static bool
graphite_initialize (isl_ctx *ctx)
{
if (number_of_loops (cfun) <= 1
/* FIXME: This limit on the number of basic blocks of a function
should be removed when the SCOP detection is faster. */
|| n_basic_blocks > PARAM_VALUE (PARAM_GRAPHITE_MAX_BBS_PER_FUNCTION))
{
if (dump_file && (dump_flags & TDF_DETAILS))
print_global_statistics (dump_file);
isl_ctx_free (ctx);
return false;
}
scev_reset ();
recompute_all_dominators ();
initialize_original_copy_tables ();
cloog_state = cloog_isl_state_malloc (ctx);
if (dump_file && dump_flags)
dump_function_to_file (current_function_decl, dump_file, dump_flags);
return true;
}
unsigned int
tree_ssa_prefetch_arrays (struct loops *loops)
{
unsigned i;
struct loop *loop;
bool unrolled = false;
int todo_flags = 0;
if (!HAVE_prefetch
/* It is possible to ask compiler for say -mtune=i486 -march=pentium4.
-mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part
of processor costs and i486 does not have prefetch, but
-march=pentium4 causes HAVE_prefetch to be true. Ugh. */
|| PREFETCH_BLOCK == 0)
return 0;
initialize_original_copy_tables ();
if (!built_in_decls[BUILT_IN_PREFETCH])
{
tree type = build_function_type (void_type_node,
tree_cons (NULL_TREE,
const_ptr_type_node,
NULL_TREE));
tree decl = lang_hooks.builtin_function ("__builtin_prefetch", type,
BUILT_IN_PREFETCH, BUILT_IN_NORMAL,
NULL, NULL_TREE);
DECL_IS_NOVOPS (decl) = true;
built_in_decls[BUILT_IN_PREFETCH] = decl;
}
/* We assume that size of cache line is a power of two, so verify this
here. */
gcc_assert ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) == 0);
for (i = loops->num - 1; i > 0; i--)
{
loop = loops->parray[i];
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Processing loop %d:\n", loop->num);
if (loop)
unrolled |= loop_prefetch_arrays (loops, loop);
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "\n\n");
}
if (unrolled)
{
scev_reset ();
todo_flags |= TODO_cleanup_cfg;
}
free_original_copy_tables ();
return todo_flags;
}
static unsigned int
tree_ssa_loop_bounds (void)
{
if (number_of_loops () <= 1)
return 0;
estimate_numbers_of_iterations ();
scev_reset ();
return 0;
}
static unsigned int
tree_ssa_loop_bounds (void)
{
if (!current_loops)
return 0;
estimate_numbers_of_iterations (current_loops);
scev_reset ();
return 0;
}
static void
graphite_finalize (bool need_cfg_cleanup_p)
{
if (need_cfg_cleanup_p)
{
scev_reset ();
cleanup_tree_cfg ();
profile_status = PROFILE_ABSENT;
release_recorded_exits ();
tree_estimate_probability ();
}
cloog_state_free (cloog_state);
free_original_copy_tables ();
if (dump_file && dump_flags)
print_loops (dump_file, 3);
}
static bool
fini_copy_prop (void)
{
unsigned i;
/* Set the final copy-of value for each variable by traversing the
copy-of chains. */
for (i = 1; i < num_ssa_names; i++)
{
tree var = ssa_name (i);
if (!var
|| !copy_of[i].value
|| copy_of[i].value == var)
continue;
/* In theory the points-to solution of all members of the
copy chain is their intersection. For now we do not bother
to compute this but only make sure we do not lose points-to
information completely by setting the points-to solution
of the representative to the first solution we find if
it doesn't have one already. */
if (copy_of[i].value != var
&& TREE_CODE (copy_of[i].value) == SSA_NAME)
{
basic_block copy_of_bb
= gimple_bb (SSA_NAME_DEF_STMT (copy_of[i].value));
basic_block var_bb = gimple_bb (SSA_NAME_DEF_STMT (var));
if (POINTER_TYPE_P (TREE_TYPE (var))
&& SSA_NAME_PTR_INFO (var)
&& !SSA_NAME_PTR_INFO (copy_of[i].value))
{
duplicate_ssa_name_ptr_info (copy_of[i].value,
SSA_NAME_PTR_INFO (var));
/* Points-to information is cfg insensitive,
but alignment info might be cfg sensitive, if it
e.g. is derived from VRP derived non-zero bits.
So, do not copy alignment info if the two SSA_NAMEs
aren't defined in the same basic block. */
if (var_bb != copy_of_bb)
mark_ptr_info_alignment_unknown
(SSA_NAME_PTR_INFO (copy_of[i].value));
}
else if (!POINTER_TYPE_P (TREE_TYPE (var))
&& SSA_NAME_RANGE_INFO (var)
&& !SSA_NAME_RANGE_INFO (copy_of[i].value)
&& var_bb == copy_of_bb)
duplicate_ssa_name_range_info (copy_of[i].value,
SSA_NAME_RANGE_TYPE (var),
SSA_NAME_RANGE_INFO (var));
}
}
bool changed = substitute_and_fold (get_value, NULL, true);
if (changed)
{
free_numbers_of_iterations_estimates ();
if (scev_initialized_p ())
scev_reset ();
}
free (copy_of);
return changed;
}
void
linear_transform_loops (struct loops *loops)
{
unsigned int i;
compute_immediate_uses (TDFA_USE_OPS | TDFA_USE_VOPS, NULL);
for (i = 1; i < loops->num; i++)
{
unsigned int depth = 0;
varray_type datarefs;
varray_type dependence_relations;
struct loop *loop_nest = loops->parray[i];
struct loop *temp;
VEC (tree) *oldivs = NULL;
VEC (tree) *invariants = NULL;
lambda_loopnest before, after;
lambda_trans_matrix trans;
bool problem = false;
bool need_perfect_nest = false;
/* If it's not a loop nest, we don't want it.
We also don't handle sibling loops properly,
which are loops of the following form:
for (i = 0; i < 50; i++)
{
for (j = 0; j < 50; j++)
{
...
}
for (j = 0; j < 50; j++)
{
...
}
} */
if (!loop_nest->inner)
continue;
depth = 1;
for (temp = loop_nest->inner; temp; temp = temp->inner)
{
flow_loop_scan (temp, LOOP_ALL);
/* If we have a sibling loop or multiple exit edges, jump ship. */
if (temp->next || temp->num_exits != 1)
{
problem = true;
break;
}
depth ++;
}
if (problem)
continue;
/* Analyze data references and dependence relations using scev. */
VARRAY_GENERIC_PTR_INIT (datarefs, 10, "datarefs");
VARRAY_GENERIC_PTR_INIT (dependence_relations, 10,
"dependence_relations");
compute_data_dependences_for_loop (depth, loop_nest,
&datarefs, &dependence_relations);
if (dump_file && (dump_flags & TDF_DETAILS))
{
unsigned int j;
for (j = 0; j < VARRAY_ACTIVE_SIZE (dependence_relations); j++)
{
struct data_dependence_relation *ddr =
(struct data_dependence_relation *)
VARRAY_GENERIC_PTR (dependence_relations, j);
if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
{
fprintf (dump_file, "DISTANCE_V (");
print_lambda_vector (dump_file, DDR_DIST_VECT (ddr),
DDR_SIZE_VECT (ddr));
fprintf (dump_file, ")\n");
fprintf (dump_file, "DIRECTION_V (");
print_lambda_vector (dump_file, DDR_DIR_VECT (ddr),
DDR_SIZE_VECT (ddr));
fprintf (dump_file, ")\n");
}
}
fprintf (dump_file, "\n\n");
}
/* Build the transformation matrix. */
trans = lambda_trans_matrix_new (depth, depth);
lambda_matrix_id (LTM_MATRIX (trans), depth);
trans = try_interchange_loops (trans, depth, dependence_relations,
datarefs, loop_nest);
if (lambda_trans_matrix_id_p (trans))
{
if (dump_file)
fprintf (dump_file, "Won't transform loop. Optimal transform is the identity transform\n");
continue;
}
/* Check whether the transformation is legal. */
if (!lambda_transform_legal_p (trans, depth, dependence_relations))
{
if (dump_file)
fprintf (dump_file, "Can't transform loop, transform is illegal:\n");
continue;
}
if (!perfect_nest_p (loop_nest))
need_perfect_nest = true;
before = gcc_loopnest_to_lambda_loopnest (loops,
loop_nest, &oldivs,
&invariants,
need_perfect_nest);
if (!before)
continue;
if (dump_file)
{
fprintf (dump_file, "Before:\n");
print_lambda_loopnest (dump_file, before, 'i');
}
after = lambda_loopnest_transform (before, trans);
if (dump_file)
{
fprintf (dump_file, "After:\n");
print_lambda_loopnest (dump_file, after, 'u');
}
lambda_loopnest_to_gcc_loopnest (loop_nest, oldivs, invariants,
after, trans);
if (dump_file)
fprintf (dump_file, "Successfully transformed loop.\n");
oldivs = NULL;
invariants = NULL;
free_dependence_relations (dependence_relations);
free_data_refs (datarefs);
}
free_df ();
scev_reset ();
rewrite_into_loop_closed_ssa ();
#ifdef ENABLE_CHECKING
verify_loop_closed_ssa ();
#endif
}
unsigned int
tree_ssa_prefetch_arrays (void)
{
loop_iterator li;
struct loop *loop;
bool unrolled = false;
int todo_flags = 0;
if (!HAVE_prefetch
/* It is possible to ask compiler for say -mtune=i486 -march=pentium4.
-mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part
of processor costs and i486 does not have prefetch, but
-march=pentium4 causes HAVE_prefetch to be true. Ugh. */
|| PREFETCH_BLOCK == 0)
return 0;
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Prefetching parameters:\n");
fprintf (dump_file, " simultaneous prefetches: %d\n",
SIMULTANEOUS_PREFETCHES);
fprintf (dump_file, " prefetch latency: %d\n", PREFETCH_LATENCY);
fprintf (dump_file, " prefetch block size: %d\n", PREFETCH_BLOCK);
fprintf (dump_file, " L1 cache size: %d lines, %d kB\n",
L1_CACHE_SIZE_BYTES / L1_CACHE_LINE_SIZE, L1_CACHE_SIZE);
fprintf (dump_file, " L1 cache line size: %d\n", L1_CACHE_LINE_SIZE);
fprintf (dump_file, " L2 cache size: %d kB\n", L2_CACHE_SIZE);
fprintf (dump_file, "\n");
}
initialize_original_copy_tables ();
if (!built_in_decls[BUILT_IN_PREFETCH])
{
tree type = build_function_type (void_type_node,
tree_cons (NULL_TREE,
const_ptr_type_node,
NULL_TREE));
tree decl = add_builtin_function ("__builtin_prefetch", type,
BUILT_IN_PREFETCH, BUILT_IN_NORMAL,
NULL, NULL_TREE);
DECL_IS_NOVOPS (decl) = true;
built_in_decls[BUILT_IN_PREFETCH] = decl;
}
/* We assume that size of cache line is a power of two, so verify this
here. */
gcc_assert ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) == 0);
FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Processing loop %d:\n", loop->num);
unrolled |= loop_prefetch_arrays (loop);
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "\n\n");
}
if (unrolled)
{
scev_reset ();
todo_flags |= TODO_cleanup_cfg;
}
free_original_copy_tables ();
return todo_flags;
}
