases
internal_assert(!can_prove(likely(f)));
internal_assert(!can_prove(!likely(t)));
internal_assert(!can_prove(!likely(x == 2)));
}
}
void check_math() {
Var x = Var("x");
check(sqrt(4.0f), 2.0f);
check(log(0.5f + 0.5f), 0.0f);
check(exp(log(2.0f)), 2.0f);
check(pow(4.0f, 0.5f), 2.0f);
check(round(1000.0f*pow(exp(1.0f), log(10.0f))), 10000.0f);
check(floor(0.98f), 0.0f);
check(ceil(0.98f), 1.0f);
check(round(0.6f), 1.0f);
check(round(-0.5f), 0.0f);
check(trunc(-1.6f), -1.0f);
check(floor(round(x)), round(x));
check(ceil(ceil(x)), ceil(x));
}
void check_overflow() {
Expr overflowing[] = {
make_const(Int(32), 0x7fffffff) + 1,
make_const(Int(32), 0x7ffffff0) + 16,
(make_const(Int(32), 0x7fffffff) +
make_const(Int(32), 0x7fffffff)),
make_const(Int(32), 0x08000000) * 16,
(make_const(Int(32), 0x00ffffff) *
make_const(Int(32), 0x00ffffff)),
make_const(Int(32), 0x80000000) - 1,
0 - make_const(Int(32), 0x80000000),
make_const(Int(64), (int64_t)0x7fffffffffffffffLL) + 1,
make_const(Int(64), (int64_t)0x7ffffffffffffff0LL) + 16,
(make_const(Int(64), (int64_t)0x7fffffffffffffffLL) +
make_const(Int(64), (int64_t)0x7fffffffffffffffLL)),
make_const(Int(64), (int64_t)0x0800000000000000LL) * 16,
(make_const(Int(64), (int64_t)0x00ffffffffffffffLL) *
make_const(Int(64), (int64_t)0x00ffffffffffffffLL)),
make_const(Int(64), (int64_t)0x8000000000000000LL) - 1,
0 - make_const(Int(64), (int64_t)0x8000000000000000LL),
};
Expr not_overflowing[] = {
make_const(Int(32), 0x7ffffffe) + 1,
make_const(Int(32), 0x7fffffef) + 16,
make_const(Int(32), 0x07ffffff) * 2,
(make_const(Int(32), 0x0000ffff) *
make_const(Int(32), 0x00008000)),
make_const(Int(32), 0x80000001) - 1,
0 - make_const(Int(32), 0x7fffffff),
make_const(Int(64), (int64_t)0x7ffffffffffffffeLL) + 1,
make_const(Int(64), (int64_t)0x7fffffffffffffefLL) + 16,
make_const(Int(64), (int64_t)0x07ffffffffffffffLL) * 16,
(make_const(Int(64), (int64_t)0x00000000ffffffffLL) *
make_const(Int(64), (int64_t)0x0000000080000000LL)),
make_const(Int(64), (int64_t)0x8000000000000001LL) - 1,
0 - make_const(Int(64), (int64_t)0x7fffffffffffffffLL),
};
for (Expr e : overflowing) {
internal_assert(!is_const(simplify(e)))
<< "Overflowing expression should not have simplified: " << e << "\n";
}
for (Expr e : not_overflowing) {
internal_assert(is_const(simplify(e)))
<< "Non-everflowing expression should have simplified: " << e << "\n";
}
}
void check_ind_expr(Expr e, bool expect_error) {
Expr e2 = simplify(e);
const Call *call = e2.as<Call>();
bool is_error = call && call->is_intrinsic(Call::indeterminate_expression);
if (expect_error && !is_error)
internal_error << "Expression should be indeterminate: " << e << " but saw: " << e2 << "\n";
else if (!expect_error && is_error)
internal_error << "Expression should not be indeterminate: " << e << " but saw: " << e2 << "\n";
}
void check_indeterminate_ops(Expr e, bool e_is_zero, bool e_is_indeterminate) {
Expr b = cast<bool>(e);
Expr t = const_true(), f = const_false();
Expr one = cast(e.type(), 1);
Expr zero = cast(e.type(), 0);
check_ind_expr(e, e_is_indeterminate);
check_ind_expr(e + e, e_is_indeterminate);
check_ind_expr(e - e, e_is_indeterminate);
check_ind_expr(e * e, e_is_indeterminate);
check_ind_expr(e / e, e_is_zero || e_is_indeterminate);
check_ind_expr((1 / e) / e, e_is_zero || e_is_indeterminate);
// Expr::operator% asserts if denom is constant zero.
if (!is_zero(e)) {
check_ind_expr(e % e, e_is_zero || e_is_indeterminate);
check_ind_expr((1 / e) % e, e_is_zero || e_is_indeterminate);
}
check_ind_expr(min(e, one), e_is_indeterminate);
check_ind_expr(max(e, one), e_is_indeterminate);
check_ind_expr(e == one, e_is_indeterminate);
check_ind_expr(one == e, e_is_indeterminate);
check_ind_expr(e < one, e_is_indeterminate);
check_ind_expr(one < e, e_is_indeterminate);
check_ind_expr(!(e == one), e_is_indeterminate);
check_ind_expr(!(one == e), e_is_indeterminate);
check_ind_expr(!(e < one), e_is_indeterminate);
check_ind_expr(!(one < e), e_is_indeterminate);
check_ind_expr(b && t, e_is_indeterminate);
check_ind_expr(t && b, e_is_indeterminate);
check_ind_expr(b || t, e_is_indeterminate);
check_ind_expr(t || b, e_is_indeterminate);
check_ind_expr(!b, e_is_indeterminate);
check_ind_expr(select(b, one, zero), e_is_indeterminate);
check_ind_expr(select(t, e, zero), e_is_indeterminate);
check_ind_expr(select(f, zero, e), e_is_indeterminate);
check_ind_expr(e << one, e_is_indeterminate);
check_ind_expr(e >> one, e_is_indeterminate);
// Avoid warnings for things like (1 << 2147483647)
if (e_is_indeterminate) {
check_ind_expr(one << e, e_is_indeterminate);
check_ind_expr(one >> e, e_is_indeterminate);
}
check_ind_expr(one & e, e_is_indeterminate);
check_ind_expr(e & one, e_is_indeterminate);
check_ind_expr(one | e, e_is_indeterminate);
check_ind_expr(e | one, e_is_indeterminate);
if (!e.type().is_uint()) {
// Avoid warnings
check_ind_expr(abs(e), e_is_indeterminate);
}
check_ind_expr(log(e), e_is_indeterminate);
check_ind_expr(sqrt(e), e_is_indeterminate);
check_ind_expr(exp(e), e_is_indeterminate);
check_ind_expr(pow(e, one), e_is_indeterminate);
// pow(x, y) explodes for huge integer y (Issue #1441)
if (e_is_indeterminate) {
check_ind_expr(pow(one, e), e_is_indeterminate);
}
check_ind_expr(floor(e), e_is_indeterminate);
check_ind_expr(ceil(e), e_is_indeterminate);
check_ind_expr(round(e), e_is_indeterminate);
check_ind_expr(trunc(e), e_is_indeterminate);
}
void check_indeterminate() {
const int32_t values[] = {
int32_t(0x80000000),
-2147483647,
-2,
-1,
0,
1,
2,
2147483647,
};
for (int32_t i1 : values) {
// reality-check for never-indeterminate values.
check_indeterminate_ops(Expr(i1), !i1, false);
for (int32_t i2 : values) {
{
Expr e1(i1), e2(i2);
Expr r = (e1 / e2);
bool r_is_zero = !i1 || (i2 != 0 && !div_imp((int64_t)i1, (int64_t)i2)); // avoid trap for -2147483648/-1
bool r_is_ind = !i2;
check_indeterminate_ops(r, r_is_zero, r_is_ind);
// Expr::operator% asserts if denom is constant zero.
if (!is_zero(e2)) {
Expr m = (e1 % e2);
bool m_is_zero = !i1 || (i2 != 0 && !mod_imp((int64_t)i1, (int64_t)i2)); // avoid trap for -2147483648/-1
bool m_is_ind = !i2;
check_indeterminate_ops(m, m_is_zero, m_is_ind);
}
}
{
uint32_t u1 = (uint32_t)i1;
uint32_t u2 = (uint32_t)i2;
Expr e1(u1), e2(u2);
Expr r = (e1 / e2);
bool r_is_zero = !u1 || (u2 != 0 && !div_imp(u1, u2));
bool r_is_ind = !u2;
check_indeterminate_ops(r, r_is_zero, r_is_ind);
// Expr::operator% asserts if denom is constant zero.
if (!is_zero(e2)) {
Expr m = (e1 % e2);
bool m_is_zero = !u1 || (u2 != 0 && !mod_imp(u1, u2));
bool m_is_ind = !u2;
check_indeterminate_ops(m, m_is_zero, m_is_ind);
}
}
}
}
}
} // namespace
void simplify_test() {
Expr x = Var("x"), y = Var("y"), z = Var("z"), w = Var("w"), v = Var("v");
Expr xf = cast<float>(x);
Expr yf = cast<float>(y);
Expr t = const_true(), f = const_false();
check_indeterminate();
check_casts();
check_algebra();
check_vectors();
check_bounds();
check_math();
check_boolean();
check_overflow();
// Check bitshift operations
check(cast(Int(16), x) << 10, cast(Int(16), x) * 1024);
check(cast(Int(16), x) >> 10, cast(Int(16), x) / 1024);
check(cast(Int(16), x) << -10, cast(Int(16), x) / 1024);
// Correctly triggers a warning:
//check(cast(Int(16), x) << 20, cast(Int(16), x) << 20);
// Check bitwise_and. (Added as result of a bug.)
// TODO: more coverage of bitwise_and and bitwise_or.
check(cast(UInt(32), x) & Expr((uint32_t)0xaaaaaaaa),
cast(UInt(32), x) & Expr((uint32_t)0xaaaaaaaa));
// Check constant-folding of bitwise ops (and indirectly, reinterpret)
check(Let::make(x.as<Variable>()->name, 5, ((~x) & 3) | 16), (~5 & 3) | 16);
check(Let::make(x.as<Variable>()->name, 5, ((~cast<uint8_t>(x)) & 3) | 16), make_const(UInt(8), (~5 & 3) | 16));
// Check that chains of widening casts don't lose the distinction
// between zero-extending and sign-extending.
check(cast(UInt(64), cast(UInt(32), cast(Int(8), -1))),
UIntImm::make(UInt(64), 0xffffffffULL));
v = Variable::make(Int(32, 4), "v");
// Check constants get pushed inwards
check(Let::make("x", 3, x+4), 7);
// Check ramps in lets get pushed inwards
check(Let::make("v", ramp(x*2+7, 3, 4), v + Expr(broadcast(2, 4))),
ramp(x*2+9, 3, 4));
// Check broadcasts in lets get pushed inwards
check(Let::make("v", broadcast(x, 4), v + Expr(broadcast(2, 4))),
broadcast(x+2, 4));
// Check that dead lets get stripped
check(Let::make("x", 3*y*y*y, 4), 4);
check(Let::make("x", 0, 0), 0);
// Check that lets inside an evaluate node get lifted
check(Evaluate::make(Let::make("x", Call::make(Int(32), "dummy", {3, x, 4}, Call::Extern), Let::make("y", 10, x + y + 2))),
LetStmt::make("x", Call::make(Int(32), "dummy", {3, x, 4}, Call::Extern), Evaluate::make(x + 12)));
// Test case with most negative 32-bit number, as constant to check that it is not negated.
check(((x * (int32_t)0x80000000) + (y + z * (int32_t)0x80000000)),
((x * (int32_t)0x80000000) + (y + z * (int32_t)0x80000000)));
// Check that constant args to a stringify get combined
check(Call::make(type_of<const char *>(), Call::stringify, {3, string(" "), 4}, Call::Intrinsic),
string("3 4"));
check(Call::make(type_of<const char *>(), Call::stringify, {3, x, 4, string(", "), 3.4f}, Call::Intrinsic),
Call::make(type_of<const char *>(), Call::stringify, {string("3"), x, string("4, 3.400000")}, Call::Intrinsic));
{
// Check that contiguous prefetch call get collapsed
Expr base = Variable::make(Handle(), "buf");
check(Call::make(Int(32), Call::prefetch, {base, x, 4, 1, 64, 4, min(x + y, 128), 256}, Call::Intrinsic),
Call::make(Int(32), Call::prefetch, {base, x, min(x + y, 128) * 256, 1}, Call::Intrinsic));
}
// Check min(x, y)*max(x, y) gets simplified into x*y
check(min(x, y)*max(x, y), x*y);
check(min(x, y)*max(y, x), x*y);
check(max(x, y)*min(x, y), x*y);
check(max(y, x)*min(x, y), x*y);
// Check min(x, y) + max(x, y) gets simplified into x + y
check(min(x, y) + max(x, y), x + y);
check(min(x, y) + max(y, x), x + y);
check(max(x, y) + min(x, y), x + y);
check(max(y, x) + min(x, y), x + y);
// Check max(min(x, y), max(x, y)) gets simplified into max(x, y)
check(max(min(x, y), max(x, y)), max(x, y));
check(max(min(x, y), max(y, x)), max(x, y));
check(max(max(x, y), min(x, y)), max(x, y));
check(max(max(y, x), min(x, y)), max(x, y));
// Check min(max(x, y), min(x, y)) gets simplified into min(x, y)
check(min(max(x, y), min(x, y)), min(x, y));
check(min(max(x, y), min(y, x)), min(x, y));
check(min(min(x, y), max(x, y)), min(x, y));
check(min(min(y, x), max(x, y)), min(x, y));
// Check if we can simplify away comparison on vector types considering bounds.
Scope<Interval> bounds_info;
bounds_info.push("x", Interval(0,4));
check_in_bounds(ramp(x, 1, 4) < broadcast( 0, 4), const_false(4), bounds_info);
check_in_bounds(ramp(x, 1, 4) < broadcast( 8, 4), const_true(4), bounds_info);
check_in_bounds(ramp(x, -1, 4) < broadcast(-4, 4), const_false(4), bounds_info);
check_in_bounds(ramp(x, -1, 4) < broadcast( 5, 4), const_true(4), bounds_info);
check_in_bounds(min(ramp(x, 1, 4), broadcast( 0, 4)), broadcast(0, 4), bounds_info);
check_in_bounds(min(ramp(x, 1, 4), broadcast( 8, 4)), ramp(x, 1, 4), bounds_info);
check_in_bounds(min(ramp(x, -1, 4), broadcast(-4, 4)), broadcast(-4, 4), bounds_info);
check_in_bounds(min(ramp(x, -1, 4), broadcast( 5, 4)), ramp(x, -1, 4), bounds_info);
check_in_bounds(max(ramp(x, 1, 4), broadcast( 0, 4)), ramp(x, 1, 4), bounds_info);
check_in_bounds(max(ramp(x, 1, 4), broadcast( 8, 4)), broadcast(8, 4), bounds_info);
check_in_bounds(max(ramp(x, -1, 4), broadcast(-4, 4)), ramp(x, -1, 4), bounds_info);
check_in_bounds(max(ramp(x, -1, 4), broadcast( 5, 4)), broadcast(5, 4), bounds_info);
// Collapse some vector interleaves
check(interleave_vectors({ramp(x, 2, 4), ramp(x+1, 2, 4)}), ramp(x, 1, 8));
check(interleave_vectors({ramp(x, 4, 4), ramp(x+2, 4, 4)}), ramp(x, 2, 8));
check(interleave_vectors({ramp(x-y, 2*y, 4), ramp(x, 2*y, 4)}), ramp(x-y, y, 8));
check(interleave_vectors({ramp(x, 3, 4), ramp(x+1, 3, 4), ramp(x+2, 3, 4)}), ramp(x, 1, 12));
{
Expr vec = ramp(x, 1, 16);
check(interleave_vectors({slice(vec, 0, 2, 8), slice(vec, 1, 2, 8)}), vec);
check(interleave_vectors({slice(vec, 0, 4, 4), slice(vec, 1, 4, 4), slice(vec, 2, 4, 4), slice(vec, 3, 4, 4)}), vec);
}
// Collapse some vector concats
check(concat_vectors({ramp(x, 2, 4), ramp(x+8, 2, 4)}), ramp(x, 2, 8));
check(concat_vectors({ramp(x, 3, 2), ramp(x+6, 3, 2), ramp(x+12, 3, 2)}), ramp(x, 3, 6));
// Now some ones that can't work
{
Expr e = interleave_vectors({ramp(x, 2, 4), ramp(x, 2, 4)});
check(e, e);
e = interleave_vectors({ramp(x, 2, 4), ramp(x+2, 2, 4)});
check(e, e);
e = interleave_vectors({ramp(x, 3, 4), ramp(x+1, 3, 4)});
check(e, e);
e = interleave_vectors({ramp(x, 2, 4), ramp(y+1, 2, 4)});
check(e, e);
e = interleave_vectors({ramp(x, 2, 4), ramp(x+1, 3, 4)});
check(e, e);
e = concat_vectors({ramp(x, 1, 4), ramp(x+4, 2, 4)});
check(e, e);
e = concat_vectors({ramp(x, 1, 4), ramp(x+8, 1, 4)});
check(e, e);
e = concat_vectors({ramp(x, 1, 4), ramp(y+4, 1, 4)});
check(e, e);
}
// Now check that an interleave of some collapsible loads collapses into a single dense load
{
Expr load1 = Load::make(Float(32, 4), "buf", ramp(x, 2, 4), Buffer<>(), Parameter(), const_true(4));
Expr load2 = Load::make(Float(32, 4), "buf", ramp(x+1, 2, 4), Buffer<>(), Parameter(), const_true(4));
Expr load12 = Load::make(Float(32, 8), "buf", ramp(x, 1, 8), Buffer<>(), Parameter(), const_true(8));
check(interleave_vectors({load1, load2}), load12);
// They don't collapse in the other order
Expr e = interleave_vectors({load2, load1});
check(e, e);
// Or if the buffers are different
Expr load3 = Load::make(Float(32, 4), "buf2", ramp(x+1, 2, 4), Buffer<>(), Parameter(), const_true(4));
e = interleave_vectors({load1, load3});
check(e, e);
}
// Check that concatenated loads of adjacent scalars collapse into a vector load.
{
int lanes = 4;
std::vector<Expr> loads;
for (int i = 0; i < lanes; i++) {
loads.push_back(Load::make(Float(32), "buf", x+i, Buffer<>(), Parameter(), const_true()));
}
check(concat_vectors(loads), Load::make(Float(32, lanes), "buf", ramp(x, 1, lanes), Buffer<>(), Parameter(), const_true(lanes)));
}
// This expression doesn't simplify, but it did cause exponential
// slowdown at one stage.
{
Expr e = x;
for (int i = 0; i < 100; i++) {
e = max(e, 1)/2;
}
check(e, e);
}
// These expressions are used to cause infinite recursion.
check(Broadcast::make(-16, 2) < (ramp(Cast::make(UInt(16), 7), Cast::make(UInt(16), 11), 2) - Broadcast::make(1, 2)),
Broadcast::make(-16, 2) < (ramp(make_const(UInt(16), 7), make_const(UInt(16), 11), 2) - Broadcast::make(1, 2)));
check((ramp(-71, 39, 2)/Cast::make(Int(32).with_lanes(2), ramp(Expr((uint16_t)1), Expr((uint16_t)1), 2))) >= Broadcast::make(23, 2),
(Cast::make(Int(32).with_lanes(2), ramp(Expr((uint16_t)1), Expr((uint16_t)1), 2)) * Broadcast::make(23, 2)) <= ramp(-71, 39, 2));
{
Expr pred = ramp(x*y + x*z, 2, 8) > 2;
Expr index = ramp(x + y, 1, 8);
Expr value = Load::make(index.type(), "f", index, Buffer<>(), Parameter(), const_true(index.type().lanes()));
Stmt stmt = Store::make("f", value, index, Parameter(), pred);
check(stmt, Evaluate::make(0));
}
{
// Verify that integer types passed to min() and max() are coerced to match
// Exprs, rather than being promoted to int first. (TODO: This doesn't really
// belong in the test for Simplify, but IROperator has no test unit of its own.)
Expr one = cast<uint16_t>(1);
const int two = 2; // note that type is int, not uint16_t
Expr r1, r2, r3;
r1 = min(one, two);
internal_assert(r1.type() == halide_type_of<uint16_t>());
r2 = min(one, two, one);
internal_assert(r2.type() == halide_type_of<uint16_t>());
// Explicitly passing 'two' as an Expr, rather than an int, will defeat this logic.
r3 = min(one, Expr(two), one);
internal_assert(r3.type() == halide_type_of<int>());
r1 = max(one, two);
internal_assert(r1.type() == halide_type_of<uint16_t>());
r2 = max(one, two, one);
internal_assert(r2.type() == halide_type_of<uint16_t>());
// Explicitly passing 'two' as an Expr, rather than an int, will defeat this logic.
r3 = max(one, Expr(two), one);
internal_assert(r3.type() == halide_type_of<int>());
}
{
Expr x = Variable::make(UInt(32), "x");
Expr y = Variable::make(UInt(32), "y");
// This is used to get simplified into broadcast(x - y, 2) which is
// incorrect when there is overflow.
Expr e = simplify(max(ramp(x, y, 2), broadcast(x, 2)) - max(broadcast(y, 2), ramp(y, y, 2)));
Expr expected = max(ramp(x, y, 2), broadcast(x, 2)) - max(ramp(y, y, 2), broadcast(y, 2));
check(e, expected);
}
check(min(x, 63) - min(x, 3), clamp(x, 3, 63) + (-3));
check(min(x, 3) - min(x, 63), 3 - clamp(x, 3, 63));
check(min(63, x) - min(x, 3), clamp(x, 3, 63) + (-3));
check(min(x, 3) - min(63, x), 3 - clamp(x, 3, 63));
// This used to throw the simplifier into a loop
simplify((min((min(((x*64) + y), (z + -63)) + 31), min((((x*64) + y) + 63), z)) -
min((min((((x*64) + y) + 63), z) + -31), (min(((x*64) + y), (z + -63)) + 32))));
check(min(x * 4 + 63, y) - min(x * 4, y - 3), clamp(y - x * 4 + (-63), -60, 0) + 63);
check(min(x * 4, y - 3) - min(x * 4 + 63, y), -3 - clamp(y - x * 4 + (-3), 0, 60));
check(min(y, x * 4 + 63) - min(x * 4, y - 3), 63 - clamp(x * 4 - y + 63, 0, 60));
check(min(x * 4, y - 3) - min(y, x * 4 + 63), -3 - clamp(y - x * 4 + (-3), 0, 60));
check(max(x, 63) - max(x, 3), 63 - clamp(x, 3, 63));
check(max(x, 3) - max(x, 63), clamp(x, 3, 63) + (-63));
check(max(63, x) - max(3, x), 63 - clamp(x, 3, 63));
check(max(3, x) - max(x, 63), clamp(x, 3, 63) + (-63));
check(max(x * 4 + 63, y) - max(x * 4, y - 3), 3 - clamp(y - x * 4 + (-63), -60, 0));
check(max(x * 4, y - 3) - max(x * 4 + 63, y), clamp(y - x * 4 + (-3), 0, 60) + (-63));
check(max(x * 4 + 63, y) - max(y - 3, x * 4), 3 - clamp(y - x * 4 + (-63), -60, 0));
check(max(y - 3, x * 4) - max(y, x * 4 + 63), -63 - clamp(x * 4 - y + 3, -60, 0));
// Check that provably-true require() expressions are simplified away
{
Expr result(42);
check(require(Expr(1) > Expr(0), result, "error"), result);
check(require(x == x, result, "error"), result);
}
std::cout << "Simplify test passed" << std::endl;
}
}
}