/*
- * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
- *
* This file is part of libFirm.
- *
- * This file may be distributed and/or modified under the terms of the
- * GNU General Public License version 2 as published by the Free Software
- * Foundation and appearing in the file LICENSE.GPL included in the
- * packaging of this file.
- *
- * Licensees holding valid libFirm Professional Edition licenses may use
- * this file in accordance with the libFirm Commercial License.
- * Agreement provided with the Software.
- *
- * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
- * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
- * PURPOSE.
+ * Copyright (C) 2012 University of Karlsruhe.
*/
/**
* @file
* @brief Peephole optimisation framework keeps track of which registers contain which values
* @author Matthias Braun
- * @version $Id$
*/
#include "config.h"
#include "irprintf.h"
#include "ircons.h"
#include "irgmod.h"
+#include "heights.h"
#include "error.h"
#include "beirg.h"
#include "belive_t.h"
#include "bearch.h"
+#include "beintlive_t.h"
#include "benode.h"
#include "besched.h"
#include "bemodule.h"
static const arch_env_t *arch_env;
static be_lv_t *lv;
static ir_node *current_node;
-ir_node ***register_values;
+ir_node **register_values;
static void clear_reg_value(ir_node *node)
{
- const arch_register_t *reg;
- const arch_register_class_t *cls;
- unsigned reg_idx;
- unsigned cls_idx;
+ const arch_register_t *reg;
+ unsigned reg_idx;
if (!mode_is_data(get_irn_mode(node)))
return;
- reg = arch_get_irn_register(node);
+ reg = arch_get_irn_register(node);
if (reg == NULL) {
panic("No register assigned at %+F", node);
}
if (reg->type & arch_register_type_virtual)
return;
- cls = arch_register_get_class(reg);
- reg_idx = arch_register_get_index(reg);
- cls_idx = arch_register_class_index(cls);
+ reg_idx = reg->global_index;
- //assert(register_values[cls_idx][reg_idx] != NULL);
DB((dbg, LEVEL_1, "Clear Register %s\n", reg->name));
- register_values[cls_idx][reg_idx] = NULL;
+ register_values[reg_idx] = NULL;
}
static void set_reg_value(ir_node *node)
{
- const arch_register_t *reg;
- const arch_register_class_t *cls;
- unsigned reg_idx;
- unsigned cls_idx;
+ const arch_register_t *reg;
+ unsigned reg_idx;
if (!mode_is_data(get_irn_mode(node)))
return;
}
if (reg->type & arch_register_type_virtual)
return;
- cls = arch_register_get_class(reg);
- reg_idx = arch_register_get_index(reg);
- cls_idx = arch_register_class_index(cls);
+ reg_idx = reg->global_index;
DB((dbg, LEVEL_1, "Set Register %s: %+F\n", reg->name, node));
- register_values[cls_idx][reg_idx] = node;
+ register_values[reg_idx] = node;
}
static void clear_defs(ir_node *node)
{
/* clear values defined */
- if (get_irn_mode(node) == mode_T) {
- const ir_edge_t *edge;
- foreach_out_edge(node, edge) {
- ir_node *proj = get_edge_src_irn(edge);
- clear_reg_value(proj);
- }
- } else {
- clear_reg_value(node);
- }
+ be_foreach_value(node, value,
+ clear_reg_value(value);
+ );
}
static void set_uses(ir_node *node)
}
}
-void be_peephole_new_node(ir_node * nw)
-{
- be_liveness_introduce(lv, nw);
-}
-
/**
* must be called from peephole optimisations before a node will be killed
* and its users will be redirected to new_node.
- * so bepeephole can update it's internal state.
+ * so bepeephole can update its internal state.
*
- * Note: killing a node and rewiring os only allowed if new_node produces
+ * Note: killing a node and rewiring is only allowed if new_node produces
* the same registers as old_node.
*/
static void be_peephole_before_exchange(const ir_node *old_node,
ir_node *new_node)
{
- const arch_register_t *reg;
- const arch_register_class_t *cls;
- unsigned reg_idx;
- unsigned cls_idx;
+ const arch_register_t *reg;
+ unsigned reg_idx;
+ bool old_is_current = false;
DB((dbg, LEVEL_1, "About to exchange and kill %+F with %+F\n", old_node, new_node));
+ assert(sched_is_scheduled(skip_Proj_const(old_node)));
+ assert(sched_is_scheduled(skip_Proj(new_node)));
+
if (current_node == old_node) {
+ old_is_current = true;
+
/* next node to be processed will be killed. Its scheduling predecessor
* must be processed next. */
current_node = sched_next(current_node);
assert (!is_Bad(current_node));
+
+ /* we can't handle liveness updates correctly when exchange current node
+ * with something behind it */
+ assert(value_dominates(skip_Proj(new_node), skip_Proj_const(old_node)));
}
if (!mode_is_data(get_irn_mode(old_node)))
assert(reg == arch_get_irn_register(new_node) &&
"KILLING a node and replacing by different register is not allowed");
- cls = arch_register_get_class(reg);
- reg_idx = arch_register_get_index(reg);
- cls_idx = arch_register_class_index(cls);
-
- if (register_values[cls_idx][reg_idx] == old_node) {
- register_values[cls_idx][reg_idx] = new_node;
+ reg_idx = reg->global_index;
+ if (register_values[reg_idx] == old_node || old_is_current) {
+ register_values[reg_idx] = new_node;
}
be_liveness_remove(lv, old_node);
be_peephole_before_exchange(old, nw);
sched_remove(old);
exchange(old, nw);
- be_peephole_new_node(nw);
+ be_liveness_introduce(lv, nw);
}
/**
*/
static void process_block(ir_node *block, void *data)
{
- unsigned n_classes;
- unsigned i;
- int l;
(void) data;
/* construct initial register assignment */
- n_classes = arch_env->n_register_classes;
- for (i = 0; i < n_classes; ++i) {
- const arch_register_class_t *cls = &arch_env->register_classes[i];
- unsigned n_regs = arch_register_class_n_regs(cls);
- memset(register_values[i], 0, sizeof(ir_node*) * n_regs);
- }
+ memset(register_values, 0, sizeof(ir_node*) * arch_env->n_registers);
- assert(lv->nodes && "live sets must be computed");
+ assert(lv->sets_valid && "live sets must be computed");
DB((dbg, LEVEL_1, "\nProcessing block %+F (from end)\n", block));
- be_lv_foreach(lv, block, be_lv_state_end, l) {
- ir_node *node = be_lv_get_irn(lv, block, l);
+ be_lv_foreach(lv, block, be_lv_state_end, node) {
set_reg_value(node);
}
DB((dbg, LEVEL_1, "\nstart processing\n"));
}
}
-static void kill_node_and_preds(ir_node *node)
+/**
+ * Check whether the node has only one user. Explicitly ignore the anchor.
+ */
+bool be_has_only_one_user(ir_node *node)
{
- ir_graph *irg = get_irn_irg(node);
- int arity, i;
+ int n = get_irn_n_edges(node);
+ int n_users;
- arity = get_irn_arity(node);
- for (i = 0; i < arity; ++i) {
- ir_node *pred = get_irn_n(node, i);
+ if (n <= 1)
+ return 1;
- set_irn_n(node, i, new_r_Bad(irg));
- if (get_irn_n_edges(pred) != 0)
+ n_users = 0;
+ foreach_out_edge(node, edge) {
+ ir_node *src = get_edge_src_irn(edge);
+ /* ignore anchor and keep-alive edges */
+ if (is_Anchor(src) || is_End(src))
continue;
-
- kill_node_and_preds(pred);
+ n_users++;
}
- if (!is_Proj(node))
- sched_remove(node);
- kill_node(node);
+ return n_users == 1;
}
-static void keep_alive_barrier_operand(ir_node* operand)
+static inline bool overlapping_regs(const arch_register_t *reg0,
+ const arch_register_req_t *req0, const arch_register_t *reg1,
+ const arch_register_req_t *req1)
{
- ir_node *schedpoint = skip_Proj(operand);
-
- do {
- schedpoint = sched_next(schedpoint);
- } while (is_Phi(schedpoint));
-
- /* There already is a keep in the schedule. */
- if (be_is_Keep(schedpoint)) {
- const arch_register_class_t *cls = arch_get_irn_reg_class_out(operand);
- be_Keep_add_node(schedpoint, cls, operand);
- } else {
- ir_node *block = get_nodes_block(operand);
- ir_node *in[1] = {operand};
- ir_node *keep = be_new_Keep(block, 1, in);
- sched_add_before(schedpoint, keep);
- }
+ if (reg0 == NULL || reg1 == NULL)
+ return false;
+ return reg0->global_index < (unsigned)reg1->global_index + req1->width
+ && reg1->global_index < (unsigned)reg0->global_index + req0->width;
}
-/**
- * Walk through the block schedule and skip all barrier nodes.
- */
-static void skip_barrier(ir_node *block, ir_graph *irg)
+bool be_can_move_down(ir_heights_t *heights, const ir_node *node,
+ const ir_node *before)
{
- ir_node *irn;
-
- sched_foreach_reverse(block, irn) {
- int arity;
- unsigned *used;
- size_t n_used;
- const ir_edge_t *edge, *next;
-
- if (!be_is_Barrier(irn))
- continue;
-
- /* track which outputs are actually used, as we have to create
- * keep nodes for unused outputs */
- arity = get_irn_arity(irn);
- rbitset_alloca(used, arity);
-
- foreach_out_edge_safe(irn, edge, next) {
- ir_node *proj = get_edge_src_irn(edge);
- int pn;
- ir_node *pred;
-
- if (is_Anchor(proj))
+ assert(get_nodes_block(node) == get_nodes_block(before));
+ assert(sched_get_time_step(node) < sched_get_time_step(before));
+
+ int node_arity = get_irn_arity(node);
+ ir_node *schedpoint = sched_next(node);
+
+ while (schedpoint != before) {
+ /* schedpoint must not use our computed values */
+ if (heights_reachable_in_block(heights, schedpoint, node))
+ return false;
+
+ /* schedpoint must not overwrite registers of our inputs */
+ for (int i = 0; i < node_arity; ++i) {
+ ir_node *in = get_irn_n(node, i);
+ const arch_register_t *reg = arch_get_irn_register(in);
+ if (reg == NULL)
continue;
-
- pn = (int) get_Proj_proj(proj);
- pred = get_irn_n(irn, pn);
-
- rbitset_set(used, pn);
-
- /* We may need to reschedule be_Keeps to keep live-ranges short. */
- if (get_irn_n_edges(proj) == 1) {
- const ir_edge_t *proj_edge = get_irn_out_edge_first(proj);
- ir_node *proj_succ = get_edge_src_irn(proj_edge);
-
- if (be_is_Keep(proj_succ)) {
- int succ_arity = get_irn_arity(proj_succ);
- ir_node *operand = get_irn_n(irn, pn);
-
- keep_alive_barrier_operand(operand);
-
- /* Disconnect old be_Keep. */
- if (succ_arity > 1) {
- int edge_pos = get_edge_src_pos(proj_edge);
- int new_arity = succ_arity - 1;
- int pos;
- int new_pos = 0;
- ir_node **ins;
-
- NEW_ARR_A(ir_node *, ins, succ_arity);
- for (pos = 0; pos < succ_arity; ++pos) {
- if (pos != edge_pos)
- ins[new_pos++] = get_irn_n(proj_succ, pos);
- }
-
- set_irn_in(proj_succ, new_arity, ins);
- }
- else {
- sched_remove(proj_succ);
- kill_node(proj_succ);
- }
- }
+ const arch_register_req_t *in_req
+ = arch_get_irn_register_req_in(node, i);
+ be_foreach_out(schedpoint, o) {
+ const arch_register_t *outreg
+ = arch_get_irn_register_out(schedpoint, o);
+ const arch_register_req_t *outreq
+ = arch_get_irn_register_req_out(schedpoint, o);
+ if (overlapping_regs(reg, in_req, outreg, outreq))
+ return false;
}
-
- edges_reroute_kind(proj, pred, EDGE_KIND_NORMAL, irg);
- edges_reroute_kind(proj, pred, EDGE_KIND_DEP, irg);
}
- /* the barrier also had the effect of a Keep for unused inputs.
- * we now have to create an explicit Keep for them */
- n_used = rbitset_popcount(used, arity);
- if (n_used < (size_t) arity) {
- int i;
-
- for (i = 0; i < arity; ++i) {
- ir_node *operand;
- if (rbitset_is_set(used, i))
- continue;
-
- operand = get_irn_n(irn, i);
- keep_alive_barrier_operand(operand);
- }
- }
-
- kill_node_and_preds(irn);
- break;
+ schedpoint = sched_next(schedpoint);
}
+ return true;
}
-/**
- * Kill the Barrier nodes for better peephole optimization.
- */
-static void kill_barriers(ir_graph *irg)
+bool be_can_move_up(ir_heights_t *heights, const ir_node *node,
+ const ir_node *after)
{
- ir_node *end_blk = get_irg_end_block(irg);
- ir_node *start_blk = get_irg_start_block(irg);
- int i;
+ const ir_node *node_block = get_nodes_block(node);
+ const ir_node *after_block = get_block_const(after);
+ const ir_node *schedpoint;
+ if (node_block != after_block) {
+ /* currently we can move up exactly 1 block */
+ assert(get_Block_cfgpred_block(node_block, 0) == after_block);
+ ir_node *first = sched_first(node_block);
+
+ /* do not move nodes changing memory */
+ if (is_memop(node)) {
+ ir_node *meminput = get_memop_mem(node);
+ if (!is_NoMem(meminput))
+ return false;
+ }
- /* skip the barrier on all return blocks */
- for (i = get_Block_n_cfgpreds(end_blk) - 1; i >= 0; --i) {
- ir_node *be_ret = get_Block_cfgpred(end_blk, i);
- ir_node *ret_blk = get_nodes_block(be_ret);
+ /* make sure we can move to the beginning of the succ block */
+ if (node != first && !be_can_move_up(heights, node, sched_prev(first)))
+ return false;
- if (ret_blk == start_blk)
- continue;
+ /* check if node overrides any of live-in values of other successors */
+ ir_graph *irg = get_irn_irg(node);
+ be_lv_t *lv = be_get_irg_liveness(irg);
+ foreach_block_succ(after_block, edge) {
+ ir_node *succ = get_edge_src_irn(edge);
+ if (succ == node_block)
+ continue;
- skip_barrier(ret_blk, irg);
+ be_lv_foreach(lv, succ, be_lv_state_in, live_node) {
+ const arch_register_t *reg = arch_get_irn_register(live_node);
+ const arch_register_req_t *req = arch_get_irn_register_req(live_node);
+ be_foreach_out(node, o) {
+ const arch_register_t *outreg
+ = arch_get_irn_register_out(node, o);
+ const arch_register_req_t *outreq
+ = arch_get_irn_register_req_out(node, o);
+ if (overlapping_regs(outreg, outreq, reg, req))
+ return false;
+ }
+ }
+ sched_foreach(succ, phi) {
+ if (!is_Phi(phi))
+ break;
+ const arch_register_t *reg = arch_get_irn_register(phi);
+ const arch_register_req_t *req = arch_get_irn_register_req(phi);
+ be_foreach_out(node, o) {
+ const arch_register_t *outreg
+ = arch_get_irn_register_out(node, o);
+ const arch_register_req_t *outreq
+ = arch_get_irn_register_req_out(node, o);
+ if (overlapping_regs(outreg, outreq, reg, req))
+ return false;
+ }
+ }
+ }
+ schedpoint = sched_last(after_block);
+ } else {
+ schedpoint = sched_prev(node);
}
- /* skip the barrier on the start block */
- start_blk = get_irg_start_block(irg);
- skip_barrier(start_blk, irg);
-}
-
-/**
- * Check whether the node has only one user. Explicitly ignore the anchor.
- */
-static int has_only_one_user(ir_node *node)
-{
- int n = get_irn_n_edges(node);
- const ir_edge_t *edge;
-
- if (n <= 1)
- return 1;
-
- if (n > 2)
- return 0;
+ /* move schedule upwards until we hit the "after" node */
+ while (schedpoint != after) {
+ /* TODO: the following heights query only works for nodes in the same
+ * block, otherwise we have to be conservative here */
+ if (get_nodes_block(node) != get_nodes_block(schedpoint))
+ return false;
+ /* node must not depend on schedpoint */
+ if (heights_reachable_in_block(heights, node, schedpoint))
+ return false;
+
+ /* node must not overwrite registers used by schedpoint */
+ int arity = get_irn_arity(schedpoint);
+ for (int i = 0; i < arity; ++i) {
+ const arch_register_t *reg
+ = arch_get_irn_register_in(schedpoint, i);
+ if (reg == NULL)
+ continue;
+ const arch_register_req_t *in_req
+ = arch_get_irn_register_req_in(schedpoint, i);
+ be_foreach_out(node, o) {
+ const arch_register_t *outreg
+ = arch_get_irn_register_out(node, o);
+ const arch_register_req_t *outreq
+ = arch_get_irn_register_req_out(node, o);
+ if (overlapping_regs(outreg, outreq, reg, in_req))
+ return false;
+ }
+ }
- foreach_out_edge(node, edge) {
- ir_node *src = get_edge_src_irn(edge);
- if (is_Anchor(src))
- return 1;
+ schedpoint = sched_prev(schedpoint);
}
-
- return 0;
+ return true;
}
/*
if (!be_is_IncSP(pred))
return node;
- if (!has_only_one_user(pred))
+ if (!be_has_only_one_user(pred))
return node;
pred_offs = be_get_IncSP_offset(pred);
curr_offs = be_get_IncSP_offset(node);
-
- if (pred_offs == BE_STACK_FRAME_SIZE_EXPAND) {
- if (curr_offs != BE_STACK_FRAME_SIZE_SHRINK) {
- return node;
- }
- offs = 0;
- } else if (pred_offs == BE_STACK_FRAME_SIZE_SHRINK) {
- if (curr_offs != BE_STACK_FRAME_SIZE_EXPAND) {
- return node;
- }
- offs = 0;
- } else if (curr_offs == BE_STACK_FRAME_SIZE_EXPAND ||
- curr_offs == BE_STACK_FRAME_SIZE_SHRINK) {
- return node;
- } else {
- offs = curr_offs + pred_offs;
- }
+ offs = curr_offs + pred_offs;
/* add node offset to pred and remove our IncSP */
be_set_IncSP_offset(pred, offs);
void be_peephole_opt(ir_graph *irg)
{
- unsigned n_classes;
- unsigned i;
-
- /* barrier nodes are used for register allocations. They hinders
- * peephole optimizations, so remove them here. */
- kill_barriers(irg);
-
- /* we sometimes find BadE nodes in float apps like optest_float.c or
- * kahansum.c for example... */
- be_liveness_invalidate(be_get_irg_liveness(irg));
- be_liveness_assure_sets(be_assure_liveness(irg));
+ be_assure_live_sets(irg);
arch_env = be_get_irg_arch_env(irg);
lv = be_get_irg_liveness(irg);
- n_classes = arch_env->n_register_classes;
- register_values = XMALLOCN(ir_node**, n_classes);
- for (i = 0; i < n_classes; ++i) {
- const arch_register_class_t *cls = &arch_env->register_classes[i];
- unsigned n_regs = arch_register_class_n_regs(cls);
- register_values[i] = XMALLOCN(ir_node*, n_regs);
- }
+ register_values = XMALLOCN(ir_node*, arch_env->n_registers);
irg_block_walk_graph(irg, process_block, NULL, NULL);
- for (i = 0; i < n_classes; ++i) {
- xfree(register_values[i]);
- }
xfree(register_values);
}
-BE_REGISTER_MODULE_CONSTRUCTOR(be_init_peephole);
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_peephole)
void be_init_peephole(void)
{
FIRM_DBG_REGISTER(dbg, "firm.be.peephole");
projects / libfirm / blobdiff