/* Array prefetching.
- Copyright (C) 2005, 2007, 2008, 2009, 2010 Free Software Foundation, Inc.
+ Copyright (C) 2005-2013 Free Software Foundation, Inc.
This file is part of GCC.
#include "coretypes.h"
#include "tm.h"
#include "tree.h"
+#include "stor-layout.h"
#include "tm_p.h"
#include "basic-block.h"
-#include "output.h"
#include "tree-pretty-print.h"
-#include "tree-flow.h"
-#include "tree-dump.h"
-#include "timevar.h"
+#include "gimple.h"
+#include "gimplify.h"
+#include "gimple-iterator.h"
+#include "gimplify-me.h"
+#include "gimple-ssa.h"
+#include "tree-ssa-loop-ivopts.h"
+#include "tree-ssa-loop-manip.h"
+#include "tree-ssa-loop-niter.h"
+#include "tree-ssa-loop.h"
+#include "tree-into-ssa.h"
#include "cfgloop.h"
#include "tree-pass.h"
#include "insn-config.h"
-#include "recog.h"
#include "hashtab.h"
#include "tree-chrec.h"
#include "tree-scalar-evolution.h"
#include "diagnostic-core.h"
-#include "toplev.h"
#include "params.h"
#include "langhooks.h"
#include "tree-inline.h"
between the GIMPLE and RTL worlds. */
#include "expr.h"
#include "optabs.h"
+#include "recog.h"
/* This pass inserts prefetch instructions to optimize cache usage during
accesses to arrays in loops. It processes loops sequentially and:
nontemporal one. */
};
-/* Dumps information about reference REF to FILE. */
-
+/* Dumps information about memory reference */
static void
-dump_mem_ref (FILE *file, struct mem_ref *ref)
+dump_mem_details (FILE *file, tree base, tree step,
+ HOST_WIDE_INT delta, bool write_p)
{
- fprintf (file, "Reference %p:\n", (void *) ref);
-
- fprintf (file, " group %p (base ", (void *) ref->group);
- print_generic_expr (file, ref->group->base, TDF_SLIM);
+ fprintf (file, "(base ");
+ print_generic_expr (file, base, TDF_SLIM);
fprintf (file, ", step ");
- if (cst_and_fits_in_hwi (ref->group->step))
- fprintf (file, HOST_WIDE_INT_PRINT_DEC, int_cst_value (ref->group->step));
+ if (cst_fits_shwi_p (step))
+ fprintf (file, HOST_WIDE_INT_PRINT_DEC, int_cst_value (step));
else
- print_generic_expr (file, ref->group->step, TDF_TREE);
+ print_generic_expr (file, step, TDF_TREE);
fprintf (file, ")\n");
-
fprintf (file, " delta ");
- fprintf (file, HOST_WIDE_INT_PRINT_DEC, ref->delta);
+ fprintf (file, HOST_WIDE_INT_PRINT_DEC, delta);
+ fprintf (file, "\n");
+ fprintf (file, " %s\n", write_p ? "write" : "read");
fprintf (file, "\n");
+}
- fprintf (file, " %s\n", ref->write_p ? "write" : "read");
+/* Dumps information about reference REF to FILE. */
- fprintf (file, "\n");
+static void
+dump_mem_ref (FILE *file, struct mem_ref *ref)
+{
+ fprintf (file, "Reference %p:\n", (void *) ref);
+
+ fprintf (file, " group %p ", (void *) ref->group);
+
+ dump_mem_details (file, ref->group->base, ref->group->step, ref->delta,
+ ref->write_p);
}
/* Finds a group with BASE and STEP in GROUPS, or creates one if it does not
/* If step is an integer constant, keep the list of groups sorted
by decreasing step. */
- if (cst_and_fits_in_hwi ((*groups)->step) && cst_and_fits_in_hwi (step)
+ if (cst_fits_shwi_p ((*groups)->step) && cst_fits_shwi_p (step)
&& int_cst_value ((*groups)->step) < int_cst_value (step))
break;
}
step = iv.step;
if (TREE_CODE (ibase) == POINTER_PLUS_EXPR
- && cst_and_fits_in_hwi (TREE_OPERAND (ibase, 1)))
+ && cst_fits_shwi_p (TREE_OPERAND (ibase, 1)))
{
idelta = int_cst_value (TREE_OPERAND (ibase, 1));
ibase = TREE_OPERAND (ibase, 0);
}
- if (cst_and_fits_in_hwi (ibase))
+ if (cst_fits_shwi_p (ibase))
{
idelta += int_cst_value (ibase);
ibase = build_int_cst (TREE_TYPE (ibase), 0);
if (TREE_CODE (base) == ARRAY_REF)
{
stepsize = array_ref_element_size (base);
- if (!cst_and_fits_in_hwi (stepsize))
+ if (!cst_fits_shwi_p (stepsize))
return false;
imult = int_cst_value (stepsize);
step = fold_build2 (MULT_EXPR, sizetype,
if (may_be_nonaddressable_p (base))
return false;
- /* Limit non-constant step prefetching only to the innermost loops. */
- if (!cst_and_fits_in_hwi (step) && loop->inner != NULL)
- return false;
+ /* Limit non-constant step prefetching only to the innermost loops and
+ only when the step is loop invariant in the entire loop nest. */
+ if (!cst_fits_shwi_p (step))
+ {
+ if (loop->inner != NULL)
+ {
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, "Memory expression %p\n",(void *) ref );
+ print_generic_expr (dump_file, ref, TDF_TREE);
+ fprintf (dump_file,":");
+ dump_mem_details (dump_file, base, step, delta, write_p);
+ fprintf (dump_file,
+ "Ignoring %p, non-constant step prefetching is "
+ "limited to inner most loops \n",
+ (void *) ref);
+ }
+ return false;
+ }
+ else
+ {
+ if (!expr_invariant_in_loop_p (loop_outermost (loop), step))
+ {
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, "Memory expression %p\n",(void *) ref );
+ print_generic_expr (dump_file, ref, TDF_TREE);
+ fprintf (dump_file,":");
+ dump_mem_details (dump_file, base, step, delta, write_p);
+ fprintf (dump_file,
+ "Not prefetching, ignoring %p due to "
+ "loop variant step\n",
+ (void *) ref);
+ }
+ return false;
+ }
+ }
+ }
/* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP
are integer constants. */
bool backward;
/* If the step size is non constant, we cannot calculate prefetch_mod. */
- if (!cst_and_fits_in_hwi (ref->group->step))
+ if (!cst_fits_shwi_p (ref->group->step))
return;
step = int_cst_value (ref->group->step);
int align_unit;
/* If the step is non constant we cannot calculate prefetch_before. */
- if (!cst_and_fits_in_hwi (ref->group->step)) {
+ if (!cst_fits_shwi_p (ref->group->step)) {
return;
}
prefetch_before = (hit_from - delta_r + step - 1) / step;
/* Do not reduce prefetch_before if we meet beyond cache size. */
- if (prefetch_before > (unsigned) abs (L2_CACHE_SIZE_BYTES / step))
+ if (prefetch_before > absu_hwi (L2_CACHE_SIZE_BYTES / step))
prefetch_before = PREFETCH_ALL;
if (prefetch_before < ref->prefetch_before)
ref->prefetch_before = prefetch_before;
for (ap = 0; ap < n_prefetches; ap++)
{
- if (cst_and_fits_in_hwi (ref->group->step))
+ if (cst_fits_shwi_p (ref->group->step))
{
/* Determine the address to prefetch. */
delta = (ahead + ap * ref->prefetch_mod) *
int_cst_value (ref->group->step);
- addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node,
- addr_base, size_int (delta));
+ addr = fold_build_pointer_plus_hwi (addr_base, delta);
addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, NULL,
true, GSI_SAME_STMT);
}
forward = fold_build2 (MULT_EXPR, sizetype,
fold_convert (sizetype, ref->group->step),
fold_convert (sizetype, size_int (ahead)));
- addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, addr_base,
- forward);
+ addr = fold_build_pointer_plus (addr_base, forward);
addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true,
NULL, true, GSI_SAME_STMT);
}
/* Create the prefetch instruction. */
- prefetch = gimple_build_call (built_in_decls[BUILT_IN_PREFETCH],
+ prefetch = gimple_build_call (builtin_decl_explicit (BUILT_IN_PREFETCH),
3, addr, write_p, local);
gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT);
}
static void
emit_mfence_after_loop (struct loop *loop)
{
- VEC (edge, heap) *exits = get_loop_exit_edges (loop);
+ vec<edge> exits = get_loop_exit_edges (loop);
edge exit;
gimple call;
gimple_stmt_iterator bsi;
unsigned i;
- FOR_EACH_VEC_ELT (edge, exits, i, exit)
+ FOR_EACH_VEC_ELT (exits, i, exit)
{
call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0);
bsi = gsi_after_labels (exit->dest);
gsi_insert_before (&bsi, call, GSI_NEW_STMT);
- mark_virtual_ops_for_renaming (call);
}
- VEC_free (edge, heap, exits);
+ exits.release ();
update_ssa (TODO_update_ssa_only_virtuals);
}
is a suitable place for it at each of the loop exits. */
if (FENCE_FOLLOWING_MOVNT != NULL_TREE)
{
- VEC (edge, heap) *exits = get_loop_exit_edges (loop);
+ vec<edge> exits = get_loop_exit_edges (loop);
unsigned i;
edge exit;
- FOR_EACH_VEC_ELT (edge, exits, i, exit)
+ FOR_EACH_VEC_ELT (exits, i, exit)
if ((exit->flags & EDGE_ABNORMAL)
- && exit->dest == EXIT_BLOCK_PTR)
+ && exit->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
ret = false;
- VEC_free (edge, heap, exits);
+ exits.release ();
}
return ret;
if ((unsigned) loop_depth (aloop) <= min_depth)
continue;
- if (host_integerp (step, 0))
- astep = tree_low_cst (step, 0);
+ if (tree_fits_shwi_p (step))
+ astep = tree_to_shwi (step);
else
astep = L1_CACHE_LINE_SIZE;
{
tree stride, access_fn;
HOST_WIDE_INT *strides, astride;
- VEC (tree, heap) *access_fns;
+ vec<tree> access_fns;
tree ref = DR_REF (dr);
unsigned i, ret = ~0u;
strides = XCNEWVEC (HOST_WIDE_INT, n);
access_fns = DR_ACCESS_FNS (dr);
- FOR_EACH_VEC_ELT (tree, access_fns, i, access_fn)
+ FOR_EACH_VEC_ELT (access_fns, i, access_fn)
{
/* Keep track of the reference corresponding to the subscript, so that we
know its stride. */
if (TREE_CODE (ref) == ARRAY_REF)
{
stride = TYPE_SIZE_UNIT (TREE_TYPE (ref));
- if (host_integerp (stride, 1))
- astride = tree_low_cst (stride, 1);
+ if (tree_fits_uhwi_p (stride))
+ astride = tree_to_uhwi (stride);
else
astride = L1_CACHE_LINE_SIZE;
/* Determines the distance till the first reuse of each reference in REFS
in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other
- memory references in the loop. */
+ memory references in the loop. Return false if the analysis fails. */
-static void
+static bool
determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs,
bool no_other_refs)
{
struct loop *nest, *aloop;
- VEC (data_reference_p, heap) *datarefs = NULL;
- VEC (ddr_p, heap) *dependences = NULL;
+ vec<data_reference_p> datarefs = vNULL;
+ vec<ddr_p> dependences = vNULL;
struct mem_ref_group *gr;
struct mem_ref *ref, *refb;
- VEC (loop_p, heap) *vloops = NULL;
+ vec<loop_p> vloops = vNULL;
unsigned *loop_data_size;
unsigned i, j, n;
unsigned volume, dist, adist;
ddr_p dep;
if (loop->inner)
- return;
+ return true;
/* Find the outermost loop of the loop nest of loop (we require that
there are no sibling loops inside the nest). */
We use this to estimate whether the reference is evicted from the
cache before its reuse. */
find_loop_nest (nest, &vloops);
- n = VEC_length (loop_p, vloops);
+ n = vloops.length ();
loop_data_size = XNEWVEC (unsigned, n);
volume = volume_of_references (refs);
i = n;
if (volume > L2_CACHE_SIZE_BYTES)
continue;
- aloop = VEC_index (loop_p, vloops, i);
- vol = estimated_loop_iterations_int (aloop, false);
- if (vol < 0)
+ aloop = vloops[i];
+ vol = estimated_stmt_executions_int (aloop);
+ if (vol == -1)
vol = expected_loop_iterations (aloop);
volume *= vol;
}
for (gr = refs; gr; gr = gr->next)
for (ref = gr->refs; ref; ref = ref->next)
{
- dr = create_data_ref (nest, ref->mem, ref->stmt, !ref->write_p);
+ dr = create_data_ref (nest, loop_containing_stmt (ref->stmt),
+ ref->mem, ref->stmt, !ref->write_p);
if (dr)
{
ref->reuse_distance = volume;
dr->aux = ref;
- VEC_safe_push (data_reference_p, heap, datarefs, dr);
+ datarefs.safe_push (dr);
}
else
no_other_refs = false;
}
- FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr)
+ FOR_EACH_VEC_ELT (datarefs, i, dr)
{
dist = self_reuse_distance (dr, loop_data_size, n, loop);
ref = (struct mem_ref *) dr->aux;
ref->independent_p = true;
}
- compute_all_dependences (datarefs, &dependences, vloops, true);
+ if (!compute_all_dependences (datarefs, &dependences, vloops, true))
+ return false;
- FOR_EACH_VEC_ELT (ddr_p, dependences, i, dep)
+ FOR_EACH_VEC_ELT (dependences, i, dep)
{
if (DDR_ARE_DEPENDENT (dep) == chrec_known)
continue;
fprintf (dump_file, " ref %p distance %u\n",
(void *) ref, ref->reuse_distance);
}
+
+ return true;
}
/* Determine whether or not the trip count to ahead ratio is too small based
return false;
ahead = (PREFETCH_LATENCY + time - 1) / time;
- est_niter = estimated_loop_iterations_int (loop, false);
+ est_niter = estimated_stmt_executions_int (loop);
+ if (est_niter == -1)
+ est_niter = max_stmt_executions_int (loop);
/* Prefetching is not likely to be profitable if the trip count to ahead
ratio is too small. */
if (nothing_to_prefetch_p (refs))
goto fail;
- determine_loop_nest_reuse (loop, refs, no_other_refs);
+ if (!determine_loop_nest_reuse (loop, refs, no_other_refs))
+ goto fail;
/* Step 3: determine unroll factor. */
unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc,
unsigned int
tree_ssa_prefetch_arrays (void)
{
- loop_iterator li;
struct loop *loop;
bool unrolled = false;
int todo_flags = 0;
initialize_original_copy_tables ();
- if (!built_in_decls[BUILT_IN_PREFETCH])
+ if (!builtin_decl_explicit_p (BUILT_IN_PREFETCH))
{
tree type = build_function_type_list (void_type_node,
const_ptr_type_node, NULL_TREE);
BUILT_IN_PREFETCH, BUILT_IN_NORMAL,
NULL, NULL_TREE);
DECL_IS_NOVOPS (decl) = true;
- built_in_decls[BUILT_IN_PREFETCH] = decl;
+ set_builtin_decl (BUILT_IN_PREFETCH, decl, false);
}
/* 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)
+ FOR_EACH_LOOP (loop, LI_FROM_INNERMOST)
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Processing loop %d:\n", loop->num);
free_original_copy_tables ();
return todo_flags;
}
+
+/* Prefetching. */
+
+static unsigned int
+tree_ssa_loop_prefetch (void)
+{
+ if (number_of_loops (cfun) <= 1)
+ return 0;
+
+ return tree_ssa_prefetch_arrays ();
+}
+
+static bool
+gate_tree_ssa_loop_prefetch (void)
+{
+ return flag_prefetch_loop_arrays > 0;
+}
+
+namespace {
+
+const pass_data pass_data_loop_prefetch =
+{
+ GIMPLE_PASS, /* type */
+ "aprefetch", /* name */
+ OPTGROUP_LOOP, /* optinfo_flags */
+ true, /* has_gate */
+ true, /* has_execute */
+ TV_TREE_PREFETCH, /* tv_id */
+ ( PROP_cfg | PROP_ssa ), /* properties_required */
+ 0, /* properties_provided */
+ 0, /* properties_destroyed */
+ 0, /* todo_flags_start */
+ 0, /* todo_flags_finish */
+};
+
+class pass_loop_prefetch : public gimple_opt_pass
+{
+public:
+ pass_loop_prefetch (gcc::context *ctxt)
+ : gimple_opt_pass (pass_data_loop_prefetch, ctxt)
+ {}
+
+ /* opt_pass methods: */
+ bool gate () { return gate_tree_ssa_loop_prefetch (); }
+ unsigned int execute () { return tree_ssa_loop_prefetch (); }
+
+}; // class pass_loop_prefetch
+
+} // anon namespace
+
+gimple_opt_pass *
+make_pass_loop_prefetch (gcc::context *ctxt)
+{
+ return new pass_loop_prefetch (ctxt);
+}
+
+
projects / thirdparty / gcc.git / blobdiff