/* Loop transformation code generation
- Copyright (C) 2003, 2004, 2005 Free Software Foundation, Inc.
+ Copyright (C) 2003, 2004, 2005, 2006 Free Software Foundation, Inc.
Contributed by Daniel Berlin <dberlin@dberlin.org>
This file is part of GCC.
#include "tree-scalar-evolution.h"
#include "vec.h"
#include "lambda.h"
+#include "vecprim.h"
/* This loop nest code generation is based on non-singular matrix
math.
Fourier-Motzkin elimination is used to compute the bounds of the base space
of the lattice. */
-DEF_VEC_I(int);
-DEF_VEC_ALLOC_I(int,heap);
-
static bool perfect_nestify (struct loops *,
struct loop *, VEC(tree,heap) *,
VEC(tree,heap) *, VEC(int,heap) *,
static lambda_lattice lambda_lattice_compute_base (lambda_loopnest);
static tree find_induction_var_from_exit_cond (struct loop *);
+static bool can_convert_to_perfect_nest (struct loop *);
/* Create a new lambda body vector. */
return ret;
}
-/* Compute the greatest common denominator of two numbers (A and B) using
- Euclid's algorithm. */
-
-static int
-gcd (int a, int b)
-{
-
- int x, y, z;
-
- x = abs (a);
- y = abs (b);
-
- while (x > 0)
- {
- z = y % x;
- y = x;
- x = z;
- }
-
- return (y);
-}
-
-/* Compute the greatest common denominator of a VECTOR of SIZE numbers. */
-
-static int
-gcd_vector (lambda_vector vector, int size)
-{
- int i;
- int gcd1 = 0;
-
- if (size > 0)
- {
- gcd1 = vector[0];
- for (i = 1; i < size; i++)
- gcd1 = gcd (gcd1, vector[i]);
- }
- return gcd1;
-}
-
/* Compute the least common multiple of two numbers A and B . */
static int
LN_LOOPS (target_nest)[i] = target_loop;
/* Computes the gcd of the coefficients of the linear part. */
- gcd1 = gcd_vector (target[i], i);
+ gcd1 = lambda_vector_gcd (target[i], i);
/* Include the denominator in the GCD. */
gcd1 = gcd (gcd1, determinant);
}
/* Find the gcd and divide by it here, rather than doing it
at the tree level. */
- gcd1 = gcd_vector (LLE_COEFFICIENTS (target_expr), depth);
- gcd2 = gcd_vector (LLE_INVARIANT_COEFFICIENTS (target_expr),
- invariants);
+ gcd1 = lambda_vector_gcd (LLE_COEFFICIENTS (target_expr), depth);
+ gcd2 = lambda_vector_gcd (LLE_INVARIANT_COEFFICIENTS (target_expr),
+ invariants);
gcd1 = gcd (gcd1, gcd2);
gcd1 = gcd (gcd1, LLE_CONSTANT (target_expr));
gcd1 = gcd (gcd1, LLE_DENOMINATOR (target_expr));
}
/* Find the gcd and divide by it here, instead of at the
tree level. */
- gcd1 = gcd_vector (LLE_COEFFICIENTS (target_expr), depth);
- gcd2 = gcd_vector (LLE_INVARIANT_COEFFICIENTS (target_expr),
- invariants);
+ gcd1 = lambda_vector_gcd (LLE_COEFFICIENTS (target_expr), depth);
+ gcd2 = lambda_vector_gcd (LLE_INVARIANT_COEFFICIENTS (target_expr),
+ invariants);
gcd1 = gcd (gcd1, gcd2);
gcd1 = gcd (gcd1, LLE_CONSTANT (target_expr));
gcd1 = gcd (gcd1, LLE_DENOMINATOR (target_expr));
lambda_loopnest
gcc_loopnest_to_lambda_loopnest (struct loops *currloops,
- struct loop * loop_nest,
+ struct loop *loop_nest,
VEC(tree,heap) **inductionvars,
- VEC(tree,heap) **invariants,
- bool need_perfect_nest)
+ VEC(tree,heap) **invariants)
{
lambda_loopnest ret = NULL;
- struct loop *temp;
- int depth = 0;
+ struct loop *temp = loop_nest;
+ int depth = depth_of_nest (loop_nest);
size_t i;
VEC(lambda_loop,heap) *loops = NULL;
VEC(tree,heap) *uboundvars = NULL;
VEC(int,heap) *steps = NULL;
lambda_loop newloop;
tree inductionvar = NULL;
-
- depth = depth_of_nest (loop_nest);
- temp = loop_nest;
+ bool perfect_nest = perfect_nest_p (loop_nest);
+
+ if (!perfect_nest && !can_convert_to_perfect_nest (loop_nest))
+ goto fail;
+
while (temp)
{
newloop = gcc_loop_to_lambda_loop (temp, depth, invariants,
&lboundvars, &uboundvars,
&steps);
if (!newloop)
- return NULL;
+ goto fail;
+
VEC_safe_push (tree, heap, *inductionvars, inductionvar);
VEC_safe_push (lambda_loop, heap, loops, newloop);
temp = temp->inner;
}
- if (need_perfect_nest)
+
+ if (!perfect_nest)
{
if (!perfect_nestify (currloops, loop_nest,
lboundvars, uboundvars, steps, *inductionvars))
fprintf (dump_file,
"Successfully converted loop nest to perfect loop nest.\n");
}
+
ret = lambda_loopnest_new (depth, 2 * depth);
+
for (i = 0; VEC_iterate (lambda_loop, loops, i, newloop); i++)
LN_LOOPS (ret)[i] = newloop;
+
fail:
VEC_free (lambda_loop, heap, loops);
VEC_free (tree, heap, uboundvars);
for (i = 0; VEC_iterate (tree, old_ivs, i, oldiv); i++)
{
imm_use_iterator imm_iter;
- use_operand_p imm_use;
+ use_operand_p use_p;
tree oldiv_def;
tree oldiv_stmt = SSA_NAME_DEF_STMT (oldiv);
+ tree stmt;
if (TREE_CODE (oldiv_stmt) == PHI_NODE)
oldiv_def = PHI_RESULT (oldiv_stmt);
oldiv_def = SINGLE_SSA_TREE_OPERAND (oldiv_stmt, SSA_OP_DEF);
gcc_assert (oldiv_def != NULL_TREE);
- FOR_EACH_IMM_USE_SAFE (imm_use, imm_iter, oldiv_def)
- {
- tree stmt = USE_STMT (imm_use);
- use_operand_p use_p;
- ssa_op_iter iter;
+ FOR_EACH_IMM_USE_STMT (stmt, imm_iter, oldiv_def)
+ {
+ tree newiv, stmts;
+ lambda_body_vector lbv, newlbv;
+
gcc_assert (TREE_CODE (stmt) != PHI_NODE);
- FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
- {
- if (USE_FROM_PTR (use_p) == oldiv)
- {
- tree newiv, stmts;
- lambda_body_vector lbv, newlbv;
- /* Compute the new expression for the induction
- variable. */
- depth = VEC_length (tree, new_ivs);
- lbv = lambda_body_vector_new (depth);
- LBV_COEFFICIENTS (lbv)[i] = 1;
-
- newlbv = lambda_body_vector_compute_new (transform, lbv);
-
- newiv = lbv_to_gcc_expression (newlbv, TREE_TYPE (oldiv),
- new_ivs, &stmts);
- bsi = bsi_for_stmt (stmt);
- /* Insert the statements to build that
- expression. */
- bsi_insert_before (&bsi, stmts, BSI_SAME_STMT);
- propagate_value (use_p, newiv);
- update_stmt (stmt);
-
- }
- }
+
+ /* Compute the new expression for the induction
+ variable. */
+ depth = VEC_length (tree, new_ivs);
+ lbv = lambda_body_vector_new (depth);
+ LBV_COEFFICIENTS (lbv)[i] = 1;
+
+ newlbv = lambda_body_vector_compute_new (transform, lbv);
+
+ newiv = lbv_to_gcc_expression (newlbv, TREE_TYPE (oldiv),
+ new_ivs, &stmts);
+ bsi = bsi_for_stmt (stmt);
+ /* Insert the statements to build that
+ expression. */
+ bsi_insert_before (&bsi, stmts, BSI_SAME_STMT);
+
+ FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
+ propagate_value (use_p, newiv);
+ update_stmt (stmt);
}
}
VEC_free (tree, heap, new_ivs);
{
tree use = USE_FROM_PTR (use_p);
tree step = NULL_TREE;
- tree access_fn = NULL_TREE;
-
-
- access_fn = instantiate_parameters
- (loop, analyze_scalar_evolution (loop, use));
- if (access_fn != NULL_TREE && access_fn != chrec_dont_know)
- step = evolution_part_in_loop_num (access_fn, loop->num);
+ tree scev = instantiate_parameters (loop,
+ analyze_scalar_evolution (loop, use));
+
+ if (scev != NULL_TREE && scev != chrec_dont_know)
+ step = evolution_part_in_loop_num (scev, loop->num);
+
if ((step && step != chrec_dont_know
&& TREE_CODE (step) == INTEGER_CST
&& int_cst_value (step) == xstep)
}
}
-/* Return TRUE if STMT uses tree OP in it's uses. */
-
-static bool
-stmt_uses_op (tree stmt, tree op)
-{
- ssa_op_iter iter;
- tree use;
-
- FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
- {
- if (use == op)
- return true;
- }
- return false;
-}
-
/* Return true if STMT is an exit PHI for LOOP */
static bool
-/* Return TRUE if LOOP is an imperfect nest that we can convert to a perfect
- one. LOOPIVS is a vector of induction variables, one per loop.
- ATM, we only handle imperfect nests of depth 2, where all of the statements
- occur after the inner loop. */
+/* Return TRUE if LOOP is an imperfect nest that we can convert to a
+ perfect one. At the moment, we only handle imperfect nests of
+ depth 2, where all of the statements occur after the inner loop. */
static bool
-can_convert_to_perfect_nest (struct loop *loop,
- VEC(tree,heap) *loopivs)
+can_convert_to_perfect_nest (struct loop *loop)
{
basic_block *bbs;
tree exit_condition, phi;
{
for (bsi = bsi_start (bbs[i]); !bsi_end_p (bsi); bsi_next (&bsi))
{
- size_t j;
tree stmt = bsi_stmt (bsi);
- tree iv;
-
+
if (stmt == exit_condition
|| not_interesting_stmt (stmt)
|| stmt_is_bumper_for_loop (loop, stmt))
continue;
- /* If the statement uses inner loop ivs, we == screwed. */
- for (j = 1; VEC_iterate (tree, loopivs, j, iv); j++)
- if (stmt_uses_op (stmt, iv))
- goto fail;
-
- /* If this is a simple operation like a cast that is
- invariant in the inner loop, or after the inner loop,
- then see if we can place it back where it came from.
- This means that we will propagate casts and other
- cheap invariant operations *back* into or after
- the inner loop if we can interchange the loop, on the
- theory that we are going to gain a lot more by
- interchanging the loop than we are by leaving some
- invariant code there for some other pass to clean
- up. */
- if (TREE_CODE (stmt) == MODIFY_EXPR
- && is_gimple_cast (TREE_OPERAND (stmt, 1))
- && (can_put_in_inner_loop (loop->inner, stmt)
- || can_put_after_inner_loop (loop, stmt)))
- continue;
+
+ /* If this is a scalar operation that can be put back
+ into the inner loop, or after the inner loop, through
+ copying, then do so. This works on the theory that
+ any amount of scalar code we have to reduplicate
+ into or after the loops is less expensive that the
+ win we get from rearranging the memory walk
+ the loop is doing so that it has better
+ cache behavior. */
+ if (TREE_CODE (stmt) == MODIFY_EXPR)
+ {
+ use_operand_p use_a, use_b;
+ imm_use_iterator imm_iter;
+ ssa_op_iter op_iter, op_iter1;
+ tree op0 = TREE_OPERAND (stmt, 0);
+ tree scev = instantiate_parameters
+ (loop, analyze_scalar_evolution (loop, op0));
+
+ /* If the IV is simple, it can be duplicated. */
+ if (!automatically_generated_chrec_p (scev))
+ {
+ tree step = evolution_part_in_loop_num (scev, loop->num);
+ if (step && step != chrec_dont_know
+ && TREE_CODE (step) == INTEGER_CST)
+ continue;
+ }
+
+ /* The statement should not define a variable used
+ in the inner loop. */
+ if (TREE_CODE (op0) == SSA_NAME)
+ FOR_EACH_IMM_USE_FAST (use_a, imm_iter, op0)
+ if (bb_for_stmt (USE_STMT (use_a))->loop_father
+ == loop->inner)
+ goto fail;
+
+ FOR_EACH_SSA_USE_OPERAND (use_a, stmt, op_iter, SSA_OP_USE)
+ {
+ tree node, op = USE_FROM_PTR (use_a);
+
+ /* The variables should not be used in both loops. */
+ FOR_EACH_IMM_USE_FAST (use_b, imm_iter, op)
+ if (bb_for_stmt (USE_STMT (use_b))->loop_father
+ == loop->inner)
+ goto fail;
+
+ /* The statement should not use the value of a
+ scalar that was modified in the loop. */
+ node = SSA_NAME_DEF_STMT (op);
+ if (TREE_CODE (node) == PHI_NODE)
+ FOR_EACH_PHI_ARG (use_b, node, op_iter1, SSA_OP_USE)
+ {
+ tree arg = USE_FROM_PTR (use_b);
+
+ if (TREE_CODE (arg) == SSA_NAME)
+ {
+ tree arg_stmt = SSA_NAME_DEF_STMT (arg);
+
+ if (bb_for_stmt (arg_stmt)->loop_father
+ == loop->inner)
+ goto fail;
+ }
+ }
+ }
+
+ if (can_put_in_inner_loop (loop->inner, stmt)
+ || can_put_after_inner_loop (loop, stmt))
+ continue;
+ }
/* Otherwise, if the bb of a statement we care about isn't
dominated by the header of the inner loop, then we can't
tree stmt;
tree oldivvar, ivvar, ivvarinced;
VEC(tree,heap) *phis = NULL;
-
- if (!can_convert_to_perfect_nest (loop, loopivs))
- return false;
-
- /* Create the new loop */
-
+
+ /* Create the new loop. */
olddest = loop->single_exit->dest;
- preheaderbb = loop_split_edge_with (loop->single_exit, NULL);
+ preheaderbb = loop_split_edge_with (loop->single_exit, NULL);
headerbb = create_empty_bb (EXIT_BLOCK_PTR->prev_bb);
/* Push the exit phi nodes that we are moving. */
{
use_operand_p use_p;
imm_use_iterator imm_iter;
+ tree imm_stmt;
tree stmt = bsi_stmt (bsi);
if (stmt == exit_condition
}
/* Make copies of this statement to put it back next
- to its uses. */
- FOR_EACH_IMM_USE_SAFE (use_p, imm_iter,
+ to its uses. */
+ FOR_EACH_IMM_USE_STMT (imm_stmt, imm_iter,
TREE_OPERAND (stmt, 0))
{
- tree imm_stmt = USE_STMT (use_p);
if (!exit_phi_for_loop_p (loop->inner, imm_stmt))
{
block_stmt_iterator tobsi;
newname = SSA_NAME_VAR (newname);
newname = make_ssa_name (newname, newstmt);
TREE_OPERAND (newstmt, 0) = newname;
- SET_USE (use_p, TREE_OPERAND (newstmt, 0));
+
+ FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
+ SET_USE (use_p, newname);
+
bsi_insert_before (&tobsi, newstmt, BSI_SAME_STMT);
update_stmt (newstmt);
update_stmt (imm_stmt);
continue;
}
- replace_uses_equiv_to_x_with_y (loop, stmt,
- oldivvar,
- VEC_index (int, steps, 0),
- ivvar);
+ replace_uses_equiv_to_x_with_y
+ (loop, stmt, oldivvar, VEC_index (int, steps, 0), ivvar);
+
bsi_move_before (&bsi, &tobsi);
/* If the statement has any virtual operands, they may
bool
lambda_transform_legal_p (lambda_trans_matrix trans,
int nb_loops,
- varray_type dependence_relations)
+ VEC (ddr_p, heap) *dependence_relations)
{
unsigned int i, j;
lambda_vector distres;
/* When there is an unknown relation in the dependence_relations, we
know that it is no worth looking at this loop nest: give up. */
- ddr = (struct data_dependence_relation *)
- VARRAY_GENERIC_PTR (dependence_relations, 0);
+ ddr = VEC_index (ddr_p, dependence_relations, 0);
if (ddr == NULL)
return true;
if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
distres = lambda_vector_new (nb_loops);
/* For each distance vector in the dependence graph. */
- for (i = 0; i < VARRAY_ACTIVE_SIZE (dependence_relations); i++)
+ for (i = 0; VEC_iterate (ddr_p, dependence_relations, i, ddr); i++)
{
- ddr = (struct data_dependence_relation *)
- VARRAY_GENERIC_PTR (dependence_relations, i);
-
/* Don't care about relations for which we know that there is no
dependence, nor about read-read (aka. output-dependences):
these data accesses can happen in any order. */