--- /dev/null
+// { dg-do compile }
+// { dg-options "-std=c++14 -O2 -ftemplate-depth=1000000" }
+
+template <class T, int Dim0, int Dim1, int Dim2> struct Tensor3;
+template <class A, class T, int Dim0, int Dim1, int Dim2, char i, char j,
+ char k>
+struct Tensor3_Expr;
+
+template <class T, int Dim0, int Dim1, int Dim2, int Dim3> struct Tensor4;
+template <class A, class T, int Dim0, int Dim1, int Dim2, int Dim3, char i,
+ char j, char k, char l>
+struct Tensor4_Expr;
+
+template <char i, int Dim> struct Index
+{};
+template <const int N> struct Number
+{
+ Number(){};
+ operator int() const { return N; }
+};
+
+template <class T, int Tensor_Dim0, int Tensor_Dim1, int Tensor_Dim2>
+struct Tensor3
+{
+ T data[Tensor_Dim0][Tensor_Dim1][Tensor_Dim2];
+
+ T operator()(const int N1, const int N2, const int N3) const
+ {
+ return data[N1][N2][N3];
+ }
+
+ template <char i, char j, char k, int Dim0, int Dim1, int Dim2>
+ Tensor3_Expr<const Tensor3<T, Tensor_Dim0, Tensor_Dim1, Tensor_Dim2>, T,
+ Dim0, Dim1, Dim2, i, j, k>
+ operator()(const Index<i, Dim0>, const Index<j, Dim1>,
+ const Index<k, Dim2>) const
+ {
+ return Tensor3_Expr<const Tensor3<T, Tensor_Dim0, Tensor_Dim1, Tensor_Dim2>,
+ T, Dim0, Dim1, Dim2, i, j, k>(*this);
+ }
+};
+
+template <class A, class T, int Dim0, int Dim1, int Dim2, char i, char j,
+ char k>
+struct Tensor3_Expr
+{
+ A iter;
+
+ Tensor3_Expr(const A &a) : iter(a) {}
+ T operator()(const int N1, const int N2, const int N3) const
+ {
+ return iter(N1, N2, N3);
+ }
+};
+
+template <class A, class T, int Tensor_Dim0, int Tensor_Dim1, int Tensor_Dim2,
+ int Dim0, int Dim1, int Dim2, char i, char j, char k>
+struct Tensor3_Expr<Tensor3<A, Tensor_Dim0, Tensor_Dim1, Tensor_Dim2>, T, Dim0,
+ Dim1, Dim2, i, j, k>
+{
+ Tensor3<A, Tensor_Dim0, Tensor_Dim1, Tensor_Dim2> &iter;
+
+ Tensor3_Expr(Tensor3<A, Tensor_Dim0, Tensor_Dim1, Tensor_Dim2> &a) : iter(a)
+ {}
+ T operator()(const int N1, const int N2, const int N3) const
+ {
+ return iter(N1, N2, N3);
+ }
+};
+
+template <class A, class B, class T, class U, int Dim0, int Dim1, int Dim23,
+ int Dim4, int Dim5, char i, char j, char k, char l, char m>
+struct Tensor3_times_Tensor3_21
+{
+ Tensor3_Expr<A, T, Dim0, Dim1, Dim23, i, j, k> iterA;
+ Tensor3_Expr<B, U, Dim23, Dim4, Dim5, k, l, m> iterB;
+
+ template <int CurrentDim>
+ T eval(const int N1, const int N2, const int N3, const int N4,
+ const Number<CurrentDim> &) const
+ {
+ return iterA(N1, N2, CurrentDim - 1) * iterB(CurrentDim - 1, N3, N4)
+ + eval(N1, N2, N3, N4, Number<CurrentDim - 1>());
+ }
+ T eval(const int N1, const int N2, const int N3, const int N4,
+ const Number<1> &) const
+ {
+ return iterA(N1, N2, 0) * iterB(0, N3, N4);
+ }
+
+ Tensor3_times_Tensor3_21(
+ const Tensor3_Expr<A, T, Dim0, Dim1, Dim23, i, j, k> &a,
+ const Tensor3_Expr<B, U, Dim23, Dim4, Dim5, k, l, m> &b)
+ : iterA(a), iterB(b)
+ {}
+ T operator()(const int &N1, const int &N2, const int &N3,
+ const int &N4) const
+ {
+ return eval(N1, N2, N3, N4, Number<Dim23>());
+ }
+};
+
+template <class A, class B, class T, class U, int Dim0, int Dim1, int Dim23,
+ int Dim4, int Dim5, char i, char j, char k, char l, char m>
+Tensor4_Expr<Tensor3_times_Tensor3_21<A, B, T, U, Dim0, Dim1, Dim23, Dim4,
+ Dim5, i, j, k, l, m>,
+ T, Dim0, Dim1, Dim4, Dim5, i, j, l, m>
+operator*(const Tensor3_Expr<A, T, Dim0, Dim1, Dim23, i, j, k> &a,
+ const Tensor3_Expr<B, U, Dim23, Dim4, Dim5, k, l, m> &b)
+{
+ using TensorExpr = Tensor3_times_Tensor3_21<A, B, T, U, Dim0, Dim1, Dim23,
+ Dim4, Dim5, i, j, k, l, m>;
+ return Tensor4_Expr<TensorExpr, T, Dim0, Dim1, Dim4, Dim5, i, j, l, m>(
+ TensorExpr(a, b));
+};
+
+template <class T, int Tensor_Dim0, int Tensor_Dim1, int Tensor_Dim2,
+ int Tensor_Dim3>
+struct Tensor4
+{
+ T data[Tensor_Dim0][Tensor_Dim1][Tensor_Dim2][Tensor_Dim3];
+
+ Tensor4() {}
+ T &operator()(const int N1, const int N2, const int N3, const int N4)
+ {
+ return data[N1][N2][N3][N4];
+ }
+
+ template <char i, char j, char k, char l, int Dim0, int Dim1, int Dim2,
+ int Dim3>
+ Tensor4_Expr<Tensor4<T, Tensor_Dim0, Tensor_Dim1, Tensor_Dim2, Tensor_Dim3>,
+ T, Dim0, Dim1, Dim2, Dim3, i, j, k, l>
+ operator()(const Index<i, Dim0>, const Index<j, Dim1>, const Index<k, Dim2>,
+ const Index<l, Dim3>)
+ {
+ return Tensor4_Expr<
+ Tensor4<T, Tensor_Dim0, Tensor_Dim1, Tensor_Dim2, Tensor_Dim3>, T, Dim0,
+ Dim1, Dim2, Dim3, i, j, k, l>(*this);
+ };
+};
+
+template <class A, class T, int Dim0, int Dim1, int Dim2, int Dim3, char i,
+ char j, char k, char l>
+struct Tensor4_Expr
+{
+ A iter;
+
+ Tensor4_Expr(const A &a) : iter(a) {}
+ T operator()(const int N1, const int N2, const int N3, const int N4) const
+ {
+ return iter(N1, N2, N3, N4);
+ }
+};
+
+template <class A, class T, int Dim0, int Dim1, int Dim2, int Dim3, char i,
+ char j, char k, char l>
+struct Tensor4_Expr<Tensor4<A, Dim0, Dim1, Dim2, Dim3>, T, Dim0, Dim1, Dim2,
+ Dim3, i, j, k, l>
+{
+ Tensor4<A, Dim0, Dim1, Dim2, Dim3> &iter;
+
+ Tensor4_Expr(Tensor4<A, Dim0, Dim1, Dim2, Dim3> &a) : iter(a) {}
+ T operator()(const int N1, const int N2, const int N3, const int N4) const
+ {
+ return iter(N1, N2, N3, N4);
+ }
+
+ template <class B, class U, int Dim1_0, int Dim1_1, int Dim1_2, int Dim1_3,
+ char i_1, char j_1, char k_1, char l_1>
+ auto &operator=(const Tensor4_Expr<B, U, Dim1_0, Dim1_1, Dim1_2, Dim1_3, i_1,
+ j_1, k_1, l_1> &rhs)
+ {
+ for(int ii = 0; ii < Dim0; ++ii)
+ for(int jj = 0; jj < Dim1; ++jj)
+ for(int kk = 0; kk < Dim2; ++kk)
+ for(int ll = 0; ll < Dim3; ++ll)
+ {
+ iter(ii, jj, kk, ll) = rhs(ii, jj, kk, ll);
+ }
+ return *this;
+ }
+};
+
+int main()
+{
+ Tensor3<float, 100, 100, 1000> t1;
+ Tensor3<float, 1000, 100, 100> t2;
+
+ Index<'l', 100> l;
+ Index<'m', 100> m;
+ Index<'k', 1000> k;
+ Index<'n', 100> n;
+ Index<'o', 100> o;
+
+ Tensor4<float, 100, 100, 100, 100> res;
+ res(l, m, n, o) = t1(l, m, k) * t2(k, n, o);
+ return 0;
+}
+
interface between the GIMPLE and RTL worlds. */
/* The infinite cost. */
-#define INFTY 10000000
+#define INFTY 1000000000
/* Returns the expected number of loop iterations for LOOP.
The average trip count is computed from profile data if it
comp_cost (): cost (0), complexity (0), scratch (0)
{}
- comp_cost (int cost, unsigned complexity, int scratch = 0)
+ comp_cost (int64_t cost, unsigned complexity, int64_t scratch = 0)
: cost (cost), complexity (complexity), scratch (scratch)
{}
/* Returns true if COST1 is smaller or equal than COST2. */
friend bool operator<= (comp_cost cost1, comp_cost cost2);
- int cost; /* The runtime cost. */
+ int64_t cost; /* The runtime cost. */
unsigned complexity; /* The estimate of the complexity of the code for
the computation (in no concrete units --
complexity field should be larger for more
complex expressions and addressing modes). */
- int scratch; /* Scratch used during cost computation. */
+ int64_t scratch; /* Scratch used during cost computation. */
};
static const comp_cost no_cost;
-static const comp_cost infinite_cost (INFTY, INFTY, INFTY);
+static const comp_cost infinite_cost (INFTY, 0, INFTY);
bool
comp_cost::infinite_cost_p ()
if (cost1.infinite_cost_p () || cost2.infinite_cost_p ())
return infinite_cost;
+ gcc_assert (cost1.cost + cost2.cost < infinite_cost.cost);
cost1.cost += cost2.cost;
cost1.complexity += cost2.complexity;
return infinite_cost;
gcc_assert (!cost2.infinite_cost_p ());
+ gcc_assert (cost1.cost - cost2.cost < infinite_cost.cost);
cost1.cost -= cost2.cost;
cost1.complexity -= cost2.complexity;
if (infinite_cost_p ())
return *this;
+ gcc_assert (this->cost + c < infinite_cost.cost);
this->cost += c;
return *this;
if (infinite_cost_p ())
return *this;
+ gcc_assert (this->cost - c < infinite_cost.cost);
this->cost -= c;
return *this;
comp_cost
comp_cost::operator/= (HOST_WIDE_INT c)
{
+ gcc_assert (c != 0);
if (infinite_cost_p ())
return *this;
if (infinite_cost_p ())
return *this;
+ gcc_assert (this->cost * c < infinite_cost.cost);
this->cost *= c;
return *this;
comp_cost cand_use_cost;
/* Total cost of candidates. */
- unsigned cand_cost;
+ int64_t cand_cost;
/* Number of times each invariant variable is used. */
unsigned *n_inv_var_uses;
if we're optimizing for speed, amortize it over the per-iteration cost.
If ROUND_UP_P is true, the result is round up rather than to zero when
optimizing for speed. */
-static unsigned
-adjust_setup_cost (struct ivopts_data *data, unsigned cost,
+static int64_t
+adjust_setup_cost (struct ivopts_data *data, int64_t cost,
bool round_up_p = false)
{
if (cost == INFTY)
return cost;
else if (optimize_loop_for_speed_p (data->current_loop))
{
- HOST_WIDE_INT niters = avg_loop_niter (data->current_loop);
- return ((HOST_WIDE_INT) cost + (round_up_p ? niters - 1 : 0)) / niters;
+ int64_t niters = (int64_t) avg_loop_niter (data->current_loop);
+ return (cost + (round_up_p ? niters - 1 : 0)) / niters;
}
else
return cost;
struct ainc_cost_data
{
- unsigned costs[AINC_NONE];
+ int64_t costs[AINC_NONE];
};
static comp_cost
if (scale_factor == 1)
return cost;
- int scaled_cost
+ int64_t scaled_cost
= cost.scratch + (cost.cost - cost.scratch) * scale_factor;
if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "Scaling cost based on bb prob "
- "by %2.2f: %d (scratch: %d) -> %d\n",
+ fprintf (dump_file, "Scaling cost based on bb prob by %2.2f: "
+ "%" PRId64 " (scratch: %" PRId64 ") -> %" PRId64 "\n",
1.0f * scale_factor, cost.cost, cost.scratch, scaled_cost);
cost.cost = scaled_cost;
|| group->cost_map[j].cost.infinite_cost_p ())
continue;
- fprintf (dump_file, " %d\t%d\t%d\t",
+ fprintf (dump_file, " %d\t%" PRId64 "\t%d\t",
group->cost_map[j].cand->id,
group->cost_map[j].cost.cost,
group->cost_map[j].cost.complexity);
determine_iv_cost (struct ivopts_data *data, struct iv_cand *cand)
{
comp_cost cost_base;
- unsigned cost, cost_step;
+ int64_t cost, cost_step;
tree base;
gcc_assert (cand->iv != NULL);
unsigned i;
comp_cost cost = iv_ca_cost (ivs);
- fprintf (file, " cost: %d (complexity %d)\n", cost.cost,
+ fprintf (file, " cost: %" PRId64 " (complexity %d)\n", cost.cost,
cost.complexity);
- fprintf (file, " cand_cost: %d\n cand_group_cost: %d (complexity %d)\n",
- ivs->cand_cost, ivs->cand_use_cost.cost,
- ivs->cand_use_cost.complexity);
+ fprintf (file, " cand_cost: %" PRId64 "\n cand_group_cost: "
+ "%" PRId64 " (complexity %d)\n", ivs->cand_cost,
+ ivs->cand_use_cost.cost, ivs->cand_use_cost.complexity);
bitmap_print (file, ivs->cands, " candidates: ","\n");
for (i = 0; i < ivs->upto; i++)
struct iv_group *group = data->vgroups[i];
struct cost_pair *cp = iv_ca_cand_for_group (ivs, group);
if (cp)
- fprintf (file, " group:%d --> iv_cand:%d, cost=(%d,%d)\n",
- group->id, cp->cand->id, cp->cost.cost,
- cp->cost.complexity);
+ fprintf (file, " group:%d --> iv_cand:%d, cost=("
+ "%" PRId64 ",%d)\n", group->id, cp->cand->id,
+ cp->cost.cost, cp->cost.complexity);
else
fprintf (file, " group:%d --> ??\n", group->id);
}
if (dump_file && (dump_flags & TDF_DETAILS))
{
- fprintf (dump_file, "Original cost %d (complexity %d)\n\n",
+ fprintf (dump_file, "Original cost %" PRId64 " (complexity %d)\n\n",
origcost.cost, origcost.complexity);
- fprintf (dump_file, "Final cost %d (complexity %d)\n\n",
+ fprintf (dump_file, "Final cost %" PRId64 " (complexity %d)\n\n",
cost.cost, cost.complexity);
}
projects / gcc.git / commitdiff