/* Perform simple optimizations to clean up the result of reload.
- Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998,
- 1999, 2000, 2001, 2002, 2003, 2004, 2005 Free Software Foundation, Inc.
+ Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997,
+ 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009,
+ 2010, 2011 Free Software Foundation, Inc.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
-Software Foundation; either version 2, or (at your option) any later
+Software Foundation; either version 3, or (at your option) any later
version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
for more details.
You should have received a copy of the GNU General Public License
-along with GCC; see the file COPYING. If not, write to the Free
-Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
-02110-1301, USA. */
+along with GCC; see the file COPYING3. If not see
+<http://www.gnu.org/licenses/>. */
#include "config.h"
#include "system.h"
#include "recog.h"
#include "output.h"
#include "cselib.h"
-#include "real.h"
-#include "toplev.h"
+#include "diagnostic-core.h"
#include "except.h"
#include "tree.h"
+#include "target.h"
#include "timevar.h"
#include "tree-pass.h"
+#include "df.h"
+#include "dbgcnt.h"
static int reload_cse_noop_set_p (rtx);
static void reload_cse_simplify (rtx, rtx);
static int reload_cse_simplify_operands (rtx, rtx);
static void reload_combine (void);
-static void reload_combine_note_use (rtx *, rtx);
-static void reload_combine_note_store (rtx, rtx, void *);
+static void reload_combine_note_use (rtx *, rtx, int, rtx);
+static void reload_combine_note_store (rtx, const_rtx, void *);
-static void reload_cse_move2add (rtx);
-static void move2add_note_store (rtx, rtx, void *);
+static bool reload_cse_move2add (rtx);
+static void move2add_note_store (rtx, const_rtx, void *);
/* Call cse / combine like post-reload optimization phases.
FIRST is the first instruction. */
void
reload_cse_regs (rtx first ATTRIBUTE_UNUSED)
{
+ bool moves_converted;
reload_cse_regs_1 (first);
reload_combine ();
- reload_cse_move2add (first);
+ moves_converted = reload_cse_move2add (first);
if (flag_expensive_optimizations)
- reload_cse_regs_1 (first);
+ {
+ if (moves_converted)
+ reload_combine ();
+ reload_cse_regs_1 (first);
+ }
}
/* See whether a single set SET is a noop. */
if (REG_P (value)
&& ! REG_FUNCTION_VALUE_P (value))
value = 0;
- delete_insn_and_edges (insn);
+ if (check_for_inc_dec (insn))
+ delete_insn_and_edges (insn);
return;
}
if (i < 0)
{
- delete_insn_and_edges (insn);
+ if (check_for_inc_dec (insn))
+ delete_insn_and_edges (insn);
/* We're done with this insn. */
return;
}
rtx insn;
rtx testreg = gen_rtx_REG (VOIDmode, -1);
- cselib_init (true);
+ cselib_init (CSELIB_RECORD_MEMORY);
init_alias_analysis ();
for (insn = first; insn; insn = NEXT_INSN (insn))
int did_change = 0;
int dreg;
rtx src;
- enum reg_class dclass;
+ reg_class_t dclass;
int old_cost;
cselib_val *val;
struct elt_loc_list *l;
#ifdef LOAD_EXTEND_OP
enum rtx_code extend_op = UNKNOWN;
#endif
+ bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn));
dreg = true_regnum (SET_DEST (set));
if (dreg < 0)
return 0;
#endif
- val = cselib_lookup (src, GET_MODE (SET_DEST (set)), 0);
+ val = cselib_lookup (src, GET_MODE (SET_DEST (set)), 0, VOIDmode);
if (! val)
return 0;
/* If memory loads are cheaper than register copies, don't change them. */
if (MEM_P (src))
- old_cost = MEMORY_MOVE_COST (GET_MODE (src), dclass, 1);
+ old_cost = memory_move_cost (GET_MODE (src), dclass, true);
else if (REG_P (src))
- old_cost = REGISTER_MOVE_COST (GET_MODE (src),
+ old_cost = register_move_cost (GET_MODE (src),
REGNO_REG_CLASS (REGNO (src)), dclass);
else
- old_cost = rtx_cost (src, SET);
+ old_cost = set_src_cost (src, speed);
for (l = val->locs; l; l = l->next)
{
/* ??? I'm lazy and don't wish to handle CONST_DOUBLE. Other
constants, such as SYMBOL_REF, cannot be extended. */
- if (GET_CODE (this_rtx) != CONST_INT)
+ if (!CONST_INT_P (this_rtx))
continue;
this_val = INTVAL (this_rtx);
this_rtx = GEN_INT (this_val);
}
#endif
- this_cost = rtx_cost (this_rtx, SET);
+ this_cost = set_src_cost (this_rtx, speed);
}
else if (REG_P (this_rtx))
{
if (extend_op != UNKNOWN)
{
this_rtx = gen_rtx_fmt_e (extend_op, word_mode, this_rtx);
- this_cost = rtx_cost (this_rtx, SET);
+ this_cost = set_src_cost (this_rtx, speed);
}
else
#endif
- this_cost = REGISTER_MOVE_COST (GET_MODE (this_rtx),
+ this_cost = register_move_cost (GET_MODE (this_rtx),
REGNO_REG_CLASS (REGNO (this_rtx)),
dclass);
}
}
#endif
- validate_change (insn, &SET_SRC (set), copy_rtx (this_rtx), 1);
+ validate_unshare_change (insn, &SET_SRC (set), this_rtx, 1);
old_cost = this_cost, did_change = 1;
}
}
if (! constrain_operands (1))
fatal_insn_not_found (insn);
- alternative_reject = alloca (recog_data.n_alternatives * sizeof (int));
- alternative_nregs = alloca (recog_data.n_alternatives * sizeof (int));
- alternative_order = alloca (recog_data.n_alternatives * sizeof (int));
+ alternative_reject = XALLOCAVEC (int, recog_data.n_alternatives);
+ alternative_nregs = XALLOCAVEC (int, recog_data.n_alternatives);
+ alternative_order = XALLOCAVEC (int, recog_data.n_alternatives);
memset (alternative_reject, 0, recog_data.n_alternatives * sizeof (int));
memset (alternative_nregs, 0, recog_data.n_alternatives * sizeof (int));
cselib_val *v;
struct elt_loc_list *l;
rtx op;
- enum machine_mode mode;
CLEAR_HARD_REG_SET (equiv_regs[i]);
continue;
op = recog_data.operand[i];
- mode = GET_MODE (op);
#ifdef LOAD_EXTEND_OP
if (MEM_P (op)
- && GET_MODE_BITSIZE (mode) < BITS_PER_WORD
- && LOAD_EXTEND_OP (mode) != UNKNOWN)
+ && GET_MODE_BITSIZE (GET_MODE (op)) < BITS_PER_WORD
+ && LOAD_EXTEND_OP (GET_MODE (op)) != UNKNOWN)
{
rtx set = single_set (insn);
&& SET_DEST (set) == recog_data.operand[1-i])
{
validate_change (insn, recog_data.operand_loc[i],
- gen_rtx_fmt_e (LOAD_EXTEND_OP (mode),
+ gen_rtx_fmt_e (LOAD_EXTEND_OP (GET_MODE (op)),
word_mode, op),
1);
validate_change (insn, recog_data.operand_loc[1-i],
continue;
}
#endif /* LOAD_EXTEND_OP */
- v = cselib_lookup (op, recog_data.operand_mode[i], 0);
+ if (side_effects_p (op))
+ continue;
+ v = cselib_lookup (op, recog_data.operand_mode[i], 0, VOIDmode);
if (! v)
continue;
int regno;
const char *p;
- op_alt_regno[i] = alloca (recog_data.n_alternatives * sizeof (int));
+ op_alt_regno[i] = XALLOCAVEC (int, recog_data.n_alternatives);
for (j = 0; j < recog_data.n_alternatives; j++)
op_alt_regno[i][j] = -1;
for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
{
- int class = (int) NO_REGS;
+ enum reg_class rclass = NO_REGS;
if (! TEST_HARD_REG_BIT (equiv_regs[i], regno))
continue;
- REGNO (testreg) = regno;
+ SET_REGNO_RAW (testreg, regno);
PUT_MODE (testreg, mode);
/* We found a register equal to this operand. Now look for all
case '*': case '%':
case '0': case '1': case '2': case '3': case '4':
case '5': case '6': case '7': case '8': case '9':
- case 'm': case '<': case '>': case 'V': case 'o':
+ case '<': case '>': case 'V': case 'o':
case 'E': case 'F': case 'G': case 'H':
case 's': case 'i': case 'n':
case 'I': case 'J': case 'K': case 'L':
case 'M': case 'N': case 'O': case 'P':
- case 'p': case 'X':
+ case 'p': case 'X': case TARGET_MEM_CONSTRAINT:
/* These don't say anything we care about. */
break;
case 'g': case 'r':
- class = reg_class_subunion[(int) class][(int) GENERAL_REGS];
+ rclass = reg_class_subunion[(int) rclass][(int) GENERAL_REGS];
break;
default:
- class
+ rclass
= (reg_class_subunion
- [(int) class]
+ [(int) rclass]
[(int) REG_CLASS_FROM_CONSTRAINT ((unsigned char) c, p)]);
break;
alternative yet and the operand being replaced is not
a cheap CONST_INT. */
if (op_alt_regno[i][j] == -1
- && reg_fits_class_p (testreg, class, 0, mode)
- && (GET_CODE (recog_data.operand[i]) != CONST_INT
- || (rtx_cost (recog_data.operand[i], SET)
- > rtx_cost (testreg, SET))))
+ && recog_data.alternative_enabled_p[j]
+ && reg_fits_class_p (testreg, rclass, 0, mode)
+ && (!CONST_INT_P (recog_data.operand[i])
+ || (set_src_cost (recog_data.operand[i],
+ optimize_bb_for_speed_p
+ (BLOCK_FOR_INSN (insn)))
+ > set_src_cost (testreg,
+ optimize_bb_for_speed_p
+ (BLOCK_FOR_INSN (insn))))))
{
alternative_nregs[j]++;
op_alt_regno[i][j] = regno;
}
j++;
- class = (int) NO_REGS;
+ rclass = NO_REGS;
break;
}
p += CONSTRAINT_LEN (c, p);
/* The maximum number of uses of a register we can keep track of to
replace them with reg+reg addressing. */
-#define RELOAD_COMBINE_MAX_USES 6
+#define RELOAD_COMBINE_MAX_USES 16
-/* INSN is the insn where a register has ben used, and USEP points to the
- location of the register within the rtl. */
-struct reg_use { rtx insn, *usep; };
+/* Describes a recorded use of a register. */
+struct reg_use
+{
+ /* The insn where a register has been used. */
+ rtx insn;
+ /* Points to the memory reference enclosing the use, if any, NULL_RTX
+ otherwise. */
+ rtx containing_mem;
+ /* Location of the register withing INSN. */
+ rtx *usep;
+ /* The reverse uid of the insn. */
+ int ruid;
+};
/* If the register is used in some unknown fashion, USE_INDEX is negative.
If it is dead, USE_INDEX is RELOAD_COMBINE_MAX_USES, and STORE_RUID
- indicates where it becomes live again.
+ indicates where it is first set or clobbered.
Otherwise, USE_INDEX is the index of the last encountered use of the
- register (which is first among these we have seen since we scan backwards),
- OFFSET contains the constant offset that is added to the register in
- all encountered uses, and USE_RUID indicates the first encountered, i.e.
- last, of these uses.
+ register (which is first among these we have seen since we scan backwards).
+ USE_RUID indicates the first encountered, i.e. last, of these uses.
+ If ALL_OFFSETS_MATCH is true, all encountered uses were inside a PLUS
+ with a constant offset; OFFSET contains this constant in that case.
STORE_RUID is always meaningful if we only want to use a value in a
register in a different place: it denotes the next insn in the insn
- stream (i.e. the last encountered) that sets or clobbers the register. */
+ stream (i.e. the last encountered) that sets or clobbers the register.
+ REAL_STORE_RUID is similar, but clobbers are ignored when updating it. */
static struct
{
struct reg_use reg_use[RELOAD_COMBINE_MAX_USES];
- int use_index;
rtx offset;
+ int use_index;
int store_ruid;
+ int real_store_ruid;
int use_ruid;
+ bool all_offsets_match;
} reg_state[FIRST_PSEUDO_REGISTER];
/* Reverse linear uid. This is increased in reload_combine while scanning
and the store_ruid / use_ruid fields in reg_state. */
static int reload_combine_ruid;
+/* The RUID of the last label we encountered in reload_combine. */
+static int last_label_ruid;
+
+/* The RUID of the last jump we encountered in reload_combine. */
+static int last_jump_ruid;
+
+/* The register numbers of the first and last index register. A value of
+ -1 in LAST_INDEX_REG indicates that we've previously computed these
+ values and found no suitable index registers. */
+static int first_index_reg = -1;
+static int last_index_reg;
+
#define LABEL_LIVE(LABEL) \
(label_live[CODE_LABEL_NUMBER (LABEL) - min_labelno])
+/* Subroutine of reload_combine_split_ruids, called to fix up a single
+ ruid pointed to by *PRUID if it is higher than SPLIT_RUID. */
+
+static inline void
+reload_combine_split_one_ruid (int *pruid, int split_ruid)
+{
+ if (*pruid > split_ruid)
+ (*pruid)++;
+}
+
+/* Called when we insert a new insn in a position we've already passed in
+ the scan. Examine all our state, increasing all ruids that are higher
+ than SPLIT_RUID by one in order to make room for a new insn. */
+
+static void
+reload_combine_split_ruids (int split_ruid)
+{
+ unsigned i;
+
+ reload_combine_split_one_ruid (&reload_combine_ruid, split_ruid);
+ reload_combine_split_one_ruid (&last_label_ruid, split_ruid);
+ reload_combine_split_one_ruid (&last_jump_ruid, split_ruid);
+
+ for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
+ {
+ int j, idx = reg_state[i].use_index;
+ reload_combine_split_one_ruid (®_state[i].use_ruid, split_ruid);
+ reload_combine_split_one_ruid (®_state[i].store_ruid, split_ruid);
+ reload_combine_split_one_ruid (®_state[i].real_store_ruid,
+ split_ruid);
+ if (idx < 0)
+ continue;
+ for (j = idx; j < RELOAD_COMBINE_MAX_USES; j++)
+ {
+ reload_combine_split_one_ruid (®_state[i].reg_use[j].ruid,
+ split_ruid);
+ }
+ }
+}
+
+/* Called when we are about to rescan a previously encountered insn with
+ reload_combine_note_use after modifying some part of it. This clears all
+ information about uses in that particular insn. */
+
+static void
+reload_combine_purge_insn_uses (rtx insn)
+{
+ unsigned i;
+
+ for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
+ {
+ int j, k, idx = reg_state[i].use_index;
+ if (idx < 0)
+ continue;
+ j = k = RELOAD_COMBINE_MAX_USES;
+ while (j-- > idx)
+ {
+ if (reg_state[i].reg_use[j].insn != insn)
+ {
+ k--;
+ if (k != j)
+ reg_state[i].reg_use[k] = reg_state[i].reg_use[j];
+ }
+ }
+ reg_state[i].use_index = k;
+ }
+}
+
+/* Called when we need to forget about all uses of REGNO after an insn
+ which is identified by RUID. */
+
+static void
+reload_combine_purge_reg_uses_after_ruid (unsigned regno, int ruid)
+{
+ int j, k, idx = reg_state[regno].use_index;
+ if (idx < 0)
+ return;
+ j = k = RELOAD_COMBINE_MAX_USES;
+ while (j-- > idx)
+ {
+ if (reg_state[regno].reg_use[j].ruid >= ruid)
+ {
+ k--;
+ if (k != j)
+ reg_state[regno].reg_use[k] = reg_state[regno].reg_use[j];
+ }
+ }
+ reg_state[regno].use_index = k;
+}
+
+/* Find the use of REGNO with the ruid that is highest among those
+ lower than RUID_LIMIT, and return it if it is the only use of this
+ reg in the insn. Return NULL otherwise. */
+
+static struct reg_use *
+reload_combine_closest_single_use (unsigned regno, int ruid_limit)
+{
+ int i, best_ruid = 0;
+ int use_idx = reg_state[regno].use_index;
+ struct reg_use *retval;
+
+ if (use_idx < 0)
+ return NULL;
+ retval = NULL;
+ for (i = use_idx; i < RELOAD_COMBINE_MAX_USES; i++)
+ {
+ struct reg_use *use = reg_state[regno].reg_use + i;
+ int this_ruid = use->ruid;
+ if (this_ruid >= ruid_limit)
+ continue;
+ if (this_ruid > best_ruid)
+ {
+ best_ruid = this_ruid;
+ retval = use;
+ }
+ else if (this_ruid == best_ruid)
+ retval = NULL;
+ }
+ if (last_label_ruid >= best_ruid)
+ return NULL;
+ return retval;
+}
+
+/* After we've moved an add insn, fix up any debug insns that occur
+ between the old location of the add and the new location. REG is
+ the destination register of the add insn; REPLACEMENT is the
+ SET_SRC of the add. FROM and TO specify the range in which we
+ should make this change on debug insns. */
+
+static void
+fixup_debug_insns (rtx reg, rtx replacement, rtx from, rtx to)
+{
+ rtx insn;
+ for (insn = from; insn != to; insn = NEXT_INSN (insn))
+ {
+ rtx t;
+
+ if (!DEBUG_INSN_P (insn))
+ continue;
+
+ t = INSN_VAR_LOCATION_LOC (insn);
+ t = simplify_replace_rtx (t, reg, replacement);
+ validate_change (insn, &INSN_VAR_LOCATION_LOC (insn), t, 0);
+ }
+}
+
+/* Subroutine of reload_combine_recognize_const_pattern. Try to replace REG
+ with SRC in the insn described by USE, taking costs into account. Return
+ true if we made the replacement. */
+
+static bool
+try_replace_in_use (struct reg_use *use, rtx reg, rtx src)
+{
+ rtx use_insn = use->insn;
+ rtx mem = use->containing_mem;
+ bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (use_insn));
+
+ if (mem != NULL_RTX)
+ {
+ addr_space_t as = MEM_ADDR_SPACE (mem);
+ rtx oldaddr = XEXP (mem, 0);
+ rtx newaddr = NULL_RTX;
+ int old_cost = address_cost (oldaddr, GET_MODE (mem), as, speed);
+ int new_cost;
+
+ newaddr = simplify_replace_rtx (oldaddr, reg, src);
+ if (memory_address_addr_space_p (GET_MODE (mem), newaddr, as))
+ {
+ XEXP (mem, 0) = newaddr;
+ new_cost = address_cost (newaddr, GET_MODE (mem), as, speed);
+ XEXP (mem, 0) = oldaddr;
+ if (new_cost <= old_cost
+ && validate_change (use_insn,
+ &XEXP (mem, 0), newaddr, 0))
+ return true;
+ }
+ }
+ else
+ {
+ rtx new_set = single_set (use_insn);
+ if (new_set
+ && REG_P (SET_DEST (new_set))
+ && GET_CODE (SET_SRC (new_set)) == PLUS
+ && REG_P (XEXP (SET_SRC (new_set), 0))
+ && CONSTANT_P (XEXP (SET_SRC (new_set), 1)))
+ {
+ rtx new_src;
+ int old_cost = set_src_cost (SET_SRC (new_set), speed);
+
+ gcc_assert (rtx_equal_p (XEXP (SET_SRC (new_set), 0), reg));
+ new_src = simplify_replace_rtx (SET_SRC (new_set), reg, src);
+
+ if (set_src_cost (new_src, speed) <= old_cost
+ && validate_change (use_insn, &SET_SRC (new_set),
+ new_src, 0))
+ return true;
+ }
+ }
+ return false;
+}
+
+/* Called by reload_combine when scanning INSN. This function tries to detect
+ patterns where a constant is added to a register, and the result is used
+ in an address.
+ Return true if no further processing is needed on INSN; false if it wasn't
+ recognized and should be handled normally. */
+
+static bool
+reload_combine_recognize_const_pattern (rtx insn)
+{
+ int from_ruid = reload_combine_ruid;
+ rtx set, pat, reg, src, addreg;
+ unsigned int regno;
+ struct reg_use *use;
+ bool must_move_add;
+ rtx add_moved_after_insn = NULL_RTX;
+ int add_moved_after_ruid = 0;
+ int clobbered_regno = -1;
+
+ set = single_set (insn);
+ if (set == NULL_RTX)
+ return false;
+
+ reg = SET_DEST (set);
+ src = SET_SRC (set);
+ if (!REG_P (reg)
+ || hard_regno_nregs[REGNO (reg)][GET_MODE (reg)] != 1
+ || GET_MODE (reg) != Pmode
+ || reg == stack_pointer_rtx)
+ return false;
+
+ regno = REGNO (reg);
+
+ /* We look for a REG1 = REG2 + CONSTANT insn, followed by either
+ uses of REG1 inside an address, or inside another add insn. If
+ possible and profitable, merge the addition into subsequent
+ uses. */
+ if (GET_CODE (src) != PLUS
+ || !REG_P (XEXP (src, 0))
+ || !CONSTANT_P (XEXP (src, 1)))
+ return false;
+
+ addreg = XEXP (src, 0);
+ must_move_add = rtx_equal_p (reg, addreg);
+
+ pat = PATTERN (insn);
+ if (must_move_add && set != pat)
+ {
+ /* We have to be careful when moving the add; apart from the
+ single_set there may also be clobbers. Recognize one special
+ case, that of one clobber alongside the set (likely a clobber
+ of the CC register). */
+ gcc_assert (GET_CODE (PATTERN (insn)) == PARALLEL);
+ if (XVECLEN (pat, 0) != 2 || XVECEXP (pat, 0, 0) != set
+ || GET_CODE (XVECEXP (pat, 0, 1)) != CLOBBER
+ || !REG_P (XEXP (XVECEXP (pat, 0, 1), 0)))
+ return false;
+ clobbered_regno = REGNO (XEXP (XVECEXP (pat, 0, 1), 0));
+ }
+
+ do
+ {
+ use = reload_combine_closest_single_use (regno, from_ruid);
+
+ if (use)
+ /* Start the search for the next use from here. */
+ from_ruid = use->ruid;
+
+ if (use && GET_MODE (*use->usep) == Pmode)
+ {
+ bool delete_add = false;
+ rtx use_insn = use->insn;
+ int use_ruid = use->ruid;
+
+ /* Avoid moving the add insn past a jump. */
+ if (must_move_add && use_ruid <= last_jump_ruid)
+ break;
+
+ /* If the add clobbers another hard reg in parallel, don't move
+ it past a real set of this hard reg. */
+ if (must_move_add && clobbered_regno >= 0
+ && reg_state[clobbered_regno].real_store_ruid >= use_ruid)
+ break;
+
+#ifdef HAVE_cc0
+ /* Do not separate cc0 setter and cc0 user on HAVE_cc0 targets. */
+ if (must_move_add && sets_cc0_p (PATTERN (use_insn)))
+ break;
+#endif
+
+ gcc_assert (reg_state[regno].store_ruid <= use_ruid);
+ /* Avoid moving a use of ADDREG past a point where it is stored. */
+ if (reg_state[REGNO (addreg)].store_ruid > use_ruid)
+ break;
+
+ /* We also must not move the addition past an insn that sets
+ the same register, unless we can combine two add insns. */
+ if (must_move_add && reg_state[regno].store_ruid == use_ruid)
+ {
+ if (use->containing_mem == NULL_RTX)
+ delete_add = true;
+ else
+ break;
+ }
+
+ if (try_replace_in_use (use, reg, src))
+ {
+ reload_combine_purge_insn_uses (use_insn);
+ reload_combine_note_use (&PATTERN (use_insn), use_insn,
+ use_ruid, NULL_RTX);
+
+ if (delete_add)
+ {
+ fixup_debug_insns (reg, src, insn, use_insn);
+ delete_insn (insn);
+ return true;
+ }
+ if (must_move_add)
+ {
+ add_moved_after_insn = use_insn;
+ add_moved_after_ruid = use_ruid;
+ }
+ continue;
+ }
+ }
+ /* If we get here, we couldn't handle this use. */
+ if (must_move_add)
+ break;
+ }
+ while (use);
+
+ if (!must_move_add || add_moved_after_insn == NULL_RTX)
+ /* Process the add normally. */
+ return false;
+
+ fixup_debug_insns (reg, src, insn, add_moved_after_insn);
+
+ reorder_insns (insn, insn, add_moved_after_insn);
+ reload_combine_purge_reg_uses_after_ruid (regno, add_moved_after_ruid);
+ reload_combine_split_ruids (add_moved_after_ruid - 1);
+ reload_combine_note_use (&PATTERN (insn), insn,
+ add_moved_after_ruid, NULL_RTX);
+ reg_state[regno].store_ruid = add_moved_after_ruid;
+
+ return true;
+}
+
+/* Called by reload_combine when scanning INSN. Try to detect a pattern we
+ can handle and improve. Return true if no further processing is needed on
+ INSN; false if it wasn't recognized and should be handled normally. */
+
+static bool
+reload_combine_recognize_pattern (rtx insn)
+{
+ rtx set, reg, src;
+ unsigned int regno;
+
+ set = single_set (insn);
+ if (set == NULL_RTX)
+ return false;
+
+ reg = SET_DEST (set);
+ src = SET_SRC (set);
+ if (!REG_P (reg)
+ || hard_regno_nregs[REGNO (reg)][GET_MODE (reg)] != 1)
+ return false;
+
+ regno = REGNO (reg);
+
+ /* Look for (set (REGX) (CONST_INT))
+ (set (REGX) (PLUS (REGX) (REGY)))
+ ...
+ ... (MEM (REGX)) ...
+ and convert it to
+ (set (REGZ) (CONST_INT))
+ ...
+ ... (MEM (PLUS (REGZ) (REGY)))... .
+
+ First, check that we have (set (REGX) (PLUS (REGX) (REGY)))
+ and that we know all uses of REGX before it dies.
+ Also, explicitly check that REGX != REGY; our life information
+ does not yet show whether REGY changes in this insn. */
+
+ if (GET_CODE (src) == PLUS
+ && reg_state[regno].all_offsets_match
+ && last_index_reg != -1
+ && REG_P (XEXP (src, 1))
+ && rtx_equal_p (XEXP (src, 0), reg)
+ && !rtx_equal_p (XEXP (src, 1), reg)
+ && reg_state[regno].use_index >= 0
+ && reg_state[regno].use_index < RELOAD_COMBINE_MAX_USES
+ && last_label_ruid < reg_state[regno].use_ruid)
+ {
+ rtx base = XEXP (src, 1);
+ rtx prev = prev_nonnote_nondebug_insn (insn);
+ rtx prev_set = prev ? single_set (prev) : NULL_RTX;
+ rtx index_reg = NULL_RTX;
+ rtx reg_sum = NULL_RTX;
+ int i;
+
+ /* Now we need to set INDEX_REG to an index register (denoted as
+ REGZ in the illustration above) and REG_SUM to the expression
+ register+register that we want to use to substitute uses of REG
+ (typically in MEMs) with. First check REG and BASE for being
+ index registers; we can use them even if they are not dead. */
+ if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], regno)
+ || TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS],
+ REGNO (base)))
+ {
+ index_reg = reg;
+ reg_sum = src;
+ }
+ else
+ {
+ /* Otherwise, look for a free index register. Since we have
+ checked above that neither REG nor BASE are index registers,
+ if we find anything at all, it will be different from these
+ two registers. */
+ for (i = first_index_reg; i <= last_index_reg; i++)
+ {
+ if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], i)
+ && reg_state[i].use_index == RELOAD_COMBINE_MAX_USES
+ && reg_state[i].store_ruid <= reg_state[regno].use_ruid
+ && (call_used_regs[i] || df_regs_ever_live_p (i))
+ && (!frame_pointer_needed || i != HARD_FRAME_POINTER_REGNUM)
+ && !fixed_regs[i] && !global_regs[i]
+ && hard_regno_nregs[i][GET_MODE (reg)] == 1
+ && targetm.hard_regno_scratch_ok (i))
+ {
+ index_reg = gen_rtx_REG (GET_MODE (reg), i);
+ reg_sum = gen_rtx_PLUS (GET_MODE (reg), index_reg, base);
+ break;
+ }
+ }
+ }
+
+ /* Check that PREV_SET is indeed (set (REGX) (CONST_INT)) and that
+ (REGY), i.e. BASE, is not clobbered before the last use we'll
+ create. */
+ if (reg_sum
+ && prev_set
+ && CONST_INT_P (SET_SRC (prev_set))
+ && rtx_equal_p (SET_DEST (prev_set), reg)
+ && (reg_state[REGNO (base)].store_ruid
+ <= reg_state[regno].use_ruid))
+ {
+ /* Change destination register and, if necessary, the constant
+ value in PREV, the constant loading instruction. */
+ validate_change (prev, &SET_DEST (prev_set), index_reg, 1);
+ if (reg_state[regno].offset != const0_rtx)
+ validate_change (prev,
+ &SET_SRC (prev_set),
+ GEN_INT (INTVAL (SET_SRC (prev_set))
+ + INTVAL (reg_state[regno].offset)),
+ 1);
+
+ /* Now for every use of REG that we have recorded, replace REG
+ with REG_SUM. */
+ for (i = reg_state[regno].use_index;
+ i < RELOAD_COMBINE_MAX_USES; i++)
+ validate_unshare_change (reg_state[regno].reg_use[i].insn,
+ reg_state[regno].reg_use[i].usep,
+ /* Each change must have its own
+ replacement. */
+ reg_sum, 1);
+
+ if (apply_change_group ())
+ {
+ struct reg_use *lowest_ruid = NULL;
+
+ /* For every new use of REG_SUM, we have to record the use
+ of BASE therein, i.e. operand 1. */
+ for (i = reg_state[regno].use_index;
+ i < RELOAD_COMBINE_MAX_USES; i++)
+ {
+ struct reg_use *use = reg_state[regno].reg_use + i;
+ reload_combine_note_use (&XEXP (*use->usep, 1), use->insn,
+ use->ruid, use->containing_mem);
+ if (lowest_ruid == NULL || use->ruid < lowest_ruid->ruid)
+ lowest_ruid = use;
+ }
+
+ fixup_debug_insns (reg, reg_sum, insn, lowest_ruid->insn);
+
+ /* Delete the reg-reg addition. */
+ delete_insn (insn);
+
+ if (reg_state[regno].offset != const0_rtx)
+ /* Previous REG_EQUIV / REG_EQUAL notes for PREV
+ are now invalid. */
+ remove_reg_equal_equiv_notes (prev);
+
+ reg_state[regno].use_index = RELOAD_COMBINE_MAX_USES;
+ return true;
+ }
+ }
+ }
+ return false;
+}
+
static void
reload_combine (void)
{
- rtx insn, set;
- int first_index_reg = -1;
- int last_index_reg = 0;
- int i;
+ rtx insn, prev;
basic_block bb;
unsigned int r;
- int last_label_ruid;
int min_labelno, n_labels;
HARD_REG_SET ever_live_at_start, *label_live;
- /* If reg+reg can be used in offsetable memory addresses, the main chunk of
- reload has already used it where appropriate, so there is no use in
- trying to generate it now. */
- if (double_reg_address_ok && INDEX_REG_CLASS != NO_REGS)
- return;
-
/* To avoid wasting too much time later searching for an index register,
determine the minimum and maximum index register numbers. */
- for (r = 0; r < FIRST_PSEUDO_REGISTER; r++)
- if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], r))
- {
- if (first_index_reg == -1)
- first_index_reg = r;
+ if (INDEX_REG_CLASS == NO_REGS)
+ last_index_reg = -1;
+ else if (first_index_reg == -1 && last_index_reg == 0)
+ {
+ for (r = 0; r < FIRST_PSEUDO_REGISTER; r++)
+ if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], r))
+ {
+ if (first_index_reg == -1)
+ first_index_reg = r;
- last_index_reg = r;
- }
+ last_index_reg = r;
+ }
- /* If no index register is available, we can quit now. */
- if (first_index_reg == -1)
- return;
+ /* If no index register is available, we can quit now. Set LAST_INDEX_REG
+ to -1 so we'll know to quit early the next time we get here. */
+ if (first_index_reg == -1)
+ {
+ last_index_reg = -1;
+ return;
+ }
+ }
/* Set up LABEL_LIVE and EVER_LIVE_AT_START. The register lifetime
information is a bit fuzzy immediately after reload, but it's
if (LABEL_P (insn))
{
HARD_REG_SET live;
+ bitmap live_in = df_get_live_in (bb);
- REG_SET_TO_HARD_REG_SET (live,
- bb->il.rtl->global_live_at_start);
- compute_use_by_pseudos (&live,
- bb->il.rtl->global_live_at_start);
+ REG_SET_TO_HARD_REG_SET (live, live_in);
+ compute_use_by_pseudos (&live, live_in);
COPY_HARD_REG_SET (LABEL_LIVE (insn), live);
IOR_HARD_REG_SET (ever_live_at_start, live);
}
}
/* Initialize last_label_ruid, reload_combine_ruid and reg_state. */
- last_label_ruid = reload_combine_ruid = 0;
+ last_label_ruid = last_jump_ruid = reload_combine_ruid = 0;
for (r = 0; r < FIRST_PSEUDO_REGISTER; r++)
{
- reg_state[r].store_ruid = reload_combine_ruid;
+ reg_state[r].store_ruid = 0;
+ reg_state[r].real_store_ruid = 0;
if (fixed_regs[r])
reg_state[r].use_index = -1;
else
reg_state[r].use_index = RELOAD_COMBINE_MAX_USES;
}
- for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
+ for (insn = get_last_insn (); insn; insn = prev)
{
+ bool control_flow_insn;
rtx note;
+ prev = PREV_INSN (insn);
+
/* We cannot do our optimization across labels. Invalidating all the use
information we have would be costly, so we just note where the label
is and then later disable any optimization that would cross it. */
if (LABEL_P (insn))
last_label_ruid = reload_combine_ruid;
else if (BARRIER_P (insn))
- for (r = 0; r < FIRST_PSEUDO_REGISTER; r++)
- if (! fixed_regs[r])
+ {
+ /* Crossing a barrier resets all the use information. */
+ for (r = 0; r < FIRST_PSEUDO_REGISTER; r++)
+ if (! fixed_regs[r])
reg_state[r].use_index = RELOAD_COMBINE_MAX_USES;
+ }
+ else if (INSN_P (insn) && volatile_insn_p (PATTERN (insn)))
+ /* Optimizations across insns being marked as volatile must be
+ prevented. All the usage information is invalidated
+ here. */
+ for (r = 0; r < FIRST_PSEUDO_REGISTER; r++)
+ if (! fixed_regs[r]
+ && reg_state[r].use_index != RELOAD_COMBINE_MAX_USES)
+ reg_state[r].use_index = -1;
- if (! INSN_P (insn))
+ if (! NONDEBUG_INSN_P (insn))
continue;
reload_combine_ruid++;
- /* Look for (set (REGX) (CONST_INT))
- (set (REGX) (PLUS (REGX) (REGY)))
- ...
- ... (MEM (REGX)) ...
- and convert it to
- (set (REGZ) (CONST_INT))
- ...
- ... (MEM (PLUS (REGZ) (REGY)))... .
-
- First, check that we have (set (REGX) (PLUS (REGX) (REGY)))
- and that we know all uses of REGX before it dies.
- Also, explicitly check that REGX != REGY; our life information
- does not yet show whether REGY changes in this insn. */
- set = single_set (insn);
- if (set != NULL_RTX
- && REG_P (SET_DEST (set))
- && (hard_regno_nregs[REGNO (SET_DEST (set))]
- [GET_MODE (SET_DEST (set))]
- == 1)
- && GET_CODE (SET_SRC (set)) == PLUS
- && REG_P (XEXP (SET_SRC (set), 1))
- && rtx_equal_p (XEXP (SET_SRC (set), 0), SET_DEST (set))
- && !rtx_equal_p (XEXP (SET_SRC (set), 1), SET_DEST (set))
- && last_label_ruid < reg_state[REGNO (SET_DEST (set))].use_ruid)
- {
- rtx reg = SET_DEST (set);
- rtx plus = SET_SRC (set);
- rtx base = XEXP (plus, 1);
- rtx prev = prev_nonnote_insn (insn);
- rtx prev_set = prev ? single_set (prev) : NULL_RTX;
- unsigned int regno = REGNO (reg);
- rtx const_reg = NULL_RTX;
- rtx reg_sum = NULL_RTX;
-
- /* Now, we need an index register.
- We'll set index_reg to this index register, const_reg to the
- register that is to be loaded with the constant
- (denoted as REGZ in the substitution illustration above),
- and reg_sum to the register-register that we want to use to
- substitute uses of REG (typically in MEMs) with.
- First check REG and BASE for being index registers;
- we can use them even if they are not dead. */
- if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], regno)
- || TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS],
- REGNO (base)))
- {
- const_reg = reg;
- reg_sum = plus;
- }
- else
- {
- /* Otherwise, look for a free index register. Since we have
- checked above that neither REG nor BASE are index registers,
- if we find anything at all, it will be different from these
- two registers. */
- for (i = first_index_reg; i <= last_index_reg; i++)
- {
- if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS],
- i)
- && reg_state[i].use_index == RELOAD_COMBINE_MAX_USES
- && reg_state[i].store_ruid <= reg_state[regno].use_ruid
- && hard_regno_nregs[i][GET_MODE (reg)] == 1)
- {
- rtx index_reg = gen_rtx_REG (GET_MODE (reg), i);
+ control_flow_insn = control_flow_insn_p (insn);
+ if (control_flow_insn)
+ last_jump_ruid = reload_combine_ruid;
- const_reg = index_reg;
- reg_sum = gen_rtx_PLUS (GET_MODE (reg), index_reg, base);
- break;
- }
- }
- }
-
- /* Check that PREV_SET is indeed (set (REGX) (CONST_INT)) and that
- (REGY), i.e. BASE, is not clobbered before the last use we'll
- create. */
- if (prev_set != 0
- && GET_CODE (SET_SRC (prev_set)) == CONST_INT
- && rtx_equal_p (SET_DEST (prev_set), reg)
- && reg_state[regno].use_index >= 0
- && (reg_state[REGNO (base)].store_ruid
- <= reg_state[regno].use_ruid)
- && reg_sum != 0)
- {
- int i;
-
- /* Change destination register and, if necessary, the
- constant value in PREV, the constant loading instruction. */
- validate_change (prev, &SET_DEST (prev_set), const_reg, 1);
- if (reg_state[regno].offset != const0_rtx)
- validate_change (prev,
- &SET_SRC (prev_set),
- GEN_INT (INTVAL (SET_SRC (prev_set))
- + INTVAL (reg_state[regno].offset)),
- 1);
-
- /* Now for every use of REG that we have recorded, replace REG
- with REG_SUM. */
- for (i = reg_state[regno].use_index;
- i < RELOAD_COMBINE_MAX_USES; i++)
- validate_change (reg_state[regno].reg_use[i].insn,
- reg_state[regno].reg_use[i].usep,
- /* Each change must have its own
- replacement. */
- copy_rtx (reg_sum), 1);
-
- if (apply_change_group ())
- {
- rtx *np;
-
- /* Delete the reg-reg addition. */
- delete_insn (insn);
-
- if (reg_state[regno].offset != const0_rtx)
- /* Previous REG_EQUIV / REG_EQUAL notes for PREV
- are now invalid. */
- for (np = ®_NOTES (prev); *np;)
- {
- if (REG_NOTE_KIND (*np) == REG_EQUAL
- || REG_NOTE_KIND (*np) == REG_EQUIV)
- *np = XEXP (*np, 1);
- else
- np = &XEXP (*np, 1);
- }
-
- reg_state[regno].use_index = RELOAD_COMBINE_MAX_USES;
- reg_state[REGNO (const_reg)].store_ruid
- = reload_combine_ruid;
- continue;
- }
- }
- }
+ if (reload_combine_recognize_const_pattern (insn)
+ || reload_combine_recognize_pattern (insn))
+ continue;
note_stores (PATTERN (insn), reload_combine_note_store, NULL);
{
unsigned int i;
unsigned int start_reg = REGNO (usage_rtx);
- unsigned int num_regs =
- hard_regno_nregs[start_reg][GET_MODE (usage_rtx)];
- unsigned int end_reg = start_reg + num_regs - 1;
+ unsigned int num_regs
+ = hard_regno_nregs[start_reg][GET_MODE (usage_rtx)];
+ unsigned int end_reg = start_reg + num_regs - 1;
for (i = start_reg; i <= end_reg; i++)
if (GET_CODE (XEXP (link, 0)) == CLOBBER)
{
reg_state[i].use_index = -1;
}
}
-
}
- else if (JUMP_P (insn)
- && GET_CODE (PATTERN (insn)) != RETURN)
+
+ if (control_flow_insn && GET_CODE (PATTERN (insn)) != RETURN)
{
/* Non-spill registers might be used at the call destination in
some unknown fashion, so we have to mark the unknown use. */
else
live = &ever_live_at_start;
- for (i = FIRST_PSEUDO_REGISTER - 1; i >= 0; --i)
- if (TEST_HARD_REG_BIT (*live, i))
- reg_state[i].use_index = -1;
+ for (r = 0; r < FIRST_PSEUDO_REGISTER; r++)
+ if (TEST_HARD_REG_BIT (*live, r))
+ reg_state[r].use_index = -1;
}
- reload_combine_note_use (&PATTERN (insn), insn);
+ reload_combine_note_use (&PATTERN (insn), insn, reload_combine_ruid,
+ NULL_RTX);
+
for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
{
- if (REG_NOTE_KIND (note) == REG_INC
- && REG_P (XEXP (note, 0)))
+ if (REG_NOTE_KIND (note) == REG_INC && REG_P (XEXP (note, 0)))
{
int regno = REGNO (XEXP (note, 0));
-
reg_state[regno].store_ruid = reload_combine_ruid;
+ reg_state[regno].real_store_ruid = reload_combine_ruid;
reg_state[regno].use_index = -1;
}
}
}
/* Check if DST is a register or a subreg of a register; if it is,
- update reg_state[regno].store_ruid and reg_state[regno].use_index
- accordingly. Called via note_stores from reload_combine. */
+ update store_ruid, real_store_ruid and use_index in the reg_state
+ structure accordingly. Called via note_stores from reload_combine. */
static void
-reload_combine_note_store (rtx dst, rtx set, void *data ATTRIBUTE_UNUSED)
+reload_combine_note_store (rtx dst, const_rtx set, void *data ATTRIBUTE_UNUSED)
{
int regno = 0;
int i;
GET_MODE (dst));
dst = SUBREG_REG (dst);
}
+
+ /* Some targets do argument pushes without adding REG_INC notes. */
+
+ if (MEM_P (dst))
+ {
+ dst = XEXP (dst, 0);
+ if (GET_CODE (dst) == PRE_INC || GET_CODE (dst) == POST_INC
+ || GET_CODE (dst) == PRE_DEC || GET_CODE (dst) == POST_DEC
+ || GET_CODE (dst) == PRE_MODIFY || GET_CODE (dst) == POST_MODIFY)
+ {
+ regno = REGNO (XEXP (dst, 0));
+ mode = GET_MODE (XEXP (dst, 0));
+ for (i = hard_regno_nregs[regno][mode] - 1 + regno; i >= regno; i--)
+ {
+ /* We could probably do better, but for now mark the register
+ as used in an unknown fashion and set/clobbered at this
+ insn. */
+ reg_state[i].use_index = -1;
+ reg_state[i].store_ruid = reload_combine_ruid;
+ reg_state[i].real_store_ruid = reload_combine_ruid;
+ }
+ }
+ else
+ return;
+ }
+
if (!REG_P (dst))
return;
regno += REGNO (dst);
/* note_stores might have stripped a STRICT_LOW_PART, so we have to be
careful with registers / register parts that are not full words.
Similarly for ZERO_EXTRACT. */
- if (GET_CODE (set) != SET
- || GET_CODE (SET_DEST (set)) == ZERO_EXTRACT
+ if (GET_CODE (SET_DEST (set)) == ZERO_EXTRACT
|| GET_CODE (SET_DEST (set)) == STRICT_LOW_PART)
{
for (i = hard_regno_nregs[regno][mode] - 1 + regno; i >= regno; i--)
{
reg_state[i].use_index = -1;
reg_state[i].store_ruid = reload_combine_ruid;
+ reg_state[i].real_store_ruid = reload_combine_ruid;
}
}
else
for (i = hard_regno_nregs[regno][mode] - 1 + regno; i >= regno; i--)
{
reg_state[i].store_ruid = reload_combine_ruid;
+ if (GET_CODE (set) == SET)
+ reg_state[i].real_store_ruid = reload_combine_ruid;
reg_state[i].use_index = RELOAD_COMBINE_MAX_USES;
}
}
*XP is the pattern of INSN, or a part of it.
Called from reload_combine, and recursively by itself. */
static void
-reload_combine_note_use (rtx *xp, rtx insn)
+reload_combine_note_use (rtx *xp, rtx insn, int ruid, rtx containing_mem)
{
rtx x = *xp;
enum rtx_code code = x->code;
case SET:
if (REG_P (SET_DEST (x)))
{
- reload_combine_note_use (&SET_SRC (x), insn);
+ reload_combine_note_use (&SET_SRC (x), insn, ruid, NULL_RTX);
return;
}
break;
case PLUS:
/* We are interested in (plus (reg) (const_int)) . */
if (!REG_P (XEXP (x, 0))
- || GET_CODE (XEXP (x, 1)) != CONST_INT)
+ || !CONST_INT_P (XEXP (x, 1)))
break;
offset = XEXP (x, 1);
x = XEXP (x, 0);
return;
}
+ /* We may be called to update uses in previously seen insns.
+ Don't add uses beyond the last store we saw. */
+ if (ruid < reg_state[regno].store_ruid)
+ return;
+
/* If this register is already used in some unknown fashion, we
can't do anything.
If we decrement the index from zero to -1, we can't store more
if (use_index < 0)
return;
- if (use_index != RELOAD_COMBINE_MAX_USES - 1)
- {
- /* We have found another use for a register that is already
- used later. Check if the offsets match; if not, mark the
- register as used in an unknown fashion. */
- if (! rtx_equal_p (offset, reg_state[regno].offset))
- {
- reg_state[regno].use_index = -1;
- return;
- }
- }
- else
+ if (use_index == RELOAD_COMBINE_MAX_USES - 1)
{
/* This is the first use of this register we have seen since we
marked it as dead. */
reg_state[regno].offset = offset;
- reg_state[regno].use_ruid = reload_combine_ruid;
+ reg_state[regno].all_offsets_match = true;
+ reg_state[regno].use_ruid = ruid;
}
+ else
+ {
+ if (reg_state[regno].use_ruid > ruid)
+ reg_state[regno].use_ruid = ruid;
+
+ if (! rtx_equal_p (offset, reg_state[regno].offset))
+ reg_state[regno].all_offsets_match = false;
+ }
+
reg_state[regno].reg_use[use_index].insn = insn;
+ reg_state[regno].reg_use[use_index].ruid = ruid;
+ reg_state[regno].reg_use[use_index].containing_mem = containing_mem;
reg_state[regno].reg_use[use_index].usep = xp;
return;
}
+ case MEM:
+ containing_mem = x;
+ break;
+
default:
break;
}
for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
{
if (fmt[i] == 'e')
- reload_combine_note_use (&XEXP (x, i), insn);
+ reload_combine_note_use (&XEXP (x, i), insn, ruid, containing_mem);
else if (fmt[i] == 'E')
{
for (j = XVECLEN (x, i) - 1; j >= 0; j--)
- reload_combine_note_use (&XVECEXP (x, i, j), insn);
+ reload_combine_note_use (&XVECEXP (x, i, j), insn, ruid,
+ containing_mem);
}
}
}
information about register contents we have would be costly, so we
use move2add_last_label_luid to note where the label is and then
later disable any optimization that would cross it.
- reg_offset[n] / reg_base_reg[n] / reg_mode[n] are only valid if
- reg_set_luid[n] is greater than move2add_last_label_luid. */
+ reg_offset[n] / reg_base_reg[n] / reg_symbol_ref[n] / reg_mode[n]
+ are only valid if reg_set_luid[n] is greater than
+ move2add_last_label_luid. */
static int reg_set_luid[FIRST_PSEUDO_REGISTER];
/* If reg_base_reg[n] is negative, register n has been set to
- reg_offset[n] in mode reg_mode[n] .
+ reg_offset[n] or reg_symbol_ref[n] + reg_offset[n] in mode reg_mode[n].
If reg_base_reg[n] is non-negative, register n has been set to the
sum of reg_offset[n] and the value of register reg_base_reg[n]
before reg_set_luid[n], calculated in mode reg_mode[n] . */
static HOST_WIDE_INT reg_offset[FIRST_PSEUDO_REGISTER];
static int reg_base_reg[FIRST_PSEUDO_REGISTER];
+static rtx reg_symbol_ref[FIRST_PSEUDO_REGISTER];
static enum machine_mode reg_mode[FIRST_PSEUDO_REGISTER];
/* move2add_luid is linearly increased while scanning the instructions
#define MODES_OK_FOR_MOVE2ADD(OUTMODE, INMODE) \
(GET_MODE_SIZE (OUTMODE) == GET_MODE_SIZE (INMODE) \
|| (GET_MODE_SIZE (OUTMODE) <= GET_MODE_SIZE (INMODE) \
- && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (OUTMODE), \
- GET_MODE_BITSIZE (INMODE))))
+ && TRULY_NOOP_TRUNCATION_MODES_P (OUTMODE, INMODE)))
-static void
+/* This function is called with INSN that sets REG to (SYM + OFF),
+ while REG is known to already have value (SYM + offset).
+ This function tries to change INSN into an add instruction
+ (set (REG) (plus (REG) (OFF - offset))) using the known value.
+ It also updates the information about REG's known value.
+ Return true if we made a change. */
+
+static bool
+move2add_use_add2_insn (rtx reg, rtx sym, rtx off, rtx insn)
+{
+ rtx pat = PATTERN (insn);
+ rtx src = SET_SRC (pat);
+ int regno = REGNO (reg);
+ rtx new_src = gen_int_mode (INTVAL (off) - reg_offset[regno],
+ GET_MODE (reg));
+ bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn));
+ bool changed = false;
+
+ /* (set (reg) (plus (reg) (const_int 0))) is not canonical;
+ use (set (reg) (reg)) instead.
+ We don't delete this insn, nor do we convert it into a
+ note, to avoid losing register notes or the return
+ value flag. jump2 already knows how to get rid of
+ no-op moves. */
+ if (new_src == const0_rtx)
+ {
+ /* If the constants are different, this is a
+ truncation, that, if turned into (set (reg)
+ (reg)), would be discarded. Maybe we should
+ try a truncMN pattern? */
+ if (INTVAL (off) == reg_offset [regno])
+ changed = validate_change (insn, &SET_SRC (pat), reg, 0);
+ }
+ else
+ {
+ struct full_rtx_costs oldcst, newcst;
+ rtx tem = gen_rtx_PLUS (GET_MODE (reg), reg, new_src);
+
+ get_full_set_rtx_cost (pat, &oldcst);
+ SET_SRC (pat) = tem;
+ get_full_set_rtx_cost (pat, &newcst);
+ SET_SRC (pat) = src;
+
+ if (costs_lt_p (&newcst, &oldcst, speed)
+ && have_add2_insn (reg, new_src))
+ changed = validate_change (insn, &SET_SRC (pat), tem, 0);
+ else if (sym == NULL_RTX && GET_MODE (reg) != BImode)
+ {
+ enum machine_mode narrow_mode;
+ for (narrow_mode = GET_CLASS_NARROWEST_MODE (MODE_INT);
+ narrow_mode != VOIDmode
+ && narrow_mode != GET_MODE (reg);
+ narrow_mode = GET_MODE_WIDER_MODE (narrow_mode))
+ {
+ if (have_insn_for (STRICT_LOW_PART, narrow_mode)
+ && ((reg_offset[regno] & ~GET_MODE_MASK (narrow_mode))
+ == (INTVAL (off) & ~GET_MODE_MASK (narrow_mode))))
+ {
+ rtx narrow_reg = gen_rtx_REG (narrow_mode,
+ REGNO (reg));
+ rtx narrow_src = gen_int_mode (INTVAL (off),
+ narrow_mode);
+ rtx new_set
+ = gen_rtx_SET (VOIDmode,
+ gen_rtx_STRICT_LOW_PART (VOIDmode,
+ narrow_reg),
+ narrow_src);
+ changed = validate_change (insn, &PATTERN (insn),
+ new_set, 0);
+ if (changed)
+ break;
+ }
+ }
+ }
+ }
+ reg_set_luid[regno] = move2add_luid;
+ reg_base_reg[regno] = -1;
+ reg_mode[regno] = GET_MODE (reg);
+ reg_symbol_ref[regno] = sym;
+ reg_offset[regno] = INTVAL (off);
+ return changed;
+}
+
+
+/* This function is called with INSN that sets REG to (SYM + OFF),
+ but REG doesn't have known value (SYM + offset). This function
+ tries to find another register which is known to already have
+ value (SYM + offset) and change INSN into an add instruction
+ (set (REG) (plus (the found register) (OFF - offset))) if such
+ a register is found. It also updates the information about
+ REG's known value.
+ Return true iff we made a change. */
+
+static bool
+move2add_use_add3_insn (rtx reg, rtx sym, rtx off, rtx insn)
+{
+ rtx pat = PATTERN (insn);
+ rtx src = SET_SRC (pat);
+ int regno = REGNO (reg);
+ int min_regno = 0;
+ bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn));
+ int i;
+ bool changed = false;
+ struct full_rtx_costs oldcst, newcst, mincst;
+ rtx plus_expr;
+
+ init_costs_to_max (&mincst);
+ get_full_set_rtx_cost (pat, &oldcst);
+
+ plus_expr = gen_rtx_PLUS (GET_MODE (reg), reg, const0_rtx);
+ SET_SRC (pat) = plus_expr;
+
+ for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
+ if (reg_set_luid[i] > move2add_last_label_luid
+ && reg_mode[i] == GET_MODE (reg)
+ && reg_base_reg[i] < 0
+ && reg_symbol_ref[i] != NULL_RTX
+ && rtx_equal_p (sym, reg_symbol_ref[i]))
+ {
+ rtx new_src = gen_int_mode (INTVAL (off) - reg_offset[i],
+ GET_MODE (reg));
+ /* (set (reg) (plus (reg) (const_int 0))) is not canonical;
+ use (set (reg) (reg)) instead.
+ We don't delete this insn, nor do we convert it into a
+ note, to avoid losing register notes or the return
+ value flag. jump2 already knows how to get rid of
+ no-op moves. */
+ if (new_src == const0_rtx)
+ {
+ init_costs_to_zero (&mincst);
+ min_regno = i;
+ break;
+ }
+ else
+ {
+ XEXP (plus_expr, 1) = new_src;
+ get_full_set_rtx_cost (pat, &newcst);
+
+ if (costs_lt_p (&newcst, &mincst, speed))
+ {
+ mincst = newcst;
+ min_regno = i;
+ }
+ }
+ }
+ SET_SRC (pat) = src;
+
+ if (costs_lt_p (&mincst, &oldcst, speed))
+ {
+ rtx tem;
+
+ tem = gen_rtx_REG (GET_MODE (reg), min_regno);
+ if (i != min_regno)
+ {
+ rtx new_src = gen_int_mode (INTVAL (off) - reg_offset[min_regno],
+ GET_MODE (reg));
+ tem = gen_rtx_PLUS (GET_MODE (reg), tem, new_src);
+ }
+ if (validate_change (insn, &SET_SRC (pat), tem, 0))
+ changed = true;
+ }
+ reg_set_luid[regno] = move2add_luid;
+ reg_base_reg[regno] = -1;
+ reg_mode[regno] = GET_MODE (reg);
+ reg_symbol_ref[regno] = sym;
+ reg_offset[regno] = INTVAL (off);
+ return changed;
+}
+
+/* Convert move insns with constant inputs to additions if they are cheaper.
+ Return true if any changes were made. */
+static bool
reload_cse_move2add (rtx first)
{
int i;
rtx insn;
+ bool changed = false;
for (i = FIRST_PSEUDO_REGISTER - 1; i >= 0; i--)
- reg_set_luid[i] = 0;
+ {
+ reg_set_luid[i] = 0;
+ reg_offset[i] = 0;
+ reg_base_reg[i] = 0;
+ reg_symbol_ref[i] = NULL_RTX;
+ reg_mode[i] = VOIDmode;
+ }
move2add_last_label_luid = 0;
move2add_luid = 2;
/* Check if we have valid information on the contents of this
register in the mode of REG. */
if (reg_set_luid[regno] > move2add_last_label_luid
- && MODES_OK_FOR_MOVE2ADD (GET_MODE (reg), reg_mode[regno]))
+ && MODES_OK_FOR_MOVE2ADD (GET_MODE (reg), reg_mode[regno])
+ && dbg_cnt (cse2_move2add))
{
/* Try to transform (set (REGX) (CONST_INT A))
...
(set (STRICT_LOW_PART (REGX)) (CONST_INT B))
*/
- if (GET_CODE (src) == CONST_INT && reg_base_reg[regno] < 0)
+ if (CONST_INT_P (src)
+ && reg_base_reg[regno] < 0
+ && reg_symbol_ref[regno] == NULL_RTX)
{
- rtx new_src = gen_int_mode (INTVAL (src) - reg_offset[regno],
- GET_MODE (reg));
- /* (set (reg) (plus (reg) (const_int 0))) is not canonical;
- use (set (reg) (reg)) instead.
- We don't delete this insn, nor do we convert it into a
- note, to avoid losing register notes or the return
- value flag. jump2 already knows how to get rid of
- no-op moves. */
- if (new_src == const0_rtx)
- {
- /* If the constants are different, this is a
- truncation, that, if turned into (set (reg)
- (reg)), would be discarded. Maybe we should
- try a truncMN pattern? */
- if (INTVAL (src) == reg_offset [regno])
- validate_change (insn, &SET_SRC (pat), reg, 0);
- }
- else if (rtx_cost (new_src, PLUS) < rtx_cost (src, SET)
- && have_add2_insn (reg, new_src))
- {
- rtx tem = gen_rtx_PLUS (GET_MODE (reg), reg, new_src);
- validate_change (insn, &SET_SRC (pat), tem, 0);
- }
- else if (GET_MODE (reg) != BImode)
- {
- enum machine_mode narrow_mode;
- for (narrow_mode = GET_CLASS_NARROWEST_MODE (MODE_INT);
- narrow_mode != VOIDmode
- && narrow_mode != GET_MODE (reg);
- narrow_mode = GET_MODE_WIDER_MODE (narrow_mode))
- {
- if (have_insn_for (STRICT_LOW_PART, narrow_mode)
- && ((reg_offset[regno]
- & ~GET_MODE_MASK (narrow_mode))
- == (INTVAL (src)
- & ~GET_MODE_MASK (narrow_mode))))
- {
- rtx narrow_reg = gen_rtx_REG (narrow_mode,
- REGNO (reg));
- rtx narrow_src = gen_int_mode (INTVAL (src),
- narrow_mode);
- rtx new_set =
- gen_rtx_SET (VOIDmode,
- gen_rtx_STRICT_LOW_PART (VOIDmode,
- narrow_reg),
- narrow_src);
- if (validate_change (insn, &PATTERN (insn),
- new_set, 0))
- break;
- }
- }
- }
- reg_set_luid[regno] = move2add_luid;
- reg_mode[regno] = GET_MODE (reg);
- reg_offset[regno] = INTVAL (src);
+ changed |= move2add_use_add2_insn (reg, NULL_RTX, src, insn);
continue;
}
&& MODES_OK_FOR_MOVE2ADD (GET_MODE (reg),
reg_mode[REGNO (src)]))
{
- rtx next = next_nonnote_insn (insn);
+ rtx next = next_nonnote_nondebug_insn (insn);
rtx set = NULL_RTX;
if (next)
set = single_set (next);
&& SET_DEST (set) == reg
&& GET_CODE (SET_SRC (set)) == PLUS
&& XEXP (SET_SRC (set), 0) == reg
- && GET_CODE (XEXP (SET_SRC (set), 1)) == CONST_INT)
+ && CONST_INT_P (XEXP (SET_SRC (set), 1)))
{
rtx src3 = XEXP (SET_SRC (set), 1);
HOST_WIDE_INT added_offset = INTVAL (src3);
+ base_offset
- regno_offset,
GET_MODE (reg));
- int success = 0;
+ bool success = false;
+ bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn));
if (new_src == const0_rtx)
/* See above why we create (set (reg) (reg)) here. */
success
= validate_change (next, &SET_SRC (set), reg, 0);
- else if ((rtx_cost (new_src, PLUS)
- < COSTS_N_INSNS (1) + rtx_cost (src3, SET))
- && have_add2_insn (reg, new_src))
+ else
{
- rtx newpat = gen_rtx_SET (VOIDmode,
- reg,
- gen_rtx_PLUS (GET_MODE (reg),
- reg,
- new_src));
- success
- = validate_change (next, &PATTERN (next),
- newpat, 0);
+ rtx old_src = SET_SRC (set);
+ struct full_rtx_costs oldcst, newcst;
+ rtx tem = gen_rtx_PLUS (GET_MODE (reg), reg, new_src);
+
+ get_full_set_rtx_cost (set, &oldcst);
+ SET_SRC (set) = tem;
+ get_full_set_src_cost (tem, &newcst);
+ SET_SRC (set) = old_src;
+ costs_add_n_insns (&oldcst, 1);
+
+ if (costs_lt_p (&newcst, &oldcst, speed)
+ && have_add2_insn (reg, new_src))
+ {
+ rtx newpat = gen_rtx_SET (VOIDmode, reg, tem);
+ success
+ = validate_change (next, &PATTERN (next),
+ newpat, 0);
+ }
}
if (success)
delete_insn (insn);
+ changed |= success;
insn = next;
reg_mode[regno] = GET_MODE (reg);
reg_offset[regno] =
}
}
}
+
+ /* Try to transform
+ (set (REGX) (CONST (PLUS (SYMBOL_REF) (CONST_INT A))))
+ ...
+ (set (REGY) (CONST (PLUS (SYMBOL_REF) (CONST_INT B))))
+ to
+ (set (REGX) (CONST (PLUS (SYMBOL_REF) (CONST_INT A))))
+ ...
+ (set (REGY) (CONST (PLUS (REGX) (CONST_INT B-A)))) */
+ if ((GET_CODE (src) == SYMBOL_REF
+ || (GET_CODE (src) == CONST
+ && GET_CODE (XEXP (src, 0)) == PLUS
+ && GET_CODE (XEXP (XEXP (src, 0), 0)) == SYMBOL_REF
+ && CONST_INT_P (XEXP (XEXP (src, 0), 1))))
+ && dbg_cnt (cse2_move2add))
+ {
+ rtx sym, off;
+
+ if (GET_CODE (src) == SYMBOL_REF)
+ {
+ sym = src;
+ off = const0_rtx;
+ }
+ else
+ {
+ sym = XEXP (XEXP (src, 0), 0);
+ off = XEXP (XEXP (src, 0), 1);
+ }
+
+ /* If the reg already contains the value which is sum of
+ sym and some constant value, we can use an add2 insn. */
+ if (reg_set_luid[regno] > move2add_last_label_luid
+ && MODES_OK_FOR_MOVE2ADD (GET_MODE (reg), reg_mode[regno])
+ && reg_base_reg[regno] < 0
+ && reg_symbol_ref[regno] != NULL_RTX
+ && rtx_equal_p (sym, reg_symbol_ref[regno]))
+ changed |= move2add_use_add2_insn (reg, sym, off, insn);
+
+ /* Otherwise, we have to find a register whose value is sum
+ of sym and some constant value. */
+ else
+ changed |= move2add_use_add3_insn (reg, sym, off, insn);
+
+ continue;
+ }
}
for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
reg_set_luid[regno] = 0;
}
}
- note_stores (PATTERN (insn), move2add_note_store, NULL);
+ note_stores (PATTERN (insn), move2add_note_store, insn);
/* If INSN is a conditional branch, we try to extract an
implicit set out of it. */
allocation if possible. */
&& SCALAR_INT_MODE_P (GET_MODE (XEXP (cnd, 0)))
&& hard_regno_nregs[REGNO (XEXP (cnd, 0))][GET_MODE (XEXP (cnd, 0))] == 1
- && GET_CODE (XEXP (cnd, 1)) == CONST_INT)
+ && CONST_INT_P (XEXP (cnd, 1)))
{
rtx implicit_set =
gen_rtx_SET (VOIDmode, XEXP (cnd, 0), XEXP (cnd, 1));
- move2add_note_store (SET_DEST (implicit_set), implicit_set, 0);
+ move2add_note_store (SET_DEST (implicit_set), implicit_set, insn);
}
}
}
}
}
+ return changed;
}
-/* SET is a SET or CLOBBER that sets DST.
+/* SET is a SET or CLOBBER that sets DST. DATA is the insn which
+ contains SET.
Update reg_set_luid, reg_offset and reg_base_reg accordingly.
Called from reload_cse_move2add via note_stores. */
static void
-move2add_note_store (rtx dst, rtx set, void *data ATTRIBUTE_UNUSED)
+move2add_note_store (rtx dst, const_rtx set, void *data)
{
+ rtx insn = (rtx) data;
unsigned int regno = 0;
+ unsigned int nregs = 0;
unsigned int i;
enum machine_mode mode = GET_MODE (dst);
GET_MODE (SUBREG_REG (dst)),
SUBREG_BYTE (dst),
GET_MODE (dst));
+ nregs = subreg_nregs (dst);
dst = SUBREG_REG (dst);
}
return;
regno += REGNO (dst);
+ if (!nregs)
+ nregs = hard_regno_nregs[regno][mode];
if (SCALAR_INT_MODE_P (GET_MODE (dst))
- && hard_regno_nregs[regno][mode] == 1 && GET_CODE (set) == SET
+ && nregs == 1 && GET_CODE (set) == SET)
+ {
+ rtx note, sym = NULL_RTX;
+ HOST_WIDE_INT off;
+
+ note = find_reg_equal_equiv_note (insn);
+ if (note && GET_CODE (XEXP (note, 0)) == SYMBOL_REF)
+ {
+ sym = XEXP (note, 0);
+ off = 0;
+ }
+ else if (note && GET_CODE (XEXP (note, 0)) == CONST
+ && GET_CODE (XEXP (XEXP (note, 0), 0)) == PLUS
+ && GET_CODE (XEXP (XEXP (XEXP (note, 0), 0), 0)) == SYMBOL_REF
+ && CONST_INT_P (XEXP (XEXP (XEXP (note, 0), 0), 1)))
+ {
+ sym = XEXP (XEXP (XEXP (note, 0), 0), 0);
+ off = INTVAL (XEXP (XEXP (XEXP (note, 0), 0), 1));
+ }
+
+ if (sym != NULL_RTX)
+ {
+ reg_base_reg[regno] = -1;
+ reg_symbol_ref[regno] = sym;
+ reg_offset[regno] = off;
+ reg_mode[regno] = mode;
+ reg_set_luid[regno] = move2add_luid;
+ return;
+ }
+ }
+
+ if (SCALAR_INT_MODE_P (GET_MODE (dst))
+ && nregs == 1 && GET_CODE (set) == SET
&& GET_CODE (SET_DEST (set)) != ZERO_EXTRACT
&& GET_CODE (SET_DEST (set)) != STRICT_LOW_PART)
{
{
base_reg = XEXP (src, 0);
- if (GET_CODE (XEXP (src, 1)) == CONST_INT)
+ if (CONST_INT_P (XEXP (src, 1)))
offset = INTVAL (XEXP (src, 1));
else if (REG_P (XEXP (src, 1))
&& (reg_set_luid[REGNO (XEXP (src, 1))]
&& (MODES_OK_FOR_MOVE2ADD
(dst_mode, reg_mode[REGNO (XEXP (src, 1))])))
{
- if (reg_base_reg[REGNO (XEXP (src, 1))] < 0)
+ if (reg_base_reg[REGNO (XEXP (src, 1))] < 0
+ && reg_symbol_ref[REGNO (XEXP (src, 1))] == NULL_RTX)
offset = reg_offset[REGNO (XEXP (src, 1))];
/* Maybe the first register is known to be a
constant. */
else if (reg_set_luid[REGNO (base_reg)]
> move2add_last_label_luid
&& (MODES_OK_FOR_MOVE2ADD
- (dst_mode, reg_mode[REGNO (XEXP (src, 1))]))
- && reg_base_reg[REGNO (base_reg)] < 0)
+ (dst_mode, reg_mode[REGNO (base_reg)]))
+ && reg_base_reg[REGNO (base_reg)] < 0
+ && reg_symbol_ref[REGNO (base_reg)] == NULL_RTX)
{
offset = reg_offset[REGNO (base_reg)];
base_reg = XEXP (src, 1);
case CONST_INT:
/* Start tracking the register as a constant. */
reg_base_reg[regno] = -1;
+ reg_symbol_ref[regno] = NULL_RTX;
reg_offset[regno] = INTVAL (SET_SRC (set));
/* We assign the same luid to all registers set to constants. */
reg_set_luid[regno] = move2add_last_label_luid + 1;
if (reg_set_luid[base_regno] <= move2add_last_label_luid)
{
reg_base_reg[base_regno] = base_regno;
+ reg_symbol_ref[base_regno] = NULL_RTX;
reg_offset[base_regno] = 0;
reg_set_luid[base_regno] = move2add_luid;
reg_mode[base_regno] = mode;
/* Copy base information from our base register. */
reg_set_luid[regno] = reg_set_luid[base_regno];
reg_base_reg[regno] = reg_base_reg[base_regno];
+ reg_symbol_ref[regno] = reg_symbol_ref[base_regno];
/* Compute the sum of the offsets or constants. */
reg_offset[regno] = trunc_int_for_mode (offset
}
else
{
- unsigned int endregno = regno + hard_regno_nregs[regno][mode];
+ unsigned int endregno = regno + nregs;
for (i = regno; i < endregno; i++)
/* Reset the information about this register. */
static bool
gate_handle_postreload (void)
{
- return (optimize > 0);
+ return (optimize > 0 && reload_completed);
}
static unsigned int
rest_of_handle_postreload (void)
{
+ if (!dbg_cnt (postreload_cse))
+ return 0;
+
/* Do a very simple CSE pass over just the hard registers. */
reload_cse_regs (get_insns ());
- /* reload_cse_regs can eliminate potentially-trapping MEMs.
+ /* Reload_cse_regs can eliminate potentially-trapping MEMs.
Remove any EH edges associated with them. */
- if (flag_non_call_exceptions)
+ if (cfun->can_throw_non_call_exceptions)
purge_all_dead_edges ();
+
return 0;
}
-struct tree_opt_pass pass_postreload_cse =
+struct rtl_opt_pass pass_postreload_cse =
{
+ {
+ RTL_PASS,
"postreload", /* name */
gate_handle_postreload, /* gate */
rest_of_handle_postreload, /* execute */
0, /* properties_provided */
0, /* properties_destroyed */
0, /* todo_flags_start */
- TODO_dump_func, /* todo_flags_finish */
- 'o' /* letter */
+ TODO_df_finish | TODO_verify_rtl_sharing |
+ 0 /* todo_flags_finish */
+ }
};
-
OSDN Git Service
